Columbia Economics Review: Fall 2019 by Columbia Economics Review

COLUMBIA ECONOMIC REVIEW F FACULTY ADVISOR Wouter Vergote, Ph.D.

EDITORIAL BOARD Editor-in-Chief & President: Mathieu K. Sabbagh Managing Editor (Journal): Neel Puri Deputy Managing Editor (Journal): Sinet Chelagat Deputy Managing Editor (Journal): Ignacio Lopez Geffney Executive Editor (Online): Hallie Gruder Junior Editor (Online): Devyani Goel Junior Editor (Online): Philip Jang Executive Director (Operations): Jenna Karp Publisher: Gabriel Multedo Art Director: Jessica Lu

STAFF EDITORS

OPERATIONS TEAM

Michael Chu Sinet Chelagat Edmund Shen Ariunzaya Oktyabri Jeremy Liu Jeffrey Luo Bennett Smith-Worthington Nurasyl Shรณkeyev Bennett Bookstein Charlie Munns Ignacio Lopez Gaffney Blake Jones Noah Love Akshat Singh

Jennifer Lee Nathan Chen Reyna Yerim Choi Vidhima Shetty Surina Goel Sybil Fu Samantha Syme Cara Vo Cindy Gao Donna Qi Phoenix Chen Kristina Hadjipetkov Samantha Hochstat Lillian Zhang

PUBLISHING STAFF

ONLINE STAFF

Genevieve Spencer Naina Durga Lavakare

Joseph Cambpell Lenny Ciotti Shreya Ganguly Linh Hoang Nicolas Hortiguera Alqaim Lalani Andrew Liu Casey Li Leon Lu Eshita Sangal Serdil Yalcin

Cover Art: Nicholas A. DiCostanzo

Cover Design: Gabriel Multedo

Welcome to the eleventh year of the Columbia Economic Review (“CER”), the leading undergraduate student-run economics journal in the country. The CER continues to see growth in readers and papers, all while the quality of submissions remains high. This fall/winter edition presents a nice balance of articles on a range of important topics—from domestic credit in emerging markets to bankruptcy and health care. The articles were selected through a peer review process involving undergraduate students. Each article was then reviewed by my colleague and the journal’s faculty advisor, Wouter Vergote. Taken together, the articles published tell a story about our department’s commitment to training students to carefully analyze economic and social issues, preparing them for a host of careers. In closing, let me thank you for being a member of our vibrant research community. I invite you to visit our Department website to learn more about our programs and the ways we are creating new opportunities for learning. With best wishes, Miguel Urquiola Chair, Department of Economics Columbia University

It is a pleasure to review the papers in this issue of the Columbia Economic Review (“CER”). The research represents a wide range of perspectives, policy concerns, and highlight the significant contributions of undergraduate researchers in the field. The Program for Economic Research (“PER”) is happy to support the journal. PER works with undergraduates across Columbia University to promote discourse and research at the intersection of economics, business, politics, and society. This issue presents us with five papers by current undergraduate students or recent graduates in economics or areas studies, and include: •

Okel, Caroline, “Exchange Rate Regimes, Capital Inﬂows, and Domestic Credit in Emerging Market Economics”

•

Dickens, Jessie, “Changing Traditions: Modeling the Probability of Bidding to Host the Olympic Games Over Time”

•

Barkay, Shaked, “Determinants of Health Care Utilization in Germany: A Latent Class Analysis

•

Zwiefel, Noah, “What is the Impact of a Medicaid Expansion on the Bankruptcy Rate? Evidence from the ACA”

•

Caldwell, Marissa, “The Heuristics of Obesity: Inﬂuences on Physician Decision-Making”

A special thanks to the journal’s faculty advisor Wouter Vergote. Harrison Hong John R. Eckel, Jr. Professor of Financial Economics, Columbia University Executive Director, Program for Economic Research

Dear Readers, It is with great pleasure that we present to you the latest biannual journal of the Columbia Economic Review. We are delighted to report continued successes in our goal to increase the discipline’s visibility among fellow undergraduates on campus and beyond. Indeed, with a historic decade soon coming to an end, it is important to reflect upon the prevailing economic topics that have dominated our world. This semester’s edition of our journal attempts to do just that. Though implemented right before the start of the 2010s, the Affordable Care Act (ACA) and its various modifications over later legislative terms can be seen as a flashpoint for the ideological divide which currently grips the United States. More generally, the matter of healthcare can be seen as a litmus test for political-economic orientations so often pinned against one another by aspiring policy makers. Three papers this semester fall within this broad category. Shaked Barkay explores healthcare utilization through the lens of class in her piece, exploring whether the expected number of physician visits differs among socio-economic classes using survey data from Germany. Notable is her discussion of the effect of private coverage in creating heterogenous visit rates among the general populace. Marissa Caldwell discusses the effects of obesity status on medical decisions. In particular, she explores how diagnosis rates for coronary disease, diabetes, and high blood pressure differ on either side of a BMI cutoff. Noah Zwiefel assesses the impact of the ACA on household finances across the country. In particular, he estimates the number of bankruptcies which were avoided as a result of Medicaid expansion. Rounding out this semester’s edition are two impressive papers respectively covering monetary policy and sport economics. Caroline Okel offers a study of capital inflow on credit provision within countries. In particular, she impressively models the risk of credit bubbles developing in fixed exchange-rate regimes and in floating exchange-rate regimes. Jessie Dickens models Olympic Games bid likelihood, using a novel model that focuses on the relative contributions of economic, geopolitical, and environmental country-specific factors. We would be remiss if we did not take the time to acknowledge the great members of our Columbia community who have actively supported CER over

the past semester. In particular, we would like to thank Dr. Sophia N. Johnson, Assistant Director of the Program for Economic Research (PER) which provides most of our funding, and Prof. Wouter Vergote, who is our long-standing faculty advisor and a source of invaluable guidance. It is thanks to the exceptional support of the Columbia Department of Economics and PER that we are able to continue fulďŹ lling our mission of promoting engagement with undergraduate economics research. We hope that the issue you now hold in your hands will live up to the standards of excellence that deďŹ ne our university. Warm regards, Neel Puri, Managing Editor Mathieu K. Sabbagh, Editor-in-Chief and President

Letter From the Chair of the Department of Economics Letter From the Program for Economic Research Letter From the Editors

Determinants Of Health Care Utilization In Germany: A Latent Class Analysis Shaked Barkay

What is the Impact of a Medicaid Expansion on the Bankruptcy Rate? Evidence from the ACA Noah Zwiefel

The Heuristics of Obesity: InďŹ&#x201A;uences on Physician Decision-Making Marissa Caldwell

133

Changing Traditions: Modeling the Probability of Bidding to Host the Olympic Games Over Time Jessie Dickens

171

Exchange Rate Regimes, Capital InďŹ&#x201A;ows, and Domestic Credit in Emerging Market Economies Elena M. Stacy

DETERMINANTS OF HEALTH CARE UTILIZATION IN GERMANY: A LATENT CLASS ANALYSIS

University of Pennsylvania Abstract: This paper models consumption of health care, as measured by physician visits and hospitalizations, using data from the German Socioeconomic Panel (SOEP) for the years 2006-2015. A latent class framework is used to account for individual unobserved heterogeneity. Two latent class models are used within each utilization type: a binary model predicting if a doctor visit/hospitalization occurred and a latent class aggregated count model predicting the count of physician visits/ hospitalizations. A special emphasis is placed on the role of private insurance in determining health care utilization. In particular, the paper examines the impact of moral hazard on health care utilization by attempting to disentangle its effects from those of adverse selection. The results show signiﬁcant evidence of heterogeneity across classes. Additionally, the holding of private insurance is found to be insigniﬁcant in determining health care utilization.

Introduction he rise in health care costs in the developed world has emerged as one of the foremost problems with which governments will need to contend over the coming decades. A 2015 Organization for Economic Co-operation and Development (OECD) study found that health spending has risen faster than economic growth across all member countries. It showed that the percentage of GDP allocated toward health care spending will rise by almost 50% in the next 15 years and will more than double over the next few decades. These rising costs will lead to increasingly strained government budgets across member countries and a higher tax burden on their citizens. While these increases reflect both medical inflation and increased consumption, medical inflation accounts for a smaller share of the health care expenditure growth (with the possible exception of the United States). As such, attaining a clearer understanding of the factors that impact consumption of health care is an essential avenue of research. Moreover, it is important to distinguish between the usage of different types of care which can have different short and long-term budgetary implications. In this study, a latent class analysis was used to examine both the observed and latent determinants of Germany’s health care consumption from

Columbia Economic Review | 15

Shaked Barkay

the years 2006 and 2008 to 2015 using longitudinal data from the SOEP. The number of hospitalizations as well as doctor visits was measured, allowing for a more nuanced picture of utilization patterns across different types of care. Finally, developing previous research on the association of various demographic factors and the consumption of health care, the latent class analysis presents the different interactions of these factors within latent groups. Latent class analysis identifies unobserved discrete or “latent” class membership in the underlying population; these classes are heterogeneous and react differently to the covariates. As such, a latent class regression provides a more accurate model of the relationship between the dependent and independent variables, where the underlying population is divided into discrete and heterogeneous classes. A particular emphasis is placed on the effect of private insurance on consumption. This provides an opportunity to study the effects of moral hazard alone by including a wide range of controls for adverse selection. These controls would pick up adverse selection that is due to unhealthy, old or less risk-averse individuals purchasing private insurance at higher rates, so that the effect of the private insurance variable would be largely confined to moral hazard. However, it cannot be ruled out that some heterogeneity between those with public and private insurance is unaccounted for. Literature Review The rapid development of healthcare technology, growth of income, expansion of health insurance and increased longevity have all led to higher levels of health care consumption and consumption of more expensive types of care. To better understand this phenomenon, there has been a proliferation of research on the demand and supply factors that determine the degree of utilization. Research has focused broadly on examining the influence of the following categories: I. II. III. IV.

Personal factors (demographics) Level of health insurance coverage Availability of health care Latent individual characteristics

The latter in particular has been the focus of recent research, aided by the relatively recent introduction of latent class analysis into the health economics literature. This has led to a richer understanding of the interplay between the three preceding categories and to a more nuanced portrait of what factors can impact the degree of healthcare consumption. The following is a summary of the research pertinent to my paper found within and across these four categories.

Determinants of health care utilization in Germany

I) Personal factors: Personal factors cover a wide category which includes gender, race, marital status, income, level of insurance and more. Recently, immigration and refugee status have also been examined, with a particular interest in the growing effects of refugee waves on western European health care consumption. Research on the interplay of gender and health care consumption has found that women are overwhelmingly more likely to use more health care services even when correcting for services specific to women such as gynecology and obstetrics (Briscoe, 1987; Corney, 1990). However, more recent research has shown that some of these differences may stem from higher reported somatic morbidity and mental distress (Koopmans & Lamers, 2007). Moreover, not all consumption indicators are similarly influenced by gender (Glaesmer et. al, 2012). Glaesmer et. al have found evidence for a consumption propensity gap across the genders so that women have a higher propensity for utilization. To account not only for the difference in the consumption of males and females, but also the interaction of sex with other factors, this paper conducts the analysis of men and women separately. Evidence on the effect of race on health care consumption is far more mixed and may be much more prone to geographic variation. In the US, there is some evidence that African-Americans and Latinos are less likely to utilize certain types of health care even when controlling for level of insurance and income (Cooper & Ford, 1995). Research has found that while hospitalization (when controlling for level of insurance and income) stays were not significantly correlated with race, physician visits were (Johnson-Lans & Bellemore, 1997). Yet, a limitation of these findings is that they don’t consider the impacts of the higher rate of uninsurance within African-American communities, which may extend to African-Americans with insurance. This may occur if parents without insurance impact the health care behavior of their children, even those who go on to purchase insurance. However, in the United Kingdom (UK) where health care provision is governmental and thus no such confounding variables exist, belonging to an ethnic minority is associated with a higher number of general practitioner visits along with an increased likelihood that one has visited a general practitioner over the previous year. This includes Black British residents in direct contrast to the results in the US (Bago d’Uva, 2005). The American result is especially surprising given that African-Americans have, on average, more physical ailments (Johnson-Lans & Bellemore 1997). Furthermore, the studies on the consumption patterns of immigrants have also been mixed. Weinstein et al. found that adult refugees within the US have utilization rates below the average rate of the population (2000). However, other studies such as that of McMahon et al. in Ireland have found that refugees have higher utilization than the general populace (2007).

Columbia Economic Review | 17

Shaked Barkay

Education is often a confounding variable in measuring the between native and immigrant utilization, as it has been found to overwhelmingly correlate with a higher utilization of health care (Bago d’Uva, 2005; Olusanya et al., 2016, Ali & Elsayed, 2018). Income also consistently correlates with higher levels of consumption, even when including countries whose government is the primary health care provider and have a “zero price” at point of utilization (Bago d’Uva, 2005; Curtis & MacMinn 2008). However, this is not consistent across all types of consumption; the gaps are particularly prominent in visits to specialists and general practitioners and less so with hospitalizations (Curtis & MacMinn 2008). This differentiation motivates the inclusion of four metrics of utilization in my paper: binary doctor visits1, number of doctor visits, binary hospitalization2 and number of hospitalizations. These metrics capture the different trends that emerge across different types of consumption. There are also consequential differences between types of utilization, for example, a low health capital stock in early life due to infrequent visits to physicians and/or specialists may lead to more frequent utilization later in life. Thus, it is important to differentiate between types of health care. Research has found that working long hours is a barrier to consuming more health care (Fell et al., 2007). Similarly, unemployment is associated with higher degrees of utilization (Åhs et al., 2012), though it is unclear if this due to the greater morbidity associated with unemployment. Another factor that has a significant influence on the degree of consumption is marital status. Research has found that being divorced or widowed is associated with increased consumption of hospital care (Joung & Mackenback, 1995). However, married individuals of both genders are more likely to visit general practitioners (Bago d’Uva, 2005), even when including a full set of demographic and health status controls. The results diverge on types of health care use and stress the importance of including more than one measure of utilization of health care.Age is widely known to be the predominant determinant of health care spending, and the growth of spending among the oldest members of society has furthered this phenomenon (Wong et al., 2012). This is due to both a higher life expectancy leading to a higher amount of consumption and the proliferation of expensive medical technology especially geared to the elderly which has led to more expensive care. By measuring only metrics of utilization rather than cost, this study avoids the effects of expensive new technologies along with medical inflation that can skew comparisons across time. Finally, a variable that has not been widely studied but has been suggested as a possible explanation for the difference in consumption between men and women and across races is risk aversion (Rosen et al., 2003). For this reason, this paper includes a variable for self-reported willingness to take risk as a measure of risk aversion. 18

Determinants of health care utilization in Germany

II) Level of Insurance Coverage The effect of health insurance and the comprehensiveness of insurance coverage on the level of consumption is perhaps the most often researched question in health economics. That those with insurance would utilize higher quantities of health care is supported by both the theories of moral hazard and adverse selection. Moral hazard refers to the decreased need to take preventative action against a certain outcome when one is insured for that outcome. In terms of health, it refers to the decreased precautions one takes in order to protect his/her health (ex-ante moral hazard) and a decreased incentive to “bargain” once one utilizes healthcare (ex-post moral hazard) (Rowell & Connelly, 2012). Adverse selection, as famously described in George Akerlof ’s 1970 paper on the market for used cars, is the phenomenon where under asymmetric information only “lemons”3 would be left in the market. In the case of health insurance, it means only the sickly would be insured since the healthy would be driven out given that their expected health care costs are lower than the market premium, which results from the pooling of risk types (Akerlof, 1970). The evidence demonstrates the effect of insurance on healthcare consumption is, as theorized, a positive one, with the expansion or introduction of insurance leading to greater consumption (Card et al., 2008; Boyoung & Soonman, 2013; Wagstaff & Lindelow 2008). However, the magnitude is dependent on the type of healthcare utilized: more elective procedures, visits to specialists and outpatient care are all more strongly affected than inpatient care (Card et al., 2008; Boyoung & Soonman, 2013). This may occur because dire procedures are price inelastic, or close to it, as well as because higher earners, who purchase more comprehensive insurance, would have a greater opportunity cost to inpatient care, which is typically time consuming (Xin & Wei, 2013). Additionally, with the exception of Card et al. which studied seniors automatically enrolled in Medicare and thus who did not suffer from adverse selection, these studies do not enable a clear differentiation of moral hazard from adverse selection. Even Card’s study is limited in that it does not take into account the pre-enrollment deferment of healthcare seniors may opt for when waiting until they turn 65 and are eligible for Medicare. In Germany, there is universal health insurance through publicly funded sickness funds with an option for higher earners to purchase private insurance plans. The use of latent class analysis and a large set of controls allow for a better separation of health care consumption due to adverse selection and consumption due to moral hazard among those with private insurance. The effect picked up by the private insurance variable is almost entirely due to moral hazard.

Columbia Economic Review | 19

Shaked Barkay

III) Health Care Availability A large body of research has demonstrated that the proximity of an individual to health care providers increases the consumption of health care services (Bosanac et al., 1976; Shanon et al., 1979). Similarly, those living in dense population centers visit general practitioners and specialists more frequently (though this is not true of hospital stays) (McDonald & Conde, 2010). The limitations of this line of thinking are revealed when one considers reverse causality: it may be that health-conscious individuals are purposefully choosing to live in urban centers in order to utilize more healthcare. A latent class analysis wherein individuals are divided by use types mitigates this problem, as under this framework geographic differences are less likely to result from underlying differences in the use tendencies of the populations. IV) Latent Characteristics Factor analysis and particularly latent class analysis have been used in recent research to isolate latent variables that impact health care utilization. Bago d’Uva (2005) uses a latent class regression for both a binary variable 4 and a count variable5 of visits to a general practitioner among UK residents. In the UK, general practitioner visits is an adequate measure of utilization given that all other types of utilization must ﬁrst be approved by the general practitioner (GP). However, it still leaves out important distinctions between the types of care used. In contrast, Liu et al. (2012) used three different measures of utilization: ambulatory care, overall cost and inpatient care, thus providing a more nuanced description of health care utilization. For Germany’s health care system, where there are private providers and no required gatekeeping, distinguishing between types of utilization is especially important. Thus, in this study I include two different measures of utilization, doctor visits and hospitalizations. In addition, I make the same inclusion of binary and count variables as in Bago d’Uva (2005). Latent Class Model While traditional regression models only describe relationships between observed variables, latent class models (or ﬁnite mixture models) include one or more unobserved or latent variables. A latent class regression predicts the dependent variable as a function of the covariates, and also includes a K-category latent variable, where each category represents a homogenous population. Different regressions are run for each segment of the population that falls into a different latent class. By allowing the parameters of the model to vary across classes, a latent class regression does not rely on the assumption of homogeneity across a population, and can thus provide a far more accurate and nuanced description of the data. Along with not assuming homogeneity across the individuals in the dataset, latent class models do not assume linear 20

Determinants of health care utilization in Germany

relationships nor a normal distribution in the underlying population, as is assumed in simple regression models. As these properties are often violated in the real data, latent class models are less susceptible to biases arising from the data not conforming to the model assumptions. The latent class model used in this study is based on that of Bago d’Uva (2005) and is modiﬁed in the case of the count model to account for the slightly different speciﬁcations made. In the model, each individual is numbered i = 1 ,2, …, 82190 (for female, for male it would be i =1, 2, …,72512) and are each observed Ti times where Ti =1, 2, …,9 for all i. The dependent variable (either the binary or categorical variable for doctor visits or hospital admissions) is called yit, such that the observations of an individual over the entire panel are yit= (yi1, yi2, … , yiT). In addition, each individual has a vector of covariates xit that includes a constant. Further, let each individual be a member of latent class j, where j = 1, 2, …, C. Individuals are heterogenous across classes, meaning they differ with respect to the relative importance placed on each of the covariates and in the intercept but are homogenous within each class. Given the class an individual belongs to, the dependent variable has a probability density function of fj (yit |xit , θj ), where θj is the set of parameters of the model of class j. Based on the preceding, the joint density function of the dependent variable across the observed t time periods is ∏Ti fj (yit |xit , θj ). Furthermore, the probability of belonging to class j is equal to π ij such that 0 < πij < 1 and ∑Cj=1 πij = 1. Therefore, the unconditional density function of y is equal to:

Following the Bago d’Uva (2005) the set of θj = (θ1, … , θC ) and j = (1, 2, …, C) are parameterized as functions of the time invariant individual characteristics zi. Therefore, class membership can be estimated using a multinomial logit function as follows:

Further, let zi = . Since the zi’s are the determinants of class membership, this specification allows for the observed regressors to be correlated with individual effects. Once the vectors of the parameters θj and j are estimated jointly using maximum-likelihood, it is possible to find the posterior probability that an individual belongs to each class.

Columbia Economic Review | 21

Shaked Barkay

The posterior probability of membership in class j is given by:

Based on this, the individual should be assigned to the class for which he/she has the highest posterior probability. This model offers a great deal of ﬂexibility by assuming no distribution for the latent individual effects and allowing for correlation between the latent heterogeneity and the covariates, as well as both intercept and slope heterogeneity. Binary Regression Model In the binary case, in which an individual has either visited a hospital/doctor over the past twelve months or not, yit can either take the values 0 or 1. Using a logit model to estimate this:

where βj is the set of coefﬁcients of the covariates for class j. The conditional joint distribution of yi is:

and the unconditional joint distribution of yi is:

Count Regression Model Since a latent class regression requires categorical variables, the number of hospital stays and doctor visits must be aggregated so that yit is a categorical variable. Letting yit * be the true number of doctor/hospital visits, yit is deﬁned for hospital visits as:

Determinants of health care utilization in Germany

For doctor visits, I deﬁned yit as:

Bago d’Uva (2005) assumes the underlying distribution of yit * is a negative binomial distribution. In contrast, this study assumes a Poisson distribution, the method used in the XLSTAT Latent package, which is nested in the negative binomial distribution. The limitation of this is that the standard errors become less accurate in the case of overdispersion. Because a Poisson distribution has only one free parameter, variance is completely determined by the mean. The conditional distribution of yit * on j classes is:

This is the same as the negative binomial distribution:

with αj = 0.

and θj = (αj , βj ), where k = 1 and

The distribution for yit in the case of hospital visits conditional on j is obtained by:

The distribution for yit in the case of doctor visits conditional on j is obtained by:

Columbia Economic Review | 23

Shaked Barkay

The conditional joint distribution for yt is equal to: and the unconditional joint distribution is given by: Specifying the model as such allows for the estimation of the true number of visits. Model Estimation Estimation was done using the XLSTAT-LG package. A drawback of latent class models is that there is no guarantee the solution is the maximum likelihood solution. XLSTAT-LG employs either the ExpectationMaximization or Newton Raphson algorithm, which may converge to a local rather than the global maximum likelihood solution. However, it starts the algorithm at randomized starting values with multiple iterations to increase the probability of converging to the global solution. The method used in this study is the EM algorithm with 650 iterations. Data The data used in this study is from the SOEP during the years 2006-2015, except for the year 2007 which did not include a question about personal willingness to take risk, one of the variables included in the regression. The SOEP is a national longitudinal survey conducted annually of a sample of German households and all household members aged 17 or older. It is a rich dataset with information that includes demographic, economic, occupational, education and health variables. Table 1 lists all the variables included in the paper and provides their description and their mean values for each sex. There are four different dependent variables used, two of which are count variables indicating the number of hospital visits and doctor visits an individual has had over the last twelve months. However, the count values do not reďŹ&#x201A;ect the number of times visited but one of six categories of visitation, which are described in the Latent Class Model section. Secondly, there are two binary variables indicating whether or not one has been hospitalized/visited a doctor over the past year. The covariates can be divided into four parts. The ďŹ rst of these are the controls for time, a dummy variable for each year included in the survey. The second are geographic controls, with a dummy variable for each state in Germany and another dummy variable for whether or not one resides in East Germany. These attempt to control for supply-side factors related to health care such as the number and quality of the doctors and hospitals wherein one resides. 24

Determinants of health care utilization in Germany

The third category is the health status controls. These include a variable for age, personal health satisfaction on a 1-10 scale and one’s self reported health status with the options being: very good, good, acceptable, less good, and bad. The category also includes two dummy variables for individuals who are mildly handicapped (at 0-50% handicap status, as determined by the German government) and for those who are more severely handicapped (at 51-100% handicap status). Finally, the fourth category is of other personal characteristics found significant in the literature. These include variables for personal willingness to take risk (on a 1-10 scale), the natural log of household income, number of hours worked over the last year, whether or not one is employed, number of years of education, marital status, if one has a foreign nationality, and if one has private insurance. The latter, private insurance, is the primary focus of this study, which seeks to ascertain whether there is evidence for moral hazard when including many controls for both supply side factors, health status and a set of personal characteristics. The means are separate for males and females to capture the differences in utilization across the sexes which is well noted in the literature. As seen in Table 1, females utilize more of every metric of care included. Other notable differences are that females work significantly fewer hours. However, because household rather than personal income was included, the income of males is higher but not dramatically so. Males are also more willing to take risks, far more likely to hold private insurance and are more likely to be handicapped but are less likely to rate their health as bad. Thus, there is significant heterogeneity between the sexes that is best captured by running a separate analysis for males and females, as is done in the subsequent analysis.

Columbia Economic Review | 25

Shaked Barkay

Table 1. Variable Description and Mean Values

Determinants of health care utilization in Germany

Results In total eight regressions were run: 1. Female Hospitalization (Binary) 2. Male Hospitalization (Binary) 3. Female Doctor Visit (Binary) 4. Male Doctor Visit (Binary) 5. Female Hospitalization (Count) 6. Male Hospitalization (Count) 7. Female Doctor Visit (Count) 8. Male Doctor Visit (Count) I selected the number of latent classes using the Schwarz Criterion (BIC), testing two to four classes for each regression. In some of the regressions, the heterogeneity across classes reﬂects differences in overall consumption. For example, in the ‘Male Hospitalization Binary’ regression, the mean probability of hospitalization in the ﬁrst latent class is much greater than that of the second (0.224 and 0.029). In other regressions, the heterogeneity Columbia Economic Review | 27

Shaked Barkay

across classes is due to differences in reaction to the covariates. For regressions in which heterogeneity is due to differences in consumption, classes are labeled high or low use. Otherwise, classes are just numbered. In addition to estimating the coefficients of the covariates, I conducted two Wald’s tests. The first measured whether a variable is significant assuming homogeneity across classes. The second measured whether the differences in the coefficients across classes is significant. Regression #1: Female Hospitalization (Binary) Table 2. Female Hospitalization Binary Regression Results

Determinants of health care utilization in Germany

For the ‘Binary Male Hospitalization’ regression, two separate classes minimized the Schwarz Criterion. The percentage hospitalized across the classes is almost identical, as can be seen in Table 3. Table 3. Estimated Percentage Hospitalized by Class

As expected, consumption of healthcare for class 1 generally increases over time, with no significant variation from year to year besides this. For class 2 there is no significant effect of time, except for the year 2013. With respect to the effects of geography – the control for supply side factors – though some states are significantly correlated with an increased likelihood of hospitalization, there are no clear patterns across both classes. Additionally, residency in East Germany is not significant in either class beyond what was picked up by the state dummies.

Columbia Economic Review | 29

Shaked Barkay

The results of the controls for health status are more interesting. In class 1, the rate of hospitalization increases with age and degree of handicap and decreases with health satisfaction and better reported health. In class 2 both handicap and current health are insignificant while the pattern for age and health satisfaction is reversed, with hospitalizations negatively correlated with age and positively correlated with health satisfaction. A possible explanation for this is that class 2 has a disproportionately large number of pregnant women or women who were recently pregnant. This would explain why younger women are more likely to be hospitalized, and why other measures of health (current health and handicap status) are not significant. This hypothesis also explains why unmarried women in this class are significantly less likely to be hospitalized (the inverse of the relationship in class 1). Lack of employment is significant and positive for both classes. Working more hours had a significant negative effect in class 1 but did not have a significant effect in class 2. In contrast, education and income had a positive impact in class 2 but no impact on class 1. Willingness to take risk had a significant positive effect on class 1 and a significant negative effect on class 2. Finally, the variable of primary interest, private insurance, is insignificant for both classes. Regression #2: Male Hospitalization (Binary) Table 4. Male Hospitalization Binary Regression Results

Determinants of health care utilization in Germany

In this latent class regression, the classes represent different levels of consumption, where class 1 (labeled ‘High Utilizers’) has a higher mean probability of hospitalization than class two (labeled ‘Low Utilizers’) (see Table 5).

Columbia Economic Review | 31

Shaked Barkay

Table 5. Estimated Percentage Hospitalized by Class

In this regression, time and geography are not significant. While certain years and states are associated with increased hospitalization, these results are not consistent across classes. The controls for health status generally conform to expectations (in contrast to the female hospitalization regression). Health satisfaction has a significant and negative effect for both classes and degree of handicap has a significant positive effect across both classes. Moreover, reported worse current health is generally correlated with a higher likelihood of hospitalization, the exception being the “Bad” category in the low utilization class. This may be due to a lack of individuals in the class. The only measure that showed unexpected results is age in the low use class, having a negative effect on hospitalization (as opposed to a positive effect as in the high use class). This may result from the other health controls fully accounting for differences in health status. Both employment and marital status have an insignificant effect on hospitalization. Working more hours has a significant negative effect in the high utilization class but no significant effect in the low utilization class. Education and income on the other hand, have a negative impact on the low utilization class and no impact on the high utilization class. Willingness to take risk has a significant positive effect on both classes. In this case foreign nationality has a negative effect in the high utilization class while it has a positive effect in the low utilization class. This may reflect different underlying populations, heterogenous in their rate of hospital care consumption. Finally, private insurance is insignificant across both classes in predicting whether or not one is hospitalized.

Determinants of health care utilization in Germany

Regression #3: Female Doctor Visits (Binary) Table 6. Female Doctor Visits Binary Regression Results

Columbia Economic Review | 33

Shaked Barkay

In this regression, the classes are separated by the magnitude of consumption in health care services (see Table 7). Table 7. Estimated Percentage Visited Doctor by Class

Consequently, most of the variation in the propensity to visit a doctor is due to latent factors rather than covariates. This latent factor is presumably whether or not an individual had become sick. If this had been a one-time occurrence (as opposed to chronic) it would not be picked up by measures of health. Nevertheless, there are many signiďŹ cant covariates within both classes. As is the case in the hospitalization regression, the effect of the time variables shows no discernable pattern. For the geographic controls, it seems that most of the states are signiďŹ cant in the low utilization class but only Berlin is for the high utilization class. 34

Determinants of health care utilization in Germany

The results for the controls for health status conform to expectations. Worse current health and degree of handicap are both associated with a higher likelihood of doctor visitation across both classes. Age is only significant in the low utilization class where it has positive coefficient. Health satisfaction is only significant in the high utilization class, where as expected it has a negative effect. Foreign nationality and income are insignificant across both classes. Both lack of employment and education have a significant positive correlation with doctor visitation in the high consumption class. Hours worked has a significant negative effect on doctor visitation in the low use class, while it is insignificant in the high use class. Willingness to take risk is significant in both classes, but it has a positive correlation in the high consumption class and a negative correlation in the low use class. This may be due to the fact that lower risk aversion increases the likelihood of that one would engage in activities that necessitate doctor visits, but also decreases the likelihood that one would visit a doctor for any given health state. Interestingly, private insurance has a significant negative effect on doctor visitation in the high utilization category. While difficult to explain theoretically, this may simply reflect heterogeneity between private and public insurance holders that is not captured by the other covariates or latent classes. Regression #4: Male Doctor Visits (Binary) Table 8. Male Doctor Visits Binary Regression Results

Columbia Economic Review | 35

Shaked Barkay

Once again, the classes represent levels of use, with class 2 being composed almost entirely of individuals who have seen a doctor over the last year and class 1 having just under 40% of such individuals (see Table 9).

Determinants of health care utilization in Germany

Table 9. Estimated Percentage Visited Doctor by Class

The low use class is far more sensitive to the covariates and demonstrates all the expected patterns for the controls: no clear pattern in time, geography, and age; handicap and a compromised health having a positive effect on use; and health satisfaction having a negative effect. Education has a significant positive effect, while being unmarried has a significant negative effect on use. Probability of doctor visitation also has a significant negative correlation with risk and hours worked. Foreign nationality is not significant. Unexpectedly, the effect of possession of private insurance on doctor visitation is significant and negative rather than positive as expected. This may arise from the covariates and latent classes not fully picking up the heterogeneity between those with private insurance and those without it. For the high utilization class, very few of the variables are significant. In fact, only health satisfaction, risk and foreign nationality are significant at the 95% level. The effect of health satisfaction on doctor visitation is, as expected, a negative one. The effect of risk is positive, while that of foreign nationality is negative. Regression #5: Female Hospitalization (Count) Table 10. Female Hospitalization Count Regression Results

Columbia Economic Review | 37

Shaked Barkay

Determinants of health care utilization in Germany

In this regression model there is a hybrid of classes: some differ on overall levels of consumption along with the coefﬁcients of the covariates, while others differ almost exclusively on the latter. Class 1 is a low utilization class that includes the majority of individuals, whereas classes 2-4 all have higher mean levels of hospitalizations (Table 11). Table 11. Estimated Mean Category of Doctor Visits by Class

Across all four classes, neither geography nor time variables were significant (apart from the year 2015 for class 1). The effects of the health controls largely conformed to expectations, though not all were significant. As expected, health satisfaction has a significant negative effect on number of hospitalizations for three of the four classes. At least one of the handicap categories is significant and positive for each class. The personal assessment of current health is only significant for two classes. Finally, age is only significant for the low utilization class, having (surprisingly) a negative effect on hospitalizations. This may be due to two reasons: health variables picking up morbidity associated with age, and/or a higher rate of pregnant women in that class. Of the remaining variables, only employment and income are significant across any of the classes. Lack of employment has a significant positive effect on number of hospitalizations across all four classes. Income is significant only for class four, wherein it has a negative effect on number of hospitalizations. Private insurance is insignificant across all classes.

Columbia Economic Review | 39

Shaked Barkay

Regression #6: Male Hospitalization (Count) Table 12. Male Hospitalization Count Regression Results

Determinants of health care utilization in Germany

As in the case for females, class one is a low utilization class that includes the majority of the individuals in the dataset, while classes 2-4 have higher mean levels of hospitalization. Table 13. Estimated Mean Category of Hospitalizations by Class

Again, the controls for geography and time were not significant apart from the year 2015 for class 1. When significant, the controls for health conform to expectations. Age has a positive and significant correlation with number of hospitalizations in three classes (all but class 2) and is particularly strong in the high use class. The same pattern is true of health satisfaction, the effect of which is significant and Columbia Economic Review | 41

Shaked Barkay

negative in all but class 2. At least one of the two categories of handicap status are significant – with a positive effect – for all classes except for class 2. Both the “Bad” and “Less good” personal assessments of health are significant in class 3 and 4. Lack of employment, marital status and income are insignificant across all classes. Hours worked and education are both significant and negatively correlated for class 4. Unlike in the case for females, risk is significant in predicting the number of hospitalizations in two of the four classes, having a positive effect for both. Once again private insurance is insignificant in all classes. Regression #7: Female Doctor Visits (Count) Table 14. Female Doctor Visits Count Regression Results

Determinants of health care utilization in Germany

For the ‘female doctor count’, three classes minimize the BIC. These classes differ along level of use, where class one is the higher use class, while classes two and three are both lower use classes that differ in their sensitivity to the covariates. Table 15. Estimated Mean Category of Doctor Visits by Class

The high use class is the only class wherein there are any significant variations across time or geography (the 2009, 2014, Bremen and North Rhine-Westphalia variables are all significant). The health controls are mostly significant and conform to expectations, the exception being health satisfaction and the highest level of handicap, which are insignificant in class 3. Columbia Economic Review | 43

Shaked Barkay

Income, foreign nationality, marital status and risk are insigniďŹ cant across all classes. Hours worked and employment are only signiďŹ cant in the high utilization class, where they both have a positive effect on doctor visitation. This goes against what the literature predicts. My result may be due to a higher rate of preventive care undertaken by high use women in order to prevent later health problems that can disrupt work in the long run. Finally, private insurance is insigniďŹ cant across all classes of utilization. Regression #8: Male Doctor Visits (Count) Table 16. Male Doctor Visits Count Regression Results

Determinants of health care utilization in Germany

Here – as in the case for females – there are three signiﬁcant classes where the ﬁrst class is a high use class and the latter two are low use classes with different sensitivity to the covariates (see Table 17). Table 17. Estimated Mean Category of Doctor Visits by Class

The time and geographic controls are insignificant across all the classes, except for the high use class where the variables for the years 2009 and 2011 are significant along with the state of Berlin. The health controls are generally significant across all classes and are consistent with expectations. Age and handicap have a significant positive effect on doctor visits for all three classes, while health satisfaction has a significant negative effect for all three classes. Finally, all the categories of self-reported Columbia Economic Review | 45

Shaked Barkay

with the exception of “Bad” in class 2 and “Acceptable” in class 3 are significant with positive coefficients in comparison to the reference class (“Very Good”). Income and education are all significant (with positive effects) across all classes. Lack of employment is significant and positively correlated in class 3 and the high use class. Being unmarried is only significant in the lowest use category (where it has a negative coefficient). As in the case for females, both income and risk are insignificant across all classes. Private insurance is only significant for class 3 where – unexpectedly – it is negatively correlated with doctor visits. Again, this may be due to unaccounted heterogeneity between private and public insurance holders. Conclusion A summary of all the regressions and significant variables (with the exception of the time, geography and health controls) is provided below (Table 18). First, the results show a significant amount of heterogeneity across classes within every measure of health care use, demonstrating the importance of using latent class analysis to observe individual heterogeneity. Second, they show that the ownership of private insurance is not positively correlated with use. This contrasts sharply with research demonstrating that fuller insurance increases the amount of health care consumption. In this study, no evidence for moral hazard is found. Even more unusual, in the few cases where insurance coverage is a significant determinant of health care use (high use class of the female doctor visit binary regression, low use class of the male doctor visit binary regression, and class three of the male doctor visit count regression), it has a negative effect. This would occur if there is advantageous selection, where individuals with less propensity to utilize health care opt for private insurance. This can arise if the included controls do not sufficiently account for the full extent of heterogeneity between those with private and public insurance. However, since the controls do include a wide range of the theorized causes for advantageous selection (income, education, age and risk aversion) (Fang et al., 2008; Schmitz, 2011), this result is nonetheless significant and warrants further exploration. Other significant findings are that lack of employment is generally correlated with greater use of health care while working longer hours is generally correlated with less use (with some exceptions), which is consistent with literature suggesting that working long hours is a barrier to the consumption of health care. Income and education are more significant factors in determining the quantity of male physician- use than female physician-use, demonstrating significant health care use patterns between sexes. The willingness to take risk is a significant determinant of the probability of utilizing health care as well as positive effect on the frequency of male hospitalization. This demonstrates that the adverse effects of risk-taking 46

Determinants of health care utilization in Germany

behavior outweigh the reduced propensity to seek hospital care for males. Therefore, differences in risk aversion between the sexes cannot be used to explain greater female use of health care as was suggested by Rosen et al, 2003, at least for the frequency of hospitalization. Foreign nationality is not found to be broadly significant in determining health care use with the exception of the male hospitalization probability where it showed mixed results. This is consistent with literature demonstrating mixed effects of nationality on health care use. The study’s finding of an absence of moral hazard across all use types has important policy implications. The assumption of moral hazard due to increased health insurance coverage has led to concerns over expanded health insurance coverage and rise of deductibles in an effort to address this. The goal of these deductibles is to provide a safety net for costlier and typically more urgent care but disincentive less necessary care by placing the initial costs entirely on the consumer of the health care, so that he does not utilize more health care than he otherwise would in the absence of moral hazard or at the very least minimize the distortionary effects of increased health insurance coverage. German private insurers have attempted to tackle the assumed moral hazard problem by allowing lower premiums for those who accept deductibles. In the presence of moral hazard this would be welfare improving. However, the finding of no moral hazard in this study, while warranting further research, implies that this policy may not be welfareimproving. Table 18. Regression Results Summary

Columbia Economic Review | 47

Shaked Barkay

Notes 1 Binary doctor visits refers to whether or not one has visited a doctor over the past twelve months. 2 Binary hospitalization refers to whether or not one has been hospitalized over the past twelve months. 3 An expression used to the describe the lowest quality used cars 4 Whether or not one has visited a general practitioner over the past year 5 The number of times one has visited a general practitioner 6 Refers to anyone not currently employed rather than unemployed as deďŹ ned by Bundesministerium fĂźr Arbeit und Soziales

Determinants of health care utilization in Germany

References Åhs, A., Burell, G. & Westerling, R. (2012). Care or Not Care—that is the Question: Predictors of Healthcare Utilisation in Relation to Emploment Status. International Journal of Behavioral Medicine, 19(1), 29-38 Ali F., Elsayed M. (2018). The effect of parental education on child health: Quasi-experimental evidence from a reduction in the length of primary schooling in Egypt. Health Economics, 27(4), 649–662. Akerlof, G. A. (1970). The Market for ‘Lemons’: Quality Uncertainty and the Market Mechanism. Quarterly Journal of Economics, 84(3), 488–500 Bago d’Uva, T. (2005), Latent class models for use of primary care: evidence from a British panel. Health Econ., 14(9), 873–892. Boman, E. K. O., & Walker, G. A. (2010). Predictors of men’s health care utilization. Psychology of Men & Masculinity, 11(2), 113-122. Bosanac, E. M., & Parkinson, R. C. (1976). Geographic access to hospital care: a 30-minute travel time standard. Medical Care, 14(7), 616-624. Boyoung, J., & Soonman, K. (2013) Effect of private health insurance on health care utilization in a universal public insurance system: A case of South Korea. Health Policy, 113(1–2), 69-76. Briscoe, M. E. (1987) Why do people go to the doctor? Sex differences in the correlates of GP consultation. Social Science & Medicine, 25(5), 507-513 Card, D., Dobkin, C., & Maestas, N. (2008). The Impact of Nearly Universal Insurance Coverage on Health Care Utilization: Evidence from Medicare. The American Economic Review, 98(5), 2242-2258. Consedine, N. S., & Butler, H. F. (2014). Mindfulness, health symptoms and healthcare utilization: Active facets and possible affective mediators. Psychology, Health & Medicine, 19(4), 392-401. Corney, R. H. (1990) Sex differences in general practice attendance and help seeking for minor illness. Journal of Psychosomatic Research, 34(5), 525-534 Curtis, L. J., & MacMinn, W. J. (2008) Health Care Utilization in Canada: Twenty-ﬁve Years of Evidence. Canadian Public Policy, 34(1), 65-87. Fang, H., Keane, M., & Silverman, D. (2008). Sources of Advantageous Selection: Evidence from the Medigap Insurance Market. Journal of Political Economy, 116(2), 303-350. Fell, D. B., Kephart, G., Curtis, L. J., Bower, K. , Muhajarine, N. , Reid, R. and Roos, L. (2007), The Relationship between Work Hours and Utilization of General Practitioners in Four Canadian Provinces. Health Services Research, 42(4), 1483-1498. Ford, E. S., & Cooper, R. S. (1995). Racial/ethnic differences in health care utilization of cardiovascular procedures: a review of the evidence. Health Services Research, 30(1), 237–252. Glaesmer, H., Brähler, E., Martin, A., Mewes, R. & Rief, W. (2012), Gender Differences in Healthcare Utilization: The Mediating Effect of Utilization Propensity. Journal of Applied Social Psychology, 42(5), 1266–1279. Johnson-Lans, S. & Bellemore, F. (1997) Gender and race as factors in health care utilization. International Advances in Economic Research, 3(2), 193-205. Columbia Economic Review | 49

Shaked Barkay

Joung, I. M., van der Meer, J B, & Mackenbach, J. P. (1995). Marital status and health care utilization. International Journal of Epidemiology, 24(3), 569. Koopmans, G. T., & Lamers, L. M. (2007) Gender and health care utilization: The role of mental distress and help-seeking propensity. Social Science & Medicine, 64(6), 1216-1230. Liu, L. F., Tian, W. H., & Yao, H. P. (2012). Utilization of health care services by elderly people with National Health Insurance in Taiwan: The heterogeneous health profile approach. Health Policy, 108(2–3), 246-255. McDonald, J. T., & Conde, H. (2010). Suppl. Special Issue from the Social and Economic Dimensions of an Aging Population. Canadian Journal on Aging, 29(1), 23-37. McMahon, J. D., Macfarlane A., Avalos, G. E., & Cantillon, P., Murphy, A. W. (2007) A survey of asylum seekers’ general practice service utilisation and morbidity patterns. Irish Medical Journal, 100(5), 461–4. OECD (2016). Health expenditure and financing: Health expenditure indicators (Edition 2015). OECD Health Statistics (database) Olusanya, B., Ashaye, A., Owoaje, E., Baiyeroju, A., & Ajayi, B. (2016). Determinants of utilization of eye care services in a rural adult population of a developing country. Middle East African Journal of Ophthalmology, 23(1), 96. Höfter, R. H. (2007) Private health insurance and utilization of health services in Chile. Applied Economics, 38(4), 423-439. Rosen, A., Tsai, J., Downs, S. (2003). Variations in Risk Attitude across Race, Gender, and Education. Medical Decision Making., 23(6), 511-517. Rowell, D., & Connelly, L.B. (2012) A history of the term “moral hazard.” Journal of Risk and Insurance, 79(4), 1051–75. Shannon, G. W., Lovett, J., & Bashur (1979). Travel for primary care: expectation and performance in a rural setting. Journal of Community Health, 5(2),113-125. Schmitz H. (2011). Direct evidence of risk aversion as a source of advantageous selection in health insurance. Economic Letters, 113(2), 180-182. Wagstaff, A., & Lindelow, M. (2008). Can insurance increase financial risk? The curious case of health insurance in China. Journal of Health Economics, 27(4), 990-1005. Weinstein, H. M., Sarnoff, R. H., Gladstone, E., Lipson, J. G. (2000) Physical and psychological health issues of resettled refugees in the United States. Journal of Refugee Studies, 13(3), 303–27. Wong, A., Wouterse, B., Slobbe, L. C. J., Boshuizen, H. C., & Polder, J. J. (2012). Medical innovation and age-specific trends in health care utilization: Findings and implications. Social Science & Medicine, 74(2), 263-272. Xin, L., & Wei, Z. (2013). The impacts of health insurance on health care utilization among the older people in China. Social Science & Medicine, 85, 59-65. 50

WHAT IS THE IMPACT OF A MEDICAID EXPANSION ON THE BANKRUPTCY RATE? EVIDENCE FROM THE ACA

Noah Zwiefel Macalester College Abstract: The ACA Medicaid expansion increased the number of individuals enrolled in Medicaid, reducing the risk that many will have to pay for medical procedures out of pocket. While this decreased the medical bankruptcy rate, the exact magnitude of this impact is unknown. Using a difference-in-differences model with controls and a fixed effects model with an instrumental variable, I estimate that 75,000-125,000 bankruptcies were prevented by the ACA Medicaid expansion between 2014 and 2016 among low-income individuals, which is an approximate 6-8% decrease in the bankruptcy rate. This effect is stronger among those earning below the median income; there is no evidence of a similar effect among those earning above the median income. These results are strengthened through a series of falsification tests and robustness checks. This research provides yet another reason to support Medicaid expansions, given these clear financial benefits.

I. Introduction

ccording to Himmelstein (2009), 60% of bankruptcies can be attributed to medical costs. The U.S. medical system is one of the most expensive health systems in the world (Papanicolas et al, 2018). The median cost of an emergency appendectomy in the United States is $33,000 (Hsia, 2012), and 20.6% of non-elderly Americans who were uninsured in 2013 – 36 million individuals (KFF, 2017) – must pay this entire expense out of pocket, or risk bankruptcy. A similar procedure would cost only $4,000 in Australia (Kliff & Oh, 2018), with none of this cost charged to the patient. This difference in costs and increase in debt has real impacts on individuals, with over half (53%) of the uninsured reporting that they struggled with medical debt, and more than 125,0001 bankruptcies in 2015 were directly attributed to medical debts among the uninsured (Kaiser Family Foundation, 2016). Medicaid is the U.S. government’s public health insurance program follow income and disabled individuals, is designed to ensure that the poor and disabled have access to medical care. With over $592 billion in spending by individual states and the federal government during FY 2018 and over 37 million adults enrolled (KFF, 2018), this program provides coverage to many Americans (KFF, 2018). Columbia Economic Review | 51

Noah Zwiefel

The Affordable Care Act (ACA) Medicaid expansion increased the number of individuals en-rolled in Medicaid by 11.9 million (KFF, 2018), reducing the risk that many will have to pay out-of-pocket for an emergency medical procedure. However, the significance of this impact in terms of actual bankruptcies prevented is unknown. In light of the passage of Medicaid expansion ballot initiatives in Idaho, Nebraska and Utah, and the election of pro-Medicaid expansion governors in Kansas, Maine, and Wisconsin in 2018, this question assumes an even more relevant role in the literature. Prior research finds an 8% decrease in bankruptcies as a result of a 10% increase in Medicaid eligibility under the CHIP and Medicaid expansions of the 1990s (Gross et al 2011). Gross’s (2011) study is somewhat limited by the state-level nature of Gross’s data, which neglects within state trends. In addition, they do not test for heterogeneous impacts based on the pre-bankruptcy income of the filer. More recently, Brevoort et al (2018) estimate using a difference-indifferences approach that 125,0001 bankruptcies a year were prevented by the ACA Medicaid expansion in expansion-implementing states. However, Brevoort et al (2018) do not test the robustness of this estimate, nor do they ensure that this decrease in bankruptcies is localized specifically to those who become eligible for Medicaid under the ACA, as their focus is on the impacts of the Medicaid expansion on credit availability. To my knowledge, there have been no studies focused specifically on estimating the number of bankruptcies prevented by the ACA Medicaid expansion. In this paper, I contribute by estimating the number of bankruptcies prevented by the ACA Medicaid expansion. I utilize a novel dataset which allows a county-level analysis; in addition, I perform a series of robustness checks, including a previously unutilized check which ensures that my results are localized among those who became eligible for Medicaid. I utilize two econometric models: first, a county-level difference-in-differences approach, which exploits the fact that some states expanded their Medicaid programs while others held their criteria constant, given the op- tionality of this expansion as a result of a 2012 Supreme Court decision. Secondly, I use a county level fixed-effects model. In both cases I find evidence that the ACA Medicaid expansion prevented approximately 75,000-125,000 bankruptcies in Medicaid expansion implementing states between 2014 and 20162. This comprises between one-third and two-thirds of the decrease in bankruptcies over this time period, and a reduction by 6-8% in the overall bankruptcy rate. This effect is stronger when examining only those who earn below the median income, and non-existent among those earning above the median income.

Impact of Medicaid on Bankruptcy Rate

II. Bankruptcy Background and Definitions3 Within the United States, bankruptcies for individuals generally fall into two categories: Chapter 7 and Chapter 13. Under Chapter 7, individuals must surrender all assets above an exemption limit, but have their debts fully forgiven. Meanwhile, under Chapter 13, individuals are allowed to keep all of their assets, but must work with the courts and their lenders to develop a plan to repay at least part of their debts over a period of time. Both chapters of the bankruptcy code leave the same mark on an individual’s credit report. However, under Chapter 13 that mark only exists for 7 years, while under Chapter 7 the mark remains for 10 years. In some cases, as a result of the 2005 bankruptcy code reform, an individual is forced to file under Chapter 13 if their income is above the median income in their state of residence (Li, 2007). In other cases, even if an individual would qualify for and benefit from filing for bankruptcy under Chapter 7, they may be unable to afford a lawyer, which given the complex nature of bankruptcy forms means that they are unable to file for bankruptcy (Kiel, 2018). Since a low-income individual seeking bankruptcy is unlikely to have access to enough cash to cover the $1,000 or greater attorney fee upfront, their lawyer may instead offer a payment plan, and induce them to file under Chapter 13. The lawyer then becomes one of the individual’s creditors, rolling their fee into a payment plan under Chapter 13 (Fresques, 2017). This barrier results in many individuals that would benefit the most under Chapter 7 instead filing under Chapter 13. To address this distortion, I categorize consumer bankruptcy filings into low and high-income filers based on the whether the filer’s average monthly income for the past 12 months, as declared at filing on Line 16, Schedule I of their bankruptcy filing, is above or below their state’s median income. Because individuals that became eligible for Medicaid under the ACA Medicaid expansion will almost certainly have income below the median income, any change in the bankruptcy rate as a result of the ACA Medicaid expansion should be entirely localized to this group. I exploit this fact in a series of robustness checks, to substantiate my claim of identification.

III. Empirical Approach I utilize two econometric models to identify the impact of the ACA Medicaid Expansion on the bankruptcy rate. First, I utilize a difference-indifferences approach deﬁned below:

Columbia Economic Review | 53

Noah Zwiefel

Yit = β0+β1Expandedi +β2PostExpansiont +β3PostExpansiont × Expandedi +β4Xit +εit (1) where Yit is the bankruptcy rate per 1,000 residents in county i during year t; β 1 is the treatment group term, which accounts for the average constant differences between expansion and non- expansion counties; β 2 is the average time trend post-2014 in expansion and non-expansion states; β 3 is the average treatment effect; β4 is a vector of control variables for each county i in year t; and εit is an error term. Under this approach, the treatment and control groups must remain the same over the entire time period. Unfortunately, several states chose to expand their Medicaid programs post-2014, as shown in Figure 1a. To ensure that my treatment and control groups remain the same, I re-categorize states as presented in Figure 1b. I lose all observations from Pennsylvania, Indiana, Alaska, Montana, and Louisiana, however, I do not exclude Michigan or New Hampshire, as both states expanded their Medicaid programs in 2014. Furthermore, I do not exclude any state that expanded after 2016. The vector of controls, X, is necessary because there are clear differences in the overall bankruptcy patterns between states that expanded and those that did not, as proven by Figure 4’s non- parallel trends in bankruptcy pre-expansion. Non-parallel trends are similarly found when splitting bankruptcy filers by income, as shown in Figure 3. In all of these instances, I assume that parallel trends are created through the use of appropriate controls 4 and use a series of robustness checks to corroborate my results. Because of the strength of assumptions needed for an accurate estimate of average treatment effect under difference-in-differences, I attempt to confirm these results by estimating the change in bankruptcy rate as a result of a change in Medicaid enrollment. I modify a log-linear fixed effects specification employed by previous researchers (Gross et al, 2011): Inverse Hyperbolic Sinea(Yit) = β0 + β1MedicaidEnrolledit + αi + αt + εit (2) where Yit is the number of consumer bankruptcies per 1,000 filed in county i in year t; MedicaidEnrolledit is the percent of the 18-64 population in county i in year t that is enrolled in Medicaid; α i and αt are fixed effects for year and county respectively; and εit is an error term. With this specification, I am able to include all states, even those that expanded after 2014. Unfortunately, this specification introduces the possibility of omitted variable bias. Medicaid enrollment is influenced by factors outside of the change in eligibility under the ACA. A change in enrollment within a county could occur because of an unobservable shock – not associated with the Medicaid a

The inverse hyperbolic sine is a form of a log transformation, deﬁned as log(y + jy2 + 1). This transformation is not undeﬁned where y = 0, and thus ensures that all counties are included in my regression.

Impact of Medicaid on Bankruptcy Rate

expansion – which impacts the health, beliefs or behaviors of residents, inducing them to enroll in Medicaid, even if they were previously eligible. Causality cannot be inferred with certainty, and there is a high likelihood that the estimate of the average treatment effect is inconsistent and biased. To address this, I use Medicaid eligibility as an instrumental variable. Given that a change in Medicaid eligibility criteria is what produced increased Medicaid enrollment under the ACA, I will be able to make a direct causal inference of the ACA Medicaid expansion’s impact on the bankruptcy rate. I use a fixed effects two-stage least squares estimator to instrument Medicaid enrollment. During the first stage, an instrumented version of Medicaid enrollment is calculated, using Medicaid eligibility as the instrumental variable. MedicaidEnrollmentit = γ0 + γ1MedicaidEligibilityit + αi + αt + uit (3) Then, in the second stage, that instrumented version of Medicaid enrollment is utilized in my originally specified regression. Inverse Hyperbolic Sine(Yit) = β0 + β1MedicaidEnrollmentit + αi + αt + εit (4) I assume that bankruptcy rates would have had a similar pattern post-Medicaid expansion had the eligibility criteria for Medicaid not changed. I further assume that the fixed effects of a county absorb all local characteristics which impact the bankruptcy rate, and that these characteristics remain constant during my sample period; I also assume that the year fixed effects will absorb all time-specific U.S. macroeconomic shocks that may influence the bankruptcy rate. Finally, I assume that there are no systematic drivers of a change in the percent of individuals eligible for Medicaid within a county, besides the change in Medicaid rules under the ACA. Under these assumptions, instrumented Medicaid enrollment will be the only variable influencing the bankruptcy rate in this regression. To test the validity of these assumptions, I examine several specifications with controls for confounding variables 5. When including these controls, my system of equations becomes: MedicaidEnrollmentit = γ0 + γ1MedicaidEligibilityit + γ2X + αi + αt + uit (5) Inverse Hyperbolic Sine(Yit) = β0 + β1MedicaidEnrollmentit + β2Xit + αi + αt + εit (6)

where X is a vector of control variables.

Columbia Economic Review | 55

Noah Zwiefel

To validate this model, as with my difference-in-differences, I use a series of falsification tests designed to ensure that my results are specifically localized to bankruptcy filings among low income individuals.

IV. Data Ideally, I would have an annual count of bankruptcies specifically caused by medical costs, and a measure of Medicaid enrollment, absent any unmeasured or unobserved shocks, aside from the change in policy. Direct causality could be readily established were this data available. Unfortunately, these ideal data do not exist, since measuring and collecting such data would be both prohibitively expensive, and likely impossible in the case of medical bankruptcy filings. As my measure of bankruptcy, I utilize a county-level derived annual count of bankruptcies by chapter and type of filer between 2011-2016 (Federal Judicial Center, 2018). This is a novel, previously unutilized dataset that enables a more localized analysis when compared with previous literature, which has tended to utilize a state level count of bankruptcies from the U.S. Court system’s F-2 annual report (2018). For Medicaid enrollment, I use the one-year ACS PUMS to derive a weighted estimate of the percent of the 18-64 year old population in a Public Use Microdata Area (PUMA) that is enrolled in Medicaid. I then use a weighted geocorrelation average to convert the PUMA-level estimate into a county-level estimate. As my instrumental variable, I utilize a simulation detailed in Appendix B to derive an estimate of the 18-64 year old population that is Medicaid eligible through the income- qualifying pathway at a county level. This estimate of eligibility may lack precision6. I address this through several robustness checks, in addition to my difference-in-differences approach. I utilize a number of control variables which prior studies have used as potential predictors of bankruptcies (Domowitz, 1999; Gross 2011). As economic predictors of bankruptcy, I utilize county level data on median income, the bottom and second quintile of income distribution, the business bankruptcy rate, per-capita personal income, and the unemployment rate. As demographic control variables, I utilize the percent of population aged 25-44, one of the populations most likely to file for bankruptcy (Domowitz, 1999), the percent of the population at or below the poverty level, the percent of the population with subprime credit scores, the percent of the population that owns their own home, and the percent of the population that is black or Hispanic. These controls were all retrieved from GeoFRED (St Louis Federal Reserve, 2018). I test the robustness of these controls by estimating a model with fixed effects, in an attempt to control for unobserved or unmeasured heterogeneities. My summary statistics are presented in Table 1. For my summary statistics

Impact of Medicaid on Bankruptcy Rate

figures, I omit Pennsylvania, Indiana, Alaska, Montana, and Louisiana unless otherwise specified – all states that chose to expand their Medicaid program after 2014 but before 2016. Figure 2 compares the estimated percentage of the 18-64 population in expansion and non-expansion states that are eligible or enrolled in Medicaid. The Medicaid eligibility rate greatly increased post-2014 in Medicaid expansion-implementing states, while the eligibility rate among non-Medicaid expansion-implementing states remained relatively flat. The enrollment rate increased in expansion states, however it did not increase by as much as eligibility, which is to be expected, given that take up, for nearly all government programs, tends to lag behind eligibility (Currie, 2004). Controlling for no other economic characteristics of a community, this large increase in eligibility seems to have had a limited impact on the number of bankruptcies. Figure 3 stratifies bankruptcy by pre-bankruptcy income of the filer. Post-Medicaid expansion, it appears as if Medicaid expansion states had a larger decrease in low-income bankruptcy filings when compared with non-Medicaid expansion states. Meanwhile, the post-Medicaid expansion time period is associated with a relatively parallel trend in high-income bankruptcy filings in expansion and non-expansion states. This is consistent with the ACA Medicaid expansion impacting the bankruptcy rate, as we would expect to see a change in bankruptcy filing patterns to be localized to low-income bankruptcy filers. Theoretically, these trends could be attributable to the recovery from the 2008 recession. The 2008 financial crisis resulted in a massive spike in bankruptcies, which led to a steep decrease in bankruptcies after the recession ended. Thus, examining only the raw bankruptcy rate does not enable us to extrapolate pre-treatment trends as a “but-for”, or counterfactual, bankruptcy rate. Given this, I attempt to estimate a more accurate average treatment effect by attempting to control for economic characteristics that would be driving a change in bankruptcy as a result of a recovering economy , such as the median income of an area, the income distribution, and the unemployment rate. Figure 4 explores the trend in aggregate consumer bankruptcies filings when compared with aggregate business bankruptcies. Under the hypothesis that the ACA Medicaid expansion impacted bankruptcy filings, we should observe that consumer bankruptcy filings fall relative to business bankruptcies and this is, indeed, what we note. Consumer bankruptcies appear to have fallen by a greater amount than business bankruptcies in expansion states, when compared with non-expansion states. More quantitatively, I present summary statistics on the bankruptcy rate in Table 2. Bankruptcies fell more post-expansion in Medicaid expansion states, when compared with non-Medicaid expansion states. Overall, a visual inspection of these figures and tables implies that the bankruptcy rate fell by

Columbia Economic Review | 57

Noah Zwiefel

slightly more in expansion states than in non-expansion states. This is, of course, a naive approach, given that we are controlling for no other macroeconomic or community level features that may impact the bankruptcy rate. Thus, a more thorough econometric approach is needed to analyze this question.

V. Results 1.Difference-in-Differences I first utilize a difference-in-differences approach to estimate the effect of the ACA Medicaid expansion on the bankruptcy rate. I initially estimate a model with no control variables, the results of which are presented in Table 3. This model implies that the ACA Medicaid expansion is associated with a decrease of approximately 27 bankruptcies per 100,000 residents in a county, or a 10% decrease from the 2011-2013 mean. To ensure that this statistical significance is not the product of statistical abnormalities or heteroskedasticity, I utilize robust standard errors in Model 2, and cluster standard errors by state in Model 3. The coefficient on the interaction terms remains statistically significant in all cases. Given the selection bias, non-parallel trends, and contamination from the 2008 recession recovery, a degree of skepticism ought to be applied to these results. To address these shortcomings, I next estimate this model using a set of control variables. These estimates are presented in Table 4. The coefficient of interest attributes an approximate decrease in the bankruptcy rate of 18 per 100,000 to the ACA Medicaid expansion, or a 5.9% decrease over pre-treatment bankruptcy rates. This coefficient remains statistically significant, regardless of whether I utilize heteroskedastic robust or clustered standard errors. I explore several different specifications with controls for how liberal a state’s government may be, given that this may have influenced the implementation of the ACA Medicaid expansion. States with more liberal political environments may have invested increased resources in advertising the ACA Medicaid expansion, or may have had additional navigators available to help individuals sign up. Furthermore, liberal states may have had more consumer protections, or better social safety net programs available, confounding the impacts of the ACA Medicaid expansion. If not controlled for, this political heterogeneity may lead to a corrupted estimate of average treatment effect. Thus, I test several different controls, including the percent Democratic margin for president in 2008, 2012, 2016, and whether or not the state’s governor in 2014 was a Democrat. As we see in Models 3, 4, and 5 in Table 4, these each have little impact on the average treatment effect. To further test this model, I utilize several robustness checks, presented in Table 5. As a first falsification test, I estimate my model with business

Impact of Medicaid on Bankruptcy Rate

bankruptcies per 1,000 as my dependent variable. As we see in Model 1, the interaction term loses all significance – or, in other words, the ACA Medicaid expansion is not associated with a change in the business bankruptcy rate. I next re-estimated my model twice: once with the bankruptcy filers which I classified as low-income and once with high-income bankruptcy filers. We expect the impacts of the ACA Medicaid expansion on the bankruptcy rate to be localized to low-income filers, and indeed that is what we observe. We see a coefficient that lacks any form of statistical significance in Model 2, while in Model 3 we see a statistically significant coefficient. I next examine the degree to which states that exercised an option to expand their Medicaid programs early may be contaminating my results. To test this, I excluded any states that took advantage of this waiver from my regression.7 The results of this are presented in Table 5, Models 4 and 5. While the standard error rises8, the estimate of average treatment effect remains similar to prior estimates. I thus assume that any contamination of results from including these states is minimal. I confirm this assumption in Section 6.2, using a regression specification that should not be impacted by these early expanders. Finally, I investigate if there were additional unmeasured or uncontrolled for heterogenities between the treated and control states by adding county and year fixed effects. As we see in Table 5, Model 6, adding fixed effects had little impact on the average treatment effect. My control variables thus seem to be adequately accounting for the idiosyncrasies between the two groups, lending credence to this being an accurate estimate of average treatment effect. 2. Fixed Effects with Instrumental Variable To further establish causality and strengthen this conclusion, I estimate a fixed effects specification. My results from this specification are presented in Table 6. In this model, a 1% increase in instrumented Medicaid enrollment is associated with a 1.483% decrease in the bankruptcy rate. When we consider that instrumented Medicaid eligibility increased by 5% on average in expansion states, we can attribute a decrease by approximately 22.56 bankruptcies per 100,000 residents to the ACA Medicaid expansion. This estimate falls to 15.3 bankruptcies per 100,000 residents when adding controls. These estimates are almost identical as the estimates from my difference-in-differences model, and remain significant when clustering standard errors by state. To analyze the robustness of this model, I tested a series of confounding predictors of bankruptcy to ensure that fixed effects were properly accounting for community demographics and macroeconomic trends. These results are presented in Table 7. The coefficient of interest – the coefficient on instrumented Medicaid enrollment – changes only slightly, regardless of which controls I utilize. While it is impossible to test every possible confounding factor,

Columbia Economic Review | 59

Noah Zwiefel

given this representative selection of variables, it seems likely that fixed effects roughly proxy for any observed or unobserved heterogeneity which could be influencing my results. Finally, as with my difference-in-differences model, I utilize a series of robustness checks to ensure that these results are localized to low income bankruptcy filers. The results of these falsification tests are presented in Table 8. Once again, I find no evidence that an increase in enrollment under the ACA is associated with a change in the business or high-income consumer bankruptcy rates; however, it is strongly associated with bankruptcy filings among those that are earning below the median income, even when clustered by state. This corroborates my identifying assumptions, given that any change in the bankruptcy rate as a result of the ACA Medicaid expansion should be entirely localized to low income bankruptcy filers.

VI. Discussion 1. Overall Bankruptcy Rate In Figures 5a and 5b, I present the number and percent of bankruptcies in 2014 that were prevented by the ACA Medicaid expansion among all Americans in Medicaid expansion states, based on the results of my Fixed Effects model. Among all bankruptcy filers, the ACA Medicaid Expansion decreased the bankruptcy rate by between 1-8% in each county. A comparison of how many bankruptcies each of my approaches predicts were prevented between 2014-2016 by the ACA Medicaid expansion is presented in Figure 6. Each of my approaches produces fairly consistent point estimates of between 75,000-125,000, as well as roughly equivalent 95% confidence intervals, with the exception of my fixed-effects model, which is afflicted with large standard errors. As we see in Figure 7, this implies that the ACA Medicaid expansion is responsible for between one-third to two-thirds of the decrease in bankruptcies between 2014-2016. This is a fairly significant impact, and one that is similar to estimates from Brevoort et al (2018), although given the large confidence intervals, a degree of uncertainty remains. 2. Low Income Bankruptcy Rate One reason for the large confidence interval is that we are measuring the impact on the over-all bankruptcy rate, for both rich and poor Americans. When we examine only the change in bankruptcy rate for Americans earning below their states median income in Figure 8, we see a much starker impact. This model implies that the ACA Medicaid expansion resulted in a 1-13% decrease in each county’s bankruptcy rate year-over-year in 2014 among those earning less than the median income. Figure 6 reports the number of bankruptcy filings among those earning below the median income that each

Impact of Medicaid on Bankruptcy Rate

approach predicts were prevented between 2014-2016 due to the ACA Medicaid expansion. As we see, these models consistently provide a point estimate of between 75,000 and 125,000 bankruptcies, with very similar 95% conﬁdence intervals. Given the similarity between the estimates from the overall bankruptcy model and this model, my case for causality is further validated. This estimate is, again, in line with estimates from Brevoort et al (2018). These estimates imply that the ACA Medicaid expansion is responsible for the majority of the decrease in bankruptcy rates among the low income between 2014-2016, as we observe in Figure 7. This is, of course, assuming that there were no factors inﬂuencing bankruptcy upwards; it is possible that there could be macroeconomic trends forcing bankruptcies among the low income upwards, and but-for this trend the impact of the ACA on the bankruptcy rate would have been relatively lower. Regardless of this potential, however, if my estimates are accurate, the bankruptcy rate would have been at least 6-8% higher in expansion implementing states in a counterfactual scenario.

VII. Limitations There are several factors that could threaten the validity of my study, which I briefly summarize in Table 9 and address in detail in this section. I am first limited by my data. The overall bankruptcy rate, and even the low-income bankruptcy rate I calculate may be problematic. I cannot be certain whether these bankruptcy filers filed for bankruptcy for medical reasons, what their credit scores were, or if they were even eligible for Medicaid. Individuals may misreport their average monthly income, or there could be underlying clerical errors. It is further not uncommon for the wealthy to hide their income prior to bankruptcies (Goldstein, 2013). This could induce downwards bias in my results, given that the ultra-wealthy may end up included in the ”low income” bankruptcy group. While there is nothing I can do to address this given my resources, I assume that any bias that is introduced due to my bankruptcy data is a very limited downwards bias. These are official court house records. Significant discrepancies or inaccuracies would be concerning for the stability of our legal system. I am equally limited by my estimates of Medicaid eligibility and enrollment, which may lack precision. Because I am relying on the annual American Community Survey PUMS, sampling bias in either direction is a likely possibility; however, I utilize the ACS provided person weights, which greatly mitigates this potential inaccuracy9. My difference-in-differences model is limited by the fact that treatment was not randomly assigned. As we observe in Table 10, states in the expansion and non-expansion groups were demo- graphically and economically heterogeneous prior to the Medicaid expansion. On average, people in non-expansion states tended

Columbia Economic Review | 61

Noah Zwiefel

to be poorer, and were more likely to have subprime credit scores. This implies that all else constant, were non-expansion states to implement the Medicaid expansion, the impact would likely be much greater. Complicating this conclusion, however, is the fact that counties in non-expansion states simultaneously tended to have fewer low-income bankruptcy filers, due perhaps to the poverty trap described earlier, where low income individuals are unable to file for bankruptcy because they are too poor (Kiel, 2018). It is thus possible that an extrapolation of these results may result in either an understatement or an overstatement of the number of bankruptcies prevented, depending on which of these two is a greater factor. More significantly, I am limited by non-parallel pre-treatment trends in bankruptcy filings between the two groups pre-treatment. As with the selection bias, if left unmitigated, this has the potential to significantly bias my estimates in either direction. To address both the selection bias and non-parallel trends, I explore Difference-in-Differences specifications with control variables, county and year fixed effects, and falsification tests. I further use an alternative fixed effects regression approach. Given the consistency among all of these estimates, I believe that the risk that selection bias and non-parallel trends poses to the validity of my results is significantly lessened, as long as I utilize control variables or fixed effects. This model is additionally limited by the fact that Medicaid enrollment increased in both expansion and non-expansion states, because of increased publicity and a new enrollment process. This results in a contamination of my untreated group, which may bias the average treatment effect downward. I believe that the impacts of this are limited, as the difference in mean Medicaid enrollment between 2013 and 2014 in non-expansion counties was a statistically insignificant 1-3% (see Table 2), compared to 10% or greater in expansion states (Wehby, 2018; Wachino, 2014). Some of the aforementioned limitations in my difference-in-differences models are addressed through a fixed effects approach. However, this model is limited as well, mainly by the fact that there are many unobservable shocks that may impact enrollment beyond the change in eligibility under the ACA Medicaid expansion. I utilize several strategies in an attempt to mitigate any bias introduced. I use county-level fixed effects, which theoretically captures all within-county heterogeneity, thus resulting in a more accurate estimation of the treatment effect when compared with prior studies (Gross, 2011). I further use Medicaid eligibility as an instrumental variable, in an attempt to isolate the change in Medicaid enrollment to just that which is due to the change in Medicaid eligibility, the driver under the causality I posit.

Impact of Medicaid on Bankruptcy Rate

More generally, the Medicaid expansion causality I propose, and the specific estimate for bankruptcies prevented I calculate may be biased upwards by other components of the Affordable Care Act. These components include a ban on health insurance policies that cap the amount of benefits given out in a year or over a lifetime, a requirement that insurance companies offer coverage to those with pre-existing conditions, and a mandate that all individuals must purchase health insurance. Given that these were national policies, in theory the impact of these changes should be homogeneous across states within a year. However, it is entirely possible that some localities may have been more impacted. In some regions pre-treatment and benefit caps may have been common, while in other regions, benefit caps may have been uncommon. I believe that the impact of this is limited, given my quasi-experimental design 10, the relative consistency between the estimated treatment effect from both my fixed effects model and difference-in-differences model, the set of control variables I use11, and that the change in bankruptcies is localized to those earning below the median income. However, despite these mitigation attempts, the risk of a slight upwards bias remains. I do not evaluate whether this decrease in bankruptcies was driven by a decrease in medical debt, or whether it was driven by an increase in earnings due to less sick time. It is plausible that preventing the shock to income from a prolonged or serious illness may be a major driver of medical bankruptcy, especially given the fact that many low income individuals work jobs where they are not eligible for sick-time pay. For the 78% of Americans who reported living paycheck to paycheck in 2017 (CareerBuilder), missing even a week of pay for medical reasons is enough to place them in precarious financial position. Finally, I am limited in what conclusions can be drawn regarding the actual medical bankruptcy rate. While a 3% decrease in the overall bankruptcy rate may seem more in line with Dobkin’s (2018) estimate of a 6% medical bankruptcy rate, rather than Himmelstein and Warren’s (2005) estimate of 50% of bankruptcies being medically caused, it is essential to realize that the vast majority of individuals in the United States receive their insurance through the private sector. I evaluate only one portion of the Affordable Care Act, and fully ignore the impact of its private health insurance reforms. A far greater driver of the medical bankruptcy rate could be the privately insured with a high deductible plan who face unexpected medical costs, or who discover that their insurance company refuses to pay claims because a medical provider was out of network, or an emergency procedure was not pre-approved (Rosenthal, 2015). Moreover, while an individual’s medical bills from an extended hospital stay may be covered, they may still have lost income during their illness, and thus been unable to meet other financial obligations. These other expenses may result in an individual still having to declare bankruptcy, even if covered by health insurance.

Columbia Economic Review | 63

Noah Zwiefel

VIII. Conclusion I have found evidence that the ACA Medicaid expansion prevented between 75,000-125,000 bankruptcies between 2014-2016 in Medicaid expansion-implementing states. Without this program, the bankruptcy rate would have likely been around 8% higher between 2014-2016 among those earning below the median income in Medicaid-expansion states, and approximately 6% higher overall. This is a significant impact for just one portion of the ACA, and yet another way that the ACA Medicaid expansion has left poor individuals in an improved state of being. While this impact may seem numerically small, it should be noted that I do not examine how medical debt burden was impacted by this program. Bankruptcies in general are relatively rare in the population: a recent New York Times survey found that for every 100 people who reported struggling with medical debt, only 2 had filed for bankruptcy that year (Sanger-Katz, 2016). Extrapolating this to my findings, it is conceivable that 2.5 million or more individuals may have avoided a crippling level of medical debt between 20142016 as a result of the ACA Medicaid expansion. This program is impacting real people, such as Alex Andrews, who was shot by a home intruder. Medical coverage under the ACA Medicaid expansion protected him from a shock of $500,000 in medical bills (Campbell, 2017), which would have likely resulted in an eventual bankruptcy filing. From a policy perspective, this provides yet another reason to support the ACA Medicaid expansion. Ignoring the improved health and wellness outcomes among those who become eligible for Medicaid, the financial benefits are clear. This incentives all states to fully implement the ACA Medicaid expansion, given the expenses on the court systems, businesses and individuals within these states that could be avoided through this program. These effects would likely be even more pronounced were Southern states to implement this program, given that these states are somewhat poorer on average than those that implemented the ACA Medicaid expansion. Future research should focus on how the recently approved Medicaid expansions in Virginia, Idaho, Maine, Nebraska, and Utah impact the bankruptcy rate, to examine if consistent results are noted. Further research should also focus on the impacts of private insurance reform, as well as the implications of recent Medicaid-for-all proposals. Finally, future research should not lose sight of what should be the fundamental reason behind a Medicaid expansion: improving the health and well-being of Americans.

Impact of Medicaid on Bankruptcy Rate

Notes 1% of the 53% who reported that they struggled with medical debt. This does not include bankruptcies declared more than 12 months before the survey date, nor bankruptcies among the insured, who may also struggle with medical bills. 2 Approximately 25,000-40,000 a year 3 A brief literature review can be found in Appendix A 4 See Section 5 for a description of the controls that I utilize. 5 I discuss this in Section 5; I use the same controls as in my difference-indifferences models. 6 See Appendix B. My simulation does not exclude adults already eligible for health insurance (for example under their parents’ plan), or who might have insurance through their workplace, but still have an income which qualifies them for Medicaid; in addition, the derivation process may introduce some slight bias. As a result, my estimate may be somewhat imprecise. 7 Specifically California, Connecticut, D.C, Minnesota, New Jersey, and Washington (Kaiser Family Foundation, 2012). 8 This likely occurs because of the loss of observations leading to a loss of predictive power. 9 See Appendix B for more details. 10 It is possible that this too may be compromised by the other components which came into force in 2014 (e.g. the health insurance exchanges and the requirement to buy health insurance). I assume any heterogeneity in the impact of these programs is captured by either my fixed effects or my control variables. 11 Specifically the controls for how Democratic a state was, given that heterogeneity between how these components were implemented or supported on a state-level would likely be driven by how conservative or liberal a state is. 1

References Antonisse, Larisa, Rachel Garﬁeld, and Robin Rudowitz (Mar. 2018). The Effects of Medicaid Expansion under the ACA: Updated Findings from a Literature Review. url: https://www.kff.org/medicaid/issue-brief/ the-effects-of-medicaid-expansion-under-the-aca- updated-ﬁnd ings-from-a-literature-review-march-2018/. Arnold, Emily A. et al. (July 2012). “A Qualitative Study of Provider Thoughts on Implement- ing Pre-Exposure Prophylaxis (PrEP) in Clinical Settings to Prevent HIV Infection”. English. In: PLoS One 7.7, e40603. doi: 10 . 1371 / journal. pone. 0040603. url: https:// search.proquest.com/docview/1325497230.

Columbia Economic Review | 65

Noah Zwiefel

Banegas, Matthew P. et al. (Jan. 2016). “For Working-Age Cancer Survivors, Medical Debt And Bankruptcy Create Financial Hardships”. English. In: Health Affairs 35.1, pp. 54–61. doi: 10. 1377/hlthaff.2015.0830. url: http://www.ncbi.nlm.nih.gov/pubmed/26733701. Banthin, Jessica S. and Thomas M. Selden (July 2003). “The ABCs of Children’s Health Care: How the Medicaid Expansions Affected Access, Burdens, and Coverage between 1987 and 1996”. English. In: Inquiry 40.2, pp. 133–145. doi: 10.5034/inquiryjrnl_40.2.133. url: https://www.jstor.org/ stable/29773064. Bernick, Ethan M. and Nathan Myers (Aug. 2008). “Treatment or Placebo: Are State Programs Decreasing the Proportion of Uninsured?” English. In: Policy Studies Journal 36.3, pp. 367–384. doi: 10.1111/j.1541-0072.2008.00272.x. url: https://onlinelibrary.wiley.com/ doi/abs/10.1111/j.1541-0072.2008.00272.x. Blewett, Lynn A. (Oct. 2015). Medicaid “Early Opt-In” States. url: https:// www.shadac.org/ news/medicaid-%E2%80%9Cearly-opt-in %E2%80%9D-states. Boudreaux, Michel et al. (Feb. 2011). “Counting Uninsurance and Means-Tested Coverage in the American Community Survey: A Comparison to the Current Population Survey”. English. In: Health Services Research 46.1p1, pp. 210–231. doi: 10.1111/j.1475-6773.2010.01193.x. Brevoort, Kenneth, Daniel Grodzicki, and Martin B. Hackmann (Jan. 2017). Medicaid and Financial Health. English. Tech. rep. w24002. Brevoort, Kenneth et al. (2018). “The Credit Consequences of Unpaid Medical Bills”. In: NBER Working Paper Series. Brown, Clare C., J. Mick Tilford, and T. Mac Bird (May 2018). “Improved Health and Insurance Status Among Cigarette Smokers After Medicaid Expansion, 2011-2016”. English. In: Public Health Reports 133.3, pp. 294–302. doi: 10.1177/0033354918763169. url: http://journals. sage pub.com/doi/full/10.1177/0033354918763169. Bureau, Consumer Financial Protection (2018). Mortgages 30–89 days delinquent. (Visited on ). Bureau, U.S. Census (Dec. 2010). “Weighting and Estimation”. In: American Community Survey Design and Methodology. Washington, D.C.: U.S. Government Printing Office. Campbell, Alexia F. (July 2018). These 2 Medicaid provisions prevent medical debts from ruining people’s lives. url: https://www.vox.com/policy -and-politics/2017/7/19/15949250/ medicaid-medical-bankruptcy. Center, Federal Judicial (2018). Bankruptcy cases filed, terminated, and pending from FY 2008 to present. url: https://www.fjc.gov/research/idb/ bankruptcy-cases-filed-terminated- and-pending-fy-2008-present.// onlinelibrary.wiley.com/doi/abs/10.1111/jrh.12258.

Impact of Medicaid on Bankruptcy Rate

Chavez, Laura J. et al. (2018). “Mental Health and Substance Use Care Among Young Adults Before and After Affordable Care Act (ACA) Implementation: A Rural and Urban Comparison”. English. In: The Journal of Rural Health 34.1, pp. 42–47. doi: 10.1111/jrh.12258. url: https://onlinelibrary.wiley.com/doi/abs/10.1111/jrh.12258. Chokshi, Niraj (Mar. 2014). Historians take note: What America looked like before Obamacare. url: https:// www. washingtonpost. com/ blogs/ govbeat/ wp/ 2014 / 03 / 26 / historians- take- note- what- americalooked- like- before- obamacare/?noredirect=on&utm_term= .036bc75c68d9. Clemens, Michael A and Jennifer Hunt (May 2017). The Labor Market Effects of Refugee Waves: Reconciling Conﬂicting Results. Working Paper 23433. National Bureau of Economic Research. doi: 10.3386/w23433. url: http://www.nber.org/papers/w23433. Courtemanche, Charles et al. (2017). “Early Impacts of the Affordable Care Act on Health In- surance Coverage in Medicaid Expansion and Non-Expansion States”. English. In: Journal of Policy Analysis and Management 36.1, pp. 178–210. doi: 10.1002/pam.21961. url: https: //onlinelibrary.wiley.com/doi/abs/10.1002/pam.21961. Courts, U.S. (2018). U.S. Bankruptcy Courts - Business and Nonbusiness Cases Filed, by Chapter of the Bankruptcy Code, District, and County (Table F-2). url: http://www.uscourts.gov/ report-name/bankruptcy-ﬁlings. Currie, Janet M. and Jonathan Gruber (1995). Health insurance eligibility, utilization of medical care, and child health. Vol. 5052. Cambridge, Mass. Dobkin, Carlos et al. (Feb. 2018). “The Economic Consequences of Hospital Admissions”. English. In: American Economic Review 108.2, pp. 308–352. doi: 10.1257/aer.20161038. url: https://search.proquest.com/ docview/1991174089. Domowitz, Ian and Robert L. Sartain (Feb. 1999). “Determinants of the Consumer Bankruptcy Decision”. English. In: The Journal of Finance 54.1, pp. 403–420. doi: 10.1111/0022- 1082. 00110. url: https://www. jstor.org/stable/222422. Dranove, David and Michael L. Millenson (Mar. 2006). “Medical Bankruptcy: Myth Versus Fact”. English. In: Health Affairs 25.2, w83. doi: 10.1377/hlthaff.25.w74. url: http://content. healthaffairs.org/cgi/ content/abstract/25/2/w74. Fay, S., E. Hurst, and M. J. White (June 2002). “The Household Bankruptcy Decision”. English. In: The American Economic Review 92.3, pp. 706–718. doi: 10.1257/00028280260136327. url: http://www.ingentaconnect.com/content/aea/aer/2002/00000092/00000003/art00014.

Columbia Economic Review | 67

Noah Zwiefel

Fay, Scott A., Erik Hurst, and Michelle J. White (n.d.). “The Bankruptcy Decision: Does Stigma Matter?” English. In: SSRN Electronic Journal (). doi: 10.2139/ssrn.70915. Finkelstein, Amy et al. (Aug. 2012). “The Oregon Health Insurance Experiment: Evidence from the First Year”. English. In: The Quarterly Journal of Economics 127.3, pp. 1057–1106. doi: 10.1093/qje/qjs020. url: http:// www.jstor.org/stable/23251981. Fisher, Jonathan D. (2005). “The effect of unemployment beneﬁts, welfare beneﬁts, and other income on personal bankruptcy”. English. In: Contemporary economic policy 23.4, pp. 483–492. url: http://www. econis.eu/PPNSET?PPN=503405531. Flavin, Patrick (June 2018). “State Medicaid Expansion and Citizens’ Quality of Life”. English. In: Social Science Quarterly 99.2, pp. 616–625. doi: 10 .1111/ssqu.12452. url:https://onlinelibrary.wiley.com/doi/ abs/10.1111/ssqu.12452. Frean, Molly, Jonathan Gruber, and Benjamin D. Sommers (May 2017). “Premium subsidies, the mandate, and Medicaid expansion: Coverage effects of the Affordable Care Act”. English. In: Journal of Health Economics 53, pp. 72–86. doi: 10.1016/j.jhealeco.2017.02.004. url: https://www.sciencedirect.com/science/article/pii/S0167629616302272. Fresques, Hannah and Paul Kiel (Sept. 2017). In the South, Bankruptcy is Different, Especially for Black Debtors. url: https://projects.propublica. org/graphics/bankruptcy- chapter- 13. Fuller, Sirius (Apr. 2018). Using American Community Survey (ACS) Estimates and Margins of Error. Goldstein, Cliff (Sept. 2013). Suspicious Bankruptcy Declarations and Other Ways The Ultra-Wealthy Hide Their Money. url: https://www.nasdaq .com/article/suspicious-bankruptcy- declarations-and-other-ways-theultrawealthy-hide-their-money-cm273088. Gomez, Javier (Feb. 2017). Introduction to the American Community Survey Public Use Microdata Files (PUMS) Files. Gross, Tal and Matthew J. Notowidigdo (31 Jan 2011). “Health insurance and the consumer bankruptcy decision: Evidence from expansions of Medicaid”. English. In: Journal of Pub- lic Economics 95.7, pp. 767– 778. doi: 10 . 1016 / j . jpubeco . 2011 . 01 . 012. url: https :// www.sciencedirect.com/science/article/pii/S0047272711000168. docview/1820778592. Hackney, Donald D., Daniel Friesner, and Matthew Q. McPherson (Jan. 2015). “Do Debtors Have an Obvious Financial Rationale for Filing a Chapter 13 Bankruptcy Petition?” English. In: Economics Bulletin 35.3, pp. 1572–1588. Haley, Jennifer M., Victoria Lynch, and Genevieve M. Kenney (Mar. 2014). The Urban Institute Health Policy Center’s Medicaid/CHIP Eligibility

Impact of Medicaid on Bankruptcy Rate

Simulation Model. English. Tech. rep. Urban Institute. url: https:// search.proquest.com/docview/1820778592. Hill, Steven C. et al. (Apr. 2014). “Adults in the income range for the Affordable Care Act’s Medicaid expansion are healthier than pre-ACA enrollees”. English. In: Health affairs (Project Hope) 33.4, pp. 691–699. doi: 10.1377/hlthaff.2013.0743. url: http://www.ncbi.nlm.nih. gov/pubmed/24670269. Himmelstein, David U., Deborah Thorne, and Stefﬁe Woolhandler (2011). “Medical Bankruptcy in Massachusetts: Has Health Reform Made a Difference?” English. In: American Journal of Medicine, The 124.3, pp. 224–228. doi: 10 . 1016 / j . amjmed . 2010 . 11 . 009. url: https://www. clinicalkey.es/playcontent/1-s2.0-S0002934310009915. Himmelstein, David U. et al. (Jan. 2005). “Illness and injury as contributors to bankruptcy”. En- glish. In: Health affairs (Project Hope) Suppl Web Exclusives, W5. url: http://www.ncbi. nlm.nih.gov/pubmed/15689369. Himmelstein, David U. et al. (2009). “Medical Bankruptcy in the United States, 2007: Results of a National Study”. English. In: American Journal of Medicine, The 122.8, pp. 741–746. doi: 10.1016/j.amjmed.2009.04.012. url: https://www.clinicalkey.es/playcontent/1- s2.0-S0002934309004045. Hollingworth, William et al. (Aug. 2007). “The Risk of Bankrupcy before and after Brain or Spinal Cord Injury: A Glimpse of the Iceberg’s Tip”. English. In: Medical Care 45.8, pp. 702–711. doi: 10.1097/ MLR.0b013e318041f765. url: https://www.jstor.org/stable/40221496. Jacoby, Melissa B. and Mirya Holman (2010). “Managing Medical Bills on the Brink of Bankruptcy”. English. In: Yale Journal of Health Policy, Law, and Ethics, 10.2, p. 239. url: http://www. ncbi.nlm.nih.gov/pubmed/20681437. Kaiser Family Foundation (Jan. 2014). How Will the Uninsured in Wisconsin Fare Under the Affordable Care Act? url: http://www.kff.org/healthreform/fact-sheet/state-proﬁles-uninsured-under-aca-wisconsin/. Kaiser Family Foundation (Nov. 2017). Key Facts About the Uninsured Population. English. Tech. rep. url: https://www.kff.org/uninsured/ fact- sheet/key- facts- about- the- uninsured-population/. Kaiser Family Foundation (Jan. 2018a). Medicaid Income Eligibility Limits for Other Non-Disabled Adults, 2011-2018. url: https:// www. kff. org/ medicaid/state-indicator/medicaid-income-eligibility-limits-for-other -non-disabled-adults/?currentTimeframe=0& sortModel=%5C %7B%5C%22colId%5C%22:%5C%22Location%5C%22,%5C%22 sort%5C% 22:%5C%22asc%5C%22%5C%7D. (Jan. 2018b). Medicaid Income Eligibility Limits for Parents, 2002-2018. url: https://www. kff.org/medicaid/state-indicator/medicaid -income-eligibility-limits-for-parents/?currentTimeframe=0& sortModel=%5C%7B%5C%22colId%5C%22:%5C%22Location% 5C%22,%5C%22sort%5C%22:%5C%22asc%5C%22%5C%7D. (Apr. 2012). States Getting a Jump Start on Health Reform’s Medicaid Expansion. url: https://www.kff.org/health-reform/issue-brief/statesgetting-a-jump- start-on-health/.

Columbia Economic Review | 69

Noah Zwiefel

Karpman, Michael and Kyle Caswell (Mar. 2017). Past-Due Medical Debt among Nonelderly Adults, 2012–15. Tech. rep. Urban Institute. url: https://www.urban.org/research/publication/ past-due-medical-debtamong-nonelderly-adults-2012-15. Kiel, Paul (Feb. 2018). When You Can’t Afford to Go Bankrupt. url: https:// www.propublica. org/article/when-you-cannot-afford-to-go-bankrupt. Kliff, Sarah and Soo Oh (May 2018). America’s health care prices are out of control. These 11 charts prove it. url: https://www.vox.com/a/ health-prices. Kruzikas, Denise T. (2004). Preventable hospitalizations : a window into primary and preventive care, 2000. English. Rockville, MD : U.S. Dept. of Health, Human Services, Agency for Health- care Research, and Quality, color map. url: Google%20http://books.google.com/ books?id= zWBrAAAAMAAJ%20HathiTrust%20Digital%20Library, %20Full%20view%20http://catalog. hathitrust.org/api/volumes/ oclc/58043715.html%20http://purl.access.gpo.gov/GPO/ LPS119313. Leip, David (2018). David Leip’s Atlas of U.S. Presidential Elections. url: https://uselectionatlas. org/. Li, Wenli (Jan. 2007). “What Do We Know about Chapter 13 Personal Bankruptcy Filings?” English. In: Federal Reserve Bank of Philadelphia Business Review, pp. 19–26. url: https: //search.proquest.com/docview /231399779. Mann, Ronald J. and Katherine Porter (Jan. 2010). “Saving up for bankruptcy”. English. In: George- town Law Journal 98.2, p. 289. url: https://search. proquest.com/docview/231579822. McArdle, Megan (Jan. 2017). The Myth of the Medical Bankruptcy. url: https:// www.bloomberg. com/opinion/articles/2017-01-17/the-myth-of-themedical-bankruptcy. McElwee, Sean (Aug. 2016). Enough to Make You Sick: The Burden of Medical Debt. url: http://www.demos.org/publication/enough-make-you-sickburden-medical-debt. Miller, Sarah (Dec. 2012). “The Effect of the Massachusetts Reform on Health Care Utilization”. English. In: Inquiry 49.4, pp. 317–326. doi: 10.5034/ inquiryjrnl_49.04.05. url: https: //www.jstor.org/stable/23480769. Miller, Sarah et al. (Sept. 2018). “The ACA Medicaid Expansion in Michigan and Financial Health”. In: NBER Working Paper Series. Missouri Census Data Center (Sept. 2010). Geocorr 2014: Geographic Correspondence Engine. url: http://mcdc.missouri.edu/applications/ geocorr2014.html. Moore, Brian, Katharine Levit, and Anne Elixhauser (Oct. 2014). Costs for Hospital Stays in the United States, 2012. Tech. rep. Agency for Healthcare Research and Quality. url: https ://www.hcup-us.ahrq.gov/ reports/statbriefs/sb181-Hospital-Costs-United-States- 2012.jsp.

Impact of Medicaid on Bankruptcy Rate

Myerson, Rebecca et al. (Aug. 2018). “Medicaid Eligibility Expansion May Address Gaps in Access to Diabetes Medications”. In: Health Affairs (Project Hope) 37.8, pp. 1200–1207. doi: 10.1377/ hlthaff.2018.0154. Papanicolas, Irene, Liana R. Woskie, and Ashish K. Jha (Mar. 2018). “Health Care Spending in the United States and Other High-Income Countries”. English. In: JAMA 319.10, pp. 1024–1039. doi: 10.1001/jama.2018.1150. url: http://dx.doi.org/10.1001/jama.2018.1150. Pollitz, Karen and Cynchia Cox (Jan. 2004). Medical Debt Among People With Health Insurance. url: https://www.kff.org/report-section/ medical-debt-among-people-with-health-insurance-how-does-medicaldebt-become-a-problem-for-people-with-health-insurance/view/print/. Princeton (2018). Eviction Lab. url: evictionlab.org. Ramsey, Scott et al. (June 2013). “Washington State cancer patients found to be at greater risk for bankruptcy than people without a cancer diagnosis”. English. In: Health affairs (Project Hope) 32.6, pp. 1143– 1152. doi: 10.1377/hlthaff.2012.1263. url: http://www.ncbi.nlm.nih. gov/pubmed/23676531. Report on the Economic Well-Being of U.S. Households in 2017 (May 2018). Tech. rep. Federal Reserve. url: https://www.federalreserve.gov/ publications/ﬁles/2017-report-economic-well-being-us-households201805.pdf. Saloner, Brendan et al. (May 2015). “Most Uninsured Adults Could Schedule Primary Care Appointments Before The ACA, But Average Price Was $160”. English. In: Health affairs (Project Hope) 34.5, pp. 773– 780. doi: 10.1377/hlthaff.2014.1258. url: https://www.ncbi.nlm.nih.gov /pubmed/25941278. Sanger-Katz, Margot (Jan. 2016). ‘I Am Drowning.’ The Voices of People With Medical Debt. url: https://www.nytimes.com/interactive/2016/01/11/ upshot/12up- medicaldebt.html? mtrref=www.google.com. Seifert, R. W. (2006). “Home sick: How medical debt undermines housing security”. In: Louis ULJ 51, p. 325. Selden, T. M., J. S. Banthin, and J. W. Cohen (May 1998). “Medicaid’s problem children: eligible but not enrolled”. English. In: Health Affairs 17.3, pp. 192–200. doi: 10.1377/hlthaff.17.3.192. url: http://www.ncbi. nlm.nih.gov/pubmed/9637975 Shrime, Mark G. et al. (Jan. 2018). “Trading Bankruptcy for Health: A DiscreteChoice Experiment”. English. In: Value in Health 21.1, pp. 95–104. doi: 10.1016/j.jval.2017.07.006. url: https://www.sciencedirect.com/science/ article/pii/S1098301517303273. Sommers, Benjamin D., Katherine Baicker, and Arnold M. Epstein (Sept. 2012). “Mortality and access to care among adults after state Medicaid

Columbia Economic Review | 71

Noah Zwiefel

expansions”. English. In: The New England journal of medicine 367.11, pp. 1025–1034. doi: 10.1056/NEJMsa1202099. url: http://www. ncbi. nlm.nih.gov/pubmed/22830435. Sommers, Benjamin D. and Donald Oellerich (Sept. 2013). “The povertyreducing effect of Medi-caid”. English. In: Journal of Health Economics 32.5, pp. 816–832. doi: 10.1016/j.jhealeco. 2013.06.005. url: https://www.sciencedirect.com/science/article/pii/S016762961300091X. Sommers, Benjamin D. et al. (July 2015). “Changes in Self-reported Insurance Coverage, Access to Care, and Health Under the Affordable Care Act”. English. In: JAMA 314.4, pp. 366–374. doi: 10.1001/jama.2015.8421. url: http://dx.doi.org/10.1001/jama.2015.8421. Sommers, Benjamin D. et al. (Oct. 2016). “Changes in Utilization and Health Among Low-Income Adults After Medicaid Expansion or Expanded Private Insurance”. English. In: JAMA Internal Medicine 176.10, pp. 1501–1509. doi: 10.1001/jamainternmed.2016.4419. url: http://dx. doi.org/10.1001/jamainternmed.2016.4419. U.S. Dept of Health and Human Services and Assistant Secretary for Planning and Evaluation (n.d.). U.S. Federal Poverty Guidelines Used to Determine Financial Eligibility for Certain Federal Programs. url: https://aspe.hhs.gov/poverty-guidelines. U.S. Dept. of Health and Human Services (2018). Prior HHS Poverty Guidelines and Federal Reg- ister References. url: https://aspe.hhs.gov/ prior-hhs-poverty-guidelines-and-federal-register-references. Wachino, Vikki, Samantha Artiga, and Robin Rudowitz (2014). “How is the ACA impacting Med- icaid enrollment”. In: Kaiser Family Foundation, May. Wehby, George L. and Wei Lyu (Apr. 2018). “The Impact of the ACA Medicaid Expansions on Health Insurance Coverage through 2015 and Coverage Disparities by Age, Race/Ethnicity, and Gender”. English. In: Health Services Research 53.2, pp. 1248–1271. doi: 10.1111/1475- 6773.12711. url: https://onlinelibrary.wiley.com/doi/abs/10.1111/1475-6773.12711. Wherry, Laura R. and Sarah Miller (June 2016). “Early Coverage, Access, Utilization, and Health Effects Associated With the Affordable Care Act Medicaid Expansions: A Quasi-experimental Study”. English. In: Annals of internal medicine 164.12, p. 795. doi: 10.7326/M15-2234. url: http://www.ncbi.nlm.nih.gov/pubmed/27088438. White, Michelle J. (Oct. 1998). “Why Don’t More Households File for Bankruptcy?” English. In: Journal of Law, Economics, & Organization 14.2, pp. 205–231. doi: 10.1093/jleo/14.2.205. url: https://www.jstor.org/stable/765103. Zhu, Ning (Jan. 2011). “Household Consumption and Personal Bankruptcy”. English. In: The Journal of Legal Studies 40.1, pp. 1–37. doi: 10 . 1086 / 649033. url: https:// www. jstor. org/stable/10.1086/649033.

Impact of Medicaid on Bankruptcy Rate

Literature Review 1. Uniqueness of Medical Debt There are many reasons behind bankruptcies that range from overconsumption to excessive usage of credit cards (Zhu, 2011). However, an overarching cause of bankruptcy is the inability to cover emergency expenses. A Federal Reserve survey found that only 59% of Americans in 2017 would be able to produce $400 of cash in an emergency (Federal Reserve, 2018). This, combined with a lack of health insurance, are the major drivers of medical bankruptcies. In 2015, one in five adults between 18-64 years old had some form of medical debt (Karpman, 2017). According to the Kaiser Family Foundation (2018), 56% of Americans had private insurance, 36% had public insurance, and 9% were uninsured in 2017. Medical emergencies are unique when compared with other types of emergencies that might incur large costs. Broadly speaking, a broken down car can be repaired later; a damaged washer can be replaced when you can afford it. However, when an individual is facing a life threatening medical condition, they must go to a hospital or risk death. Further complicating the problem, while you can often estimate the price of repairing your car or buying a new washer in advance, you can never be certain what you will be billed when you are admitted to the hospital. Moreover, even if an individual avoids a doctor visit for a symptom, medical issues may become worse (and more expensive) over time. Once an uninsured individual receives the bill, there is no “returning” the health care received. They can choose to pay the bill in full, appeal to charity for a reduction in cost, or utilize a payment plan. Because of these factors, the presence of medical debt is the largest signifier of a potential future bankruptcy filing (Domowitz, 1999). This issue of medical debt only becomes compounded when one considers that uninsured and low-income individuals often avoid preventative care because they cannot afford it (Kaiser Family Foundation, 2017). Low-income individuals with chronic but manageable health conditions are more likely to be admitted to the hospital for something that could have been prevented given primary care intervention (Kruzikas 2004). In this way, medical problems that could have been addressed by a primary care provider for $100$200 (Saloner, 2015) may instead result in thousands of dollars of costs if a hospital stay is required (Moore et al, 2014). Such an expense is enough to push a financially vulnerable individual into bankruptcy. Furthermore, this fails to consider any lost earnings due to missing work during illness, which could be even more detrimental to the poor (Himmelstein, 2005).

Columbia Economic Review | 73

Noah Zwiefel

2. Bankruptcy Strategy The decision to file for bankruptcy is generally thought to be driven by a strategic cost/benefit analysis on the part of the filer (Fay, 2002). Among the costs is a social stigma (Fay et al, 1998; Hackney, 2015), a decrease in credit score, the potential loss of property (Domowitz, 1999), and the shame of needing to appeal to charity in order to afford the nearly $1,500 expense in order to actual file bankruptcy (Mann, 2010; Kiel, 2018). The benefits of filing for bankruptcy include forgiveness of debt and the end of harassment by debt collectors (Domowitz 1999). The costs of filing for bankruptcy have a discernible impact on the choices individuals make with respect to their healthcare; one in twelve individuals in a discrete choice experiment were found to value preventing bankruptcy over having an illness cured (Shrime, 2018). Research suggests that the pressure from debt collectors plays an outsized role in an individuals decision to file for bankruptcy. Individuals might decide against filing for bankruptcy in some cases if they have not faced aggressive debt collection efforts (White, 1998). Mann (2010) supports this through a study of judicial filings and interviews, finding that regardless of the amount of debt, individuals who do not face aggressive debt collection efforts tend not to file bankruptcy when compared with their peers who faced pugnacious debt collectors. This implies that the social costs of bankruptcy are heavily weighted when an individual decides whether or not to file for bankruptcy. 3. Medical Bankruptcy There is a debate regarding what percentage of bankruptcies are actually “medical bankruptcies”. Using survey data, Himmelstein and Warren (2005) suggest that 46.5% of bankruptcies in 2001 were caused by medical costs. This figure is controversial in the literature, with other scholars suchas Dranove (2006)12 and Hackney (2015)13 using slightly different approaches to measuring medical bankruptcies, and finding lower rates14 as a result. Despite this, Himmelstein et al. continue to find consistent percentages of around 50%-60% in survey data, finding 62.1% of bankruptcies in 2007 were caused by medical expenses (2009) and that 52.9% of bankruptcies in Massachusetts are caused by medical costs (2011). Jacoby (2010) proposes that the disparity between survey-based and bankruptcy filing analysis approaches to estimating the medical bankruptcy rate occurs because medical debt owed to care providers is often shifted to more generic payment methods such as credit cards. A 2012 survey found that 47% of low-income households with medical debt had transferred part of it to a credit card (McElwee, 2016). However, this argument has been challenged in recent research from Dobkin et al (2018). Using a novel dataset that combines survey data and credit reports of patients aged 18-64 at California hospitals, Dobkin finds that

Impact of Medicaid on Bankruptcy Rate

approximately 6% of uninsured adults admitted to a hospital proceed to file for bankruptcy as a result. It is important to note that this study only looked at medical events that resulted in a hospital admission, ignoring all other types of visits. Given that many may not be admitted to a hospital after visiting an emergency room, but still incur costly medical bills or have to miss work (and thus lose income), it seems likely that this figure is a lower-bound to the true number of medical bankruptcies. While much of the literature surrounding medical bankruptcies is observational, there are a few quasi-experimental studies which examine how specific medical events change the likelihood of filing for bankruptcy. Both cancer (Ramsey, 2013) and spinal cord injury (Hollingsworth, 2007) increase the likelihood that one will file for bankruptcy; however, Medicaid, Medicare, and Social Security mediate this impact (Ramsey, 2013; Hollingsworth, 2007). Using data on cancer patients from 1995- 2009, Ramsey (2013) finds that being over the age of 65 (and thus eligible for Medicare and Social Security) is associated with a fourfold decrease in the bankruptcy rate, when compared with younger cancer patients15. Similarly, among individuals hospitalized for a brain or spinal cord injury, being on Medicaid is associated with an approximately 50% lower bankruptcy rate, as compared with those who have private or no insurance (Hollingworth, 2007), a drop of approximately 3% in the overall bankruptcy rate. Given these facts, it follows that the ACA Medicaid expansion would decrease the bankruptcy rate, however there has been little direct research to confirm this. 4.I mpacts of Public Health Expansions Public health expansions are socially beneficial. Using an estimate of Medicaid eligibility, Currie (1995) finds that the Medicaid expansions of the 1990s are associated with a decrease in child mortality. Similar effects have been noted as a result of the ACA Medicaid expansion, including a decrease in death, an increase in self-reported health and an increase in subjective well-being among poor and non-white individuals in Medicaid-expanding states (Sommers, 2012; Flavin et al 2018). The ACA Medicaid expansion also increases the likelihood that people will receive preventative care (Wherry, 2016), have better health (Brown, 2018), get needed supplies to manage their diabetes (Myerson et al, 2018), and have access to care, particularly among young adults (Chavez 2018). The impacts of the ACA Medicaid expansion with respect to the uninsurance rate is similarly clear in the literature. The uninsurance rate is lower in Medicaid expansion states than in non- expansion states. (Sommers, 2015; Courtemanche, 2017). More quantitatively, Frean et al. (2017) finds that the ACA Medicaid expansion decreased the uninsurance rate by between 9-19%.

Columbia Economic Review | 75

Noah Zwiefel

This increase in the insurance rate happened regardless of whether or not the Medicaid expansion was driven by an expansion of true public health insurance, as in Kentucky, or whether it was through a private insurance company supported with public dollars, as in Arkansas (Sommers, 2016). Further- more, the Medicaid take up rate increased by between 1-3% post-ACA expansion in non-expansion states (Wehby, 2018), which is potentially due to a new simplified enrollment process implemented nationwide that makes it easier for individuals to apply for Medicaid, regardless of their state of residence (Wachino, 2014). Public health insurance expansions improve the financial well-being of individuals. Sommers (2013) finds that Medicaid decreases the poverty rate by 1%, thus keeping between 2.6-3.4 million individuals out of poverty, with the greatest impact localized among disabled adults, elderly, children, and non-white minorities. Similar effects are seen in government programs at large, with payments directly to individuals through TANF and unemployment benefit being associated with a decrease in bankruptcies (Fisher, 2005). In this regard, public health expansions are no different. The CHIP and Medicaid expansions of the 1990s are associated with an 8% decrease in bankruptcies as a result of a 10% increase in Medicaid eligibility (Gross et al 2011). Unfortunately Gross’s (2011) study is somewhat limited by the state-level nature of Gross’s data, the fact that they do not examine how the impacts vary based on the pre-bankruptcy income of the filer, and the fact that it was published before the ACA Medicaid expansion went into effect. The Massachusetts reform of 2006 lowered out of pocket costs for individuals (Miller, 2012), which aligns with Finkelstein’s (2016) findings that the Oregon Health Insurance Medicaid lottery resulted in a decrease in the probability of having unpaid bills sent to collectors. However, Finkel- stein (2016) did not find a statistically significant reduction in bankruptcies among participants in this experiment; however this study was performed only a year after the expansion. It is possible that changes in bankruptcies lag behind new health insurance legislation, since health insurance expansions do not retroactively cover individuals’ expenses. Individuals in Michigan that were eligible and enrolled in Medicaid were found to have better credit scores (Miller et al, 2018). This generalizes to the U.S. at large, with Brevoort (2017), finding that the ACA Medicaid Expansion decreased the amount of unpaid medical bills by $3.4 billion over two years, increased credit availability for individuals by $520 million a year, and prevented approximately 25,000 bankruptcies a year (Brevoort, 2018). While Brevoort et al (2018) estimate using a difference-in-differences approach that 25,000 bankruptcies a year were prevented by the ACA Medicaid expansion in expansion states, they do not examine the robustness of this estimate, nor do they ensure that this decrease in bankruptcies is localized

Impact of Medicaid on Bankruptcy Rate

specifically to those who become eligible for Medicaid under the ACA, as their focus is on the impacts of the Medicaid expansion on credit availability. To my knowledge, there have been no studies to specifically determine the number of bankruptcies prevented by the ACA Medicaid expansion. Furthermore, the majority of prior analyses of the ACA utilize state level data. While state level data enables an evaluation of overall trends across states, it misses trends that may be present within states. While a state on aggregate may have low bankruptcy and poverty rates, areas within a state may have high bankruptcy and poverty rates. It is these areas that may provide the strongest evidence as to the ramifications of the ACA Medicaid expansion on the bankruptcy rate. In this study, I provide an estimate of the ACA Medicaid expansion’s impact on the bankruptcy rate using county level data. I additionally stratify bankruptcies based on the pre-bankruptcy income of the filer, to ensure that my results are localized to those who actually became eligible for Medicaid. Finally, I perform a series of robustness checks to ensure the accuracy of my results.

Medicaid Simulation I utilize a protocol developed and enhanced by Currie et al (1995), Selden (1998), Banthin (2003), and most recently the Urban Institute (Haley, 2014). I use the American Community Survey (ACS) Public Use Microdata Sample (PUMS) and a Python script to simulate the number of individuals eligible for Medicaid in a public use microdata area (PUMA). This differs slightly from prior approaches which tend to use the Current Population Survey (CPS); however, this approach is supported in the literature (Boudreaux 2011) and is utilized by the Urban Institute (Haley 2014). My script models the income pathway of Medicaid eligibility at a per-person level in my sample. I ﬁrst determine an individual’s age. If they are under the age of 18, I use their family’s income as listed on the household level ACS data to estimate whether or not they are eligible for CHIP and Medicaid under their state’s rules, based on what percent poverty that income falls in for their family size, as deﬁned by the U.S. Dept. of Health and Human Services for that year (2018). If the individual is over the age of 18 and listed as unmarried in the PUMS, I use their own income as listed on the individual level ACS PUMS to calculate whether or not they are eligible for Medicaid. If the individual is married, I use their family income as listed in the ACS PUMS to determine whether or not they are eligible for Medicaid. My output is separated into three categories: children eligible for CHIP, children eligible for Medicaid, and the percentage of adults 18-64 eligible for Medicaid.

Columbia Economic Review | 77

Noah Zwiefel

Finally, I convert the data from a PUMA level to a county level using a geo-spatial crosswalk. I use geocorr 2014 from the Missouri Census Data Center (2010) to build a weighted county-level average of Medicaid eligibility. This is necessary because PUMAs, unlike counties, are a population- based geospatial area. Each PUMA contains between 100,000-200,000 individuals, whereas a county can have anywhere between thousands to millions of people. A PUMA may be composed of multiple counties in rural areas, or a county may be split into multiple PUMAs in urban areas. Furthermore, PUMAs rarely nest nicely within counties, instead slicing seemingly arbitrarily through counties. There are several potential sources of error in this process. The ACS yearly estimates rely on a minuscule sample of approximately 1% of US households (Gomez, 2017) and thus can have a quite large margin of error (Fuller, 2018). I attempt to correct for this using the ACS provided weights (U.S. Census Bureau, 2010), however a slight upwards or downwards bias may remain. Further bias may be introduced during the PUMA to county crosswalk, given that multiple counties may exist in a single PUMA. Despite my attempts to correct this via a geocorrelation weighted average, my estimates may overstate or understate the true percentage eligible for Medicaid depending on the county. Given these weaknesses, I present several alternative tests and speciﬁcations which do not rely on my estimate of Medicaid eligibility to boost the robustness of my conclusions.

Impacts of the ACA Medicaid Expansion on Other Financial Measures To further validate these results, I briefly explore the impact of the ACA Medicaid expansion on several precursors to bankruptcies.16 I focus only on the fixed effects model, given that this model does not rely on the assumption of parallel trends in each of my dependent variables pre-treatment. 1. Eviction Filings Evictions filings are often a precursor for bankruptcy. Under U.S. law, individuals are generally granted an automatic eviction stay if they file for bankruptcy, assuming that their only reason for eviction is non-payment of rent; individuals may thus be induced to file for bankruptcy to stay an eviction. Moreover, medical debt drives housing insecurity, including the eviction rate (Seifert, 2006; Pollitz, 2014). Given these two factors, evictions should decrease in areas where there is an increase in individuals that become eligible for (and enroll in) Medicaid.

Impact of Medicaid on Bankruptcy Rate

I used my fixed effects model with instrumented Medicaid enrollment to estimate the impact of an increase in Medicaid enrollment on the Inverse Hyperbolic Sine of the eviction rate per 100 renters filed within a county (Priceton, 2018). The results of this regression are presented in Table 11. There does, indeed, appear to be a connection between an increase in Medicaid enrollment as a result of the ACA Medicaid expansion and a decrease in eviction filings. This relationship becomes insignificant when clustering standard errors by state, which I attribute to within-state variations in the impacts of an increase in Medicaid on the eviction rate. It is conceivable that some areas within a state might have experienced a large increase in Medicaid eligibility, but there exists a local culture of owning ones’ home, and thus there would be little change in the eviction rate. A more thorough analysis is needed to test this hypothesis. Thus, we have preliminary evidence that the ACA Medicaid expansion decreased eviction filings, substantiating my previous findings; however, further research is needed to confirm the specific magnitude of this impact. 2. Mortgage Delinquency I next examine the impacts of the ACA Medicaid expansion on the rate of mortgages that are 30-89 days delinquent. Individuals who have missed this many mortgage payments may still avert a bankruptcy, but are clearly in financial trouble, making this an clear precursor to bankruptcy. If the ACA Medicaid expansion was preventing bankruptcies, we would expect to see a decrease in the number of homes with delinquent mortgages. To test this hypothesis, I retrieved data from the Consumer Financial Protection Bureau (2018) on the number of homes that were 30-89 days delinquent 17, and then performed my fixed effects analysis. My results are presented in Table 12. We again see a statistically significant relationship between an increase in instrumented Medicaid enrollment and a decrease in mortgage delinquency. This is consistent with what we would expect to find given our prior results; however, further research is needed to confirm the specific treatment effect on mortgage delinquency. 3. Subprime Credit Poor credit is an obvious predictor of impending or future bankruptcy (Domowitz, 1999). Individ- uals with poor credit may have fallen behind on bill payments, or may have had prior bankruptcies, which increases the likelihood of a future bankruptcy. Furthermore, those struggling with medical debt hypothetically have a higher likelihood of missing payments, and prior research has found that the ACA Medicaid expansion resulted in a decrease in subprime credit scores (Antonisse, 2018).

Columbia Economic Review | 79

Noah Zwiefel

In Table 13, I examine the impact of an increase in Medicaid enrollment on the inverse hyperbolic sine of individuals per 1,000 with subprime credit (GeoFRED, 2018). Unfortunately, there is little to be concluded, given the large standard errors. This could be due to unobserved and unmeasurable heterogeneities between counties that may not be captured in year or county fixed effects, or due to an incorrect model specification for analyzing this relationship; I believe that could we control for these factors or structure the model appropriately, we would see a statistically significant relationship, given prior (Antonisse, 2018) research.

Impact of Medicaid on Bankruptcy Rate

Tables and Figures

Columbia Economic Review | 81

Noah Zwiefel

Impact of Medicaid on Bankruptcy Rate

Columbia Economic Review | 83

Noah Zwiefel

Impact of Medicaid on Bankruptcy Rate

Columbia Economic Review | 85

Noah Zwiefel

Impact of Medicaid on Bankruptcy Rate

Columbia Economic Review | 87

Noah Zwiefel

Impact of Medicaid on Bankruptcy Rate

Columbia Economic Review | 89

Noah Zwiefel

THE HEURISTICS OF OBESITY:INFLUENCES ON PHYSICIAN DECISION-MAKING

Marissa Caldwell Wellesley College Abstract: Obesity status, which is based on arbitrary cutoffs along the BMI distribution, may be used heuristically by physicians when making medical decisions. Using a regression discontinuity design, I explore whether there are differences in diagnoses and treatments among individuals close to the obesity cutoff (BMI=30). Women just above the threshold are less likely to be told they are diabetic, have high blood pressure, or have coronary heart disease. Men are more likely to be told they have a heart condition, but are less likely to report being advised to change their behaviors to reduce risk of developing heart disease.

I. Introduction

he ideal interaction between a physician, their patient, and the patient’s resulting medical outcome can be described using a classical economic assumption that the decision-maker – the doctor – is perfectly rational and unbiased. A patient supplies his physician with all pertinent information about their health. The physician, a rational, unbiased decision-maker, uses that information and their medical knowledge to customize the optimal treatment plan for the patient. With the abundance of medical tests now available, a physician is likely able to make well-informed decisions about how to treat patients. . However, this ideal interaction assumes that the physician’s decision is not influenced by any factors other than those which are applicable to the patient’s health, such as preconceived notions about the patient, a recent interaction with another patient, or the profit potential of possible procedures. It is also easy to imagine that a physician may occasionally be unsure because of inconclusive test results, and thus must rely on other types of information to make a diagnosis and determine the treatment path. In fact, research in health and behavioral economics suggests it is inappropriate to model a physician as a purely rational decision-maker: there is substantial variation in physician decision-making, and the optimal outcome is not always achieved (Abuluck et al., 2016; Johnson and Rehavi, 2016; Finkelstein et al., 2014). I seek to expand on this literature by exploring one type of influence on a physician’s treatment recommendations: heuristics, or rules of thumb. by exploring one type of influence on a physician’s treatment recommendations: heuristics, or rules of thumb. Columbia Economic Review | 91

Marissa Caldwell

Heuristics are commonly used in a variety of contexts, from determining financial aid eligibility to deciding a student’s grade in a class, and are simple ways to make decisions more efficiently. Like any form of estimation, heuristics are not perfect and will not always lead to the “best” or “correct” decision. The potential consequences of using heuristics in medical practice make them particularly worthy of investigation. For example, if doctors use an age cutoff such as “age 40” to determine whether a patient should be tested for heart disease, a patient who visits the doctor just a few days before turning 40 may remain untested and untreated, while a patient whose 40th birthday was just a few days earlier may be tested and given a treatment they do not necessarily need. (For example, a test may show a false positive, which may cause the doctor to prescribe a medication with harmful side effects. In some cases, the test itself can be risky and lead to worse health. I define this as “over-treatment”.) In both cases, the heuristic could make a patient worse off and have detrimental health effects. Heuristics are commonly used in a variety of contexts, from determining financial aid eligibility to deciding a student’s grade in a class, and are simple ways to make decisions more efficiently. Like any form of estimation, heuristics are not perfect and will not always lead to the “best” or “correct” decision. The potential consequences of using heuristics in medical practice make them particularly worthy of investigation. For example, if doctors use an age cutoff such as “age 40” to determine whether a patient should be tested for heart disease, a patient who visits the doctor just a few days before turning 40 may remain untested and untreated, while a patient whose 40th birthday was just a few days earlier may be tested and given a treatment they do not necessarily need. In both cases, the heuristic could make a patient worse off and have detrimental health effects. In this paper, I look for evidence of heuristic decision-making around the obesity threshold (BMI=30) using a regression discontinuity design and two nationally representative datasets. There are two key identifying assumptions of the empirical strategy. First, the obesity threshold must be arbitrarily defined. Second, individuals close to the threshold on either side must not have systematically different underlying health and healthcare utilization characteristics; they must differ only in whether they are classified as “overweight” or “obese.” If the underlying health of individuals near both sides of the obesity threshold is very similar, there should be no difference in their health outcomes. In other words, the threshold should not create any systematic differences in health. I hypothesize, however, that in cases in which a doctor faces uncertainty about how to treat her patient, she may rely on the obesity threshold to aid in her decision-making— this creates a jump, or discontinuity, in BMI-associated health outcomes which is only due to the difference in the physician’s decision at the obesity threshold. Notes: Except BMI, all reported means are from indicator variables equal to 1 if true. “Self report in good health” is equal to 1 if the respondent indicated being in good, very good, or excellent health. Universe includes adults age 18 or older.

The Heuristics of Obesity

To test my hypothesis, I use the National Health Interview Surveys (NHIS) from 1980 to 2016 and the Behavioral Risk Factor Surveillance System (BRFSS) from 1987 to 2016. Both the NHIS and BRFSS are cross-sectional datasets that ask for respondents’ height and weight at the time of the survey and collect information about their interactions with the health care system, including health insurance status, frequency of doctors’ visits, and diagnosis of certain medical conditions. The detailed health status information allows me to approximate, with some assumptions, both the respondents’ BMI status at the time they saw a doctor and the subsequent medical decisions the doctor made. This also allows me to predict whether patients with a BMI just under the obesity threshold have different healthcare interactions than patients with a BMI just above the threshold. (The survey-data format, however, imposes some limitations on this causal interpretation, which I address in detail in Section III.) The remainder of the paper is organized as follows: Section II provides background on the body mass index and obesity in the United States, reviews recent literature documenting heuristics in medical decisions, and motivates why the obesity threshold is a plausible heuristic. Section III describes the data sources used in my analysis, and Section IV explains the empirical strategy. Section V presents the results, and Section VI concludes my ﬁndings.

II. Background and Literature Review BMI and Obesity in the United States Body mass index, or BMI, is a function of an individual’s height and weight, and can be calculated two ways:

The calculated value is then discretized into categories based on arbitrary cutoffs to determine obesity status. Individuals with a BMI less than 18.5 are considered underweight; between 18.5 and 25, normal; between 25 and 30, overweight; and 30 and above, obese. (The Body Mass Index (BMI) was developed by statistician Adolphe Quetelet in the 19th century, whose goal was to ﬁnd a relationship between human height and weight that ﬁt a Gaussian distribution. Post-WWII, insurance companies began using it as a determinant of health insurance premiums after noting their heavier clients had higher mortality rates (Eknoyan, 2008). Current cutoffs for the obesity categories were established from observational and epidemiologic studies evaluating BMI and risk of morbidity and mortality (National Institutes of Health, 1998). ) BMI is often used as a proxy for total body fat. However, as a measure of body weight, it may falsely identify a healthier individual with high muscle mass as being obese, and conclude that individuals with low BMI but high fat percentage are healthy. Despite the fact that researchers developed the index by

Columbia Economic Review | 93

Marissa Caldwell

studying predominantly white populations and setting arbitrary obesity cutoffs, the measure is commonly applied to patients of all demographics. It is pertinent to investigate, then, how physicians use BMI and obesity cutoffs in medical decisions and what effects those decisions have on patients. In 2016, 36.5% of U.S. adults were classified as obese, compared to 15% four decades earlier (Ogden et al., 2015; National Center for Health Statistics, 2008). This reflects an upward trend in the average adult BMI in the United States that can be attributed to increases in the average body weight. The probability density plots in Figure 1, constructed using height, weight, and BMI data from the NHIS from 1980 to 2016, highlight this trend: while the height distribution has not changed in each decade since the late 20th century, the weight distribution has shifted right in each consecutive period, leading to an overall shift in the BMI distribution. This rightward shift is concerning given the physical, emotional, and financial costs of obesity. In 2008 dollars, the medical care costs of obesity were estimated to be $147 billion, and productivity losses due to obesity were estimated to be between $3 billion and $6 billion (Finkelstein et al., 2009; Trogdon et al., 2008). These striking figures, combined with the proposed link between obesity and health outcomes, have made obesity a topic of national conversation, and are cited as reasons for why BMI to be integrated into standard medical practice (National Institutes of Health, 1998). Figure 1. Height, Weight, and BMI distribution over time Source: NHIS, 1980-2016 A. Height Distribution, in inches

Note: Epanechnikov univariate kernel density estimation with a bandwidth of 1 and population weights 94

The Heuristics of Obesity B. Weight Distribution, in pounds

Note: Epanechnikov univariate kernel density estimation with a bandwidth of 10 and population weights. The distribution begins at 100 for the 2000s and 2010s because starting in 2000, the minimum weight recorded was 100 pounds. C. BMI Distribution

Note: Epanechnikov univariate kernel density estimation with a bandwidth of 0.25 and population weights. Black and red vertical lines denote overweight and obese cutoffs at BMI=25 and BMI=30, respectively.

Columbia Economic Review | 95

Marissa Caldwell

Modeling Medical Decision-Making Classical economic theory assumes that agents in a transaction have perfect information and act rationally, which leads to the optimal outcome for all participants. As Avorn (2018) points out, this theory is fundamentally flawed in the context of medicine: it is an uncertain environment with high associated costs, and both physicians and patients can act irrationally. In fact, a considerable body of research has revealed substantial variation in physician decision-making when one would expect physicians to reach similar conclusions. There is evidence, for example, that reimbursement incentives drive physicians to recommend C-sections to their non-physician patients (Johnson & Rehavi, 2016). Finkelstein et al. (2014) estimate that half of the geographic variation in healthcare costs is due to physician-specific factors rather than patient demand. Using data on CT scans ordered for Medicare patients visiting the ER, Abaluck et al. (2016) finds that (1) physician experience level is negatively correlated with ordering CT scans for patients, and (2) physicians do not efficiently use the information available to decide whether to order a test: riskier patients are tested less often, while lower-risk patients are tested more often. Acknowledging and investigating the role heuristics play could provide a useful framework for understanding, in part, how clinicians make decisions (For a more detailed framework of heuristic decision-making, see Tversky and Khaneman (1974), who propose three types of heuristics that individuals commonly employ when facing uncertainty.) The biggest concern, though, is not that heuristics are used at all in medicine, but that their improper use can lead to detrimental systematic errors. In an extreme but hypothetical case, a patient who falls just below the obesity cutoff and is not assessed for cardiovascular risk could later suffer from a heart attack that could have been prevented if the doctor had not used the heuristic and tested the patient. The potential for missed diagnoses alone should drive efforts to fully understand how heuristics influence physician-patient interactions and health outcomes. The possible impact these heuristics and guidelines can have on patients’ treatment paths is highlighted by the ease with which they can be changed. For example, in November of 2017, the American College of Cardiology and American Heart Association announced new standards for the assessment and treatment of hypertension. Previously, hypertension was defined as having blood pressure at or above 140/90 mmHg; the change lowered the threshold for hypertension to 130/80 mmHg (Welton et al., 2017). On the one hand, this could allow more patients who would benefit from the treatment to be tested for hypertension, but may have been missed under the previous guidelines. On the other hand, it could lead to over-medication of patients who are perhaps mildly hypertensive or happened to have a higher blood pressure at the time it was measured. Regardless of the policy change, the cutoff itself could lead to 96

The Heuristics of Obesity

under-treatment of individuals just below the cutoff, or over-treatment of individuals just above it. This is the fundamental issue with heuristic decision-making in medicine: if these arbitrary guidelines are strictly adhered to, some patients may be under diagnosed while others are over-treated, leading to nontrivial disparities in overall health outcomes. How can heuristic thinking be avoided? One possible solution, proposed by Thaler and Sunstein (2008), is to use a “nudge,” something that pushes an individual towards an optimal choice without mandating it or limiting her options. Electronic decision support systems could be used in situations where heuristic thinking has proven problematic. Avorn (2018) suggests that putting the best drug as the default setting in a drug order system could nudge physicians towards ordering the “right” prescription. This has its own limitations, though; a drug that may be appropriate for the average patient may not be appropriate for all. These systems also tend to have unnecessarily sensitive warning systems in place to reduce liability, which in turn cause physicians to override most warnings. (Kesselheim et al., 2011). Still, as Coussens (2017) points out, for physicians pressed for time or under substantial uncertainty, technology can be an effective way to reach the optimal outcome while allowing the physician to focus on other aspects of medicine that cannot be automated. Vickrey, et al. (2010) has a simpler proposal: merely making physicians aware of heuristics and providing them with mental exercises to analyze their decisions may be sufficient to identify their biases and limit flawed decision-making. Empirical strategies to estimate the effects of heuristics Despite the prevalence of heuristics in medicine, only recently have econometric methods been applied to determine how they might affect patient outcomes. Since these heuristics are often related to a continuous variable with an arbitrary threshold, they lend themselves well to regression discontinuity (RD) design strategies. The key identifying assumption is that individuals close to an arbitrary threshold on either side are otherwise identical in underlying observable and unobservable characteristics. Successful implementation of this strategy has shown compelling evidence of heuristic decision-making in modern medicine. Coussens (2017) looks for evidence of heuristics in diagnosis of ischemic heart disease (IHD) among ER patients near age 40, since patients above age 40 are thought to be at higher risk. He finds that patients who visit the ER just after their 40th birthday are 0.89 percentage points more likely to be tested for and 0.13 percentage points more likely to be diagnosed with ischemic heart disease than those who are just under age 40. These estimates are nontrivial; they translate to a 10% and 20% increase in testing and diagnosis rates, respectively, suggesting there are patients just below age 40 whose IHD is going undiagnosed and untreated. Columbia Economic Review | 97

Marissa Caldwell

Almond et al. (2010) find differences in mortality rates, costs, and health outcomes for babies born on either side of the Very Low Birth Weight (VLBW) threshold (1500 grams or less). They report that infants born weighing just below 1500 grams, qualifying them for increased medical attention, have lower mortality rates and higher medical costs than infants born just above the VLBW cutoff. The infant mortality rate is estimated to be 1.21 percentage points lower for infants below the cutoff, which translates to a 22% decrease relative to the mean mortality rate of 5.53% for infants within just 3 ounces of the threshold. They estimate this additional care adds $4,000 in hospital costs for infants born below the VLBW threshold relative to average hospital costs of $40,000 just above it. The quality of the hospital may also affect the discontinuity: they find that the reduction in mortality rate just below the threshold is greater in magnitude for hospitals with lower-level care and those lacking an intensive care unit. Not only do these dramatic estimates suggest the treatment for VLBW babies is effective, it also implies that the heuristic can play a significant role in a baby’s survival: there are infants just above the threshold who may have survived had they been given the treatments, rather than being screened out by the heuristic decision-making. Clearly, decisions based on these thresholds can have critical impacts on patients’ health. Still, the current body of research documenting their effects and proposing strategies to reduce their consequences is insufficient, especially compared to the prevalence of heuristics in medicine. Evidence of Obesity as a Heuristic Similar to the Very Low Birth Weight cutoff or age threshold for cardiovascular risk, there is a clear opportunity for the obesity threshold to be used heuristically in medical decision-making. There is little evidence, however, that the current cutoffs predict increased risk for conditions associated with higher BMI (NIH, 2013). In addition, no known studies have used the regression discontinuity (RD) design to examine the effect of the threshold in the clinical setting. Here, I motivate why the obesity cutoff could be a heuristic by examining how it is discussed in the medical field. A patient’s BMI and obesity status are prominent features of medical charts. Beyond this, standard medical practices suggest that obesity thresholds do factor into a physician’s treatment recommendations. The National Institutes of Health have developed clinical guidelines based on obesity cutoffs for assessing patients’ risk for obesity-related diseases, including a treatment algorithm. (See Appendix Figure 1 for the treatment algorithm proposed by the NIH in their Clinical Guidelines on the Identification, Evaluation, and Treatment of Overweight and Obesity in Adults (1998).) For example, if a patient has a BMI greater than or equal to 25, it is suggested that his physician assess for other risk factors (such as coronary heart disease); if his BMI is greater than or equal to 30 the physician is further advised to suggest a weight 98

The Heuristics of Obesity

loss strategy and develop a plan to lower risk factors. However, these are guidelines rather than strict regulations, so there may be substantial variation in the way hospitals, private practices, or specific specialties treat obesity. In fact, there is evidence of differences in the way different fields of medicine treat obesity status. Descriptive studies suggest the categories play a prominent role in the decision for a pregnant woman to deliver via Caesarean section (C-section). Berendzen and Howard (2013) find that overweight and obese pregnant women are more likely to have a C-section than women in the “healthy weight” category, and Abenhaim and Benjamin (2011) find that the decision to deliver via C-section is made at an earlier stage in labor for women with higher BMI. Organ transplants are another area in which obesity may play a role. The Johns Hopkins Comprehensive Transplant Center lists a BMI above the obesity threshold to be a “relative contraindication” for a patient being considered for a lung transplant. Research on adverse outcomes of heart transplants has led to a standard requirement that patients have a BMI of less than 35 in order to qualify (Wever-Pinzon et al., 2015). These medical procedures can be life-saving and are often performed as a last resort, motivating the need to understand if there are important differences in the care and health outcomes of individuals near thresholds. Correlational studies have also repeatedly found positive associations between BMI and morbidity for hypertension, type II diabetes, coronary heart disease (CHD), stroke, osteoarthritis, and some types of cancer (NIH, 1998). In sum, the evidence provides ample opportunity for an obesity heuristic to be employed. If a physician is faced with uncertainty about how to treat a patient suffering from these obesity-related conditions, they may rely on the obesity cutoff. However, a systematic review conducted by the NIH (2013) finds no evidence that the current obesity cutoffs predict increased risk for these conditions relative to other potential cutoffs. This provides perhaps the largest motivation for my study: if there is little empirical evidence that these well-established cutoffs are important, how are they impacting the health outcomes of patients at this threshold?

III. Data and Descriptive Statistics My research question considers whether the obesity status of a patient, particularly if very close to the BMI threshold, inﬂuences the way they interact with the health care system and the types of medical procedures their physician requests. Medical records are the ideal data source to test whether physicians make different decisions for patients at the obesity threshold. Height, weight, and BMI are typically determined at the time of the visit—at the end of the visit, doctors order particular tests. . Later, diagnoses based on those tests and Columbia Economic Review | 99

Marissa Caldwell

treatments can be linked to the BMI measurements. Since I do not have access to medical records data, I rely on nationally representative survey data. In these datasets, individuals are asked about their height and weight at the time of the survey, allowing for a BMI calculation. Information about medical tests and diagnoses, however, is obtained through retrospective questions. This could lead to misclassification of BMI relative to the obesity threshold if patients have significantly gained or lost weight since the time of their diagnosis. This means there is likely measurement error in my BMI and obesity measures relative to the information that the physician sees when making their decision. This will be discussed in more detail below as I describe the datasets I employ; the consequences of the measurement error for my findings will be addressed. The two data sets I use are the National Health Interview Survey (NHIS) and the Behavioral Risk Factor Surveillance System (BRFSS). NHIS I use the National Health Interview Survey (NHIS) from 1980 to 2016 for the majority of my analysis. The NHIS is an individual-level, cross-sectional dataset that provides a wide range of health and demographic information for adults age 18 and above in the United States. Each year of the sample contains approximately 100,000 individuals from 42,000 households. I use this dataset to obtain individuals’ height, weight, BMI, demographic characteristics, and health care interactions. I focus on health outcomes most likely correlated with BMI: diabetes, hypertension, coronary heart disease, angina, heart attack, and “other heart conditions”. (Defined by the NHIS to be any condition other than angina, coronary heart disease, or heart attack.) Respondents report their height in inches and weight in pounds. For BMI, I use a calculated variable, available through the NHIS, that is constructed using the height and weight data. The health outcomes data used is imperfect in that it is generally retrospective and asks respondents whether a doctor told them they have a particular condition, not when or if they were tested. While some survey questions do ask if the patient currently has a particular condition, this may only resolve measurement error if the individual was recently diagnosed. Additionally, many of these questions are asked in only a few years of the survey and have a small sample size—I therefore do not use them in my analysis. (For more information on the health outcomes variables, see Appendix Table 1. ) I assume that the respondents’ BMI is the same as their BMI at the time they were told they had a particular condition. This assumption has a considerable limitation in that respondents may have changed their weight, and therefore BMI, as a result of having been told they have a certain BMI-related health condition. This would make it more likely to find individuals with a condition below the threshold than 100

The Heuristics of Obesity

individuals above it. This limitation impacts the interpretation of my results, which is discussed in section V. A second limitation is that this data cannot isolate a physician’s decision to order a test. For example, one survey question asks whether a physician has ever told the respondent she is diabetic. An afﬁrmative response means that (1) the respondent visited a doctor, (2) the doctor tested the respondent for diabetes, and (3) the respondent has diabetes. Someone with a BMI just above the obesity threshold could affect the ﬁrst two points: individuals who know they are categorized as obese may make a different decision about seeing a doctor—perhaps because they think the doctor will only talk about their weight. Or, conditional on seeing a doctor, the doctor may use the heuristic and make different decisions about testing for obesity-related diseases. Regarding the third point, whether a patient actually has diabetes should not differ based on whether his BMI is just below the cutoff, such as 29.9, versus at or just above the cutoff at 30. I present evidence that underlying health, health insurance status, and visits to the doctor do not differ for those just below and just above the obesity threshold. Thus, I will interpret differences in being told one has diabetes on either side of the threshold as coming from differences in being tested, even though it is the combined effect of going to the doctor, the doctor ordering a test, and the patient being diagnosed. Table 1 shows demographic statistics of the NHIS sample. All analyses use population weights and are separated by gender. I drop observations that have a BMI less than or equal to 10, or greater than or equal to 65. Of note, 25.5% of women and 41.9% of men are overweight, whereas 17.7% of women and 16.9% of men are obese over the entire period. Final sample sizes for my outcomes of interest vary because the questions in the survey were not consistent across the years. Table 1 also reports the sample means of the health outcomes I investigate. A higher percentage of women in my sample have health insurance than men, and a higher percentage of men report being in good health (86.1% and 88.0% for women and men, respectively). (“Self report in good health” encompasses respondents who rank themselves as generally being in good, very good, or excellent health on a Likert scale ranging from poor to excellent.) A higher proportion of men are diabetic and have or had angina, coronary heart disease, or a heart attack; heart disease and heart attack rates are almost double the rates for women. More women than men are hypertensive or have another heart condition.

Columbia Economic Review | 101

Marissa Caldwell

Notes: Except BMI, all reported means are from indicator variables equal to 1 if true. “Self report in good health” is equal to 1 if the respondent indicated being in good, very good, or excellent health. Universe includes adults age 18 or older. BRFSS

The Behavior and Risk Factor Surveillance System (BRFSS) is another individual-level, cross-sectional dataset that provides similar health outcomes data as the NHIS, as well as additional behavioral health information. Roughly 400,000 individuals are surveyed by phone each year from all 50 states, the District of Columbia, and three U.S. territories. I use all available data for adults age 18 and above from 1987-2016. As with the NHIS, respondents report their height and weight only, not their BMI. I focus my analysis on six health outcomes: diabetes, high blood pressure, coronary heart disease (CHD), heart attack, and physicians’ advice to change diet or exercise to reduce the risk of developing heart disease. Detailed information about the survey questions used to create these outcomes is available in Appendix Table 2. The BRFSS faces the same limitation as the NHIS in that the observed BMI might be different from what the physician observed when making a decision, leading to bias in the estimated effect of crossing the obesity threshold. Table 2 shows descriptive statistics for adult women and men from the BRFSS sample. Again, all analyses are separated by gender and weighted by population. Note that race is categorized differently in the BRFSS than in the NHIS. Individuals with a BMI less than or equal to 10 or greater than or equal to 65 are dropped from the sample.

102

The Heuristics of Obesity

27.9% of women and 43.0% of men in the sample are overweight, while 21.3% of women and 21.9% of men are classiﬁed as obese. Similar to the NHIS, more women have health insurance than men (87.0% versus 86.5%), and more men report being in good health than women (85.0% versus 83.0%). Prevalence of diabetes and high blood pressure is about 8.0% and 20.1% respectively for both men and women. Heart attacks and coronary heart disease rates are higher in men than women, but more women than men report being told to change their diet or exercise to reduce heart disease risk.

IV. Empirical Design A. Identiﬁcation Strategy The four major obesity categories—underweight, normal, overweight, and obese—are determined solely by thresholds along the BMI index, which is a continuous variable that can take on any value according to an individual’s height and weight. While BMI may be correlated with health and other individual-speciﬁc characteristics, it is unclear how the obesity categories are used in both patients’ and physicians’ healthcare decisions. However, because these cutoffs are arbitrary, whether an individual’s BMI falls just above or just below one of these arbitrary cutoffs is essentially random. I exploit the random variation near these cutoffs with a regression discontinuity (RD) design to examine whether there is a discontinuity, or a jump, in health outcomes for individuals on either side of the cutoff. ( See Lee (2008) for more information on random variation in an RD design.)

Columbia Economic Review | 103

Marissa Caldwell

More specifically, I focus on the obesity threshold at BMI=30 to look for evidence of heuristic decision-making based on a patient’s proximity to being classified as obese. This method is similar to the methods employed in previous studies of heuristics (Bharadwaj & Neilson, 2011; Almond, Doyle, Kowalski, & Williams, 2008; Coussens, 2017). Formally, the regression specification I use is as follows: The dependent variable Health Outcome is a binary variable that is equal to one if the individual i was told she has a particular condition or had a particular procedure, and zero otherwise. On the right-hand side, Cutoff is an indicator variable equal to 1 if the individual has a BMI greater than or equal to 30, and BMI is a continuous variable representing BMI at the time of the survey, scaled to 0 at BMI=30. BMI*Cutoff allows the slope of the regression to vary on either side of the cutoff, and X is a vector of individual-level characteristics including age, race, ethnicity, and education. The coefficient of interest is β, which describes the difference in the outcome variable for individuals just above the obesity threshold compared to individuals just below the threshold. Causal interpretation of the estimated coefficient βˆ rests on three assumptions. First, the assignment variable must be balanced on either side of the cutoff. Since respondents from the NHIS and BRFSS reported their height and weight separately and not their BMI, they would not have had exact control over their reported BMI. It is unlikely that they were able to sort non-randomly onto one side of the obesity cutoff. I check that the density of the BMI distribution across the full sample of the NHIS (Figure 2) does not show evidence of hollowing out at the obesity threshold. While the distribution is not entirely smooth, it does not appear to be disproportionately jagged at BMI=30. ( Since individuals have a unique combination of height and weight (and therefore a unique BMI), one would expect the BMI distribution to be smooth. However, since this distribution reflects self-reports of height and weight, the jaggedness may be due to whole number rounding of height and weight values. For example, a BMI of 27.1 corresponds to a height of 6 feet and a weight of 200 pounds. Restricted to men, there are 10,518 observations with a BMI of 27.1 in the NHIS, but only 4,598 observations at 27.2, indicating that men may be disproportionately rounding to “6 feet, 200 pounds” in their survey response.) This plot supports the first assumption that BMI is smooth across the obesity cutoff and that individuals are not able to manipulate whether they are on one side of the cutoff or the other. Thus, conditional on BMI, I treat whether one is on one side of the arbitrary cutoff or the other as random.

104

The Heuristics of Obesity

Figure 2: BMI Distribution, full sample

Î˛

Columbia Economic Review | 105

Marissa Caldwell

may make it difficult to identify an effect (if there is one). Second, individuals’ BMI could change as a result of having been told, for example, they are diabetic, and this could put them on a different side of the obesity cutoff. Then, the outcome variable reflects not only going to the doctor and being tested for it, but also the effect on BMI of being diagnosed with the condition. If people tend to lose weight when they are told they have diabetes, I would be more likely to find people with diabetes below the obesity threshold. Thus, for my analysis, I interpret my results assuming BMI does not vary systematically over time in response to being told that one has a particular condition, but I discuss the implications if it does. As a final note, it is important to recognize that the effect I expect to find may be small given that there may only be a minority of cases in which the physician faces enough uncertainty to rely on heuristics. To mitigate this issue, I use many years of data to increase the sample size and power. B. Methodology The specification listed in (1) implies a linear relationship between the dependent variables and BMI. However, it is possible this relationship is nonlinear and differs depending on the outcome variable. This is critical because an incorrect specification could mechanically produce a discontinuity when there is none. The stylized example shown in Figure 3 illustrates this—imagine there is a cubic relationship between BMI and a health outcome that has an inflection point at the obesity threshold. A linear fit would estimate a discontinuity at the BMI threshold as shown in the first figure, while a cubic specification, shown in the second figure, would accurately depict the relationship between BMI and the outcome without predicting a discontinuity at the threshold. In order to avoid this error, it is important to visually examine the results of various specifications. I look for a functional form that fits the data well and is relatively easy to replicate. Figure 3: Functional Form and Predicted Discontinuities in a RD Design

106

The Heuristics of Obesity

Here, I describe my method for finding a robust functional form to estimate β or the discontinuity in health outcomes at the BMI threshold. While I am interested in the differences in health outcomes for diabetes, C-sections, hypertension, and other aspects of cardiac health, I use NHIS data on individuals’ reports of being told they are diabetic as a model for finding a robust functional form. I choose this variable for two reasons: diabetes is commonly discussed as a correlate of obesity, and it affects both men and women. First, I use an ordinary least squares (OLS) specification to examine the correlates of diabetes, starting with BMI and adding the demographic variables that have been associated with increased risk of diabetes. I then explore the functional form relationship between BMI and diabetes using specification (1) with the full sample and restricted bandwidths. I choose the final specification

Notes: Outcome variable is an indicator equal to 1 if a respondent has been diagnosed with diabetes. Race and ethnicity include dummy variables for white, black, Asian, other, and Hispanic. All regressions use population weights.

Columbia Economic Review | 107

Marissa Caldwell

by comparing estimates of the discontinuity from different functional forms across similar bandwidths. All analyses separate men and women. (Physicians may use heuristics different between their male and female patients, as Coussens (2017) shows. Thus, I separate men and women for all analyses.) Next, I explore the functional relationship between BMI and diabetes with linear, quadratic, and cubic variations of BMI in specification (1). BMI is scaled to 0 at BMI=30, the obesity cutoff, and all specifications include interaction terms between BMI and the cutoff to allow the slope of the curve to vary on either side of the threshold. Table 4 shows the results of these regressions using the full NHIS sample. βˆ, or the estimates of the discontinuity at the cutoff, are inconsistent in magnitude and sign for both men and women and for specifications with and without covariates. For example, the linear specification

Notes: BMI is scaled to 0 at the cutoff, BMI=30. All regressions include an interaction between the cutoff and BMI to allow the slope to vary on either side of the obesity threshold and include the full BMI distribution. Outcome variable is an indicator equal to 1 if a respondent has been diagnosed with diabetes. Each column of each panel represents a different functional form of BMI: linear (BMI), quadratic (BMI2), or cubic (BMI3), with or without covariates. Regressions with covariates include age, age2, and age3, as well as indicators for race, ethnicity, and level of education (high school or above). All regressions use population weights. 108

The Heuristics of Obesity

for women with covariates (Table 4, Panel A) suggests that women with a BMI at least equal to the obesity cutoff are 1.32 percentage points more likely to be told they are diabetic, while the quadratic specification suggests they are 1.07 percentage points less likely to be told they are diabetic with both significant at the 1% level. One explanation for the inconsistent estimates is that the regressions above estimate the cutoff effect using the entire BMI distribution, even though the region of interest is just around the obesity cutoff. Therefore, I re-estimate the discontinuity with the same linear and quadratic specifications, but limiting the bandwidth or the region of the assignment variable to within 3, 4, or 5 units above and below the cutoff at BMI=30. (As a starting point to determine bandwidths, I used the rdbwselect (regression discontinuity bandwidth selection) Stata package from Calonico, Cattaneo and Titiunik (2014), which calculates the optimal bandwidth by minimizing the sum of squared errors. The calculated optimal bandwidth for these specifications varied, but tended to be in the range of 2 to 7 units away from the cutoff. I tested 3, 4, and 5 in hopes of improving precision without violating the assumption that individuals included in the regression are essentially identical.) The results of these regressions, shown in Table 5, give more consistent estimates for the linear and quadratic specifications with and without the controls added. For women, the coefficient on the cutoff is negative and statistically significant across nearly all bandwidth and functional form specifications. Columns (7) and (10) in particular, which are linear and quadratic specifications at a bandwidth of 3 respectively, have similar coefficients that are significant at the 10% level. For men, none of the estimates are statistically significant. This makes it difficult to determine the best functional form. Visually, Figure 4, which shows the linear specifications for men and women, confirms that the relationship between diabetes diagnosis and BMI can be estimated linearly with a bandwidth of three sufficiently for men and women. Thus, my preferred specification for the health outcomes of interest is identical to (1) but limited to a bandwidth of 3 on each side of the BMI cutoff.

Columbia Economic Review | 109

Marissa Caldwell

Notes: BMI is scaled to 0 at the cutoff, BMI=30. All regressions include an interaction between the cutoff and BMI to allow the slope to vary on either side of the obesity threshold. Outcome variable is an indicator equal to 1 if a respondent has been diagnosed with diabetes. Each column of each panel represents a different test of functional form, bandwidth, and covariates: linear or quadratic and a bandwidth of 3, 4, or 5 on either side of the cutoff. Regressions with covariates include age, age2, and age3, as well as indicators for race, ethnicity, and level of education (high school or above). All regressions use population weights.

110

The Heuristics of Obesity

Columbia Economic Review | 111

Marissa Caldwell

Figure 4: Visual Check of Preferred SpeciďŹ cation A. Women

B. Men

Notes: Plots show the relationship between the residuals of the outcome regressed on the covariates in speciďŹ cation (1) at a bandwidth of 3 and BMI. Dots are the averages within each bin.

112

The Heuristics of Obesity

C. Balancing Tests As discussed in Section IVA, causal interpretation of βˆ relies in part on the assumption that the sample is balanced across the cutoff on characteristics that may be correlated with obesity and the outcome variables of interest—in other words, the threshold is not correlated with the error term. In theory, since the obesity cutoff is arbitrary, it should not be correlated with demographic characteristics in the error term, including age, race, ethnicity, or education. Even if these observable characteristics are not balanced across the threshold, they can be removed from the error term by controlling for them in the regression. The internal validity of the estimate βˆ is threatened more if the variables related to healthcare interactions are also discontinuous at the obesity threshold. In this case, the interpretation of a discontinuity could not be causally attributed to heuristic decision-making. To explore this further, imagine that individuals who know they are classified as obese tend to think of themselves as being less healthy. Or, perhaps insurance companies are able to discriminate such that obese individuals are less likely to have health insurance compared to those just barely below the obesity threshold. In these cases, an estimate of a discontinuity at the threshold in a health outcome such as heart disease diagnoses would capture the physician’s heuristic decision—making as well as the patient’s tendencies and ability to go to the doctor. However, if it is clear that there are no discontinuities in underlying health status or insurance coverage, then I can more confidently attribute a discontinuity in health outcomes at the threshold to heuristic decision-making. To address these threats to validity, I use equation (1), with a bandwidth of three and without additional controls, to examine whether the sample is balanced across the obesity threshold in self-reports of health status, health insurance coverage, and demographic characteristics. I also test whether there are underlying differences in the frequency of visiting the doctor, although this test could be considered an outcome rather than a control since people who are told they are obese versus not may respond by going to the doctor less (or more). Table 6 shows the results of these regressions for women and men. I first use the same linear specification used for diabetes, discussed above, and find no significant discontinuity in health insurance coverage for women and men (column 1 of panels A and B). The coefficient for the obesity cutoff is insignificant as an indicator for women’s self-reports of good health, but significant at the 10% level for men (column 2). However, this coefficient becomes insignificant when a quadratic specification is used instead. Figures 5 and 6 give a visual approximation of these discontinuities and affirm the regression results.( Note that the intercepts and the predicted discontinuities differ slightly from the results reported in Table 6 because the plots are not constructed with population weights.) This shows there are not any observable underlying differences in health at the threshold, and no differences in access to Columbia Economic Review | 113

Marissa Caldwell

care that would lower the probability of being diagnosed with a disease. I also find no evidence of underlying differences in frequency of interactions with physicians; a binary variable indicating if it has been more than five years since the respondent’s last visit to the doctor is insignificant for both men and women (Panel A, Column 3; Panel B, Column 4). Thus, my interpretation of βîs not invalidated. An indicator for white race shows significant discontinuities for men and women with different polynomial terms, which may indicate a different functional relationship between this variable and BMI (Columns 5-7 and 6-8 of Panels A and B, Table 6). While this nonlinearity and lack of balance at the threshold is a threat to the validity of my causal interpretation, it is difficult to understand what kind of underlying characteristics it would lead to, for example, a jump in the fraction of white women just over the obesity threshold. It could be that BMI has a nonlinear relationship with race in this bandwidth; however, for my analysis I am more concerned that the functional form is the right one for the health outcome variables (such as diabetes). When race and other demographic variables are used as controls (shown in the previous section), a linear versus quadratic specification for diabetes versus BMI does not meaningfully change the estimate of the discontinuity. By controlling for these observable characteristics in the regression, only unobservable factors in the error term that are correlated with the obesity threshold are potential sources of bias in the estimated discontinuity.

114

The Heuristics of Obesity

Notes: BMI is scaled to 0 at the cutoff, BMI=30. All regressions include an interaction between the cutoff and BMI to allow the slope to vary on either side of the obesity threshold. Regressions use the linear, quadratic, or cubic regression discontinuity speciďŹ cations with respect to BMI, without any additional covariates. All outcomes except age are indicator variables equal to 1 if true. Self-report in good health indicates if respondents think they are in good, very good, or excellent health. All regressions use population weights and a bandwidth of 3 on either side of the cutoff.

Columbia Economic Review | 115

Marissa Caldwell

Figure 6. Visual Balance Test, NHIS Men

Note: Plots show the ﬁtted values of the outcome and BMI using speciﬁcation (1) with a bandwidth of 3, similar to column (2) of Table 6A and column (3) of Table 6B. The intercepts and direction of discontinuity differ from the regression results because the plot is constructed without population weights. Dots are the averages within each bin.

V. Results A. NHIS I focus my analysis from the NHIS on six health outcomes likely to display evidence of heuristic decision making at the obesity threshold: diabetes, hypertension, angina, coronary heart disease (CHD), heart attack and other heart conditions (besides angina, coronary heart disease, or heart attack). Table 7 presents OLS regressions of these outcome variables on BMI using the full sample of data, along with demographic controls. As expected, for both men and women, there is a statistically signiﬁcant, positive relationship between all six health outcomes and BMI. For example, a one-unit increase in BMI is estimated to increase likelihood of having diabetes by about 0.8 percentage points for both men and women. The correlation between these outcomes and BMI appear slightly higher for men than women, but are very similar in magnitude.

116

The Heuristics of Obesity

Notes: In all regressions, the outcome is regressed on BMI with covariates added. All outcomes are indicator variables equal to 1 if true. Covariates include age, age2, and age3, as well as indicators for race, ethnicity, and level of education (high school or above). All regressions use population weights.

Table 8 shows the results of the regression discontinuity specification (Equation 1) with a bandwidth of three, for women and men. For women, the estimate of the discontinuity is only marginally significant for diabetes. It is also opposite in sign and magnitude from what would be expected given that the OLS regression shows a higher likelihood of being told one is diabetic as BMI increases. Women whose BMI places them just beyond the obesity cutoff are 0.7 percentage points less likely to have been told they are diabetic, or 7% relative to the mean across the specified bandwidth (Panel A). Figure 7A shows this discontinuity visually. Note that this image is identical to the one produced in Figure 4. This negative coefficient could reflect that women just over the threshold are being tested more often but do not have the condition, that they react to being diagnosed with diabetes by losing weight, or are less likely to get tested. The first explanation can be disregarded because the balance tests show no underlying difference in health across the threshold. The third explanation, also, is improbable; it seems unlikely that the physician would test a patient less often if she is at higher risk.

Columbia Economic Review | 117

Marissa Caldwell

The second explanation, however, is consistent with heuristic decision-making by the physician. To illustrate, if a physician is more likely to test for diabetes if a woman is above the threshold, they are more likely to diagnose her with diabetes; the physician may then advise the patient to lose weight, or the patient may lose weight on her own accord, lowering her BMI and placing her on the other side of the threshold by the time she is observed in the data. (This may be true in particular for patients with diabetes; some medications for Type II diabetes treatments have been associated with weight loss that can last for over ten years (Diabetes Prevention Program Research Group, 2009). For men, I find positive discontinuities at the BMI cutoff for coronary heart disease and other heart conditions that are significant at the 10% and 1% levels, respectively. Being just over the obesity threshold is associated with a 0.5 percentage-point increase in likelihood of being diagnosed with CHD and a 0.8 percentage-point increase in the likelihood of being diagnosed with a heart condition other than CHD, angina, or heart attack, which respectively correspond to 9% and 11% increases relative to the mean around the threshold. Figures 7B and C show the discontinuities visually. These estimates are consistent with heuristic decision-making: given that men may be at high risk for cardiac health issues, physicians may be even more inclined to test and diagnose male patients for heart conditions if they are obese. The lack of a significant discontinuity for angina or heart attack diagnoses may reflect that these conditions may be more easily diagnosed because they have more distinguishable symptoms, such as chest pain, so physicians may not need to resort to heuristics as often. While speculating on why discontinuities are found for different measures for men and women is beyond the scope of this analysis, it is worth considering why the observed discontinuities are in different directions. The results may merely expose the limitations of the dataset, since an unknown amount of time passed between diagnosis and observed BMI, and the timeframe may differ with diagnosis. Alternatively, the negative discontinuity for women versus the positive discontinuity for men may indicate that men and women respond differently to diagnoses: women may be more motivated than men to lose weight once diagnosed, putting them on the other side of the cutoff by the time of the survey, and creating the negative discontinuity observed here. A third but related possibility is that treatments for heart conditions and diabetes have differential effects on weight, which would again account for the discontinuities in opposite directions. Hypertension, which appears to be unaffected by the obesity threshold, is an interesting case because it is defined by another cutoff: a blood pressure reading of 130/80 mmHg (or 140/90 mmHg under older guidelines). A physician may diagnose a patient with hypertension as soon as she sees blood 118

The Heuristics of Obesity

pressure reading at this value or higher, and may rarely need to consider other factors. This would make it harder to ďŹ nd an effect at the obesity threshold, consistent with my ďŹ ndings.

Notes: BMI is scaled to 0 at the cutoff, BMI = 30. All regressions are linear with a bandwidth of 3 on either side of the cutoff, and include an interaction between the cutoff and BMI to allow the slope to vary on either side of the obesity threshold. Outcome variable is an indicator equal to 1 if true. Covariates include age, age2, and age3, as well as indicators for race, ethnicity, and level of education (high school or above). All regressions use population weights.

Columbia Economic Review | 119

Marissa Caldwell

Figure 7. Visual Representation of SigniďŹ cant RD Results A. Women: Diabetic

B. Men: Coronary Heart Disease (CHD)

120

The Heuristics of Obesity

B. BRFSS I use the BRFSS dataset to test additional obesity-related health-related outcomes. Additional Health Outcomes I test six outcomes from the BRFSS for evidence of heuristic decision-making. Four of the outcome variables, diabetes, high blood pressure, coronary heart disease (CHD), and heart attack, are constructed from questions similar to the NHIS, allowing me to test for reproducibility. Two additional variables provide information on whether a doctor has told a patient to change their diet or exercise in order to reduce their risk of heart disease. These two outcomes are of particular interest since they may be the most sensitive to heuristics: if a physician sees a patient with a BMI over the obesity threshold, it is reasonable that she may suggest adjusting diet or exercise to lower risk of weight and BMI-related health concerns, regardless of whether the patient actually has a condition. I first use an OLS specification, as with the NHIS, to see the correlation between the health outcomes and BMI using the full sample. (See Appendix Table 4 for OLS results. OLS estimates are similar in magnitude and direction to the NHIS). Then, I run specification (1) with a bandwidth of three. (Note that the BRFSS provides age as a categorical rather than continuous variable, so there are no age2 and age3 terms in these regressions.) Next, I present the regression discontinuity results, reported in Table 9. There are a few key observations to note about these RD results. First, the estimate of the discontinuity in diabetes at the obesity threshold is almost identical to the estimate obtained from the NHIS: -0.00746 versus -0.00702 from the NHIS, and is statistically significant at the 1% level. The coefficient indicates that individuals whose BMI is just over the obesity threshold are 0.746 percentage points less likely to be told they have diabetes than those just under the BMI cutoff, a 6% increase relative to the mean across the threshold (Panel A, column 1). Again, the negative discontinuity may be consistent with heuristic decision making if women diagnosed with diabetes tend to lose weight and are on the other side of the BMI threshold when their BMI is observed in the data. There are a few differences among the estimates for women between the NHIS and BRFSS. While a positive, insignificant discontinuity is estimated for hypertension in the NHIS, the BRFSS data suggests a negative, significant discontinuity of 0.434 percentage points, implying a 1.5% decrease relative to the mean. While the NHIS did not exhibit a significant discontinuity for coronary heart disease, the BRFSS estimates a 0.229 percentage-point, or 5.0%, decrease in the likelihood of being diagnosed with CHD if above the obesity threshold. Again, this could potentially reflect a tendency of women to react to obesity-related diagnoses by losing weight.

Columbia Economic Review | 121

Marissa Caldwell

The results for men, reported in Panel B, are similar to those in the NHIS. Men over the threshold have a 0.5 percentage-point increase (about 2% relative to the mean) in the probability of having high blood pressure, which is statistically significant at the 5% level; they are also 0.294 percentage-points more likely to be diagnosed with coronary heart disease, which is a 5% increase relative to the mean and significant at the 1% level. The most intriguing results are that the discontinuities for exercise and diet outcomes (columns 5 and 6, panel B) are statistically significant and negative. Being at or above the obesity threshold is associated with a 3.19 percentage-point (or 7%) reduction in the likelihood of being told to exercise to reduce heart disease risk and a 1.34 percentage-point (or 4%) reduction in the likelihood of being told to alter diet. Through the framework of heuristic decision-making, this result seems surprising: one might expect physicians to recommend diet or exercise more often to patients that are above the obesity threshold. However, if a doctor does recommend weight loss and the patient subsequently heeds the advice, the observed BMI may be below the threshold. Again, this is a limitation of the data because I cannot observe a patient’s exact BMI at the time of the encounter with the doctor.

122

The Heuristics of Obesity

Notes: BMI is scaled to 0 at the cutoff, BMI=30. All regressions are linear with a bandwidth of 3 and include an interaction between the cutoff and BMI to allow the slope to vary on either side of the obesity threshold. Outcome variable is an indicator equal to 1 if true. Covariates include age (categorical) and indicators for race, ethnicity, and level of education (high school or above). All regressions use population weights.

VI. Conclusion BMI and obesity categories feature prominently in medical charts, yet relatively little is known about how doctors use this information to treat patients. It is unknown whether obesity thresholds lead to discontinuous treatment and health outcomes of individuals near these cutoffs. Here, I explore whether patients on either side of the obesity threshold at BMI=30 are more or less likely to have been diagnosed with diabetes, hypertension, or other cardiac conditions. I also explore whether doctors are more or less likely to suggest diet or exercise to reduce heart disease risk for patients at this threshold. I ďŹ nd evidence that women just above the obesity threshold are signiďŹ cantly less likely to be told they are diabetic, have high blood pressure, or have coronary heart disease. Men just above the obesity threshold are more likely to be told they have a heart condition, but are less likely to report being told to change their diet or exercise habits in order to reduce their risk of developing heart disease. Columbia Economic Review | 123

Marissa Caldwell

These results suggest that physicians may use the obesity threshold at BMI=30 to decide how to treat their patients. This implies there may be stark differences in health outcomes on either side of the cutoff, even when underlying differences in health are negligible. Given the nature of the data sources, it is difficult to interpret why some health outcomes show a positive discontinuity and some show a negative discontinuity at the obesity threshold. Negative discontinuities may reflect either that individuals at the threshold are being tested just as often or more, but are less likely to have the condition. This explanation is unlikely, given that the underlying health across the threshold appears similar. Instead, since BMI is observed after the diagnosis, negative discontinuities could reflect that people who were obese, tested, and diagnosed subsequently lost weight (perhaps as part of their treatment), putting them on the lower side of the threshold. This could explain the negative discontinuities associated with diagnoses for women and diet and exercise advice for men. For example, after being told they were diabetic (and perhaps that obesity played a role), women may have lost weight and their obesity categorization may have changed. The positive discontinuities are consistent with a model in which physicians test patients above the obesity threshold more often and are therefore more likely to diagnose them with a condition. This is similar to the model tested by Coussens (2017) for diagnosis of ischemic heart disease for patients near age 40. This could be beneficial or problematic; if patients on either side of the cutoff are in fact identical in other characteristics, important diagnoses could be missed for patients under the cutoff. Or, patients above the cutoff could be given treatment when they do not need it—either because of false positives from the subsequent tests, or because BMI is an imperfect measure of health that may not accurately identify a patient’s true risk factors. Ultimately, these effects will depend on the patient’s health condition; in some cases, it may be better to over-treat patients, while in others, costs associated with treatment may make it less desirable. Future Steps There are limitations to this study that could be improved in future research. First, I assume that the observed BMI is the same as the BMI of the patient at the time of medical treatment. Second, I use diagnosis of a condition, such as diabetes or heart disease, as a proxy for having been tested for it. Both of these limitations affect causal interpretation of the effect of the obesity threshold on physicians’ treatment decisions. However, both could be mitigated with medical records data, which would show patients’ BMI at the time of a doctor’s visit and the tests the doctor subsequently ordered. Future research should use medical records to not only explore discontinuities in medical care at the obesity threshold, but also to investigate whether patient 124

The Heuristics of Obesity

outcomes are affected by the discontinuity. If patients are more likely to be tested for coronary heart disease if they are above the threshold, the effectiveness of the test and treatments could be evaluated by comparing the health outcomes of those in the vicinity of the threshold. Finally, the specification I use for all health outcomes is based on one “test” outcome, diabetes. I did this primarily for convenience; there is no reason to assume that all health outcomes have the same functional relationship with BMI as diabetes. I could improve the validity of my results by systematically testing each health outcome with different functional forms and bandwidths to determine if the discontinuities I find here are robust to alternate specifications. The extent to which physicians rely on heuristics may be dependent on their experience level. Doctors with more experience could be less uncertain about the proper tests to recommend, so they may not rely on heuristics. This hypothesis could be tested by using medical records data from hospitals with different experience ratings or in different regions, or by modeling the experience level of the physician, such as by age or status (resident or attending). One aspect of obesity that may separate it from other heuristics is that the thresholds are relatively common knowledge, especially compared to other cutoffs like the very low birth weight cutoff. BMI charts are fixtures of medical offices as well as medical records, and children are introduced to it at a young age in physical education tests in school. Individuals may ascribe some value to their obesity status, especially if they pass from one category to another and if their physicians discuss it often. Thus, the effect of the obesity threshold should be investigated from the patient perspective, such as mental health status or opinions of the medical advice they receive. Exploring this question could give valuable insight to the unintended consequences of an inherently meaningless threshold. A physician’s role is to diagnose and treat her patient. Some interactions with patients may not provide the physician with the information necessary to make informed, rational decisions about treatment options, so she may employ heuristics, such as the obesity threshold, to help. Even though the heuristic itself could be trivial, using it to determine the treatment a patient receives could have important impacts on patient health. The results presented provide evidence that physicians use the obesity threshold heuristically, potentially leading to missed diagnoses or over-treatment of patients who are otherwise identical. Given that this threshold is just one of many possibly used in the medical field, it is pertinent to continue investigating the effects of heuristics to improve a physician’s decision-making skills and allow them to better choose the optimal treatment for their patients.

Columbia Economic Review | 125

Marissa Caldwell

VII. Acknowledgements I would like to thank my advisor, Kristin Butcher, for her unwavering support and incredible insight throughout this process.

VIII. References Abaluck, J., L. Agha, C. Kabrhel, A. Raja, and A. Venkatesh. (2016). The Determinants of Productivity in Medical Testing: Intensity and Allocation of Care. American Economic Review 106(12), 3730-3764. Abenhaim, H. A. and A. Benjamin. (2011). Higher Caesarean Section Rates in Women With Higher Body Mass Index: Are We Managing Labour Differently? Journal of Obstetrics and Gynecology Canada 33(5), 443-448. Almond, D., J. J. Doyle, Jr., A. E. Kowalski, and H. Williams. (2010). Estimating Marginal Returns to Medical Care: Evidence from At-risk Newborns. The Quarterly Journal of Economics 125(2), 591-634. Avorn, J. (2018). The Psychology of Clinical Decision Making and Implications for Medication Use. The New England Journal of Medicine 378(8), 689-691. Berendzen, J. A. and B. C. Howard. (2013) Association between cesarean delivery rate and body mass index. Tennessee Medicine 106(1), 35-37. Bharadwaj, P. and C. Neilson. (2011). The Role of Early Life Health Interventions on Mortality and Academic Achievement. Blewett, L. A., J. A. Rivera Drew, R. GrifďŹ n, M. K. King, & K. C.W. Williams. IPUMS Health Surveys: National Heath National Health Interview Survey, Version 6.2. Minneapolis, MN: University of Minnesota, 2016. http://doi.org/10.18128/D070.V6.2. Bureau of Labor Statistics, U.S. Department of Labor. National Longitudinal Survey of Youth 1979 cohort, 1979-2014 (rounds 1-26). Center for Human Resource Research, The Ohio State University. Columbus, OH: 2014. https://www.nlsinfo.org/investigator/pages/search.jsp. Calonico, S., M. D. Cattaneo, and R. Titiunik. (2014). Robust data-driven inference in the regression-discontinuity design. The Stata Journal 14(4), 909-946. Centers for Disease Control and Prevention, Behavioral Risk Factor Surveillance System. (1987-2016). Annual Survey Data. https://www.cdc.gov/brfss/annual_data/annual_data.htm. Coussens, S. (2017). Behaving Discretely: Heuristic Thinking in the Emergency Department. Diabetes Prevention Program Research Group. (2009). 10-year follow-up of diabetes incidence and weight loss in the Diabetes Prevention Program Outcomes Study. Lancet, 374: 1677-86.

126

The Heuristics of Obesity

Eknoyan, G. (2008). Adolphe Quetelet (1796-1874)—the average man and indices of obesity. Nephrology Dialysis Transplantation 23, 47-51. Finkelstein, A., M. Gentzkow, H. Williams. (2014). Sources of Geographic Variation in Health Care: Evidence from Patient Migration. National Bureau of Economic Research Working Paper No. 20789. Finkelstein, E. A., J. G. Trogdon, J. W. Cohen, and W. Dietz. (2009). Annual Medical Spending Attributable to Obesity: Payer-and Service-Specific Estimates. Health Affairs 28(5), w822-w831. Hoynes, H., D. W. Schanzenbach, and D. Almond. (2016). Long-Run Impacts of Childhood Access to the Safety Net. American Economic Review 106(4), 903-934. Johnson, E. M. and M. M. Rehavi. (2016). Physicians Treating Physicians: Information and Incentives in Childbirth. American Economic Journal: Economic Policy 8(1), 115-141. Kesselheim, A. S., K. Cresswell, S. Phansalkar, D. W. Bates, and A. Sheikh. (2011). Clinical Decision Support Systems Could Be Modified To Reduce ‘Alert Fatigue’ While Still Minimizing The Risk Of Litigation. Health Affairs 30(12), 2310-2317. Kling, R. Jeffrey, Jeffrey B. Liebman, and Lawrence F. Katz. 2007. Experimental Analysis of Neighborhood Effects. Econometrica 75 (1), 83–119. Lee, D. S. (2008) Randomized experiments from non-random selection in U.S. House elections. Journal of Econometrics 142, 675-697. Lung Transplant Patient Selection Criteria. Johns Hopkins Comprehensive Transplant Center. https://www.hopkinsmedicine.org/transplant/ referring_physicians/patient_selection_criteria/lung.html. National Center for Health Statistics. (2008). Prevalence of overweight, obesity, and extreme obesity among adults: United States, trends 1976-80 through 2005-2006. National Institutes of Health: National Heart, Lung, and Blood Institute. (1998). Clinical Guidelines on the Identification, Evaluation, and Treatment of Overweight and Obesity in Adults, The Evidence Report. NIH Publication No. 98-4083. National Institutes of Health: National Heart, Lung, and Blood Institute. (2013). Managing Overweight and Obesity in Adults. Ogden, C., M.D. Carroll, C. D. Fryar, and K.M. Flegal. (2015). Prevalence of Obesity Among Adults and Youth: United States, 2011-2014. NCHS Data Brief No. 219. Thaler, R. H. and C. R. Sunstein. (2008). Nudge: Improving Decisions About Health, Wealth, and Happiness. New Haven, CT: Yale University Press.

Columbia Economic Review | 127

Marissa Caldwell

Trogdon, J. G., E. A. Finkelstein, T. Hylands, P. S. Dellea, and S. J. Kamal-Bahl. (2009). Indirect costs of obesity: a review of the current literature. Obesity Reviews 9, 489-500. Tversky, A. and D. Kahneman. (1974). Judgment under Uncertainty: Heuristics and Biases. Science 185(4157), 1124-1131. Vickrey, B. G., M. A. Samuels, and A. H. Ropper. (2010). How Neurologists Think: Cognitive Psychology Perspective on Missed Diagnoses. Annals of Neurology 67, 425-433. Wever-Pinzon, J., S. G. Drakos, and J. C. Fang. (2015, June 29). Con: The Obese Heart Failure Patient as a Candidate for Mechanical Circulatory Support: Itâ&#x20AC;&#x2122;s Rarely Appropriate. American College of Cardiology. http://www.acc.org/latest-in-cardiology/ articles/2015/06/29/08/49/con-the-obese-heart-failure-patient-ascandidate-for-mechanical-circulatory-support-its-rarely-appropriate. Whelton, P. K., R. M. Carey, W. S. Aronow, D. E, Casey Jr., K. J. Collins, C. Dennison Himmelfarb,â&#x20AC;Śand J. T. Write, Jr. (2017). 2017 ACC/AHA/ AAPA/ABC/ACPM/AGS/APhA/ASH/ASPC/NMA/PCNA Guideline for the Prevention, Detection, Evaluation, and Management of High Blood Pressure in Adults. Journal of the American College of Cardiology. doi: 10.1016/j.jacc.2017.11.006.

128

The Heuristics of Obesity

IX. Appendix Appendix Figure 1: Treatment Algorithm for Overweight and Obese Patients

Source: National Institutes of Health (1998). Clinical Guidelines on the IdentiďŹ cation, Evaluation, and Treatment of Overweight and Obesity in Adults. Appendix Table 1: Health Outcomes constructed using NHIS Survey Questions

Columbia Economic Review | 129

Marissa Caldwell

Appendix Table 2: Health Outcomes constructed using BRFSS Survey Questions

Notes: Regressions are identical to those reported in Table 8,but include year ďŹ xed effects in addition to all other covariates (age, age2, and age3, and indicators for race, ethnicity, and level of education). BMI is scaled to 0 at the cutoff, BMI=30. All regressions are linear with a bandwidth of 3 on either side of the cutoff, and include an interaction between the cutoff and BMI to allow the slope to vary on either side of the obesity threshold. Outcome variable is an indicator equal to 1 if true. All regressions use population weights. 130

The Heuristics of Obesity

Columbia Economic Review | 131

CHANGING TRADITIONS: MODELING THE PROBABILITY OF BIDDING TO HOST THE OLYMPIC GAMES OVER TIME

Jessie Dickens University of Central Florida

Abstract: In September 2017, the International Olympic Committee (IOC) announced that, for the first time in its history, it would award the hosting rights for two different Olympic Games at the same time, giving the 2024 and 2028 Summer Olympic Games to Paris and Los Angeles respectively, due to a lack of available candidates. As a result, the question is raised as to why the IOC broke tradition in its host city selection process. The Olympics are one of the largest, most economically impactful mega-events in modern culture. However, with prior host cities reporting astronomical costs and high debt balances associated with hosting, many cities have retracted their bids or have refused to submit candidature bids altogether. While hosting does provide a nation with the opportunity to show off its infrastructure and culture before millions of people worldwide, the costs to host the event have steadily risen as the Olympics have grown in popularity. To predict the probability that a select city would bid to host the Olympics at a given time, we develop a multiple logistic regression model that evaluates a panel data set. It incorporates data from 78 cities across 13 time periods and investigates 23 explanatory variables which cover city-specific economic factors, factors related to the Olympic bidding process of and geopolitical factors. We then use post estimation techniques to isolate the probability of bidding across time periods, regions and individual cities and we explain the variability of these probabilities using the results of our model. Our results yield three main findings- that bid probabilities are dynamic across time and fluctuate cyclically, that regional probabilities have shifted dramatically in recent years, and that a variety of economic, geographic and Olympics-related factors play a role in the changing probabilities. The results provide insight into the profile of the typical bidder and suggest ways that bidding processes for the Olympics can be made more accessible to all prospective host cities so that there is increased competition in the process.

n 2017, for the ﬁrst time in history, the hosting of two Olympic Games was awarded at the same time during the 131st International Olympic Committee (IOC) session in Lima, Peru. The hosting rights for the 2024 and 2028 Summer Olympic Games were formally awarded to Paris and Los Angeles respectively, which raises the question as to why the IOC broke tradition in its host city selection process. The answer lies in the fact that the bidding contest to host the Games, usually tightly contested by a myriad of potential hosts, saw no legitimate bidders besides the two aforementioned cities for either of the Columbia Economic Review | 133

Jessie Dickens

Games. This study aims to provide an explanation for the fluctuation in the number of bidding candidates over time and examine the various factors that determine the probability that a given city will bid. To determine the probability that a given city will bid for the Olympics in a given year, we develop a multiple logistic regression model that evaluates a panel data set which incorporates data from 78 cities across 13 time periods (1962-2010 with 4-year gaps to represent the Olympic bidding cycles). We investigate 23 explanatory variables in our model which cover city-specific economic factors, factors related to the Olympic bidding process, and geopolitical factors. We then iteratively simplify our model to isolate the determinants that play the biggest role in whether or not a given city will bid. Finally, post estimation techniques are used to isolate the probability of bidding across time periods, regions and individual cities and we explain the variability of these probabilities using the results of our model. Our findings confirm that bid probabilities for hosting the Olympics are dynamic across time and fluctuate in a cyclical manner, impacted by the number of bids in previous cycles as well as the fluctuating reported cost levels incurred by hosts. We also show that aggregate regional bid probabilities have changed drastically in recent years as the composition of bidding cities has changed, with cities from developing economies now more likely to elect to bid. We also show that a bid committees’ decision to bid is complex and is influenced by a variety of different variables. Economic, geographic, sociocultural and Olympic-related variables all play a role in influencing bid probabilities. Our results allow us to develop the profile of a typical bidder and provide us with information that is useful for understanding policy decisions relating to bidding to host the Olympic Games. By understanding the factors that drive governments’ decisionmaking process for submitting a bid, we are better able to identify variables over time that impact changing bid probabilities and can use our results to predict future trends. The layout of our study is as follows: Section 1 details background information surrounding the modern state of hosting the Olympic Games and the bidding process. Section 2 reviews the relevant literature pertaining to the economics of bidding for and hosting the Olympics. Section 3 lays out the methodology that was utilized and specifies the economic model. Section 4 presents the results, Section 5 analyzes and discusses them, and Section 6 concludes.

I. Background The modern-day Olympics were revived at the end of the late 19th century with the formation of the International Olympic Committee (IOC), a non-proﬁt organization that serves as the governing body of the Games, by Frenchman Baron Pierre de Coubertin. The Games have evolved drastically 134

Changing Traditions and the Olympic Games

throughout modern history with the development of technology, radio, television and the quality of infrastructure. For the purposes of this study, all Games before 1972 are not examined due to a lack of available economic data as well as the fact that the structuring of the games, and the expenditures and revenues aren’t necessarily analogous to those that arise from the Games that are held today. In the research novel The Economics of Staging the Olympics, sports economist Holger Preuss divides the history of the modern-day Olympics into four phases (Preuss, 2004). Phase I covers the advent of the Modern Olympics through 1968, Phase II covers 1968-1980, Phase III covers 1981-2000 and Phase IV covers 2001 through today. The circumstances surrounding the financing of the Games and the resulting economic impacts have varied widely across the evolution of these Olympic phases. There have been Games that have been financed almost completely publicly and that have resulted in massive long-term debts [Montreal 1976], while others that have been financed privately, as a joint effort between the Organizing Committee for the Olympic Games (OCOG) and the IOC, have resulted in profits for the host city [Los Angeles 1984]. One constant fixture that has been observed over time is the ballooning of the overall cost level associated with the hosting of the Games, with Beijing reportedly spending over $40 billion USD on the 2008 Olympic Games (Riley, 2015). Contemporary Olympics are characterized by lavish spending and the creation of sports facilities that showcase luxury, wealth and a nationalistic sense of culture. These Games are marked by high cost overruns and many have experienced economic losses as a result. Our study investigates whether this impacts the probability of a city submitting a bid in a given year. The process of submitting the bid itself can also be costly. The bidding process begins about nine years before the actual staging of the Games with the National Olympic Committee (NOC) of each interested country submitting an initial application that establishes that they meet the IOC’s minimum requirements to host on behalf of their chosen prospective city. Once a city passes this initial phase, it is officially considered a candidate. The city then completes a more detailed questionnaire that documents logistics regarding the city’s plans to successfully host. Candidate cities must provide information on their vision and concept for the Games, their envisioned Games experience, plans for the Paralympic Games, and the sustainability of the Games including their intended legacy (The IOC, 2014). The city must address factors including transportation, lodging, infrastructure, safety, security, public support and perhaps most importantly- plans for the financing of the Games. The IOC Evaluation Commission, comprised of IOC officials, NOCs, athletes and international experts, then evaluates these questionnaires and makes personal visits to inspect each of the cities. The evaluation occurs based on a Columbia Economic Review | 135

Jessie Dickens

fixed set of criteria including general infrastructure, the Olympic Village, transportation and government support among other things. Finally, the IOC Evaluation Commission votes on a shortlist of up to five cities until one receives the absolute majority and is chosen as the host country. This typically occurs about seven years before the hosting of the Games. Once the host city is selected, they sign a contract with the IOC that determines the conditions for the organization of the Games. The host city is then responsible for forming the OCOG that assumes responsibility for the planning and financing of the Games. Economic liability is transferred to the host city and the OCOG which are both jointly liable for all financial aspects of the Games. The contract lays out what amount of financial contribution the IOC will provide to the host city and also lays out how any potential revenues will be split up. Our study also investigates the impact that hosting conditions have on the probability of bidding in a current cycle.

II. Literature Review There is an extensive body of literature that investigates mega-events, including the Olympic Games, and the economic impact that a host city or nation experiences from hosting them. Mega-events, including the Summer Olympics, generally refer to “short lived collective cultural actions with long-lived pre- and post-event social dimensions” and take the form of specially constructed and staged international cultural and sports events (Roche, 2003). Local politicians are often eager to pursue bids to host mega-events and use a variety of justifications to support the decision to bid, including positive economic prospects, embracing the opportunity to improve infrastructure and facilities, engaging in the opportunity to showcase the city/country on the global stage, and increasing the level of national pride and happiness within the nation’s citizens (de Nooii and van den Berg, 2018). Before bidding, governments often commission economic impact studies that are used to project massive increases in economic activity and economic prosperity within the region resulting from hosting. Economists, however, are skeptical of these findings and have pointed out many issues pertaining to the methods used to project these economic gains. We summarize the methodologies used to quantify economic impact, existing critiques of these methodologies and the results of a variety of studies pertaining to the impact of hosting mega-events within a host nation. We then consider these findings in the context of making the decision to bid. Methodologies Used to Capture Economic Impact and Critiques. The two main methods for determining the impact that the hosting of a mega-event has on a city’s economy are economic impact analysis (EIA) and cost benefit analysis (CBA). EIA generally consists of aggregating all economic activity — direct, 136

Changing Traditions and the Olympic Games

indirect and induced — that results from a mega-event and using it to demonstrate the overall impact of the event. Induced effects, such as those created by infrastructure investments and tourism, are approximated using multipliers (Taks et al., 2011; de Nooji and van den Berg, 2018). The EIA methodology doesn’t differentiate between costs and benefits and simply presents the overall impact without determining whether it is a net positive or negative impact. Conversely, CBA categorizes economic flows into distinct cost/benefit categories and measures these against each other to determine the total net impact. While costs and benefits can be difficult to quantify, benefits are generally categorized as short-run benefits (tourism), long-run benefits (infrastructure, trade, etc.), and intangible or feel-good benefits, while costs are generally categorized as either direct capital costs (sports infrastructure), indirect capital costs (general infrastructure), and operational or administrative costs (Baade and Matheson, 2016). Theoretically, a bid committee would only choose to proceed with a bid if the estimated marginal benefits outweigh the marginal costs. However, bid committees and supporters of bidding typically prefer the EIA methodology to the CBA methodology as it often provides a more optimistic result (de Nooji and van den Berg, 2018).This is understandable as a mega-event such as the Olympics will undoubtedly generate a large amount of economic activity, which the EIA captures. Nonetheless, a majority of the economic activity is often composed of the costs that are incurred, which the CBA, but not the EIA, measures and factors into a net profit or loss calculation. Exaggerated economic impact studies from bid committees are intuitive due to the inherent conflict of interest that exists given the nature of their position, as they try to gain public support for bidding. However, while ex post economic research has contributed mixed findings, the overall consensus is that hosting the Olympics tends to be an economic risk and there has been “no consistent evidence of positive economic impacts even remotely approaching estimates in economic impact studies” (Baade and Matheson, 2016; Owen, 2005). The disparity between the findings of ex ante and ex post studies results from a variety of theoretical and methodological shortcomings that have been documented in critiques of economic impact studies (Preuss, 2004). These shortcomings include the use of exaggerated or inappropriate multipliers (Matheson, 2009; Owen, 2005; Crompton, 1995), counting economic transfers as benefits (Owen, 2005), ignoring opportunity costs (Giesecke and Madden, 2011; Barclay, 2009; Crompton, 1995), and not accounting for the crowding out effect (Taks et al., 2011; Barclay, 2009). When conducting an economic impact study, massive costs such as stadium construction costs are Columbia Economic Review | 137

Jessie Dickens

counted as economic benefits because they boost overall economic activity. However, the counting of this activity as an economic benefit is misguided as these are simply transfers of funds that would have been spent on other projects such as road building, public housing, etc. and the money spent is typically collected from taxes on consumers who likely would have spent the money within the local economy anyway. The impact studies then take these expenditures and extrapolate them to account for indirect costs, the recirculation of new money that has been injected into the local economy, ignoring the fact that much of this money originated locally. Thus, multipliers used to estimate indirect costs further magnify faulty implementations of economic benefit counting. Finally, estimates of increased tourism spending end up being overstated as the spectacle of the mega-event displaces regular tourism within the region. Thus, empirical studies have found that actual increases in tourism within the local region end up being negligible as consumers would prefer to stay outside the region where rates are lower and then travel into the local area for the mega-event. The result is overstated benefits as well as understated costs, which affect stakeholders at both a macroeconomic and microeconomic level. Ultimately, taxpayers are burdened with maintaining sports facilities that could have been useful buildings such as hospitals or schools, and business owners are affected adversely as evidenced by low traffic/occupancy rates during the event and elevated bankruptcy rates due to oversupply after the event (Baade and Matheson, 2016; Tiegland, 1999). Findings of Economic Impact Studies. The findings of economic impact studies (both ex ante and ex post) are often cited by bid supporters and detractors in supporting their respective arguments. A major issue faced by prospective bidders are the massive costs that are associated with even submitting a bid. For example, Chicago’s unsuccessful bid to host the 2016 Olympics cost nearly $70.6 million and other bids have cost up to $100 million (Pletz, 2010; Zimbalist, 2015). Thus, submitting a bid is a high-risk proposition with potentially no reward if the city is not selected to host. On the other hand, if a city is selected to host, they face the prospect of spending billions of dollars on preparations for the Games. Rising costs have been accelerated by high levels of cost overrun, with the Summer Olympics averaging a cost overrun level that is 156% of budgeted cost in real terms on average (Flyvbjerg and Stewart 2012; Flyvbjerg and Stewart 2016). Thus, Olympic budgets serve as baseline spending points rather than hard spending caps, which contribute to rampant overspending in order to complete the hosting accommodations on time. Given the massive costs associated with hosting the Olympics, one would expect a 138

Changing Traditions and the Olympic Games

multitude of benefits to result from hosting which make the investment economically feasible. There are mixed findings relating to a variety of key economic performance indicators. One justification that is often used for the large expenditures taken on is the increases in employment that will result from hosting. However, research shows that increases in employment are “transitory” and are often lower than projected (Baade and Matheson, 2016; Jasmand and Maennig, 2008). Hotchkiss et al. (2003), using a differences-in-differences technique demonstrates an increase in employment levels in Georgia’s venue counties following the 1996 Olympics, but also notes no significant increase in wage levels. Coates and Humphreys (2003) suggest that spending on a mega-event and spending on other forms of recreation are substitutes as employment gains in certain sectors (amusement/recreation) are negated by employment losses in others (hotels, retail). Politicians who back Olympic bids also usually hope to use the opportunity to revamp infrastructure within the city. However, Lipsitz (1984) posits that the construction of new stadia for the Olympics generally only benefits a few, wealthy investors and doesn’t benefit most citizens. The increased cost of living around the city center forces citizens of lower economic status outward, causing a series of “concentric slums”, described by the Warner Model (Lipsitz, 1984). Thus, in order for a stadium to be economically effective, it must be integrated into the fabric of the city (Maennig and du Plessis, 2009). For cities in developing economies, hosting the Olympics does create an opportunity for urban development through improved facilities and transportation, as evidenced by Beijing and the 2008 Summer Olympics, although it experienced a low return on investment from infrastructure projects (Singh and Zhou, 2015). Nonetheless, economic benefits from infrastructure investments are generally overstated and can have adverse effects such as displacement of citizens and poor accessibility (Levy and Berger, 2013; Maennig and du Plessis, 2009). Coates and Matheson (2009) model citizens’ willingness-to-pay for increased infrastructure created by mega-events by measuring changes in housing rental prices but fail to find consistent evidence of any change. There is evidence that the large global spotlight that is placed upon host cities serves as a conduit for more favorable perceptions of the city after the Games are complete. Both Beijing, whose 2008 opening ceremony is widely lauded as one of the most extravagant ever, and South Africa, which hosted the FIFA World Cup in 2010, reported favorable responses from citizens around the world resulting from hosting (Singh and Zhou, 2015; Bob and Potgieter, 2013). Local officials report, however, that care must be taken in presenting a city’s image before the world as all aspects of a city are on display, not just the positive ones. Host cities in the past have struggled with negative factors such Columbia Economic Review | 139

Jessie Dickens

as terrorism (Munich, 1972), crime (South Africa, 2010), and human rights (Beijing 2008, Qatar 2022 World Cup). The literature notes that host cities have had trouble converting favorable perceptions into sustained tourism post-event, which is another expected benefit suggested by ex-ante economic impact studies (Tiegland, 1999; Singh and Zhou, 2015; Giesecke and Madden, 2011). Other economic variables have been studied in relation to the hosting of the Olympics. As expected, there is a lot of economic activity that results from preparing for the Games. Thus, hosting has an expansionary effect and can bolster a sluggish economy, especially in the years leading up to the Games (Kasimati and Dawson, 2009). Bruckner and Pappa (2015) find that bidding for the Olympics increases economic activity 7-9 years prior to the event and cities that were selected to host receive an additional boost 2-5 years prior, as the announcement of a mega-event serves as a “news shock” to the local economy. The expansionary effect resulting from hosting has also been found to increase consumer confidence, consumption spending and government investment levels (Sterken, 2006). Additionally, hosting a mega event has a signaling effect to other countries around the world that the host is pursuing an economic liberalization policy. Rose and Spiegel (2009) demonstrate this through a variety of trade models and conclude that host countries experience a trade output increase of 36% that is permanent after hosting the Olympics, which could make hosting attractive to cities looking to enhance economic partnerships. Finally, hosting the Olympics has an intangible “feel good” effect that is difficult to quantify, but boosts feelings of culture and pride throughout the host city (Porsche and Maennig, 2008). Factors that Make a City More Likely to be Selected to Host. Given the high costs of bidding, a bid committee would likely only decide to bid if they were confident that their city would ultimately be selected to host. Maennig and Vierhaus (2017) develop a rank-ordered logistic regression model that investigates 147 variables and determines which play a role in determining whether a bidding city will ultimately be selected to host. This study was instructive in selecting variables for the current study. The authors find that cities with large populations, high growth economies with large markets, and high public support for a bid are most likely to be selected. Additionally, the IOC seems to prefer cities with a large number of existing stadiums and cities that use political liberalization as a theme in their bidding materials. The IOC likely prefers cities with a large amount of infrastructure built at the time of bidding to mitigate risk that construction projects will not be completed on time. Feddersen, Maennig and Zimmermann (2007) conduct a similar study and find that the average distance between sports venues and the Olympic village as well as the city’s average temperature play a role. 140

Changing Traditions and the Olympic Games

Shoval (2002) notes a paradox that exists surrounding the types of cities that bid and that are selected to host as cities that are in developing economies stand the most to gain economically from hosting, but it is the larger cities from developed economies that are usually selected. Baade and Matheson (2016) note however that the IOC has attempted to make the hosting rights more accessible to developing economies in recent years as indicated by the changing composition of cities that are hosting. Given the dubious economic benefits associated with hosting the Olympics, the question is raised as to what extent social factors play a role in support for a bid. The literature finds that politicians are quick to back a bid despite the economic ramifications because they hope to make their people “proud and happy” by building feelings of national pride, all while garnering popular support for future elections and enhancing their own legacy (de Nooij and van den Burg, 2018). Streicher et al. (2016) surveyed citizens that supported a bid through public referenda on the Games and found that while citizens do consider economic factors to be important, they are primarily driven in their support by social factors such as building a sense of community, boosting their city’s reputation in the global view, and advancing the culture of the city. Despite the social and economic effects of hosting the Olympics, economists generally still view the risk as not worth the potential rewards, referring to bidding as “fool’s gold” with the winner gaining a “winner’s curse”.

III. Methodology Sample Selection. Given the findings of the literature review above and the stark decrease in the number of cities that have submitted bids to host the Olympic Games in previous years, the research question becomes, “How does the probability of a given city submitting a bid to host the Olympic Games fluctuate over time and what variables determine that fluctuation?” We model the bid decision process using panel data. This allows us to analyze a variety of cities and how their probability of submitting a bid changes over time. The time period for the present study includes Games hosted between 1972-2020 (bid cycles 1962-2010). The bidding cycle generally occurs about 7-10 years before the year that the Olympics is actually held in. Previous literature has typically used data from 10 years prior to the year that the Games were held as the standard for examining variables during a given bidding cycle (Bruckner and Pappa, 2015). We follow this convention in our dataset. To provide an example, consider that we are evaluating the bidding cycle for the 2016 Summer Olympics that recently occurred. The bidding period for this Olympic Games occurred in the years 2007-2009. Thus, prospective bidders most likely examined economic data from the 2000, 2004 and 2008 Summer Olympics Olympics when making their decision on whether or not to bid. To account for this and to provide consistent notation throughout the paper, when Columbia Economic Review | 141

Jessie Dickens

examining a specific Olympic Games (say 2016), that year will be designated by the time value t. In developing the explanatory variables, time lagged variables from the t-2 (2008), t-3 (2004), and t-4 (2000), etc. cycles would be used as well as economic benchmarks from the year 2006. We do not examine economic variables from the t-1 period (the 2012 Olympics cycle, which would have occurred before 2016) because the bidding and selection of the host city had already occurred by this point. This allows us to accurately examine the bidding decision through the same lens that decision makers would have used. Data is collected in four-year increments for convenience (Olympics related data by its nature is only available in four-year increments) and because decision makers examine economic impact data from economies that hosted the Games in the periods where the Games occurred, as economic conditions during these periods are irregular relative to other years. The sample contains data from three types of cities: cities that have hosted the Olympics, cities that have bid but not been selected to host and cities that have never bid at all. Cities were selected randomly from their respective pools of cities meeting the above criteria. The final pool contained 19 cities that had previously hosted, 16 cities that had previously bid but never hosted, and 43 cities that had never bid. The listing of cities for the Summer Olympics regression can be found in Appendix A. Data Compilation. In an effort to create a holistic study and to avoid omitted variable bias to the maximum extent possible, three categories of explanatory variables were used during the analysis. The categories of variables are general economic variables, geopolitical variables, and Olympics-specific variables. Each one of these categories contains several variables that are possible factors in determining whether a city is likely to submit a bid. A listing of variable names, descriptions, and data sources can be found in Appendix B. General economic variables are used as potential determinants of a city’s likelihood to bid because bidding for the Olympics is a huge economic undertaking that has potentially massive economic impacts. The chosen variables were selected because they provide an insight not only into the economic state of a nation but also into the quality of life for the citizens of that city. All economic variables were transformed using the natural logarithm (i.e. log.RealGDP = LN(RealGDP)). This was done to reduce outliers in the data and to promote linearity throughout the model. Geopolitical variables were also chosen because official bidding and financing of the Olympics involves the governmental regime and system of the specific nation that is considering a bid. The IOC has established baseline requirements for bidding and will generally only select cities that have governments that will legitimately be able to host the Games. Thus, the political state of a country possibly plays an impact on whether or not it will bid. Additionally, the IOC tends to rotate the continent that the Games is hosted in which may play a role in whether a city will bid or not. 142

Changing Traditions and the Olympic Games

Finally, and perhaps most importantly, reported statistics regarding previous Olympics also are likely to play a crucial role in whether or not a city will bid. These variables are split into two categories: bidding and hosting history variables and economic performance variables. This allows us to determine if cities that recently hosted are less likely to bid and if those that have bid recently but were not selected are more likely to bid again. These variables also allow us to draw conclusions over whether cities truly take the past economic performance reported by other host cities into account or whether there are other motivations at play. The outcome variable is a binary variable that is unity if the ofďŹ cials of the city chose to submit a bid in the corresponding time period. Summary statistics for all variables are presented in the table below:

Columbia Economic Review | 143

Jessie Dickens

Missing Data. Overall, there were 438 missing data values out of 28,392 total values (1.54%) and missing data points were generally missing at random. A majority of missing data values came from economic variables. Missing data points were imputed where possible using two primary methods. For monotonic sequences of economic data, missing values at the beginning or end of sequences were extrapolated using a growth rate computed across the entire set of observations for the given city. Missing values in the middle of the sequence were interpolated using a similar method. If the sequence of data was not monotonic, the data value for the missing time period was imputed using a moving average of the three prior time periods. For irregular data sequences with no clear trend, the observation containing the missing data point was omitted using listwise deletion. After imputation of data, there were 78 (0.27% missing data values. The Model. Due to the binary nature of the outcome variable, we use a random effects logit regression model to predict the probabilities for bidding. A Hausman Test (P > Ď&#x2021;^2 = 1.00) conďŹ rms that a random effects model is appropriate by demonstrating a lack of correlation between regressors and effects: We then use clustering analysis to further analyze the results. The random effects model allows for analysis of the variables that are constant within country clusters across time and not just those that vary across time (for example, index variables such as region and Human Development Index are constant across time). The general form of the random effects model is: Using our explanatory and outcome variables, we ďŹ t the general form of the model:

where i = the country observed and t = the year in which it is being observed. The AvgNumBids, AvgRevenue, Avg_Venues and AvgCost variables were averages of time lagged variables representing the number of bids submitted in recent cycles (t-2 through t-4), the percentage share of revenue previous host cities received from the IOC (t-2 through t-3), the percentage of venues that were 144

Changing Traditions and the Olympic Games

newly built in recent Olympic Games (t-2 through t-3) and the total reported cost of hosting recent Games (t-2 through t-3). Multicollinearity Checks. While estimating the logit model, we analyze our dataset for multicollinearity. We test for multicollinearity by determining a variance inďŹ&#x201A;ation factor (VIF) for each variable in the dataset. As a result, we omit the Globalization variable from our regression due to multicollinearity (VIF=10.65). After omitting the collinear predictor variable, we conduct the analysis again and conďŹ rm that no multicollinearity exists within any of our variables, as all variables have a VIF lower than 10. Table 2 summarizes the variance inďŹ&#x201A;ation factors for each variable. A variance-covariance matrix for the variables is presented in Appendix D. Additionally, we account for heteroskedasticity by clustering standard errors in the model at the city level, ensuring that they are heteroskedasticity-robust across each cluster.

Columbia Economic Review | 145

Jessie Dickens

IV. Results Model Results. With our set of independent variables comprising a coefficient vector for β, we obtain the logit regression results establishing a relationship between the variable coefficients and the dependent variable Bid. From our original list of variables, we use Gaussian model pruning as described by Rojas (2015) to simplify our model and obtain efficient results. The listed independent variables were omitted from the regression models in the following order, with the sign of their β-coefficients listed in parentheses: NumBids_t2 (+), Avg_Venues (-), logRealGDP (+), i.HDI (-), NumBids_t4 (-), AvgTemp (+), FIFA_Prev10 (-), PrevBid (+), LargestCity.(-). Both Akaikes’s information criterion (AIC) and the Bayesian information criterion (BIC) support the use of the nested model (AIC=312.84, BIC=424.78) as opposed to the full model (AIC=329.27, BIC=488.98). The results of the nested model are presented in Table 3. To further investigate the effects that the chosen independent variables have on the outcome variable Bid, we compute marginal effects associated with the logistic regression. We compute average marginal effects for our explanatory variables, using the formula:

146

Changing Traditions and the Olympic Games

Our measure of marginal effects provides us with the change in our outcome (the probability of a city submitting a bid in time) as a function of the change in an explanatory variable, holding all other explanatory variables constant. Marginal effects are presented below in Table 4.

Columbia Economic Review | 147

Jessie Dickens

After establishing relationships between our outcome variable and explanatory variables, we use our regression results to predict bid probabilities, clustering our results by bidding year and examining them at the regional level and the individual country level. Prediction results clustered by bid year and summarized at the regional level are presented in Table 5. Figure 1 depicts aggregate bid probabilities against the number of bids actually submitted in each bid cycle. Additionally, Figures 2-8 provide graphical depictions of regional bidding trends and are displayed in Appendix C.

148

Changing Traditions and the Olympic Games

Figure 1: Aggregate Bid Probability and Number of Bids Submitted

Columbia Economic Review | 149

Jessie Dickens

Robustness Checks. After obtaining the logit model results, we analyze our dataset for predictive accuracy. We test the predictive capacity of the model by examining predictions for the probability of bidding for each city in each time period and comparing the predicted probabilities to actual bidding results. As indicated by Table 5, the probability that a randomly chosen city from any region will bid for hosting rights in a randomly chosen time period is approximately 5.02%. Based on the aggregate results in Table 5, we use 10% as a benchmark probability indicating a positive result- that a city is relatively likely to submit a bid to host. It is also important to note that a bid probability of 10% is considered a positive outcome because it is a high probability relative to the average probability of bidding, not because it is necessarily a high absolute measure. In absolute terms, we would expect a city with a 0.10 bid probability not to bid a majority (90%) of the time, which contributes to the false positive rate. We then compare the positive/negative probability results predicted by the model in relation to actual bidding results. The results are summarized in Table 6. Using 10% as the chosen standard for predicting a positive outcome (i.e. the city bidding in the selected time period t), the model correctly predicts bidding cities 87.72% of the time, which provides insight into the modelâ&#x20AC;&#x2122;s accuracy and robustness over time and region. In particular, when the model predicts a positive result, it is accurate 25.68% of the time and when the model predicts a negative result, it is accurate 98.48% of the time.

The accuracy rate for negative outcomes is particularly illuminating. This is due to the fact that there is a larger sample size for negative outcomes (847 negative predictions compared to 148 positive ones) and that very few cities elect to ultimately bid for the Olympic Games, which would naturally result in low number of true positive results. Furthermore, at least one city ranked in the top 4 out of 78 cities observed in terms of bid probability, actually bid in 12 150

Changing Traditions and the Olympic Games

out of 13 (92.3%) bid cycles and at least two cities bid in 7 out of 13 cycles. Overall, the 87.72% accuracy rate is indicative of the model’s predictive power and robustness over time-specific and city-specific factors. As a further robustness check, we use the model to predict cities that would bid for the 2024 and 2028 Olympics. We predict the probability by compiling 2014 and 2018 data and plugging it into the logit model, with the resulting output being represented by .We then use the output value to find the predicted probability of bidding by substituting it into the equation: Using the 10% threshold as the standard for a positive outcome, the model correctly predicts that Los Angeles (36.97%) and Paris (11.48%) would bid for the 2024 Summer Olympics and that Los Angeles (52.87%) would bid for the 2028 Summer Olympics. In both bidding cycles, Los Angeles is ranked in the top 4 in terms of bid probabilities. This provides positive evidence of the model’s ability to successfully predict bidders in future years. However, we are limited in drawing conclusions about the model’s ability to predict future cycles due to the fact that there were only 3 total bidders for the 2024 and 2028 cycles. While the model successfully predicted these three cities as bidders, the sample size is relatively small and thus the model should be tested on cycles beyond the 2028 cycle (once they have been completed) to determine its predictive capacity for future cycles. V. Analysis Our results can be summarized into three main findings: aggregate bid probabilities fluctuate cyclically, regional differences in bid probabilities have changed over time, and a diverse array of factors play a role in determining a city’s bid probability during a given cycle. We discuss these three findings below. I. Aggregate Bid Probabilities are Dynamic Across Time and Fluctuate Cyclically. Our study demonstrates that the probability of an arbitrarily chosen city bidding for the Olympic Games across a range of bidding cycles is not constant and fluctuates significantly due to a variety of factors. We first examine the fluctuations in probabilities at the global level by aggregating bid probabilities from all regions studied. The 1980 Summer Olympics bidding cycle possesses the lowest aggregate bid probability out of all bidding cycles studied with a probability of approximately 2.88%, when only two cities (Moscow and Los Angeles) submitted bids. The aggregate bid probability then begins to increase until it reaches a probability of 9.71% for the 2004 Summer Olympics where 11 cities submitted bids. The aggregate probability then decreases again, landing at 4.85% and 5.28% for the 2016 and 2020 Summer Olympics respectively, where 7 and then 5 cities submitted bids. Columbia Economic Review | 151

Jessie Dickens

The cyclical nature of aggregate bid probability fluctuation supports a finding posited by Preuss (2004), relating the number of bidders in bidding cycles, the cost of the corresponding Olympics, and the number of bidders in future cycles. Preuss notes that in a cycle where there are many bidders, the IOC possesses more bargaining power in dictating the terms of the hosting arrangement. Thus, the IOC can request that the selected host city take on additional expenditures such as building new venues (even if there are existing venues already in place). In arranging the hosting terms, they can also use their bargaining power to request a higher share of revenue from the event, while bearing little to no liability for any of the associated costs. If a city shows apprehension regarding the terms of the arrangement, the IOC is simply able to choose another city to host from the supply of bidders. This is coupled with the fact that more bidders represent increased competition for the right to ultimately host the Games. By introducing more supply side agents, cities must essentially outbid each other, offering more luxurious amenities than other bidders. This drives up costs for the event, resulting in some of the outlandish costs and cost overruns discussed in the literature. The high costs experienced by previous hosts result in future bidders facing more pressure not to host the Olympics, which results in less bidders in future cycles. The AvgCost variable (significant at the α=0.05 threshold) in our model captures the impact of previous event costs on future bidding probability with its negative β-coefficient, implying that higher event costs drive down future aggregate bid probability. Conversely, when fewer cities bid, the selected bidding city has more relative bargaining power when determining terms of hosting with the IOC. These cities are more likely to be able to use existing infrastructure and retain a higher portion of Olympic-related revenue. The IOC has relatively few options to choose from in terms of selecting from alternative bidders, so they must be more willing to make compromises with the selected bidder. This results in lower hosting costs for the selected city and even the potential for a profit to be made off of the event, as demonstrated by the hosting of the 1984 Summer Olympics in Los Angeles, which was the only city to bid. In future cycles, more cities are encouraged to bid by the prospect of turning an economic profit while also experiencing the legacy effects associated with hosting. Thus, lower costs in previous cycles lead to higher aggregate bid probabilities in future cycles. II. Regional Differences in Bid Probabilities Have Changed Over Time. The predicted bid probabilities stemming from our model highlight the changing composition of bidders on a regional level. Prior to the 1986 bid cycle (which ultimately awarded the 1996 Games to Atlanta), the regions with the highest bid probabilities were consistently Europe, North America and, to a lesser extent, the Middle East (which started experiencing elevated bid probabilities in the 152

Changing Traditions and the Olympic Games

in the 1974 bid cycle).1 Having the composition of bidders primarily consisting of cities from Europe and North America is largely unsurprising given the fact that a majority of bids and a majority of host cities have come from these two regions since the start of the modern Olympics. Since the advent of the modern Olympics in 1896, Asian cities have hosted 5 times (although Tokyo counted for 3 of these occasions), Oceanian cities twice (ďŹ rst in 1956, both from Australia), and South American cities once (Rio de Janeiro in 2016), with no African or Middle Eastern cities ever hosting the Games. This is a massive disparity compared to the combined 26 times that a European or North American city has hosted. Thus, it is intuitive that cities from these regions are less likely to bid given that the IOC has demonstrated a general bias toward selecting cities from Europe and North America to ultimately host. Given the high expenditures associated with bidding, a city would likely not bid unless the bid committee felt that there was a strong probability of being selected to ultimately host. However, the trend of European and North American cities having the top bid probabilities has changed in recent bid cycles. Across bid cycles 1962-1982, the mean difference between the average bid probabilities of North American/European cities and cities from the other regions (excluding the Middle East) is approximately 3.29% compared to 0.88% across bid cycles 1986-2010. This aligns with the selection of host cities such as Sydney in 2000, Beijing in 2008 and Rio de Janeiro in 2016, indicating that cities from regions besides North America and Europe now have a stronger probability of being selected. The selection of Rio de Janeiro and Beijing also indicates that the IOC is willing to award the hosting of the Games to cities from developing economies which makes hosting more accessible to countries throughout the world. Thus, it is reasonable to conclude that bid probabilities from these developing regions have increased as a result of the IOC being more regionally inclusive in its selection process. Since 1986, aggregate regional bid probabilities from Oceania, South America or Asia have surpassed the mean global bid probability in every bid cycle with the exception of the 2002 bid cycle. The results also suggest that the regional rotation of host cities impacts the likelihood of bidding in certain cycles. In our sample of bidding cycles, consecutive host cities have never come from the same region. Although there is no ofďŹ cial regional rotation system, the IOC likely attempts to cycle the event regionally to maintain global interest by continuously introducing it into new markets and to demonstrate to cities from all regions that there is an equal opportunity of being selected to host. The model suggests that prospective cities are less likely to submit bids given that the previous host city came from the same region. On average, if a city from a given region is selected to bid in a bidding cycle t, the aggregate bid probability for that region decreases by 2.82% in bidding cycle t+1. The logit model captures this trend with the RegionPrev Columbia Economic Review | 153

Jessie Dickens

variable (significant at the α=0.05 level) which has a negative β-coefficient, implying that a city’s bid probability is lower if that city’s region hosted the previous Games. Thus, the selection of a city serves as implicit information from the IOC that it will look to a different region when selecting a host in the next cycle, which drives out supply side agents and leads to decreased bid probabilities within that region. Therefore , there is evidence that a de facto regional rotation system does exist and that bid committees from prospective cities take this into consideration when deciding whether to bid. III. A Diverse Array of Factors Determine Bid Probabilities. The literature primarily discusses the impact that hosting a mega-event has on the economic state of a host city. However, our model illustrates the impact that the economic state of a prospective city has on the likelihood it will submit a bid to host in the first place. Furthermore, the model demonstrates that geographic and Olympics-related factors have a significant role in determining probability of bidding along with the economic factors that are generally focused on. By analyzing the multi-faceted determinants of bid probability, we are able to gain insight into the decision-making process of a bid committee and develop a profile of a typical high-likelihood bidder. The most likely bidders tend to be cities that are economically liberal and that have characteristics that are indicative of economies that are relatively developed and industrialized as well as in expansionary phases. As indicated by a positive logistic regression β-coefficient and by the fact that countries with higher exports (as a % of GDP) have higher bid probabilities on average. Rose and Spiegel (2009) conclude that cities host the Olympics to provide a signaling effect to other economies that they are interested in pursuing an economic liberalization policy which in turn boosts levels of trade ex post. Our model supports this conclusion but also demonstrates that economies that already adhere to liberal economic policies in terms of exports and trade are likely to pursue the opportunity to host. These economies are also likely at an advantage in terms of meeting the economic and logistical burdens of hosting as they are able to rely on other economies for goods and services that will be necessary to develop the infrastructure necessary to support an event with the magnitude of the Olympics. The positive β-coefficients and marginal effects rates associated with GDP per Capita variables and CO2 emissions rates also signify that economically developed and highly industrialized cities have higher bid probabilities on average. Log.GDPperCap was designed to serve as an indicator of the economic quality of life within cities that bid to host the Olympic Games. It is intuitive that GDP per capita levels would be positively 154

Changing Traditions and the Olympic Games

related to bid probability, as developed economies likely have more resources to cover the massive expenditures associated with bidding and hosting. Log.CO2 was also included to proxy for economic growth and industrialization levels. Saidi and Hammami (2015) demonstrate links between CO2 emissions and energy consumption, and in turn positive relationships between energy consumption and economic growth variables. Countries with high CO2 emissions generally have more industrialized economies and higher growth levels which explain why cities from these countries have higher bid probabilities, as they are similarly able to produce infrastructure and goods/ services necessary to successfully plan for and host the event. The positive relationship between our inflation variable and bid probabilities also supports the notion that expanding economies bid for the Olympics. Hosting a mega-event has been shown to have an expansionary effect on the economy of the host so these cities likely view hosting as an opportunity to continue to boost economic activity as well as to boost consumption levels and consumer confidence within the city’s economy. The Human Capital Index (HCI), which is based on the average years of schooling and an assumed rate of return to education, suggests that cities with a more educated and skilled population are less likely to submit bids (Penn World Table Version 9.0, The Human Capital Index, 2018). A higher human capital index suggests that the individual citizens of a nation are able to provide “more global economic value” through their knowledge and skills (World Economic Forum, 2017).This result is a bit surprising as one might expect that a city with more human capital would be more capable of and likely to bid. However, we offer the counterpoint that cities with highly educated citizens and skilled labor forces likely have less to gain from hosting the Olympics as they don’t need the influx of low-skilled, labor intensive employment opportunities that come with preparing for the Games. Additionally, nations with more skilled labor forces (and thus higher wage rates) have more to lose if the Games are an economic disaster as evidenced by the 1976 Summer Olympics in Montreal. Thus, citizens from a highly educated city may be less likely to support a bid for the Olympics. This was highlighted by Rome’s failed 2020 bid, which lost support from the municipal administration who felt that it was not a responsible use of money, and Hamburg’s failed 2024 bid, which did not survive a public referendum. This is a key insight that explains why the probability of submitting a bid fluctuates over time and has decreased in recent years as the developed economies that once submitted bids to host are no longer doing so. The results also indicate that prospective bidding cities use information from previous cycles, such as the IOC’s behavior in selecting a host as well as the relative success of the previous host, in order to weigh marginal benefits andcosts and ultimately make a decision on bidding. A major finding is that cities weigh the outright costs incurred by previous hosts while staging the Games and that rising costs result in decreased bid probabilities in future cycles. Columbia Economic Review | 155

Jessie Dickens

It is widely documented in the literature that hosting costs and cost overruns are exorbitant and have rapidly increased in recent cycles (Flyvbjerg and Stewart, 2016; Zimbalist, 2015). Our model demonstrates that higher costs have an adverse effect on bid probabilities, likely because they represent increased risk levels for a potential bidder. Hosting the Olympics is a huge economic gamble and lots of publicity follows if the event is not successful financially (Zimbalist, 2015). Since large quantities of public funds are used to finance many of the preparations needed to stage the event, governments and bid committees must be able to justify the expenditures as a worthy use of taxpayer dollars. As reported costs increase, this becomes a more difficult task, driving down the probability of ultimately bidding. Finally, a city’s prior behavior also dictates future behavior in terms of bid probabilities. This number of previous bids a city has submitted has a negative relationship with bid probability. This would suggest that each time a city bids for the Olympics it decreases the probability that it will bid again in future cycles. This is likely for two reasons. If a city bids and is selected to host the Games, it is highly unlikely to submit another bid in the close future (a unity value for our HostPrev20 variable perfectly correlates with Bid=0). Very few cities have hosted the Games more than once and those that have typically do so decades (if not centuries in the case of Athens) apart. This, coupled with the cost of bidding also likely decreases the probabilities of bidding for multiple Games (Giesceke and Madden, 2011). Thus, a city that pours a significant amount of resources into an unsuccessful bid may not continually submit bids if they do not feel that they can be successful. However, there have been many cases of cities that will bid in the very next cycle after submitting an unsuccessful bid. This is evidenced by the significance of the BidPrevCycle variable. In recent history, large, developed cities such as Tokyo, Paris and Los Angeles have submitted bids in cycles directly seceding a cycle where they bid unsuccessfully. It is worth noting that the second bids have been generally successful (as evidenced by their selection to host the 2020, 2024 and 2028 Olympic Games) and thus these cities were likely confident in their abilities to be selected as hosts. These cities also likely had the resources to maintain and build on the preparations that they had made in the previous bidding cycle.

VI. Conclusion Our random effect, panel logit model provides key insights regarding the ﬂuctuation of bid probabilities for hosting the Summer Olympic Games, one of the largest, most extravagant and economically impactful mega-events in modern history. Using a panel of 78 cities across 13 time periods, we examine 23 explanatory variables and their impact on bid probability. The model suggests that the ﬂuctuating number of bids across time, including years where 156

Changing Traditions and the Olympic Games

there may be just one bidder, do not occur simply randomly and many factors impact the increasing or decreasing probability rates across cycles. To our knowledge, the present study is the first to examine bidding for the Olympics through the lens of probabilistic modeling and the model provides insight into the typical profile of a likely bidder as well as a variety of universal and city-specific factors that affect bidding probability in a given bid cycle. The resulting information is useful for scholars and practitioners as it sheds light on the thought process bid committees may engage in while deciding whether or not to bid, which could ultimately aid in policy creation that will make bidding for the Olympics more accessible and economically feasible. While there is a vast amount of literature that studies the ex-post economic effects of hosting the Olympic Games, literature that studies the bidding process and the ex-ante variables is sparse. Additionally, there are very few studies in any realm that examine the ex post or ex ante effects surrounding the bidding and hosting of the Olympic Games over time, with many studies using a case study approach on a solitary occurrence of the Games. The longitudinal structure of our data provides a unique opportunity to study bo th of these little-studied areas. While it is notably difficult to draw conclusions as to the single most important factors that determine the probability of submitting a bid at any one point in time given the diverse composition of potential cities and changing dynamics over time, our model proposes three main findings: that bid probabilities are dynamic across time and fluctuate cyclically, that regional bid probability rates have shifted in recent years, and that many different factors play a role in determining bid probability. We show that development of a city both socially and economically as well as many of the inherent economic risks and potential benefits associated with hosting that are discussed in previous literature surely play a role in each committee’s decision to submit a bid. Our model is not without limitations. It is a difficult undertaking to capture exactly how the variables impact the probability of submitting a bid across time in the aggregate because each potential bidder and each Olympic bidding cycle is unique. The bidding environment that exists today is not exactly the same as it was a few decades ago and this makes it difficult to draw conclusions about what factors are at the forefront of impacting the probability to bid. Technological advancements also make it difficult to attain data that is present for all Games by which you can compare their decision-making process. For example, a large consideration in the decision to host today’s Olympic Games is the level of revenue share from television and media rights deals that go to the IOC versus the share that goes to the host city. This isn’t a consideration that would have been taken into account several decades ago, so this makes comparisons difficult and also limits the ability to find a perfect model that determines the factors that make up a country’s decision to bid. Columbia Economic Review | 157

Jessie Dickens

However, the variables chosen for our model are variables that are common across the time frame selected and that provide an accurate basis for comparison. Another challenge in determining bid probabilities is the inclusion of qualitative factors in the modeling process. Cities that are considering a bid to host the Olympic Games often take a perceived “legacy effect” into consideration when making the decision to bid or not (Kasimati and Dawson, 2009). Hosting the Olympics provides a large stage on which a host city can showcase itself to the world. Given the historical significance of the Games, cities are often eager to have the opportunity to carry on the tradition of hosting and showcase their culture to the world. Governments also likely take the “feel good” effect into consideration when pushing to submit a bid for the Olympic Games or any mega-event (Maennig and Porsche, 2008). Governments also seek to boost public support for future elections and have this as an ulterior motive they consider when making the decision to submit a bid. These factors are difficult to quantify which makes it likely that they play a role that our model fails to capture. While we comment on qualitative factors as a limitation of our model, future research should attempt to capture these effects. Future research should also examine bid probabilities for other largescale mega-events, such as the FIFA World Cup or Winter Olympic Games, to see if the variables that impact the probability of submitting a bid for the Summer Olympic Games also impact the probability of bidding for these events. Hosting these events provides similar challenges as hosting the Summer Olympics but the preparations and staging of each event are unique (Mills and Rosentraub, 2013). Additionally, this model should be applied to future Summer Olympics bidding cycles as they occur, to continue to test its robustness and predictive capacity. In order to maintain the long-term success of the Olympics, it will be necessary to address the rising costs of hosting the Games which IOC is already aware of, as evidenced by the strategic planning information contained within their Annual Review and their Olympic Agenda 2020 (International Olympic Committee, IOC Annual Report 2017, September 24, 2018). Finding ways to decrease the costs of bidding and hosting will likely entice more nations to bid from all parts of the world and make hosting the Olympics more accessible and financially feasible. Our model suggests that increasing the probability of bidding and alleviating the situation that occurred with the 2024 and 2028 Olympic bidding cycles will come if the IOC is able to make necessary adjustments and provide more economic as well as sociocultural benefits to nations that are considering a bid.

158

Changing Traditions and the Olympic Games

References Baade, R., and V. Matheson (2016). Going for the Gold: The Economics of the Olympics. The Journal of Economic Perspectives, 30(2), 201-218. Baade, R., and V. Matheson (2002). Bidding for the Olympics: Fool’s Gold. Transatlantic Sport: The Comparative Economics of North American and European Sports, 127-151. Barclay, J. (2009). Predicting the Costs and Benefits of Mega-Sporting Events: Misjudgment of Olympic Proportions? Economic Affairs, 29(2), 62-66. Bob, U., and C. Potgieter (2013). Mega-events and Tourism Impacts: Foreign Visitor Perceptions of the 2010 FIFA World Cup in South Africa. Journal of Human Ecology, 43(1), 71-82. Bruckner, M., and E. Pappa (2015). News Shocks in the Data: Olympic Games and Their Macroeconomic Effects. Journal of Money, Credit and Banking, 47(7), 1339-1368. Burgo, E., and F. Cromartie (2018). The Benefits of Bidding and Hosting the Olympic Games are Difficult to Justify Due to the Overall Costs. The Sports Journal, 20, 1-1. Coates, D., and B. Humphreys (2003). The Effect of Professional Sports on Earnings and Employment in the Services and Retail Sectors in US Cities. Regional Science and Urban Economics, 33(2), 175-198. Coates, D. and V. Matheson (2011). Mega-events and housing costs: raising the rent while raising the roof? The Annals of Regional Science, 46(1), 119-138. Crompton, J. (1995). Economic Impact Analysis of Sports Facilities and Events: Eleven Sources of Misapplication. Journal of Sport Management, 9(1), 1435. de Nooij, M. and M. van den Burg (2018). The Bidding Paradox: Why Politicians Favor Hosting Mega Sports Events Despite the Bleak Economic Prospects. Journal of Sport & Social Issues, 42(1), 68-92. Feddersen, A., W. Maennig, and P. Zimmermann (2007). How to Win the Olympic Games – The Empirics of Key Success Factors of Olympic Bids. Hamburg Contemporary Economic Discussions, 2, 1-21. Flyvbjerg, B. and A. Stewart (2012). Olympic Proportions: Cost and Cost Overrun at the Olympics 1960-2012. Said Business School, 1-23. Flyvbjerg, B. and A. Stewart (2016). The Oxford Olympics Study 2016: Cost and Cost Overrun at the Games. Said Business School, WP 2016-20, 1-28. Giesecke, J. and J. Madden (2011). Modelling the Economic Impacts of the Sydney Olympics in Retrospect - Game Over for the Bonanza Story? Economic Papers: A Journal of Applied Economics and Policy, 30(2), 218-232. Grohmann, Karolos. “IOC Will Not Exclude Asian Cities from 2026 Games Bid.” Reuters. February 06, 2018. Accessed December 1, 2018. https://www.reuters.com/article/us-olympics-2018-bids/ioc-will-notexclude-asian-cities-from-2026-games-bid-idUSKBN1FQ15M. Columbia Economic Review | 159

Jessie Dickens

Groningen Growth and Development Centre. “Penn World Table V9.0”. October 30, 2017. Accessed November 17, 2018. https://www.rug.nl/ ggdc/productivity/pwt/ Gylgi, S., F. Haelg, N. Potrafke, and J. Sturm (2018). The KOF Globalisation Index- Revisited. Review of International Organizations (forthcoming). Hotchkiss, J., R. Moore, and S. Zobay (2003). Impact of the 1996 Summer Olympic Games on Employment and Wages in Georgia. Southern Economic Journal,69(3), 691-704. International Olympic Committee. “IOC Commissions Reports”. January 18, 2018. Accessed November 15, 2018. https://www.olympic.org/docu ments/ioc-commissions. International Olympic Committee. “Olympic Games Candidature Process”. January 26, 2018. Accessed November 15, 2018. https://www.olympic. org/all-about-the-candidature-process International Olympic Committee. “IOC Annual Report 2017”. September 24, 2018. Accessed November 15, 2018. https://stillmed.olympic.org/ media/Document%20Library/OlympicOrg/Documents/IOC-Annu al-Report/IOC-annual-report-2017.pdf#_ga=2.51522665.1307754014.1 548645402-1435244120.1548433017 Jasmand, S. and W. Maennig (2008). Regional Income and Employment Effects of the 1972 Munich Summer Olympic Games. Regional Studies, 42(7), 991-1002. Kasimati, E. and Dawson, P. (2009). Assessing the impact of the 2004 Olympic Games on the Greek economy: A small macroeconometric model. Economic Modelling, 26(1), 139-146. Levy, B. and P. Berger (2013). On the Financial Advantage of Hosting the Olympics. International Journal of Humanities and Social Science, 3(1), 11-20. Lipsitz, G. (1984). Sports Stadia and Urban Development: A Tale of Three Cities. Journal of Sport & Social Issues, 8(2), 1-18. Maennig, W. and S. du Plessis (2009). Sports Stadia, Sporting Events and Urban Development: International Experience and Ambitions of Durban. Urban Forum, 20(1), 61-77. Maennig, W. and C. Vierhaus (2017). Winning the Olympic Host City Election: KeySuccess Factors. Journal of Applied Economics, 49(31), 3086-3099. Matheson, V. (2009). Economic Multipliers and Mega-Event Analysis. International Journal of Sport Finance, 4(1), 63-70. Mills, B., and M. Rosentraub (2013). Hosting mega-events: A guide to the evaluation of development effects in integrated metropolitan regions. Tourism Management, 34, 238-246. Nakamura, A. and M. Nakamura (1981). On the Relationships Among Several Speciﬁcation Error Tests Presented by Durbin, Wu, and Hausman. Econometrica, 49(6), 1583-1588. 160

Changing Traditions and the Olympic Games

Owen, J. (2005). Estimating the Cost and Benefit of Hosting Olympic Game: What Can Beijing Expect from Its 2008 Games? The Industrial Geographer, 3(1), 1-18. Pletz, J. “Chicago 2016’s final tally: $70.6M spent on Olympics effort.” Chicago Business. May 17, 2010. Accessed November 15, 2018. http://www. chicagobusiness.com/article/20100517/NEWS02/200038265/chicago2016s-final-tally-70-6m-spent-on-olympics-Effort. Porsche, M. and W. Maennig (2010). The feel-good effect at Mega Sport Events. Public and Private Management Problems informed by the experiences of the FIFA World Cup. International Journal of Business Research, 10(4), 15-29. Preuss, H. (2004). Calculating the regional economic impact of the Olympic Games. European Sport Management Quarterly, 4(4), 234-253. Preuss, Holger. The Economics of Staging the Olympics: A Comparison of the Games 1972-2008. Edward Elgar Publishing, 2004. Roche, M. (2003). Mega-events, time and modernity- On time structures in global society. Time & Society, 12(1), 99-126. Rojas, Raul. “Gaussian Modell Pruning for Linear Regression”. February 20, 2015. Accessed November 15, 2018. https://www.inf.fu-berlin.de/inst/ ag-ki/rojas_home/documents/tutorials/ModelPruning.pdf Rose, A. and M. Spiegel (2011). The Olympic Effect. Economic Journal, Royal Economic Society, 121(553), 652-677. Saidi, K. and S. Hammami (2015). The Effect of Energy Consumption and Economic Growth on Co2 Emissions: Evidence from 58 Countries. Bulleting of Energy Economics, 3(3), 91-104. Shoval, N. (2002). A New Phase in the Competition for the Olympic Gold: The London and New York Bids for the 2012 Games. Journal of Urban Affairs, 24(5), 583-599. Singh, N. and H. Zhou (2016). Transformation of Tourism in Beijing after the 2008 Summer Olympics: An Analysis of the Impacts in 2014. International Journal of Tourism Research, 18(4), 277-285. Sterken, E. (2006). Growth Impact of Major Sporting Events. European Sport Management Quarterly, 6(4), 375-389. Streicher, T., S. Schmidt, D. Schreyer, and B. Torgler (2017). Is it the econo my, stupid? The role of social versus economic factors in people’s support for hosting the Olympic Games: evidence from 12 democratic countries. Applied Economics Letters, 24(3), 170-174. Taks, M., S. Kesenne, L. Chalip and C. Green (2011). Economic Impact Analysis Versus Cost Benefit Analysis: The Case of a Medium-Sized Sport Event. International Journal of Sport Finance, 6(3), 187-203. Tiegland, J. (1999). Mega-events and impacts on tourism; the predictions and realities of the Lillehammer Olympics. Impact Assessment and Project Appraisal, 17(4), 305-317. Columbia Economic Review | 161

Jessie Dickens

World Bank. “World Development Indicators”. The World Bank. November 14, 2018. Accessed November 17, 2018. https://datacatalog.worldbank. org/dataset/world-development-indicators World Economic Forum. “The Global Human Capital Report 2017”. World Economic Forum. September 13, 2017. Accessed November 16, 2018. https://www.weforum.org/reports/the-global-human-capital-re port-2017 Zimbalist, Andrew. Circus Maximus: The Economic Gamble Behind Hosting the Olympic Games. Washington D.C.: The Brookings Institution, 2015.

Appendix A: List of Cities Sampled

162

Changing Traditions and the Olympic Games

Note Host: City has hosted the Olympics Bid: City has bid for but never hosted the Olympics Non Bid: City has never bid for the Olympics.

Columbia Economic Review | 163

Jessie Dickens

Appendix B: Variable and Data Source Listing

Note The Region variable indicates what region/continent the city is located in to capture rotational trends in bidding at the regional level. The following values 164

Changing Traditions and the Olympic Games

were assigned for each region. 1: North America 2: Europe 3: Oceania 4: Asia 5: South America 6: Africa 7: Middle East

Appendix C: Bidding Trends Figure 2: Aggregate Bid Probabilities vs. North American Bid Probabilities

Figure 3: Aggregate Bid Probabilities vs. European Bid Probabilities

Columbia Economic Review | 165

Jessie Dickens

Figure 4: Aggregate Bid Probabilities vs. Asian Bid Probabilities

Figure 5: Aggregate Bid Probabilities vs. Oceanian Bid Probabilities

166

Changing Traditions and the Olympic Games

Figure 6: Aggregate Bid Probabilities vs. South American Bid Probabilities

Figure 7: Aggregate Bid Probabilities vs. African Bid Probabilities

Columbia Economic Review | 167

Jessie Dickens

Figure 8: Regional Bid Probability Summary

Appendix D: Variance-Covariance Matrix

168

Changing Traditions and the Olympic Games

Footnotes Elevated Middle East bid probabilities are likely due to a relatively high proportion of previous bidders in the randomly selected pool of Middle Eastern cities. 1

Columbia Economic Review | 169

EXCHANGE RATE REGIMES, CAPITAL INFLOWS, AND DOMESTIC CREDIT IN EMERGING MARKET ECONOMIES

Caroline Okel Davidson College Abstract: Capital inflows can increase credit, leading to subsequent financial instability. Blanchard, Ostry, Ghosh, & Chamon (2017) extend a Mundell-Fleming model to show that certain capital flows to emerging markets increase domestic credit, a finding that contradicts the standard Mundell-Fleming prediction. I incorporate fixed exchange rate regimes into their framework, predicting that the effect of capital inflows on domestic credit is larger in fixed regimes than in floating regimes. My empirical findings align with these predictions. This research strengthens the view that fixed regime policymakers should be more concerned than policymakers in floating regimes about the relationship between capital inflows and credit bubbles.

I would like to express my appreciation to Dr. Caleb Stroup and Dr. Mark Foley for their guidance, support, and encouragement throughout this research.

I. Introduction

uring the past ten years, emerging market economies (EMEs) have experienced recurrent periods of large foreign capital inflows. After the Great Recession, interest rates in advanced economies dropped, so investors increased their interest in EMEs to take advantage of the higher rates of return in these nations. This led to a surge in foreign capital flows to these countries, a phenomenon referred to as a capital inflow “bonanza” in the literature. Some EME policymakers believe that capital flow cycles play a role in credit cycles, and bonanzas have been associated with credit booms (Mendoza and Terrones (2008, 2012), Elekdag and Wu (2013), Calderón and Kubota (2012)). Many scholars have suggested that credit booms increase the probability of a subsequent financial or banking crisis (Elekdag and Wu (2013), Abiad, Li, & Dell’Ariccia (2011), Mendoza and Terrones (2012)). Recent examples of funds flowing out of countries such as Argentina and Turkey and the economic turbulence that these outflows have caused serve as examples of this chain reaction (Cecchetti and Schoenholtz 2018). Thus, if capital inflow bonanzas cause credit booms and credit booms cause crises, understanding the relationship between capital flows and credit may lead to more effective policy to prevent recessions. Columbia Economic Review | 171

Caroline Okel

A traditional Mundell-Fleming model predicts that an increase in capital inflows will decrease a country’s ratio of domestic credit to GDP (henceforth domestic credit). Incorporating both safe and risky assets into a traditional Mundell-Fleming model, Blanchard et al. (2017) attempt to reconcile theory with the beliefs of EME policymakers. The authors refer to safe assets as bonds and too risky assets as non-bonds. They find that both bond and non-bond inflows appreciate a country’s currency, which puts downward pressure on domestic credit. The authors label this mechanism the “appreciation channel,” and it is the mechanism observed in the traditional Mundell-Fleming model. However, Blanchard et al.’s (2017) model also incorporates a second mechanism: the “financial intermediation” channel. An increase in demand for risky assets decreases their rate of return. Thus, the risk premium decreases, and firms can borrow at a reduced cost. This stimulates demand for credit, increasing domestic credit in the economy. The authors presume that for risky flows, the financial intermediation channel will dominate the appreciation channel in EMEs and hypothesize that these non-bond flows will increase domestic credit. But, bond inflows impact domestic credit only through the appreciation channel, implying that bond inflows will decrease domestic credit. My model produces two novel predictions. First, non-bond flows are expected to increase domestic credit more in fixed regimes than in floating regimes. This prediction arises from the fact that by definition, the exchange rate cannot appreciate in fixed regimes. Thus, non-bond inflows only affect domestic credit through financial intermediation channel and do not cause any downward pressure on domestic credit through the appreciation channel. Second, in fixed regimes, bond inflows are not expected to influence domestic credit. Through their theory, Blanchard et al. (2017) argue that bond flows to EMEs only affect domestic credit through the appreciation channel but not through the financial intermediation channel. Thus, in fixed regimes, if there is no impact through the appreciation channel, then there is no mechanism for bond flows to influence domestic credit. Using a sample of 34 EMEs from 2000 to 2017, I find that non-bond flows to fixed regimes are associated with an increase in domestic credit. The effect of non-bond inflows on domestic credit in fixed regimes is greater than the effect in floating regimes, and this difference is statistically significant. This finding provides suggestive evidence in favor of prediction (1). The effect of bond inflows on domestic credit in fixed regimes is statistically indistinguishable from zero, providing preliminary evidence in favor of prediction (2). Therefore, the predictions of my model are consistent with the findings of previous research, and these predictions appear to hold in my empirical analysis. These findings thus align with the view that fixed regimes are more susceptible to credit booms and busts than floating regimes. This supplements Magud, Reinhart, & Vesperoni’s (2014) and Magud and Vesperoni’s (2015) 172

Emerging Market Economies

assertions that fixed regimes benefit relatively more from capital controls and measures to contain excessive credit growth during capital inflow bonanzas. My research suggests that EME policymakers should be concerned about the relationship between capital inflows and domestic credit. Blanchard et al. (2017) state that EME policymakers should consider policies such as capital controls and macro-prudential tools to manage the effects that capital inflows and subsequent credit booms can have on EMEs. However, these authors assume perfect flexibility of exchange rates, an implausible assumption for many developing economies. One might thus wonder whether their prescription of stronger capital controls extends to cases where exchange rates are not perfectly flexible. My research theoretically demonstrates that capital inflows can increase domestic credit in fixed regimes, and my empirical findings are consistent with this result. Therefore, the implications of Blanchard et al.’s (2017) model are robust to assumptions about exchange rate regime. Section II gives an overview of studies about capital inflows, credit booms, and exchange rate regimes in EMEs. Section III presents my modified version of Blanchard et. al’s (2017) model of capital flows to EMEs. Section IV presents my data and empirical findings. Sections V and VII are my conclusion and references list, respectively.

II. Background I describe three strands of literature: (1) research about capital inflow dynamics in EMEs, (2) studies about the relationship between credit booms and capital inflow cycles, and (3) research about how exchange rate regimes impact the relationship between credit booms and capital inflow cycles. The first and second strands of literature are summarized to demonstrate how capital inflow bonanzas and credit booms may negatively affect EMEs. These negative effects motivate the need to study the relationship between inflows and credit. In the third section, I summarize research about how exchange rate regimes influence this relationship between inflows and credit, and this research most closely correlates with my research question. A. Why might EME policymakers restrict capital flows? Capital inflow surges to EMEs can exacerbate vulnerabilities associated with already fragile economies. Capital inflows are often cyclical and can have detrimental results upon reversal. Reinhart and Reinhart (2008) examine the pattern associated with these cycles: typically, foreign investors target a certain developing country, and capital surges into the market. These inflows appreciate the currency and increase asset prices. Policymakers in EMEs tend to treat these inflow periods as permanent phenomena, but when investors lose interest in the economy, the flows reverse, asset prices fall, and the economy faces a painful recovery (Reinhart and Reinhart 2008). Columbia Economic Review | 173

Caroline Okel

On one hand, capital inflows are beneficial for EMEs because they finance investment and foster economic growth. However, periods of heavy inflows may also reinforce shallow financial sectors (Magud and Vesperoni 2015). Specifically, inflow periods may result in excessive monetary and credit expansions, vulnerabilities associated with currency and maturity mismatches, and distortions in asset prices (Magud et al. 2014). These bonanzas are especially problematic in developing countries because policymakers and investors are more prone to treat inflow periods as permanent phenomena. Surges in government spending during bonanzas serve as evidence of this attitude (Reinhart and Reinhart 2008). Furthermore, Reinhart and Reinhart (2008) find empirical evidence that associates bonanza periods with a higher incidence of banking, currency, and inflation crises in all but high-income nations. This demonstrates the fragility of emerging markets during inflow periods. Using a panel of developed and emerging economies from 1970 to 2007, Furceri, Guichard, & Rusticelli (2011) find that a large capital inflow episode almost doubles the probability of a banking or currency crisis occurring in the following two years. However, capital inflow surges are not inherently problematic. If economies avoid excessive credit growth, economic overheating, and currency overvaluation and regulate inflows through capital controls, surges are less likely to end in crisis (Ghosh, Ostry, & Qureshi 2016). Capital inflow bonanzas can provoke crises through several channels. First, bonanzas result in large exchange rate appreciations due to increased demand for domestic currency, and these appreciations can lead to overvaluations and put downward pressure on international trade. Second, large inflows of capital set the stage for an inflow reversal (Furceri et al. 2011). These inflow reversals, known as “sudden stops,” can be especially damaging in economies with policymakers who behave as if the bonanza will last forever. Sudden stops trigger large exchange rate depreciations and financial disruptions and decrease rates of return, investment, and growth (Calvo, Izquierdo, & Meíja 2004). Agosin and Huaita (2011) find that a surge in capital inflows increases the probability of a sudden stop by between 5.7 and 9.3 probability points in the following year. Third, there are reasons to believe that capital inflows to EMEs are associated with credit booms. I will cover this relationship in the following section. B. Do capital inflows to EMEs cause credit booms? Credit booms are periods of time in which credit to the private sector grows more than the typical rate, and they have been shown to be associated with inflow bonanzas. Using a panel of 49 countries over the period 1960-2006, Mendoza and Terrones (2008) find that large capital inflows preceded over 50 percent of credit booms in EMEs but only 27 percent of credit booms in developed economies. This finding suggests that the relationship between credit booms and capital inflows is especially pronounced in developing countries. 174

Emerging Market Economies

Several scholars have conducted empirical work demonstrating the frequency of credit booms following capital inflow bonanzas. Mendoza and Terrones (2008, 2012), Elekdag and Wu (2013), and Calderón and Kubota (2012) all find evidence suggesting that capital inflow booms usually precede credit booms. Calderón and Kubota (2012) specify that inflows consisting of portfolio investments are more predictive of a credit boom than FDI. Furthermore, these authors find that credit booms preceded by private other investment or portfolio investment booms are more likely to end in a banking crisis. Blanchard et al. (2017) find that FDI flows actually have a significant negative effect on credit. They hypothesize that when FDI increases, FDI financing replaces intermediation that would have taken place through banks. The authors separate portfolio flows into two categories: bonds and non-bonds. They find that bond inflows have a negative and insignificant effect on credit growth but that non-bond inflows have a positive and significant effect on credit growth. These results correspond with my model, outlined in the following section. It is important for policymakers to understand if inflow bonanzas cause credit booms because of the negative effects that credit booms can have on EMEs. Credit booms in EMEs can be problematic because they can exacerbate vulnerabilities in already fragile economies. In a sample of sixty emerging market credit booms, Elekdag and Wu (2013) find empirical evidence that sharp reversals in macroeconomic variables such as real GDP, domestic demand, and investment follow credit surges. Abiad et al. (2011) find that economies in economic recovery are two times more likely to have negative real credit growth following the recession if a credit boom precedes the recession.1 Even though all credit booms do not end in crisis, a significant amount do. Using a panel of 70 credit booms, Mendoza and Terrones (2012) find that roughly 20-25% of credit booms in both emerging markets and developed countries were followed by a banking or currency crisis, and 14% of credit booms were followed by a sudden stop. Finally, surges in private sector lending during credit booms increases corporate leverage, and this leverage can increase the economy’s sensitivity to slower growth, higher interest rates, and weaker exchange rates (Sahay, Arora, Arvanitis, Faruqee, N’Diaye, & Mancini-Griffoli 2014). Therefore, if a sudden stop occurs, the economy may face an even more severe recovery period. Because there is evidence suggesting that credit booms cause crises, it is important to understand the factors that lead to credit booms. It has been suggested that capital inflow bonanzas cause credit booms, so understanding the mechanisms behind this relationship may allow policymakers to make more informed decisions. If capital inflow bonanzas to EMEs cause credit booms and credit booms cause crises, capital controls on EMEs or tools to manage credit expansions may be effective policy tools to prevent recessions. Columbia Economic Review | 175

Caroline Okel

C. Why might exchange rate regime influence domestic credit? Empirical findings suggest that the relationship between inflows and credit is stronger in fixed than in floating regimes (Magud et. al (2014); Magud and Vesperoni (2015)). If exchange rate regimes play a role in the relationship between capital flows and credit, appropriate policy responses to capital inflows may vary from fixed regimes to floating regimes. In this section, I provide an overview of exchange rate regimes in EMEs and summarize research on the relationship between capital flows and credit in the context of exchange rate regime. Studies suggest that exchange rate regimes impact domestic credit during the capital flow cycle, and economists have posed multiple explanations for this relationship. Some believe that this relationship exists solely because exchange rate regime impacts capital flow dynamics. Because fixed regimes signal guaranteed exchange rate stability to investors, fixed economies may attract larger volumes of capital inflows (Magud et al 2014). However, empirical evidence rejects this hypothesis. Boudias (2015) examines the entire capital flow cycle, finding that exchange rate regime does not impact the cycle. Furthermore, Ghosh et. al (2014) find empirical evidence that ceteris paribus, changes in domestic credit over a three-year period are almost twice as large in hard peg regimes as in intermediate regimes. Obstfeld et. al (2017) find that global financial shocks affect domestic credit and banking sector leverage more in fixed regimes than floating regimes. These findings imply that exchange rate regime does not alter credit dynamics solely through capital flows, so exchange rate regime must impact domestic credit through a different channel. Therefore, all three factors, exchange rate regime, capital flows, and credit dynamics, must be considered jointly. Magud et al. (2014) examine this relationship empirically. They use a panel of data on twenty-five emerging markets in Asia, Europe, and Latin America during periods of large capital inflows. The authors find that countries with less flexible exchange rate regimes have larger shares of domestic bank credit. They briefly cite possible explanations for this phenomenon. Committing to a peg limits the central bank’s ability to curb monetary expansions that follow capital inflow booms. This absence of monetary policy tools limits the central bank’s ability to intervene before a credit boom occurs. Furthermore, during inflows, a fixed regime requires the central bank to accumulate foreign currency reserves to prevent currency appreciation. The intermediation of these flows by banks may increase credit expansions. The authors also cite Montiel and Reinhart (2001), who claim that pegged exchange rates signal a guarantee of foreign currency claims to banks, increasing their scope for lending abroad. These are not detailed explanations, but they justify the authors’ empirical results. 176

Emerging Market Economies

Magud and Vesperoni (2015) expand on this research by studying the entire capital flow cycle. They examine the impact of capital inflow reversals on credit behavior across different exchange rate regimes. Their findings suggest that more flexible regimes experience milder credit swings during capital flow cycles, and the authors contribute this result to the milder credit growth observed in bonanza periods. This flexibility does not completely protect these economies from credit reversals during a sudden stop. Flexible regimes also face the problem of minimal credit growth in the period following an inflow reversal.3 Empirically, the authors find that economies with fixed regimes have a milder adjustment in consumption during reversals, validating this theory. These two pieces of research aligns with my research the closest. Both works provide evidence that capital inflow bonanzas and sudden stops impact the credit cycle more in fixed regimes than in floating regimes. In the next section I develop a model to connect macroeconomic theory to the results observed by Magud et al. (2014), Magud and Vesperoni (2015), and the other economists discussed above.

III. Theory I extend a model developed by Blanchard et al. (2017) to examine how different types of capital inflows affect domestic credit. Blanchard et al.’s (2017) model assumes a floating exchange rate regime, so I expand on their model to test if it is robust to assumptions about exchange rate flexibility. The predictions of a traditional Mundell-Fleming model suggest that capital inflows should decrease domestic credit. As capital inflows increase, demand for the currency increases. This appreciates the currency, which decreases net exports and output. Assuming that the central bank determines the policy rate exogenously and does not cut rates in response to this decrease in output, investment will also decrease. Thus, demand for credit will decrease, which leads to a decrease in the ratio of domestic credit to GDP, despite the decrease in GDP. Empirical findings and the observations of policymakers suggest that this prediction does not apply to EMEs. Blanchard et al. (2017) attempt to build on the Mundell-Fleming framework to account for these empirical observations. They extend the set of assets to include both bonds and non-bonds. After making this extension, the model suggests that capital inflows influence domestic credit through two channels: (1) an appreciation channel, and (2) a financial intermediation channel. The “appreciation channel” is the mechanism observed in a traditional Mundell-Fleming model. Through this channel, capital inflows decrease domestic credit due to a currency appreciation. The “financial intermediation” channel occurs as following. An increase in demand for risky assets decreases their rate of return. Thus, the risk premium decreases, and firms can borrow at a reduced cost. This stimulates demand for credit, increasing domestic credit in the economy. Columbia Economic Review | 177

Caroline Okel

Blanchard et al. (2017) argue that in EMEs, a decrease in the cost of financial intermediation will have a big effect due to the need for development in the banking sector. Thus, their theory suggests that non-bond flows will increase domestic credit and bond flows will decrease domestic credit. Blanchard et al.’s (2017) empirical results are consistent with predictions about non-bond flows, but the estimated effect of bond flows is not statistically distinguishable from zero. I extend this model to see if the effects of inflows on credit differ in fixed regimes from floating regimes. Blanchard et al.’s (2017) model implicitly assumes a floating exchange rate regime; there can be no appreciation channel if the exchange rate regime is fixed. My extension is necessary because EMEs are more likely to be on the fixed end of the exchange rate regime spectrum than developed economies. I solve for the ratio of domestic credit to GDP (λ), my key endogenous variable, in a fixed exchange rate regime and in a floating regime, differentiating between the two by taking into account their differences in central bank intervention. My comparative statics exercise explores how λ reacts to increases in both bond inflows and non-bond inflows. The model predicts that non-bond inflows will increase domestic credit more in fixed regimes than in floating regimes. This prediction arises from the fact that by definition, the exchange rate cannot appreciate in fixed regimes. Thus, non-bond inflows only affect domestic credit through the financial intermediation channel and do not cause any downward pressure on domestic credit through the appreciation channel. The model also predicts that bond inflows will not influence domestic credit. Blanchard et al. (2017) argue that bond flows to EMEs only affect domestic credit through the appreciation channel but not the financial intermediation channel. Thus, in fixed regimes, if there is no impact through the appreciation channel, then there is no mechanism for bond flows to influence domestic credit. These predictions rationalize the empirical findings of Magud et al. (2014) and Magud and Vesperoni (2015). These authors find that fixed regimes are more susceptible to credit booms in the face of capital inflow bonanzas. Thus, my model contributes to the existing literature because it produces a framework to understand why certain types of capital inflows may especially increase domestic credit in fixed regimes. A. Financial Market To develop a financial market, I build on Blanchard et al. (2017). In the model, there are four assets which are all imperfect substitutes: money, domestic bonds, foreign bonds, and domestic non-bonds. Monetary policy controls the rate of return on bonds. Bonds include non-risky assets such as sovreign 178

Emerging Market Economies

debt. Non-bonds include domestic equities, FDI, domestic bank liabilities, and short-term corporate bonds. These are riskier assets than bonds. The policy rate does not affect the rate of return on non-bonds, and this rate of return can be impacted by capital inﬂows. Let RB be the rate of return on domestic bonds, RN be the rate of return on non-bonds, and RFB be the rate of return on foreign bonds in foreign currency. Let E be the exchange rate with . Thus, an increase in E denotes an appreciation of the homecurrency. Let E+1 be the expected exchange rate. The rate of return on foreign bonds in the home currency will be . W denotes domestic wealth. Following Blanchard et al. (2017), let demand for money by domestic investors be given by: The sensitivity of investors to changes in the rate of return on domestic bonds is given by α1. As the rate of return on domestic bonds increases, investors save more and their demand for money decreases. Autonomous money demand is given by α0. Demand for domestic bonds by domestic investors is: Demand for domestic non-bonds by domestic investors is:

Demand for foreign bonds by domestic investors is:

Domestic investors spend all their net wealth on domestic bonds, domestic non-bonds, and foreign bonds. Thus, a + b + c = 1 because these constants designate the autonomous portion of net wealth allocated to each asset. Demand for each asset is proportional to domestic wealth net of money demand and also depends on the rate of return relative to the other two asset classes. For simplicity, the sensitivity of each investor to changes in relative rates of return equals β for all asset classes. Let initial holdings of each asset class equal M̅D , B̅D, N̅D, and B̅*D. It is only important to consider foreign investors’ demand for domestic bonds and domestic non-bonds. Demand for domestic bonds by foreign investors is: Demand for domestic non-bonds by foreign investors is:

Columbia Economic Review | 179

Caroline Okel

Demand for domestic assets by foreign investors depends on wealth net of foreign money demand and the relative rates of return of each asset. For simplicity, sensitivity of foreign investors to changes in relative rates of return is β. However, foreign demand for bonds and non-bonds differs from domestic demand. Shifts in foreign demand forbonds, SB, and shifts in foreign demand for non-bonds, SN, capture this difference. These terms serve as the source of capital inflows in the model. Let initial foreign holdings of each asset class equal B̅F and N̅F. A model for a floating exchange rate regime is developed, and then the model is expanded to account for fixed exchange rate regimes. In a floating regime, the central bank does not manage the exchange rate by accumulating foreign reserves. The central bank chooses the money supply, MS, and its supply of domestic bonds, BCB, such that

1. Equilibrium in Floating Exchange Rate Regimes

In equilibrium, money supply must equal money demand, domestic bond markets will clear, domestic non-bond markets will clear, and the capital account will balance. These conditions are represented with the following equations:

Walras Law permits me to drop the equation concerning the bond market. Note from equation (1) that money demand solely depends on R B. Substituting equations (1) and (7) into equation (8), This equation demonstrates that foreign capital inﬂows do not impact RB. Thus, equation (8) is dropped as well. The equations for non-bonds and the capital account are used to solve for equilibrium RN and E. For simplicity, net wealth for foreign and domestic investors are the same, so (W - MD)=(W* - M*D). Note from the equations above that b represents the autonomous share of net wealth that domestic investors designate to domestic non-bonds, and f represents the autonomous share of net wealth that foreign investors designate to domestic non-bonds. If net wealth is equivalent for both foreign and domestic investors, then (b+ f)(W - MD) will equal the autonomous volume of non-bonds demanded by the market. This autonomous demand for non-bonds equals the initial holdings of non-bonds. 180

Emerging Market Economies

Similarly, the autonomous volume of domestic assets demanded by foreign investors, (d + f)(W + MD), equals the initial holdings of domestic assets by foreign investors. Lastly, the autonomous volume of foreign bonds demanded by domestic investors equals the initial holdings of foreign bonds by domestic investors.

Using these assumptions, equations (10) and (11) simplify to:

Solving for RN and E yields

In a ďŹ&#x201A;oating exchange rate regime, the central bank determines RB exogenously. RN depends positively on RB and SB but negatively on SN. E depends positively on E+1, RB, SB, and SN but negatively on RFB. Equation (18) has an intuitive interpretation: if the interest rate on domestic bonds increases, the rate of return on riskier non-bonds will also increase. If foreign demand for domestic non-bonds increases, then the price of non-bonds will increase and their rate of return will decrease. If foreign demand for domestic bonds increases, then demand for domestic non-bonds will decrease, prices will drop, and rates of return on domestic non-bonds will increase. Equation (19) also has an intuitive interpretation. When foreign demand for domestic assets increases, foreign investors demand more domestic currency, leading to an appreciation. Similarly, if the policy rate, RB, increases, more investors will demand currency to take advantage of this rate of return.

Columbia Economic Review | 181

Caroline Okel

2. Equilibrium in Fixed Exchange Rate Regimes

In a fixed regime, the government or the central bank sets a rate at which it will buy or sell foreign exchange. Therefore, the central bank fixes the price of the home currency against some foreign currency. The central bank maintains this fixed rate, or “peg”, through the mechanism of foreign intervention. The central bank intervenes through large scale sales and purchases of foreign exchange. For example, if the nominal exchange rate appreciates and the home currency becomes stronger, the central bank will buy foreign currency and sell domestic currency so that the fixed rate remains unchanged. If the nominal exchange rate depreciates, the central bank will sell foreign currency and buy domestic currency. Thus, an identical shift in the supply of the currency counteracts any change in the demand for currency that alters the nominal exchange rate. Central banks use foreign bonds to buy and sell foreign currency. Therefore, in a fixed exchange rate regime, the central bank’s demand for foreign bonds impacts the balance of payments equation. Let B* CB denote the central bank’s demand for foreign bonds and B*CB denote the central bank’s initial holdings of foreign bonds. In equilibrium, the equation concerning the non-bond market will remain unchanged. However, the equation concerning the capital account differs for a fixed regime because the central bank contributes to capital outflows when they purchase foreign bonds. Let denote the capital outflows made by the central bank. Therefore, the capital account equilibrium condition will equal

Solving these two equations yields:

The exchange rate will equal the exogenous rate determined by the central bank. Let E denote this pegged rate. The exchange rate never ﬂuctuates, so E = Ê. Because the central bank guarantees that E = Ê, expected depreciation for the exchange rate will always equal zero. Therefore, E = Ê = E+1. Using equation (23), outﬂows by the central bank must ensure that E = E +1. In order for this equation to hold, 182

Emerging Market Economies

This requirement implies that in equilibrium,

In a fixed regime, the central bank determines the policy rate and the exchange rate exogenously. RN depends positively on R* and RB and negatively on SN. Equation (26) makes intuitive sense. If the central bank increases the policy rate, the rate of return of riskier assets will also increase. Furthermore, an increase in the rate of return of foreign bonds will increase the rate of return of domestic non-bonds, as investors will move away from domestic non-bonds and towards foreign bonds. The rate of return on non-bonds depends negatively on foreign demand for non-bonds because as demand for non-bonds increases, the price of non-bonds will increase and interest rates will decrease. The findings for a fixed regime primarily differ from floating regimes because the rate of return on non-bonds does not depend on foreign demand for bonds. An increase in foreign demand for bonds appreciates the currency. Under the fixed regime model, the central bank will respond to this appreciation by matching the bond inflow with an equivalent outflow of capital. This outflow counteracts the impact that foreign demand for bonds would have had on the rate of return on non-bonds. However, in response to a non-bond inflow, the central bank only responds with a capital outflow half as large, allowing non-bond inflows to have 15 an effect on the rate of return of non-bonds. B. Goods Market 1. The Goods Market

Investment, consumption, government spending, and net exports comprise the goods market. Let I denote equilibrium investment.

In the investment sector of the economy, businesses purchase capital and individuals buy houses. The expected value of ﬁrm sales and the cost of capital both affect future investment. Firms only purchase capital if they expect to make sales, so the expected value of sales positively impacts investment. Firms will purchase less capital if capital becomes more expensive, so investment depends negatively on the cost of capital.

Columbia Economic Review | 183

Caroline Okel

The expected value of sales is equivalent to the total output of the economy. The interest rate charged to borrowers serves as an approximation of the cost of capital because capital becomes more expensive when interest rates increase. Businesses receive capital through risky ďŹ nancing methods and face interest rates RN. Substituting (29) and (28) into (27) gives: This relationship is represented with the equation: Let C denote equilibrium consumption. Consumption decisions depend on disposable income and interest rates. Disposable income is the amount of income remaining after taxes. When it increases, individuals have 16 more funds to use for purchases, so consumption increases. The policy rate determines interest rates paid to savers. An increase in interest rates paid to savers, RB, decreases consumption because individuals will choose to save more to take advantage of these more favorable rates. This relationship is represented by the equation: Let Yd = Y - T (lump sum taxes). Government spending, G, is an exogenous variable. Let X denote equilibrium domestic exports, and let Y* denote world output. Exports from the domestic country depend on world output and the exchange rate. When world output increases, world income increases, so foreign individuals have the means to consume more goods from the domestic economy. An increase in E denotes an appreciation of the domestic currency. This appreciation means that domestic goods are more expensive for foreigners, so as the exchange rate increases, individuals in foreign economies will import fewer domestic goods. The functional form for this relationship is: Let M denote equilibrium domestic imports.

184

Emerging Market Economies

Imports to the domestic country depend on domestic output and the exchange rate. When domestic output increases, domestic income increases, and citizens have the means to demand more foreign goods. When the exchange rate increases, domestic citizens can buy foreign goods at more favorable prices, so foreign goods become more attractive relative to home goods. The functional form of this equation is:

2. Goods Market Equilibrium

Equations for investment, consumption, government spending, imports, and exports have been derived. In equilibrium, Substituting equations (31), (34), (36), and (38) into (39), This equation simpliﬁes to:

This equilibrium represents combinations of the interest rate and output at which domestic output equals the planned expenditure on goods and services. C. Credit Market Total bank credit equals the quantity of loans in equilibrium. The quantity of loans demanded equals equilibrium investment. Total credit in an economy will be equal to the sum of credit supplied by domestic and foreign banks. Denote domestic credit as CR D and foreign credit as CRF.

λ is the ratio of (44) to (41). D. Solving the Model in Floating Regimes In a ﬂoating regime, if the goods market and ﬁnancial market are in equilibrium, then

Columbia Economic Review | 185

Caroline Okel

Substituting equations (46) and (47) into (45),

To simplify the equation for domestic credit, substitute equations (48) and (46) into (44).

Finally, the ratio of (49) to (48) yields

1. Comparative Statics: Floating Regimes

yields:

Taking the partial derivative of equation (50) with respect to SN and SB

where

b1+c1 < 1, so γ < 1. Because of the size of G and x1Y *,δ> 1. The sign of ε does not impact the results. Note that these two derivatives are identical except (51) is negative and their numerators are slightly different. This difference in numerators is not sufﬁcient to result in different signs within the parentheses. The numerators are positive due to the negativity of γ and the magnitude of δ. The two denominators are equivalent and positive because they are squared.

186

Emerging Market Economies

Hence, because

is negative,

Thus, the model predicts that when a floating regime faces bond inflows, the ratio of domestic credit to output will decrease. This is consistent with Blanchard et. al’s (2017) theory that the appreciation effect will decrease λ. On the other hand, the ratio of domestic credit to output will increase in response to an increase in non-bond inflows. This suggests that the financial intermediation effect dominates the appreciation effect. E. Solving the Model in Fixed Regimes Fixed regimes differ from floating regimes because the central bank participates in the foreign bond market. If the goods market and financial market are in equilibrium, then

Substituting (59) and (60) into (58) yields:

Thus,

The ratio of (62) to (61) yields λ.

1. Comparative Statics: Fixed Regimes

Taking the partial derivative of λﬁ with respect to SN and SB,

Columbia Economic Review | 187

Caroline Okel

where

Because RN does not depend on foreign demand for domestic bonds, changes in SN do not impact λﬁ. Due to the negativity of γ, is positive.

Thus, in fixed regimes, only non-bond inflows impact the ratio of domestic credit to GDP, and the model suggests that an increase in non-bond flows will increase this ratio. Thus, the financial intermediation channel is positive, as expected. F. Testable Implications In Blanchard et. al’s (2017) research, the authors argue that capital inflows affect economies through two channels: currency appreciation and financial intermediation. A currency appreciation will decrease λ, but cheaper financial intermediation will increase λ. The authors argue that in EMEs, financial intermediation will have a bigger effect due to the need for development in the banking sector. They conclude that in EMEs, bond inflows only impact domestic credit conditions through the currency appreciation channel, leading to a decrease in domestic credit. However, while non-bond inflows also appreciate the currency, the impact of cheaper financial intermediation dominates, which increases λ. I extend Blanchard et. al’s (2017) model by relaxing exchange rate assumptions. My model predicts that floating regimes will respond to capital inflows in the same way as predicted by Blanchard et al (2017): bond flows will decrease λ, and non-bond flows will increase λ. In fixed regimes, my model predicts that bond inflows will not impact the ratio of domestic credit to GDP. This result is unsurprising because a government eliminates the appreciation channel when fixing the exchange rate. In fixed regimes, non-bond inflows are expected to increase domestic credit.

188

Emerging Market Economies

I now compare the magnitude of the impact of non-bond inflows on domestic credit in each regime. I hypothesize that the effect of non-bond inflows will be more pronounced in fixed exchange rate regimes than in floating regimes. During non-bond flows to floating regimes, the appreciation effect nets out part of the positive impact of the financial intermediation effect. In fixed regimes, non-bond inflows only impact credit through the financial intermediation channel, so there is no downward pressure on domestic credit as a result of the appreciation channel. Thus, the effect of non-bond inflows on domestic credit will be larger in magnitude in fixed regimes than in floating regimes.

Hypothesis (1): Non-bond inflows increase domestic credit more in fixed regimes than floating regimes. Building on Blanchard et al.’s (2017) framework, non-bond inflows to fixed regimes only influence domestic credit through the financial intermediation channel because there is no appreciation channel. Thus, I expect that non-bond inflows will lead to a more pronounced effect on credit in fixed regimes than in floating regimes. Hypothesis (2): Bond inflows do not influence domestic credit in fixed regimes. Bond inflows to countries with fixed exchange rate regimes do not impact domestic credit through either channel. Hence, I expect that the effect of bond inflows on domestic credit in fixed regimes will be insignificant and small in magnitude. Hypotheses (1) and (2) are consistent with the empirical findings of Magud et al. (2015) and Magud and Vesperoni (2016). These authors find that fixed Columbia Economic Review | 189

Caroline Okel

regimes are more prone to credit booms during inflow bonanzas than floating regimes. If non-bond flows to fixed regimes are associated with a large increase in domestic credit and if bond inflows are associated with a small, non-negative effect on domestic credit, then it is possible that Blanchard et al.’s (2017) framework can be used to explain the empirical findings of Magud et al. (2015) and Magud and Vesperoni (2016). Hypothesis (3a): Non-bond inflows increase domestic credit in floating regimes. Hypothesis (3b): Bond inflows decrease domestic credit in floating regimes. The predictions about floating regimes are consistent with the findings of Blanchard et al. (2017).

IV. Empirics A. Methodology I test the hypotheses of my theoretical model using data on 34 EMEs from 2000 to 2017. I estimate the following regression:

where i indexes countries and t indexes years. The variable λit remains unchanged from my theoretical model, and it takes the value of the ratio of domestic credit to GDP of country i in year t. Following Blanchard et al. (2017), ∆λit represents the year-over-year percent change in λ. I represent country fixed effects with γi and time fixed effects with α t. My parameters of interest are β1, β2, β3, and β4. The parameters β1 and β2 measure the effect of bond inflows on domestic credit, and β3 and β4 measure the effect of non-bond inflows on domestic credit. FBit and FNBit denote bond inflows and non-bond inflows to country i in year t, respectively. Both of these variables are scaled by country i’s nominal GDP. The variable Rit is a binary variable that equals one if country i falls to the fixed side of the exchange rate regime spectrum and equals zero if the country maintains a floating exchange rate regime. Following Blanchard et al. (2017), five control variables are included. The variable λit−1 is lagged domestic credit, and ∆GDP*it is the change in nominal GDP of country i’s trading partners in year t. The variable ∆TOTit controls for the change in i’s terms of trade in year t, DRit is the country’s central bank policy interest rate, and RFit denotes i’s outflow of reserve assets in year t.

190

Emerging Market Economies

Country fixed effects, γi, control for country-specific unobservable factors that do not vary over time. For example, this term may capture aspects of a country such as financial sector sophistication that may impact both domestic credit and capital inflows. Time fixed effects, αt, control for yearspecific aspects of the global economy that impact all countries. Time fixed effects may capture unobservables such as worldwide financial conditions and investor sentiment towards EMEs. B. Data My dataset includes capital inflow and credit data on 34 EMEs from 2000 to 2017. The sample includes the 23 countries grouped as EMEs by the International Monetary Fund: Argentina, Bangladesh, Brazil, Bulgaria, Chile, China, Colombia, Hungary, India, Indone- sia, Malaysia, Mexico, Pakistan, Peru, Philippines, Poland, Romania, Russia, South Africa, Thailand, Turkey, Ukraine, and Venezuela. To further diversify the sample, eleven more EMEs are included that are listed by FTSE, MSCI, S&P, or the EM bond index. These additions include the Czech Republic, Egypt, Greece, Israel, Kuwait, Nigeria, Oman, Qatar, Saudi Arabia, South Korea, and Vietnam. In total, the sample includes 34 countries facing a variety of development conditions and exchange rate regimes. The data on these countries span from the year 2000 to 2017. Due to data restrictions and to reduce noise, I use annual data. Furthermore, annual data provide an extensive enough time period for capital flows to actually impact credit conditions (Blanchard et. al 2017). This strengthens the potential for causal interpretations in the empirical analysis. In total, the sample is comprised of 612 country-year pairs. 1. Domestic Credit

Data on domestic credit come from the World Bank’s World Development Indicators database. The World Bank defines domestic credit as financial resources with a legal obligation for re-payment provided by financial corporations. These financial resources can include loans, purchases of non equity securities, and other accounts receivable. Financial corporations include banks and other monetary authorities. I focus on lending from these institutions to the domestic private sector. The World Bank receives this data on domestic credit from estimates made by the International Monetary Fund. These estimates are the aggregated nominal value of credit instruments provided by domestic banks to the country’s private sector. These values are scaled by each country’s nominal GDP to allow the magnitude of domestic credit to be interpreted relatively. Following Blanchard et al. (2017), the year-over-year percent change in domestic credit is calculated as the dependent variable.

Columbia Economic Review | 191

Caroline Okel

2. Capital Inﬂows

Capital flows data come from the International Monetary Fund’s Balance of Payments Statistics Database. Each country in this database submits financial account data to the IMF in accordance with the IMF’s BPM6 rules. Financial account flows are divided into five different flow types: direct investment (FDI), portfolio investment, financial derivatives, other investment, and reserves. Each flow type is further disaggregated by instrument type and sector. At each aggregation level, capital flows are measured as asset flows, liability flows, and net flows (outflows - inflows). Asset flows differ from liability flows because while asset flows record capital flows from residents to foreigners, liability flows record flows from foreign investors to the country. Positive and negative signs on these flow types do not align with capital inflows or outflows. Rather, a positive asset flow signals that capital has left the country on net by domestic residents, but a positive liability flow signals that capital entered the country on net by foreigners. Therefore, capital inflows result in a negative sign for asset flows but a positive sign for liability flows. I illustrate with an example. In 2000, Brazil recorded FDI asset flows of roughly 2.5 billion USD. This positive sign indicates that Brazilian residents purchased more foreign assets through FDI than they divested or let mature, resulting in capital outflow. The country faced FDI liability flows of over 33 billion USD, which signals a capital inflow. Foreign citizens invested more money through FDI in Brazil than they divested. Researchers typically refer to asset flows as “gross outflows” and to liability flows as “gross inflows.” However, these labels are deceiving because gross inflows and outflows should reflect the total amount of capital that flowed into or out of a country without netting the reverse direction out. Because Brazil’s 2000 FDI asset flows and liability flows subtract flows occurring in the opposite direction, it is more appropriate to label these flows as “net inflows” and “net outflows.” In the BOP dataset, net total flows are calculated as the difference between net inflows and net outflows. Accordingly, Brazil recorded a net FDI flow of -30.5 billion USD (the 2.5 billion asset outflow - the 33 billion liability inflow). Therefore, Brazil faced a net 30.5 billion USD capital inflow through FDI in 2000, and this number was primarily driven by investment in Brazil by foreigners. I use net liability flows as the measure of capital flows. This variable reports the net amount of capital that flowed into each country through the sale of domestic assets to foreign investors each year, aligning with my thought experiment. Ideally, I would have access to gross inflow data, but a comprehensive dataset with this information does not exist. Thus, the readily available net inflow variable is used as the metric for capital inflows. To test the predictions of the model in the previous section, liability capital flows are divided into safe and risky assets. In line with the model, bond 192

Emerging Market Economies

inflows are safe assets and non-bond inflows are risky assets. To distinguish between bond flows and non-bond flows in the BOP Database, I follow Blanchard et al. (2017). The BOP database labels bond flows as net portfolio inflows through debt securities, and these flows generate the bond inflows variable in my dataset. Non-bond flows are defined as the sum of net FDI inflows, net equity inflows, and net other inflows. The BOP database labels equity flows as the net portfolio inflows through equity and investment fund shares, FDI flows as net direct investment inflows, and other flows as net other inflows. My research primarily focuses on the effect of bond inflows and non-bond inflows on domestic credit conditions; however, the impact of non-bond inflows divided into their separate categories is also studied. If countries did not disclose net inflow data, the BOP Database left these values blank. For example, Qatar did not report any flows from the year 2000 to 2010, and Venezuela did not report financial account data in the year 2017. These observations are dropped from the sample along with the 62 country-year pairs with missing bond-inflow data. The non-bond inflow variable is generated by adding net FDI inflows, net equity inflows, and net other inflows. To limit the number of dropped observations, it is assumed that missing FDI, equity, or other inflow values are equal to zero. Assuming that at least one inflow value is not missing, I sum FDI, equity, and other inflow values to generate a non-zero value for non-bond inflows. When the effect of each disaggregated non-bond inflow type on ∆λ is estimated, all observations with missing values are dropped. 3. Exchange Rate Regime

The International Monetary Fund’s Annual Report of Exchange Arrangements and Exchange Restrictions (AREAER) is used as a standardized resource to assign each country’s yearly exchange rate regime. Exchange rate regimes can vary from year to year, so the annual nature of this report allows fluctuations in each country’s regime choice to be uncovered. Furthermore, the IMF uses a classification that is based on each country’s actual, de facto arrangement, rather than focusing on the de jure arrangements claimed to be held by each country’s government. Levy-Yeyati and Sturzenegger (2003) argue that relying on announced de jure arrangements leads to bias because countries may claim to have a floating rate but still interfere in the currency market to increase currency stability. In the AREAER, IMF staff determines de facto arrangements based on developments in the foreign exchange market. Exchange rate regimes exist on a spectrum, so the AREAER provides eleven potential categories of classification: no separate legal tender, currency boards, conventional pegs, stabilized arrangements, crawling bands, crawling pegs, crawl-like arrangements, other managed arrangements, pegs within horizontal bands, floating rates, and free-floating rates. I code exchange rate Columbia Economic Review | 193

Caroline Okel

arrangements as dummy variables by hand. According to the AREAER, free-floating arrangements differ from floating arrangements because the IMF limits government intervention to at most three instances in the previous six months to be classified as a free-floating regime. Floating regimes can experience slightly more management. Because the AREAER did not provide the category of free-floating rates until 2009, floating and free-floating arrangements are grouped together as floating regimes. To generate a fixed versus floating dummy variable, all other regime types are grouped together as fixed regimes. The fixed regime dummy equals one for fixed exchange rate regimes and zero for floating regimes. The effect of exchange rate regimes on the relationship between capital inflows and credit is isolated empirically by interacting the fixed regime dummy with each inflow variable. These interaction terms allow me to study how the relationship between capital inflows and credit differs in fixed regimes relative to floating regimes. Empirical findings focus on the binary fixed versus flexible variable. It would be interesting to study the nuances the relationship between capital inflows and credit across the full exchange rate regime spectrum, but conducting this analysis with the eleven different regime dummies is a task for future research. 4. Lagged Domestic Credit

The previous year’s ratio of domestic credit to GDP is included to control for credit cycles. Previous levels of domestic credit correlate with current domestic credit conditions due to the cyclical nature of the credit cycle. Furthermore, past credit booms may lead to economic development in the private sector, which may encourage interest from foreign investors. Therefore, I hypothesize that omitting a lagged domestic credit variable would lead to an upward bias on the estimated coefﬁcients on the variables of interest. 5. International Trading Position

To account for each country’s trading position, each country’s year-overyear change in terms of trade and the growth in each country’s trading partners are included as control variables. The World Bank’s World Integrated Trade Solution (WITS) database is used as the information source for the change in terms of trade variable. This database reports the yearly net barter terms of trade index for each country with the base year of 2000. To study terms of trade change, each country’s year-over-year percent change in terms of trade is calculated. This control is necessary because surges in capital inflows can appreciate exchange rates, increasing the terms of trade. Furthermore, deterioration in the terms of trade may put financial pressure on the private sector, leading to an increased demand for domestic credit. Thus, failing to control for this variable may put downward pressure on the estimated coefficients on my variables of interest. 194

Emerging Market Economies

I also control for trading partner GDP growth. Annual information on the percentage of exports that each country sends to each of their trading partners is collected from WITS. This information is partnered with data on the real GDPs in 2010 USD of every country in the world. The World Bank provides this data as well. Using this real GDP data, a weighted average of trading partner GDP is calculated by export share. Finally, the year-over-year percentage change in these weighted averages is calculated to reach the desired control, trading partner growth. This variable impacts domestic credit conditions because as trading partners grow, they will demand more exports and lead to private sector growth within the country. This expansion in the private sector will decrease the demand for credit. Furthermore, this economic growth will also encourage foreign investment, leading to increases in capital inﬂows. Hence, I hypothesize that omitting this variable will lead to a downward bias on the estimated coefﬁcients on my variables of interest. 6. Monetary Policy Controls

Finally, I control for monetary policy responses to capital inflows. Information on interest rates comes from the IMF’s International Financial Statistics (IFS) database. The policy interest rate is used as the control variable, but the discount rate is used for countries where the policy rate is unavailable. As seen in the model, interest rate decisions made by the central bank serve as a key determinant in foreign investment decisions. Assuming that the policy rate influences both the rate of return of bonds and non-bonds, an increase in the policy rate will increase foreign interest in both assets. Changes in the policy rate will decrease the share of domestic credit because increases in the policy rate will increase the cost of borrowing for the private sector. Therefore, I believe that failing to account for the policy rate will lead to downward pressure on the estimated coefficients on my variables of interest. Central banks can also buy and sell reserve assets as a monetary policy tool. This variable is constructed by using information on annual reserve flows from the IMF BOP database. These reserve flows are recorded in the dataset as net purchases of external assets by the central bank. Thus, a positive value signals that the central bank purchased more foreign assets than they sold, leading to a net outflow of capital. These reserve outflows are scaled by nominal GDP, and this serves as the final control variable in my analysis. This control is necessary because in response to large inflows, central banks may manipulate reserves to prevent exchange rate fluctuations. Reserve flows may also be correlated with domestic credit conditions because central banks may transact with domestic banks when buying and selling reserve assets. If domestic banks serve as the counter-party in these large transactions, they may increase their supply of loanable funds in response to large reserve outflows. I hypothesize that omitting this variable will lead to a downward bias on the estimated coefficients on my variables of interest. Columbia Economic Review | 195

Caroline Okel

C. Descriptive Statistics This thought experiment requires a sample of countries of various exchange regime classifications, so a list of countries of various regime types is compiled. The Annual Report of Exchange Arrangements and Exchange Restrictions, the source for exchange rate regime data, provides eleven potential categories of regime classification. The distribution of the sample is presented in Figure 1. For simplicity, these eleven classifications are condensed into two categories: fixed regimes versus floating regimes. All floating or free-floating arrangements are considered to be floating, and all other nine categories are grouped together as fixed regimes. In Figure 1, this binary is accounted for because floating regimes are highlighted in green and fixed regimes are highlighted in blue. Of the 612 country-year pairs in my sample, the IMF classifies 395 as floating exchange rate regimes and 217 as fixed regimes. I use annual data to classify each country into an exchange rate regime, and there is some variation in regime classification within countries over time. In the sample, the IMF classified 11 countries as fixed regimes for all 18 years and 4 countries as floating regimes across the whole time period. Nineteen countries received both a fixed and floating classification in at least one year. Thus, country fixed effects do not capture this variation. Table 1 reports summary statistics for the full sample and for the categories of fixed regimes and floating regimes. This table contains sample averages, with standard deviations in parentheses. In the full sample, the average value for the dependent variable, Change in Domestic Credit/GDP, is 4.346%. Thus, on average, countries experience growth in domestic credit of 4.346% yearover-year. This change is slightly greater in fixed regimes. The variance in the year-over-year change in λ is significant. The most dramatic decline in the share of λ occurred in Poland when domestic credit declined by roughly 45% from 2001 to 2002. Poland also faced the most dramatic growth in λ, with domestic credit growing by roughly 96% from 2003 to 2004. However, all countries are not this extreme. For example, Bangladesh and India have faced a more stable growth rate in λ, with year-over-year change hovering around 4%. 196

Emerging Market Economies

Although change in λ serves as the dependent variable in my analysis, my dataset also contains information on each country’s yearly value of domestic credit. The average value of domestic credit is 55.080% of GDP for all the observations in my sample. Using the average nominal GDP of 534 billion USD, this correlates to an average annual value of domestic credit of 294 billion USD. Table 1 also reports the average capital flows of the sample. In the full sample, the average value of bond flows is 0.813% of GDP, and the average value of non-bond flows is 4.525% of GDP. The average value of non-bond flows is over four times larger than the average value of bond flows. The non-bond inflows category aggregates FDI inflows, equity inflows, and other inflows, and this variety of flow types partially drives this difference. Furthermore, as noted in Section II, FDI inflows are the largest source of capital inflows to EMEs. Thus, it is unsurprising that EMEs receive a higher share of non-bond flows than bond flows. In Table 1, the non-bond flows variable is also disaggregated into its three subcomponents: equity flows, FDI flows, and other flows. On average, fixed regimes experience a larger share of capital inflows than floating regimes, with the exception of equity inflows.

Columbia Economic Review | 197

Caroline Okel

D. Empirical Findings Using the preferred functional form, I find suggestive evidence to support my hypotheses of interest, hypothesis (1) and hypothesis (2). These findings suggest that non-bond flows to fixed regimes increase domestic credit more than non-bond flows to floating regimes and that bond flows to fixed regimes do not affect domestic credit conditions. Furthermore, empirical results support hypothesis (3a), a finding consistent with Blanchard el al.’s (2017) empirical work. This finding suggests that non-bond flows to floating regimes increase domestic credit, giving further evidence to support the existence of the financial intermediation channel. Insufficient evidence is found to support hypothesis (3b). Table 2 presents preliminary empirical findings. These findings do not account for the possible endogeneity of capital flows, but this concern is addressed in the following section. Column (1) is a naïve ordinary least squares regression of the four variables of interest on Change in Domestic Credit/GDP, ∆λ. In Column (2), time and country fixed effects are added. Column (3) includes fixed effects and controls for lagged domestic credit, trading partner growth, and the change in terms of trade. In Column (4), controls for the policy rate and reserve flows are added. Finally, in Column (5) the Non-Bond Flows/GDP variable is disaggregated into its three components: FDI inflows, equity inflows, and other inflows. In fixed exchange rate regimes, estimates in Column (4) predict that the impact of non-bond flows will be more pronounced in fixed regimes than in floating regimes. The positive and statistically-significant estimates of β 4 in this table are consistent with hypothesis (1). On average, ceteris paribus, the estimates of β3 and β4 suggest that an increase in non-bond flows of 1% of GDP results in a 0.497 percentage point increase in λ for fixed regimes. Estimates suggest that the effect of non-bond flows is stronger in fixed regimes than in floating regimes, as the estimate of β4 is statistically-significant at the 1% level. Furthermore, estimates of β4 and β3 are jointly statistically different from zero, as seen by the F-statistic of 13.05 with p-value of 0.00 in Table 3. The average year-over-year percentage change in λ for the fixed regimes in my sample was 6.11%. Thus, this finding suggests that if the share of non-bond flows increased by 1%, domestic credit would increase from being 6.11% of GDP to 6.61%. The average nominal GDP of fixed regimes in my sample was 602 billion USD. Thus, these findings suggest that on average, an increase in domestic non-bond flows of 6.02 billion USD would be associated with an increase in domestic credit of 3.01 billion USD. As predicted by the model, estimates in Column (4) suggest that for fixed regimes, the effect of bond inflows on domestic credit is statistically indistinguishable from zero. The estimates of β1 and β2 are not jointly significant, as seen by the F-statistic of 0.62 and the p-value of 0.54 in Table 3. 198

Emerging Market Economies

However, this provides limited evidence for the hypothesis that bond inflows do not impact λ in fixed regimes because the estimated effect of bond flows to fixed regimes is large in magnitude. The findings in Column (4) suggest that for fixed regimes, an increase in bond flows of 1% of GDP will lead to a 0.670 percentage point decrease in domestic credit on average, ceteris paribus. In the sample, the year-over-year percentage change in λ for fixed regimes was 6.11%. Thus, these findings suggest that if bond flows increased by 1%, domestic credit would decrease from being 6.11% of GDP to being 5.44% of GDP. Using the average fixed regime nominal GDP of 602 billion USD, the findings in Table 2 suggest that on average, an increase in domestic bond flows of 6.02 billion USD will lead to a decrease in domestic credit of 4.03 billion USD. Statistical significance aside, this finding does not suggest that non-bond inflows have no impact on domestic credit in fixed regimes.

Columbia Economic Review | 199

Caroline Okel

The positive and statistically-significant estimates of β3 show that non-bond flows to floating regimes are expected to increase domestic credit. This finding is consistent with hypothesis (3a). As seen in Column (4), estimates suggest that in a floating exchange rate regime, an increase in non-bond flows of 1% of GDP will increase domestic credit by 0.204 percentage points on average, ceteris paribus. This finding is both statistically-significant and large in magnitude. In the sample of floating exchange rate regimes, the average year-over-year percentage change in domestic credit was 3.40%. Thus, the estimate of β3 suggests that if the share of non-bond flows increased by 1%, domestic credit would increase from being 3.40% of nominal GDP to 3.60%. The average nominal GDP of floating regimes in the sample was 511 billion USD. This indicates that on average, an increase in domestic non-bond flows of 5.11 billion USD would be associated with an increase in domestic credit of 1.40 billion USD. Estimates in Column (4) indicate that in a floating exchange rate regime, an increase in bond flows of 1% of GDP will increase λ by 0.038 percentage points on average, ceteris paribus. This finding contradicts hypothesis (3b) that bond flows to floating regimes decrease domestic credit and is not consistent with the predictions of Blanchard et al. (2017). However, this estimate of β 1 is statistically indistinguishable from zero and is not large in magnitude, and Blanchard et al.’s (2017) empirical work faces a similar finding. The average year-over-year percentage change in λ for the floating regimes in the sample was 3.40%. Thus, these empirical findings suggest that if the share of bond-flows increases by 1%, domestic credit would increase from being 3.40% of nominal GDP to 3.44%. Following Blanchard et al. (2017), five control variables are included in this regression to account for past credit conditions, the country’s trading position, and the country’s monetary policy. Empirically, an increase in the monetary policy rate, the terms of trade, or last year’s domestic credit are all expected to decrease domestic credit. The negative sign on the coefficients on the policy rate and change in terms of trade variables were expected; however, I did not expect this negative sign on the lagged domestic credit coefficient. All of these coefficients are significant at the 1% level and large in magnitude. The coefficient on trading partner growth is negative, as expected, but it is small in magnitude, although statistically significant at the 1% level. Finally, the impact of reserve flows is neither statistically-significant nor large in magnitude.

200

Emerging Market Economies

The coefficients on the variables of interest do not notably vary across Columns (1) through (5). Column (1) presents the most naïve regression, an ordinary least squares regression without any fixed effects or control variables. Including country and time fixed effects in Column (2) does not remarkably change the estimates from Column (1), as the signs on coefficients remain the same and do not change notably in magnitude. However, it is worth noting that β4 was significant at the 10% level in Column (1) but becomes significant at the 5% level in Column (2), as the standard error decreased from 0.159 to 0.120, a substantial increase in precision. Country and time fixed effects are included to control for unobservables. Country fixed effects control for unobservable factors within countries that do not vary over time. For example, this term may capture factors such as financial sector sophistication. Time fixed effects control for year-specific factors that impact all countries. Time fixed effects may capture unobservables such as worldwide financial conditions and investor sentiment towards EMEs. In Column (3), controls for past domestic credit conditions and international trade position are added. In sections IV.B.4 and IV.B.5, I hypothesize that omitting a lagged domestic credit variable may bias the coefficients of interest upward. I also argue that omitting a change in terms of trade variable and a trading partner growth variable possibly leads to a downward bias on the coefficients of interest. Including the controls decreases the magnitude of all four coefficients of interest. Column (4) adds monetary policy controls. In section IV.B.6, I hypothesize that omitting a control for interest rates will lead to a downward bias on the variables of interest. I also speculate that omitting a reserve asset control variable will lead to an upward bias on all coefficients of interest. In Column (4), both the bond flow and non-bond flow variables decrease in magnitude. Including fixed effects and the control variables decreases the number of observations from 512 in Column (1) to 347 in Column (4). Furthermore, Columbia Economic Review | 201

Caroline Okel

Greece, Argentina, Ukraine, and the Czech Republic are not included in Column (4) due to missing interest rate data. As a robustness check, in Column (5), the Non-Bond Flows/GDP variable is disaggregated into its three components: FDI inflows, equity inflows, and other inflows. These estimated coefficients indicate that “other” flows are the main driving force behind the positive impact of non-bond inflows on λ in floating regimes. The estimated coefficients in Column (5) also predict that the impact of “other” flows is smaller in magnitude in a fixed regime than in a floating regime. However, according to these findings, FDI inflows have a positive effect on domestic credit conditions in fixed regimes. Table 3 demonstrates that the Other Inflows/GDP variable and the interaction between this variable and the fixed regime dummy are jointly significant. The FDI inflow variable and the FDI interaction term are also jointly significant. Thus, while “other” flows appear to be the only driver in the impact of non-bond flows on λ in floating regimes, both FDI and “other” flows appear to influence domestic credit conditions in fixed regimes.

V. Discussion and Conclusions Should policymakers in emerging market economies be concerned about the relationship between capital inflows and credit booms? It depends: If capital inflows cause credit booms (and financial crises, in turn), then capital or credit controls seem like attractive policy options. Specifically, macro-prudential tools would limit credit expansions, while capital controls could prevent inflow bonanzas in the first place. Given the costs of implementing such policies, a clear theoretical and empirical link between inflows and credit bonanzas would need to be established. The standard Mundell-Fleming framework predicts that capital inflows actually decrease credit booms. Thus, in this framework, policymakers should not be concerned about capital inflow surges; however, many economists and policymakers believe that capital inflows do cause credit booms. Blanchard et al. (2017) provide theoretical support for such beliefs by extending the standard Mundell-Fleming framework with flexible exchange rates to incorporate risky financial assets. Their model predicts that certain types of capital inflows cause credit booms in EMEs. Given that credit booms are regularly followed by busts, the authors conclude that policymakers in EMEs should consider strengthening capital controls to manage non-bond inflow bonanzas, thus popping credit bubbles. Yet, the framework of Blanchard et al. (2017) assumes flexibility of exchange rates, an implausible assumption for many EMEs. One might thus 202

Emerging Market Economies

wonder whether their prescription of stronger capital controls extends to cases where exchange rates are not perfectly flexible. In this paper, I extend their model and show that capital inflows that involve purchases of risky assets increase domestic credit even in fixed exchange rate regimes. In fact, I find that non-bond inflows increase domestic credit more in fixed regimes than in floating regimes. Blanchard et al.’s (2017) theory predicts that bond flows to floating regimes will decrease domestic credit, potentially netting out the positive effect that non-bond inflows have on credit. My model suggests that bond flows to fixed regimes will not influence domestic credit. Hence, these predictions suggest that a capital inflow bonanza comprised of both flow types will lead to even larger credit booms when exchange rates are fixed. My results imply even greater concern about capital inflows to EMEs when exchange rates are fixed - beyond those identified by Blanchard et. al (2017). These results rationalize the purely empirical findings of Magud et al. (2015) and Magud and Vesperoni (2016), who find that fixed regimes are more susceptible to credit booms in the face of capital inflow bonanzas. I also empirically examine my extended model’s predictions and find suggestive evidence that the relationship between capital inflows and domestic credit differs between fixed regimes and floating regimes. Empirical findings suggest that non-bond inflows may increase domestic credit more in fixed regimes than in floating regimes. I find that the effect of bond inflows on domestic credit in fixed regimes is statistically indistinguishable from zero. Although these findings support of the hypotheses of my model, future research is necessary to address concerns such as possible endogeneity of capital flows and to make causal claims. Despite this need for future research, both my theoretical and empirical work potentially suggest that fixed regimes are more vulnerable to credit booms than floating regimes. Even though these findings cannot be interpreted as causal, overall, my research suggests that capital flows to emerging markets may lead to credit booms and that policymakers in fixed regimes may have reason to be especially concerned. In this light, EME policymakers should consider implementing tools that mitigate excessive credit growth in the face of capital inflow booms.

Columbia Economic Review | 203

Caroline Okel

Notes Page Negative credit growth is undesirable because while average output growth is roughly 6.3% in recoveries with credit, average output growth in recoveries with negative credit growth is 4.5% (Abiad et al 2011). 1

Magud and Vesperoni attribute the difference in domestic credit swings between fixed and flexible regimes to differences in the response of consumption to capital flow cycles. For flexible regimes, a capital inflow can trigger a nominal exchange rate appreciation, and this result barely impacts the relative prices of tradable and non-tradable goods. However, in fixed regimes, the nominal exchange rate cannot appreciate, so demand for non-tradable goods increases non-tradable inflation. This non-tradable inflation increases the supply of non-tradable goods above its full employment level. Thus, total consumption during an inflow in fixed regimes grows faster than in flexible regimes, leading to a harsher adjustment process for fixed regimes during an inflow reversal. 2

204

Emerging Market Economies

VI. References Abiad, Ad., Li, B., & Dell’Ariccia, G. (2011). Creditless recoveries. International Monetary Fund, Washington. IMF Working Paper No. 11/58. Addvjiev, S., Hardy, B., Kalemli-Özcan, Ş., & Servén, L. (2017). Gross capital inflows to banks, corporates and sovereigns. NBER Working Paper No. 23116. Agosin, M.R., & Huaita, F. (2011). Capital flows to emerging economies: Minsky in the tropics. Cambridge Journal of Economics, 35, pp.663-683. Ahmed, S., Coulibaly, B., & Zlate, A. (2017). International financial spillovers to emerging market economies: How important are economic fundamentals? Federal Reserve Bank of Boston, Boston. Risk & Policy Analysis Unit Working Paper. RPA 17-02. Amri, P.D., Richey, G.M., & Willet, T.D. (2016). Capital surges and credit booms: How tight is the relationship? Open Economy Review, 27, pp. 637-670. Blanchard, O. (2017). Macroeconomics (7th ed.). Boston, MA: Pearson. Blanchard, O., Ostry, JD., Ghosh, A.R., & Chamon, M. (2017). Are capital inflows expansionary or contractionary? Theory, policy implications, and some evidence. IMF Economic Review, 65(3), pp. 563-585. Blanchard, O., Ostry, JD., Ghosh, A.R., & Chamon, M. (2015). Expansionary or contractionary effects of capital inflows: It depends what kind. VOX CEPR Policy Portal. Retrieved from https://voxeu.org/article/macro -effects-capital-inflows-capital-type- matters. Boudias, R. (2015). Capital inflows, exchange rate regimes and credit dynamics in emerging market economies. International Economics, 143, pp.80-97. Calderón, C., & Kubota, M. (2012). Gross inflows gone wild: Gross capital inflows, credit booms and crises. The World Bank Latin America and the Caribbean Region. Policy Research Working Paper No. 6270. Carlin, W., & Soskice, D. (2006). Macroeconomics: Imperfections, institutions & policies. Oxford: Oxford University Press. Calvo, G.A., Izquierdo, A., & Mejía, L.F. (2004). On the empirics of sudden stops: The relevance of balance-sheet effects. Inter-American Development Bank, Washington. Inter-American Development Bank Working Paper No 509. Cecchetti, S., & Schoenholtz, K. (2018). Sudden stops: A primer on balance-of-payments crises. VOX CEPR Policy Portal. Retrieved from https://voxeu.org/content/sudden- stops-primer-balance-payments-crises. Elekdag, S., & Wu, Y. (2013). Rapid credit growth in emerging markets: Boon or boom-bust? Emerging Markets Finance & Trade, 49(5), pp. 45-62. Furceri, D., Guichard, S., & Rusticelli, E. (2011). Episodes of large capital inflows and the likelihood of banking and currency crises and sudden

Columbia Economic Review | 205

Caroline Okel

stops. OECD Economics Department Working Papers, No. 865, OECD Publishing, Paris. Ghosh, A.R., Ostry, J.D., & Qureshi, M.S. (2016). When do capital inflow surges end in tears? American Economic Review: Papers and Proceedings, 106(5), pp. 581-585. International Monetary Fund (2018). Annual report on exchange arrangements and exchange restrictions. International Monetary Fund, Washington. Levy-Yeyati, E., $ Sturzenegger, F. (2003). To float or to fix: Evidence on the impact of exchange rate regimes on growth. AmericanEconomic Re view, 93(4), pp. 1173-1193. Magud, N.E., Reinhart, C.M., & Rogoff, K.S. (2018). Capital controls: Myth and reality. Annals of Economics and Finance, 19-1, pp.1-47. Magud, N.E., Reinhart, C.M., & Vesperoni, E.R. (2014). Capital inflows, exchange rate flexibility and credit booms. Review of Developmental Economics, 18(3), pp.415-430. Magud, N.E., & Vesperoni, E.R. (2015). Exchange rate flexibility and credit during capital inflow reversals: Purgatory ... not paradise. Journal of International Money and Finance, 55, pp.88-110. Mendoza, E.G, & Terrones, M.E. (2008). An anatomy of credit booms: Evidence from macro aggregates and micro data. NBER Working Paper No. 14049. Mendoza, E.G, & Terrones, M.E. (2012). An anatomy of credit booms and their demise. NBER Working Paper No. 18379. Montiel, P., & Reinhart, C. (2001). The dynamics of capital movements to emerging economies during the 1990s. In Griffith-Jones, S., Montes, M., and Nasution, A. (Eds.), Short-term Capital Flows and Economic Crises (pp. 3-28). Oxford: Oxford University Press. Obstfeld, M., Ostry, J.D., & Quereshi, M.S. (2017). A tie that binds: Revisiting the trilemma in emerging market economies. Review of Economics and Statistics. Reinhart, C.M., & Montiel, P. (2001). The dynamics of capital movements to emerging economies during the 1990s. In Griffith-Jones, M. (Eds.), Short-term Capital Movements and Balance of Payments Crises. Oxford: Oxford University Press. Reinhart, C.M., & Reinhart, V.R. (2008). Capital flow bonanzas: An encompassing view of the past and present. In J. Frankel & C. Pissarides (Eds.), NBER International Seminar in Macroeconomics (pp. 1-54). Chicago: Chicago University Press. Sahay, R., Arora, V., Arvanitis, T., Faruqee, H., N’Diaye, P., & ManciniGriffoli, T. (2014). Emerging market volatility: Lessons from the taper tantrum. IMF Staff Discussion Note. 206

Emerging Market Economies

Sivabalan, S. (2018). Not a happy anniversary: EMâ&#x20AC;&#x2122;s taper tantrum began 5 years ago. Bloomberg News. Retrieved from https://www.bloomberg.com/ news/articles/2018-05- 22/not-a-happy-anniversary-em-s-tapertantrum-began-5-years-ago. The World Bank Data. Domestic credit to private sector (% of GDP). Retrieved from https://data.worldbank.org/indicator/FS.AST.PRVT. GD.ZS?end=2017&locations=XP &start=2008.

Columbia Economic Review | 207

Founded in 2009 as the ďŹ rst undergraduate economic journal in the United States, the Columbia Economic Review (CER) aims to promote discourse and research at the intersection of economics, business, politics, and society by publishing a rigorous selection of research papers in our print journal. We further strive to engage individuals on campus, locally, and globally through speaker series, symposia, competitions, and other events established to promote dialogue and encourage deeper insights on economic issues. CER is sponsored by the Program for Economic Research (PER) at Columbia University, and is entirely led, organized, and operated by undergraduate students at Columbia University across a range of academic disciplines.

econreview@columbia.edu columbiaeconreview.com

COLUMBIA ECONOMIC REVIEW | FALL 2019 ISSUE