Columbia Economics Review: Fall 2017 by Columbia Economics Review

Columbia Economics Review

Crossing Borders, Pouring Mortar Primetime Crime Barreling Down the Highway Penny for Your Shots The Default in Our Stars A City without Uber?

Vol. IX No. I Fall 2017

Fall 2017

COLUMBIA ECONOMICS REVIEW PUBLICATION INFORMATION Columbia Economics Review (CER) aims to promote discourse and research at the intersection of economics, business, politics, and society by publishing a rigorous selection of student essays, opinions, and research papers. CER also holds the Columbia Economics Forum, a speaker series established to promote dialogue and encourage deeper insights into economic issues.

2017-2018 E D I T O R I A L B O A R D EDITOR-IN-CHIEF

Alan Lin

JOURNAL STAFF EDITORS

MANAGING EDITORS

PUBLISHER

SENIOR EDITORS

DESIGN DIRECTOR

LAYOUT EDITOR

Ben Titlebaum

Michael Crapotta Alex Whitman

Jessica Lu

Jessica Bai Michael Chu Neel Puri Michelle Yan Elise Gout

Bennie Chen Jack Hannah Brian Ho Elizabeth Lee Manuel Perez Archila Ojaswee Rajbhandary Rajiv Ram William Sammartino Kirk Wu John Zhou

CONTRIBUTING ARTISTS

Sirena Khanna Jennifer Fan Nicholas DiCostanzo Lipa Schmelczer Stacy Tao Yingzi Huang

ONLINE EXECUTIVE EDITORS

Guillermo Carranza Jordan Mathieu Sabbagh

WEBMASTER

Frank Zhu

Hallie Gruder Phillip Jang Elise Parrish Matt Simons

CONTRIBUTORS

Guillermo Castro Dominguez Roxanne Farhad Alec Howe

Jerry Shi Mitchell Mikinski Jack Zhang

OPERATIONS EXECUTIVE DIRECTOR

Makenzie Nohr TREASURER

Michelle Yan

SENIOR ADVISORS

Zoey Chopra Bryan Li Randy Zhong

CEC MEMBERS

Alex Shek Maria Thompson Zachary Rotman

EPC MEMBERS

Yael Cohen Blaine Helloloid Samantha Syme

HEC MEMBERS

Jenna Karp Derek Zhang

A special thanks to the Columbia University Economics Department for their help in the publication of this issue.

Columbia Economics Review would like to thank its alumni and donors for their generous support of the publication.

We welcome your comments. To send a letter to the editor, please email: econreview@columbia. edu We reserve the right to edit and condense all letters.

Licensed under Creative Commons Attribution-NonCommercial-NoDerivs 3.0 Unported License

Columbia Economics | Program for Economic Research

Printed with generous support from the Columbia University Program for Economic Research Columbia Economics Review

Fall 2017

TABLE OF CONTENTS

Immigration 6

Crossing Borders, Pouring Mortar

Housing Demand and Labor Supply: The 1962 Algerian Repatriates to France

Sports and City Welfare 16

Primetime Crime

Winning and Losing: The Socioeconomic Impact of Sports on Crime

Transit Demand 22

Barreling Down the Highway

Understanding Public Transit Ridership through Gasoline Demand: A Case Study in the San Francisco Bay Area, CA

Microeconomics of the NBA 36

Penny for Your Shots

The Underpaid Superstar: The Max Contractâ&#x20AC;&#x2122;s Effect on Parity within the 2015-2016 NBA Microeconomy

Debt Policy in Latin America 42

The Default in Our Stars

The Evolution of Collective Action Clauses and Their Impact on Recent Latin American Debt Crises

The Sharing Economy 46

A City without Uber?

An Online Piece Discussing the Economic Implications of Londonâ&#x20AC;&#x2122;s Proposed Uber Ban

For a complete list of papers cited by our authors and a full version of all editorials, please visit our website at columbiaeconreview.com

Opinions expressed herein do not necessarily reflect the views of Columbia University or Columbia Economics Review, its staff, sponsors, or affiliates.

Columbia Economics Review

Fall 2017

COLUMBIA ECONOMICS REVIEW

Call for Submissions Columbia Economics Review is interested in your article proposals, senior theses, seminar papers, editorials, art and photography.

GUIDELINES CER is currently accepting pitches for its upcoming issue. You are encouraged to submit your article proposals, academic scholarship, senior seminar papers, editorials, art and photography broadly relating to the field of economics. You may submit multiple pitches. Your pitch or complete article should include the following information: 1. Name, school, year, and contact information (email and phone number) 2. If you are submitting a pitch, please state your argument clearly and expand upon it in one brief paragraph. List sources and professors/industry professionals you intend to contact to support your argument. Note that a full source list is not required. 3. If you are submitting a completed paper, please make sure the file is accessible through MS Word. Pitches will be accepted throughout the fall and are due on a rolling basis. Send all pitches to econreview@columbia.edu with the subject â&#x20AC;&#x153;CER Pitch- Last Name, First Name.â&#x20AC;? If you have any questions regarding this process, please do not hesitate to e-mail us at econreview@columbia.edu. We look forward to reading your submissions!

Columbia Economics Review

Fall 2017

A LETTER FROM THE EDITORS Dear Readers, As the year draws to a close, we are excited to present to you our Fall 2017 issue. Around the world, 2017 has been a year of both economic and social upheaval. On the one hand, nativist reactions to issues like immigration have superseded common sense economic truths in the policymaking realm, placing self-preserving protectionism above forward-looking growth. On the other hand, sharp criticism of private and public institutions has shed light on the many burdens borne by society, which had been largely ignored or inadequately addressed by both policymakers and academic circles in the past. As we attempt to make sense of an increasingly reactionary social agenda, we must reevaluate the roles that institutions play in our daily lives. By utilizing a rich arsenal of economic thought, we are able to guide our solutions toward an understanding of society that, rather than merely reacting to social ills, addresses the plethora of externalities that we are now facing. As economists, we spend the majority of our time modeling how decisions shape the world around us. We aim to present world-class research that bridges theory with empirically observable truths. However, in an age that increasingly monetizes skepticism, economists must bridge the gap between academia, policymakers, and the public, rather than simply obtain authority only by asserting data-derived facts. We must be able to provide answers to those who miss out on the benefits of the economic globalization. Rather than articulating a long-term state that is beneficial for all, economists must answer to the pains that confront society in the current environment. The Columbia Economics Review is fortunate to have received an abundance of remarkable submissions from across the nation and the world. We are therefore pleased to feature five papers and one online editorial that address social themes with economic rigor, ingenuity, and foresight. The pieces in this edition address relevant social issues from a range of different vantage points. They distinguish between questions that affect a majority of the population and questions that affect a select cohort of society. Despite differing approaches and subject matters, all of our selections represent a conflict in our understanding of the social good. By including these papers in our journal, we hope to demonstrate that the public good can be clearly analyzed through economics, and ultimately provide us with a means of maximizing social welfare in a global economy. Of all the contemporaneous topics of 2017, we found the economics of immigration to be particularly relevant to our experience. Analyzing the effect of Algerian repatriates on France’s housing and labor markets, Olivia Briffault (6) utilizes an excellent case study that helps debunk the myth of widespread wage depression thanks to a mass influx of immigrants entering the labor market. We also enjoy the nuanced approach that Columbia alumna Isabella Santandreu (42) takes in analyzing debt crises in Latin America through the lens of collective action clauses; Santandreu finds that while sovereign debt restructurings may take a lasting toll on economic recovery, collective action clauses can prevent serious defaults if proper reforms are considered. Ultimately, because the debt burden often befalls the most vulnerable members of society, reform within the financial sector is crucial to mitigating crippling inequality. We have also selected papers that employ economic theory to address social issues from nontraditional angles. While crime is often seen as a function of structural or situational poverty, Columbia student Brendan Moore (16) takes the unique approach of analyzing the relationship that sports wins may have on decreasing incidences of criminal offense. We are excited by the possibilities that isolation and further investigation of unexpected externalities can offer the field of economics, especially as new social quandaries emerge with each passing day. Dylan Robbins-Kelly (36) analyzes the microeconomy of the NBA to highlight the economic inefficiencies generated by salary caps. Ultimately, the disjuncture between market price and talent level results in inequality among NBA teams, harming the sport as an institution, while demonstrating larger truths about market failures. Finally, we found Hansen Sun’s analysis of public transit ridership (22) to be an important consideration of how the prices of gasoline affect our everyday decisions of how to get from point A to point B. With a rigorous case study of San Francisco’s public transportation system, Sun’s analysis sheds light on the importance of correcting the frequent failures of public goods markets. As we proceed toward 2018, we at The Columbia Economics Review remain convinced that novel and insightful research has the ability to generate solutions to economic troubles that can, in turn, benefit humanity. We believe that the selection of papers in our Fall 2017 issue demonstrates the ability of academic research to delve into the personal and the public realms in a way that teaches us more about each other, while also showing us how to transmit our knowledge of specific economic contexts. We hope that our commitment to academic inquiry will continue to support a platform connecting students and economists through the realm of economic research. Cheers, Michael Crapotta CC’19 | Managing Editor Alex Whitman CC’19 | Managing Editor Alan Lin CC’18 | Editor-in-Chief Benjamin Titlebaum CC’19 | Publisher

Columbia Economics Review

Fall 2017

Crossing Borders, Pouring Mortar Housing Demand and Labor Supply: The 1962 Algerian Repatriates to France Olivia Briffault Yale University This article contributes to solving the puzzle that is the economics of immigration. Briffault examines how the arrival of Algerian repatriates during the 1960s affected France’s housing and labour markets. She expands on the current literature by using new data and investigating the demand and the supply side of the market. Her findings reveal that repatriates had a minimal impact on unemployment and wages — even in regions that welcomed the highest number. The positive labour supply effects were offset by an increase in housing demand. As immigration has recently become a hot topic in public policy discussions, Briffault offers a rigorous and informative treatment of this natural experiment. -W.L.S. BACKGROUND Algerian Repatriation in the 1960s French settlement of Algeria began during the 1830s when France conquered the city of Algiers. Soon after, Christian and Jewish settlers, known as the Pieds-Noirs, came from all over the Mediterranean region to claim land (Naylor, 2000). By 1962, when Algeria gained independence from France, the Pieds-Noirs made up as much as ten percent of the population, held over forty percent of the land, encompassing a wide range of socioeconomic strata (Grenville, 2005). By 1962, when it became evident that Algeria would become independent, the Pieds-Noirs were forced out of Algeria and began a mass exodus to mainland France. Close to one million Pieds-Noirs ended up settling in France by 1970 (Hubbell, 2015). The repatriates settled predominantly in Le Midi area of France where the generally warm and sunny climate was similar to that of Algeria (See Figure 1). Before their arrival, Le Midi had not been a region of strong economic growth and had much higher unemployment rates than other areas of France (Baillet, 1975). Immigration and Labor Market Outcomes There is a vast literature on the relation-

ship between immigration and labor market outcomes. Most studies find that immigration tends to have a minimal effect on the local labor market (Borjas, 2015). The influx of the Pieds-Noirs into France provides a natural experiment for understanding the impact on local economies of large waves of immigration. These immigrants settled together in specific regions of France determined by weather rather than economic conditions. There is little selection bias amongst these immigrants, as almost all Algerians of French origin returned to France by the end of the war, regardless of economic status. In the analysis that follows, Hunt’s analysis is repeated using the INSEE Employment Survey - a more frequent data set than the Census data used by Hunt. In addition to Hunt’s measure, two different measures of Pieds-Noirs are also compared. Most important, Hunt’s analysis is expanded by testing the hypothesis that an increase in aggregate labor supply is offset by an increase in aggregate demand for consumption goods by analyzing housing conditions and housing market outcomes in France following the arrival of the repatriates.

Columbia Economics Review

DATA The principal dataset used is the Employment Survey administered by INSEE (Institut Nationale de la Statistique et des Études Économiques) as part of their yearly data collection surveys in France. Unlike most employment surveys, this survey includes detailed information on the characteristics of housing for each individual, including when the house was built, the number of rooms, and who lives in the house, as well as detailed labor market information about the sector in which the individual is employed. The data are individual responses to the survey, the unit of analysis is the home, and individuals over age 15 years old are surveyed. The survey asks about employment and hours worked, but it does not include wage data. There are four steps to this analysis: first, determining the appropriate way to measure the percent of the population that are Algerian repatriates; second, analyzing the impact of an influx of repatriates on local labor markets; and third, determining whether there was construction undertaken to house repatriates. MEASURING ALGERIAN REPATRIATES France is broken into 21 administrative

Fall 2017 regions, and then further into 96 departments – which are below the region level but above the commune level. Regions are comprised of between six and thirteen departments depending on population size. Regions mask some of the clumping in the settlement of repatriates – mean repatriate by region varies from 0 to about .02 in some regions, while the percent repatriate in the department varies from 0 to about .04 in some departments. To further analyze the settlement pattern of repatriates, percent repatriate is compared across department. However, the department variable is only present in the 1968, 1969 and 1970 Employment Survey years, and so there is less variation across years. Results There are 488,534 observations over

the 5 years (1964, 1965, 1968, 1969, 1970) included in the sample for ages 22 to 65 (these are prime working years that are consistently available in all the data sets). There are approximately 40,000 observations from each of the 1964 and 1965 surveys, and 90,000 observations in each of the 1969 and 1970 surveys. There are closer to 250,000 in the 1968 survey as 1968 was a census year and so more households were surveyed. Two different measures of repatriate are utilized to estimate the number of repatriates in France over the time period. The “Algerian” variable varies from about 1% in 1964 to closer to 1.5% in 1970. The second measure of repatriate used is “Returned”. It is defined as having a nationality of Algerian or having lived outside of the French Metropole one year earlier

Columbia Economics Review

7 and having a nationality of French. Those “Returned” make up about 1.75% of the population in 1965 and that falls to 1.47% in 1970. One reason it may have fallen is that over time as more repatriates move into France, those who identified their region of residence one year prior as outside of the French Metropole also decreases. Repatriates settled disproportionately in the Midi region of France ProvenceCôte D’Azur, Languedoc, Rhône-Alpes and Midi-Pyrénées have the highest mean repatriates per region throughout the time period using both measures of repatriate. The paper compares the percent repatriate by region and year for each of my measures of repatriate to the percent repatriate by region and year that Hunt (1992) reports. Results show that

8 the same regions have high repatriated portions of the population. MEASURING THE EFFECTS OF IMMIGRATION ON THE LABOR MARKET Methods The local labor market effect is estimated by using both unemployment rates and hours worked per week. “Unemployed” is defined as an individual who is in the labor force and is actively seeking work. Unemployment is a binary variable, so the model is estimated using a logit regression and marginal effects at the means are reported. Unemployed is regressed on a dummy for repatriate, the percentage repatriate in the region, and an interaction between the dummy for repatriate and the percentage repatriate in the region to capture the effect of being a repatriate in a region with higher portion of the population repatriated. The labor force is defined as those aged 22 to 65 who are employed or actively seeking work. The unemployment analyses are repeated for each measure of “repatriate”: “returned” and “Algerian”, as well as using Hunt’s measure of the mean repatriates by region. The analysis is repeated separately for men and women. To model the effects of differing ages rather than assuming age is linear, controls for age and age squared are included. The model includes region fixed effects to control for any possible regional variations in unemployment rates and year fixed effects to capture business cycle variation in unemployment rates by year. Another important factor that contributes to unemployment is education, but the education variable is only available for the years 1969-1970. The unemployment regressions are only repeated on the years the education variable is available with and without the education variable as a robustness check to compare how much of the variation in unemployment is explained by education, and not just by restricting the years of the sample. The unemployment analyses are repeated separately with education controls using the regressors defined by department (mean repatriates by department) and with fixed effects for department, to see how the clumping of repatriates by department is reflected in local unemployment rates. The equation is estimated below (on the region level): Unemploymentijt = β1% repatriatejt + β2 repatriatei

Fall 2017 + β3 (repatriatei*%repatriatejt) + β4Agei...+β5Age2i...ΣγjJ +ΣδtT + ϵ Where i = individual; j = region; T=time. The β1 term describes the effect of the local share of the population at time t who are repatriates on the probability that an individual in that locality is unemployed. The β2 term describes the effect of the individual’s own repatriate status on the probability that he or she is unemployed. The β3 term is an interaction between the repatriate dummy and percent repatriate by region to estimate the impact of being a repatriate in a repatriate-dense region. Another way to measure the impact of the influx of repatriates on the local labor market is with number of hours worked per week. This analysis is only conducted on the portion of the population that is employed (or has more than 0 hours worked per week). Only observations with fewer than 70 hours worked are included in the analysis due to the lack of observations with fewer hours worked. The hours worked variable is only pre-

sent in the 1968, 1969, and 1970 data sets. The analysis for men and women is repeated separately. Using OLS estimates, hours worked on mean repatriate is regressed by region, a repatriate person dummy, and interaction between the repatriate dummy and the mean repatriate per region. Like the unemployment analyses, controls for age, age squared, region fixed effects and year fixed effects are included. I repeat the hours worked regressions restricting the sample to the years the education variable is available (1969 and 1970) with and without the education variable to compare how much of the variation in unemployment is explained by education, and not just by restricting the years of the sample. I repeat the analysis using mean repatriate by department with department level fixed effects and education. I do not include observations for women. Results Unemployment rates average about 1.3% over the time period (see Table 1).

Table 1: Mean and Standard Deviations of Key Variables by Year.

Standard deviations in parenthesis. These data are taken from the 1964-1970 employment survey. % Algerian denotes the percent of people that identify their nationality as Algerian in the nationality question present in every employment survey (1964, 1965, 1968, 1969, 1970.) In 1968 they changed the coding of Algerian from “Algerian Musulman” (Muslim Algerian) to just “Algerian.” % Returned denotes the percent of people who either identify their nationality as “Algerian” or identified their nationality as French and their region of residence one year earlier as outside of mainland France. The “returned” variable is only present in the 1965, 1969, and 1970 surveys. Unemployment is defined as the percent of the labor force without work and looking for work. The labor force does not include those who are students, who perform housework, who are part of the military or who are retired. Hours worked is the number of hours worked per week for those who are employed (<70). Crowding is defined as the number of people per number of rooms. New Home is defined as whether a person lives in a home built since 1962. Construction is defined as the percent of employed people working in the building or public works industries. Some high school education or more is whether a person has begun high school. Age varies from 22-65. *I do not include Paris in analyses with “Returned.”

Columbia Economics Review

Fall 2017 Figures vary from 0.64% in 1964 to about 1.5% unemployed in 1969 and 1.3% in 1970. These unemployment rates appear remarkably low, even if France was going through a period of economic growth. I compare the unemployment rates calculated from the survey data to French unemployment rates collected from the St. Louis Federal Reserve (FRED) over the same time period to check the accuracy of these rates. FRED rates are similar to the survey rates calculated from the employment survey, they also vary from about 1-2% from 1964 to 1970. Hours worked (found only in the 1968, 1969, and 1970 datasets) for employed men are about 47 per week from 1968-1970 and about 39 for women from 1968-1970. There is little variation across the years. Education rates are low throughout the period. Only about 17% of the population has a high school education or more. Table 3 reports the marginal effects at the mean of a logit regression of whether the respondent is employed (separately for men and women) on the three measures of repatriate (“Algerian”, “Returned”, and Hunt’s means by region) with means of each variable reported. Unemployment is quadratic and decreases in age until age 40 for men and age 43 for women, at which point it begins to rise again. I do not see substantial effects of being in a repatriate-dense region on unemployment rates. Using the “Algerian” measure of repatriate, areas with more Algerians do not have consistently higher unemployment rates. There is a penalty for being Algerian: male “Algerians” have 3.4% higher unemployment than non-Algerian males. However being “Algerian” in regions with a higher proportion “Algerian” erases that penalty. Using the “returned” measure of repatriate, areas with more recently “returned” men have about 0.3% higher unemployment rates (p<0.05). Those recently “returned” have about 3.8% higher likelihood of being unemployed. Again, being a recent “returnee” in a region with a higher proportion “returned” erases about 95% of the penalty of being a “returnee” on its own. I do not include Paris in analyses using the “returned” variable because of the discrepancy with 1969 data explained above. When I repeat the analysis including Paris, there is no effect of being in an area with a high proportion of the population “returned” on unemployment, the penalty for being recently “returned” remains about the same (3.6% higher unemployment rates) and the interaction term also remains about the same. The effect of being an “Algerian”

Table 3: The impact of the influx of repatriate on individual unemployment

Unemployed is a dummy variable for whether the person is employed and is fitted with a logit model. Marginal fixed effects are reported at the means. There are three measures of repatriate: “Algerian” is defined as those who indicate their nationality is Algerian; “Returned” is defined as those who indicate their nationality is French and their previous region of residence as outside of the French Metropole or who indicate their nationality is Algerian) and is only available in the 1965, 1969 and 1970 data sets; Hunt measure of Algerian is the percent Algerian by region taken from Hunt (1992) and only available for the years 1962 and 1968. Analyses with the “Returned” variable exclude the region of Paris. I regress unemployed on mean repatriate by region for each of these three measures, a dummy for repatriate for each of the three measures and an interaction between “mean repatriate by region” and the dummy. Means for each of the measures of “repatriate” are reported, as are years in the sample. Age ranges from 22-65. There are region fixed effects for each of the 21 regions, and year fixed effects for each of the years. The analysis is reported separately for men and women.

on own unemployment is similar using both the “Algerian” and “returned” variables with and without including Paris. Using mean repatriates by region collected by Hunt (1992) and taken from the census data, there is no effect of being in a repatriate-dense region on likelihood of being unemployed. Next, I report results controlling for education in Table 3.1. I conduct these analyses only for men as there are few women who report education level. Focusing only on 1969 and 1970, the years in which the education variable is available, I find no effect of being in a repatriatedense region on unemployment rates for either “Algerian” or “returned” measures of repatriation. There is little further change from controlling for education. Restricting the sample does change the unemployment effects of the respondent himself being “Algerian”. With the “Algerian” measure of repatriate variable, restricting the sample erases any increase in unemployment rates for being Columbia Economics Review

“Algerian.” Further controlling for education has no effect. For the “Returned” variable, however, unemployment rates for those who are recently returned stay about the same even with the restricted sample and education controls. Table 4 shows the results of the analyses on hours worked. Areas with more repatriates using the “Algerian” measure of repatriate have about 0.45 more hours worked per week (p<0.01) than areas with no repatriates. “Algerians”, however, work about one fewer hour per week than the rest of the population. Using the “returned” measure of repatriate, there is no result on mean repatriate by region, the repatriate dummy or the interaction. Repeating the “returned” analyses with Paris included, areas with more repatriates have about 0.03 fewer hours worked per week, and those recently “returned” work on average one fewer hour per week. The hours worked analyses with controls for education are reported in Table

Fall 2017

Table 3.1: The impact of immigration on individual unemployment rates with restricted years (1969, 1970) controlling for education

Unemployed is a dummy variable for whether the person is unemployed and is fitted with a logit model. Marginal fixed effects are reported at the means. There are two measures of repatriate: “Algerian” is defined as those who indicate their nationality is Algerian and “Returned” is defined as those who indicate their nationality is French and their previous region of residence as outside of the French Metropole or who indicate their nationality is Algerian and is only available in the 1965, 1969 and 1970 data sets. Analyses with the “Returned” variable exclude the region of Paris. I regress unemployed on mean repatriate by region for each of these three measures, a dummy for repatriate for each of the three measures and an interaction between “mean repatriate by region” and the dummy. I repeat the analyses only for the years where the education variable is available (1969, 1970) with and without education. Some high school education or more is a dummy for whether the person has begun high school. Means for each of the measures of “repatriate” are reported, as are the years in the sample. Age ranges from 22-65. There are region fixed effects for each of the 21 regions, and year fixed effects for each of the years. Women are not included because there are few education observations for women.

4.1. Changing the years of the sample removes any effect on hours worked using either measure of repatriate, but adding education controls to the restricted sample has no effect. Repeating the labor market analysis using mean repatriate by department, I find that departments with more repatriates– using either measure of repatriate –have higher unemployment rates (see Table 4.2). Departments with the highest proportion repatriates–about 4.2%–have less than a 0.01 percentage point higher likelihood that an individual is unemployed. Repatriates themselves have about 5% higher likelihood of being unemployed. Again, being a repatriate in a repatriate-

dense region erases about 10% of the penalty for being a repatriate. The hours effect at the department level is opposite in sign to what might be expected. In departments with the highest proportion repatriates, hours worked per week are about 5.4 hours higher than in departments with no repatriates. There is no effect of being a repatriate, or of the interaction of the repatriate dummy and mean repatriate by department on hours worked. Combining the unemployment and hours worked analyses, at the department level, my results suggest that in departments with a higher proportion of repatriates, individuals are slightly less likely to be working, but if they are

Columbia Economics Review

working, they are working more hours per week. DEMAND SIDE ANALYSIS Methods The arrival of the Pieds-Noirs had three stages of impact on the local housing market. Following the arrival of the Pieds-Noirs, existing homes should become more crowded. This should lead to the construction of new homes. Lastly, an increase in demand for housing construction should lead to an increase in demand for construction workers. I calculate crowding as the number of people per room for each household. Using an OLS model, I regress crowding on percent repatriate by region, a dummy for whether the person is a repatriate, and the interaction term for each measure of repatriate. I repeat the analysis with Hunt’s repatriate means by region. To make sure that there is only one observation for each household, and so that more crowded homes are not being over-counted, the data set is restricted to only include observations for the head of each household. I repeat the regressions for the years 1964, 1968, 1969 and 1970 separately to observe changes in the coefficient on the repatriate variables over time. Finally, the analysis is repeated with all years and controls for region and year fixed effects. In all analyses, controls for age, age squared, and sex are included. I repeat the analysis with mean repatriate by department for each year and then with department level fixed effects. Next I compare the age of homes in areas with more repatriates to the rest of France. Living in a new home is defined as living in a home built after 1962. I use a logit model to regress whether the home is new on mean repatriate by region, a repatriate dummy, and the interaction term for each measure of repatriate. Marginal effects are reported at the means. Control for age, age squared, and sex are included. I repeat the analysis separately by year (1965, 1969, and 1970) and then with all years and fixed effects. I repeat the analysis with mean repatriate by department and department fixed effects. Lastly, if more homes are being built, there should be an increased demand for construction workers. I use a logit model to regress whether a person is a construction worker on percent repatriate by region, a repatriate dummy and an interaction term and report marginal effects at the means. Only those employed are included in the analysis. Controls for age, age squared, sex, and education are included. I repeat the analysis individu-

Fall 2017 Table 4. The impact of the influx of repatriates on individual level on hours worked per week with mean repatriate by region.

Results are OLS estimates. Hours worked is the number of hours worked per week and is capped at 70. The sample is restricted to only those who are employed. The hours worked variable is only present in the 1968, 1969 and 1970 data sets. There are three measures of repatriate: “Algerian” is defined as those who indicate their nationality is Algerian; “Returned” is defined as those who indicate their nationality is French and their previous region of residence as outside of the France Metropole or who indicate their nationality is Algerian; Hunt measure of Repatriate is the percent Repatriate by region taken from Hunt (1992) and is only available for the years 1962 and 1968. Analyses with the “Returned” variable exclude the region of Paris. I regress hours worked on mean repatriate by region for each of these three measures, a dummy for repatriate for each of the three measures and an interaction between “mean repatriate by region” and the dummy. Means for each of the measures of “repatriate” are reported, as are years the years in the sample. Age ranges from 22-65. There are region fixed effects for each of the 21 regions, and year fixed effects for each of the years. The analysis is reported separately for men and women.

ally by year and then again for all years with fixed effects. The analysis is then repeated using mean repatriate by department and department fixed effects to see if the region level masks some clumping. Results In Table 1, I show housing and construction patterns across France over the time period. Crowding (number of people per number of rooms) decreases slightly from 1.4 in 1964 to 1.17 in 1970. The number of people living in a new home increases dramatically over the time period. In 1964, only 4.8% of the homes are new, compared to 24.3% in 1970. Part of the reason for this dramatic increase is that the baseline for new homes is 1962, so since construction is ongoing, the share of homes that are identified as new will necessarily increase. The percent of those employed who work in the building and public works industries remains at about 10% throughout the period. Table 5 shows the results of the regression of crowding by region across years for each individual year on each measure of repatriate. Crowding is quadratic in

Table 4.1. The impact of French repatriates on individual level hours worked per week with a restricted years and education controls.

Results are OLS estimates. Hours worked is the number of hours worked per week and is capped at 70. The sample is restricted to only those who are employed. The variable is only present in the 1968, 1969, and 1970 data sets. I repeat the analyses for each of the measures of mean repatriate by region: “Algerian” is defined as those who indicate their nationality is Algerian, “Returned” is defined as those who indicate their nationality is French and their previous region of residence as outside of the French Metropolis or who indicate their nationality is Algerian (it is only available in the 1965, 1969 and 1970 data series). Analyses with “returned” do not include Paris. I regress hours worked on mean repatriate by region for each of these three measures, a dummy for repatriate for each of the three measures and an interaction between “mean repatriate by region” and the dummy. Some high school education is a dummy for whether the person has begun high school (it is only included in the 1965, 1969 and 1970 data sets) so I repeat the analysis with and without the education variable for the same set of years. There are region dummies for each of the 21 regions, and year fixed effects for the two years. Age ranges from 22-65. The analysis only includes men as there are few education observations for women.

age and reaches its maximum at about age 32. Using the “Algerian” measure of repatriation, the coefficient on crowding in repatriate-dense regions decreases from 19.47 to -5.53 between 1964 and 1969 and then jumps to 3.35 in 1970, but remains below the 1964 high. “Algerians” consistently live in more crowded homes than non-Algerians, and “Algerians” in regions with a higher proportion “Algerian” live in significantly more crowded homes. There is only data from 1969 and 1970 for the “returned” measure of repatriate and so, while crowding in repatriatedense regions does increase from 1969 to 1970, that change takes place after most of the wave of repatriation has already ended. Additionally, identifying changes off two sequential years of data does not provide much variation in the data. There is a similar problem with mean repatriate by department, for which only three years of applicable data are available. Table 5.1 shows that the effect of repatriates on crowding is less consistent with means by department, but there are also fewer years in the sample and thus less variation in the data. Crowding in “Algerian”Columbia Economics Review

dense departments decreases from 1968 to 1969 and then rises from 1969 to 1970 using both measures of repatriate. In general, repatriates tend to have

“‘Algerians’ consistently live in more crowded homes than non-Algerians and “Algerians” in regions with a higher proportion “Algerian” live in significantly more crowded homes.” higher levels of crowding than non-repatriates. (See table 5.2) Repatriates in departments with a higher proportion of repatriates have consistently decreas-

Fall 2017

Table 5. The impact of the influx of repatriates on regional crowding (year by year).

Table 5.1. The impact of the influx of repatriates on department level crowding (year by year). Crowding is the number of people per number of rooms. There is one observation per household taken from the head of the household for the years 1968, 1969 and 1970. Results are OLS estimates. “Algerian” is defined as those who indicate their nationality is Algerian. “Returned” is defined as those who indicate their nationality is French and their previous region of residence as outside of the French Metropole or who indicate their nationality is Algerian (only included in the 1965, 1969, and 1970 data sets). Analyses with Returned do not include Paris. Age ranges from 22-65. Sex is 0 if male and 1 if female. I regress crowding on mean repatriate by department, a dummy for repatriate and an interaction between the dummy and mean repatriate by department for each measure of repatriate. The analysis is repeated by year for each measure of repatriate.

Columbia Economics Review

ing levels of crowding using both measures of repatriate. Using the “Algerian” measure of repatriate, the coefficient on crowding for “Algerians” in “Algerian”dense departments decreases from 24.26 in 1968 to -5.198 in 1970. I repeat the analysis across all years with region and department level fixed effects (Table 5.2). Across region and department for both measures of repatriate, repatriates have significantly higher levels of crowding (p<0.01) and repatriates living in regions with a higher proportion repatriate have significantly higher levels of crowding (p<0.01). The coefficient on the interaction is inconsistent between regions and departments. Table 6 shows the results of the new home regressions. For the new home variable, I regress each year separately to see how the coefficient changes over time and then look at results across years including fixed effects. Using the “Algerian” measure of repatriate, the likelihood of living in a new home in a repatriate-dense region increases from 0.721 (p<0.01) in 1964 to 4.976 in 1970 (p<0.01). Using the “returned” measure of repatriate, the likelihood of living in a new home falls between 1969 and 1970. However, there are only two years in the sample and the mean number of “returned” in France is fairly stable. Using the Hunt measure, being in a repatriate-dense region also significantly increases the likelihood of living in a new home in 1968 by 0.01

“In departments with a higher proportion of repatriates, individuals are slightly less likely to be working, but if they are working, they are working more hours per week.” percentage points (p<0.01). For all measures, repatriates are about equally likely to live in a new home as non-repatriates. The likelihood of living in a new home if you are a repatriate in a repatriate-dense region is also increasing over the time period for both measures of repatriate. Using mean repatriate by department (see Table 6.1), some of the effect disappears but there are only two years of data

Fall 2017 (1969 and 1970) and little variation in the number of repatriates in France across those years. The likelihood of living in a new home if you live in a repatriatedense region is somewhat ambiguous, as is the likelihood of living in a new home if you are a repatriate. However, for both measures of repatriate, the likelihood of living in a new home if you are a repatriate in a repatriate-dense region is increasing from 1969 to 1970. Repeating the analysis with region and department fixed effects across all years

Table 5.2. The impact of the influx of repatriates on crowding for all years with region, year and department fixed effects.

“[In areas with more repatriates,] Crowding [...] generally decreased, more new homes were built [...] and the number of construction workers increased.” (see Table 6.2), repatriates are consistently and significantly less likely to live in a new home than non-repatriates. This may be because there are only a few years in the sample and the fixed effects are over-counting the variation between two sequential years. The last part of the demand side analysis is the impact of immigration on the number of construction workers in a region. Table 7 shows regressions on construction worker by region and year. The likelihood of being a construction worker if you live in a repatriate-dense region is consistently increasing from 1968 to 1970 using both measures of repatriate. The coefficient increases from -1.072 in 1968 to 0.504 in 1970 using the “Algerian” measure of repatriate. Using the “returned” measure of repatriate, there is no effect in 1969, and there are 5% (p<0.05) more construction workers in “returned” heavy regions in 1970. In general, repatriates also have a higher likelihood of being construction workers. However, there is a somewhat ambiguous of effect of being a repatriate in a repatriate-dense region on likelihood of being a construction worker. Using mean repatriate by department, I find similar results (Table 7.1). Using the “Algerian” measure of repatriate, the likelihood of being in the construction industry if residing in a region with a higher

Crowding is the number of people per number of rooms. There is one observation per household taken from the head of the household for the years 1964, 1968, 1969 and 1970. Results are OLS estimates. Algerian is defined as those who indicate their nationality is Algerian. “Returned” is defined as those who indicate their nationality is French and their previous region of residence as outside of the French Metropole or who indicate their nationality is Algerian (only included in the 1965, 1969, and 1970 data sets). Analyses with the Returned variable do not include Paris. Hunt measure of Algerian is the percent Algerian by region taken from Hunt (1992) and only available for the years 1962 and 1968. I regress crowding on mean repatriate by region, a dummy for repatriate and an interaction between the dummy and mean repatriate by region for each measure of repatriate. Age ranges from 22-65. Sex is 0 if male and 1 if female. In columns 1, 2 and 3, I include, year and region fixed effects. In columns 4 and 5 I include year and department fixed effects.

Table 6. The impact of the influx of repatriates on likelihood of living in a new home with mean repatriate by region (year by year). New home is a dummy variable for whether the home was built after 1962. There is one observation per household taken from the head of the household for the years 1964, 1969 and 1970. Results are marginal fixed effects at the means from a logit model. “Algerian” is defined as those who indicate their nationality is Algerian. “Returned” is defined as those who indicate their nationality is French and their previous region of residence as outside of the French Metropole or who indicate their nationality is Algerian (only included in the 1965, 1969 and 1970 data sets). Analyses with the Returned variable do not include Paris. Hunt measure of repatriate is the percent repatriate by region taken from Hunt (1992) and is only reported for 1962. Age ranges from 22-65. Sex is 0 if male and 1 if female. I regress new home on mean repatriate by region, a dummy for repatriate and an interaction between the dummy and mean repatriate by region for each measure of repatriate. The analysis is repeated by year for each measure of repatriate.

Columbia Economics Review

Fall 2017

Table 6.1. The impact of the influx of repatriates on likelihood of living in a new home with mean repatriate by department (year by year)

Table 6.2. The impact of the influx of repatriates on likelihood of living in a new home with region, year and department fixed effects.

New home is a dummy variable for whether the home was built after 1962. There is one observation per household taken from the head of the household for the years 1969 and 1970. Results are marginal fixed effects reported at the means from a logit model.” Algerian” is defined as those who indicate their nationality is Algerian. “Returned” is defined as those who indicate their nationality is French and their previous region of residence as outside of the French Metropole or who indicate their nationality is Algerian (only included in the 1965, 1969 and 1970 data sets). Analyses with Returned do not include departments in the region of Paris. Age ranges from 22-65. New Home is a dummy variable for whether the home was built after 1962. There is one obSex is 0 if male and 1 if female. I regress new home on mean repatriate by department, a dummy for repatriate and an interaction between the dummy and mean repatriate by department for each servation per household taken from the head of the household for the years 1964, 1968, 1969 and 1970. Results are marginal fixed effects reported at the means from a logit model. “Algerian” is measure of repatriate. The analysis is repeated by year for each measure of repatriate. defined as those who indicate their nationality is Algerian. “Returned” is defined as those who indicate their nationality is French and their previous region of residence as outside of the French Table 7. The impact of the influx of repatriates on likelihood Metropole or who indicate their nationality is Algerian (only included in the 1965, 1969 and 1970 data sets). Analyses with Returned do not include the region of Paris. Hunt measure of repatriate of being a construction worker with mean repatriate by region is the percent repatriate by region taken from Hunt (1992) and only available for the years 1962 and 1968. I regress new home on mean repatriate by region, a dummy for repatriate and an in(year by year). teraction between the dummy and mean repatriate by region for each measure of repatriate. Age ranges from 22-65. Sex is 0 if male and 1 if female. In columns 1, 2 and 3, I include, year and region fixed effects. In columns 4 and 5 I include year and department fixed effects.

Table 7.1. The impact of the influx of repatriates on likelihood of being a construction worker with mean repatriate by department (year by year).

Construction worker is a dummy for being employed in the building or public works industry. The analysis only includes those employed. There is one observation per household taken from the head of the household for the years 1964, 1969 and 1970. Results are marginal fixed effects at the means from a logit model. Algerian” is defined as those who indicate their nationality is Algerian. “Returned” is defined as those who indicate their nationality is French and their previous region of residence as outside of the French Metropole or who indicate their nationality is Algerian (only included in the 1965, 1969 and 1970 data sets). Analyses with Returned do not include the region of Paris. Hunt measure of repatriate is the percent repatriate by region taken from Hunt (1992) and is only reported for 1968. Age ranges from 22-65. Sex is 0 if male and 1 if female. I regress construction on mean repatriate by region, a dummy for repatriate and an interaction between the dummy and mean repatriate by region for each measure of repatriate. Some high school education is a dummy for whether the person has begun high school. The analysis is repeated by year for each measure of repatriate.

Construction worker is a dummy for being employed in the building or public works industry. The analysis only includes those employed. There is one observation per household taken from the head of the household for the years 1969, and 1970. Results are marginal fixed effects at the means from a logit model. “Algerian” is defined as those who indicate their nationality is Algerian. “Returned” is defined as those who indicate their nationality is French and their previous region of residence as outside of the French Metropole or who indicate their nationality is Algerian (only included in the 1965, 1969 and 1970 data sets). Analyses with Returned do not include the region of Paris. Age ranges from 22-65. Sex is 0 if male and 1 if female. I regress construction on mean repatriate by department, a dummy for repatriate and an interaction between the dummy and mean repatriate by department for each measure of repatriate. Some high school education is a dummy for whether the person has begun high school. The analysis is repeated by year for each measure of repatriate.

Columbia Economics Review

Fall 2017 Table 7.2. The impact of the influx of repatriates on likelihood of being a construction worker with region, year and département fixed effects.

Construction worker is a dummy for being employed in the building or public works industry. The analysis only includes those employed. There is one observation per household taken from the head of the household for the years 1964, 1968, 1969 and 1970. Results are marginal fixed effects reported at the means from a logit model. “Algerian” is defined as those who indicate their nationality is Algerian. “Returned” is defined as those who indicate their nationality is French and their previous region of residence as outside of the French Metropole or who indicate their nationality is Algerian (only included in the 1965, 1969 and 1970 data sets). Analyses with “Returned” do not include the region of Paris. I regress new home on mean repatriate by region, a dummy for repatriate and an interaction between the dummy and mean repatriate by region for each measure of repatriate. Age ranges from 22-65. Some high school education is a dummy for whether the person has begun high school. Sex is 0 if male and 1 if female. In columns 1 and 2, I include, year and region fixed effects. In columns 3 and 4, I include year and department fixed effects.

proportion “Algerian” increases by about 2.5%, and the likelihood of being in the construction industry if you are “Algerian” increases by about 0.6. However, the likelihood of being in construction if you are an “Algerian” in a department with a higher proportion “Algerian” erases about 25% of the effect. The opposite holds true using the “returned” measure for these two years. Last, I repeat the construction analysis for all years and include region, year, and department fixed effects (Table 7.2). Those living in “Algerian”-dense regions are more likely to be construction workers, but not those living in “Algerian”dense departments. “Algerians” are more likely to be construction workers, but being in a region with a higher proportion “Algerian” erases 85% of that effect. Using the “returned” measure, the opposite is true. There is no effect of being recently “returned” on likelihood of being a construction worker, however being a recent “returnee” in a region or department with many recent “returnees” does make it more likely that you would be in the construction industries.

15 Figure 1. Hunt Map of Repatriates as percent of the labor force by region.

Figure is taken from Hunt (1992). This map of France shows repatriates as a proportion of the labor force by region in 1968. The repatriates are the largest portions of the population in the Languedoc and Provence-Côte D’Azur regions. The Midi area of France encompasses Languedoc, Provence and Aquitaine.

CONCLUSION The influx of Pieds-Noirs in the 1960s constitutes an excellent natural experiment to examine whether a supply shock to labor also brings about a shock to the demand for consumption goods. As in Hunt’s prior study of this immigration (1992), I find that the presence of repatriates had a minimal impact on the local labor market. The largest labor market effects I find are those using the “returned” measure of repatriate. Even with these estimates, my results suggest that the region with the most repatriates (about 4.3% of the workforce) had only about 0.01 percentage point higher unemployment rates than regions with no recently “returned.” My findings suggest that this modest result may be because those repatriates had a countervailing effect on local labor demand. I show evidence that is consistent with the hypothesis that housing demand was elevated in areas where there were more repatriates. From 1964 to 1970, crowding in areas with more repatriates generally decreased, more new homes were built in regions with more repatriates, and the number of construction Columbia Economics Review

workers increased. My analysis is limited because of the difficulty of measuring who exactly was an Algerian repatriate and where the repatriates were located. The complicated sense of identity of the Algerian repatriates in France has made them particularly difficult to identify in the data. In this study, I use three different measures of repatriates: one taken from a previous study, one based on nationality and a third based on previous region of residence. All three of these measures are correlated, and the “Algerian” and “Returned” measure are quite highly correlated, but they also differ significantly across region and year. The magnitudes of my results varies depending on which measure of repatriation I use. In the future, I would want to repeat the analysis using more years of data to test these differing measures of repatriation. I would also want to repeat the demandside analysis with a more diverse basket of goods in order to see how immigrants affect different parts of the local economies. n

Fall 2017

Primetime Crime Winning and Losing: The Socioeconomic Impact of Sports on Crime Brendan Moore Columbia University Professional sports have long since captivated America’s attention; regardless of experts’ and pundits’ predictions, the result of a game unfurls before its audience in real-time. This study expands the reach of the relatively small field of sports economics by linking the outcomes of games to crime. Through a numerically-driven analysis of the three most popular major sports leagues, Cooper suggests that unexpected or lucky wins, particularly in Major League Baseball, may decrease the crime levels in the sports teams’ host cities throughout an entire sports season. Cooper’s analysis reminds us that a proper accounting of sports’ intangible externalities must include measures of city welfare, such as crime. -K.W.

INTRODUCTION Five years ago, it was purported that local Baltimore sports star Aquille Carr, nicknamed the Crimestopper, had a magical effect on crime in his neighborhood. The Baltimore Police Department reported that crime would completely halt on nights when his high school basketball team would play. Similarly, in an interview with ESPN in May 2011, renowned Baltimore Ravens linebacker Ray Lewis remarked, “Watch how much crime picks up if you take away our game … there’s nothing else to do.” This paper will explore whether this effect of sports purported by these famous Baltimoreans still holds water. LITERATURE REVIEW The conversation surrounding the link between crime and sports that had only surfaced in sports folklore and local newspapers has been abstracted into formal academic studies in recent years. While the aforementioned Baltimore anecdote captures the influence of sports success on lowering crime, a number of scholars have found the reverse effect as sports have accelerated crime. These papers each get to the root of the social externalities and external public

consumption benefits that these franchises provide, but fail to identify a testable metric for ascertaining these intangible benefits. I believe that city crime is one such variable, often overlooked, that is a good way of determining a city’s general welfare and psyche.

“Winning in sports towns generates a reduction in crime, while losing produces more crime over time.” While past literature on the effect of sports teams and stadia is vast, those that study the consequence of winning and losing are rather sparse. Card and Dahl (2009) study the impact of unexpected NFL game outcomes on domestic abuse, analyzing police reports of family violence on Sundays during the NFL season. Using time-varying controls and a Poisson variable to define home upsets, Columbia Economics Review

Card and Dahl conclude that upset losses by the home team led to an 8% increase in police reports of at-home male-onfemale intimate partner violence. More recently, Eren and Mocan (2016) find that unexpected college football losses increase disposition length on juvenile defendants imposed by the judges by 6.4% among several southern universities in the United States. Kalist and Lee (2016) discover that there is a 2.6% surge in total crime in a city on the day of an NFL regular season home game. And finally, Janhuba (2016) examines the relationship between college football results and life satisfaction, measured by Facebook “likes”, and concludes that unexpected wins have positive effects on life satisfaction. My paper is unique in studying the impact of winning and losing on city crime over the course of the sports season, as opposed to the immediate temporal effects of a game. Harcourt and Ludwig (2006) consider such a correlation between violence and the success of the New York Yankees, but they mention such a theory only as a counterexplanation for another crime theory and acknowledge the vulnerability of constructing such single-city time-series regressions.

Fall 2017 My theory is simple: Winning in sports towns generates a reduction in crime, while losing produces more crime over time. I will test this prediction on a holistic level – covering every professional sports team in the three major sports (baseball, football, and basketball) for every index crime – as well as on a local level for Memphis, Tennessee that may prove a more demonstrable effect. DESCRIPTION OF DATA In constructing regressions to test the impact of sports on urban crime, I gathered three types of datasets: crime statistics, league standings, and crime controls. Overall, the national analysis spans forty-three cities and thirty-one years of data, producing roughly 1,300 observations. In cases in which sports teams had more than one primary host city, I include the team’s record for both city crime districts. The only teams that had crossover between cities were the Texas Rangers (Dallas, Arlington, Fort Worth); New Jersey Nets, New York Giants, and New York Jets (New York City, Newark); and Golden State Warriors (Oakland, San Francisco). I accessed annual city crime data from the United Crime Reporting (UCR) statistics, using specifically the seven primary index crimes in each city from 1985 to 2015. (Portland and Cleveland’s crime data were unavailable in 2015, and Cleveland and Chicago’s rape statistics have large gaps throughout the thirty-year period.) Furthermore, for crime in Memphis, I reached out to their local police department to collect monthly crime statistics. Since Canada has its own distinct system of reporting crime, I exclude Toronto from my analysis. The numerous control variables for this investigation are derived from typically utilized crime control variables – city

population, size of police force, weather effects, and the young African-American male population – as well as those that may correlate with city crime on an annual basis, like unemployment. Some of the controls were only available over the last decade and are dropped in most regressions (See chart below). The control variables for city crime stem from a variety of sources. The UCR provided annual data for city population and police force size. For annual unemployment across cities, I extracted data from the Bureau of Labor Statistics, and for monthly unemployment rate in Memphis, I used the Federal Reserve Bank of St. Louis’ Economic Research. The American Community Survey (ACS) supplied annual data for the population of African-American males in each city from the age of 18 to 24.8 Finally, I included weather controls – average temperature and monthly precipitation – courtesy of the NOAA National Climatic Data Center.9 When running regressions for all sports leagues simultaneously, I averaged the monthly values over the year to convert them to annual averages. Nevertheless, for the regressions that I run for each sport separately, I parsed the weather controls into three sets of averages based on the months in which the leagues are in play: November to May for the NBA, September to January for the NFL, and April-October for the MLB. There were several control variables that I considered, but did not incorporate into my analysis for a number of reasons. One control variable that is prominent in the crime literature is prison or jail population. However, such data is only available on a state level and is not readily available for the number of years that would make it a useful control for my study. Other possible controls that I observed

Columbia Economics Review

17 but did not include are city interactions with gun control laws and a populationweighted variable that accounts for lead exposure and legalized abortion (Reyes (2007) analyzed the effect of lead exposure to criminal activity later in life, while Donohue & Levitt (2001) studied the impact of legalized abortion on crime.) Ultimately, the exclusion of these many controls may lead to omitted variable bias, but the city and time effects, as well as the many crime controls, should help eliminate such bias.

“The only trick was to match the annual crime data to sports data, considering the fact that sports seasons do not directly coincide with the calendar year.” Gathering sports data came mostly from extracting season-by-season standings from sports websites.12 The only trick was to match the annual crime data to sports data, considering the fact that sports seasons do not directly coincide with the calendar year. The NBA Analytics Department supplied me with regular season and playoff standings that followed the calendar year from 1977 to the present. For the NFL data, to find the 2015 calendar record of a team, for instance, I looked at the standings on January 1, 2016, and manually add the wins and losses of games played in January and February of 2015. In my Memphis monthly regression, I also included a win percentage variable that captures the Grizzlies’ monthly win total cumulatively throughout the season, as fans are more sensitive to a team’s cumulative standing than simply the team’s incremental win rate. I choose not to distinguish between home and away games, as this paper tracks the contribution of wins and losses generally and less so the immediate effects of a home win or loss. Lastly, I debated whether to include NHL data, as well as college basketball and football for prominent college towns like Chapel Hill and Auburn, but decided against both because those who follow professional

18 hockey and non-professional sports are less likely to live in a large city or major crime area than are NBA, NFL and MLB fans. To demonstrate stronger causality, I applied an element of luck in the team’s win percentage. For MLB data, BaseballReference already has a built-in luck percentage of wins each season that relates to expectation. Specifically, the MLB luck variable is equal to the difference between a team’s actual wins and their expected record based on number of runs scored and allowed. For NBA and NFL teams, I searched for betting odds in an attempt to construct a regressor that would equal the difference between a team’s actual number of wins and their preseason expected number of wins. While such odds are not available on an annual basis in the years covered by these regressions, percentage of close games won serves as a worthy substitute for measuring

Fall 2017 luck. The close game calendar standings, defined as those decided by five points or fewer, from 1997 to the present, suggest a certain randomness that lends itself well to a causal exogenous effect on crime that is unaffected by other city factors.

“Strong results that show a negative effect on crime into the summer would suggest a possible lag in the effect of crime.”

Columbia Economics Review

EMPIRICAL METHODOLOGY In every regression, I use a team’s winning percentage as the key determinant and crime asthe dependent variable. The regressions are broken down into two sets: the national panel of data and the Memphis monthly tests. For each regression, I run the Ordinary Least Squares (OLS) method to estimate the effect of the success of sports teams on city crime. To best capture crime elasticity, I scale all crime and sports regressors to their natural logs. The first bucket of regressions includes year-by-year crime data and win-loss records of every NBA, MLB, and NFL team, both within the regular season and the postseason. Each index crime – murder, rape, robbery, aggravated assault, burglary, auto theft, and larceny – is included as its own regressor, as are the general categories of violent and property crime. The tests are first carried

Fall 2017 out with all three sports as simultaneous regressors, thus comprising only the eighteen cities that have all three teams. I then test each sport independently, so I can include observations for every NBA, NFL and MLB team, though I recognize that doing so hides the interaction of the other sports variables for cities with multiple sports. Models (1) and (2) are broken down as follows: lCrimeky,c = αNBAlRecordNBA,y,c + αNFLlRecordNFL,y,c + αMLBlRecordMLB,y,c + βXy,c + γy + γc +ε lCrimeky,c = αleaguelRecordleague,y,c + βXleague,y,c + γy + γc +ε

(1)

(2)

In these regressions, k represents type of crime, y designates the year and c the city. Xy,c denotes a combination of yearly and city controls that may affect crime, including population size, size of police force, average temperature, and average precipitation. For (2), the weather effects in Xleague,y,c change based on the average temperature and precipitation in months in which the sport is played. And finally, γy and γc represent the year and city fixed effects. For cities with more than one NBA, NFL, or MLB team, Recordleague,y,c denotes that of the team with the highest win percentage within the respective sport. For example, in Chicago, for any given year, whichever of the Cubs and White Sox had the better MLB record would be used for RecordMLB,y,Chicago. The rationale is that most sports fans in a city will root for whichever team has the greatest success, given the “fair-weather fan” phenomenon. Next, I change the crime-related variables to per capita numbers, transforming population from a control to an implicit component of the dependent regressor. Though crime does not follow a perfect per capita trend and the coefficient of crime on population is relatively inconsistent, a per capita regression may help corroborate results from the first two tests. The following equation models the effect on per capita crime, where Xy,c no longer includes population as a distinct variable: l(Crimek/Population)y,c = αleaguelRecordleague,y,c + βXleague,y,c + γy + γc + ε

(3)

In the second bundle of regressions, I look at Memphis exclusively and use monthly crime in order to study the dips and spikes in crime on a local level. The goal of this set of regressions is to capture whether city crime fluctuates on a monthly basis based

on the Grizzlies’ performance. (Memphis is an ideal case study, as the Memphis Grizzlies are its only professional sports team and feature prominently in the city’s identity.) The analysis covers months in which the Grizzlies play at least five games – January, February, March, April, May, November and December. I include year and month fixed effects and unemployment and weather effects as the controls. In addition, I consider the Grizzlies’ win percentage within the month specifically, as well as its cumulative standing. This city-specific model is shown below: lCrimeky,m = αrawlRecordGrizzlies,y,m + αcumlCumRecordGrizzlies,y,m + βXy,m + γm + γy + ε

(4)

While this regression may demonstrate variation in crime during an NBA season, I recognize that crime heats up most during the summer months when the Grizzlies do not play (Anderson (1987) finds that violent crime particularly picks up in the summer.) Thus, in addition, I run several regressions that contrast the effect during the NBA season and the summer months, the latter using the Grizzlies’ final record as of June. Strong results that show a negative effect on crime into the summer would suggest a possible lag in the effect of crime.

“I found statistically significant results that demonstrate that the successes and failures of sports teams help explain changes in urban crime on both an annual and monthly level.” Alternatively, if the summer crime regression shows a positive correlation, the story would follow that once a successful NBA season has ended, the opportunity cost for committing crime has declined in the summer. Memphis could then be emblematic of any franchise in which fans and the general city residents avert crime due to the team’s success during the year, but jump on the opportunity to commit crime in the summer. To set up the direct Columbia Economics Review

comparison, only the Grizzlies’ cumulative record is used for the primary explanatory variable: lCrimeky,m = αlCumRecordGrizzlies,y,m + βXy,m + ε

(5)

lCrimeky,summer m = αlCumRecordGrizzlies,y,season + βXy,summer m + ε

(6)

Finally, in both regression sets, I implement a “luck” factor that would prove an even stronger effect of sports on crime. Whereas cumulative team records are susceptible to autocorrelation, the luck variables show no serial correlation both across and within the sports season. By looking at team records specifically for close games in the NBA and the difference between actual and expected wins for MLB teams, the α below should capture the direct effect of lucky bounces in sports on crime. The formulae for the “lucky” regressions are as follows, the former spanning all cities with an NBA and MLB team, and the latter for the Memphisspecific monthly data: Crimeky,c = αLeagueRecordLuck,y,c + βXleague,y,c+ γy + γc + ε

(7)

Crimeky,m = αRecordGrizzliesLuck,y,m + βXy,m + γm + γy + ε

(8)

FINDINGS As anticipated, I found statistically significant results that demonstrate that the successes and failures of sports teams help explain changes in urban crime on both an annual and monthly level. Nevertheless, the source of the effect that tested with significance on the national panel – MLB teams – is rather surprising and certainly demands further exploration. Arguably more notably, the results show significance when testing for a luck factor in Memphis, as well as with baseball teams nationally, plausibly suggesting a causal effect on crime. In organizing my findings, I match the tables to the equations above. (Though not indicated in the tables, every sports and crime variable is converted to a natural log, as are the police officer and population controls.) Table 1 shows that the successes and failures of MLB and NBA teams help explain changesin urban crime on an annual level – baseball club success lowering crime and basketball victories counter-intuitively triggering a higher crime rate. Such results are promising in demonstrating a negative correlation between a city’s MLB team success and its crime, but they also warrant

Fall 2017 Table 1: All Cities, Annual, Fixed Effects

Table 2: All Cities, Annual, Fixed Effects, Sports Tested Independently

Table 3: All Cities, Annual, Time Fixed Effects, Per Capita Crime

Columbia Economics Review

an explanation of why NBA team success should drive a higher crime rate. When each sport is examined independently, the statistical significance of the positive NBA correlations is greatly diminished. Table 2 verifies the strength of the MLB team effect, with violent and property criminal activity dropping by almost 2% for a 10% increase in a baseball team’s record. (It is interesting to note that when testing with only time fixed effects and not including city fixed effects at all, NFL teams showed a significantly high negative correlation, while NBA teams again showed the reverse. This suggests that cities with strong NFL teams happened to be low crime rate districts, while those with strong NBA teams tended to be zones with high criminal activity.)

“[T]he results show significance when testing for a luck factor in Memphis, as well as with baseball teams nationally, plausibly suggesting a causal effect on crime.” To ensure that the above findings are not just a fluke, that the best MLB teams do not simply correlate with the lowest crime rates and vice versa, I run regression (3) that adjusts the dependent crime variable to criminal incidents per capita, as well as the officer control to police officers per capita. As Table 3 demonstrates, the NBA effect has completely disappeared, while the initial findings for baseball remain, albeit weakly. The per capita results show that the number of murders and larceny felonies fluctuates in the opposite direction of a city’s MLB team success. Undoubtedly, an essential limitation in the above regressions is the reality that the seasonality of the sports does not match up necessarily with the high peaks of crime. Certain crimes may be more or less pervasive in certain months, but without monthly effects, such cannot be monitored above. Such a realization prompted my decision to choose an individual city as a case study to be able to test the sports effect on crime on a monthly level, specifically during the sports season.

Fall 2017 Transitioning to the Memphis set of regressions, I aim to ascertain whether such a sports effect exists on a more granular level. Though Memphis is host to a basketball club, as opposed to baseball, it is worthwhile to gauge whether there is a discernible effect of sports teams on crime on a monthly basis and in cities that are passionate about their teams. Though the controls I use differ slightly from the national set, I also test both sets of regressions without controls and with only fixed effects to ensure an apples-to-apples comparison. Table 7 shows some significant effects for baseball in line with the original findings, but only for aggravated assault and larceny. However, the mere fact that such a negative correlation is significant at all may be worthy to acknowledge, considering the inherent randomness of the MLB Win% variable. The existence of such a correlation between lucky wins and criminal activity conveys the impression that city crime may be somewhat rooted in the unexpected success and demise of their baseball clubs. At the very least, these results serve as fuel for the next regression that regresses crime on lucky wins in Memphis.

“When there is only one show in town, the sports success may play a monumental role in shaping the city’s wellbeing and happiness.” CONCLUSION This paper presents strong evidence of my initial hypothesis that sports teams affect their host cities’ crime levels, but in ways that have been most surprising. The fact that baseball is the only sport that tested significantly is ironic considering the MLB’s dying fanbase. One can attempt to explain these results by pointing to the commitment and hardline loyalties of baseball fans, but this factor is counterbalanced by the fact that NBA and NFL’s fan bases are more traditionally classified as representing the inner city and the younger population that is more susceptible to crime. My findings fascinatingly corroborate a “Broken Yankees Hypothesis” proposed by Harcourt and Ludwig (2006) as they

Table 3: All Cities, Annual, Time Fixed Effects, Per Capita Crime

consider such a correlation between violence and the success of the New York Yankees, despite the fact such a theory was only used as an amusing counterexplanation for the “Broken Windows Hypothesis.” Nevertheless, the more important conclusion that this paper has drawn is the strong effect of a sports team’s success on various crimes for a city like Memphis. I would be cautious, nonetheless, to draw inferences from Memphis to other sports towns, since Memphis is one of few cities that have only one sports team. Such a negative correlation between sports and crime may be more likely in cities with only one professional sports team that represents the “heart and soul” of the city. (When running the national set of regressions, I also conducted a separate test for the seven cities with just an NBA team, though the results were not telling. To properly test the “one show in town” hypothesis, I would need to execute the same extensive monthly analysis for NBA cities like Orlando, Oklahoma City, Portland, Sacramento, Salt Lake City, and San Antonio as I did for Memphis) When there is only one show in town, the sports success may play a monumental role in shaping the city’s well-being and happiness. The polarization of a win or loss is much more palpable when the team serves as a central passion for a city’s residents. But even so, I would have been careful in drawing any conclusion from Regressions (1) - (6) were it not for the results in the “lucky win” regressions. The relationship on both the national and local levels between a team’s lucky wins and criminal Columbia Economics Review

activity provides strong ammunition for the thesis that as a city’s sports team goes, so goes crime. In spite of the strong results, it is

“The relationship on both the national and local levels between a team’s lucky wins and criminal activity provides strong ammunition for the thesis that as a city’s sports team goes, so goes crime.” essential to recognize the limitations of this study. Attempting to find determinants of crime on an annual level is a tall task, particularly trying to put a finger on a specific regressor. But even if it were possible to detect such a variable, the effect in the real world of any one crime determinant is close to negligible, as crime has so many explanatory variables. Thus, a key takeaway for those looking to replicate this project would be to analyze more individual city case studies, like Memphis, to better test the sports effect. Only then could one truly discern between a “Broken A’s” and a “Broken Rays” Hypothesis and propose new ones that forever alter the way we perceive sports. n

Fall 2017

Barreling Down the Highway Understanding Public Transit Ridership through Gasoline Demand: A Case Study in the San Francisco Bay Area, CA Hansen Ouquan Sun University of California, Berkeley Sunâ&#x20AC;&#x2122;s paper seeks to ascertain the general relationship between gas prices and public transit ridership, as well as estimate a precise value for the cross demand elasticity of gasoline specific to the San Francisco Bay Area. Expanding upon past studies of public transportation, she takes into account a greater range of control variables, such as transit fare increases, and provides a more updated analysis that examines data after the dramatic decline in gas prices around 2008. Sun also thoroughly checks her findings against those of past papers to ensure additional validity of results, and posits cogent explanations where they differ. Her research poses important implications for transit agencies in determining how to adjust their service schedules in response to fluctuating demand, allowing for more efficient allocation of resources. - B.L.H.

INTRODUCTION Research Question This study aims to revisit past studies to ascertain and quantify the relationship between gasoline prices and its impacts on public transit ridership. The price of gasoline affects the total cost of travel incurred by consumers. Based on market observation and standard economic intuition, different travel options will become relatively more attractive at different gas prices. The working hypothesis for this paper, as is consistent with past papers, would be that an increase in gas prices increases public transit ridership and conversely a decrease in gas prices decreases public transit ridership. In determining this relationship, the paper also estimates a cross demand elasticity of gasoline. Motivation Public transportation in the United States has become a significant part to solving economic, energy and environmental challenges that the US faces. The many benefits to public transportation

such as fuel and monetary savings, congestion reduction and lessening carbon footprints are just the main benefits that the American Public Transportation Association (APTA) advertises. According to APTA Statistics, over the 18 year period from 1995 to 2013, average public transportation ridership grew 37.2% while average population growth was 20.3%, a long term trend showing that more Americans are using public transportation. The highest recorded total public transit ridership, as of 2014, was also logged in 2013 (APTA). On average, public transit has risen across US over the years (Figure 1), with the highest number of public transport trips being made in 2013. This trend is also observed in this paperâ&#x20AC;&#x2122;s area of study in the San Francisco Bay area, as shown in the ridership data charts below (Figure 2). The significance of this research lies in isolating key factors of public transit ridership that is potentially useful for policymakers to consider on improving overall ridership for other bigger aims.

Columbia Economics Review

It would be particularly useful to determine the optimal level of subsidies and taxation in order to determine the amount of resources that are dedicated to the development of public transportation. Other papers have also looked into factors that affect public transit ridership, such as highway system characteristics, regional geography, metropolitan economy and population characteristics (Taylor et al, 2008). However, fundamentally, standard economic intuition would point us to the most basic component of all transportation systems, the price of gasoline. Furthermore, since around 2008, there has been a steep drop in oil prices, with a culmination of many factors such as the onset of increased supply of oil domestically from sources like oil fracking, the decision of OPEC not to manipulate oil prices and a weaker global demand for oil. With such changes in the macroeconomic environment, past papers that investigated the trend at the height of gas prices around 2008 might be time out-

Fall 2017

23 changes of gasoline.

Figure 1: Trend of Total American Public Transit Ridership Over Period of Investigation dated. Choosing the Research Location California is known to have higher gas prices than most other states in the US, with approximately 30 cents extra per gallon on average (Fueling California, 2016) The many unique features of California such as it being a â&#x20AC;&#x2DC;fuel islandâ&#x20AC;&#x2122; and not being linked directly to petroleum or crude oil supplies decrease the available supply compared to other states. Furthermore, there are many California regulations on the blends of petroleum that also cause the prices to increase. With the higher prices, these make the study of the gasoline price impacts more significant in California. In determining the specific location within California to be investigated, a key selection criteria was the quality of service that is provided for by the public transportation network. Regardless of the city size, a transit system that does not provide adequate service to prove a viable alternative to driving cannot expect travelers to change modes, no matter the expense savings, as this disrupts their mobility (Haire & Machemehl, 2007). The converse is also true, where driving is no longer a necessity in localities like New York City, where the population is too heavily dependent on public transit. They are also demand inelastic with regards to changing gas prices because of the lack of viable alternatives available (Yanmaz-Tuzel & Ozbay (2010). Hence,

as proposed in Yanmaz-Tuzel and Ozbay (2010), metropolitan areas with established transit systems and large numbers of automobile commuters are the best locations to test such a hypothesis of the substitutability of public transport. Indeed, this is the main rationale for choosing the San Francisco Bay Area, where there possibly could be a substitution effect being observed due to the price

The Significance of Gas Prices Following the regression models of past papers, this paper will run similar regressions aligned with the models used in previous studies, using a time-series dataset that spans monthly data from January 2002 to August 2015, through the following different models that utilize the Ordinary Least Squares estimates. More will be discussed in the Data and Methodology section of the paper. Constant elasticity, where we assume a constant elasticity throughout the period of study. Event elasticity, where we assume that elasticity values as events such as fare price increase and structural upgrades occur within the transport system. Instrumental variables analysis, where we account for a possible endogeneity within the model, and use an instrument variable to correctly predict the effect. Lagged models, in which we assume that consumers respond to the gas prices with a lag time of 6 months. Furthermore, the fixed effects control variables that are being used within the regression methods are more comprehensive than past studies. To account for possible omitted variable biases, control variables such as the yearly and monthly fixed effects, system upgrades, fare increases and unemployment are included. Details and the reasoning for inclusion of

Figure 2: Trend of San Francisco Public Transit Ridership Over Period of Investigation

Columbia Economics Review

Fall 2017

Columbia Economics Review

Fall 2017

Columbia Economics Review

26 the variables are elaborated in the Data and Methodology components of the paper. Agency Responses to Gasoline Price changes After determining the significant relationship between gasoline prices and ridership, Transit/Vehicle Revenue Hours is being tested for significant relationship with gasoline prices. Using the most appropriate regression model decided for each transit mode in the analysis above, we change the dependent variable from Log(Ridership) to that of Log(Revenue Hours) in this segment of analysis. Providing evidence of the substitution effect To justify the elasticity, or inelasticity of the values obtained of public transit, evidence of the substitution is being investigated along critical highway segments of the San Francisco Bay Area, where service such as the BART and MUNI operate, such as the CA-24E, I-80W and US101 highways. Data of average speeds of the highway at the peak hours (5-7pm) at these freeway segments was collected. Using the most appropriate regression model determined in the first segment of analysis, the highway speed is used as the dependent variable to predict and test for a corresponding significant relationship with the independent variables. The underlying foundation of such analysis is that a higher gasoline price would deter driving, and in doing so with less drivers on the roads will lead to an increase in average travelling speeds which are picked up by these sensors along the freeway. Similarly, the contrary, whereby gasoline prices are lower would encourage more people to drive and hence result in a lower average travelling speeds because of the traffic congestion caused. Key Contributions Past papers have provided this paper with the fundamental models of regression that will be covered in detail in the next section. The understanding of the methodology behind the papers have elucidated three key areas of renewal that this paper hopes to achieve. Firstly, the time relevance of past research is to be questioned, particularly after the significant decrease in crude oil prices in 2008. All of the research mentioned above were conducted during the time where prices of oil have been on a steady upward trend. Given the significant drop of prices from the high of

Fall 2017 ~USD100/barrel to the current ~USD45/ barrel, this update of macroeconomic circumstances will be an important contribution to the literature that follows beyond this paper, to ascertain if the relationship between gas prices and public transit ridership is still statistically significant and time relevant. This is particularly pertinent as many research reports, including one from the National Public Radio (2015) that point to the likely possibility of an environment with consistently lower gasoline prices than before. Under such an environment, policymakers will have to perhaps revise their judgement and policy decisions. Secondly, upon careful inspection of

â&#x20AC;&#x153;There seems to be an overall positive effect of gas prices on public transit ridership, statistically significant in all transportation modes investigated, of the Heavy Rail, Light Rail and Bus services.â&#x20AC;? the econometric methods utilized in past studies such as Currie and Phung (2007), it occurs that the regressions in those studies have not accounted nor controlled fully for other important variables that are determinants of public transit ridership. This paper posits that there could be phenomenon of Omitted Variable Bias (OVB) present. In the past papers, the only control variables utilized in the models were the monthly fixed effects, which this paper also takes into account, but other control variables such as yearly and monthly fixed effects, Population Unemployment Rate, Fare Price Increases (FI) and System Upgrades were not included. This paper gathered all of these important variables and will include them in the regressions as determinants as to whether a person would take public transport. Hence, this paper will show the value of adding these control variables into the regression and test for their significance. Columbia Economics Review

Finally, previous studies have used aggregated data from many cities differing in size and density across the entire United States, which has limited policy implications given the decentralized political nature of the various transit agencies. Other studies that covered on one locality were on other metropolitan areas such as New Jersey, Philadelphia etc. The focus on San Francisco Bay Area is novel and serves a local purpose to shed insights for the San Francisco Transportation planning agencies and public transit operators. Key Findings The ridership trends were evaluated and compared with the California gas prices for each mode of public transit in the San Francisco Bay Area. Based on the studyâ&#x20AC;&#x2122;s regression, there seems to be an overall positive effect of gas prices on public transit ridership, statistically significant in all transportation modes investigated, of the Heavy Rail, Light Rail and Bus services. The elasticity values obtained from the regressions range from 0.0581 to 0.147. It was also discovered that different regression models best suited different transportation modes, with the Constant Model applicable for both the MUNI Bus and MUNI Light Rail, whereas the Behavioral Model applying for the BART. Using the Vehicle Revenue Hours to relate to the same set of independent and control variables, it appears that the transit agencies do not currently take into account fluctuations in gasoline prices to factor for the transit service schedules. Upon further investigation, there are preliminary relationships between the gasoline prices and highway speed at selected key highway segments within the Bay Area, near the operation of the BART, MUNI Light Rail and MUNI Bus modes. The following sections will elaborate these key findings in more detail. LITERATURE REVIEW Relevant Transportation Research Amongst the few papers that have investigated the relationships between gasoline prices and public transit ridership, the consensus from these papers seems that there is a small but significant amount of the variability in transit ridership that is attributable to fluctuations in gasoline prices (Lane, 2010). This is indeed the relationship that this paper wishes to ascertain given the changing gasoline price environment.

Fall 2017 The Study of Time and Seasonal Effects The time and seasonal effects in studying public transportation behavior is highly vital and accounted for in most relevant literature. Haire and Machemehl (2007) showed that the seasonal average of trips within the US from 1992-2001 differed greatly from seasons, with the Spring months (March-May) and Fall (Sept-Oct) showing much more trips being made. Furthermore, controlling for seasons also yielded higher t-statistics on the gas price coefficient, giving evidence that this is an important variable to be controlled for. Similarly, in Currie and Phung (2007) and Yanmaz-Tuzel and Ozbay (2010), monthly dummy variables were included to account for these cross month differences. Lagged Temporal Choices As for behavioral choices, Goodwin (1992) found that behavioral response to cost changes in transportation is a response that takes place over time. Hence, people take time to adapt to the change in prices. This was similarly talked about in Yanmaz-Tuzel & Ozbay (2010) in the New Jersey context, where there was evidence that several months elapse before travelers respond to gasoline price changes. Currie and Phung (2007) also agrees on the same point by specifying evidence of 7-8 months time lag between changes in gas price and public transit demand. This gives the theoretical underpinning of the Behavioral Lag model that will be elabo-

rated further in the Data and Methodology segment of the paper. Miley and Weinberger (2007) ends off their paper with future considerations with a question that this paper hopes to find some information on. Towards the end of their study period in Summer 2008, gas prices had decreased substantially in Philadelphia. However, ridership had increased, which contradicts standard economic theory, or our intuition, under conditions of ceteris paribus. Could habit formation play a role in determining public transport ridership? With renewed data, it is possible to find more evidence for the above hypotheses, and this paper institutes a six-month behavioral lag gas price variable to test this hypothesis. New Areas of Consideration The contributions of this paper are threefold. Firstly, this paper seeks to establish and confirm the time relevance of the gas-transit relationship. Secondly, this paper utilizes a more robust regression method in determining this relationship by expanding on past regression models through including more control variables. Thirdly, this paper also carries out the analysis in the San Francisco locality, providing novel insights in an area that has not been previously investigated. EMPIRICAL STRATEGY AND DATA METHODOLOGY Transport Choice Modelling As a foundation to frame the public transportation choice issue into an economics problem, this paper refers to the

Columbia Economics Review

27 standard model, which represents the costs structure and considerations from the perspective of an urban dwelling individual (O’Sullivan). This model brings into perspective on how fuel costs enter the decision making process, and we make a few assumptions to explicate the model, namely: The individual has three options to choose from: driving alone, riding a bus or riding a train (both heavy and light rail). This essentially reflects a choice between private and public transportation. The individual seeks to cost minimize, as represented in the equation below: Trip Cost = m + Tada + Tv dv (1) Full cost of the trip involves both monetary and time costs, where time costs are split into both access time (Ta) and travel time (Tv ). O’Sullivan showed with empirical studies that the marginal disutility of access time is larger than marginal disutility of in-vehicle time. The model hence shows the implications of fluctuating gas price can have significant impacts on “m” that will affect commuter decisions in deciding to use private or public transportation. DESCRIPTION OF DATA Data Collection and Sources Sources Of Data: Given the novel nature of this study, the dataset was self constructed using various sources. The two most important datasets used to construct this dataset were the independent variable of gasoline prices and the dependent variable of public transit ridership information. Ridership Information was obtained from the American Public Transport Administration’s (APTA) Ridership Report section, where information about the Average Monthly Ridership for various transport systems around America are documented since January 2002. It specifically also provides breakdown into the various modes of transport within each of the transport system, allowing users to focus on the subset of data required. Specifically for the San Francisco geographic subset, there are 164 data points in the latest monthly dataset that spans January 2002 to August 2015 data on the Unlinked Passenger Trips (UPT) figures for the Bus, Light rail and Heavy rail services that this paper analyses, in the forms of Bay Area Rapid Transit (BART) and the San Francisco Municipal Railway Bus (MUNI Bus) and Light Rail (MUNI). This dataset is updated quarterly to include the latest figures, which presents opportunities for

28 the study to update the regressions as more data become available. The gasoline prices were obtained from the US Energy Information Administration (EIA), which provides a weekly California All Grades All Formulations Retail Gasoline Prices (Dollars per Gallon). The dataset extends back into 2000, offering approximately 180 monthly data points. This paper hence constructs the dataset from the time period of overlap between January 2002 and August 2015. In using the Behavioral Model of analysis, where the paper accounts for a delayed response in public transit to gas prices, we shift the scope of the gas prices scope from July 2001 to February 2015, which is available in the dataset provided by the EIA. This dataset is updated weekly, and also presents opportunities for the study to update the regressions as more data become available. Control Variables: To ensure regression accuracy, this paper constructs many fixed effects that are included in the regression to reduce the omitted variable bias effects. Month and Year dummy variables were created manually based on the information obtained from the above two datasets to account for time and seasonal effects. These monthly dummies will help to account for the inherent seasonal differences experienced in travel patterns between different time periods of the year. 11 dummy variables were created to dictate January to November, with December (dropped) being the base value. Similarly, Years 2003 to 2015 were created, with 2002 (dropped) being the base value to prevent perfect multicollinearity. The unemployment rate gives us an overview of the demographic profile of the population we are studying, and possibly explain for any sudden shocks when there are macroeconomic restructuring that affects consumption patterns. Yanmaz-Tuzel & Ozbay (2010) also mentioned the significant positive impact of economic growth on transit ridership, hence including the unemployment rates in the regression helps to control for the differences in economy state. Unemployment rates in the Bay Area was gathered from datasets obtained from the Bureau of Labor Statistics (BLS), whereby the economic situation of San Francisco-Oakland-Fremont was listed, and unemployment rates of each month going back to 1990 was available. The intuition is that when unemployment rate increases, we might possibly see higher public transit as the people choose public transit to reduce cost incurred of travel. Also, at time

Fall 2017

Figure 8: Trend of CA State Avg and San Francico Gas Prices over the past 11 years periods of higher unemployment, it is likely that work-related commute will be reduced, resulting in decreased total trips of travel. Fare Increases for each mode of transport were carefully constructed in the following ways. For the MUNI Bus and Light Rail systems, they follow a consistent and fixed fare structure, and given that the MUNI website did not specifically present information on fare increases, the author collated the fare increase data through searching for Press Releases and News Archives that documented the fare price increases from the dollar value of $1 in 2002 to $2.25 today. Fare Increases may cause intuitively a decrease in ridership Columbia Economics Review

as the cost of riding public transit could possibly become more similar to that of private transportation. For the BART, as the fare prices were calculated dependent on the stops one travels, an average Fare Increase percentage was obtained from the BART Management itself, and the average percentage increase of the fare was considered for this variable. These system upgrades listed were sourced from the key openings within the systems that would predictably have an increased ridership influence during the period of study. For the BART, there were three system upgrades considered in this research. Firstly, in June 2003, the

Fall 2017

Figure 9: Trend of Unemployment and Gas Prices in San Francisco over period

Figure 10: Deviations of Monthly ridership from Monthly Average San Francisco International Airport (SFO) Extension, with South San Francisco, San Bruno and Millbrae stations were connected. Secondly, in February 2011, the West Dublin/Pleasanton station was opened. Thirdly, and finally, the Oakland International Airport (OAK) Connector service was opened in November 2014. For the MUNI, there was only one significant line extension which is the opening of the Third Street (T line) in April 2007 was marked as a significant upgrade within the MUNI system. For the MUNI Bus, it is much harder to model using discrete events such as the above given the numerous lines that are within the system that have been expanded or collapsed during this period of investigation. Hence, this variable is not included for the MUNI Bus analysis. Finally, the vehicle revenue hours helps to control for the variance experienced from month to month in terms of operating hours and days, and is also provided in the APTA dataset where the UPT values were extracted.

gression utilizes the state gas average, the graphical analysis provided here from data supplied by ‘Gasbuddy.com’ shows us that the usage of State Average does not generally cause a loss of generality, as both follow very similar price movement trajectories. DESCRIPTIVE STATISTICS Interpretation Gas Prices: We can see from the graph in Figure 9 that the gas prices have been on a steady increase up until a point in around 2008 when it had a steep drop. The prices following that time period has been largely volatile, showing much fluctuations in price changes. However, it has since shown a significant decrease in price which extends beyond the dataset (beyond April 2015). There was also a huge increase in the unemployment rate that coincides with around the same time of the gas price drop. As we are estimating the cross demand elasticity, the responsiveness of quantity demanded for a good to a change in the

Applicability of Data chosen: Ridership Data from the APTA segments the SF Muni Ridership into its different modal forms (Light Rail, Cable Car, Bus etc), hence choosing Bus and Light Rail subsections gives us the relevant figures needed. Similarly, as the BART only operates on the Heavy Rail segment, the figures obtained from the dataset can be directly applied onto our analysis. For the gas prices, even though the reColumbia Economics Review

29 price of another good, the terms being measured have to be in natural logarithmic terms (Ln). The same applies to the vehicle revenue hours for all three services, which are presented here. However, for ease of understanding on how the variation of the riderships vary across the months, graphs in Figure 10 depicting how much each month’s average ridership deviates from the average are plotted, split into the various transportation modes. As evident, there is higher ridership throughout the spring and summer months, and lesser ridership over winter months. This shows the presence of seasonal fluctuations of ridership across the modes. Empirical Strategy This paper considers four different models to estimate the cross elasticity. The Constant Elasticity, Event Elasticity and Behavioral Lag models estimates the elasticities using the Ordinary Least Squares (OLS) method whereas the Instrumental Model utilizes a two-stage least squares (2SLS) instrumental regression to estimate the coefficients. In the 2SLS method, the dependent variable’s error terms are correlated with the independent variable, causing an endogeneity problem, and an instrument variable is used to estimate the correct coefficient. The Constant and Event models’ rudimentary structure were adapted from Currie and Phung (2007), whereas the Instrumental and Behavioral Lag were derivations created by the author. For each mode of transport, the four models are applied: 1. Constant Elasticity Model:

Fall 2017

This model measures elasticity by assuming constant elasticity for the entire duration of investigation from 2002-2015. 2. Event Elasticity Model:

Figure 11: Testing for Instrument Relevance using the BART Instrumental Regression

This model measures elasticities, allowing some flexibility in that the elasticity can change as events occur, such as system upgrades and fare price increases, which are reasonable changes that can possibly affectthe demand elasticity of public transit. 3. Instrumental Model:

Figure 12: Testing for Instrument Exogeneity using the BART Instrumental Regression is viable. For instance, a January 2002 ridership figure will correspond to a July 2001 gasoline price.

To account for possible endogeneity between gas prices and public transit ridership, an instrument of crude oil prices (Western Texas Intermediate) is used. This instrument is exogenous and relevant, as there is a high correlation between crude oil prices and gasoline prices, but no direct relationship between crude oil prices and public transit ridership given that crude oil prices have an indirect effect via gasoline prices, conditional on controls. 4. Behavioral Lag Model:

This model is exactly the same as the Constant Elasticity model except for the independent variable, which is a time-lagged Log Gas Price. It might be possible that the response to gasoline prices is not immediate. As noted in both Yanmaz-Tuzel & Ozbay (2010) and Currie and Phung (2007), it takes several months before travelers respond to gasoline price changes; hence, a 6-month gasoline lag variable is used to see if this relationship

Instrument Validity The Federal Reserve Economic Data (FRED) from St. Louis Federal Reserve Bank provides data on the monthly West Texas Intermediate (WTI) prices at Cushing, Oklahoma. The WTI is a grade of crude oil typically used as a benchmark in oil pricing, and the underlying commodity of the New York Mercantile Exchangeâ&#x20AC;&#x2122;s oil futures contracts,(which, being public domain data, makes it reasonably easy to access). Prior to analyzing the results from regressions, the instrument is tested to see if it is relevant and exogenous. WTI prices are highly correlated with gasoline prices because crude oil is a key component in determining gasoline prices, along with refinery, distribution and taxes. Intuitively, there is also no direct relationship between transport ridership and crude oil price. Based on economic reasoning, the exogenous and relevance requirements are satisfied. To confirm this, using the estat firststage function on STATA, the F-statistic for the instrument choice of WTI is generated. With the F-stat from the partial F test >10, which satisfies the rule of not being a weak instrument, the instrument can be said to be valid for the IV regression (see Figure 11). Similarly, using the estat endog function on STATA, the exogeneity of the instrument is also confirmed (see Figure 12). Columbia Economics Review

Gasoline Prices and Public Transit Ridership This paper was able to find statistically significant and positive coefficients for the gasoline price coefficients in each of the modes being investigated, which satisfies the original hypothesis. This paper first replicates the Constant and Events models of what past papers have done, and then proceeds to add in the other control variables using the proposed four models described above. The results will be interpreted by the mode of transport for purposes of clarity, and are displayed in the summary of regression results tables. 6 regressions were conducted for each transportation type, as represented by the short forms in the column heading of the regression output table: the Replicate Constant Model (Rep Const), the Replicate Events Model (Rep Events), Constant Model (Constant), Events Model (Event), Instrumental Model (Instrument) and Behavior Lagged Model (Lag). BART The replicated models of Constant (0.312) and Events (0.0799) elasticities both returned statistically significant coefficients and positive figures for Log Gas. However, the main problem with this model is that because the Yearly fixed effects were not included, there are unexplained deviations in ridership numbers. As a result, the Replicate Models do not have as high of an R-squared value as that of the expanded models. Looking at the enhanced models, the Log Gas coefficient is statistically sig-

Fall 2017 nificant only in the Lagged variable of 6 months. As the four expanded models all have sufficiently high R-squared values (0.943), the choice ofe the most suitable model lies in the underlying economic reasoning. The positive and significant coefficient of 0.0581 for Log Gas Lag suggests that an increase in 1% of lagged gas prices comes with an increase of 0.06% of BART Ridership. A possible explanation for this occurrence lies in the psychology of perception that people have in favor of rail systems over bus. Studies, such as Hensher and Mulley (2015), have been conducted finding that under very similar conditions, when respondents were presented with the same images, with the only difference being bus or train as the mode of transport, the vast majority (73%) chose the train images. This biased perception of transportation essentially points to the fact that there is less demand elasticity for the heavy rail service, particularly in the context when heavy rail systems like the BART operate on their own dedicated lanes of operation, and provide a much more reliable service schedule and expectations (unlike bus and light rail, which may be stuck in traffic jams, etc). Hence, the lower elasticity values for heavy rail can be explained and accounted for in this way. This therefore also supports the Be-

havioral Lag model that is favored over the Constant elasticity model. Given the relative demand inelasticity of BART Ridership, it takes a much longer time for consumers to change their behavior in terms of choosing to ride the BART when gas prices decrease. For the reverse case, when prices increase, there are also higher â&#x20AC;?barriers to entryâ&#x20AC;? because of the relative network effects of the BART system compared to Bus and Light Rail systems. There are generally fewer accessible options for one whose access to a BART station is limited compared to bus and light rail services that are more accessible within the neighborhoods. The time-fixed effects show that there is significant variance in transport numbers at different phases of the year. The monthly fixed effects return statistically significant results for most months, from January to October. The yearly fixed effects were also statistically significant for most years, from 2007-2009 and 20122014. These results provide evidence that there are seasonal and time fluctuations accounted for by the month and yearfixed effects, adding explanatory power to the regression. Unemployment was an important dummy variable, with a negative coefficient of -0.0103, as there are many key stations along the San Francisco and Oakland segment of the service that commute downtown, where the commercial centers of

Columbia Economics Review

31 the cities are. Hence, during a period of high unemployment, there will be significantly fewer passengers who use the service to commute to work. However, the BART Upgrades Dummies and Fare Increases dummy variables were statistically insignificant. It could be the case that the train service was already reaching maximum capacity, and hence the change in passenger services would not have been accurately picked up. We also acknowledge that there perhaps could be unaccounted factors that may have confounded the results, but in summary, the results give us a good explanation of the occurrence. Model chosen: Behavioral Lag Model, Model 6 in Table 3 MUNI Light Rail In both replicate models, the coefficient for gasoline prices wassignificant. However, the sign of the coefficient was not consistent. In the Replicate Events model, the -0.107 coefficient was not intuitive, as a rise in gasoline prices was correlated with a decrease in the MUNI train ridership. It is possible that there is a faultyregression specification or omitted variable bias present. Also, in both regressions, the Rsquared values were very low (0.021 and 0.388). Therefore, the expanded models which had much higher R-squared values were selected for the explanation of Ridership values. Looking at the enhanced models, the Constant model seems to be the mostsuitable model for two reasons. First, it is the only model with a statistically significant coefficient of 0.147 for the Log Gas variable, which makes economic sense. Second, comparing the Constant model with other models, the control variables seem to be more powerfully predictive, with their statistical significance. The positive and significant coefficient of 0.147 for Log Gas suggests that an increase in 1% of gas prices comes with an increase of 0.15% of MUNI ridership, which is about twice the effect of that of the BART elasticity. Interesting, this relationship is the reverse of what previous papers have derived, given that past papers like Currie and Phung (2007) argued for higher sensitivity in elasticity values for Heavy Rail over Light Rail and Bus. The time-fixed effects show that there is significant variance in the transport numbers at different phases of time. The monthly fixed effects show statistically significant results for a few months in January, February and November. Similarly, the yearly fixed effects, also show significant results in the years 2004-2006

and 2004-2015. The MUNI Fare Increase hada strongly significant negative coefficient of -0.213. Due to the nature of the MUNI Light Rails, and the general traffic conditions in San Francisco, the fare increase likely prompted a high degree of transport substitution. Alternatives such as biking, walking or even app-based taxis such as Uber and Lyft, are all possible contributors to the shift in ridership figures. On that note, it is possible that with MUNI possibly being the cheapest form of public transportation (other than walking), when unemployment rates increase, there is an increase in ridership, as supported by the statistically significant coefficient of 0.0744. Interestingly though, the MUNI Upgrade of the T line did not seem to have caused a significant difference in ridership. Comparing this to the BART Models, it does seem that there might be unexplained deviations due to the lower R-squared values. Model chosen: Constant Model, Model 3 from Table 4.

Fall 2017

Looking at the enhanced models, there are three models (Constant, Instrument and Behavioral Lag) that return statistically significant coefficients. However, the Constant model seems to be the most suitable model for three reasons. This is chosen through the processes of elimination. First, in the Lagged model, the coefficient for Log Gas Lag is -0.110, which is significantly negative anddoes not make intuitive economic sense. Furthermore, the Yearly fixed effects dummy variablesare not statistically significant unlike those of the other models. Hence, we look at the other two possible models. We conduct the tests of endogeneity with STATA to decide if the instrumental model is viable. Mixed results were obtained, with the Durbin test showing a need for instrumental regression, and the Wu-Hausman test displaying the re-

MUNI Bus For the replicated models, only the Events model showed a statistically significant coefficient that was negative, -0.0460 which is not intuitive given that a rise in gas prices is accompanied by a decrease in MUNI Bus ridership. It is possible that there is a faulty regression specification or omitted variable bias present. Furthermore, the R-squared values for the replicated models were significantly lower than the expanded models. Hence, we look at the enhanced models for further explanation. Columbia Economics Review

verse results. Therefore, the conservative approach would be to adopt the Constant model. Furthermore, the adjusted R-squared value for the Constant model is also the highest amongst all the expanded models, at 0.686, which suggests that this model has the highest explanatory power. The positive and significant coefficient of 0.125 for Log Gas suggests that an increase in 1% of gas prices comes with an increase of 0.13% of MUNI ridership, which is about twice the effect of that of the BART elasticity. Similar to the earlier section on MUNI, this relationship is the reverse of what previous papers have derived, given that papers like Currie and Phung (2007) argued for higher sensitivity in elasticity values for Heavy Rail over Light Rail and Bus. The time fixed effects show that there is significant variance in the transport numbers at different phases of time and year. The monthly fixed effects also return statistically significant results across the different models in most months, from February to October. The yearly fixed effects were also statistically significant in most years from 2004 to 2015. These results provide evidence that there are seasonal and time fluctuations accounted for by the monthly and yearly fixed effects, adding explanatory power to the regression. The Fare Increase variable was also statistically significant with a positive coefficient of 0.106. This could be explained by the fact that the bus is already one of the most cost effective means of travel, and that the cost increase would not have a significant effect on pushing people towards alternatives. Going along with this line of logic, the Unemployment variable may not have been significant because of the inelasticity

Fall 2017 of transit to changes in the macroeconomic environment. Given that buses operate in more residential areas, they might be the best form of transportation that some residents have. Hence, even if they do not have a job to travel to, they will still depend on this mode of transportation for their daily tasks and routines. Model Chosen: Constant Model, Model 3 from Table 5. Robustness Checks As with other empirical papers, robustness checks are done to ensure the structural validity and feasibility of the coefficients for the regressors. First, this paper replicates the constant model and events model adapted from Currie & Phung (2007). These models do not involve the complete set of fixed effects controls that we eventually use, but provide us with the structure and some intuition on how the regression results will turn out eventually. The results were reported in the first two columns of each type of transportation mode, under the replication models. In investigating the model of elasticity, Currie and Phung (2007) also discovered that demand elasticities may not be constant over time, when they compared a world events model which took into account significant events such as Hurricane Katrina and 9/11 and a standard constant elasticity model. Including these interaction terms, the world events model returned a higher adjusted R2 than in a constant elasticity model. Their proposed model from past research was the Events elasticity model. Based on this paperâ&#x20AC;&#x2122;s analysis, it seemed that a Constant model was most suitable due to the following interpretations. Based on standard economic reasoning, elasticity, which is a behavioral response to change in prices, is affected by several factors. First, the type of product investigated, transportation, is a daily necessity. There are a limited range of substitutes available for transit. This favors the constant elasticity model, given that a personâ&#x20AC;&#x2122;s preferences should not change significantly within such a short span of time. Additionally, the changes in prices for public transportation usually constitute small proportions of a personâ&#x20AC;&#x2122;s income for most segments of the population. It also coincides with the rather low values of elasticity derived in the regressions for each transport mode. Based on BLS Statistics of the Consumer Price Index (CPI), the changes in proportion of transportation from 2012-2013 and 2013-2014 on average were only 0.1%

and 0.8%, respectively. Therefore, this paper has decided to follow the analysis through with the Constant Elasticity and Behavioral Lag model, which both assume a constant elasticity during the period of investigation. Subsequently, for the chosen model for each transportation type (Behavioral Lag & Constant), there were 4 separate regressions run, with the final one being the actual regression used in the result analysis, involving all of the fixed effects variables. Using the equations displayed below, we start by running regressions with fewercontrol variables to see how different the relationship is in magnitude, and logic test checks whether the estimated coefficients were consistent with expectations. The equations are listed as below: 1.

This is the most basic model that only includes Log Gas Price as regressors. 2.

This is a basic model that only includes the monthly fixed effects as control variables. 3.

This is the model without the Yearly fixed effects added. 4.

This is the complete model with all the fixed effects added. Columbia Economics Review

33 BART In all of the regressions, the coefficient of Log Gas Lag was statistically significant. For the most basic regression, the coefficient on Log Gas Lag was 0.278, with an adjusted R-squared value of 0.571. As we add more control variables, the explanatory power of the regression increases, as the adjusted R-squared value increases to 0.768, 0.905 and 0.945 in the final regression. The decrease in coefficient value also showed that we accounted for potential Omitted Variable Bias inherent in the first regression equation. MUNI Light Rail The coefficients on Log Gas variable was significant in the first, third and fourth regressions. The sign however was inconsistent, although the most basic regression also returned a positive coefficient of 0.0429. The adjusted R-squared value increased from the original 0.021 to 0.242, 0.387 and 0.663 for the final model equation. This shows that each time we add control variables to the regression there is increased explanatory power in the regression. The increase in coefficient value while maintaining the positive sign also showed that we accounted for potential Omitted Variable Bias inherent in the first regression equation. also showed that we reduced the Omitted Variable Bias inherent in the first regression equation. MUNI Bus The coefficients on Log Gas was only statistically significant in the final model, although the most basic regression also returned a positive coefficient of 0.0144. The adjusted R-squared value also increased from the original 0.034 to 0.511, 0.522 and to 0.686 in the final model equation. This shows that each time we add control variables to the regression there is increased explanatory power in the regression. The increase in coefficient value while maintaining the positive sign also showed that we accounted for potential Omitted Variable Bias inherent in the first regression equation. Newey West Standard Errors Given that this paper looks at a timeseries data where there is potential for autocorrelation or cross-autocorrelation between the adjacent time periods, the regressions with the models that are best selected are re-run on STATA using the Newey West estimator to take into account that time lag. The results are consistent with the mod-

34 el chosen for each transportation mode. For the BART, the Log Gas Lag variable is still consistent, after taking into account a possible 6 month autocorrelation, which is consistent with the model of the Lagged Gas Price of 6 months. For the MUNI, the Log Gas variable is only significant up onto a point of 1 month lag, which is consistent with that of the Constant Model. For the MUNI Bus, the Log Gas variable is significant in both the 1 month and 6 months lagged variant, which is consistent in the fact that the MUNI Bus has both the Constant and Lagged Model displaying significant results in the section above. The Correlations of Highway Speeds and Transit Ridership In an attempt to check the validity and provide extra evidence for the relationship between gasoline prices and public transit ridership, this extension was constrcuted to investigate the direct effects driving has on public transportation. Would the relationship be a direct substitution effect, or something more complicated? By analyzing the average driving speed on selected freeways around the Bay Area at their peak usage time (57pm), the correlation between the modes of transportation is studied. In order to control for the quality of the analysis, data points with quality below 50% were removed from the analysis. Although this reduced the size of the sample that was being investigated, it maintains the accuracy of the analysis because of the huge amount of data that is being aggregated. The multivariate regressions from Sections 4 and 5, with the appropriate model chosen for each mode of transport, was used for this analysis. The dependent variable was substituted for the respective highway speeds. CA-24E The studied segment of CA-24E connects San Francisco to the Inner East Bay cities towards Walnut Creek, Lafayette and Orinda. There are more cars owned on average in these cities that lie further away from San Francisco. In these cities, there is a higher proportion of drivers to work. Amongst the three transport means, only the variable for the Log Gas Lag of the BART was significant. This is reasonable as there is a BART line that operates along this stretch, unlike the MUNI and MUNI Bus that operates within the San Francisco City itself. However, the negative sign of the Log Gas Lag variable does not initially match the expec-

Fall 2017 tations, as the anticipated effect would be the opposite, as presumably people would switch to the BART when gas prices become more expensive. However, it seems that peopleâ&#x20AC;&#x2122;s choices are quite inelastic here, and perhaps they are hoping that others would make the switch due to other reasons such as convenience and comfort etc. Another possibility could also be that there is indirect substitution, like the switch from driving individually to carpooling or ride-share services like Uber and Lyft, or other public transportation means such as buses. In general, the adjusted R-squared values for each of the three transportation modes are high, and have sufficient explanatory power. Also, the significance on month variables like January, April and May also suggest that there is some time-seasonality involved on highway speeds on the CA-24E.

Interstate 80W The studied segment of the I-80W runs in the Alameda County portion of the freeway, where the BART Service also runs in the East Bay towards Berkeley and Richmond. This can be considered the alternate route taken by drivers who otherwise utilize the BART service. An important point to note would be that the road lanes in this segment of the freeway are constantly congested during the peak hours of usage, and hence that might affect the coefficients obtained because it may be hard to differentiate the speed change as caused by the difference by consumer behavior. Based on the regressions, surprisingly, only the MUNI and MUNI Bus models seem to be affected despite this route being a corridor of the BART. A possible exColumbia Economics Review

planation could be that it seems that peopleâ&#x20AC;&#x2122;s choices are quite inelastic here, and that there is indirect substitution, such as the switch from driving individually to carpooling or ride-share services like Uber and Lyft, or other public transportation means such as buses such as the AC Transit which offers a Transbay Bus service into the city. The MUNI Upgrade variable also returned a significant coefficient. The positive value means that the MUNI Light Rail upgrade is correlated with an increase in the highway speed, which means that presumably more people are substituting away for public transportation within the City. In general, the adjusted R-squared values for each of the three transportation modes are high, and have sufficient explanatory power. Also, the significance on all month variables except July and the years 2011-2013 also suggest that there is some time-seasonality involved on highway speeds on the I80W.

US-101 The US-101N portion of interest was considered a control segment of the highway because none of the major public transit modes studied reaches to Marin County, towards the North of San Francisco. Hence, presumably, the consumer behavior towards driving, which would be a main form of transportation into the city, would likely be more inelastic, and this is demonstrated by the low correlation coefficients to Gas. However, because there is a significant overlap in route near to Downtown where the BART operates, there is an expecta-

Fall 2017 tion that perhaps the BART effect might also be captured. Indeed, as expected, all the Log Gas variables are positive in nature, which represents that a higher gas price causes a reduction in the driving, hence resulting in the higher speeds on the freeway observed. In particular, the coefficients on MUNI and MUNI Bus are strongly positive, and depict this effect. The MUNI Upgrade also is strongly significant to have helped divert some of the driving away. Presumably, one living in the North would have taken the ferry or bus services before making a transfer to reach their intended destination. In general, the adjusted R-squared values for each of the three transportation modes are high, and have sufficient explanatory power. Also, the significance on most month variables from January to May and the years 2005/2006 also suggest that there is some time-seasonality involved on highway speeds on the US-101.

Takeaways from this analysis Apart from the inherent problems of the data quality, the infrastructure present could also be a reason why we were not able to test the hypothesis properly. The highway infrastructure systems in Northern California are nearing maximum capacity, and if they do reach a state like those in Southern California, minor differences as measured by our regressions will be hard to pick up through driving speeds. It could also be that there is not a simple substitution relationship we can look into between public transit and private driving. With the advent of new forms of transportation, such as Uber and Lyft, and the increasing platforms for carpooling and incentives from companies en-

couraging their employees to carpool, the substitution equation is not as discrete as initially perceived. Although we were not able to draw much insight from the I-80 and CA-24E segments, the US-101 provided a good control that indeed shows the substitutability effect, confirming the results of the previous segments, leaving us on a good concluding note for future research. Further Research Opportunities There are many potential areas of extensions to the paper to allow the analysis to be more comprehensive. Firstly, in this paper, a short term elasticity was estimated, based on how ridership figures respond to gasoline prices in the relatively short run. However, papers such as Tsai et.al (2014) have argued that the Long Run and Short Run elasticities are different, and that a trend in gasoline prices may change travel pattern and gasoline consumption pattern such as converting to electric or hybrid cars etc. This would be an interesting supplement to this research. Secondly, this paper does not consider the network effects generated by the various modes of transportation in totality. For instance, a person might choose to take public transport if the modes of transfer between the BART and MUNI were seamless and hassle free. Hence, papers such as Sorensen & Longva (2011) will be useful references to consider on expansion between the coordination effects between transport modes that will refine the estimates in this paper. Finally, to scope down on the specifics of the data, it would be useful to better understand actual consumer behavior. For instance, the effects of ridership can be separated to monthly pass users and single trip users. Given that the monthly pass offers unlimited access to the mode of public transportation over the specified duration, the relative transit demand would be rather inelastic, given the sunk cost of purchasing the monthly pass, whereas the single trip user will be more likely to consider factors like gasoline prices to decide finally on what mode of transport to utilize. With a finer dataset to work with, this will be helpful in these aspects of expansions. As mentioned in the previous section, the investigation can be expanded to include more recent developments in policy and business through the rideshare applications and carpooling policies and measures. CONCLUSION This paper has examined the relationship between gasoline prices and pubColumbia Economics Review

35 lic transit ridership in the San Francisco Bay Area using time series data analysis. Numerous past studies have suggested the relationships between the two, although research for the San Francisco locality and the time- relevant time frame were not present. The significance of this study was to provide an updated cross demand elasticity figure of gasoline on public transportation specific to the San Francisco Bay Area. Following the intuition given from the average cross elasticity values derived from past studies like Currie and Phung (2007) of 0.12, the range of values arrived at for this paper range within 0.0581 to 0.147, which is similar to other study values. A trend noticed from this study that was different from past studies was the difference in the order of magnitude in the classification of elasticity values. Past studies have determined a greater sensitivity of elasticity for Heavy Rail compared to Light Rail and Bus, but in the present study, a reversal of this trend was observed. As mentioned in the analysis, it is posited that the relative inelasticity arising from the difference in service choice and preference as the main explanation for this phenomenon. Another key insight justified in this study was the usage of a Constant elasticity model that gives the most explanatory power at least when applied to the San Francisco locality. This is in contrast to the conclusions derived from past studies which used a Events based elasticity model in determining the relationships between gasoline price and transit ridership. Furthermore, cross checking with the highway speed data around East Bay and San Francisco, we can also preliminarily see the effects of the substitution between public transportation and private driving through the changing gasoline prices that also has interesting implications to policy decision-making. It is acknowledged that there are limitations to the data, given the scope of 164 observations and the usage of monthly Unlinked Passenger Trips. Using additional data points from PEMS to analyze the driving speeds, and the revenue hours analysis, we now have a clearer interpretation of the transport ridership figures. It is hoped that these results apply to the San Francisco locality and can possibly serve in ways to the Transportation departments and service operators for more efficient resource allocation to further improve the service standards and quality of public transit. n

Fall 2017

Penny for Your Shots The Underpaid Superstar: The Max Contract’s Effect on Parity within the 2015-2016 NBA Microeconomy Dylan Robbins-Kelly University of California, Santa Barbara Robbins-Kelley’s piece details the effects of max contracts in the NBA, where superstar players’ salaries are capped on a perseason basis according to league rules. His study goes beyond the economic inefficiency caused by players signing below their free market price and investigates whether the max contract provides unfair advantages for teams recruiting top talent. With a thorough literature review and rigorous statistical analysis backing his study, Robbins-Kelley suggests that the max contract eliminates teams’ ability to recruit players via monetary incentives, and results in polarization of the league, whereby winning teams continue to attract superstar talent by virtue of their “stacked” rosters. Given the importance of sports in promoting the economic health of US cities, Robbins-Kelley’s analysis is particularly relevant. He concludes that the NBA should abandon the max contract to increase team parity in the league. - R.R. INTRODUCTION The differences in economies across sports leagues are striking. For example, Major League Baseball (MLB) operates as a relatively free market. Team payrolls across the league in 2015 ranged from $60 million to upwards of $300 million (Spotrac.com). Unlike the MLB, the NBA caps both the overall payroll of a team (a salary cap), as well as the amount that any one player can receive in a given season (a max contract that varies depending on tenure). Entering the 2015-2016 season, only 11 NBA players made $20 million or more per year — a stark contrast to the 28 MLB players that did so as of March 2015 (Badenhausen 2015). This difference can most likely be attributed to the presence of a max contract in the NBA. Even if the salary cap were kept at $70 million, its level in 2015, it is not hard to imagine that in a world without max contracts the best NBA players might make upwards of $30 million a season (NBA.com). This being the case, it is fair to say that the best players in the NBA, the “superstars,” are likely acquired below market value under the current collective bargaining agree-

ment (CBA). Thus, teams that are able to land “superstars” at a max contract may have a competitive advantage over those teams that cannot. There has been research on the excess value given to teams by the max contract, most notably by statistician Nate Silver (Paine 2015). However, this has not yet been expanded to address the impact on competitive balance. My hypothesis is that the presence of a max contract detracts from parity in the NBA, as it gives an unfair advantage to those teams that are able to sign a player for less money than he would theoretically go for in a free market. By maintaining a salary cap whilst doing away with the max contract, I believe that a greater level of competitive balance would be attainable. LITERATURE REVIEW Parity in sports is, at its heart, an economic issue. Sports franchises exist to create revenue and, as Berri, Schmidt, and Brooke (2004) suggest, wins are the greatest factor in increasing a team’s revenue within the NBA (Berri, Schmidt, & Brook 2004). Increased competitive balance not

Columbia Economics Review

only means a more equal spread of revenue across the league, but also likely provides an overall increase in profits as previously uninterested markets are tapped. Ideal deviation of winning percentages in the NBA, the deviation that should result if each team within the league has a 50% chance at victory each game, can be calculated as follows: σ = 0.5/ √82 where 82 is the number of games played by a team in a single NBA season (Zimbalist 2002). We will use this to define true parity. An empirical definition of perfect competitive balance is imperative in our attempt to determine the magnitude of the max contract’s effect on parity within the current NBA economy. Max contracts were introduced to the NBA in 1999, after a labor lockout forced a new CBA between the NBA and the NBA Players Association. As Hill and Groothuis (2001) point out, this new CBA focused on moving money from the highest paid players towards those in the middle, the “median voters”. This move-

Fall 2017

Columbia Economics Review

38 ment resulted in greater income equality, as the players in the top 40% of salaries lost earning potential to the remaining 60% (Hill & Groothuis 2001). However, promoting income equality likely came with a tax on competitive balance. Max contracts also affect competitive balance in less apparent ways. Sanderson and Siegfried (2003) note that max contracts drive a player to prioritize choosing a team with which he can win rather than choosing the team that is willing to offer him the most money. Because the contract that teams can offer a “max player” is capped, such a player will likely receive the same/similar offer(s) from a handful of teams. Thus, his decision is no longer strictly monetary: the quality of players that are already on a team becomes one of the most influential factors in making his decision. This results in a further decrease in parity. Adding this to the increased inequality created by undervaluing “superstars” through the max contract, as the talent-rich teams continue to get richer.

“[The 1999 CBA] focused on moving money from the highest paid players towards those in the middle.” Although it is clear that the max contract has some sort of effect on parity within the NBA, there has yet to be an empirical estimate of this effect. In this paper, we will provide such an estimate, and deem whether the effect significantly hinders the competitive balance of the league. THEORY In order to measure the max contract’s effect on parity, we must first understand the general microeconomy of the NBA. If we accept that a given player would rather play in the NBA than not, regardless of the contract offered to him, we can say that the supply of this player’s talents is highly inelastic (for simplicity we will assume perfect inelasticity at 1 unit). Clearly, this is a broad generalization, but it is likely applicable to the vast majority of NBA athletes. Even the minimum salary, a little over $500,000 in 2015-16, is more than enough to justify playing a

Fall 2017 game for a living (to most). Next, if we assume that a team would buy multiple players of identical talent at a given price level we can say that demand is perfectly elastic. If we replace individual players with the wins that these players provide, it is not unreasonable to believe that demand should be perfectly elastic. If a team can acquire wins below market value they will theoretically do so infinitely, just as they should not be willing to pay more than market value for a win. Hence, when a price ceiling (the max contract) is enforced on a hypothetical “superstar’s” contract, we should see a market imbalance. We will assume that in a free market this superstar would be worth $30 million per year based on his performance in previous seasons. Due to the max contract of $22.97 million imposed upon players with 10+ years of experience, our hypothetical “superstar” is unable to earn his free market salary of $30 million. This inherently creates imbalance in the market, as the player is losing out on $7.03 million in salary. In turn, the team that signs this player has consumer surplus equal to the difference between the equilibrium salary and the max contract, again $7.03 million. We will use this team surplus that price ceilings create to measure the max contract’s effect on competitive balance in the NBA. METHODOLOGY In order to calculate consumer surplus, we will need an approximation of any given player’s value. We want to use a single statistic, Value Over Replacement Player (VORP), to summarize a player’s relative talent level. VORP is a measure of the value that a player provides to his team when compared to that which would be provided by a minimum contract/non-rotation player, that is, “a replacement player” (Myers). Since NBA contracts are not given out retroactively, it would be remiss of us to simply look at a player’s VORP when determining how much he should have been paid in a season. Rather, we will need to predict VORP going into a season to evaluate how much money a player should be making. Thus, we will create a predictive regression model built from several statistics and measures from previous seasons to predict a player’s VORP prior to a given season (See Table B for a comprehensive list of covariates). Ultimately, we will use this model to predict VORP in the 2015-2016 season. To create our predictive model, we will implement a linear regression on the dependent variable of VORP in the 2014Columbia Economics Review

“Because the contract that teams can offer a ‘max player’ is capped, [...] the quality of players that are already on a team becomes one of the most influential factors in making his 2015 season, based on a host of covariates (See Table B). We will use adjusted R-squared, and mean squared prediction error (MSPE) as our criteria for model selection. MSPE is a measure of the error in the predictions provided by a model as opposed to the true values in a dataset. Thus, a lower MSPE is preferable. To measure MSPE, we will split our data into two equal subsets. We will train a linear model on one subset, using the other to test our model’s predictive power by calculating MSPE based on the predictions that our trained model renders on the test set. Then, we can merge the two subsets together again and recreate what we have determined to be our ideal model, allowing said model to be as informed as possible. The next step is to determine the price that a player would go for on an open market with no price ceiling. First, we need to translate our predicted VORP values to Wins Above Replacement (WAR), which we do by translating a player’s point differential into wins (Myers):

“Since NBA contracts are not given out retroactively [...] we will need to predict VORP going into a season to evaluate how much money a player should be making.”

Fall 2017 Table A

WAR = 2.4 * VORP. WAR is a rough estimate of the additional team wins that a player accounts for beyong the previously defined “replacement player”. Therefore, if we multiply a player’s WAR prediction by the approximate value of a win, we will get an estimate of the contract that said player should receive in competitive equilibrium. Nate Silver has developed the following process through which one can estimate the value of a win in a given NBA season. To calculate the market value of a win, we take the total payroll across the NBA in the 2015-16 season and divide by the number of teams to get the average payroll per team. We then subtract $12 million from this value, as we can assume that a team will need on average approximately $1 million per player on average just to field a replacement level 12-man roster. This leaves us with the average payroll above the minimum that teams currently spend to acquiring talent, also known as Average Payroll Over Replacement (APOR). Silver explains that the Pythagorean win expectation model predicts that an average team would win 41 games, while a replacement-level team would win 16 games. Thus, we divide the APOR by the difference between these two values, as spending the average payroll over replacement should theoretically move a team from replacement level to average. This leaves us with an approximate value of a win in the 2015-16 NBA of $2,627,275.87. Now that we have the pieces with which to estimate free market contracts, we calculate consumer surplus. Since we are primarily interested in the effect of the max contract, we will calculate the surplus achieved on signings of only the

39 Figure B

players currently making “max money.” Grouping “max players” by team will allow us to sum the surplus value in dollars that each team is receiving due to the max contract. Using the previously calculated value of a win, we can translate this dollar value back into wins, yielding the number of surplus wins that each team receives from the max contract. The final step is to definitively determine the effect of the max contract on parity within the NBA. First we will use

“If even one team significantly benefits from the max contract, we must view this benefit as a negative effect on parity, since there is ultimately one champion in an NBA season.” an F-test on equality of variance to see if the distribution of wins across all thirty NBA teams is significantly affected by the surplus wins that the max contract introduces. If this does not prove to be significant, we can alternatively perform Z-tests on the difference in winning percentages of specific teams with and without the surplus wins provided by their max contract players. If even one team significantly benefits from the max contract, we must view this benefit as a negative

Columbia Economics Review

effect on parity, since there is ultimately one champion in a NBA season. If any of these tests prove statistically significant, we can reasonably conclude that the max contract is, in fact, altering the competitive balance of the NBA. DATA Player age, contract, draft spot, minutes played, performance related statistics, and position data from 1995-2016 were extracted from BasketballReference.com. A table of performance-based statistics extracted is provided in Table A. Data was organized in both MySQL and Microsoft Excel. Analysis was performed in R. It seems intuitive to believe that those players who perform well in a given season will continue to do so the next. Examination of Figure C confirms this belief. There is a very clear positive correlation between VORP in 2014 and 2015, which leads us to believe that VORP in the previous season will likely be a key regressor in our predictive model. Similarly, we observed the distribution of VORP across age groups. Although it seems plausible that there is a peak age, or “prime”, at which an athlete’s physical and mental attributes reach an optimal combination for performance, the data suggests otherwise. VORP rises until age 26, where it seems to level off. This could be attributed to older players retiring before their skills decline, leaving behind only those older players that defy the normal aging curve in the NBA, which would skew the outcome of VORP. RESULTS After trimming and validating a model with 2014-2015 VORP as the dependent variable we settled upon a linear regres-

Fall 2017

40 sion of the form depicted in Table C. Notably, age was dropped from our model. The bias in the age curve noted in the data section likely contributed to age’s insignificance. Cross validation was implemented to test the predictive power of potential models. Models that did not include draft position yielded slightly lower MSPE values. However, because we are focused on predicting the value of the best players, and the top 4 draft pick factor is significant to our model, we decided to retain draft position as an independent variable. The lowest MSPE was obtained relying only on VORP Year – 1 and VORP Year – 2 as independent variables. However, when fitting a model that looked only at stats from Year – 1 we saw that VORP, PER, TS%, 3PAr, FTr, OWS, and draft position were all retained using stepwise model selection. Thus, we decided that they should be retained in our final model. It should be noted that USG% and TOV% were also significant when restricting to one year prior, how-

“[T]wo out of the three teams receiving statistically significant gains from the max contract (the Cavaliers and the Thunder) reached their respective conference’s finals in 2015-16” ever their significance dropped substantially in a more complex model so they were dropped. Once our final model was created, we plugged in the data for each player in the NBA at the beginning of the 2015-2016 in order to predict their performance. As the focus of our research is the surplus value created by players on max con-

Table C: Baseline 2nd Round Draft Pick (Picks 31-60)

tracts, we centered our analysis on only those players that are both receiving their maximum contract AND are predicted to outperform this contract. Results for these players are shown in Table F, separated by team. Given these results, we can now perform hypothesis tests to see whether the max contract truly influences parity in the NBA. Our initial test uses an F-statistic to determine whether or not the difference in winning percentage variance between a league in ideal parity and a league with maximum contracts is significant (See page 3 for the standard deviation of winning percentage under true parity). To find the variance in winning percentage in a league with maximum contracts (assuming parity is otherwise held constant), we add the additional variance in winning percentage that the max contract introduces to the NBA to the variance under true parity. This test has a null hypothesis that the two variances are equivalent, while the alternative is that the variance in a league with max contracts is Table D

Columbia Economics Review

Fall 2017

higher (29, 29 are the degrees of freedom; the number of teams in the league, 30, minus 1 is 29). We conclude that we are 80% confident that the league with a max contract has a higher variance. This is not a statistically significant result, though more testing is certainly warranted. We now use Z tests to determine whether any one team is benefitting from their max contract related gains in winning percentage. The null hypothesis for these tests is that there is no significant difference in the winning percentage of a team that is getting excess value from their max contract players and a team that is not. The alternative hypothesis is that the surplus wins created by the max contract significantly increase a teamâ&#x20AC;&#x2122;s winning percentage. Thus, these are onesided Z tests. To find our Z statistics we divide by the deviation of winning percentages in a league with max contracts that is otherwise held to parity, calculated the same way as in our F test. We

find that the Cleveland Cavaliers and the Houston Rockets show a slight statistically significant increase in their winning percentages because of the max contract (alpha=.10), while the Oklahoma City Thunder show a significant increase (alpha=.05). These results support our hypothesis that max contracts have a negative effect on parity within the NBA, as these effects only benefit a few teams. CONCLUSION It is impossible to ever obtain true parity within the NBA, not to mention any other sports association or league. There will always be players on contracts that prove to be worth more than they are paid, just as there will be players that underperform, resulting in negative value for their respective team. Athletes do not perform at a constant level, so the fact that they are often signed to multi-year contracts means that there is inherently going to be variation in the value that Columbia Economics Review

each team is receiving as opposed to the amount of money they are spending. This being said, the fact that the max contract itself is directly responsible for raising the winning percentages of some teams in the NBA raises questions regarding its practicality. It is telling that two out of the three teams receiving statistically significant gains from the max contract (the Cavaliers and the Thunder) reached their respective conferenceâ&#x20AC;&#x2122;s finals in 201516, with the Cavaliers eventually being crowned the NBA champions. An attempt to quantify the max contractâ&#x20AC;&#x2122;s effect on parity resulting from its effect on the salaries of non-max players could theoretically be of further interest. Quantifying such effects would be tough due to difficulty in determining causation, though this certainly presents an interesting endeavor. Nonetheless, from where we stand now it is apparent that the max contract does result in superstars that are being paid below market value.n

Fall 2017

The Default in Our Stars The Evolution of Collective Action Clauses and Their Impact on Recent Latin American Debt Crises Isabella Santandreu Columbia University Santandreu examines the introduction of Collective Action Clauses (CACs) in 2003 on the sovereign bond market and its effects on debt restructuring. Santandreu suggests that CACs reduce risk of interference by holdouts, thereby simplifying the process of bond restructuring in a manner which accommodates the debtor. Her analysis not only covers the immediate economic and legal impacts directly following the emergence of CACs, but also considers their future ramifications in subsequent events of sovereign default. In light of Venezuela’s proposal for a $60 billion bond restructure in early November, Santandreu addresses real and pressing issues within the international bond market. -E.S.L.

During the 1980s, nearly every country in Latin America was plagued with external debt crises that led to the restructuring of hundreds of billions of dollars within a few years. This instability and uncertainty continued into the 1990s, when Mexico’s near-default in 1994 “solidified public consensus that the era of bond crises had arrived, and was worse than the 1980s loan crisis” (Gelpern and Gulati 19). Further complicating the situation was the fact that sovereign bonds issued by emerging markets were considered to be “technically difficult to restructure” (Gelpern and Gulati 19), so options were fairly limited for debtors who defaulted. In the following pages, a brief background on the history of Collective Action Clauses will be followed by an analysis of their impact on recent Latin American debt crises. The focus will be on Argentina’s 2002 sovereign debt default and lengthy restructuring, Ecuador’s 2008 default due to “immoral and illegitimate” bonds, and Venezuela’s rapidly-evolving crisis. Following the Mexican Peso Crisis of

1994, international agencies and central banks began considering ways to alleviate a country’s burden following events of default. The consensus was that “distressed debtors need a fresh start, not

“[Complex restructuing processes] further intensified recessions and destroyed even the slightest possibilty of economic recovery.” just temporary assistance” (Guzman and Stiglitz 4). Complex restructuring processes hindered both creditors and debtors. Additionally, in a large majority of cases, insufficient funds triggered soverColumbia Economics Review

eign defaults. Thus, continuing to pressure an already weak economy was, quite simply, counterproductive; these restructurings further intensified recessions and destroyed even the slightest possibility of economic recovery. Before 2003, the typical procedure for carrying out sovereign debt restructurings (SDRs) under New York law consisted of “a decentralized market-based process in which the debtor [engaged] in intricate and complicated negotiations with many creditors with different interests” (Guzman and Stiglitz 4). Restructurings were often unsatisfactory for both parties––the processes took too long and yielded too little. SDRs also allowed for the creation of vulture funds. These aptlynamed funds procured huge returns for their firms by buying severely distressed or defaulted debt in the secondary market and litigating until they received full payment. As these funds were buying these bonds for cents on the dollar, potential returns were extremely rewarding. In order to capture these potential returns, vulture funds became part of a group called the

Fall 2017 “holdouts,” that in an event of default, rejected any propositions for restructurings from the sovereign until they were paid the face value of their bonds plus any accumulated interest. Holdouts could be catastrophic to sovereigns since that they had the “ability to interfere with the sovereign’s deal with other creditors” (Carletti et al. 9). Additionally, given the risk that holdouts could grab, “a disproportionate share of the pie made all creditors reluctant to enter an exchange” (Carletti et al. 9). The interference of holdouts in SDRs was possible due to the lack of Collective Action Clauses (CACs). These clauses “allow a majority of bondholders to agree to changes in bond terms that are legally binding to all bondholders, including those who vote against the restructuring” (Guzman and Stiglitz 11). Thus, whether or not a bond is issued with CACs has a defining impact on the SDR process. According to the IMF, as of June 2014 roughly 75% of the US$420 billion New

York law-governed bonds outstanding included CACs (Hagan et. al). This percentage has presumably increased in the

“Whether or not a bond is issued with CACs has a defining impact on the SDR process.” past three years when older bonds without CACs have matured and all recentlyissued bonds include these clauses. Collective Action Clauses were first proposed following Mexico’s tesobono crisis in 1994-5, which prompted the IMF and the United States to put together a $50bn loan package to prevent a default. Gelpern and Gulati note that “if Mexicostyle bailouts were no more, bond re-

Columbia Economics Review

43 structuring was inevitable” (22). In order to facilitate the debt restructuring process, central bankers of the G10 began the drafting of CACs in 1996 with the initial purpose of eliminating huge bailouts (Herman et al. 23). The take-up, however, wasn’t exactly seamless. Even though six years earlier, a group of “less prominent issuers” from Europe and the Middle East began to employ CACs in their bonds, the market impact was low given that these bonds were not SEC registered and were issued in the European market (Gelpern and Gulati 7). CACs were seen as alternatives to the Sovereign Debt Restructuring Mechanism (SDRM), which aimed to address creditor coordination issues. Interestingly enough, Mexico, the same country that inspired their creation, was the first to employ CACs in a 2003 $1bn issuance. The bond’s indenture stated that the amendment threshold was set at 75%, meaning that in the case of a restructuring, any amendment of financial terms agreed upon by holders of 75%

44 of outstanding principal would apply to 100% of creditors. Mexico was effectively able to turn “CACs into a signal of strength and [quiet] market fears that sovereigns would more readily default on their bonds” by demonstrating that their inclusion had no negative effects on their borrowing terms (Herman et al. 24). Mexico’s action had an unprecedented impact. Following their issuance, CACs became a “boilerplate” term in these complex financial contracts (Scigliuzzo). Unfortunately, the first use of CACs in Mexico’s $1bn issuance was not implemented until 2003, and Argentina was facing financial troubles in 2002, the preCAC era. What unfolded was a case of restructuring gone wrong. The now infamous incident started with the default of roughly $100bn in foreign bonds (Gelpern and Gulati 11). In 2005, following sev-

“Christina Kirchner, then-President of Argentina, refused to compromise with the holdouts, thus leaving Argentina in perpetuating default until 2016...” eral years of negotiations with the IMF and bondholders, Argentina performed a bond swap that was accepted by holders of 76% of the outstanding principal of its bonds (Herman et al.). By 2010, 93% of the defaulted debt had been swapped. These bondholders were the victims of a massive haircut, receiving a mere 25 cents on the dollar. The remaining 7%––the holdouts––consisted of 4 funds, led by NML Capital (a subsidiary of the hedge fund Elliott Management). These funds benefited from the fact that Argentina’s bonds included a pari passu clause but no CACs. Pari passu requires that all bondholders to be treated evenly among equally ranked investors. In order to exploit the simultaneous inclusion of pari passu and absence of CACs, the vultures took to the courts of New York, demanding that the pari passu clause be recognized. A US federal judge, Thomas Griesa, ruled in favor of the funds, preventing Argentina from paying interest on the newly-renegoti-

Fall 2017 ated swapped bonds until an agreement with the holdouts was reached and they received payment as well. Cristina Kirchner, then-President of Argentina, refused to compromise with the holdouts, thus leaving Argentina in perpetuating default until 2016, when President Mauricio Macri finally settled the issue, a full 15 years after the default had begun. Eventually, the holdouts suffered a haircut of 25%, a much more acceptable proposition than that of the 93% group. Argentina’s exclusion from the international capital markets exacerbated the default event. The bondholders who agreed to a bond swap suffered as well, as the holdouts’ actions prevented them from collecting interest and principal payments on their restructured bonds (Macfarlane). Had CACs been a part of these bonds’ indentures, the ordeal would have ended in 2005, when the bondholders of 76% of the outstanding principal agreed to a bond swap. Their decision would have applied to the remaining 24% of bonds, regardless of the remaining bondholders’ wishes. The lack of CACs led scholars such as Herman et. al to conclude that Argentina’s debt workout was neither “efficient [nor] timely” (Herman et al. 16). Ecuador, often described as a serial defaulter, defaulted on its debts an outstanding eight times before it finally repaid a debt obligation, the Global 2015s, in full in 2015. In late 2008, despite the fact that the country had the necessary foreign reserves to pay the interests and principal outstanding, Ecuador stopped payment on $3.2bn of debt, its second default in less than a decade. Since Ecuador’s defaulted bonds, the Global 2012s and 2030s, did not include CACs, the country was forced to find a solution that did not entail a restructuring of the terms of the bonds. Given what had happened in Argentina just a few years earlier, a restructuring did not seem like a plausible option. President Rafael Correa of Ecuador knew that a financial crisis meant Ecuador’s creditors would be under pressure to liquidate assets, and, therefore, would be unlikely to put up a nasty fight with the sovereign during the process. In order to avoid a restructuring, Ecuador concocted a strategy that worked effectively due to the global financial crisis looming over investors in 2008. The first step relied on the Banco del Pacifico, a large local bank, buying defaulted bonds in the open market at around 20 cents on the dollar. Ecuador opted to pay its creditors cash instead of offering a bond swap, which “avoided [the need] Columbia Economics Review

to go through a lengthy registration process” (“Ecuador’s Winning Strategy”). Finally, Ecuador instituted a Dutch auction, where bondholders offered a price at auction instead of a “take-it-or-leave-it transaction” (“Ecuador’s Winning Strategy”). This move was immensely successful, allowing Ecuador to buy back over 91% of its defaulted foreign bonds, repre-

“Ecuador, often described as a serial defaulter, defaulted on its debts an outstanding eight times before it finally repaid a debt obligation...” senting a third of Ecuador’s total foreign debt, at a meager 35 cents on the dollar (“Ecuador’s Winning Strategy”). Best of all, “despite the huge discount, [the repurchase] was relatively well received by the market” (“Ecuador’s Winning Strategy”), and Ecuador executed “one of the most elegant restructurings [ever] seen” (Salmon). One wonders, however, if the same route would have been taken had Ecuador been supported by CACs. The last country in this analysis, Venezuela, has struggled with its mounting debt burden for the past several years. A dreadful combination of historically low oil prices and an unsuitable government has led to much speculation about when a default will occur. Even though Venezuela has not triggered an event of default, it is included in this analysis because it has outstanding bonds both with and without CACs, allowing for a natural experiment to test how investors price the addition of CACs to bond contracts. The breakdown of Venezuela’s debt is as follows: two series of Venezuela’s sovereign debt––the 9.25% September 2027s and the 13.625% August 18s––are free of CACs because they were issued before 2003, when CACs became “boilerplate.” Venezuela also holds debt with CACs at both 85% and 75% thresholds (Scigliuzzo). The December 2018s and January 2034s, issued during the early years of CACs, 2003 and 2004, fall in the 85% threshold category. Those with a 75% threshold were issued after 2004. It follows that “the 100% bonds are going to

Fall 2017

be easier to hold out on than the 85% ones which, in turn, are going to be easier to hold out on than the 75% ones” (Carletti et al. 6). As Argentina and Greece illustrated, bonds which are easier to hold out on tend to have a higher payout (Carletti et al. 6). The maturity dates for Venezuela’s 9.25% September 2027s and 9.25% May

“[E]xisting literature has mostly studied CAC provisions in economically-sound countries - those with virtually no possibility of default....” 2028s are just eight months apart with the same exact coupon rate. What differentiates them is that the 2027s are unanimity bonds (no CACs), and the 2028s have CACs with a 75% threshold (Carletti et al. 10). Logically, then, the 2027s must be the most valuable to creditors given their large issue size, strong liquidity, and lack of CACs. The latter reason also makes 2027s easy prey for hedge funds, who “could look to borrow a page from Argentina’s playbook and try…to block a potential restructuring and sue for full repayment” (Scigliuzzo). As evidenced by the Bloomberg graph above, in the past five years the 2028s have traded at a higher yield than the

2027s, labeling the 2028s as a riskier asset, presumably due to the risk of a lower payout in the case of a restructuring. The average yield differential between the two bonds has grown as the risk of default has increased. From 2013-2014, for example, the spread between the 2027s and 2028s was only 35 points. From 20142015, the spread grew to 52 points. This trend continues throughout, with the spread totaling 161 basis points in the past year. Given that the average yield for the 2027s over the past year is 21.51%, the 161 basis point spread represents the 7% premium that investors are willing to pay for a CAC-less bond. However, this pricing premium was valued at only 3% between 2013 and 2014. These pricing premiums are currently extremely undervalued. Taking Argentina’s experience as a basis, the fact that the holdouts were able to receive 75 cents on the dollar is an impressive deal compared to the bondholders who accepted the 2005 restructuring, receiving only 25 cents on the dollar. Thus, a 7% premium is a huge bargain considering that the holder can possibly recover roughly 75% of face value in case of a default. The market has, in general, treated the 75% and 85% threshold bonds “quite similarly,” but it is hard to make such an accurate comparison with the 75% and 100% threshold bonds because there are no two Venezuelan bonds with coupons and maturities that are similar enough (Carletti et al. 10). Carletti et al. note that prior literature on CAC provisions “has largely found either no pricing difference or a pricing premium for bonds with lower vote thresholds” (Carletti et al. 15). This discrepancy arises from the fact that existing literature has mostly studied CAC provisions in economically-sound countries Columbia Economics Review

45 – those with virtually no possibility of default. Here, investors tend to be large institutions and unsophisticated retail investors seeking safe investments, but in the event of a default these investors tend to be replaced by sophisticated vulture funds that know exactly how to “parse contracts seeking an advantage for their investors” (Carletti et al. 15). In these cases, as the Venezuela 27s and 28s exemplify, there is a significant spread between bonds which require a unanimous vote and those with lower vote thresholds. The introduction of CACs as a contract term for New York law-governed bonds has had a tremendous impact on recent debt crises in Latin America. In Argentina’s case, where the sovereign default predates the incorporation of CACs, the restructuring process was a long and expensive one for all involved––much longer than it would have been if CACs had prevented holdouts. In Ecuador’s 2008 case, the lack of CACs led to the sovereign’s devising of a clever strategy that avoided an actual restructuring of the

“The introduction of CACs as a contract term for New York law-governed bonds has had a tremendous impact on recent debt crises in Latin America.” bonds. This was possible given that the country had not triggered a default event due to lack of funds, but rather because of President Correa’s belief that the bonds were “immoral.” While this strategy worked to avoid a restructuring process and, hence, any holdouts, it is one that would have traditionally been unavailable to a bankrupt borrower. Ecuador was able to have a “seamless” default even without the presence of CACs only because it had defaulted due to unconventional reasons not related to the country’s lack of reserves. Finally, Venezuela’s case provides us with a case study that illustrates the value added of CACs for a country that is on the brink of defaulting. Despite the positive effect that CACs have had on the financial world, continued reform is necessary. n

Fall 2017

A City without Uber? Roxy Farhad Columbia University Whereas New Yorkers might find Uber to be slightly preferable to traditional yellow cabs, the app-based service had tremendously eat into the black cab market in Greater London – that is up until its recent ban by city authorities on the basis of security concerns. With millions of Britons still reeling from this shock reversal, Roxy incisively explores the economic implications of inhibiting a previously disruptive element in a major market such as for-hire ground transportation. Drawing on Schumpeter’s theory of creative destruction, her article points to the fact that Uber’s tremendous performance in the London scene more so than anything revealed an unacceptable lack of efficiency in traditional London taxi services. Roxy thus argues that such a realization should undercut any sizable return in clientele to the formerly dominant city cabs. If anything, this market crisis further reveals the inadequacies of a heavily regulated industry, one which seems to only hurt itself by opposing dynamism and innovation in its midst. – M.S.

Uber’s infiltration of London in July 2012 marked the turning point at which the city’s vast and complex Transport system irreversibly began to disrupt. The app revolutionized London’s bridge between driver and consumer, through its low prices and ease of tapping for a cab from your smartphone. Since its entry into London, Uber has over 40,000 drivers and over 3.5 million users. Considering the city’s population is just below but close to 9 million, Uber’s entry can ultimately be described as a takeover. Yet it seems that Uber’s conquest has reached a standstill. On September 22nd, Uber was stripped of the possibility of renewing its operating license in London by the Transport for London Network. This was an utter shock to the city, with a petition to oppose the ban created and signed by over 800,000 people in the following days. While there have been huge pressures from the pre-Uber era taxi network to put greater regulations on Uber resulting from their capture of the taxi market, TFL has announced that the ban is actually a result of Uber’s disregard for safety. TFL declared that it believes that there is an absence of corporate responsibility around issues of security and potential public safety. These issues

have included Uber’s approach to reporting serious criminal and notably sexual offences and the way that the company obtains medical certificates and security checks for their drivers. However, the way that the Uber ban has generated such hostility among the London population generates an interesting tension: are we willing to sacrifice our safety for cheaper prices? Uber represents monumental progress not only for innovation, but also

“[T]he way that the Uber ban has generated such hostility among the London population generates an interesting tension: are we willing to sacrifice our safety for cheaper prices? Columbia Economics Review

for the gig economy. But that revolutionary quality of the app does not make it insusceptible from the law. In fact, the imminent and striding approach of the gig and technological economy makes it all the more necessary that government regulators take a tough approach from the onset, setting a precedent for future innovations to also abide to the law. Uber’s entry into London completely reformed the London taxi network, making it almost inconceivable to imagine a London without it. While in New York City Uber typically charges its customers the same prices as its yellow-cab counterpart, London reflects a totally different situation. Research into the price differences between Uber and London’s Black cabs (the equivalent of NYC’s yellow cabs) has discovered that Uber is on average 30% cheaper than the latter. A 13-minute trip in a London Black cab would cost roughly £9 while the same Uber trip can cost as little as £5.70, depending on the hour and price surges. Uber ultimately freed the population from the belief that private cars are exclusive to the rich. Moreover, this huge undercutting of prices may have flourished the market, but it also caused the slow disappearance of many traditional

Fall 2017

London has among the worst traffic in the world: TomTom rates it 25th, the highest in Western Europe, and higher than any American city other than Los Angeles (which ranks 12th). To combat this, London enacted a congestion charge in central districts; according to the Financial Times, average speeds within that district dipped last year for the first time below pre-charge levels, hitting an average of 8.3 mph in 2015. Private cars have been almost entirely priced off the roads--but with the rise of Uber, the number of licensed private cars has risen by 3 times to 81,710 in 2016. What remains to be seen, though, is the effect of the new ban on traffic levels in Central London.

minicab firms in the city. Statistics show that enrolment for Knowledge point, London’s college for acquiring the Black Cab taxi license, dropped from 7000 in 2012 to 4000 in 2015. While there has not been published statistics for years since, the sharp declining trend gives to the belief that numbers continued to fall. An Uber ban would cause a huge vacuum in the market, raising an infinite number of questions about consumer substitution of Uber, the unemployment within the taxi market, but also the efficiency of the taxi market. The presence of Uber can be connected to the Schumpeter’s famous theory of the creative force of destruction. This theory refers to the incessant product and process innovation mechanism by which new production units replace outdated ones. In a way, Uber is the embodiment of the new production unit, which has been created through transformative information technology. In this way, the presence of Uber is meant to force other taxi companies to innovate and adapt to new

The presence of Uber can be connected to the Schumpeter’s famous theory of the creative force of destruction [... forcing] other taxi companies to innovate and adapt to new consumer wants to survive consumer wants to survive and those incapable of doing so will eventually be eliminated. If Uber is creating greater ef-

Columbia Economics Review

ficiency within the taxi market, through increasing consumer surplus, is the Uber ban justified? The answer to this question lies on the value of our safety and the way in which we treat corporations that break the rules that government set out. While the ban may cause inconvenience for consumers, letting Uber run free sets a precedent for other progressive corporations to do the same, potentially generating much greater havoc for consumers and society. But the static nature of a “ban” also cannot be overlooked. If a corporation changes, learning from their mistakes, can we continue to exclude them from the market? Today’s world and systems, whether economic or not, have been created through innovation. Thus, a complete shutting of doors to progressive corporations, such as Uber, goes against the values that characterizes our society. Yes, Uber has much to learn and change but that does not necessarily mean total elimination from the market. n

CER Journal & Content O n li n e a t

columbiaeconreview.com Read, share, and discuss Keep up to date with web-exclusive content

Columbia Economics | Program for Economic Research Printed with generous support from the Columbia University Program for Economic Research