Volume 15, Issue 2 Âˇ Fall 2017
Editor in Chief Payce Madden Managing Editor Emily Rosen Acquisitions Manager Anna Vande Velde
Editors Chris Bell Nicole Gans Tobie Irvine Sara Jackson J. Alexander Killion Jennie Luptak Shivarthn Maniam Mickey McGlasson Sarah Pesi Gilbert Resendez Robert Santamaria Owen Stevenson Yiwen Wu Victoria Zuber
Climate Equity and Justice in the Post-Paris Agreement World: A New Financial Paradigm
Gender and STEM Education in Japan and the United States
Nilanjan Raghunath, Aye Myat Khine Win, and Fumiko Nishino
Analysis of Intergenerational Income Mobility for Counties Within the U.S. Using Machine Learning
Ada Tso, Gursmeep Hundal, and Vicky Mei
Big Data on a Big New Market: Insights from Suppliers and Customers in Washington Stateâ&#x20AC;&#x2122;s Legal Cannabis Market Jonathan P. Caulkins, Yilun Bao, Imane Fahli, Yutian Guo, Krista Kinnard, Mary Najewicz, and Lauren Renaud
The Heinz Journal is a student-run publication of the H. John Heinz III College at Carnegie Mellon University, dedicated to publishing works that link critical and theoretical analysis with policy implementation.
The Heinz Journal 4800 Forbes Avenue Pittsburgh, PA 15213 journal.heinz.cmu.edu
We accept submissions from professionals, policy school graduate students, and members of the Pittsburgh community. Submissions and inquiries can be emailed to TheHeinzJournal@gmail.com.
The Heinz Journal
Climate Equity and Justice in the Post-Paris Agreement World: A New Financial Paradigm Niles Guo Carnegie Mellon University – Department of Engineering and Public Policy The Paris Agreement signaled a move away from a top-down approach to climate mitigation negotiations, to a bottom-up pledge-and-review framework, where each nation submits Individual National Determined Contributions (INDCs) for their emissions targets. This approach removes the discussion of equity as one of the long-standing barriers to international climate cooperation, which many observers argued was key for success of the Paris Agreement. However, this paper will show that burden-sharing justice will remain an important factor, with the results of Paris only signaling a shift of this discussion from mitigation efforts (specifically emission reduction targets) to the global climate finance structure related to adaptation and mitigation efforts. Finally, this paper identifies additional areas of research that include the importance of understanding the market forces driving pledges and disbursements for global climate finance.
I. Introduction Signed in December 2015, the Paris Agreement was a significant milestone in the international community’s response to the dangers of climate change. Some commentators, including Savaresi, have argued that the agreement was “arguably the best that could be achieved at this time and place.”1 For the first time, representatives from all 196 countries2 have agreed and pledged towards a common goal of “holding the increase in the global average temperature to well-below 2° C above preindustrialized levels and pursuing efforts to limit the temperature increase to 1.5° C above preindustrial levels.”3 Despite its significance, the Paris Agreement mostly skirted around the issue of fairness and equity. At their cores, the definitions of fairness and 1
Savaresi Annalisa, "The Paris Agreement: An Early Assessment," Environmental Policy and Law 46, no. 1 (2016): 14. 2 While Syria initially failed to sign the Paris Agree in 2015, it has since signed in November 2017. On the other hand, while the Trump administration announced its intention to withdraw the United States from the accord, it is still currently a signatory to the accord. 3 “United Nations Framework Convention on Climate Change (UNFCCC),” (United Nations: Adoption of the Paris Agreement, Paris, 2015), Article 2.1(a).
equity would define the responsibilities and obligations for each nation state in their response to climate change. A fair and equitable response would see each nation do its “fair share” in the global response to this crisis. But while the final agreement reiterated the principle of “equity and common differentiated responsibilities and capabilities,” and mentioned the “right to development” in the preamble, it did not provide any clarity on the principles of equity and fairness.4 Instead, it used the tactic pioneered during COP15 in Copenhagen, and called for individual nations to submit voluntary emission reduction targets under a bottom-up and pledge-and-review approach.5 Commentators such as Hall believed this helped to bridge the traditional divide between the developing and the developed world,6 and was a
UNFCCC. Paris Agreement, 2. Jennifer Jacquet and Dale Jamieson, “Soft but Significant Power in the Paris Agreement,” Natural Climate Change 6, no. 7 (2016): 644. 6 The use of developing and developed nations as labels is based on the definition offered by UNFCCC for Annex I and Non-Annex I countries. While many have suggested these labels (created in 1992) are no longer accurate, given the growth of Non-Annex I countries such as South Korea and Singapore, I have decided to use them as a matter of convention. 5
Climate Equity and Justice in the Post-Paris Agreement World necessary step to secure the success of the Paris negotiations.7 As such, the Paris Agreement did not contain any new obligations by any actor on the question of distributive justice, its definition, or implementation.8 Does this mean that the question of fairness and equity in burden sharing is no longer valid in the new world where bottom-up, pledgeand-review, voluntary action is the de-facto framework for climate negotiations? This paper will show the answer is a resounding no, and while the bottom-up approach has proved to be more politically acceptable, fairness and equity will continue to have an important role in future climate discussions. Discussions on climate equity have traditionally focused on climate mitigation efforts, and there is already a wealth of research and analysis conducted in this area. Grosso (2007), Baer and Athanasiou (2008), Jamieson (2005), and Shukla (2005) have all conducted normative analysis on the various equity principles, ranging from basic needs principle, to per-capita emissions principle, to historical responsibility.9 Lange and Vogt (2003), Lange, Vogt, and Ziegler (2007), and Lange et al. (2009) provided econometric analysis of the effectiveness of each equity principle, and found Jan Erik Hall, “Paris Agreement on Climate Change: A Diplomatic Triumph–How Can It Succeed?” New Global Studies 10, no. 2 (2016): 176. 8 Annalisa, “The Paris Agreement,” 16. 9 Marco Grasso, “A Normative Ethical Framework in Climate Change,” Climatic Change 81, no. 3-4 (2007): 223-246; Paul Baer and Tom Athanasiou, “Frameworks & Proposals: A Brief, Adequacy and Equity-Based Evaluation of Some Prominent Climate Policy Frameworks and Proposals,” Global Issue Papers, no. 30 (2007); Dale Jamieson, “Adaptation, Mitigation, and Justice,” Advances in the Economics of Environmental Resources 5 (2005): 217–248; Priyadarshi R. Shukla, “Aligning Justice and Efficiency in the Global Climate Change Regime,” Perspectives on Climate Change: Science, Economics, Politics, Ethics 5: 121-145. 10 Andreas Lange, Carsten Vogt and Andreas Ziegler, “On the Importance of Equity in International Climate Policy: An Empirical Analysis,” Energy Economics 29, no.3 (2007): 545562; Andreas Lange and Carsten Vogt, “Cooperation 7
that a combination of “polluter-pays” and “egalitarian rule” can form the foundation of future climate negotiations.10 On the other hand, relatively less attention was paid to equity principles relating to global climate finances. Grosso (2010) first developed a justice framework for international climate adaption funding (focusing on inclusion, specification, and commitment), and then applied them to both Kyoto mechanisms (AF, SCCF etc.) and post-Kyoto proposals such as the World Climate Change Fund (WCCF).11 Grosso (2011) researched the role of justice in the case for the Adaptation Fund, and found that despite not being explicitly mentioned, distributive justice played a large role in Adaptation Fund (AF) negotiations.12 Ciplet, Roberts, and Khan (2013) used Amartya Sen’s “realizationfocused comparison” theory of justice and applied it to adaptation finance.13 They found adaptation finance efforts have not met the basic equity criteria (where funding is not directed towards areas most vulnerable to climate change), with those most responsible for climate change and capabilities falling short on their obligations.14 Finally, Stadelmann et. al. (2013) examined both equity and effectiveness principles for adaptation project funding and found there is not necessarily a tradeoff between equity and effectiveness, and only a in International Environmental Negotiations Due to a Preference for Equity,” Journal of public Economics 87, no.9 (2003): 2049-2067; Andreas Lange, Andreas Löschel, Carsten Vogt and Andreas Ziegler, “On the Self-interested Use of Equity in International Climate Negotiations,” European Economic Review 54, no. 3 (2010): 359–375. 11 Marco Grasso, “An Ethical Approach to Climate Adaptation Finance,” Global Environmental Change 20, no. 1 (2010): 74-81. 12 Marco Grasso, “The Role of Justice in the North– South Conflict in Climate Change: the Case of Negotiations on the Adaptation Fund,” International Environmental Agreements: Politics, Law and Economics 11, no. 4 (2011): 374-375. 13 David Ciplet, Timmons Roberts and Mizan Khan, “The Politics of International Climate Adaptation Funding: Justice and Divisions in the Greenhouse,” Global Environmental Politics 13, no. 1 (2013): 4968. 14 Ciplet, Roberts and Khan, “The Politics of International Climate Adaptation Funding,” 64-65.
The Heinz Journal “pure economic indicator for effectiveness tends to be in contradiction with major equity principles.”15 This paper will also show that the discussion around burden-sharing justice will shift away from mitigation efforts (specifically emission reduction targets), to a global climate finance structure relating to adaptation and mitigation efforts. Finally, the paper will list some of the normative barriers that will need to be resolved to satisfy the equity and fairness considerations in the global climate finance.
II. Top-Down Versus Bottom-Up: A Brief History of Equity and Fairness in Climate Negotiations Prior to the Copenhagen climate talks in COP15, international climate change negotiations concentrated on reaching a binding consensus on emissions reduction targets for the entire community.16 Drawing on the success of previous international environmental negotiations such as the Montreal Protocol, this strategy had five major principles: 1. Top down approach based on generally understood principles; 2. Attempts to develop targets and climate governance policy instruments in a comprehensive manner; 3. Applied universally, with an agreed-upon principle of burden sharing; 4. Negotiated within the United Nations framework; 5. Seeks legally binding agreement.17 15
Martin Stadelmann, Åsa Persson, Izabela Ratajczak-Juszko and Axel Michaelowa, “Equity and Cost-effectiveness of Multilateral Adaptation Finance: Are They Friends or Foes?” International Environmental Agreements: Politics, Law and Economics, 14, no. 2 (2014): 117. 16 For a comprehensive history of climate change negotiations, see Joyeeta Gupta’s The History of Global Climate Governance. 17 Robert Falkner, Hannes Stephan and John Vogler, “International Climate Policy After Copenhagen: Towards a ‘Building Blocks’ Approach,” Global Policy 1, no. 3 (2010): 253.
Despite numerous perceived benefits of this strategy,18 negotiations were consistently bogged down by different interpretations and perspectives on climate justice and equity principles.19 Article 3 of the United Nations Framework Convention on Climate Change specifically stated the principle of a solution founded “… on the basis of equity and in accordance with their common but differentiated responsibilities and respective capabilities.”20 Despite this common principle, nations have placed different interpretations on the issues of fairness and equity, leading to numerous disagreements around the appropriate equity principle that has proved to be costly during climate negotiations. For developed nations, such as the United States, their longstanding climate negotiation policy requires commitment from all “major emitters” nations, which now consists of nations with rapidly growing economies such as China.21 From their perspective of efficiency, any principle that required developed nations to “take the lead” in addressing climate change implies that developing nations do not need to do their share.22 As related by Pickering, Vanderheiden, and Miller, during the 2011 UN Climate Change Conference held in Durban, South Africa, the lead U.S. envoy stated that, “if equity is in, we are out.”23 This fear of “free riding” behavior by developing nations was evident throughout Cancun, Copenhagen, and Durban.24 On the other hand, for many developing countries, it is almost impossible to separate the issue of climate change emissions reduction from economic development, especially since emissions mitigation
For a list of benefits to this global deal strategy, see Falkner, Stephan, and Vogler’s International Climate Policy After Copenhagen in Global Policy. 19 Hall, “Paris Agreement on Climate Change,” 177. 20 United Nations Framework Convention on Climate Change, (United Nations, New York City, 1992), Article 3. 21 Jonathan Pickering, Steve Vanderheiden and Seumas Miller, “‘If Equity's In, We're Out’: Scope for Fairness in the Next Global Climate Agreement," Ethics & International Affairs 26, no.4 (2012): 424. 22 Pickering, Vanderheiden and Miller. “If Equity’s In, We’re Out,” 424. 23 Ibid., 423. 24 Ibid., 435.
Climate Equity and Justice in the Post-Paris Agreement World efforts are seen as barriers to economic growth.25 It is therefore with some justification that they believe an approach to global mitigation policy, based on consideration of efficiency rather than equity, would wipe away much of the historical contribution of the developed countries, while hampering their own economic development.26 They also argued that since it is predominantly the developing countries who are the most impacted by effects of climate change caused by the developed countries, they should bear a heavier burden.27 Finally, many public policy theorists believe that there is an opportunity for developing countries to “leapfrog” conventional carbon-intensive technologies directly to carbon neutral solutions as part of their economic development.28 However, without assistance from developed nations, it is difficult for developing nations to make that “leap,” especially when considering the high capital costs associated with those types of projects.29 Therefore, many leaders argue that it is the responsibility of those who gained the benefits of carbon emissions in the past that should bear the burden of the cost. This clash of ideology has contributed greatly to the inability of the world community to develop a coherent strategy on climate change, from the failure of the Kyoto Protocol to the lackluster results of the Copenhagen and Durban Summits, and can be considered as one of the principal impediments to arriving at more substantial outcomes.30 The push for a universally agreed-upon Priyadarshi Shukla, “Aligning Justice and Efficiency in the Global Climate Change Regime: A Developing Country Perspective,” in Perspectives on Climate Change: Science, Economics, Politics, Ethics, eds. Walter Sinnott-Armstrong and Richard B. Howarth (Bingley, United Kingdom: Emerald Group Publishing Limited, 2005), 133. 26 China Daily, “China Open to Talks on Binding Emission Cut,” May 12, 2011, http://usa.chinadaily.com.cn/china/201112/05/content_14213729.htm[Accessed 20 02 2016]. 27 Helmut Weidner, “Global Equity versus Public Interest? The Case of Climate Change Policy in Germany,” (WZB Discussion Paper SP IV 2005-102, 2005), 4. 28 Michael Hasper, "Green Technology in Developing Countries: Creating Accessibility through a Global Exchange Forum," Duke Law & Technology Review 7, no. 1 (2009): 3-4. 25
principle of burden sharing and legally binding agreement have proved to be insurmountable barriers, with countries displaying self-serving bias that prevented them to come to an agreement.31 Compounding this issue is Kyoto’s division of countries into the “developed” (Annex I), “relatively rich Annex I countries” (Annex II), and the “developing” (non-Annex I) countries with different responsibilities for each category.32 While the initial distinction was justifiable and logical (as developed nations had polluted more in their drive to economic development, and developing nations still needed room to grow), this division has reinforced the North-South division in the climate negotiations.33 Emerging economies in the nonAnnex I group did not want to stifle their future growth, while developed nations who struggled to fulfill previous commitments were reluctant to put themselves under any further rounds of rigid targets.34 Finally, despite being tasked to manage the global collective response to this issue, the United Nations Framework Convention on Climate Change (UNFCCC) lacks any form of institutional authority. Unlike domestic states, the UNFCCC has “no coercive powers at the international level” to determine how nations work together to tackle this problem.35 This made any attempt to reach a legally binding commitment to reduce greenhouse gas (GHG) emissions extremely difficult, especially when such a commitment would be against the self29
Ibid., 4. Steve Vanderheiden, “Justice in the Greenhouse: Climate Change and the Idea of Fairness.” Social Philosophy Today 19 (2004): 92-93. 31 Kjell Arne Brekke and Olof Johansson-Stenman, “The Behavioural Economics of Climate Change,” Oxford Review of Economic Policy 24, no. 2 (2008): 288. 32 Gupta, The History of Global Climate Governance, 149-150. 33 Ibid., 147, 171. 34 Robert Falkner, “The Paris Agreement and the New Logic of International Climate Politics,” International Affairs 92, no. 5 (2016): 1111. 35 Zou Ji and Fu Sha, “The Challenges of the PostCOP21 Regime: Interpreting CBDR in the INDC Context,” International Environmental Agreements: Politics, Law and Economics 15, no. 4 (2015): 422. 30
The Heinz Journal interest of the powerful members of the international community.36 In contrast to the Montreal Protocol on Ozone Depletion, where the cost of action was relatively small and the benefits were large and focused, climate change action requires high mitigation costs and its effects are unevenly distributed across the world.37 While many environmental commentators viewed the COP15 summit in Copenhagen as a failure, it did introduce a new paradigm in international climate negotiations. After two weeks of fruitless negotiations, the United States and China entered into bilateral negotiations, and hammered out a deal with support from the BASIC nations (Brazil, South Africa, India, and China).38 The final Copenhagen Accord moved away from the rigid distinctions set by Kyoto, where each country’s responsibility is based on their annex status.39 It also sidestepped the need to have agreed-upon and binding emissions reduction targets, and instead had voluntary pledges as basis for future climate action.40 The top-down approach favored by the international community since the Montreal Protocol was finally abandoned in Copenhagen. This trend continued in subsequent international climate negotiations. Utilizing a bottom-up and pledge-and-review approach in Paris allowed for each nation to submit their Individual Nationally Determined Contributions (INDC) that outlined their post 2020 climate pledges.41 Many commentators (See Falkner 2016, Savaresi 2016, and Hall 2016) have pointed to submission of 186 countries’ INDC prior to the convention as evidence that a bottom-up and voluntary approach has proved to be a more successful method than previous climate negotiation attempts. Rather than having progress blocked by historical equity arguments, parties can move forward in areas where there is consensus. Furthermore, using voluntary
INDCs, all nations have pledged to act against climate change. Emerging economies with tremendous growth no longer can hide behind the “common but differentiated responsibility (CBDR)” clause first set back in 1992.42 This finally met one of the most important demands of the U.S., and helped to foster the final agreement. As Savaresi wrote, when compared against previous attempts such as the Kyoto Accord, “the Paris Agreement is better aligned with current emission patterns and more realistically equipped to frame future action.”43
III. Fairness and Equity: Obsolete Concepts in This New World? This evolution in climate negotiation begs the question: are concepts such as fairness and equity in burden-sharing still valid in the new world where bottom-up, pledge-and-review, voluntary action is the de-facto framework for climate negotiations? After all, as described earlier, the previous climate negotiation failures can be partially attributed to the difficulty of achieving consensus in a topic as heated as climate equity, where parties could reject proposals simply due to the terms being used.44 As the lead U.S. envoy to the Durban Conference later recounted, It’s not that there’s anything wrong with… talking about equity in the context of climate negotiations, and the term appears in the framework convention, and we tend to look at the phrase as calling for fairness to all parties, and we think that’s fine. But in this context, when we’re talking about setting up a negotiation… the key element of which for us was to include all the major players in the same legal system kind of together, we just thought that that would be a distraction that would tend to drive people back into the old
Ibid. Lucas Bretschger, “Climate Policy and Equity Principles: Fair Burden Sharing in a Dynamic World,” Environment and Development Economics 18, no. 5 (2013): 517-518. 38 Falkner, Stephan, and Vogler, “International Climate Policy After Copenhagen,” 257. 39 Falkner, “The Paris Agreement and the New Logic of International Climate Politics,” 1111.
Ibid. Jacquet and Jamieson, “Soft but Significant Power in the Paris Agreement,” 645. 42 Andrew Hurrell and Sandeep Sengupta, “Emerging Powers, North–South Relations and Global Climate Politics,” International Affairs 88, no. 3 (2012): 465. 43 Savaresi, “The Paris Agreement,” 17. 44 Pickering, Vanderheiden, and Miller, “If Equity's In, We're Out,” 432-433.
Climate Equity and Justice in the Post-Paris Agreement World paradigm, if you will, and we didn’t want to go there.45 Indeed, one can argue that this is another example of what Charles Lindblom described as the “incrementalism” of policy making. In situations where parties cannot agree on basic values or even objectives, Lindblom argues that the true test of a “good” policy is “the possibility of the policy itself, which remains possible even when agreement on values is not.”46 The fact that an agreement can be reached in the climate change space means it is a good policy, even if there is no agreement on the core issues of equity and fairness. Progress therefore can be made without the need to resolve the underlying differences in values between each state actor. However, as Rajamani wrote, the question of equity remains the normative core of the international climate regime.47 If this core is removed completely from the conversation, nations’ self-interests will take over, and the future of an effective multilateral effort will be called into question. As Rajamani continued, At the heart of differential treatment lies the notion that equity or fairness dictates special or preferential treatment for certain countries, whether on the grounds of their differing capacities or their lesser contributions to the global environmental problem. It is only when countries believe themselves to be treated fairly that they will first participate in the regime, and next consider ways of enhancing their own ambition in relation to its central goals. For developing countries with serious energy poverty and developmental challenges, a climate regime built around symmetry will impose severe limitations on their ability to lift their people out of poverty or provide universal access to energy.
Ibid., 433. Charles E. Lindblom, “The Science of “Muddling” Through,” Public Administration Review 9, no. 2 (1959): 83. 47 Lavanya Rajamani, “The Changing Fortunes of Differential Treatment in the Evolution of International Environmental Law,” International Affairs 88, no. 3 (2012): 623. 46
In fact, the very foundation of participatory international climate change negotiations was based on the principles of equity and fairness. Principles 3 and 7 of the Rio Declaration stated that: Principle 3: The right to development must be fulfilled so as to equitably meet developmental and environmental needs of present and future generations. Principle 7: States shall cooperate in a spirit of global partnership to conserve, protect and restore the health and integrity of the Earth's ecosystem. In view of the different contributions to global environmental degradation, States have common but differentiated responsibilities. The developed countries acknowledge the responsibility that they bear in the international pursuit to sustainable development in view of the pressures their societies place on the global environment and of the technologies and financial resources they command.48 This principle of “common but differentiated responsibilities” was recognized as a key component to the participation of developing countries. It offered them the ability to participate in international environmental institutions, while protecting themselves against onerous obligations that they could not meet with their limited resources and capabilities.49 This principle also represented an acknowledgement of the historical and moral responsibility of the developed countries in climate change.50 By moving away from this principle, and limiting the differentiation between developed and developing nations, it can actually have the negative and unintended effect that developing nations will limit their ambition and effort to reach the goals set forth by the Paris Agreement, or bind themselves upon its requirements.51
United Nations Environment Programme, “Rio Declaration on Environment and Development,” 1992, principle 3, principle 7. Accessed from: http://www.unep.org/documents.multilingual/default. asp?documentid=78&articleid=1163. 49 Rajamani, “The Changing Fortunes,” 613. 50 Ibid., 616. 51 Ibid., 623. 48
The Heinz Journal This is already evident based on the INDCs submitted by each nation. Numerous studies demonstrate the current pledges are not enough to meet the 2° C goal stated under the Paris Agreement. For example, UNFCCC’s October 2015 Synthesis Report on the Aggregate Effect of the Intended Nationally Determined Contributions showed that aggregate GHG emission level from the unconditional INDC pledges are predicted to be 19% higher (8.7 Gigatonnes (Gt)) than the 2 °C scenario in 2025, and 35% higher (15.1 Gt) in 2030.52 Rogelj et al. also showed that the unconditional INDC pledges will be 14 Gt short in 2030 compared to the 2° C scenario.53 Regardless of the figure cited, the current INDC pledges are insufficient to meet the Paris Agreement goals, and significant gaps still exist in order to reach a 2° C scenario. It is clear then that despite a politically acceptable agreement, this will not, in its current form, reach the goal of the ultimate policy question, and Lindblom’s incrementalism approach to policy making is ultimately insufficient in this case.
IV. Climate Equity’s New Arena: Global Climate Finance While this paper has argued for the importance of retaining equity and fairness as the normative core of climate change negotiations, the equity differentiation between nations for mitigation purposes is no longer politically feasible.54 Instead, the demand for greater financial support in most INDCs suggests that the focus of future climate negotiation will migrate around the global climate finance efforts. 52
United Nations Framework Convention on Climate Change, 2015. Synthesis report on the aggregate effect of the intended nationally determined contributions, Paris: United Nations. 10. 53 Rogelj, Joeri, Michel Den Elzen, Niklas Höhne, Taryn Fransen, Hanna Femkete, Harald Winkler, Roberto Schaeffer, Fu Sha, Keywan Riahi, and Malte Meinshausen. "Paris Agreement climate proposals need a boost to keep warming well below 2 C." Nature 534, no. 7609 (2016): 632. 54 Rajamani, “The Changing Fortunes,” 623. 55 Wen Zhang, and Xun Pan. "Study on the demand of climate finance for developing countries based on submitted INDC," Advances in Climate Change Research 7, no. 1 (2016): 101. 56 Ibid., 101.
A recent study by Zhang and Pan showed that out of the 160 INDCs received by the UNFCCC in December 2015, 122 (76.3%) explicitly contained financial content.55 A further 64 INDCs contained explicit and quantitative financial estimates to achieve their targets.56 When tallied together, the total financial demand by these 64 nations will reach USD $4.6 trillion, with India, Iran, and South Africa ranking as the top three countries in demand.57 It is clear that the success of each INDC pledge is dependent on each country’s ability to fund its mitigation and adaptation efforts. Currently, global climate finance under the UNFCCC auspices is primarily administered through the Green Climate Fund (GCF), the Special Climate Change Fund (SCCF), the Least Developed Countries Fund (LDCF), and the Adaptation Fund (AF).58 Three of these, the SCCF, LDCF, and AF, are administered under the UNFCCC and established through the Kyoto Accord, while the GCF was established under the Cancun Agreements in 2010.59 Each of these funding mechanisms also targets different parties, with different funding and evaluation criteria: Special Climate Change Fund (SCCF): The SCCF finances projects relating to adaptation such as technology transfer and capacity building, especially relating to industries such as energy and agriculture.60 Least Developed Countries Fund (LDCF): The LDCF supports Least Developed Countries to
Ibid.. United Nations Framework Convention on Climate Change, Biennial Assessment and Overview of Climate Finance Flows, 2016. Accessed November 30, 2016. https://www.unfccc.int/cooperation_and_support/fina ncial_mechanism/items/2807.php. 59 Kanoko Morita and Ken’ichi Matsumoto. "Financing adaptation to climate change in developing countries." Handbook of climate change adaptation (2015): 984. 60 United Nations Framework Convention on Climate Change (2001). Decision 7/CP.7. “Funding under the Convention.” Accessed from: https://unfccc.int/files/cooperation_and_support/ldc/a pplication/pdf/13a01p43.pdf. 58
Climate Equity and Justice in the Post-Paris Agreement World “carry out preparation and implementation of national adaptation programs”.61 Adaptation Fund (AF): The AF finances specific adaptation programs and projects in developing countries, especially those who are vulnerable to the adverse effects of climate change.62 AF is financed through the sale of Clean Development Mechanisms (CDM), where if a developed country implements an emissions reduction project in a developing country, that will earn CDM credits and can be used to offset its own emissions against the Kyoto goals.63 Green Climate Fund (GCF): The GCF was created to support both the adaptation and mitigation efforts, striving for a 50:50 split between adaptation and mitigation. Furthermore, the AF also commits at least 50% of funding going to the most vulnerable nations, including Least Developed Countries, Small Island Developing States (SIDS), and other African states.64 While an initial 10 billion USD was pledged by developed nations, capitalization efforts for this fund have only recently started.65 Accounting for climate change financial flows is extremely complex and difficult, with no agreed definition of climate finance by individual actors and inconsistent reporting leading to uncertainty in the final data.66 Despite these difficulties, the UNFCCC Biennial Assessment and Overview of Climate Finance Flows in 2014 estimated the global climate flow ranged from 340 billion USD to 650 billion USD per year, with only 40 to 175 billion USD transferred from developed countries to developing countries.67 This includes the 0.6 billion USD administered through the UNFCCC administered funds (SCCF, LDCF, AF), which is a 61
United Nations Framework Convention on Climate Change (2001). Decision 7/CP.7. 62 Morita and Matsumoto, “Financing Adaptation to Climate Change in Developing Countries,” 986. ` 63 Ibid. 64 Green Climate Fund. Global Context – GCF Features. Accessed November 30, 2016. http://www.greenclimate.fund/about-gcf/globalcontext#history. 65 Jacquet and Jamieson, “Soft but significant,” 645. 66 United Nations Framework Convention on Climate Change, UNFCCC Standing Committee on Finance:
tiny portion of the overall financial resources used for climate purposes.68 When compared against the self-assessed financial needs indicated through the INDCs, there is an order of magnitude difference between the global climate financial flows and the financial demand. Paragraph 54 of the Paris Agreement pledged that the developed nations will “mobilize a floor of 100 billion USD by 2025 for meaningful mitigation and adaptation purposes,” and while this is significant, it will still not be enough to close the existing gap.69 Instead, this substantial shortfall of financial resources to meet the INDC pledges will likely result in future recriminations, as countries are struggling to fulfill their INDC goals. This competition for the limited available financial resources will then eventually raise questions on how to divide up the global pie, and who is responsible for increasing the financial supply. It is not farfetched, and in fact it is expected to see equity and fairness once again providing the normative criterion for this discussion, where developing nations (including G77 + China) will utilize the CBDR principle to argue for greater financial support, and developed nations such as U.S. and the E.U. argue for efficiency to determine distribution of the resources.
V. Challenges to the Equity Discussion in Global Climate Finance To understand what challenges equity discussions will have in international climate finance negotiations, a basic justice framework first needs to be built. Academic literature has established that the use of a procedural justice framework is appropriate when discussing climate change
2014 Biennial Assessment and Overview of Climate Finance Flows, Bonn 2014. Accessed from: http://unfccc.int/files/cooperation_and_support/financ ial_mechanism/standing_committee/application/pdf/2 014_biennial_assessment_and_overview_of_climate _finance_flows_report_web.pdf: 5-6. 67 UNFCCC, “UNFCCC Standing Committee on Finance,” 6. 68 Ibid., 46-47. 69 United Nations, Framework Convention on Climate Change. Adoption of the Paris Agreement 2015, Paris: 54.
The Heinz Journal issues.70 Grosso in fact argued that the decisionmaking process must be just to acquire any institutional legitimacy.71 Therefore, instead of focusing on what the result is, procedural justice aims to define the principles needed to have a just procedure, where if followed, the outcome will likely be correct and just.72 John Rawls’ Theory of Justice is the most wellknown example of a pure procedural justice framework. In his 1971 Theory of Justice, starting from what Rawls termed as an original position (where he utilized a veil of ignorance to argue for what is a just society), he postulated a just society will follow the following two principles of justice: 1. Everyone has an equal right to basic liberties if it is compatible with the liberties of others.73 2. Unequal access to these basic liberties is permitted if and only if it a) “offers the greatest benefit for the most disadvantaged, and b) attached to offices and positions open to all under conditions of fair equal opportunity.”74 Rawls argues that if these two principles are considered when considering the climate burdensharing issue, ultimately the outcome will likely be fair and just. Despite criticisms from Amartya Sen in his work The Idea of Justice,75 this is a useful theoretical starting point to apply ideas of justice and fairness to the global climate finance, especially in the UNFCCC rounds. However, unlike prior considerations of equity in climate mitigation context, there are two challenges that could potentially impose an even higher negotiation cost. They are institutional complexity and dualism of actors. Jouni Paavola & W. Neil Adger. “Fair adaptation to climate change.” Ecological Economics 56, no. 4 (2006): 601. 71 Grosso, “Justice in Negotiations on the Adaptation Fund,” 365. 72 Grosso, “Justice in Negotiations on the Adaptation Fund,” 365 & John, Rawls, A Theory of Justice, rev. ed., Cambridge, MA: Belknap, 1999: 75. 73 Rawls, A Theory of Justice, 54. 74 Ibid., 72. 75 For a list of Sen’s criticism of Rawls’ Theory of Justice, see Sen, A. (2011). The Idea of Justice. Harvard University Press. 70
Institutional Complexity As described in Section 2, one of the major principles for international climate talks was that negotiations exclusively occurred under the United Nations auspice, in this case the UNFCCC.76 Since participation in the UNFCCC negotiation was universal, an agreed-upon equity principle would have been universally applied to all states. However, as described in Section 4, global climate finance, even under the UNFCCC framework, is managed by different organizations and arenas. For example, the LDCF and SCCF are managed by the UNFCCC through the Global Environment Facility (GEF), an institution established by the World Bank.77 The AF, on the other hand, despite the GEF still providing secretariat service, has greater autonomy due to its funding mechanism through the CDM, and provides a more streamlined process for developing countries to access the funds.78 Furthermore, since the GCF was introduced during the Copenhagen Summit, developing countries successfully advocated that the fund be managed directed through the UN, as opposed through the GEF.79 This complex institutional structure and organizational independence means that each fund is administered separately, with little commonality between them. Therefore, instead of having to reach an agreement on burden-sharing using equity and justice in one institution as before, multiple rounds of negotiations will be needed to reach an agreement in all institutions. This could substantially increase the negotiation costs, as each institution has vastly different power dynamics (for example, developing countries have greater voices Falkner, Stephan, and Vogler, “International Climate Policy After Copenhagen,” 261. 77 Ciplet, Roberts, Khan, “The Politics of International Climate Adaptation Funding,” 62. 78 Ciplet, Roberts, Khan, “The Politics of International Climate Adaptation Funding,” 62-63. & Global Environment Facility. (2016). “Funding”. Accessed on: December 1, 2016. Accessed from: https://www.thegef.org/about/funding. 79 Ciplet, Roberts, Khan, “The Politics of International Climate Adaptation Funding,” 63. 76
Climate Equity and Justice in the Post-Paris Agreement World in AF, while developed nations holds more sway in the LDCF and SCCF).80 Dualism of Actors In traditional climate mitigation negotiations, the goal for each nation was usually defined as an emissions target or allocation.81 Equity and fairness principles were then applied to this single dimension for each nation, with the cost of mitigation folded into this dimension.82 However, in the case of climate finance, each nation now has two separate dimensions to consider: contribution to the global fund, and the distribution of the fund. Applying the two Rawlsian principles of justice to both dimensions could result in different equity principles for each. For example, one of the most popular equity principles is the “historical responsibility” framework, and using the idea of “polluters pay,”83 where the contribution and distribution of the global climate fund is based on the historical emissions of each nation.84 When applied against the Rawls principles of justice, on the contribution side, this equity framework satisfies both principles. While contribution to the global climate fund is unequal, it still satisfies the second principle of justice, as the inequality will benefit those countries who are most disadvantaged. Countries with greater historical emissions have already reaped the economic benefits associated with emissions, while countries who did not emit do not need to contribute to the fund. However, when the same logic is applied to the distribution side, the result becomes absurd. Should countries with the lowest level of historical emissions receive the biggest portion of the fund? Obviously, the answer is no, and a more equitable approach would see the funds reserved for countries who are most vulnerable to climate change. This could then 80
Ibid., 62-63. Luke Tomlinson, Procedural Justice in the United Nations Framework Convention on Climate Change. Cham: Springer International Publishing, (2016): 8. 82 Ibid., 38. 83 Lian-Biao Cui , Lei Zhu, Marco Springmann, and Ying Fan. "Design and analysis of the green climate fund." Journal of Systems Science and Systems Engineering 23, no. 3 (2014): 270. 84 For more information on the historical responsibility equity framework, or any other equity 81
satisfy the second principle of justice, where those who are most disadvantaged receive the greatest benefit of unequal distribution of resources. This dichotomy between contribution and distribution could increase the negotiation cost, since now for each fund, a player has the potential to be both a donor and a recipient. For rapidly expanding economies such as China and India, a switch from one equity principle to another could easily flip their status from a donor to a recipient, and vice versa. Given these barriers, pessimists will rightly question whether this equity discussion will fall to the wayside and future climate negotiators will sidestep once again the issue of equity and fairness in climate finance. Thankfully, there are encouraging signs. Grosso, in his study of the initial AF negotiations, finds that the financial needs of particularly vulnerable nations are emphasized and acknowledged.85 The use of CDM in the Adaptation Fund also stemmed from the principle of “common but differentiated responsibilities,” and is consistent with the equity criteria.86 Furthermore, Grosso goes one step further and argues the success of the AF demonstrated the developed nations’ acknowledgement of their responsibility of the climate impact and their willingness to work on a common governance structure for an international climate fund.87 In addition, LDCs, SIDS, and other vulnerable states have been explicitly acknowledged in various fund institutions’ governance documents, where their needs will be prioritized.88 These informal steps demonstrate that equity considerations, while not explicitly stated, are being considered, and there is a potential to find a common ground that transcends traditional NorthSouth boundaries.
frameworks, see Tomlinson, Procedural Justice in the UNFCCC, 45-49. 85 Grosso, “Justice in Negotiations on the Adaptation Fund,” 375. 86 Ibid. 87 Ibid., 375-376. 88 David Ciplet, J. Timmons Roberts, and Mizan Khan. "The politics of international climate adaptation funding: Justice and divisions in the greenhouse." Global Environmental Politics 13, no. 1 (2013): 59-60.
The Heinz Journal VI. Conclusion While the Paris Agreement moved to a bottom-up and pledge-and-review framework for climate negotiations, equity and fairness will retain their core normative importance in the realm of climate change politics. Due to the greater international demand of climate finances to address climate change, competition for the scarce resource will raise the question of how to best divide up this supply in a way that is fair and equitable. In other words, equity and fairness will once again play an important role in climate negotiations moving forward.
SCCF grants were disbursed during the same time.91 This shows that, despite a growing need in financial resources around the world for climate related activities, funds are not released for these activities. Is this because of a shortage of suitable projects around the world? Or there is a lack of capability in vulnerable countries to successfully apply for these grants? A better understanding of this question will allow funds to flow into the regions where they are most needed.
However, this paper does not attempt to address which equity framework can be used as a starting point to determine what a just climate finance regime could look like. Would it be “historical responsibility?” Or “payment capability?” Or a combination of responsibility and capacity, like the Greenhouse Development Rights Framework?89 Identifying a suitable and politically acceptable equity framework for this context will significantly reduce the cost of future climate negotiations. While this paper has addressed the normative aspects of equity in climate finance, there remains much work to be done to understand the market forces driving pledges and disbursements. After all, the creation of a just funding structure is but one of the means to the end, and the gap in overall funding – not only in pledges, but in actual funds disbursed – needs to be addressed. As Ciplet, Roberts, and Khan wrote: “A justly governed fund without sufficient money does little to promote justice for the millions of people already experiencing adverse impacts of climate change.”90 The GEF FY15 report showed that only 35.41% of LDCF grants were disbursed in 2015, while only 45.47% of 89
Paul Baer, Tom Athanasiou, and Sivan Kartha The Right to Development in a Climate Constrained World: The Greenhouse Development Rights Framework: A Report. (2007). Berlin: Heinrich Böll Foundation. 90 David Ciplet, J. Timmons Roberts, and Mizan Khan. "The politics of international climate adaptation funding: Justice and divisions in the greenhouse." Global Environmental Politics 13, no. 1 (2013): 64-65.
Global Environmental Facility. FY15 Annual Monitoring Review of the Least Developed Countries Fund and the Special Climate Change Fund.. (May 2016) Accessed from: https://www.thegef.org/sites/default/files/councilmeetingdocuments/EN_GEF.LDCF_.SCCF_.20.04_FY15_A MR_LDCF_SCCF.pdf.
The Heinz Journal
Gender and STEM Education in Japan and the United States Nilanjan Raghunath*, Aye Myat Khine Win*, and Fumiko Nishino+ * Singapore University of Technology and Design, in collaboration with MIT + Hitotsubashi University - Graduate School of Social Sciences Science, technology, engineering and mathematics (STEM) fields play an important role in a nationâ&#x20AC;&#x2122;s rapid economic development, but many universities struggle to attract women into STEM fields despite improving gender policies and practices, especially in developed countries. In the United States and Japan, women are still underrepresented in computer science education and the STEM workforce more generally. This underutilization of women's talent represents a lost potential that further affects the gender balance within information technology fields. This article analyzes data from official US and Japanese sources, and shows that women are still underrepresented in STEM education in the United States and Japan despite decades of effort from policymakers to push for better gender equality in education. We argue that policy alone cannot bridge the gender divide, and that more action is required from corporations and educational institutions to create positive environments and develop more female mentors and role models.
II. Data and Methodology
This study addresses the issue of why developed countries continue to struggle to attract women into science, technology, engineering, and mathematics (STEM) programs. Comparing female participation trends in Japan and the U.S. provides a better understanding of different perspectives toward STEM education, and gives insight into the difficulties of improving participation rates through policy. We find that low participation rates of women in STEM programs persist despite policy efforts toward gender equality in tertiary education, and argue that schools, universities, and employers need to make micro-level changes to counter negative stereotypes. We also argue that internationally standardized educational statistics must be introduced in order to provide better data to make more robust comparisons between countries. It is important that policymakers realize the potential significance of the gender gap in STEM in order to counter the current lack of public awareness, which acts as a feedback loop for lower priority gender policies.1
In this study, we apply a non-experimental research design to analyze data from the U.S. National Science Foundation (NSF), the U.S. Department of Education, the U.S. Bureau of Labor Statistics, and the Japanese Ministry of Internal Affairs and Communications Statistics Bureau. These sources provide nearly five decades of official, accurate data on education and careers related to the computer science (CS) field. The data spans a time period from the early 1960s to 2013.
United Nations Educational, Scientific and Cultural Organization. 2007. Science, Technology and Gender: An International Report. Paris: United Nations Educational, Scientific and Cultural Organization.
We focus on Japan and the U.S. due to their rapid industrialization and scientific progress in the 20th century, and their current international recognition as leaders in technology innovation. These countries have the two largest shares of Patent Cooperation Treaty international applications; in 2015, the U.S. had the highest share, at 26.2 percent, and Japan the second highest share, at 20.3 percent, by country of residence of applicant.2 The World Intellectual Property Organizationâ&#x20AC;&#x2122;s World IP Report 2015 showed that Japan and the U.S. lead
Japan Patent Office. 2016. JPO Status Report. JPO Tokyo. https://www.jpo.go.jp/shiryou/toukei/pdf/status2016/01 00.pdf.
Gender and STEM Education in Japan and the United States a small group of nations that are predominantly responsible for breakthrough technologies.3 However, Japan and the U.S. have different social norms and cultures, especially regarding gender roles. According to the World Economic Forum Gender Gap Ranking, Japan ranks 111th out of 144 countries and the worst among G7 countries in terms of its overall gender gap. In contrast, the U.S.’s ranking for overall gender gap is significantly higher, at 45th; furthermore, the U.S. ranked 1st for female educational attainment in 2016.4 According to the 2017 OECD report The Pursuit of Gender Equality, Japan has a sharp division of gender roles and a significant educational gap. The report states that “Japanese girls and women perform very well in OECD education evaluation programs PISA and PIAAC, with Japanese women performing the highest in PIAAC among women in the OECD. Despite this, Japan has the lowest share of female bachelor’s graduates in the OECD (45.4 percent).”5 Finally, comparing Japan and the U.S. provides an interesting basis on which to study women’s participation rates in STEM education due to their vastly different education systems. Our review of the literature discussing women’s participation in STEM fields identified themes including a lack of academic mentorship for females, fewer role models, the pay gap, gender stereotyping, and an overall male-dominated environment. We argue that these themes suggest that women need more community support from both genders, as well as additional networking opportunities to develop and sustain positive stereotypes during various phases of their education and careers.
World Intellectual Property Organization. 2015. WIPO Intellectual Property Report 2015. http://www.wipo.int/edocs/pubdocs/en/wipo_pub_944_ 2015.pdf. 4 World Economic Forum. 2016. The Global Gender Gap Report 2016. http://www3.weforum.org/docs/GGGR16/WEF_Global _Gender_Gap_Report_2016.pdf. 5 OECD. 2017. The Pursuit of Gender Equality: An Uphill Battle. OECD Publishing, Paris. http://dx.doi.org/10.1787/9789264281318-en.
III. Literature Review Geek Culture and Gender Stereotyping In order to understand why women are underrepresented in STEM fields, it is important to first understand the underlying reasons that lead women to find these fields unattractive. One obstacle is the so-called “geek culture” of STEM Geek culture portrays the image of an ideal computer scientist, who has an excessive obsession with and encyclopedic knowledge of computing as well as a fascination with advanced technology and algorithms, spends most of his time on a computer without involvement in other social activities, and excludes himself from usual social events. For example, Kendall describes the ideal profile of a geek as a white male who excels in school, especially in mathematics and science; has a high IQ; has hobbies such as the collection of technical products; and displays no interest in social activities or skills.6 These images, or myths, give negative impressions to women and become negative inhibitors for women who wish to join STEM fields but do not wish to be stereotyped as manly or socially inactive. Margolis and Fisher state that “while the stereotype of the CS student as someone who is myopically focused on computers is rejected by many male and female students, women report more distress and are more affected by the perceived difference between themselves and their peers.”7 Amy Yin, co-founder of Harvard Women in CS, expressed the social stigma that women considering the CS field face: “…a lot of women think you can’t be social if you’re in computer science, that people will stigmatize you as a nerd, that guys will be intimidated by you, that you won’t find a boyfriend or a husband.”8 Geek culture, then, reflects the domination of males in CS, and means that to 6
L. Kendall. 2000. Hanging out in the virtual pub: Masculinities and relationships online. Berkeley: University of California Press. 7 Jane Margolis and Allan Fisher. 2003. Unlocking the Clubhouse Women in Computing. Cambridge, Massachusetts: The MIT Press. 69. 8 Beth Romanik. 2014. Mind the Gender Gap: Getting Women into Computer Science Careers. TechWell Insights. https://www.techwell.com/techwellinsights/2014/02/mind-gender-gap-getting-womencomputer-science-careers.
The Heinz Journal succeed as a professional, one has to be both male and a geek. These stereotypes lead women to question themselves, their identity in CS, and whether they belong or fit into geek culture, and raises a further question: are there female geeks?9 In Susan Silbey’s work, women from four universities in America were interviewed in order to understand why many female engineers leave the field. The interviews revealed that gender stereotyping is a main factor, as it prevents women from being assigned important tasks in the collaborative work that is required for most engineering projects. Women in the study reported frequent assignment to routine or managerial tasks such as secretarial work, copying, and editing rather than problem-solving tasks. A second contributing factor is that women wish to do social good through engineering, but are unable to realize their goals due to organizational constraints. A third factor is that women in male-dominated fields experience sexual harassment, and that not enough is being done to socialize male engineers to treat women with respect. Finally, the authors argue that curriculum reform alone does not help women stay in the field.10 Linda Sax, a gender scholar who studies gender disparities in CS, argues that men and women experience college differently.11 Spertus finds that there are several ways in which males and females are treated differently, including through different expectations and standards. The author concludes that gender stereotyping can deter women from pursuing an interest in mathematics or engineering and can sabotage a woman’s career.12
Chuck Leddy. 2014. Closing the Gender Gap in Computer Science. Harvard Gazette. https://news.harvard.edu/gazette/story/2014/02/closingthe-gender-gap-in-computer-science/. 10 Susan Silbey. 2016. Why Do So Many Women Who Study Engineering Leave the Field? Harvard Business Review. 11 Linda J. Sax. 2008. The Gender Gap in College: Maximizing the Developmental Potential of Women and Men. San Francisco: Jossey-Bass. 12 Ellen Spertus. 1991. Why are There so Few Female Computer Scientists? Cambridge: MIT Press. 13 Ibid. 14 Mia L. Parviainen. 2008. "The Experiences of Women in Computer Science: The Importance of
Male-Dominated Environment Once women enter a STEM field of study, they may face additional barriers unique to male-dominated fields. Females pursuing CS Bachelor’s degrees in a masculine environment often feel uneasy and may have difficulties adapting to the environment. Spertus finds that men often have habits of making sexist or sexual remarks as humor, which makes women feel uncomfortable.13 Parviainen’s study describing how females interact in male-dominated environments, including the classroom, argues that the lack of females in CS could be the cause of sexism.14 In other words, a low representation of women in the classroom prevents more women from being attracted to the same classroom, and is a social barrier for females in the male-dominated CS field. As Margo I. Seltzer, the Herschel Smith Professor of Computer Science at SEAS, describes, "underrepresentation leads to continued underrepresentation. We all need to work toward making our workplaces and communities more welcoming."15 Wilson also studied the factors promoting success in CS, including gender differences. She found that, second only to mathematics and computer programming background, comfort level is the most critical and best predictor for success in the CS classroom.16 A further study by Scragg and Smith found that, out of six possible barriers to women taking undergraduate CS courses, male dominance and a lack of self-confidence were the two most significant factors.17 Margolis and Fisher pointed out that the loss of female confidence is especially severe in historically male-dominated
Awareness and Communication." Human Architecture: Jounal of the Sociology of Self-Knowledge 6 (4): 87-94. 15 Chuck Leddy. 2014. Closing the Gender Gap in Computer Science. Harvard Gazette. https://news.harvard.edu/gazette/story/2014/02/closingthe-gender-gap-in-computer-science/. 16 Brenda Cantwell Wilson. 2002. "A Study of Factors Promoting Success in Computer Science Including Gender Differences." Computer Science Education 12: 141-164. 17 Greg Scragg and Jesse Smith. 1998. "A Study of Barriers to Women in Undergraduate Computer Science." SIGCSE. New York: ACM SIGCSE Bulletin. 82-86.
Gender and STEM Education in Japan and the United States fields.18 These factors might explain why women who initially pursue CS often switch to different majors. Academic Mentorship and Role Models Earlier researchers have identified the importance of faculty in mentoring CS students.19 Studies have shown that most CS faculty are males who tend to have some subconscious bias against females.20 Good academic mentorship can be crucial to a student’s success and has a great impact on females in CS. As Gilbert et al. point out, “[f]emale graduate students who identified female professors as role models viewed themselves as more career oriented, confident, and instrumental than did female students identifying male role models.”21 However, many studies have concluded that girls in the U.S. and Japan are not encouraged to continue in mathematics classes (Dachey, 1983; Hess and Miura, 1985; Isa and Chinen,2014).22,23,24 There is therefore a need for interventions at the community and governmental levels to support women pursuing CS. These forms of support can include academic advice, scholarships, support groups, and awareness. In the U.S., institutions including Harvey Mudd College, the University of Washington, and Stanford University devote significant resources to making sure women are comfortable in the male-dominated CS environment. At Stanford University, for example, a student organization called “Women in Computer 18
Jane Margolis and Allan Fisher. 2003. Unlocking the Clubhouse Women in Computing. Cambridge, Massachusetts: The MIT Press. 19 Henry Etzkowitz, Carol Kemelgor, and Brian Uzzi. 2000. Athena Unbound The Advancement of Women in Science and Technology. Cambridge: Cambridge University Press. 20 Simeone Angela. 1987. Academic Women: Working Towards Equality. Massachusetts: Bergin and Garvey Publisher. 21 Lucia A. Gilbert, June M. Gallessich, and Sherri L. Evans. 1983. "Sex of Faculty Role Model and Students' Self Perceptions of Competency." Sex Roles: A Journal of Research. 9: 597-607. 22 Dachey, Karen. 1983. "Computing not GenderNeutral: Females More Intimidated." Vol. 183. Standford, California: The Stanford Daily, April 8. A5. http://stanforddailyarchive.com/cgibin/stanford?a=d&d=stanford19830408-01.1.5&e=------en-20--1--txt-txIN-------.
Science” is actively engaged in supporting and promoting females in CS. In Japan, however, there are very few organizations supporting women in CS, though the number has seen slight recent improvement. Furthermore, although males have numerous successful role models in CS-related fields, such as Steve Jobs, Bill Gates, and Mark Zuckerberg, successful female role models are rare. Pay Gap Another reason that females avoid pursuing STEM fields is the pay gap. Statistics show that CS has the highest pay gap of all fields, with females earning about 23 percent less than males. The pay gap is significantly less in healthcare fields and the humanities; even other STEM fields have a gender pay gap about half that of CS, at 12 percent.25 This wide pay gap certainly discourages women from entering the CS field. Furthermore, if females who graduate with CS degrees have limited opportunities to pursue their career, this might have larger social implications. Jennifer Hunt, Professor of Economics at Rutgers University, commented that “women in the engineering field are having a harder time advancing compared to other fields.” Her studies on the exit rate of females from STEM fields show that as the number of males in the field increases, the number of women leaving the field also increases, and that these women cite reasons for leaving that include the pay gap and promotion opportunities.26 23
Robert D. Hess and Irene T. Miura. 1985. "Gender Differenes in Enrollment in Computer Camps and Classes." Sex Roles 13: 193-203. 24 Natsumi Isa and Chinen Ayumu. 2014. “Rikei Kamoku ni okeru Gakuryoku to Iyoku no Jyenda Sa (Gender Gaps of Achievement and Aspirations in Mathematics).” Nihon Rodo Kenkyu Zasshi (The Japanese Journal of Labour Studies). 648: 84-93. 25 Emily Forrest Cataldi, Caitlin Green, Robin Henke, Terry Lew, Jennie Woo, Bryan Shepherd, Peter Siegel, and Ted Socha. 2011. 2008-09 Baccalaureate and Beyond Longitudinal Study (B&B:08/09). Washington, DC: National Center of Education Statistics. http://nces.ed.gov/pubs2011/2011236.pdf. 26 Jenny Marder and PBS NewsHour. 2012. Accessed October 28, 2014. http://www.scientificamerican.com/article/why-theengineering-and-science-gender-gap/.
The Heinz Journal IV. Policy Environment American Policy America is a global leader in innovation and entrepreneurship because it has been a magnet for local and global talent, scientific advances, infrastructure, and entrepreneurs Recently, President Trump signed two laws that authorize NASA and the NSF to encourage more women to enter and stay in STEM fields. The first law, the Inspire Act, gives NASA three months to form two congressional committees with the goal of developing plans to promote careers for women in aerospace as scientists, engineers, and astronauts. The second law, the Promote Women in Entrepreneurship Act, allows the NSF to support more programs that attract women into entrepreneurship.27 Historically, past American presidents have encouraged an open and diverse scientific community by increasing engagement in both formal and informal settings and providing mentoring. However, despite these policies, the gender gap in STEM fields has persisted. According to the National Science and Technology Council on STEM education, the committee on STEM education is comprised of 13 agencies, including all major science agencies and the Department of Education. The priority areas for improvement are K-12 education, undergraduate and graduate education, public engagement (for mentoring and collaboration), broadening participation by reaching out to minorities and underrepresented groups, and coordination and evaluation (analyzing the impact factor of various agencies). These areas are to be improved by implementing the Federal STEM Education Five Year Strategic Plan to drive federal investment in STEM.28
Erin Carson. 2017. Trump Signs Laws to Promote Women in STEM. CNet. https://www.cnet.com/news/trump-women-in-stemtech-laws. 28 National Science and Technology Council. 2016. Progress Report on Coordinating Federal Science, Technology, Engineering, and Mathematics (STEM) Education. https://obamawhitehouse.archives.gov/sites/default/files /microsites/ostp/stem_budget_supplement_fy_17_final_ 0.pdf.
Women account for about 47 percent of the workforce in the U.S.29 However, women’s participation rates are only 25.6 percent in computer and mathematical occupations, and 15.4 percent in architecture and engineering occupations. Additionally, women earn only 18 percent of computer science degrees. There is clearly a need to attract more STEM-trained graduates to the job market in order to drive the economy. More policies are needed to address STEM education deficits in schools, including policies which invest in more teachers and resources. Furthermore, better policies must be developed for women in the workplace, because only approximately a quarter of women with STEM degrees stay in the profession.30 Japanese Policy Dr. Maki Kubo, from the Japanese Society for the Promotion of Science, summarized the current actions of Japanese funding agencies to assist female researchers, and mentioned that “in order for the women to choose scientific and engineering research careers, the strong influence of the traditional view and the role of women must be overcome in Japan. As the media still portrays women as homemakers rather than scientists and these traditional stereotypes are problematic and need to be addressed [sic].”31 During a workshop led by Dr. Kubo, the workshop group primarily discussed issues on leadership and training. In order to fill more senior positions and have the opportunity to influence change within structures and institutions, the group concluded that both women and men need appropriate experience, mentorship and training in management, and leadership and communication skills. Leaders in all sectors must be prepared to effectively address problem situations, including communicating concerns about barriers to women in STEM. 29
Bureau of Labor Statistics. 2017. Demographic Characteristics. https://www.bls.gov/cps/demographics.htm. 30 Office of Science and Technology Policy. 2013. Women in STEM. https://obamawhitehouse.archives.gov/administration/e op/ostp/women. 31 OECD. 2006. Women in Science, Engineering and Technology (SET): Strategies for a Global Workforce Workshop Summary. Workshop Summary.
Gender and STEM Education in Japan and the United States From the workshop, three main policy recommendations were developed: 1. The production of sex-disaggregated statistics is essential, as without such data it is impossible to analyze trends and outcomes and monitor progress. 2. Performance evaluation methods need to be addressed. Even though assessment criteria may have subjective biases, evaluation methods need to be fair, transparent, and gender neutral. 3. More rigorous research is needed to assess the relationship between gender diversity and firm or research performance, and to determine whether gender diversity, particularly at senior levels, affects performance. This work may be able to demonstrate that gender equality has quantifiable benefits that can be linked to economic growth or improved social outcomes. However, this task should be carried out in conjunction with developing equitable performance evaluations.32 In the paper Japanese Women in Science and Technology, Kuwahara describes government policies established in 1958 to strengthen traditional gender roles and encourage girls to study home economics while boys study technology in junior high schools. Similarly, in 1969 the Ministry of Education announced curricula which discouraged women students.33 Through both policy and social norms, women were encouraged to be housewives between 1955 and 1975; however, the number of women in the workforce increased from 1975 to 1995. During this period, many young women lost interest in science and technology due to the combination of techno-nationalism and masculine ideology which favored the employment of men in STEM fields in the 1980s.
will support universities and public research institutions to assist female researchers in coping with childbirth, childcare, and research. The government will expect universities and public research institutions to actively promote female appointments by formulating specific plans to stimulate the activities of female researchers and achieve the numeric targets set for female researchers, and by releasing data to the public regarding the percentage of enrolled women researchers categorized by job classification and department. The government will also expect these institutions to make efforts to increase the number of female researchers in leading positions, the number of female students of natural science, and the number of quality women aiming to be professional researchers.34 In 2010, the Council for Science, Technology and Innovation reported that there is an increasing trend for students to avoid STEM departments at colleges and universities, and that many excellent researchers and engineers in Japan are reaching retirement age. To nurture and enhance the interest of children in STEM fields during elementary and secondary education, the government must develop consistent methods to identify talented children and develop their abilities.35
Japanâ&#x20AC;&#x2122;s 2010 Science and Technology Basic Policy Report states that the government will promote the activities of female researchers to enhance gender equality as well as demonstrate organizational creativity by adopting various viewpoints and ideas and stimulating research activities. The government
Global Policies Although women represent a significant portion of any nationâ&#x20AC;&#x2122;s human resource base, they are overwhelmingly underrepresented in science and technology policy, research, and development. To address these issues, the Gender Working Group of the United Nations Conference on Trade and Development conducted a series of studies over two years, and developed the following Transformative Actions: 1. Gender equity in science and technology education; 2. Providing enabling measures for addressing gender inequalities in scientific and technological careers; 3. Making science responsive to the needs of society: the gender dimension;
Ibid. Motoko Kuwahara. 2001. "Japanese Women in Science and Technology." Kluwer Academic Publishers 203-216. 33
Council for Science and Technology Policy. 2010. Japan's Science and Technology Basic Policy Report. 35 Ibid.
The Heinz Journal 4. Making the science and technology decisionmaking process more “gender aware;” 5. Relating better with “local knowledge systems;” 6. Addressing ethical issues in science and technology: the gender dimension; 7. Improving the collection of gender disaggregated data for policy makers; 8. Equal opportunity for entry and advancement into larger-scale science, technology, engineering, mathematics disciplines and innovation systems.36 The 2005 Women in Science, Engineering and Technology: Strategies for a Global Workforce workshop held by OECD member countries presented, discussed, and explored the outcomes of various solutions to overcome barriers to the education, recruitment, progression, and retention of women in STEM, from grade-school level to university studies and late careers, and to address gender and scientific excellence.37 The background paper of the workshop concluded that: …women obtain more than half of all university degrees in many countries but only around 30 percent of university degrees awarded in science and technology, OECD countries face a paradoxical situation: a feminization of the workforce in general and of university-trained graduates in particular, but continued underrepresentation of women in the research workforce… the available data tend to reinforce results from the academic literature that show women remain unevenly distributed in research occupations and underrepresented in senior positions [sic].38
V. Comparative Case Study on Educational Outcomes As Japan and the U.S. have distinct gender-related norms and cultural practices, we conducted a comparative study between these two countries to explore the outcomes of female participation in STEM, especially in CS and informational engineering. 36
United Nations Conference on Trade and Development (UNCSTD) Gender Advisory Board. Accessed September 12, 2015. http://gab.wisat.org/transfom.htm#1.
In this section, we explore and discuss the statistics on female participation in STEM in the U.S. and Japan. This will allow us to determine the main causes of STEM’s unattractiveness to women. Moreover, this analysis allows us to gain a better understanding of how developed countries organize data on female participation trends in STEM tertiary education. We organized our analysis and study of these trends on the following questions: Question 1: Are there any similar female participation trends in STEM tertiary education between the U.S. and Japan? Question 2: Does the underrepresentation of women in STEM fields occur only at the undergraduate Bachelor’s level, or at other educational levels including Master’s and Doctoral programs in American and Japanese institutions? Before analyzing female participation trends, we studied the data organization patterns of Japan and the U.S. to allow for a direct comparison between the fields of study in their respective education systems. We found that in the U.S., statistics on female participation are collected according to faculty and departments. Data is collected according to two main categories: STEM and nonSTEM. Under the STEM category, there are two different sub-categories: engineering and science (including natural sciences and social/behavioral sciences). U.S. data collection is more comprehensive, and includes yearly data on specific science and engineering indicators (starting from 2012), which makes the data easily analyzed for any particular field of study. There is a more organized collection of statistics in the U.S. than in Japan. In Japan, data is collected differently than in the U.S. In Japan’s universities, statistics are organized according to eleven main fields: humanities, social science, science, engineering, agriculture, health, mercantile marine, home economics, education and teacher training, arts, and other. We focus on four fields: humanities, social science, science, and 37
OECD. 2006. Women in Science, Engineering and Technology (SET): Strategies for a Global Workforce Workshop Summary. Workshop Summary. 38 Ibid.
Gender and STEM Education in Japan and the United States engineering. However, there is no clear breakdown of sub-fields under the science and engineering categories. Most engineering majors are offered in technical colleges, and there are approximately eighteen different engineering majors offered in Japan, including mechanical, electrical, electronics, civil, and material engineering. Hence, some data from technical colleges will be used in comparisons with the U.S. An important aspect for our analysis is the number of females who participate in science and engineering fields. The earliest available data on Japan’s education system is from the year 1960, while U.S. data is not available until about 25 years later. However, data on male or female participation in Japan was not collected until recent years. Although statistics on female participation in the U.S. are available from the year 1985, these statistics in Japan can only be obtained from the year 2009. Hence, when comparing between the
two countries, there is a large data gap of about 24 years for the percentage of female participation in STEM education. This makes it very difficult to directly compare STEM participation in Japan and the U.S. Further difficulties are due to the fact that data on CS majors in U.S. institutions can only be compared with data on information engineering majors in Japanese technical colleges. To rectify these difficulties, we suggest that there is a need for international standardization in STEM education. Figure 1 shows a general trend of female participation in CS education at the Bachelor’s level in the U.S. from 1985 to 2011. From this figure, we can see that the number of females who obtained Bachelor’s degrees in psychology, biosciences, and social sciences outnumbered males by more than 50 percent. However, the percentage of females receiving CS Bachelor’s degrees declined from 27.6 percent to 18.2 percent
Figure 1: Percentage of Bachelor’s Degrees in a Field Earned by Females in the U.S.
Women’s Share of STEM Bachelor’s Degrees by Field in the U.S, 1985-2011 100% 90% 80%
70% 60% 50% 40% 30% 20% 10% 0% 1985
Year Physical sciences
Biological/ agricultural sciences
Data source: National Science Board, 2014
The Heinz Journal within the 27-year period between 1985 and 2011.39 Although the percentage of females pursuing Bachelorâ&#x20AC;&#x2122;s degrees was either maintained or gradually increased during this time period, the percentage pursuing CS declined dramatically, particularly from 2000 to 2010. As Figure 2 illustrates, in the 53 years between 1960 and 2013 in Japan, less than 4 percent of students pursued science degrees while less than 21
percent pursued engineering degrees. This number shows that only one fourth of undergraduate students pursued science and engineering degrees. Moreover, it shows that both males and females are not as interested in science and engineering as other fields, such as social science and humanities, at the undergraduate level. Science and engineering have been unattractive to most students at the undergraduate level in Japan for more than five decades.
Figure 2: Percentage of Total Participation by Field in Japan, 1960-2012
Total Participation by Field at Undergraduate Level in Japan, 1960-2012 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0%
Data source: Statistics Bureau of Japan In Japanâ&#x20AC;&#x2122;s education system, most engineering courses are offered at Technical Colleges. Figure 3 illustrates total participation in control information, information communication, material, electrical, information, and mechanical engineering courses. Interest in control information, information communication, information, and material engineering courses seems to increase beginning in 1989. However, the percentage of participation in these courses becomes stagnant, with the highest percentage at about six percent from the year 1993 until 2012. With additional data on total participation in Japanâ&#x20AC;&#x2122;s Technical Colleges, we can confirm that interest in engineering courses is not
improving, and that there has been a decline in interest in some engineering courses, including mechanical and electrical engineering.
Minorities, and Persons with Disabilities in Science and Engineering. Arlington, VA.
National Science Foundation, National Center for Sicence and Engineering Statistics. 2013. Women,
Figure 4 presents a comparison between the percentage of participation in science and engineering at the undergraduate level in Japan and the U.S. At the undergraduate level, most students in the U.S. are interested in science rather than engineering fields. In contrast, in Japan engineering courses are more attractive to students than science courses. Overall, the total participation in science and engineering at the undergraduate level is higher in the U.S than in Japan. The maximum
Gender and STEM Education in Japan and the United States participation rate is about 27 percent in the sciences (U.S.) and the minimum participation rate is about four percent in the sciences (Japan). Nevertheless,
the general trend for both countries is that the participation rate in science and engineering declined between 1985 and 2010.
Figure 3: Percentage of Total Participation in Japanâ&#x20AC;&#x2122;s Technical Colleges
100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0%
1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2009 2010 2011 2012 2013
Total Participation in Technical Colleges in Japan by Field, 1973-2013
Year Control information engineering
Information communication engineering
Data source: Statistics Bureau of Japan Figure 4: Percentage of Total Participation in Science and Engineering Fields in Japan and U.S.
Total Participation in Science and Engineering Fields at Undergraduate Level in Japan and the U.S., 1985-2010 30.00
25.00 20.00 15.00 10.00 5.00 0.00 1985 1987 1989 1991 1993 1995 1997 2000 2001 2002 2003 2004 2005 2006 2008 2010
Year Engineering (Japan)
Data source: Statistics Bureau of Japan, National Science Board, 2008 and 2014
The Heinz Journal Analyzing the breakdown of male and female participation in science and engineering at the undergraduate level in both countries further emphasizes the fact that the gender gap in the engineering field is quite big (see Figure 5). Male participation in the engineering field is about 4 times higher than female participation in the U.S. and about 9 times higher in Japan. Females are underrepresented in the engineering field,
especially in Japan. For the science field, participation of females in the U.S. is about 1.2 times higher than that of males. However, in Japan, the trend is the opposite of the U.S., as the male participation rate is approximately 3 times the participation rate of females. Thus, the gender gap in science and engineering in Japan is wider than that in the U.S.
Figure 5: Participation in Science and Engineering Fields at University Level in Japan and the U.S.
Participation in Science and Engineering Fields at Undergraduate Level in Japan and the U.S, 2009-2010
70.00 60.00 50.00 40.00
Science (Female, US) Science (Male, US) Science (Female, Japan) Science (Male, Japan) Engineering (Female, US) Engineering (Male, US) Engineering (Female, Japan) Engineering (Male, Japan)
30.00 20.00 10.00 0.00 2009
Data source: Statistics Bureau of Japan, National Science Board, 2014 The data on CS Bachelorâ&#x20AC;&#x2122;s degrees awarded between 1985 and 2011 suggests that the gender gap is growing (see Figure 6). Females remain underrepresented in CS undergraduate studies. Although this has been true for several decades, CS as a major has remained unattractive to women, and the percentage of females in CS in American universities has even declined (see Figures 1 and 6). For over 20 years, there has been no improvement in the percentage of females at the CS undergraduate level in American universities.
Due to limited data, the male and female participation rates between Japan and the U.S. can only be compared between 2009 and 2012. In both Japan and the U.S., the male share in CS or information engineering is about four times larger than the female share. Although the female share in science fields is higher than male in the U.S. (see Figure 5), that trend does not extend to the CS major (see Figure 7). More males in the U.S. pursue CS-related majors, while in Japan more females show interest in CS-related majors. Based on this comparison, we can confirm that generally, women
Gender and STEM Education in Japan and the United States are underrepresented in CS-related majors at the undergraduate level. Moreover, males outnumber females in the engineering fields at the undergraduate level in both the U.S. and Japan.
From this analysis, we find a very similar trend of female participation in engineering fields and CSrelated tertiary education between the U.S. and Japan.
Figure 6: Percentage of Computer Science Bachelor’s Degree Earned by Males and Females
Earned Computer Science Bachelor’s Degrees in the U.S., 1985-2012 90 80
70 60 50 40
20 10 0
Data source: National Science Board, 2008 and 2014 Figure 7: Data Comparison Between Japan (Information Engineering) and U.S. (Computer Science)
Participation in Computer Science (U.S.) and Information Engineering (Japan) at Undergraduate Level, 2009-2012
70.00 60.00 50.00 40.00 30.00
Computer Science, Female (US) Computer Science, Male (US) Information Engineering, Female… Information Engineering, Male (Japan)
20.00 10.00 0.00 2009
Data source: Statistics Bureau of Japan, National Science Board, 2014
The Heinz Journal Figure 8: Participation in Science and Engineering Fields at Master's Level in Japan and the U.S.
Participation in Science and Engineering Fields at Master's Level in Japan and the U.S., 2009-2012
70.00 60.00 50.00 Science (Female, US) Science (Male, US) Science (Female, Japan) Science (Male, Japan) Engineering (Female, US) Engineering (Male, US) Engineering (Female, Japan) Engineering (Male, Japan)
40.00 30.00 20.00 10.00 0.00 2009
Data source: Statistics Bureau of Japan, National Science Board, 2014 Participation in science and engineering at the Master’s degree level in both Japan and the U.S. shows a similar trend to the undergraduate level (see Figures 5 and 8). Males still dominate the engineering fields at the Master’s degree level in both countries. Only in the U.S. do more females pursue science-related Master’s compared to males, and the gender gap in the sciences is not very significant compared to the engineering fields. The U.S. gender gap in the science fields is only 7.72 percent, whereas in Japan it is about 57.42 percent. Overall, it can be concluded that females pursue degrees in science and engineering less than males in both countries at the Master’s level. Male domination in CS majors has continued to persist for the last three decades at the Master’s level (see Figure 9). There is an annual average increase in CS Master’s degrees pursued of 0.046 percent for males, while there is an average decrease of 0.043 percent for females within the same period (1985-2012). Female interest in CS has not improved over the years. The female share of
Master’s degree graduates has remained between 26.38 percent and 33.8 percent, constituting about one third of the average cohort; females remain underrepresented in CS Master’s degrees compared to males. However, although this data is available for CS majors in the U.S., no similar data can be obtained for information engineering in Japan. This further emphasizes the need for consistent statistics on education with standardized data collection methods, so that data can be easily compared and a more concerted approach towards achieving gender balance in STEM education can be taken across different countries. At the Doctoral level in Japan and the U.S., male domination in science and engineering is very significant, especially in the engineering fields. This shows that the participation of women in science and engineering is still low not only at the undergraduate and Master’s degree level, but also at the Doctoral level. The gender gap in Figures 8 and 10 illustrates the level of male and female interest in pursuing science and engineering at the
Gender and STEM Education in Japan and the United States higher education level. As the comparison year span is not wide enough for further conclusions, it
reinforces the need for standardized data collection internationally.
Figure 9: Percentage of Computer Science Masterâ&#x20AC;&#x2122;s Degrees Earned by Males and Females
Earned Computer Science Masterâ&#x20AC;&#x2122;s Degrees in the U.S., 1985-2012 80 70
60 50 40
20 10 0
Data source: National Science Board, 2008 and 2014 Figure 10: Participation in Science and Engineering Fields at Doctoral Level in Japan and the U.S.
Participation in Science and Engineering Fields at Doctoral Level in Japan and the U.S., 2009-2012
70.00 60.00 50.00
Science (Female, US) Science (Female, Japan) Science (Male, US) Science (Male, Japan) Engineering (Female, US) Engineering (Female, Japan) Engineering (Male, US) Engineering (Male, Japan)
40.00 30.00 20.00 10.00 0.00 2009
Data source: Statistics Bureau of Japan, National Science Board, 2014
The Heinz Journal Males have constituted approximately four times the number of females earning CS Doctoral degrees for almost 30 years (see Figure 11). Although the percentage of females earning CS Doctoral degrees is increasing annually at 0.54 percent, males have significantly outnumbered females since the 1980s. Based on the data of awarded CS Bachelor’s, Master’s, and Doctoral degrees during the period between 1985 and 2012, males have outnumbered females in all three degree categories for three decades. The gender gap continues to grow in CS higher education, and females are reluctant to pursue a CS major in American institutions (see Figures 9 and 11). The general trend of female participation in CS remains the same; females are still underrepresented in CS majors with no improvement in female participation over two decades in American universities. Moreover, the underrepresentation of women in CS occurs not
only at the undergraduate Bachelor’s level, but also in Master’s and Doctoral programs in American tertiary institutions (see Figures 1, 6, 9 and 11). As no data on information engineering can be obtained at higher education levels in Japan, it is difficult to compare Japanese participation with the U.S. Based on these comparative studies, we have found that both Japan and the U.S. display similar outcomes, albeit under distinct gender-related norms and cultural practices. This could mean that the attractiveness of STEM to women is not only a problem to Japan and the U.S., but also an issue which needs attention worldwide. Being comfortable in STEM classes and jobs will encourage females to increase their self-confidence and self-efficacy, and may boost the participation rate and reduce intimidation within largely maledominated environments.
Figure 11: Percentage of Computer Science Doctoral Degrees Earned by Males and Females
Earned Computer Science Doctoral Degrees in the U.S., 1985–2012 100 90
80 70 60 50 40
20 10 0
Data source: Statistics National Science Board, 2014
Gender and STEM Education in Japan and the United States VI. Conclusion Key Differences in U.S. and Japan Data Based on the case studies, females are still a minority in the CS field in the U.S. and Japan. The proportion of females in the CS field did not exceed 40 percent in either country. Moreover, the literature shows that females are still reluctant to pursue degrees and careers in CS because they are underrepresented due to gender stereotyping, subconsciously biased behaviors, sexism, lack of self-esteem, lack of encouragement, an uncomfortable environment, and the pay gap. Key Educational Findings Macro-level policies focus on equality in CS education. However, there seems to have been little progress in attracting women to this field. Macrolevel policies can do little to change prevailing attitudes and the everyday life experiences of potential female computer scientists and engineers. Hence, more effort needs to be made at the university and departmental level to attract and retain female students. Networking events targeted at female students and cross-disciplinary support would help to integrate CS with other fields and thereby promote better social integration of both genders. The report from the American Association of University Women (AAUW) states that transition between high school and university is a crucial point where most female students are discouraged from STEM career paths. This discouragement can also be due to a lack of mathematical qualifications in comparison to male peers.40 This shows that to encourage more women to enter fields that require prior mathematics training, more needs to be done at the grade-school level and perhaps at home.
dominated fields of education. Employers should take a larger role in the gender balance of their workforce and take into consideration female preferences. Policymakers and researchers need to study the power of social influence, as studies show that groups with similar attributes have significant social influence on memberâ&#x20AC;&#x2122;s behavior patterns.41 Another issue that needs to be addressed is the labor shortage in the field: Why are employers are more willing to outsource work than hire more females locally? The answer partly lies in the fact that few females are available locally, and that outsourcing to a cheaper country continues to be the model for many firms. Overall, the findings indicate a gender gap in STEM fields which, while present in both countries, is wider in Japan compared to the U.S.
Key Career Findings The gender gap in choice of majors needs further research in terms of large-scale qualitative and quantitative studies across universities to understand why certain attitudes prevail despite progress in the overall position of women in male40
Catherine Hill, Christianne Corbett, and Andresse St. Rose. 2010. Why So Few? Women in Science, Technology, Engineering, and Mathematics. Washington: American Association of University Women (AAUW).
Christakis, Nicholas A., and James H. Fowler. 2009. Connected: The Surprising Power of Our Social Networks and How They Shape Our Lives. New York: Little, Brown and Company.
The Heinz Journal
Analysis of Intergenerational Income Mobility for Counties Within the U.S. Using Machine Learning Ada Tso, Gursmeep Hundal, and Vicky Mei Carnegie Mellon University – Heinz College of Information Systems and Public Policy This study utilizes the data on intergenerational income mobility and geographic characteristics from the team at Equality of Opportunity Project. Several machine learning models are used on these administrative records of the incomes of children born between 1980-1982 and their parents, along with other demographic variables. The research then presents the performance of several regression and classification models by highlighting the most important predictors for each model.
I. Introduction Intergenerational income mobility describes the change in income levels across generations. The concept of whether a child’s income will surpass his or her parent’s income is often discussed in broad strokes. Once described as the core foundation of the American dream, intergenerational income mobility has been declining in the United States in recent years.1 Our research looks at this metric by analyzing the variation of income mobility among the United States’ counties. The ability to determine which counties are more or less likely to support income mobility could be useful for local, state, and federal agencies as they work to effectively direct resources to community development organizations. Evaluating the distribution of income on the basis of equality studies is “conducted in static terms, where the so-called ‘snapshots’ of the income distributions are taken as the primitives of the analysis.”2 These distributions, however, can provide only an incomplete picture, since, in most instances, the social welfare would certainly depend on the dynamics of income distribution as 1
Katz, Lawrence F., and Alan B. Krueger. “Documenting Decline in U.S. Economic Mobility.” Science, vol. 356, no. 6336, 2017, pp. 382–383., doi:10.1126/science.aan3264. 2 Fields, G. S. & Ok, E. A. (1999). The measurement of income mobility: An introduction to the literature [Electronic version]. In J. Silber (Ed.) Handbook on income inequality measurement (pp. 557-596). Norwell, MA: Kluwer Academic Publishers.
well. In most instances, the social welfare would certainly depend on the dynamics of income distribution as well.3 In contrast, the team at the Equality of Opportunity Project, led by the economists Raj Chetty (Stanford), John Friedman (Brown), and Nathaniel Hendren (Harvard), is focused on geographic variation in mobility.4 Our paper builds off the work of Chetty et al. by using county-level data and several machine learning paradigms to build a model that can predict factors that affect mobility within U.S. counties. The research in this domain has evolved significantly over the years. Typically, researchers studying this topic have focused on intergenerational elasticity (IGE), which measures the percent increase in income that a child can expect to secure for every percent increase in the income of her or his parents. Under this interpretation, IGE is inversely related to mobility: a high IGE indicates less mobility, and a low IGE indicates higher mobility.5 Musick et al. (2004) explored the links between poverty and family structure from one generation to the next, to find how intergenerational inheritance affects trends in
Solon, Gary. "Intergenerational Mobility in the Labor Market." Handbook of Labor Economics, 1999, 1761800. doi:10.1016/s1573-4463(99)03010-2. 4 Chetty, Raj, Nathaniel Hendren, Patrick Kline, and Emmanuel Saez. "Where is the Land of Opportunity? The Geography of Intergenerational Mobility in the United States." 2014. doi:10.3386/w19843. 5 "Economic Mobility in the United States." July 2015. http://www.pewtrusts.org/~/media/assets/2015/07/fsmirs-report_artfinal.pdf.
Analysis of Intergenerational Income Mobility poverty and single parenthood over time.6 Mazumdar (2005) used administrative data containing the earnings histories of parents and children, ultimately estimating “that the United States is substantially less mobile than previous research indicated.”7 Mitnik et al. (2015) analyzed data from the Statistics of Income Mobility Panel, which was assembled to provide evidence on economic mobility and the implications of tax policy for economic mobility.8 Collectively, these IGE researchers have found that approximately half of parental income advantages are passed on to children. When averaged across all levels of parental income, the IGEs of total income are estimated at 0.52 for men and 0.47 for women. These estimates, which are at the high end of the existing range, imply that the United States is very immobile.9 Regarding future research, there is growing interest in the analysis of the role single mothers play, with research by Condron (2007)10 and Musick & Mare (2004),11 but the link to intergenerational mobility is yet to be explored in depth and our analysis corroborates the need to do it. The literature points to the unequal distribution of opportunities to thrive economically. In this paper, we use machine learning models to go beyond assessing mobility level and attempt to identify the key drivers of variation in mobility at the countylevel.
II. Data Our paper uses data from Chetty et al., including information on mobility and geographic characteristics, which these authors have used in several papers on economic opportunity since 2014. Using administrative records on the incomes of millions of children born between 1980-1982 and their parents, the researchers: 1. Ranked parents’ income between 1996-2000 relative to other parents with children in this same birth cohort; 2. Ranked children’s income in 2011 and 2012 relative to other children in the same birth cohort; 3. Established a mobility score based on the slope of the relationship between the two ranks, providing the correlation between where children and their parents fall on the income distribution.12 While measuring intergenerational income mobility, absolute upward mobility refers to the mean income rank of children whose parents are at the 25th percentile of the national parent income distribution (moving forward, this will be referred to as “mobility”). Fixing the mobility score for those whose parents are at specific percentile, rather than fixing across the entire spectrum of income, helps avoid the possibility that a mobility score is more reflective of worse outcomes for the rich rather than better outcomes for the poor.13 To further understand mobility scores, which were calculated by Chetty et al. and range from 23.70 to 63.80, it is useful to look at an example. The
Musick, Kelly A., and Robert D. Mare. "Family Structure, Intergenerational Mobility, and the Reproduction of Poverty: Evidence for Increasing Polarization?" Demography 41, no. 4 (2004): 629-48. doi:10.1353/dem.2004.0034. 7 Mazumder, Bhashkar. "Fortunate Sons: New Estimates of Intergenerational Mobility in the United States Using Social Security Earnings Data." Review of Economics and Statistics 87, no. 2 (2005): 235-55. doi:10.1162/0034653053970249. 8 Pablo A. Mitnik, Victoria Bryant, Michael Weber, and David B. Grusky, “New Estimates of Intergenerational Mobility Using Administrative Data,” Statistics of Income Division, Internal Revenue Service, (2015). 9 Ibid.
Condron, Dennis J. "Stratification and Educational Sorting: Explaining Ascriptive Inequalities in Early Childhood Reading Group Placement." Social Problems 54, no. 1 (2007): 139-60. doi:10.1525/sp.2007.54.1.139. 11 Musick, Kelly A., and Robert D. Mare. "Family Structure, Intergenerational Mobility, and the Reproduction of Poverty: Evidence for Increasing Polarization?" Demography 41, no. 4 (2004): 629-48. doi:10.1353/dem.2004.0034. 12 Chetty, Raj, Nathaniel Hendren, Patrick Kline, and Emmanuel Saez. "Where is the Land of Opportunity? The Geography of Intergenerational Mobility in the United States." 2014. doi:10.3386/w19843. 13 Ibid.
The Heinz Journal
Figure C.1: Timeline of Data Sources
Analysis of Intergenerational Income Mobility mobility score of Salt Lake County, UT is 45.7, while the score of for Mecklenburg County, NC (where Charlotte is located) is 38.8. This means that, among children whose parents were in the 25th percentile of income distribution, those who grew up in Salt Lake are, on average, 6.9 percentile points higher in their birth cohortâ&#x20AC;&#x2122;s income distribution at age 30, as compared to those who grew up in Mecklenburg.14 More simply, children from households of comparable wealth on average earn more at age 30 if they grow up in Salt Lake instead of Mecklenburg. The researchers assembled place-based characteristics related to segregation, K-12 education, income distribution, family structure, social capital, and other factors in order to understand correlations between geographic variation in mobility. Children were assigned to the county where they lived at age 16. To provide further detail, Figure C.1 displays an overview of the specific covariates used in this paper, as well as their sources, the year they were collected, and the age of the children. Aside from differences in methodologies applied to the dataset, this paper has a few other differences from the work of Chetty et al. The analysis in this paper is conducted at the county level, while the majority of other researchersâ&#x20AC;&#x2122; analysis is performed using commuting zones, which are groups of neighboring counties. We believe that county-level analysis is a useful approach, because federal and state government grants and funding are often awarded or distributed at the county level. Public services are also commonly administered at county level, rather than by commuting zones. Chetty et al. explain that commuting zones (CZs) are aggregations of counties, based on commuting patterns from the 1990 Census. Constructed by Tolbert and Sizer (1996), there are 741 CZs in the U.S. and each CZ contains 4 counties on average.15
Ibid. Ibid. 16 R: Missing Value Imputations by randomForest. http://math.furman.edu/~dcs/courses/math47/R/library/r andomForest/html/rfImpute.html. 15
III. Data Cleaning County covariates were available for 3,137 counties. We removed 369 counties that lacked mobility scores due to insufficient sample points of individuals. This left us with 2,768 counties. We also omitted several variables associated with college education, because the related data was missing for more than half the counties. Further, we added variables that had been collected by Chetty et al. but not used by them in their regression analysis. These variables included income growth in counties, population density, teenage birth rate, and median house price. Ultimately, 41 predictors were used in our models. There were several variables that had missing values in some counties, which can be seen in Figure C.2. We chose to impute the missing values using the rfImpute function from the randomForest package in R. The algorithm first uses na.roughfix, which imputes the missing values by using the median or mode of the variable, then imputes missing values in our predictors using proximity from randomForest. The proximity matrix from the randomForest is used to update the imputation of the NAs.16 For continuous predictors, the imputed value is the weighted average of the non-missing observations, where the weights are the proximities.17 Figure C.2: Variables with Missing Values Variable Number Percent Growth in Chinese Imports 21 0.7% School Expenditures Per Student 21 0.7% Student-Teacher Ratio 247 7.9% Test Score Percentile 20 0.6% High School Dropout Rate 566 18.0% Social Capital Index 21 0.7% Violent Crime Rate 163 5.2% Total Crime Rate 163 5.2%
"RandomForest." Function | R Documentation. https://www.rdocumentation.org/packages/randomFore st/versions/4.6-12/topics/rfImpute.
The Heinz Journal Overview of Data Figure C.3 shows mobility scores by county across 48 States in the United States. Counties with lower scores are colored yellow and green, while counties with higher scores are shaded in blue. Counties
without a mobility score, due to insufficient data, are white. The map of the continental United States18 shows that counties with the lowest scores are concentrated along the southeast, particularly in Mississippi, Alabama, Georgia, South Carolina and
Figure C.3: Variation in Absolute Upward Mobility in the Continental United States
Figure C.4: Counties with Lowest and Highest Mobility Scores Lowest 10 Counties Rank County State Score 1 Shannon South Dakota 23.7 2 Todd South Dakota 24.3 3 Nome Alaska 30.7 4 Richmond City Virginia 30.9 5 North Slope Alaska 31.2 6 Northwest Arctic Alaska 31.3 7 Tunica Mississippi 31.8 8 McKinley New Mexico 32.1 9 St Louis City Missouri 32.2 10 Humphreys Mississippi 32.3 18
Geographic Branch. (2012, September 01). TIGER/LineÂŽ with Selected Demographic and Economic Data - Geography - U.S. Census Bureau.
Highest 10 Counties Rank County State Score 1 Nemaha Kansas 63.8 2 Stark North Dakota 63.6 3 Cedar Nebraska 63.2 4 Bottineau North Dakota 62.2 5 Crook Wyoming 61.7 6 Williams North Dakota 61.6 7 Red Lake Minnesota 61.3 8 Boone Nebraska 61.3 9 Hutchinson South Dakota 60.2 10 Dewey Oklahoma 60.1 Retrieved from https://www.census.gov/geo/mapsdata/data/tiger-data.html
Analysis of Intergenerational Income Mobility
North Carolina. The counties that have among the highest scores appear to be generally located in the center of the country, in states such as Texas, Kansas, Nebraska, Wyoming, Iowa, and Minnesota. Counties on the West Coast and in the Northeast tend to have middling mobility scores. Figure C.4 lists of the counties with the lowest scores, and those with the highest scores in the country. Interestingly, South Dakota has counties on both lists, which illustrates the high level of mobility variation that can exist within a small geographic region.
IV. Models We built two types of models: regression and classification. The regression models were useful for seeing how factors can affect mobility scores at a fine level. However, if we think of the model as a tool for identifying at-risk counties, a regression model may provide more detailed than needed. Thus, we created a classification to identify counties as having either “low” or “high” mobility. We examined mobility scores using quantiles and categorized counties in the bottom third (below 40.8 points) to be “low”, while the remaining counties were categorized as “high.” Of course, being above the bottom third does not mean mobility is high. Rather, the term “high” acts as shorthand for not being among the lowest. We split our data into a training and test set, with 80 percent of the data used in the training set (2,214 counties) and the remaining 554 counties held for the test set.
Random Forest We ran a random forest regression using randomForest and tidyverse packages in R, regressing the mobility score on all the predictor variables. We determined the minimum number of trees needed to minimize the error. We also calculated the importance of each variable. The random forest model achieved the lowest standard deviation of the cross-validated RMSE estimates (2.33). Lasso We also fitted a Lasso regression on our dataset to perform regularization and variable selection. Without specifying a lambda value for our model, we ended up with model fits for 84 lambda values. We performed ten-fold cross validation on the fitted lasso model to find the value of lambda that would give the lowest cross-validated error. This resulted in a lambda value of 0.175. However, to reduce the model’s complexity and chances of overfitting, we used the lambda value 0.278, which is within one standard error of the minimum error. When we ran the model on the held-out test set, we got a RMSE of 2.52. This means that our lasso model leads to test predictions that are within 2.52 points of the true upward mobility score. Figure D.1: RMSE Comparison Across Models Model RMSE Default 5.26 Random Forest 2.33 Lasso 2.52 Regression Tree 3.26
Regression Models Classification Models We ran several regression models on our dataset. A summary of root mean squared error (RMSE) for each model is shown in Figure D.1. All of our models performed better than our default model, as shown in Figure D.1. The best performing model was the random forest model, followed by the linear model and the lasso. As the regression tree was our lowest performing model, we will not discuss it here.
In our classification models, we chose to predict which counties have “low” mobility rather than “high” mobility, with the idea that this can potentially be a tool to determine which counties have a greater need for support. Decision Tree We used the rpart package in R to build a decision tree used to classify counties as “low” or “high” in mobility. It has built-in cross-validation, which
The Heinz Journal
Figure D.2: Decision Tree Plot (Classification)
Analysis of Intergenerational Income Mobility uses a technique to evaluate predictive models by partitioning the original sample into a training set to train the model, and a test set to evaluate it.19 Since the tree is fairly shallow and performs well on test data (as will be shown in a later section), it did not appear necessary to prune the tree. The accuracy of test data on predicting “low” mobility counties turns out to be 84.3%, which is substantially above the prevalence rate of “low” counties (33.3%). The tree splits first on fraction of black population, and then considers fraction of children with single mothers, teenage birth rate, income growth, population density, test score percentile, median house price, fraction of population foreign born, population density, and school expenditures per student. Nodes 3 and 25 are particularly interesting, as the likelihood of the counties in these nodes being low or high is strongly dependent on two variables. For
the 285 counties in Node 3, which have a black population fraction greater than 0.13 and a fraction of single mothers greater than 0.257, nearly all are classified as “low” mobility. In contrast, for the 848 counties in Node 25, which have fraction of population black less than 0.13 and teenage birth rate less than 0.146, almost all are classified as “high” mobility. Random Forest We also built a random forest classification model. We used 10-fold cross-validation to tune the random forest on the number of predictors to consider at each split in a tree (m), experimenting with m values of 4, 6 (rounded-down value of the square root of the number of our predictors), 8, 10, and 41 (the total number of predictors, essentially equivalent to a bagging model). We left the number of trees grown constant at the caret package default of 500.
Figure D.3: ROC Curve Comparison with Decision Tree and Random Forest
Figure D.4: Confusion Matrices for Classification Models 19
Vanschoren, Joaquin. "OpenML." OpenML: exploring machine learning better, together. https://www.openml.org/a/estimation-procedures/1.
The Heinz Journal Confusion Matrix: Random Forest Observed Predicted Low High Low 148 24 High 41 341 Total 189 365
Total 172 382 554
Performance of Classification Models Figure D.3 shows that both the random forest and decision tree perform substantially better than an entirely constant prediction of assigning all counties to the majority class (“high”). However, the random forest still outperforms the decision tree. Next, we built confusion matrices for both models using a cutoff score of 0.5. The results are shown in Figure D.4 Using these figures, we calculated that the random forest has an accuracy of 0.883 and sensitivity of 0.934, while the decision tree has an accuracy of 0.843 and a sensitivity of 0.915. It is surprising that the decision tree performs so well, considering that decision trees are generally more known for interpretability rather than predictive accuracy.
V. Discussion Key Findings While Chetty et al. do not strive to demonstrate causality between covariates and mobility, they do identify several strong correlations and recommend them to other researchers as factors to explore further. To find correlation, they used ordinary least squares regressions, with variables normalized to a mean of 0 and a standard deviation of 1.20 Figure E.1 shows the five most important predictors for each model, with their order of importance identified by their color. Only variables that are in the top five for at least one of the models are shown. The lightest shade of green indicates that the variable was in that model’s list of variable importance but outside its top five. For the random forest models, which provided information on variable importance for every predictor in the 20
Chetty, Raj, Nathaniel Hendren, Patrick Kline, and Emmanuel Saez. "Where is the Land of Opportunity?
Confusion Matrix: Decision Tree Observed Predicted Low High Low 133 31 High 56 334 Total 189 365
Total 164 390 554
dataset, only those where the percent increase in the mean-squared error (MSE) was above average are shaded as important. When comparing variable importance between our models and Chetty et al., it is important to note that there are several variables we consider that they did not use. These are the variables shaded in gray. Our models found that these variables were important. For example, several models find that income growth in counties, population density, and teenage birth rate are important predictors; however, because they were not included in Chetty et al.’s regression model, they were not evaluated for correlation. Across all the different models however, it is clear that the fraction of children with single mothers living in an area is a very strong predictor for mobility. Since fraction of children of single mothers emerges as such a strong predictor, it is worth investigating how a model would perform in the absence of that variable, in order to discover what other variables would increase in importance. We ran our best-performing regression model, the random forest, on the training dataset with the fraction of children of single mothers omitted. The RMSE is 2.350, only a little bit higher than the RMSE with the variable included (2.329), indicating that there is not too much loss in predictive ability. We also find that fraction of population black rises to be the most important variable when fraction of children of single mothers is removed. Potential Uses In addition to shedding more light on what countylevel factors are important to mobility, these models
The Geography of Intergenerational Mobility in the United States." 2014. doi:10.3386/w19843.
Analysis of Intergenerational Income Mobility
Figure E.1: Comparison of Variable Importance by Model
Chetty et Lasso 1-SE Decision al. paper (R) Tree (R)
Random Forest Subsetted without Random Fraction of Single Decision Forest (R) Mothers (R) Tree (C)
Random Forest (C)
Fraction of children with single mothers Fraction middle class Gini coeffcient for bottom 99% households Social capital index Teenage labor force participation Income growth Fraction black Population density Fraction of adults divorced Teenage birth rate Fraction with commute under 15 minutes Fraction religious High school dropout rate Median house price Fraction of adults married Tax progressivity Migration inflow Racial segregation School expenditure per student Violent crime rate
Legend Importance rank 1 Importance rank 2 Importance rank 3 Importance rank 4 Importance rank 5 Importance rank 6+ Variable not considered
The Heinz Journal Figure E.2: Distribution of Absolute Mobility Score Among Counties Receiving CBDG
have the potential to be developed further, in order to help agencies better target their funds intended to promote economic opportunity. For instance, one of the U.S. Department of Housing and Urban Development’s (HUD) most prominent programs is its Community Development Block Grant program (CDBG), which makes annual grants aimed at enhancing services for vulnerable populations, and creating greater economic opportunities for lowand moderate-income individuals. Most of these grants are made on an entitlement basis, which means that grants are allocated non-competitively to entities based in part on a formula. This formula includes measures of community need, which CDBG named as “extent of poverty, population, housing overcrowding, age of housing, and population growth lag.”21 To assess the performance of CDBG’s formula at targeting the areas that the grant seeks to assist, we looked at the 154 counties that received a grant in 2003 (the earliest year that the necessary data was available on the HUD website) and their mobility scores. What we find is that the scores of these counties range from 33.0 to 52.4, with a mean of 41.9, which is slightly greater than the cutoff we created for “low” mobility counties. The distribution of mobility scores of counties receiving a grant is shown in Figure E.2. While there are many factors
that contribute to grant awards, our model could be developed further to help identify which counties have the greatest need, in terms of economic mobility, to direct assistance.
Housing and Urban Development (HUD). https://www.hud.gov/program_offices/comm_planning/ communitydevelopment/programs.
Community Development Block Grant Program CDBG/U.S. Department of Housing and Urban Development (HUD) | HUD.gov / U.S. Department of
Limitations Figure E.3 shows a comparison of the ten counties with the lowest mobility scores in the test set and the ten counties with the lowest predicted mobility scores for each of the regression models. Curiously, Shannon SD, the county with the lowest mobility in the test set, does not appear on the bottom ten lists for either model. In fact, the bottom ten list for the Lasso model includes just half of the true bottom ten counties, and the list for the random forest model includes only three of the true bottom ten counties. This comparison suggests that these models do not perform well at predicting extremes. Another important limitation is that the predictor values need to be refreshed with regular frequency (for example, annually, for grants awarded each year). However, some of the predictors in this dataset were carefully constructed by the researchers, sometimes borrowing from the work of other researchers (the social mobility index, fraction of population religious, EITC exposure, etc.). Other important predictors however are
Analysis of Intergenerational Income Mobility Figure E.3: Comparison of Counties with Lowest Mobility Scores Lowest 10 Counties - Test Set Rank 1 2 3 4 5 6 7 8 9 10
County Shannon Todd Wilkinson Quitman Scotland Fulton Crisp Sunflower Anson Twiggs
State Score South Dakota 23.7 South Dakota 24.3 Georgia 32.3 Mississippi 32.7 North Carolina 32.9 Georgia 33.1 Georgia 33.3 Mississippi 33.4 North 33.6 Georgia 33.7
Lowest 10 Counties - Lasso Actual Predicted Rank County State Residuals Score Score 1 Holmes Mississippi 35.6 30.7 4.9 2 Crisp Georgia 33.3 31.5 1.8 3 Todd South Dakota 24.3 31.9 -7.6 4 Fulton Georgia 33.1 32.2 0.9 5 Sunflower Mississippi 33.4 33.1 0.3 6 Quitman Mississippi 32.7 33.2 -0.5 7 Yazoo Mississippi 36.1 33.5 2.6 8 Lee South Carolina 34.3 33.5 0.8 9 Danville City Virginia 34.9 33.7 1.2 10 Peach Georgia 35.4 33.8 1.6
Lowest 10 Counties - Random Forest (Regression) Actual Predicted Rank County State Residuals Score Score 1 Sunflower Mississippi 33.4 34.0 -0.6 2 Crisp Georgia 33.3 34.5 -1.2 3 Peach Georgia 35.4 34.5 0.9 4 Holmes Mississippi 35.6 34.5 1.1 5 Lee South Carolina 34.3 34.6 -0.3 6 Haywood Tennessee 34.1 34.9 -0.8 7 Hinds Mississippi 34.3 34.9 -0.6 8 Scotland North Carolina 32.9 35.1 -2.2 9 Yazoo Mississippi 36.1 35.2 0.9 10 Danville City Virginia 34.9 35.2 -0.3
updated on a regular basis. The Current Population Survey, sponsored by the Bureau of Labor Statistics and conducted monthly by the U.S. Census Bureau22, collects data on variables, such as fraction of children with single mothers and teenage labor force participation.
VI. Conclusion Through this paper, we have shown that there are several factors that our models consider to be important when attempting to predict a countyâ&#x20AC;&#x2122;s mobility, most notably the fraction of children with single mothers. This finding is crucial since it opens the door for a plethora of more pointed future research into understanding the role of these variables better. We also find that some of the variables Chetty et al. had not included in their analyses were quite important, such as income growth in county, and population density. This information can be useful to other researchers looking for starting points in determining what geographic characteristics are causal to mobility. The classification model holds the potential to be developed further to assist government agencies in accurately targeting counties that are most in need of funding and support to create better economic opportunities for their residents. Together, these efforts can help identify impediments to economic mobility, direct efforts towards the areas most in need, and maximize derived economic potential.
"Current Population Survey (CPS)." U.S. Bureau of Labor Statistics. https://www.bls.gov/cps/.
The Heinz Journal
Big Data on a Big New Market: Insights from Suppliers and Customers in Washington State’s Legal Cannabis Market Jonathan P. Caulkins, Yilun Bao, Imane Fahli, Yutian Guo, Krista Kinnard, Mary Najewicz, and Lauren Renaud Carnegie Mellon University – Heinz College of Information Systems and Public Policy Eight states in America have legalized the large-scale commercial production of cannabis for non-medical use, and more information is becoming available about the legal marijuana market at all levels of the supply chain. This study uses data visualization, text analytics, and difference-in-difference methods to examine the legalized marijuana market in Washington State. The main analyses explore the price and potency trends of extract products, the relationship between retail and wholesale prices, and potential market preferences for college students.
I. Introduction Over the past several years, policies surrounding the production, sale and use of cannabis have become less strict in many countries in Europe and the Americas. In November 2014, voters in Colorado and Washington State approved propositions making them the first jurisdictions to state-legalize not just home cultivation and use, but also large-scale commercial production and sale of cannabis products for recreational use. After a period of regulatory design, the first licensed stores opened in January 2014 (in Colorado) and July 2014 (in Washington). In Washington State, the State Liquor and Cannabis Board (LCB) licenses and regulates the producers, processors, laboratories and retailers. The LCB also manages a “seed-to-sale” database tracking cannabis products from growers to laboratory testing, processors and ultimately retail stores. This tracking provides data that can inform policymakers of the overall dynamics of the market, elucidate the demand for certain products and consumer markets.1 The rise of this market has been a challenge for regulators. They have had to quickly design policy and regulations to best track and monitor all marijuana transactions. The regulatory role of the 1
Washington State Liquor and Cannabis Board. Medical Marijuana FAQ's; 2016.
state of Washington in the overall cannabis market depends on understanding this emerging market segment. Regulators are tasked with decisions such as assessing how many licenses to issue, what is the best way to tax these products, and many other important choices. With the speed of this regulation, policymakers were forced to make many decisions based on limited knowledge stemming mostly from the illicit marijuana market. Now that we have more facts and evidence about the market, we hope to inform policymakers’ subsequent decisions using data analytics. The seed-to-sale tracking system is not a sample; it provides the universe of data on all legal commerce in the state, which has generated about 25GB of data over the past three years. This analysis takes advantage of the depth and breadth of these data to explore details of this new legal market. This analysis, in turn, can inform research that aids policymakers in decision making. This paper explores three trends in particular: 1. Partitioning of the broad product category “extracts for inhalation” into more insightful subgroupings; 2. The relationship between wholesale and retail prices for usable marijuana and extracts for inhalation; and 3. Evaluation of seasonality in sales in college towns, as a window into consumption patterns Weekly Marijuana Report, December 12, 2016; 2016.
Big Data on a Big New Market of this particularly interesting consumer market segment. These analyses help form a preliminary understanding of marijuana product types, market price mechanism and a special group of consumers, and provide clues for future analysis.
II. Data and Measures Washington State’s “seed-to-sale” tracking system was designed to capture all transactions and conversions of marijuana products as they move from producers to processors to labs and retail stores. Item-Entry: We call the standard unit of analysis in this paper an “item-entry” and not a “transaction” because one purchase can produce multiple observations in the data set.2 For example, if a customer bought two grams of one type of cannabis flower and one gram of another at the same time, those would show up as separate observations. However, multiple copies of the same item can appear within a single observation in this data set. For example, if that person bought two separate one gram packages of the first type of flower, that could appear as a single $20 observation with a “usable weight” of 2 grams and a “2” in the “weight” field which, for retail transactions indicates the number of items in that item-entry. Nonetheless, going forward, we will abbreviate “item-entry” to “item” for brevity. Price: Each observation reports the price paid to the seller by the buyer, whether the buyer is a retail consumer, store owner, processor, etc. In July 2015, Washington changed from a 25% tax at each step of the production process to a single retail excise tax of 37%.3 The pre-July 2015 observations include those taxes, while the post-July 2015 data do not. Following Smart et al. (in submission), we increase retail prices after July 2015 by 37% to consistently reflect the effective cost to the buyer. Prices are expressed in dollars per gram, calculated as the sale price divided by the usable weight of the cannabis. Price mark ups are calculated by
Smart, et al., in submission. Washington State Department of Revenue. Taxes Due on Marijuana; 2016.
comparing the price per gram a retailer paid for a product and the price it was sold for to a consumer. Potency: Potency is defined as the amount of “Total THC” in the product. This variable is calculated as plain THC, which is not decarboxylated, plus 0.877 times THC-A, which is decarboxylated.4 Product Type: The data set contains a variable “inventorytype” that distinguishes thirty-three product types. This analysis primarily focuses on the most common three: “usable marijuana” which refers to traditional flower with minimal processing, “edibles” which refers to cannabis infused food and drink products, and “extracts for inhalation” which refers to a wide range of processed cannabis products, including but not limited to wax, kief, shatter, oils, and distillates for portable vaporizers. Note: We frequently do not test below for the statistical significance of differences for the simple reason that the data represent the universe of all legal transactions in the state. They are not a sample drawn from some larger population. So, for example, when we report that the proportion of all extracts in two college towns (Pullman and Airway Heights) that are of the shatter/wax/dab variety is much higher (~65%) than in some other places (2435%), we do not compute a p-value to assess whether those differences are statistically significant. Those proportions are simply facts, not sample statistics.
III. Analysis and Results We conduct several analyses. First, we create a typology of extracts using text analysis and look at the relationship between price and potency for these groups. We then analyze trends between retail and wholesale prices over time using a multiplicative model. Finally, we analyze seasonal trends in Pullman, Washington to explore the purchasing habits of college students compared to the general population and explore whether these trends are generalizable to other college towns. 4
Smart, et al., in submission.
The Heinz Journal Analysis of “Extracts for Inhalation” One prominent trend in the Washington legal cannabis market observed by Smart et al.) among others is the increasing proportion of the market that is made up of extracts for inhalation (hereinafter “extracts”). Flower transactions continue to be the largest part of the market, making up about 75% of transactions as of 2016, while extracts come in second and account for 12% of transactions. This differs from some other legal marijuana markets such as Colorado where extracts appear to be less popular.5 Smart et al. analyze trends in price and the relationship between potency and price for usable
marijuana (i.e., flower) products but not extracts, in part because the extracts category is a combination of different product types. Below, we extend portions of that analysis to extracts for inhalation both as a broad class and broken down into some constituent parts. Washington’s seed-to-sale database does not distinguish between different extract products. For example, cartridges and wax are both listed as inhalants, though they differ in terms of price, potency, and modality of use.6 Figure 1 plots the average price vs. average potency for each product type within the extract category in the month of June 2016. There is one plotting point for each unique product name, not one point for each item
Figure 1: Statewide Average Price and Potency Levels of Extract Products in June 2016
Daniulaityte R., Nahhas R. W., Wijeratne S., Carlson R. G., Lamy F. R., Martins S. S. et al. “Time for dabs”: Analyzing Twitter data on marijuana concentrates across the US. Drug Alcohol Depend 2015; 155: 307311. 6 Morean M. E., Kong G., Camenga D. R., Cavallo D. A., Krishnan-Sarin S. High school students’ use of
electronic cigarettes to vaporize cannabis. Pediatrics 2015; 136: 611-616. Krauss M. J., Sowles S. J., Mylvaganam S., Zewdie K., Bierut L. J., Cavazos-Rehg P. A. Displays of dabbing marijuana extracts on YouTube. Drug Alcohol Depend 2015; 155: 45-51.
Big Data on a Big New Market sold. Few patterns can be discerned. Therefore, we try to exploit the presence or absence of key words in the free text product name field to find more homogeneous sub-groups within this broad product category (after excluding the relatively small number of high-CBD observations, defined here as those with greater than 1% CBD, which tend to be more expensive and have lower THC potency than do the THC products). We identified eight subcategories that were distinct (i.e., relatively few product names spanned categories). For each we developed a set of search words. E.g., “shatter” or “budder” for shatter type products. It is important to note that we used a search function that accounted for a “wildcard” in front and behind the search word. This ensures that when we search on “wax” we can grab products called “earwax” and “wax-pucks.”
In general, only products whose names included search words for one category and not search words associated with any other category were categorized. However, special rules were written for some common overlaps. E.g., “hash oil” was placed with other oils, and “cartridge oil” with other cartridge observations. However, a product that has both “hash” and “cartridge” in its name is left uncategorized since it is unclear what kind of product it is. Some examples include “X-Tracted Hashplant Wax .5g,” “Fire Alien OG Live Resin Wax (0.5g),” and “DD Purple Jolly Rancher Dab Oil .5g.” (Each of these examples had fewer than 20 transactions in June 2016.) Approximately 74% of all product names (accounting for about 63% of all extract transactions) could be placed within a specific product type, with the remaining observations in the “other” group (see Table 1).
Table 1: Categorization Schema and Three Most Popular Product Names Within Each Category Based on June 2016 Retail Sales Category
Three most common products within category
Proportion of obs.
cart, vape, pen, vc, refill
The Clear Cartridge; Liberty Reach, 0.5g PURE Vaporizer Cartridge, Blue Dream; Willy's Wonder .5ml Cartridges
oil, rso, eso
Jesus (.5g) Oil; Berry Haze (.5g) Oil; Pineapple Super Silver Haze (.5g) Oil
Bubble Hash .5g; Sugar Hash - 1g; Monk Hash - 1g
Kief, BSH Kief 1g; BSH Kief 1g
Lucid Dabs (1g); Dabulators 0.25g; Dabz - Mt Rainier #10
Wax 1g; Blue Dream SugarWax; Supergirl Wax 1g
1g Girl Scout Crumble (grow state); Concentrate: BHO Shatter 1g; Wa Woo Cookie Shatter
Pineapple Express Live Resin (.5g); Tangie Live Resin (.5G); Middlefork Live Resin (.5g)
Dutch Hawaiian Frost R.IO6013z 0.5g Atomizer; The Clear Concentrate; Jedi Kush
The Heinz Journal Figure 2 replicates the display in Figure 1 for the eight separate categories of inhalant products. The average price per gram for an inhalant product is around $55. This is substantially higher than the average price per gram for flower products. Smart et al. reports average potency for flower products of just over 20%, compared with 67% THC for extracts. Some additional patterns in price and potency emerge with this new categorization. We explored trends over time in price and potency for all nine categories including “other” (not shown). Broadly speaking, categories that looked similar in Figure 2’s price vs. potency scatter plot also tended to have similar trends in price and potency over time. For example, hash and kief both experienced decreases in potency over time. Another example is that wax, shatter, and resin all appear to have high potency rates that are stable over time.
Based on these patterns, and also the way these products are often presented in store menus, three larger groupings emerge. It seems sensible to group cartridge and oil observations together as they are similar products and cartridges are usually filled with oil or a distillate. Another grouping combines wax, shatter, dabs, and resin as they are products that are similar in price and have the highest potency. Hash and kief are grouped together likewise due to similarities in products, price, and potency. Subsequent analysis uses these three larger groupings, as described in Table 2. The largest proportion of the extract market is comprised of products such as dabs, shatter, resin, and wax (55% of the categorized observations, which include roughly two-thirds of all extract observations), with the 2nd largest proportion being cartridge/oil products at approximately 39% of the market (Table 2).
Figure 2: Average Price and Potency for Extract Products by Extract Category, June 2016
Big Data on a Big New Market Table 2: Extract Market Broken Down by Product Category for June 2016 Observations with matched product names only. Average Price
% of Market
The dab/wax/shatter/resin group is also the fastest growing. In June of 2015 the average number of transactions per store per day of a wax/shatter/resin product was 5. One year later, in June 2016, that average had jumped to 17. Cartridge/oil transactions also increased, but more in line with the growth rate of the overall legal cannabis market in Washington (from 7 to 12 transactions per store per day). This trend illustrates the potential value in partitioning extract observations. The number of extract transactions overall grew by a substantial 50% (from 15 to 30 transactions per store per day), but if someone who had particular concerns about the health consequences of dabbing only had access to that figure, they would have underestimated the 340% growth in the dab/shatter/wax/resin submarket which includes products that are particularly associated with dabbing. Returning to trends over time, aggregate prices declined rapidly for all three of these groupings until the summer of 2015 and then began to level out at around $30 per gram for products such as wax, shatter, and resin. Prices for cartridges/oil fell the most. Average potency of the cartridge/oil category has been steadily increasing from 50% THC to closer to 75% THC over the last 2 years, but potency for the other two categories peaked and has since been 7
Smart, et al., in submission.; Caulkins, Jonathan P. (2017). Recognizing and Regulating Cannabis as a Temptation Good. International Journal of Drug Policy, 42:50-56.
decreasing slightly, albeit at quite different levels. Wax/shatter/resin potency levels remain around 65% THC, whereas hash/kief have significantly lower potency. Trends Over Time in Wholesale and Retail Prices for Usable Marijuana and Extracts In this section, we analyze the change in retail prices in cannabis over time. Using the Seed-toSale database, we find that in accordance with many other studies, retail prices have fallen and revenue prices have increased. We find that price trends in the marijuana market tend to follow a multiplicative model, where retail prices are a fixed multiple of wholesale prices. Retail prices of cannabis products in Washington State have fallen significantly since legalization.7 This may be the result of some combination of new production and processing technologies, mastery of existing processes, economies of scale, and greater competition, including at the retail level. Policymakers pay special attention to prices because they directly impact tax revenue and indirectly affect health outcomes by influencing consumption. The elasticity of demand for marijuana is thought to be somewhere in the neighborhood of -0.5, meaning that 10% decline in retail price would increase consumption by 10%.8
Gallet, Craig A. (2014). Can price get the monkey off our back? A meta-analysis of illicit drug demand. Health Economics, 23:55-68, published online in 2013 at DOI: 10.1002/hec.2902.
The Heinz Journal Experts have argued that legalization should greatly reduce cultivation costs.9 Indeed, wholesale prices have been falling throughout the period of policy liberalization dating at least to the so-called “Ogden memo,”10 and so well before formal statelegalization of production for a non-medical market. Before policy began to liberalize, highquality domestically-produced cannabis sold at wholesale for around $6,000 per pound.11 By November 2016, the Cannabis Benchmarks spot index - a commodities price index for the American market - had fallen below $1,400 per pound. Aphria, a Canadian producer, reports production costs of $400 - $700 per pound in US dollars at current exchange rates. Aphria now produces 60 grams per square foot in its greenhouses. If production costs fell to that typical of tomatoes grown in greenhouses – roughly $4 per square foot – that would be only $30 per pound. In Washington State, while prices are going down, revenues are going up. This increase in revenues could be explained by an expanding consumer base or an increase in stores, as explained at the beginning of this section. Usable marijuana makes up the largest portion of revenues, while being the cheapest product type. Extracts for inhalation are pricier, make up the second largest portion of revenues and appear to be growing proportionally over time.
enforcement on suppressing consumption.12 Caulkins laid out two extreme models that may bracket the actual relationship. Under the additive model, retail price equals wholesale price plus a constant, so if wholesale prices fall by $1 per gram then so will retail prices. Under the multiplicative model, retail prices are a fixed multiple of wholesale prices, so if wholesale prices fall by 10% then so will retail prices.13 Technical details about Washington’s seed-to-sale database complicate finding the price a producer (farmer) was paid by a processor for a particular unit of cannabis, so the present analysis is confined to the retail price per gram the customer paid the retail store and the (wholesale) price per gram the retailer paid the processor for both usable marijuana and extracts for inhalation. The following graphs plot wholesale and retail average prices by quarter, with a 45-degree dashed line, a line representing a constant 3:1 ratio of retail to processor prices and a linear regression line. Figure 1 presents price data for usable marijuana, since the market was legalized. Figure 2 presents price data for extracts for inhalation starting from Q2 2015 when this category began to gain significant market share.
What is unclear at present is how such a radical decline in production costs, if it were to transpire, would affect pre-tax retail prices. There is a long history of asking how changes in prices further up the distribution chain may filter down to affect retail prices because it bears directly on estimates of the effectiveness of interdiction and high-level
Prices tended to fluctuate initially after legalization so the lighter dot, representing Q3 2014 for usable marijuana and Q2 2015 for extracts for inhalation, is away from the trend that seems to form later. As the market matured, prices steadily decreased by more than half for usable marijuana and a third for extracts for inhalation. However, the ratio of retail price to wholesale price consistently maintained a roughly 3:1 ratio for both product types.
Kilmer, at al., 2013.; Caulkins et al., 2016. Ogden, Davie W. 2009. Memorandum for selected United State Attorneys on investigations and prosecutions in states authorizing the medical use of marijuana. Available from the U.S. Department of Justice Archives at https://www.justice.gov/archives/opa/blog/memorandu m-selected-united-state-attorneys-investigations-andprosecutions-states. 11 Western States Information Network (WSIN). 2008. Illegal Drug Price & Purity Guide. 10
Reuter, Peter and Mark Kleiman Risks and Prices: An Economic Analysis of Drug Enforcement, in Morris and Tonry (eds.) Crime and Justice: An Annual Review of Research Volume 7, Chicago: University of Chicago Press. 1986, pp.289-340. Kleiman, Mark AR. (1992) Against Excess: Drug Policy for Results. Basic Books, New York. 13 Caulkins, J.P. (1990). The Distribution and Consumption of Illicit Drugs: Some Mathematical Models and Their Policy Implications. Doctoral dissertation, MIT, Cambridge, MA. Caulkins, 2007.
Big Data on a Big New Market Figure 3: Relationship Between Retail and Processor Prices for Usable Marijuana (Left Panel) and Extracts (Right Panel)
Figure 4: Relationship Between Processor Prices and Retail Prices for Extracts for Inhalation, per Gram, by Subcategory
As noted above, extracts for inhalation is a heterogeneous category encompassing products from kief to wax and shatter, so Figure 4 replicates the plot after partitioning extracts into several more
homogeneous sub-categories. Although these subcategories vary in both price levels and the degree to which prices decreased over time, the 3:1 ratio still holds reasonably well for all of them.
The Heinz Journal The curiously stable 3:1 ratio of retail to wholesale prices merits further analysis and replication. It appears to hold across a range of products, but obviously all of our data come from just one jurisdiction and time period. Perhaps the stability of that ratio reflects not something intrinsic to the products or their production technologies, but rather something unique to Washington State’s regulatory practices, or even the oddities of tax law (Section 280(e)) during this strange time when the cannabis business is state-legal but still fully prohibited under federal law. In conclusion, price trends in Washington State’s legal cannabis market appear to be more consistent with a multiplicative model to date. It is tempting to leap to the conclusion that if, as expected, continued declines in production costs further depress wholesale prices, this might translate into proportional reductions in pre-tax retail prices. However, it is important to remember that no causal inference can be drawn from the stable-to-date ratio about the causal relationship between wholesale and retail prices. Prices at the two market levels may have declined in lock step simply because parallel forces (efficiency gains, increased competition, etc.) operated independently but with equal force to date at the wholesale and retail market levels. Analysis of Seasonal Variation in Sales in a College Town Our final analysis centers on Pullman, Washington, home of Washington State University, in order to better understand consumption patterns of college students. The analysis compares Pullman to a group of non-college, non-urban (NCNU) towns. First, we review 2016 transactions by product type to conclude that there is a noticeable jump in sales across all three product types (usable, edibles, and extracts) in Pullman when students return for the fall semester compared to steady rises in NCNU towns and statewide. Second, we further analyze the proportion of product type sold to conclude that 14
Elmer-DeWitt, Philip (2010). Big Macs on campus. Fortune. Downloaded on April 21st, 2017 from http://fortune.com/2010/08/07/big-macs-on-campus/.
students seem less likely to buy extracts and more likely to buy edibles than non-student residents. Third, we conduct a regression analysis to confirm this difference in purchasing habits between students and non-students. Fourth, we conduct parallel analyses in two additional college towns to conclude that the overall sales results from Pullman do not appear to be generalizable to other college towns. Finally, we investigate the variation in the type of extracts purchased in these other college towns to conclude that even if college students are overall less likely to use extracts, they may not necessarily be less likely to dab. College students are of particular interest because they are often predictors of future consumption trends. For example, reports that college students were rapidly switching from Dell to Apple computers may have triggered a sell-off of Microsoft stock.14 So this section searches for hints about possible future consumption patterns by contrasting legal cannabis sales in so-called “college towns” with other places, when school is in session versus summer holidays when most students are away. Purchasing patterns of college students are also of direct interest because of concerns about how cannabis may affect brain development in younger people.15 College towns – smaller, relatively isolated towns with large student populations – could present a useful natural experiment since college students might be expected to account for a sizable share of demand during the academic year and many, if not most, leave those towns during the summer. This allows for comparison of the market between the times when students are present and when they are absent. This analysis focuses on Pullman, which has the greatest estimated ratio of college students to other residents. Pullman is home to Washington State University (WSU), with an enrollment of 20,043 students, while the US Census (2010) records the population in Pullman to be just 29,799.16 15
Hall, Wayne, and Michael Lynskey. Evaluating the public health impacts of legalizing recreational cannabis use in the United States. Addiction (2016). 16 Washington State University. Academic Calendar; 2016
Big Data on a Big New Market Pullman is compared to a group of non-college, non-urban towns (NCNU) towns. These towns have no colleges or have just small schools with a low student to non-student ratio, and have a population below 90,000 people.17 The analysis encompasses January 2016 to December 2016, which captures a spring semester, summer vacation and fall semester, and avoids irregularities that were present at the initial opening of the new market and the change to the tax structure which took place on July 1, 2015. Overview of 2016 Transactions by Product Type Figure 5 compares purchases in NCNU (top panel) and Pullman (bottom) for the major product categories over time. The vertical dashed lines
demarcate WSUâ&#x20AC;&#x2122;s summer break. Linear trend lines are drawn separately for the spring, summer, and fall semesters without connected spline points to allow for discontinuous jumps at semester transitions. Sales in NCNU towns generally increase steadily over time, as they do statewide, without any strong discontinuity in level or trend at the beginning or end of summer. The overall trend in sales is also increasing in Pullman. However, edible sales fell when students left for summer vacation while extracts rose slightly. For all three product types, there is a noticeable jump in sales when students return for fall semester. (Note the difference in yaxis scales for all NCNU collectively vs. Pullman.)
Figure 5: Sales Trends in Pullman (Bottom) and NCNU Towns Vertical dashed lines demarcate summer term, when many students are away. Note: the scales in these figures are different.
U.S. Census Bureau. American Fact Finder; 2017. Washington State University. Housing and Resident Life; 2016. 17 We considered comparing Pullman to a single nearby town with similar population that does not have a college. However, such a town was not found. Though
Clarkston, which is only about thirty miles from Pullman, is similar in size and does not itself have a college, there is a college just across the border in Idaho, where cannabis sale remains illegal. Therefore stores in Clarkston may directly or indirectly be supplying those students.
The Heinz Journal Before and After Comparison in Pullman Only Simply finding that stores sell less when they have fewer customers is not particularly revelatory. Therefore, in this section we focus further analysis on the proportion of products purchased that are of one type or another. Under a null hypothesis that students are similar to non-students in their purchasing patterns, the product mix would be consistent in both spring and summer.
The actual algebra is slightly more complicated because not every student leaves Pullman for the summer. There appear to be no official data, but calls to Washington State University’s Department of Student Affairs and Enrollment suggest that about r = 25% of students remain in Pullman for the summer. Any given product type’s share among purchases for students (S) and non-students (A) can still be found by solving the following equations:
The last day of final exams in WSU’s spring 2016 semester was Friday May 6th, but a comparative analysis of sales from Thursday the 5th to Saturday the 7th would suffer from three problems: (1) sales vary by day of week, with Saturdays consistently being busier than Thursdays, (2) sales on a single day in a small city are subject to some general variability, and (3) semester breaks are not that sharp. Some students leave early, some stay later, and exams themselves may alter purchase and use of intoxicants.
f · r · S + (1 – f) · A = (f · r + 1 - f) · Psummer f · S + (1 – f) · A = Pspring where f r Psummer Pspring
Instead, we contrast two 28-day spring and summer windows, with the spring window ending three weeks before the end of final exams to avoid the periodic effect around April 20th and the symmetric summer window starting three weeks later and observe some differences. For example, extracts’ share of all sales rose from 9.7% in the pregraduation window of Mar 19, 2016 to Apr 15, 2016 to 13.7% in the summer window of May 27, 2016 to June 23, 2016. If no students remained in the summer and students accounted for two-thirds of cannabis purchases during the school year, that would suggest that 13.7% of cannabis purchases made by Pullman’s year-round residents are extracts and the corresponding proportion for students is 7.7% since (2/3) * 7.7% + (1/3) * 13.7% = 9.7%. In other words, when students bought cannabis at a state licensed store, they would be only about 7.7% / 13.7% = 0.56 times as likely to buy an extract as are Pullman’s year-round residents. (Note: That is not the same as saying 0.56 times as likely as the general population, because permanent residents of college towns may not be representative of the general non-student population statewide.)
= Students’ share of purchases during the spring, = Proportion of students who remain in Pullman during the summer, = Observed market share of that product during the summer, and = Observed market share of that product during the spring = Non-student
Table 3 solves these equations assuming that students account for 55% of purchases during the school year and 25% of students remain in Pullman for the summer. The proportions for the three main product types (usable marijuana, extracts and edibles) sum to slightly less than 100% because there are some other, minor product types (e.g., “mixed marijuana”) omitted from the table. These point estimates would suggest that edibles account for almost twice as great a share of students’ consumption as they do for other residents, and extracts’ corresponding share is much smaller. These ratios are, however, sensitive to the somewhat arbitrary assumptions about students’ share of purchases during the school year (f) and the proportion of students who remain in Pullman during the summer (r). Table 4 shows how each of these quantities vary if we consider alternative values of those two parameters within the ranges of 0.4 ≤ f ≤ 0.7 and 0.2 ≤ r ≤ 0.25.
Big Data on a Big New Market Table 3: Product Categories Broken Down by Students and Other Residents in the Basic Scenario Pspring
S/A (ratio of S to A )
Table 4: Each Product Category Broken Down by Students and Other Residents in the Sensitivity Analysis with 0.4 ≤ f ≤ 0.7 and 0.2 ≤ r ≤ 0.25 S (student)
S/A (ratio of S to A )
From the sensitivity analysis, we draw the qualitative conclusion that students seem less likely to buy extracts and more likely to buy edibles. This conclusion is robust for a range of parameters, but the extent of the differences depend considerably on those parameter values. Furthermore, seasonality could threaten this comparison. Perhaps all sorts of people, student and non-student, tend to buy more extracts in summer and more edibles in spring. We try to address these potential confounders with a difference-in-difference estimator that compares changes from spring to summer in Pullman and other college towns with corresponding changes over the same time period in non-college towns. Using Regression to Compare Pullman to NonCollege, Non-Urban Towns A difference-in-differences model allows for comparison of the change over time between a treatment and a control location. The model suggests that our qualitative conclusion that
students seem less likely to buy extracts and more likely to buy edibles holds for Pullman when compared to NCNU towns is valid. Here the treatment is the presence of college students. Pullman is the treated location and the control location is non-college, non-urban (NCNU) towns. The outcome is the proportion of sales on a given day that are of a given product type, with three separate regressions for the three product types (usable, extracts, and edibles). A dummy variable is included to indicate whether that day is a Friday or a Saturday (days for which sales are generally higher statewide) but we also run the model for a one-week window three weeks before the end of the semester and three weeks after, so that both the treated and control periods have one observation for each type of day of the week.
The Heinz Journal The empirical model used is:
sales, which is why the outcome is the proportion of sales by product type, not the sales levels.)
Yct = α + β Pullmanc + γ Springt + δDD(Pullmanc × Springt) + π Fri_Sat + εct where Yct
= Sales proportion of a certain type of products in location c on date t; Pullmanc = Indicator for whether the location is Pullman; Springt = Indicator for the Spring Semester; Fri_Sat = Indicator for the day being a Friday or Saturday; δDD = Coefficient for the difference in differences. Under the null hypothesis that students have similar purchasing habits as other residents, the proportion of purchases that are of a particular type should not change whether students are present or absent. (The same would not be true for the absolute numbers of
For the assumptions underpinning the differencein-differences model to hold, the trends in the proportion of products must be similar between Pullman and NCNU towns during the summer (when there is no treatment) and different in the spring (when the treatment occurs). In Figure 5 above it appears there are indeed similar sales trends in the summer in both groups. However, that parallel does break down if higher order polynomials, not just linear models, are used to depict those trends. (See Figure 6 below for a quadratic fit.) The results in Table 5 suggest that stores in Pullman tend to sell a greater proportion of edibles and fewer extracts in spring, when students are present, than in summer, relative to the pattern seen in other locations.
Table 5: Impact of Students’ Presence on the Proportion of the Products Sold Y1 (% Usable) (Intercept)
(0.0050) -0.038 (0.009)
Y3 (% Edible) 0.102
Y2 (% Extract)
Big Data on a Big New Market Figure 6: Quadratic Fits to Sales Trends in NCNU Towns and Three College Towns The graph omits sales May 6, 2016 to June 9, 2016 and also August 22, 2016 to September 21, 2016 because the three schools have different academic schedules, and these buffers allow for comparison of time periods when classes are either in session or not in session at all three universities.
Replication with Other College Towns To assess whether effects observed in Pullman are generalizable, parallel analyses could be performed for the towns of Ellensburg (population 18,000), which is home to Central Washington University’s 11,000 students, and Airway Heights (population 6,000) which is 13 miles from Eastern Washington University’s 13,000 students. (The store in Airway Heights is the closest one to Eastern Washington University.)18 However, Figure 6 shows that the patterns in these other towns do not appear consistent with those seen in Pullman. For example, Figure 6 illustrates that while sales of edibles decrease in Pullman during the summer, they increase in Ellensburg. There also does not appear to be much of a difference in summer sales between Airway Heights and NCNU.
Other candidates for parallel analysis seem less promising either because of their proximity to larger metropolitan areas and/or because they are Christian
Furthermore, the trends within the untreated period (summer) in the treated locations (Ellensburg and Airways Heights) do not closely parallel the summer trends in the control locations (NCNU towns). That violates the assumptions that must hold for the difference-in-differences regression analysis to be valid. There are reasons why the results for Pullman may be valid and generalizable to college students, but nonetheless not be replicable for these two other college towns. For instance, both Central Washington University and Eastern Washington University follow a quarter, not a semester, system. However, that some initial attempts to shift time windows and otherwise address those differences did not permit a successful replication gives us pause. The effect observed in Pullman may be a “Pullman effect” and not a more general “college town effect.” This may arise from some colleges whose students may have different patterns of consumption of intoxicants.
The Heinz Journal idiosyncratic factors, such as decisions by stores in Pullman as to what types of products to stock, and result in this analysis being unable to capture differences in student vs. non-student purchase patterns. Variation in the Particular Types of Extracts Used in College Towns Since the assumptions underpinning the regression model are violated for the other college towns, we close by stepping back to basic descriptive statistics. As noted above, extracts are a heterogeneous category, and trends that hold for the category broadly may not hold for all subcategories. Thus, it is of interest to ask whether the “Pullman effect” of students’ perhaps shunning extracts holds for all types of extracts. Table 6 shows the proportion of total extract sales by type of extract in spring vs. summer of 2016 for
each of the three college towns, for the non-college non-urban group (NCNU), and also for urban cities.19 It appears that consumers in Pullman purchase more wax, shatter, and resin products and fewer cartridge and oil products than do consumers in both large cities and small towns. Figure 7 plots these proportions day-by-day with a second order polynomial fit. It shows a noticeable drop in the purchases of wax, shatter, and resin in Pullman in the summer that is not seen in urban areas or small towns. Additionally, this figure indicates lower use of those products in urban areas than in the rest of the state. This detailed breakdown suggests college students are less likely to overall, they may not necessarily be use wax/shatter/resin and, hence, be dab.
that even if use extracts less likely to less likely to
Table 6: Proportion of Extract Market by Location and Product Subcategory Average from Jan 1, 2016 to Apr 30, 2016 Urban
Wax, Shatter, Resin
Hash & Kief
Average from Jun 1, 2016 to Jul 30, 2016 Urban
Wax, Shatter, Resin
Hash & Kief
This analysis considers “big cities” to be those with a population over 90,000. They are Seattle, Spokane, 19
Tacoma, Vancouver, Bellevue, Everett, Yakima, and Renton.
Big Data on a Big New Market Figure 7: Proportion of Extract Usage by Location
IV. Discussion It has been common in the past to differentiate cannabis flower from more concentrated products produced by traditional “mechanical” extraction methods (termed herbal cannabis vs. resin in much of the world and marijuana vs. hashish in North America). Sometimes there were also special names for higher potency flower products (sinsemilla vs. commercial grade in North America and skunk or hydroponic vs. general herbal in Europe). However, that typology fails badly in Washington’s legal market. Almost all of the herbal product is high-potency (averaging over 20% THC), and hash and kief account for a quite small share of the rapidly growing extract market segment, which is dominated by solvent-based extraction (particularly if one views supercritical C02 as a solvent). Furthermore, edibles constitute a nontrivial share of sales, and there are other categories (e.g. marijuana “mixes”).
It may be that post-legalization, researchers and policymakers will need to think in terms of a family of cannabis products, akin to how we think of opioids, new psychoactive substances (NPS), and amphetamine-type stimulants (ATS), not a single drug “marijuana” the way it is possible to think of cocaine and cocaine markets. This paper took a step toward creating a typology for one important part of that family, namely extraction products. There are at least two broad reasons why the legal industry is embracing extracts. First, just as prescription pills can bring opioids to the masses who shun injecting drugs, extract-based products can reach customer segments that dislike smoking, perhaps because of the negative connotations of tobacco smoking. Second, now that the exaction machinery does not expose owners to the risk of arrest and seizure, there is no reason to discard all of the THC contained in leaves and other parts of the plant where the THC exists in less concentrated form. Since most of the plant’s weight is in leaves, not flowers, a considerable share of the intoxicant
The Heinz Journal appears in parts of the plant that could not easily be brought to market before legalization. Our guess is that Washington State today manifests only the beginning, not the culmination, of the changes in product form and marketing that legalization will bring. It is not hard to imagine additional product forms emerging that might shake up any current typology, e.g., “bundled” products that combine cannabis with alcohol, tobacco, or other intoxicants. So we likewise view the present effort as just the beginning, not the culmination, of attempts to partition the rapidly evolving cannabis product space. Put simply, the results of the college town analysis are puzzling. We have the paradox that students leaving Pullman for the summer appear to have a measurable effect on the mix of products sold, but not an overwhelming impact on the volume of products sold. Although it is possible to imagine why that might occur, we cannot be sure. For example, undergraduates, who are mostly under the age of 21 and cannot purchase directly from the state-licensed stores, might still purchase most of their cannabis through the black market. (Washington State is thought to have large-scale purely black market production for export to other states that have not yet legalized.) But smoking cannabis in a dormitory might be hard to conceal from residential hall monitors, so they might obtain (via friends who are over 21) edibles from the statelicensed stores for use in dorms. And if the students who remain over the summer live in off-campus housing because dormitories close over the summer, that might account for a decline in demand for edibles during summer months. Alternatively, it could be true that Washington State University students who leave Pullman for the summer might continue to create demand for sales from stores in Pullman if their friends who stay behind for the summer are buying in Pullman and sending the product to them. Indeed, if Washington State University has a greater proportion of out of state students than do either Central or Eastern Washington University, and students going elsewhere within the state buy locally over the summer, that purchase-by-proxy effect might help explain why seasonal patterns appear to differ in Pullman vs. Ellensburg and Airway Heights.
In sum, the natural experiment created by students leaving a college town for the summer may not be quite as “clean” an exogenous shock as one might assume. There may be other natural experiments— such as the effect of legal changes in neighboring states on sales in border towns—that might be worth exploring.
V. Conclusions and Further Work The conclusions of this analysis are clear. Legalization induces dramatic changes in cannabis markets, and seed-to-sale monitoring systems offer a valuable window into those important changes. That suggests that states should consider designing these systems in ways that facilitate not only routine administrative functions that apply to individual licensees (collecting taxes, ensuring compliance with testing requirements, detecting diversion, etc.), but also the monitoring of market-level aggregates that are of interest to public health. Although we did not delve into particular technical aspects of the data system, in many ways Washington State’s present system is not welldesigned to serve those functions. For example, the system collects information from many activities throughout the market, and the database does not make it clear how the data stored in one table is connect to data another table. This vastly complicates the ability to effectively track product from “seed-to-sale” or back. The other obvious and significant limitation of these data is the absence of individually-identifying information of the customer. If customers were known, then changes in consumption patterns between college-age purchasers and older individuals would be immediately obvious. One could also look at trajectories of purchasing patterns over time for individuals, and could investigate whether a smaller number of frequent purchasers account for a disproportionate share of sales. Indeed, it may be that the next generation of market analysis after exploiting seed-to-sale regulatory monitoring systems will be studying data from private companies’ “frequent buyer” programs since they, unlike the state, will want to connect
Big Data on a Big New Market different transactions associated with the same repeat customer.
if crimes commonly associated with alcohol go up or down on immediately after a store changes its menu of cannabis prices.
Nevertheless, seed-to-sale data systems such as the one used in Washington state allow for a variety of additional analyses. For one, the partitioning analysis could be replicated with edibles, a category that is in some respects as heterogeneous as extracts, including various solid products (cookies, caramels, granola, candy cubes, peanut butter cups, etc.) and liquids (lemonade, punch, droppers, and sprays). Analysis of weight, potency, and hence potency-adjusted price per unit is more complicated, but might inform later federal regulatory decisions concerning psychoactive food additives. Testing practices are another topic of interest. In Washington State, producers pick the lab that will test their products, and there is no auditing of reported test results by any independent third party. This creates incentives for labs to produce results that are desirable to the producer, such as inflating measured THC or repeatedly sampling until a batch passes a quality control test (e.g., concerning mold or pesticide residue). Malfeasance cannot be seen directly in these data, of course, but it is clear that some producers shop around, trying out various testing labs, and it would be interesting to see if switches from one lab to another seem more associated with tangibles like price and location or with test results. More generally, aspects of the supply chain above retail can be investigated in ways that have not previously been possible, such as yields per unit area and price markups between growers and processors, not just between processors and retailers. The structure of the industry and the database complicate those analyses, but their novelty might warrant that additional effort. Correlating these fine-grained sales data with public health outcome data on traffic crashes, emergency room visits, and the like is of considerable interest. These data record down to the minute when a sale was recorded so they may expose correlations that are hidden in data reported only monthly or quarterly. That detail might be particularly valuable for trying to tease out substitution or complementarity with alcohol, e.g.,