University of Southern California Dissertations and Theses

Essays in political economy and mechanism design

(USC Thesis Other)

Essays in political economy and mechanism design

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content ESSAYSINPOLITICALECONOMY ANDMECHANISMDESIGN by Saurabh Singhal A Dissertation Presented to the FACULTY OF THE USC GRADUATE SCHOOL UNIVERSITY OF SOUTHERN CALIFORNIA In Partial Fulllment of the Requirements for the Degree DOCTOR OF PHILOSOPHY (ECONOMICS) August 2013 Copyright 2013 Saurabh Singhal Acknowledgments The six years spent at graduate school at USC have been a wonderful learning ex- perience. This dissertation would not have been possible without the encouragement and support of my advisors, faculty and sta at USC, friends and family. First and foremost, I would like to thank my advisors Juan Carrillo and John Strauss. Juan drew me to exploring experimental economics, taught me to think scientically, patiently allowed \ve-minute" meetings to run over an hour and gave me freedom to explore my research interests. John pushed me to explore my interest in development economics and my work would have not reached this stage without his painstakingly detailed comments. John has stood by me personally and profes- sionally and his unwavering faith in my ability to be a good researcher has been most reassuring. I am also tremendously grateful to Prof. Jerey Nugent for his constant support, comments, professional and personal advice over the last six years. I also thank Isabelle Brocas, Simon Wilkie, Tridib Banerjee, Adam Rose, Yilmaz Kocer and Anant Nyshadham for their comments and Olga Shemyakina and Tilman Br uck for being supportive of my research agenda. I am grateful to Utteeyo Dasgupta and Subha Mani for their invaluable friendship and support during the job market. I thank N. Raghunathan and Saugato Datta for providing the initial push to pursue further research and Rohini Somanathan for helping me develop my interest further during my time at the Delhi School of Economics. This journey would have been incomplete without the wonderful set of friends at USC. The endless cups of coee with Arya Gaduh, James Ng and Osman Abbaso glu ii over the last six years will be sorely missed. I am also grateful to Manuel Castro, Brijesh Pinto, Maggie Switek, Mehdi Majbouri, Zara Liaqat, Mohamed Saleh, Rahul Giri and Rubina Verma for their friendship and support. I thank Young Miller, Morgan Ponder and Christopher Frias for ensuring that I did not have to deal with administrative hurdles. I am also grateful for the various fel- lowships and funding provided by USC's College of Letters and Science, Department of Economics, and Graduate School. Last, but not the least, all of this would not have been possible without the love and support of my parents, Sunil Singhal and Neelam Singhal, my sister, Surbhi Singhal and my in-laws Sushum and Hema Sharma. Finally, but most importantly, I owe lifelong gratitude to Smriti Sharma, my strongest champion who has stood patiently by my side, through thick and thin over the years as my girlfriend, anc ee and wife. iii Table of Contents Acknowledgments ii List of Tables vii List of Figures ix Abstract xii 1 Introduction 1 1.1 Economic benets of counterinsurgency . . . . . . . . . . . . . . . . . . 1 1.2 Tiered housing allocation: an experimental analysis . . . . . . . . . . . 3 2 Naxalite Insurgency and the Economic Benets of a Unique Robust Security Response 1 6 2.1 Related Literature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 2.2 The Naxalite Movement . . . . . . . . . . . . . . . . . . . . . . . . . . 11 2.2.1 State Response and the Greyhounds . . . . . . . . . . . . . . . . 13 2.3 Research Methodology and Data . . . . . . . . . . . . . . . . . . . . . 15 2.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 2.4.1 Industrial Sector . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 2.4.2 Services Sector . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47 2.4.3 Agricultural Sector . . . . . . . . . . . . . . . . . . . . . . . . . 52 2.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57 1 This chapter is co-authored with Rahul Nilakantan. iv 2.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62 Appendices 2.A Data Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63 2.B Map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67 2.C Yearly Treatment Eects . . . . . . . . . . . . . . . . . . . . . . . . . . 68 3 Tiered housing allocation with pre-announced rankings: an experimental analysis 2 69 3.1 The model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76 3.1.1 Basic denitions . . . . . . . . . . . . . . . . . . . . . . . . . . 76 3.1.2 The tiered allocation mechanisms . . . . . . . . . . . . . . . . . 78 3.1.3 Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81 3.2 Experimental Design . . . . . . . . . . . . . . . . . . . . . . . . . . . 83 3.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87 3.3.1 Final allocations . . . . . . . . . . . . . . . . . . . . . . . . . . . 88 3.3.2 Eciency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89 3.3.3 Fairness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94 3.3.4 Participation Rates . . . . . . . . . . . . . . . . . . . . . . . . . 96 3.3.5 Truthful Preference Revelation . . . . . . . . . . . . . . . . . . . 103 3.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107 2 This chapter is co-authored with Juan Carrillo. v Appendices 3.A Additional analyses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109 3.A.1 Ordinal Eciency Test . . . . . . . . . . . . . . . . . . . . . . . 109 3.A.2 Preference Revelation: Full Ranking Analysis . . . . . . . . . . . 110 3.B Sample experimental instructions . . . . . . . . . . . . . . . . . . . . . 112 3.B.1 Instructions for tTTC mechanism . . . . . . . . . . . . . . . . . 112 4 Conclusion 125 4.1 Economic benets of counterinsurgency . . . . . . . . . . . . . . . . . . 125 4.2 Tiered housing allocation: an experimental analysis . . . . . . . . . . . 126 Comprehensive bibliography 128 vi List of Tables 2.1 pcNSDP Predictor Means . . . . . . . . . . . . . . . . . . . . . . . . 23 2.2 State Weights: pcNSDP . . . . . . . . . . . . . . . . . . . . . . . . . 23 2.3 Greyhounds and pcNSDP . . . . . . . . . . . . . . . . . . . . . . . . 32 2.4 State Weights: Industry . . . . . . . . . . . . . . . . . . . . . . . . . 40 2.5 pc Industry NSDP Predictor Means . . . . . . . . . . . . . . . . . . . 41 2.6 Greyhounds and Industrial NSDP in Naxalite Violence Aected States 1970-2000 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44 2.7 Greyhounds and Industrial NSDP in Naxalite Violence Aected States 1970-1997 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45 2.8 Results Summary: Greyhounds and Industrial Performance . . . . . . . . 45 2.9 Greyhounds and Registered Manufacturing Performance: 1980-1997 . 48 2.10 pc Services NSDP Predictor Means . . . . . . . . . . . . . . . . . . . 50 2.11 State Weights: Services . . . . . . . . . . . . . . . . . . . . . . . . . . 51 2.12 pc Agricultural NSDP Predictor Means . . . . . . . . . . . . . . . . . 54 2.13 State Weights: Agriculture . . . . . . . . . . . . . . . . . . . . . . . . 55 2.14 Greyhounds and Announcement Eects . . . . . . . . . . . . . . . . . 61 2.15 Yearly Treatment Eects . . . . . . . . . . . . . . . . . . . . . . . . . 68 3.1 Payo matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85 3.2 Session details . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88 3.3 Allocations under tTTC . . . . . . . . . . . . . . . . . . . . . . . . . 89 3.4 Allocations under tGS . . . . . . . . . . . . . . . . . . . . . . . . . . 90 vii 3.5 Allocations under tSD . . . . . . . . . . . . . . . . . . . . . . . . . . 91 3.6 Normalized Empirical Eciency . . . . . . . . . . . . . . . . . . . . . 92 3.7 Normalized Conditional Eciency . . . . . . . . . . . . . . . . . . . . 93 3.8 Tiered fairness by mechanism and endowment . . . . . . . . . . . . . 95 3.9 Existing tenants' participation decision . . . . . . . . . . . . . . . . . 97 3.10 Participation as a function of maximum gain in and own house payo 99 3.11 Logit models of participation decisions . . . . . . . . . . . . . . . . . 101 3.12 Frequency of participation . . . . . . . . . . . . . . . . . . . . . . . . 102 3.13 Classication of individual participation behavior . . . . . . . . . . . 103 3.14 Proportion of truthful preference revelation (relevant ranking) . . . . 104 3.15 Frequency of truth-telling (relevant ranking) . . . . . . . . . . . . . . 106 3.16 Classication of Misrepresentations . . . . . . . . . . . . . . . . . . . 107 3.17 Proportion of truthful preference revelation (full ranking) . . . . . . . 111 3.18 Misrepresentations after relevant ranking . . . . . . . . . . . . . . . . 111 viii List of Figures 2.1 Trends in pcNSDP: Andhra Pradesh vs. Control States . . . . . . . . 22 2.2 Trends in pc NSDP: Andhra Pradesh vs. Synthetic Control . . . . . . 22 2.3 pc NSDP Gap in Andhra Pradesh and placebo gaps of Control States (Upper panel has all controls, lower panel discards states with pre-treatment MSPE Two times higher than Andhra Pradesh) . . . . 26 2.4 Ratio of post-treatment RMSPE to pre-treatment RMSPE: pcNSDP 28 2.5 Leave-one-out Checks: The upper panel shows the trends in pcNSDP while the lower panel shows the gaps in pcNSDP . . . . . . . . . . . . 29 2.6 Trends in per capita industry output: Andhra Pradesh vs. Synthetic Control. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33 2.7 pc Industrial NSDP Gap in Andhra Pradesh and placebo gaps of Control States (Upper panel has all controls, lower panel discards states with pre-treatment MSPE ve times higher than Andhra Pradesh) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35 2.8 pc Manufacturing NSDP Gap in Andhra Pradesh and placebo gaps of Control States (Upper panel has all controls, lower panel discards states with pre-treatment MSPE six times higher than Andhra Pradesh) 36 2.9 pc Regristered Manufacturing NSDP Gap in Andhra Pradesh and placebo gaps of Control States (Upper panel has all controls, lower panel discards states with pre-treatment MSPE ve times higher than Andhra Pradesh) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37 ix 2.10 pc Unregristered Manufacturing NSDP Gap in Andhra Pradesh and placebo gaps of Control States (Upper panel has all controls, lower panel discards states with pre-treatment MSPE two times higher than Andhra Pradesh) . . . . . . . . . . . . . . . . . . . . . . . . . . 38 2.11 Ratio of post-treatment MSPE to pre-treatment MSPE: Industrial pcNSDP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39 2.12 Leave one out: Trends in Industrial pcNSDP . . . . . . . . . . . . . 42 2.13 Leave one out: Gaps in Industrial pcNSDP . . . . . . . . . . . . . . 43 2.14 Trends in pc Services NSDP: Andhra Pradesh vs. Synthetic Control . 49 2.15 pc Services NSDP Gap in Andhra Pradesh and placebo gaps of Control States (Upper panel has all controls, lower panel discards states with pre-treatment MSPE Three times higher than Andhra Pradesh) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50 2.16 Ratio of post-treatment RMSPE to pre-treatment RMSPE: pc Services NSDP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51 2.17 Leave-one-out Checks: The upper panel shows the trends in pc Services NSDP while the lower panel shows the gaps pc Services NSDP 53 2.18 Trends in pc Agricultural NSDP: Andhra Pradesh vs. Synthetic Control 55 2.19 pc Agriculture NSDP Gap in Andhra Pradesh and placebo gaps of Control States (Upper panel has all controls, lower panel discards states with pre-treatment MSPE Two times higher than Andhra Pradesh) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56 2.20 Ratio of post-treatment RMSPE to pre-treatment RMSPE: Agriculture 57 2.21 Leave-one-out Checks: The upper panel shows the trends in pc Agricultural NSDP while the lower panel shows the gaps pc Agricultural NSDP . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58 x 2.22 Map of India: Naxalite Aected States . . . . . . . . . . . . . . . . . 67 3.1 Sample Screenshot . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86 xi Abstract This dissertation is a two-part investigation on political economy and mechanism design. In the rst essay we estimate the economic eects of a counterinsurgency policy using the Naxalite insurgency in India as a case study. While there exists a substantial literature looking at the relationship between insurgency and economic growth, the economic eects of counterinsurgency policies remain underexplored. Us- ing the synthetic control method of analysis (a recently developed generalization of the dierence-in-dierence methodology), we provide the rst measurements of the direct economic benets of a unique robust security response to an insurgency. Of all the states aected by Naxalite violence in India, only one state i.e. Andhra Pradesh raised a specially trained and equipped police force in 1989 known as the Grey- hounds, dedicated to combating the Naxalite insurgency. Compared to a synthetic control region constructed from states aected by Naxalite violence that did not raise a specially trained anti-Naxalite police force, we nd that Andhra Pradesh gained on average 16.11% of its per capita NSDP over the period 1989 to 2000. Interestingly, we nd that these eects come through the various sub-sectors of the non-agricultural sector. The eects on the manufacturing sector and its sub-components (registered and unregistered manufacturing) are particularly strong and range from 20%-26%. Placebo tests indicate that all results are signicant. Conventional dierence-in- dierence specications using state and industry level panel data further corroborate these ndings. In the second essay we study in the laboratory a variant of the house allocation with existing tenants problem where (i) subjects are partitioned into tiers with hierar- xii chical privileges, (ii) they play multiple matches, and (iii) they know their position in the priority queue before making their decision. We evaluate the performance of the modied versions of three well-known mechanisms: Top Trading Cycle, Gale-Shapley and Random Serial Dictatorship with Squatting Rights. For all three mechanisms, we nd low rates of participation (around 40%), high rates of truth-telling conditional on participation (around 90%), high proportions of fair allocations (above 90%) and signicant eciency losses. We also observe dierences across mechanisms: Random Serial Dictatorship is ranked highest in eciency and Top Trading Cycle is ranked lowest in fairness. We then show that position in the queue has a positive and signi- cant impact on participation whereas experience and tier has little eect on behavior. Finally, the individual analysis reveals that the majority of subjects who do not play according to the theory still follow discernible patterns of participation and preference revelation. xiii Chapter 1 Introduction This dissertation is a two-part investigation on political economy and mechanism design. In the rst essay, I study the economic eects of a counterinsurgency policy in the Indian state of Andhra Pradesh. In the second essay, I use an economic experiment to examine the empirical properties of three allocation mechanisms in a particular example of a market design problem. 1.1 Economic benets of counterinsurgency The rst essay (Chapter 2), co-authored with Rahul Nilakantan, aims to estimate the economic eects of a counterinsurgency policy, in the form of a specialized police force in one of the aected states, introduced in response to the long running Nax- alite insurgency in India. For this analysis we use a recently developed econometric technique to construct counterfactuals from other states aected by Naxalite violence that did not undertake a similar counterinsurgency measure. While understanding the linkages between con ict and economic outcomes has long been considered important, the focus on the eects of counterinsurgency policies is recent and mainly due to the wars in Iraq and Afghanistan. The existing liter- ature, though, has largely been concerned with the eects of counterinsurgency on the levels of violence (for example, Berman, Shapiro and Felter (2011) and Iyengar et al. (2011)). But apart from causing considerable loss of life and property, con icts 1 also cause political instability which in turn adversely aect savings and investment resulting in loss of economic prosperity. Counterinsurgency policies, then by reducing uncertainties and boosting condence in such an environment, may enhance invest- ment and demand. To the best of our knowledge, this analysis is the rst to explore these eects of counterinsurgency policies on economic activity. The Naxalite (or Maoist) movement is an armed political movement aimed at overthrowing the state and establishing a communist regime. Since its inception in 1967, the Naxalite movement has spread rapidly through India, and currently aects 182 districts in 16 states (Ramana, 2009) and accounts for 91% of violence in India (Government of India, 2005). The Prime Minister of India has described the Naxalite problem as the single biggest internal-security challenge that the country currently faces. The Naxalites disrupt the local economy in their areas of operation by routinely kidnapping civilians for ransom, destroying public infrastructure and robbing banks among other things. The response from the aected states has been lacking, with the states unsuccessfully relying on inadequately trained regular police forces to counter the Naxalites. However, only one of the aected states, Andhra Pradesh, raised a dedicated state police force known as the Greyhounds in 1989 to counter the Naxalite insurgency. Well-equipped and specially trained in counterinsurgency methods, the Greyhounds have been widely credited with bringing the Naxalite violence in Andhra Pradesh under control. This study examines the extent to which the creation of the Greyhounds miti- gated the loss of per capita net state domestic product (pcNSDP) due to Naxalite violence in the state of Andhra Pradesh. For this analysis we use the synthetic control methodology to estimate how an economy would have grown in the absence of the counterinsurgency policy response. Developed by Abadie and Gardeazabal (2003) and Abadie et al. (2010), it generalizes the dierence-in-dierence methodology by 2 relaxing the assumption of a common trend between the treated and the control in the absence of treatment. Although this approach does not allow for inference through traditional asymptotic methods, informative inference is still possible through falsi- cation (placebo) tests. Finally, we also run standard dierence-indierence spec- ications using state and industry level panel data to check the robustness of the results. We nd that the introduction of Greyhounds in Andhra Pradesh in 1989, yielded a \security dividend" equal on average to 16.11% of its pcNSDP over the period 1989 to 2000. We further nd that this eect came through 11%-25% eects on various subsectors of the non-agricultural sector. In particular, the eects on the manufac- turing sector and its sub-components (registered and unregistered manufacturing) are particularly strong and range from 20%-26%. Most importantly, a more detailed analysis of the registered (large and medium scale enterprises) manufacturing sector nds a positive and signicant impact on various measures of investment. Since the synthetic control methodology is a case-study approach, we would like to stress that the ndings do not imply that other Naxalite aected states raising se- curity forces similar to the Greyhounds would experience eects of similar magnitude. However, the ndings of this research are still valuable as they show that output and investments respond to counterinsurgency policies. 1.2 Tiered housing allocation: an experimental analysis In Chapter 3, a joint experimental paper with my co-advisor Juan Carrillo, I consider a particular example of a market design problem. Market design is an important tool for improving allocations in cases where the exchange of money is not permitted - for example, kidney exchange, college and school admissions, health care, labor markets, 3 etc. 1 We examine a particular version which is concerned with the allocation of a set of goods to a set of agents who are partitioned into tiers of dierent privileges. Real life scenarios include the allocation of oces to faculty members, schedules to crews in transportation industry and on-campus housing to college students. For example, suppose one wishes to reallocate oces to faculty members who are tiered into full professors, associate professors, assistant professors etc. where some already have oces to start with while others are newcomers without oces and money exchange is not allowed. A senior faculty member has more privilege, in the sense that she should get her assignment before those in lower tiers and should be able to take the endowment of a lower-tiered agent if she wishes to. In this example, a full professor should get her oce before an assistant professor and should be able to take the oce currently occupied by an assistant professor if she wants to. This is a version of what is commonly called the housing allocation with existing tenants problem and as the rst part of our analysis we extend the three leading mechanisms that have been proposed in the literature - the top trading cycle (TTC), the Gale-Shapley mechanism (GS) and Random Serial Dictatorship with squatting rights (RSD) - to this mutli-tiered environment which we denote tTTC, tGS and tSD respectively. In a controlled laboratory environment we test whether the empirical behavior of agents match the theoretical predictions of these three allocation mechanisms on four criteria: (a) Pareto eciency (the houses should be optimally allocated given the preferences of the agents); (b) fairness (the assignment should respect the priority order); (c) individual rationality (an agent should be no worse-o by participating in the mechanism); and (d) strategy-proofness (agents should not benet from mis- representing their preferences). In addition to the introduction of the hierarchical structure, we make the following additional methodological changes with respect to 1 See S onmez and Unver (2011) for a survey of the literature. 4 the existing experimental literature: 2 (i) we inform the agents of their position in the queue before they make any decisions and (ii) we play multiple rounds in order to study individual behavior. We nd that performance is similar across mechanisms with the tSD performing slightly better in terms of eciency due to higher truthful revelation rates. Of the variations we introduce we nd that telling the agents their position in the queue before eliciting their participation decision and ranking of alternatives, signicantly alters their behavior in all three mechanisms while theoretically, it should not under the tTTC and tGS mechanisms. In particular, we nd that subjects lower in the priority queue are less likely to participate relative to those higher in the queue. This the rank-dependency of choices has important policy implications that deserve to be explored in further detail. As for the other variants introduced, we do not nd any systematic tier eect on the behavior of subjects after controlling for potential gains of participation and opting out payos. Additionally, we do not nd any signicant changes in behav- ior over multiple matches. Together, these results indicate that individuals require substantiative feedback in order to understand the properties of these mechanisms. 2 The experimental literature on housing markets with existing tenants is surprisingly limited and mainly consists of the seminal paper by Chen and S onmez (2002) and Guillen and Kesten (2012). 5 Chapter 2 Naxalite Insurgency and the Economic Benets of a Unique Robust Security Response 1 Since its independence in 1947, India has faced numerous insurgencies within its borders at various points in time. One of the longest running insurgencies in India is the Naxalite - also known as the Maoist - movement. Using the synthetic control method of analysis developed by Abadie and Gardeazabal (2003), we present the rst estimates of the economic gains - in terms of income per capita - of a unique, robust security response to the Naxalite insurgency in one of the aected states i.e. Andhra Pradesh. With the ultimate objective of overthrowing the state by force and establishing a communist regime (Ramana, 2009 and Gupta, 2007), the Naxalite movement started in a small village in West Bengal in 1967, and then spread steadily across the country. The rate of spread of the movement has become alarming in the recent past, from 76 districts in 9 states in 2005 (Government of India, 2006) to 182 districts in 16 states in 2007 (Ramana, 2009). By the Indian Government's own estimate it accounts for about 91% of the total violence in India and 89% of the resulting deaths (Government of India, 2005) prompting Prime Minister Manmohan Singh to observe that the Naxalite insurgency is the single biggest internal security threat facing the country. While other countries have generally relied on soft counterinsurgency policies, India has frequently resorted to security based or \coercive" responses to its insur- 1 This chapter is co-authored with Rahul Nilakantan. 6 gencies. The usual strategy is to ood the aected area with security forces in order to sti e the insurgency and once some sort of order is restored, the State normally negotiates a political settlement within the framework of the constitution. This strat- egy has borne mixed results. While this has worked in Punjab, Tripura and Mizoram, it has been unsuccessful in Assam, Nagaland and Jammu & Kashmir. 2 The Naxalite insurgency diers from the usual Indian insurgency experience on two important counts. Firstly, while most of the other insurgencies are restricted to a small region or a state, the Naxalite insurgency is spread over a large part of India, making it dicult to coordinate counterinsurgency eorts across states. Secondly, while all the other counterinsurgency eorts have seen the involvement of the Indian Army, the states' response to the Naxalite insurgency has been dependent on the state police forces. Of the several states in India that are aected by Naxalite violence, only one state raised a specially trained police force dedicated to combating the Naxalite insurgency. This explicit change in the government's counterinsurgency policy gives us a unique opportunity to measure the direct economic benets of this robust localized security response to the Naxalite insurgency. We use a recently developed generalization of the dierence-in-dierence method- ology to get around the usual problems faced by empirical studies that evaluate interventions at the aggregate level - data and sample size limitations. Using the syn- thetic control methodology we nd that the introduction of a specialized police force called Greyhounds in Andhra Pradesh in 1989, yielded a \security dividend" equal on average to 16.11% of its per capita net state domestic product (pcNSDP) over the period 1989 to 2000. We further nd that this eect came through 11%-25% eects on various subsectors of the non-agricultural sector. The eects on the manufacturing 2 See Chadha (2005) for a summary of these insurgencies and Mukherjee (2010), Rajagopalan (2008) and Ganguly and Fidler (2009) for a discussion on India's counterinsurgency experience. 7 sector and its sub-components (registered and unregistered manufacturing) are par- ticularly strong and range from 20%-26%. Placebo tests indicate that our results are signicant. Additionally, these results are robust to standard dierence-in-dierence specications using the state and industry level panel data. To the best of our knowledge, this is the rst study that directly estimates the economic benets of a security response undertaken by a state to counter an extrem- ist threat. Although there exists a substantial literature looking at the relationship between insurgency and economic growth, recent evidence from Iraq and Afghanistan suggests a more dynamic relationship between insurgency, counterinsurgency and eco- nomic outcomes. This research, though based on an example from India, contributes to this small but growing eld of research. The rest of the paper is organized as follows. Section 2.1 brie y summarizes the related literature. Section 2.2 provides a brief history of the Naxalite movement in India and the state response. Section 2.3 provides a brief overview of the synthetic control method of analysis and the data. Section 2.4 describes the results of the analysis, Section 6 discusses the ndings and Section 7 concludes. 2.1 Related Literature While understanding the linkages between con ict and socio-economic outcomes has long been considered important, the focus on the eects of counterinsurgency policies is recent and mainly due to the wars in Iraq and Afghanistan. Researchers and prac- tioners of counterinsurgency broadly classify the policies as either \carrot" or \stick". The \carrot" or \soft" counterinsurgency approach is dominated by two major mech- anisms. The rst is the \hearts and minds" approach. This aims to win the over the population by providing them public services, with the expectation that once their 8 grievances are addressed, the attitude of the population towards the government will improve. The civilians are then less likely to help or join the insurgents and more likely to share information with the counterinsurgents. The existing research provides mixed evidence. While Berman, Shapiro and Felter (2011) nd that improved service provision through the Commanders?Emergency Reconstruction Program (CERP) in Iraq reduced violence, Beath et al.(2011) nd that even though the National Solidar- ity Program (NSP) in Afghanistan improved villagers' perception of the government, it had no eect on violence levels. Furthermore, Crost, Felter and Johnston (2012) nd that the KALAHI-CIDSS development assistance program in the Philippines actually increased violence. The second mechanism is the opportunity-cost approach that builds on Becker's theory of crime (Becker 1968). Improved economic environment, access to the market, labor market conditions, etc. increase the costs of participating in the insurgency thereby reducing the supply of insurgents. While, Berman et al. (2011) nd evidence against this mechanism, Iyengar et al. (2011) nd that labor-intensive projects under CERP reduced violence levels in Iraq. Overall, even though the current evidence on the dierent mechanisms underlying the soft counterinsurgency approach is mixed, given the U.S. army's focus on this approach (U.S. Army, 2007), the base of knowledge has been rapidly expanding over the last few years. Surprisingly, this has not been the case for the coercive or \stick" measures of counterinsurgency. Although coercive counterinsurgency measures have been employed frequently, there have been few systematic empirical evaluation of these policies. Security based state responses may deter civilian support and reduce violence (Lyall, 2009) or drive up support for the insurgents (Kocher et al., 2011). Further, the evaluation of state response to insurgencies has been largely restricted to the eects on the production of violence. However, the ecacy of counterinsur- 9 gency may also be evaluated via their eect on economic outcomes (Kapstein, 2012 and Greenstone 2007). Existing research nds that markets, by eciently aggregat- ing information, can be a good indicator of civilians' security outlook and provide an unbiased evaluation of the state's security policy. This methodology is particularly useful when violence data is unreliable or unavailable, as is the case for the Nax- alite insurgency over the duration of this study. For example, Zussman and Zussman (2006) nd that the Israeli and Palestinian stock markets respond negatively to Is- rael's assassinations of senior political leaders of Palestinian terrorist organizations but positively to the assassination of senior military leaders. Similarly, some papers have used the price of Iraqi state bonds to measure the eectiveness of various coun- terinsurgency policies in Iraq. Greenstone (2007) nds a sharp decline in bond prices immediately after the \Surge" indicating worsening expectations about Iraq's future. Chaney (2008) nds evidence that the Iraqi bond market fell following the news of coalition troops withdrawal but responded positively to news of negotiations with Iran. This paper builds on this line of work that uses economic indicators to evaluate the eectiveness of security policies. Finally, despite now running in its fth decade, there exist few systematic quanti- tative studies on the Naxalite insurgency. Cross-sectional studies nd high incidence of poverty, low levels of literacy, forest cover and population share of members of scheduled castes and tribes to be the main correlates of Naxalite activity at the dis- trict level (Borooah, 2008 and Hoelscher et al., 2012). Gawande, Kapur and Satyanath (2012) in a panel data study nd that adverse natural resource shocks increase the intensity of Naxalite violence over the period 2001-08. Eynde (2011) nds that nega- tive rainfall shocks increase Naxalite violence against civilians in order to deter them from becoming police informers. 10 2.2 The Naxalite Movement The Naxalite movement traces its roots to Naxalbari, a small village in West Bengal. In March 1967, a tribal farmer was attacked by local landlords over a land dispute. A peasant uprising followed, led by revolutionaries of the Communist Party of India (Marxist) i.e. CPI(M) in several states of India, namely Andhra Pradesh, Bihar, Jammu and Kashmir, Karnataka, Kerala, Orissa, Tamil Nadu, Uttar Pradesh, and West Bengal. The West Bengal government, despite being led by CPI(M), crushed the rebellion within West Bengal. However, the revolutionaries within the CPI(M) split to form the All India Coordination Committee of Communist Revolutionaries (AICCCR) in 1968. The AICCCR rejected parliamentary elections and called for an armed uprising against the state. Due to internal con icts, the AICCCR split, and a new organization called the Communist Party of India (Marxist-Leninist) i.e. CPI(ML) was formed in 1969. At the same time, another organization, later known as the Maoist Communist Center (MCC) was formed in Bihar under the name Dakshin Desh. Response by the state security forces was swift and violent, suppressing the insurgency by 1972. Although, the CPI(ML) continued armed struggle against the Indian state throughout the 1970s, the movement was riven by internal con icts, suered from further splits, and soon disintegrated. Of the various factions to emerge from the CPI(ML), the two most prominent ones were the Communist Party of India (Marxist-Leninist) Liberation i.e. CPI(ML) Liberation in 1974 and the Communist Party of India - Marxist Leninist (People's War), also known as the People's War Group (PWG) in 1980. While the CPI(ML) Liberation did not rule out the possibility of armed revolution against the state, it did participate in the electoral process, even winning an election in Bihar in 1989. The PWG and the MCC on the other hand, completely rejected the democratic system 11 and continuously waged a \people's war for the people's government". Through the 1980s and 1990s the various factions rapidly consolidated their bases, actively engaged the state security forces but showed little inter-group coordination. Of late, however, there have been many mergers - the biggest being that of the MCC and the PWG to form the Communist Part of India - Maoist (CPI-Maoist) in 2004. 3 The Naxalite violence has imposed economic costs on the aected states through various avenues. For example, the blog Naxal Terror Watch 4 documents incidents of Naxalites destroying pipelines transporting iron ore slurry in Chattisgarh, destroying road construction machinery in Bihar, forcing the closure of bank branches in Jhark- hand, disrupting power supply by damaging hydroelectric power stations in Orissa, impeding interstate commerce by routinely preventing the repair of national highways and damaging railway infrastructure in Jharkhand and Orissa, and degrading telecom service by destroying mobile phone towers (Naxal Terror Watch, 2012). The Naxalites also impede economic growth by administering a \kidnap and ex- tortion empire" in their areas of operation. For example, in a dramatic show of force, the Naxalites held a prestigious high-speed train and its 700 passengers hostage for about ve hours in 2009 (Hindustan Times, 2009). Extortion from small and large enterprises as well as the collection of `taxes' is reportedly common in areas under Naxalite control (Singh and Diwan, 2010). 5 Joshi (2010) reports that investments of the order of Rs.130 bn. (approximately $2.83 bn. at 2012 exchange rates) were tied up in just the power and steel industries in projects that could not be completed in the state of Chattisgarh on account of Naxalite violence. Additionally, civilians caught 3 A more detailed description of the history of the movement can be found in Kujur (2008) and Gupta (2007). 4 This blog, available at naxalwatch.blogspot.com aggregates reports of Naxalite activity from popular Indian newspapers. 5 Singh and Diwan (2010) report that police investigations in 2007 revealed the revenue of CPI(Maoist) to be over Rs. 20 bn. ($ 400 m.). Another major source of revenue is reported to be poppy cultivation. 12 between the insurgents and the government forces face reduced access to health care (Solberg, 2008) and educational services. 2.2.1 State Response and the Greyhounds The state counterinsurgency response to the rapid growth of the Naxalite movement since the 1980s has been lacking and inconsistent. The aected states have primarily relied on the regular police force to maintain security. However, the state police forces being understaed and inadequately trained for counterinsurgency operations, have failed to check the Naxalite insurgency. Further, since maintaining law and order in India is the responsibility of the states, until recently, there has been little coordination between the central and the state governments on the approach towards the insurgency. The role of the central government has been largely restricted to providing reinforcements from the central police organizations when requested to do so. The deployment of central police forces is usually for limited time periods and have done little to boost counterinsurgency eorts (Oetken, 2009 and Ramana, 2009). Of the aected states only Andhra Pradesh raised a separate police force, called the Greyhounds, whose main purpose was to combat the Naxalite insurgency in the state. 6 Established in 1989 as a separate administrative unit, the Greyhounds are an elite commando force specially trained in counterinsurgency methods, well-equipped and have their own intelligence network and other support units. The personnel are recruited from the regular police force and though exact numbers are not publicly available, the size of the Greyhounds is reported to have steadily increased from 886 in 1989 (Shatrugna, 1989) to around 2000 currently (Priyadershi, 2009). It must be pointed out that, in addition to the creation of a highly trained anti- 6 Other stated responsibilities of the Greyhounds include providing assistance during natural disasters and other grave law and order situations. For more details see the Greyhounds webpage at www.apstatepolice.org. 13 Naxalite police force, the establishment of the Greyhounds has resulted in the the transformation of the counterinsurgency capabilities of the regular state police force in Andhra Pradesh. All Greyhound personnel have a tenure of three years after which they return to their respective police units. Additionally, all newly recruited police ocers are required to train with the Greyhounds before being absorbed into the dis- trict police establishment. These policies have resulted in a signicant improvement in the counterinsurgency capabilities of the local police forces, close coordination be- tween the Greyhounds and the district police and better intelligence gathering thereby increasing the eciency of the operations of the Greyhounds (Sahni, 2007 & 2008 and Achuthan, 2010). Throughout this paper, the estimation of the eect of the establish- ment of the Greyhounds is eectively an estimation of this dramatic transformation of the police setup in Andhra Pradesh. Anecdotal evidence suggests that the Greyhounds have been successful in bringing the Naxalite violence in Andhra Pradesh under control (Swami, 2010 and Tata, 2010), motivating policy makers to consider raising similar forces in other aected states. For example, in a speech on December 20, 2007 the Prime Minister noted: \. . . I believe that given the unique nature of this problem, it is time to have a dedicated force just to tackle naxalism. Aected states must set up Special Task Forces on the Andhra Pradesh pattern and the Centre will provide assistance for this purpose." 7 Currently, several Naxalite aected states such as Chhattisgarh, Jharkhand, Orissa, Madhya Pradesh, Bihar and Maharashtra are reported to be in the process of creating dedicated security forces along the lines of the Greyhounds to tackle the insurgency 7 PM's Closing Remarks at the Chief MinistersConference on Internal Security, New Delhi 2007. Full text of this speech is available at http://pmindia.nic.in/speech-details.php?nodeid=613 (ac- cessed June 1, 2012). 14 (Tiwari, 2009). This is being done even though there exists no systematic study that establishes the ecacy of the Greyhounds. The rigorous ndings of this paper should help policy makers make informed decisions. Finally, we would like to point out some issues regarding the timing and the exo- geneity of the policy intervention under consideration. Even though the Greyhounds were ocially introduced in 1989, the decision to raise the Greyhounds was announced on June 6, 1988 (Balagopal, 1988). Although, as shown in the results later, this could have led to an announcement eect, we take 1989 as the treatment year because in ad- dition to the introduction of the Greyhounds, there was an important transformation of the Andhra Police setup that occurred only after 1989. Secondly, the introduction of the Greyhounds is reported to be an idea of K. S. Vyas, an ocer with Indian Police Service (IPS), posted in Andhra Pradesh (Raju, 2010). 8 The allocation of an ocer to a particular state is done by the central government on the basis of merit and other exogenous factors. Although, this is not denitive argument for the intro- duction of the Greyhounds being a clean natural experiment, it is possibly the closest that one can get to a natural experiment in the eld of counterinsurgency. 9 2.3 Research Methodology and Data In order to answer the research questions, we use the synthetic control method de- veloped by Abadie and Gardeazabal (2003) and Abadie, Diamond and Hainmueller (2010) (AG and ADH hereafter, respectively) in a comparative case study approach. In simple terms, the methodology uses pre-treatment outcomes and their predictive characteristics to weight the unaected (control) units in such a way that they provide 8 K. S. Vyas was later assassinated by the Naxalites in 1993. 9 However, there may remain unobservables. For example, it is possible that police ocers in other states had similar ideas that did not gain acceptance. 15 an appropriate counterfactual for the exposed (treated) unit. While some studies employ individual micro data to analyze macro policy interven- tions, it must be noted that disaggregated data using standard inference techniques assume all uncertainty enters through sampling error in estimating the means. ADH point out that in such cases analyzing the policy at the macro level is preferable as using aggregate data eliminates this uncertainty. 10 At the aggregate level, however, a regression approach is usually less appropriate due to the lack of a sucient number of treated and control units for robust inference. The synthetic control methodology is useful in circumstances such as those of the present study where the the event of interest - the creation of the Greyhounds in 1989 - occurred at an aggregate level (state level) and aected aggregate entities (states of India). 11 More importantly, as further elaborated below, the synthetic control approach can be motivated as a gen- eralization of the linear panel dierence-in-dierences model where the unobserved individual specic confounders are allowed to vary with time, thereby relaxing the assumption of a similar dierence between the treated and the control in the absense of treatment. Although this approach does not allow for inference through tradi- tional asymptotic methods, informative inference is still possible through falsication (placebo) tests. We now summarize the synthetic control methodology of ADH (notation and equations are those of ADH). Suppose we have J + 1 regions with the rst region exposed to the treatment and the remaining J regions being the potential controls. There areT time periods andT 0 pre-intervention time periods such that 1<T 0 <T . Let Y N it be the outcome observed for region i2f1;:::;J + 1g if it is not exposed to 10 Other uncertainties such as the ability of the control units to replicate the treated units still remain. See Imbens and Wooldridge (2009) for a discussion of the literature. 11 For some recent applications of this methodology see Lee (2011), Montalvo (2011), Pinotti (2012) and Hinrichs (2012). 16 the treatment and Y I it be the outcome observed for the the i th region if it is exposed to the treatment in time periods T 0+1 to T. Let D it be a dummy variable that takes the value of 1 if regioni is exposed to the treatment at time periodt and 0 otherwise, i.e. D it = 8 > < > : 1 if i = 1 and t>T 0 0 otherwise (2.1) The observed outcome for region i at time t is then Y it =Y N it + it D it (2.2) where it = Y I it Y N it is the eect of the treatment on region i at time t. We are interested in estimating 1;T 0+1 ;:::; 1;T . Since we observeY I it , in order to estimate it we just need to estimate Y N it . Let Y N it be given by a generalized dierence-in- dierence (xed eects) model, where the unobserved individual specic eect is allowed to vary with time Y N it = t + t Z i + t i + it (2.3) Here, Z i is a vector of observed covariates (which may contain time varying covari- ates), i are individual specic unobserved confounders, t is a vector of unobserved common factors and it are mean 0 shocks. Let W = fw j g J+1 j=2 be a set of non- negative weights that sum up to one. Each such set of weights represents a particular weighted average of controls i.e., a particular synthetic control. Hence for a given W the outcome for the synthetic control will be 17 J+1 X j=2 w j Y N jt = t + t J+1 X j=2 w j Z j + t J+1 X j=2 w j j + J+1 X j=2 w j jt (2.4) Let there be weights (w 2 ;:::;w J+1 ) such that Z 1 = J+1 X j=2 w j Z j and Y 1t = J+1 X j=2 w j Y jt 8t2fT 0+1 ;:::;Tg (2.5) i.e., (i) the weighted average of the covariates of the controls perfectly replicates the covariates of the treated unit, and (ii) the weighted average of the pre-treatment outcomes of the controls perfectly matches the pre-treatment outcomes of the treated unit. Then, ADH show that if P T 0 t=1 0 t t is non-singular, we have Y N 1t J+1 X j=2 w j Y jt = J+1 X j=2 w j T 0 X s=1 t T 0 X n+1 0 n n ! 1 0 s ( js 1s ) J+1 X j=2 w j ( jt 1t ) (2.6) Further, they show that the mean of the right hand side of equation (2.6) is close to zero \. . . if the number of pre-intervention periods is large relative to the scale of the transitory shocks" (pg. 495). We can therefore estimate the impact of the treatment as ^ it =Y it J+1 X j=2 w j Y N jt 8t2fT 0+1 ;:::;Tg (2.7) Usually we are unable to get a perfect synthetic control because weights do not exist such that the equations in (2.5) hold exactly. The weights are then selected such that the equations in (2.5) hold approximately. Note that by not restricting t to be constant over time, the synthetic control methodology relaxes the tradi- tional dierence-in-dierence assumption that in absence of treatment, the dierence 18 between \treatment" and \control" groups is constant over time. In order to implement the synthetic control methodology, let X 1 be the vector of Z 1 and pre-treatment outcomes for the treated state and X 0 be the matrix ofZ j and pre-treatment outcomes for the J control states. The vector of weights W is chosen to minimize (X 1 X 0 W ) 0 V (X 1 X 0 W ) subject to the weightsfw j g J+1 j=2 being non- negative and summing up to 1. The weighting matrix V can be any positive denite matrix but note that the choice ofV aectsW . Following the existing literature, we allow the choice ofV to be data driven, by choosingV such that the mean square error of the outcome variable is minimized for the pre-treatment period. All calculations in this paper were performed using the software SYNTH for STATA, developed by ADH. 12 Given the small number of control units, asymptotic inferential techniques cannot be applied to comparative case studies. To estimate the \signicance" of the results in this study, we conduct placebo tests similar to those in AG and ADH. A placebo test is one where the entire analysis is performed for a control state as if the control state was treated. Since the control state was not treated, we should not expect to nd any treatment eect. If the placebo studies using control states iteratively assigned to treatment status create treatment eects of magnitude similar to the ones estimated for the actually treated state, then the conclusion is that the analysis does not provide signicant evidence of a treatment eect for the actually treated state. The primary economic indicator used to measure the eectiveness of the Grey- hounds in real per capita net state domestic product (pcNSDP). We further investi- gate the channels through which the treatment aects pcNSDP using components of pcNSDP such as industrial pcNSDP, manufacturing pcNSDP, registered manufactur- ing pcNSDP, unregistered manufacturing pcNSDP, services pcNSDP and agricultural 12 The software is available for download at http://www.mit.edu/ jhainm/software.htm 19 pcNSDP. All these outcome measures are measured in 1999 prices. We use standard predictors of economic growth such as Human Development In- dex (HDI), population density, road density, percentage of households with access to safe drinking water, per capita electricity consumption, per capita development expenditure and percentage of population below the poverty line. We also use some observed covariates (Z i ) specically for certain outcomes. For example, for the in- dustrial sector and its various subsectors we use the Labor Reform Index constructed by Besley and Burgess (2004) 13 to proxy the industrial labor relations in the state. Similarly, for the agricultural sector we include variables such as foodgrain yields (to proxy for the extent of the spread of the green revolution) and average rainfall. The X 0 andX 1 matrices consist of a combination of observed covariates (averaged over the entire pre-intervention period) and some pre-treatment values of the outcome of in- terest. 14 The full list of the variables and details regarding their sources are provided in the appendix. Finally, the period under consideration in this study is 1970-2000 which gives us 1970-1988 as the pre-treatment period and 1989-2000 as the treatment period. The treated unit is Andhra Pradesh, and the potential control units are the other Naxalite aected states: Bihar, Madhya Pradesh, Maharashtra, Karnataka, Orissa, Uttar Pradesh and West Bengal. 15 Although we primarily use the synthetic control 13 The states are allowed to amend the central government's Industrial Disputes Act of 1947. So even though all states had the same starting point, currently labor market regulations dier across states. Besley and Burgess (2004) code each state amendment as pro-worker (+1), neutral (0) and pro-employer(-1) and calculate the net direction of change for each year over the period 1947-1997. The index is arrived at by cumulating the scores over time. See Besley and Burgess (2004) for further details. 14 While the existing literature does not specify a limit on number of pre-intervention covariates to use for matching (personal communication with Alberto Abadie), adding more variables can to lead to the standard problem of dimensionality. Although we have a fairly rich set of covariates, we restrict the total number of variables to eight and use the remaining covariates in other specications as robustness checks. 15 The states of Chhattisgarh, Uttaranchal and Jharkhand were carved out of Madhya Pradesh, Uttar Pradesh and Bihar respectively in Nov. 2000. Our analysis takes into account the undivided 20 methodology, standard dierence-in-dierence estimations are also provided whenever possible. 2.4 Results We begin our analysis with the eects of the creation of the Greyhounds on the pcNSDP of Andhra Pradesh. Figure 2.1 plots the trajectories of pcNSDP of Andhra Pradesh and a simple average of the pcNSDP of the control states (called Control States) over the period 1970-2000. Recall that the Control States are the other states that are aected by Naxalite violence but which did not set up a similar anti-Naxalite police force. The treatment year is 1989, in which the Greyhounds police force was introduced. Therefore, the pre-treatment period is 1970-1988, and the treatment period is 1989-2000. The equally weighted average pcNSDP for the control states lies above that of Andhra Pradesh for most of the pre-treatment period and well below that of Andhra Pradesh in the treatment period. This divergence in pcNSDP after 1989 however, is not the true treatment eect since before the formation of the Greyhounds force, Andhra Pradesh was a consistent underperformer relative to the rest of the Naxalite aected states. As an equally weighted average of the rest of the Naxalite aected states was signicantly dierent from Andhra Pradesh in the pre-treatment period, using it as a comparison group for Andhra Pradesh would be inappropriate. The synthetic Andhra Pradesh (i.e., the synthetic control unit) is constructed as a weighted average of the states in the potential control group that most closely resemble Andhra Pradesh in terms of (i) pre-treatment values of pcNSDP and (ii) pre- treatment values of pcNSDP growth predictors. Table 2.1 compares the pre-treatment states of Madhya Pradesh, Uttar Pradesh and Bihar. A map of India with the Naxalite aected states is provided in appendix 2.B. 21 Figure 2.1: Trends in pcNSDP: Andhra Pradesh vs. Control States Figure 2.2: Trends in pc NSDP: Andhra Pradesh vs. Synthetic Control 22 characteristics of Andhra Pradesh to those of the synthetic control (appropriately weighted average of controls) and also to the simple average of controls. As can be seen from the table, the synthetic control matches actual Andhra Pradesh in terms of HDI, population density, road density, percentage of urban households with access to safe drinking water and per capita electricity consumption far more closely than the simple average of the control states. Similarly, the synthetic control is fairly close to actual Andhra Pradesh in terms of pre-treatment pcNSDP. Table 2.1: pcNSDP Predictor Means Variables Andhra Pradesh Synthetic Control Control States Mean Min Max HDI 0.30 0.28 0.29 0.24 0.36 Population Density (persons/sq. km.) 176.50 180.46 303.86 106 588.5 Urban HH with safe drinking water (%) 63.27 68.21 70.90 51.33 85.56 Per capita Electricity Consumption (KwH) 114.35 122.86 132.44 88.84 239.28 Road Density (km./1000 sq. km. area) 366.00 405.99 487.14 202 620.5 1973 pcNSDP (1999 prices) 7605.60 7634.94 7728.50 4937.43 9553.13 1983 pcNSDP (1999 prices) 7891.15 7899.00 8478.88 5084.04 11848.05 1984 pcNSDP (1999 prices) 7703.05 7759.85 8424.71 5556.31 11686.64 Table 2.2: State Weights: pcNSDP State Weight Bihar 0.19 Karnataka 0.34 Madhya Pradesh 0.42 Maharashtra 0.00 Orissa 0.05 Uttar Pradesh 0.00 West Bengal 0.00 23 Table 2.2 shows the weights given to each state in the control group when con- structing the synthetic Andhra Pradesh. The optimal weights are positive for Bihar, Karnataka, Madhya Pradesh and Orissa; and zero for the other states. Of the four states that border Andhra Pradesh (Karnataka, Madhya Pradesh, Orissa, and Maha- rashtra) three (Karnataka, Madhya Pradesh and Orissa) account for over 80% of the weight. This gives us further condence in our estimates as these neighboring states are more like Andhra Pradesh in unobserved variables like geography and culture relative to the other controls. Figure 2.2 plots the trajectory of pcNSDP of Andhra Pradesh and the synthetic control for the period 1970-2000. In the pre-treatment period 1970-1988, the pcNSDP of the synthetic control behaves very similarly to that of Andhra Pradesh till 1987. In the treatment period, the pcNSDP of the synthetic control diverges sharply from that of Andhra Pradesh, with the gap increasing rapidly over time. What does this divergence mean for the actual and potential evolution of pcNSDP for Andhra Pradesh? Recall that the evolution of pcNSDP for the synthetic control is what the evolution of pcNSDP would have been for Andhra Pradesh if it had not raised the Greyhounds force. Hence, the gap between Andhra Pradesh and its synthetic control is an estimate of the treatment eect. The main implication of Figure 2.2 is that the establishment of the Greyhounds force paid immediate and rich dividends to Andhra Pradesh in terms of raising its pcNSDP to levels higher than could have been achieved in the absence of the dedicated anti-Naxalite force. The pcNSDP \dividend" seems to be increasing steadily from year to year all the way till the end of the study period. Lastly note that the divergence starts a year before in 1988, possibly indicating an announcement eect discussed in section 2.2.1. We can quantify the gain in pcNSDP level for Andhra Pradesh in two ways: (1) percentage average gain, and (2) average percentage gain. The former is the average 24 yearly gap between pcNSDP of Andhra Pradesh and that of the synthetic control expressed as a percentage of the average yearly pcNSDP of Andhra Pradesh over the treatment period i.e. 1989-2000. The latter is the average of the yearly percentage gap between pcNSDP of Andhra Pradesh and that of the synthetic control over the treatment period. The average yearly gap between pcNSDP of Andhra Pradesh and that of the synthetic control was Rs.2,065.2 and the average yearly pcNSDP of Andhra Pradesh was Rs.12,820 during the treatment period. Andhra Pradesh consequently experienced a percentage average gain of 16.11% of its pcNSDP during this period. 16 The maximum gain observed was Rs.4,133.7 at the end of the treatment period. 17 Another way to view the benets is in terms of the change of growth rates. The synthetic control had an annual growth rate of 2.68% over the period 1989-2000 while Andhra Pradesh grew at an annual rate of 4.41%. The Greyhounds force has therefore had a large positive impact on the pcNSDP of Andhra Pradesh. Next, the signicance of our results is assessed by conducting a series of placebo tests that involve iteratively applying the synthetic control method to each of the seven control states. The upper panel of gure 2.3 shows pcNSDP gap in Andhra Pradesh and the placebo gaps for all the control states. The estimated pcNSDP gap between each control state and its synthetic counterpart is represented by the grey lines. The black line represents the estimated pcNSDP gap between the Andhra Pradesh and its synthetic counterpart. It is clear from gure 2.3 that the estimated pcNSDP gap for Andhra Pradesh States is large (positive) in comparison to the distribution of pcNSDP gaps for the control states, with the exception of one control state i.e. Maharashtra. From the upper panel of gure 2.3, it is clear that the synthetic control method 16 The average percentage gain during this period was 15.37%. 17 The yearly gains are provided in appendix 2.C. 25 Figure 2.3: pc NSDP Gap in Andhra Pradesh and placebo gaps of Control States (Upper panel has all controls, lower panel discards states with pre-treatment MSPE Two times higher than Andhra Pradesh) 26 provides a good t for pcNSDP both in Andhra Pradesh and in control states prior to the treatment period, with the exception of Maharashtra. The pre-treatment root mean squared prediction error (RMSPE) 18 for Andhra Pradesh is just 273.48, while the median pre-treatment RMSPE for control states in the placebo runs is just 377.97, indicating relatively good pre-treatment ts. The pre-treatment RMSPE for the Maharashtra placebo run is largest at 1910.02, while that for Bihar is also relatively large at 1616.18. The poor pre-treatment ts in the placebo runs for Maharashtra and Bihar cast doubt on the reliability of the post-treatment ts for Maharashtra and Bihar. As ADH (pg.502) observe, \ . . . placebo runs with poor pre-treatment t do not provide information to measure the relative rarity of estimating a large post treatment gap for a state that was well tted prior to treatment." Therefore, in the lower panel of gure 2.3, we drop placebo runs for states that give pre-treatment RMSPEs that are at least two times higher than the pre-treatment RMSPE for Andhra Pradesh. On dropping the controls with poor pre-treatment t (Bihar and Maharashtra) the pcNSDP gap for Andhra Pradesh is clearly the largest (positive) of all the pcNSDP gaps. Dropping controls with a poor pre-treatment t involves some amount of subjec- tivity. We can also check the signicance of the estimated treatment eect for Andhra Pradesh by comparing the post-treatment RMSPE to pre-treatment RMSPE ratios. Since controls with a poor pre-treatment t are weighted down one no longer needs to drop them. Figure 2.4 shows that the ratio for Andhra Pradesh is clearly dierent from that of the controls. 18 The pre-treatment RMSPE is a measure of the lack of t between Andhra Pradesh and its synthetic control and dened as: r 1 T0 P T0 t=1 Y 1t P J+1 j=2 w j Y jt 2 . The post-treatment RSMPE and the RMSPE for other states is similarly dened. 27 Figure 2.4: Ratio of post-treatment RMSPE to pre-treatment RMSPE: pcNSDP Next, we assess the robustness of our results in two ways. Firstly, we test the sensitivity of the baseline model to the states in the control pool. For this purpose we iteratively drop one of the controls that receive positive weight in the base specication and re-estimate the baseline model. 19 The results of these iterations are shown in gure 2.5. The upper panel of gure 2.5 depicts gure 2.2 superimposed with the synthetic controls estimated by iteratively leaving out one of the controls (dashed grey lines). The lower panel of gure 2.5 shows the dierence in the estimated pcNSDP gap when using all the controls (solid black line) and leaving out one of the controls (dashed grey lines). Even though the pool of the control states is relatively small, gure 2.5 shows that the results are fairly robust to the exclusion of any given control state. Secondly, we check the sensitivity of the results by using dierent combinations of 19 Dropping states that receive zero weight does not change the results of the baseline model. 28 Figure 2.5: Leave-one-out Checks: The upper panel shows the trends in pcNSDP while the lower panel shows the gaps in pcNSDP 29 predictors of pcNSDP when constructing the synthetic control. We nd that instead of percentage of urban households with access to safe drinking water, our results re- mained robust to the use of other variables such as the percentage of rural population below the poverty line, per capita credit utilization, log per capita development expen- diture, foodgrain yields, percentage of net sown area irrigated. Similarly, the results are robust to using adult literacy rate or life expectancy at birth instead of HDI or using percentage of villages electried instead of per capita electricity consumption. As mentioned earlier, the synthetic control methodology weakens the assumptions of the usual dierence-in-dierence estimator. As an additional robustness check, we present the estimated eect of the Greyhounds at the state level using the standard dierence-in-dierence methodology in table 2.3. Since all covariates are not available for the entire period we run two regressions: one for 1970-2000 (column 2) and the other for 1970-1997 (column 3). In column 2, in addition to controlling for state eects, year eects and state specic time trends, we also control for per capita electricity consumption. To this specication, in column 3 we add other covariates such as per capita development expenditure, the Besley and Burgess Labor Reform Index (to proxy the industrial labor relations in the state) and the proportion of seats held by various political groups in the state legislature. However, since these data are only available till 1997 we lose the last three years of our treatment period. 20 As the results in table 2.3 show, the estimated eects of the Greyhounds is signicant for both specications. Further, we nd that the treatment eect estimated by using the synthetic control methodology and the dierence-in-dierence methodology are quite similar: using the synthetic control methodology we nd that the estimated treatment eect is Rs.2,065.2 over 1970-2000 and Rs.1686.9 over 1970-1997 while using 20 In all the regressions reported in table 2.3 we can reject the null hypothesis that the state specic time trend is the same across all the states at the 1 percent level. 30 the dierence-in-dierence methodology it is Rs.1251.9 and Rs.2362.2, respectively. 2.4.1 Industrial Sector We begin our analysis of the channels through which the Greyhounds aected the economy by looking at the industrial sector. The industrial sector consists of mining, manufacturing, construction, gas, electricity and water supply. Within the industrial sector we also analyze the eects on the manufacturing sector and a further breakup of the manufacturing sector into registered and unregistered manufacturing. 21 Figure 2.6(a) displays the trends in industrial outputs of Andhra Pradesh and the synthetic control. The synthetic Andhra Pradesh almost perfectly replicates the per capita industrial NSDP of Andhra Pradesh over the entire pre-treatment period, followed by an immediate divergence after the introduction of the Greyhounds. This marked divergence between Andhra Pradesh and its synthetic counterpart indicates a clear positive eect of the Greyhounds on the industrial output of Andhra Pradesh. This gap translates into an average of Rs. 437.95 over the period 1989-2000 or 16.41% of the average (observed) industrial output of Andhra Pradesh during this period. 22 Similar eects are observed in the manufacturing sector. Figures 2.6(b)-2.6(d) present the eects of the Greyhounds on the manufacturing, registered manufactur- ing and unregistered manufacturing, respectively. Once again, for the manufacturing sector and the registered manufacturing sector the synthetic control closely replicates the observed output for Andhra Pradesh over the pre-treatment period. Unsurpris- ingly, the real output of the unregistered manufacturing sector uctuates a lot given 21 In India registered manufacturing sector consists of all manufacturing rms that employ more than 20 workers without using electricity or more than 10 workers and using electricity. 22 The average per capita industrial NSDP of Andhra Pradesh is Rs. 2668.74 during 1989-2000. The average percentage gap over this period is 15.66% and the synthetic control grew at an annual growth rate of 2.89% as opposed to the annual growth rate of 4.56% of Andhra Pradesh. The yearly gains are provided in appendix 2.C. 31 Table 2.3: Greyhounds and pcNSDP 1970-2000 1970-1997 Greyhounds 1251.9 2362.2 (492.8) (683.0) pc Electricity Consumption 15.22 10.71 (6.633) (4.917) log pc Dev Expenditure 491.9 (1147.4) Lag Labor Regulation -384.5 (446.8) Congress seats 1597.0 (781.0) Janta Dal parties seats 810.0 (958.2) Left parties seats -1482.2 (2888.1) Hindu parties seats 2444.2 (544.9) Regional parties seats 2455.7 (1160.2) Constant 6226.3 3533.6 (252.7) (3473.3) State eects Yes Yes Year eects Yes Yes State-time trends Yes Yes Observations 248 224 R 2 0.968 0.971 Standard errors calculated using robust standard errors clustered at the state level are reported in parentheses. The details of the variables are provided in the appendix. * signicant at 10%,** signicant at 5%,*** signicant at 1%. 32 (a) Industry (b) Manufacturing (c) Reg. Manufacturing (d) Unreg. Manufacturing Figure 2.6: Trends in per capita industry output: Andhra Pradesh vs. Synthetic Control. 33 that it largely consists of small rms with little xed capital investment. Still the synthetic control tracks the per capita unregistered manufacturing NSDP of Andhra Pradesh fairly well in the pre-treatment periods. After the introduction of Greyhounds in 1989, the dierence in the output of Andhra Pradesh and the respective synthetic controls in gures 2.6(b)-2.6(d) provide an estimate of the treatment eects. The average gaps (percentage average gaps) for the manufacturing, registered manufacturing and unregistered manufacturing sectors are Rs. 377.52 (25.73%), Rs. 202.89 (20.24%) and 112.91 (24.14%), respectively. 23 The similarity of the synthetic control to Andhra Pradesh in terms of the pre- treatment characteristics is shown in table 2.5. The synthetic control is constructed by matching on variables such as HDI, population density, percentage of households with access to safe drinking water, log per capita development expenditure, the Labor Reform Index constructed by Besley and Burgess (2004) and pre-treatment values of output. As comparisons with the simple average of the control units in table 2.5 indicate, the synthetic control approximates Andhra Pradesh far more closely. The weights assigned to the various controls in constructing the synthetic control are shown in table 2.4. We infer the signicance of the results by doing placebo tests as discussed earlier. These are shown in gures 2.7 - 2.10. The signicance of the results can also be checked by comparing the ratio of post-treatment MSPE to the pre-treatment MSPE for Andhra Pradesh to those for the placebos as shown in gure 2.11. As before, we assess the robustness of our results by using dierent combinations of predictors of industrial output (and its components) when constructing the synthetic 23 The average percentage gaps for the manufacturing, registered manufacturing and unregistered manufacturing sectors are 24.80%, 19.87% and 22.63%, respectively. Similarly, over the period 1989-2000 the manufacturing, registered manufacturing and unregistered manufacturing sectors in Andhra Pradesh (synthetic control) grew at an annual rate of 3.67% (0.92%), 2.13% (1.42%) and 4.91% (0.94%), respectively. The yearly gains are provided in appendix 2.C. 34 Figure 2.7: pc Industrial NSDP Gap in Andhra Pradesh and placebo gaps of Control States (Upper panel has all controls, lower panel discards states with pre-treatment MSPE five times higher than Andhra Pradesh) 35 Figure 2.8: pc Manufacturing NSDP Gap in Andhra Pradesh and placebo gaps of Control States (Upper panel has all controls, lower panel discards states with pre-treatment MSPE six times higher than Andhra Pradesh) 36 Figure 2.9: pc Regristered Manufacturing NSDP Gap in Andhra Pradesh and placebo gaps of Control States (Upper panel has all controls, lower panel discards states with pre-treatment MSPE five times higher than Andhra Pradesh) 37 Figure 2.10: pc Unregristered Manufacturing NSDP Gap in Andhra Pradesh and placebo gaps of Control States (Upper panel has all controls, lower panel discards states with pre-treatment MSPE two times higher than Andhra Pradesh) 38 (a) Industry (b) Manufacturing (c) Reg. Manufacturing (d) Unreg. Manufacturing Figure 2.11: Ratio of post-treatment MSPE to pre-treatment MSPE: Industrial pcNSDP 39 Table 2.4: State Weights: Industry Synthetic Control Weights State Industry Manufacturing Manufacturing Registered Unregistered Bihar 0.13 0.32 0.04 0.00 Karnataka 0.39 0.31 0.26 0.00 Madhya Pradesh 0.26 0.34 0.59 0.72 Maharashtra 0.00 0.00 0.00 0.00 Orissa 0.00 0.03 0.00 0.28 Uttar Pradesh 0.22 0.00 0.11 0.00 West Bengal 0.00 0.00 0.00 0.00 control. The results are robust to the use of adult literacy rate or life expectancy at birth instead of HDI. Instead of using the Besley and Burgess Labor Reform Index the results remain robust to the usage of other variables that capture the industrial relations environment in the state such as the number of mandays lost due to industrial disputes and the membership of labor unions that submit returns in the state. Similarly, instead of log per capita development expenditure, using other predictors of industrial activity such as the per capita consumption of industrial electricity, the percentage of population below the poverty line, road density or per capita credit utilization does not aect the signicance of the results. Sensitivity of the results to the composition of the control pool is assessed by iteratively leaving out one of the controls that is assigned positive weight and re-estimating the baseline model. The results of this test are shown in gures 2.12 and 2.13. Once again, as a further robustness check for the results, we present the estimates of the eect of the Greyhounds using the standard dierence-in-dierence methodol- ogy. Table 2.6 shows the estimated eects of the Greyhounds on industrial perfor- mance. In addition to controlling for the per capita consumption of industrial elec- 40 Table 2.5: pc Industry NSDP Predictor Means Variables Andhra Pradesh Synthetic Control Control States Industry Manufacturing Manufacturing Mean Min Max Registered Unregistered HDI 0.30 0.29 0.27 0.27 0.25 0.29 0.24 0.36 Population Density (persons/sq. km.) 176.50 271.31 211.43 185.08 119.57 303.86 106 588.5 Total HH with safe drinking water (%) 25.89 30.78 29.89 25.88 18.62 36.00 14.58 69.65 Log pc Development Expenditure 4.95 4.76 4.71 4.79 4.78 4.79 4.34 5.20 Labor Reform Index -2.26 -0.03 -0.03 -0.05 0.05 0.34 -0.05 1.42 1976 pc Ind. NSDP 1085.13 1093.30 1501.7 441.79 3302.11 1984 pc Ind. NSDP 1401.76 1430.66 1704.61 539.37 3603.37 1986 pc Ind. NSDP 1538.37 1504.56 1820.22 513.69 4072.35 1977 pc Mgf. NSDP 629.43 635.47 1081.07 332.35 2801.98 1983 pc Mgf. NSDP 730.15 768.02 1148.61 429.06 2905.30 1986 pc Mgf. NSDP 777.15 751.98 1183.81 420.94 3222.24 1979 pc Reg. Mgf. NSDP 400.31 400.70 732.54 62.90 2269.17 1982 pc Reg. Mgf. NSDP 469.91 462.67 681.04 65.86 2009.76 1986 pc Reg. Mgf. NSDP 511.44 463.84 755.40 79.04 2400.58 1974 pc Unreg. Mgf. NSDP 306.34 242.42 409.35 153.66 646.36 1983 pc Unreg. Mgf. NSDP 269.53 332.34 487.06 295.76 806.30 1988 pc Unreg. Mgf. NSDP 286.72 272.05 494.79 228.61 809.03 41 (a) Industry (b) Manufacturing (c) Reg. Manufacturing (d) Unreg. Manufacturing Figure 2.12: Leave one out: Trends in Industrial pcNSDP 42 (a) Industry (b) Manufacturing (c) Reg. Manufacturing (d) Unreg. Manufacturing Figure 2.13: Leave one out: Gaps in Industrial pcNSDP 43 tricity, we also control for state eects, year eects and state specic time trends. 24 The eect of the Greyhounds is signicant for the industrial sector and its com- ponents (manufacturing, registered manufacturing and unregistered manufacturing). Additionally, we can also control for other covariates of industrial production such as the Besley and Burgess Labor Reform Index, log per capita development expenditure and the proportion of seats help by various political groups in the state legislature. However, as mentioned earlier, as these data are only available till 1997 we lose the last three years of our treatment period. The results of adding more observed predic- tors are presented in table 2.7. Once again, the eect of Greyhounds is consistently signicant on all measures of industrial performance. Our condence in the estimated treatment eect if further boosted by comparing the results of the synthetic control and dierence-in-dierence methodologies. As the comparisons presented in table 2.8 show, the results using the two methodologies are quite similar. Table 2.6: Greyhounds and Industrial NSDP in Naxalite Violence Affected States 1970-2000 Industry Manufacturing Manufacturing Registered Unregistered Greyhounds 294.4** 238.6** 137.4* 101.6*** (124.4) (92.33) (72.03) (20.57) Industrial electricity -2.31 -0.89 -0.53 -0.61 consumption per capita (3.03) (2.42) (1.94) (1.05) State eects Yes Yes Yes Yes Year eects Yes Yes Yes Yes State time trends Yes Yes Yes Yes R 2 0.93 0.87 0.84 0.78 N 248 248 248 248 Standard errors calculated using robust standard errors clustered at the state level are reported in parentheses. The details of the variables are provided in the appendix. * signicant at 10%,** signicant at 5%,*** signicant at 1%. 24 In all the regressions reported in tables 2.6 - 2.7 we can reject the null hypothesis that the state specic time trend is the same across all the states at the 1 percent level. 44 Table 2.7: Greyhounds and Industrial NSDP in Naxalite Violence Affected States 1970-1997 Industry Manufacturing Manufacturing Registered Unregistered Greyhounds 734.8** 530.3* 310.1* 227.0** (307.8) (230.8) (164.2) (76.39) Industrial electricity 0.03 0.74 1.40 -0.71 consumption per capita (3.82) (3.29) (2.48) (0.90) Log pc dev Expenditure 270.4 155.6 150.7 -11.47 (149.5) (115.9) (84.06) (58.61) Lag labor regulation -148.3 -94.52 -47.15 -39.88 (85.37) (88.73) (56.67) (37.88) Congress Seats 291.9 249.1 -2.53 226.7** (342.3) (176.8) (146.6) (67.25) Janta Dal parties Seats 232.9 195.9 -23.04 195.0** (318.0) (162.9) (122.9) (78.19) Left parties Seats 1173.4 631.6 616.7 124.1 (779.4) (640.5) (386.5) (274.8) Hindu parties Seats 400.8 340.2 35.57 274.7*** (299.1) (182.3) (164.4) (72.19) Regional parties Seats 669.7 508.4 144.7 370.9*** (574.0) (347.2) (288.4) (86.57) State eects Yes Yes Yes Yes Year eects Yes Yes Yes Yes State time trends Yes Yes Yes Yes R 2 0.93 0.88 0.88 0.84 N 224 224 224 224 Standard errors calculated using robust standard errors clustered at the state level are reported in parentheses. The details of the variables are provided in the appendix. * signicant at 10%,** signicant at 5%,*** signicant at 1%. Table 2.8: Results Summary: Greyhounds and Industrial Performance 1970-2000 1970-1997 SCM D-i-D SCM D-i-D Industry 437.95 294.4 391.86 734.8 Manufacturing 377.52 238.6 352.54 530.3 Registered Manufacturing 202.89 137.4 208.16 310.1 Unregistered Manufacturing 112.91 101.6 92.50 227.0 45 Some disaggregated evidence We can explore the eects of the Greyhounds on the industrial sector at a more disaggregated level. For this purpose we use the Aghion et al. (2008) state-industry panel data on the registered manufacturing sector collected under the Annual Survey of Industries. The period under consideration in 1980-1997 (nine years pre- and post- treatment) and industrial variables are measured at the two-digit level. 25 We estimate the following regression: y ist = is +(Grey st ) +(X ist ) + it + ist (2.8) wherey ist is the outcome variable,Grey st is a dummy variable that takes the value of unity from 1989 onwards if the state is Andhra Pradesh,X ist are observed covariates, is are state-industry interactions to control for unobserved time-invariant factors such as location, natural resources etc., it are industry-year interactions that control for unobserved industry-year eects such as technological innovations. The outcome variables considered are logarithms of real output, number of factories, real physical capital and number of employees. Industrial licensing was implemented by the central government in order to reg- ulate the manufacturing sector. This policy was gradually reversed during the 1980s and 1990s. The variable delicensing reform indicates the fraction of existing three- digit industries within a two-digit industry that were delisenced. Another key dereg- ulation policy was the removal of restrictions on foreign direct investment (FDI). The variable FDI reform measures the fraction of existing three-digit industries within a two-digit industry that had any product opened for automatic approval of FDI (up to 25 The sampling unit in Aghion et al. (2008) is a state and three-digit industry pair. There are on average 64 three-digit industries in each state leading to a considerable amount of heterogeneity in the data. For this paper we aggregate the Aghion et al. (2008) data to the two-digit industry level. 46 51%). Further following Aghion et al. (2008), we interact the state-level measure of labor regulation with the two deregulation measures (delicensing and FDI reform) to capture the dierential eects of these policies across states. In addition to this, we also include other covariates such as log per capita development expenditure and the proportion of seats held by various political groups in the state legislature. Further details of these data are provided in appendix 2.A. The results of these regressions are in table 2.9. As the results clearly indicate, after controlling for a variety of covariates, the introduction of Greyhounds led to a signicant increase in industrial activity in Andhra Pradesh relative to other Naxalite aected states. 26 Though the eects are slightly weak for xed capital, they are particularly strong for the number of factories, employees and output. Other results broadly conrm to the existing literature. Industrial performance improved following FDI reform, industries in relatively more pro-employer states beneted more from industrial deregulation (delicensing and FDI) while states with greater representation of Left-wing parties restricted the performance of the registered manufacturing sector. 2.4.2 Services Sector We now look at the eect of Greyhounds on the services sector of Andhra Pradesh. Figure 2.14 shows the trajectory of per capita services output for Andhra Pradesh and the synthetic control. The synthetic control closely tracks the actual Andhra Pradesh from 1970 to 1987 after which there is a divergence that gradually increases over time. This gap in pc Services NSDP peaks in 2000 and is approximately equal to Rs. 1,310.7. The average yearly gap between Andhra Pradesh and the synthetic control in the treatment years is Rs. 652.80. As the average per capita services 26 Similar results are found for a specication that includes year dummies instead of industry-year dummies. 47 Table 2.9: Greyhounds and Registered Manufacturing Performance: 1980-1997 Log real output Log no. factories Log real capital Log no. employees Greyhounds 0.174 0.304 0.176 0.131 (0.0774) (0.0499) (0.0981) (0.0560) Labor Regulation 0.00374 0.0348 0.109 0.0361 (0.0462) (0.0301) (0.0549) (0.0331) Delicense -0.540 -0.143 -0.142 -0.0184 (0.282) (0.129) (0.390) (0.220) FDI reform 0.942 0.379 1.157 0.684 (0.259) (0.200) (0.401) (0.266) Delicense * Labor Regulation -0.0651 -0.0382 0.0102 -0.0827 (0.0232) (0.0141) (0.0308) (0.0178) FDI Reform * Labor Regulation -0.0802 -0.000692 -0.108 -0.0317 (0.0215) (0.0136) (0.0305) (0.0159) Congress seats -0.0619 0.00277 -0.166 0.182 (0.227) (0.159) (0.356) (0.174) Left parties seats -1.726 -1.557 -2.455 -1.085 (0.704) (0.492) (1.020) (0.581) Janta Dal parties seats 0.0414 -0.0463 -0.0908 0.312 (0.259) (0.174) (0.380) (0.200) Hindu parties seats 0.459 0.543 0.770 0.395 (0.294) (0.208) (0.502) (0.241) Regional parties seats 0.405 0.202 0.423 0.461 (0.242) (0.179) (0.351) (0.182) Log Development Expenditure 0.360 0.290 0.479 0.258 (0.298) (0.195) (0.447) (0.218) Constant 12.47 3.599 10.27 7.776 (1.427) (0.931) (2.103) (1.039) Industry-Year eects Yes Yes Yes Yes State-Industry eects Yes Yes Yes Yes P-value of joint test of: Labor Reg. interactions 0.00 0.03 0.00 0.00 Observations 2639 2639 2639 2639 Adjusted R 2 0.939 0.967 0.912 0.953 Standard errors calculated using robust standard errors clustered at the state-year level are reported in parentheses. The details of the variables are provided in the appendix. * signicant at 10%,** signicant at 5%,*** signicant at 1%. 48 output in Andhra Pradesh over this period is Rs. 5,784.81 this gap translates into a percentage average gap of 11.29%. 27 Figure 2.14: Trends in pc Services NSDP: Andhra Pradesh vs. Synthetic Control Tables 2.10 and 2.11 show the pre-treatment predictors of services NSDP and the resulting weights assigned to the controls. The treated and the controls are matched on observed covariates such as HDI, population density, road density, per capita electricity consumption, percentage of households with access to safe drinking water and log per capita development expenditure. Madhya Pradesh, Uttar Pradesh and Karnataka get assigned about a third of the weight each in constructing the synthetic control while the rest of the controls get zero weights. 27 The average percentage gap is 10.74% and the synthetic control grew at an annual growth rate of 4.39% as opposed to the annual growth rate of 5.51% of Andhra Pradesh. The yearly gains are provided in appendix 2.C. 49 Figure 2.15: pc Services NSDP Gap in Andhra Pradesh and placebo gaps of Control States (Upper panel has all controls, lower panel discards states with pre-treatment MSPE Three times higher than Andhra Pradesh) Table 2.10: pc Services NSDP Predictor Means Variables Andhra Pradesh Synthetic Control HDI 0.30 0.29 Population Density (persons/sq. km.) 176.50 281.73 Road Density (km./1000 sq. km. area) 366.00 432.30 Per capita electricity consumption (KwH) 114.35 126.13 Total HH with safe drinking water (%) 25.89 30.54 Log pc Development Expenditure 4.95 4.82 1971 pc Services NSDP (1999 prices) 2226.45 2277.91 1984 pc Services NSDP (1999 prices) 3102.50 3224.33 50 Table 2.11: State Weights: Services State Weight Bihar 0.00 Karnataka 0.46 Madhya Pradesh 0.24 Maharashtra 0.00 Orissa 0.00 Uttar Pradesh 0.30 West Bengal 0.00 Figure 2.16: Ratio of post-treatment RMSPE to pre-treatment RMSPE: pc Services NSDP 51 Signicance of the estimated treatment eect is again determined through placebo tests. The upper panel in gure 2.15 shows the estimated gaps in pc services NSDP for Andhra Pradesh and all the controls. Once again, Maharashtra has a poor t before treatment. The lower panel in gure 2.15 shows the estimated gaps once we drop the controls that have an MSPE three times greater than that of Andhra Pradesh in the pre-treatment period. Alternatively, the signicance of the result can be seen by comparing the ratios of post-treatment MSPE to pre-treatment MSPE in gure 2.16. Both tests indicate a signicant eect of Greyhounds on the services sector of Andhra Pradesh. Robustness of the results is further conrmed by (1) using other predictors of pc Services NSDP such as adult literacy rate or life expectancy at birth instead of HDI or the percentage of total population below the poverty line and per capita credit utilized instead of log of per capita development expenditure; and (2) iteratively dropping one of the controls that is assigned positive weight and re-estimating the baseline model (gure 2.17). 2.4.3 Agricultural Sector Lastly, we look at the eect of the Greyhounds on the agricultural sector. The eect on per capita agricultural NSDP is not signicant. Figure 2.18 shows the trends in per capita agricultural NSDP between Andhra Pradesh and its synthetic control. As the trajectory of pc agricultural NSDP indicates, the agricultural sector in India is still heavily dependent on rainfall. For example, the sharp dip in agricultural output in 1997 was caused due to El Ni~ no Southern Oscillation related events. Still, the syn- thetic control does a fairly good job in tracking the treated unit till 1991 after which there is a steady divergence leading to a percentage average gain of 10.99% of its per 52 Figure 2.17: Leave-one-out Checks: The upper panel shows the trends in pc Services NSDP while the lower panel shows the gaps pc Services NSDP 53 capita agricultural NSDP. 28 Once again tables 2.12 and 2.13 show the pre-treatment predictors of agricultural NSDP and the resulting weights assigned to the controls. Apart from HDI and population density, other predictors of agricultual output used are percentage of rural population under the poverty line, standard deviation of av- erage monthly rainfall and foodgrain yields (to proxy for the extent of the spread of the green revolution). Madhya Pradesh and Karnataka receive bulk of the weight in the construction of the synthetic control while the remaining weight is assigned to Bihar and Maharashtra. Table 2.12: pc Agricultural NSDP Predictor Means Variables Andhra Pradesh Synthetic Control HDI 0.30 0.30 Population Density (persons/sq. km.) 176.50 178.54 Rural population under poverty line (%) 37.68 52.99 Foodgrain Yield 960.06 798.94 Average monthly Rainfall (s.d.) 43.52 54.41 1974 pc Agricultural NSDP (1999 prices) 4409.73 4238.96 1978 pc Agricultural NSDP (1999 prices) 3338.53 3312.59 1987 pc Agricultural NSDP (1999 prices) 3129.55 3361.12 The eect, however, is not signicant as it fails the placebo tests. As gure 2.19 shows, the eect on the agricultural sector of Andhra Pradesh is indistinguishable from the eects on the control states. Similarly, comparing the ratio of post-treatment MSPE to the pre-treatment MSPE of Andhra Pradesh and the placebos indicates the insignicance of the result (gure 2.20). Figure 2.21 shows that this result is fairly insensitive to the composition of the control pool. 28 The gap reaches a maximum of Rs. 1269.23 in 2000 and the average percentage gap over the treatment period is 10.45%. Andhra Pradesh grew at an annual rate of 2.85% while the synthetic control grew at 1.47%. The yearly gains are provided in appendix 2.C. 54 Table 2.13: State Weights: Agriculture State Weight Bihar 0.15 Karnataka 0.51 Madhya Pradesh 0.34 Maharashtra 0.01 Orissa 0.00 Uttar Pradesh 0.00 West Bengal 0.00 Figure 2.18: Trends in pc Agricultural NSDP: Andhra Pradesh vs. Synthetic Control 55 Figure 2.19: pc Agriculture NSDP Gap in Andhra Pradesh and placebo gaps of Control States (Upper panel has all controls, lower panel discards states with pre-treatment MSPE Two times higher than Andhra Pradesh) 56 Figure 2.20: Ratio of post-treatment RMSPE to pre-treatment RMSPE: Agriculture 2.5 Discussion Implicit in the calculation of the economic benets of the Greyhounds is the assump- tion of no spillover between units i.e., outcomes of control units are not aected by the treatment administered to the treated unit. This assumption may be violated in the following three ways in the present context. Firstly, there may be a spillover through the security forces. For example, the Greyhounds could oer support to the police forces of other states in the form of training, joint operations, material supply, deputing ocers to other state police forces etc., to tackle the Naxalite insurgency. To the extent that the Greyhounds oered such support, the eect of the Greyhounds on pcNSDP would not have been conned to Andhra Pradesh, but would have extended to all other states whose police forces were supported by the Greyhounds. This would have articially raised the output of the other Naxalite aected states and thus biased 57 Figure 2.21: Leave-one-out Checks: The upper panel shows the trends in pc Agricultural NSDP while the lower panel shows the gaps pc Agricultural NSDP 58 the estimate of the treatment eect downwards. As the Greyhounds started oering training to the police forces of other Naxalite aected states from 2000 onwards, 29 we terminate the analysis at the year 2000 to avoid this source of bias. Secondly, anti-Naxalite operations by the Greyhounds in Andhra Pradesh could result in the movement of Naxalites from Andhra Pradesh to the surrounding Nax- alite aected states, where the absence of specially trained and equipped security forces would have meant a less threatening security environment for Naxalites. An example of such an instance occurred in 2007 when Naxalites from North Telangana and Nallamala regions in Andhra Pradesh state retreated to Dantewada district (in Chattisgarh state) in response to the increased tempo of Greyhounds operations in Andhra Pradesh (Tata, 2010). Although this incident occurred after the study period ended, such instances may have occurred during the study period as well. To the ex- tent that such displaced Naxalites indulged in insurgency activity in their temporary refuge across state borders, the per capita output of the other Naxalite aected states would have decreased (the decrease being a direct consequence of the activities of the Greyhounds force in Andhra Pradesh). This would result in an over estimate of the Greyhounds treatment eect. Unfortunately, there is no publicly available data to check the extent of such bias in the calculated treatment eect. In this case, the calculated treatment eect must therefore be interpreted as an upper bound on the true treatment eect. Lastly, spillover between the units may also happen through the migration of civilians from other Naxalite aected states to Andhra Pradesh. This migration may be selective in the sense that the civilians most likely to have the resources to migrate are more productive on average. This could then raise the per capita output in Andhra Pradesh and lower it other Naxalite aected states again leading to an upward bias in 29 This information is available on the Greyhounds webpage at www.apstatepolice.org 59 the estimated treatment eect. However, we nd that migration in India is low and is mostly intra-state in nature. Lusome and Bhagat (2006) using Indian census data report that approximately 80% of the internal migration in India is in the form of intra-district and inter-district (but within the same state) migration over the period 1971-2001. Further, in 2001 only 33.7% of the inter-state migration is for employment or business purposes. Finally, as mentioned before in section 2.2.1, that although the decision to intro- duce the Greyhounds was announced in 1988, the analysis presented in this paper takes 1989 to be the treatment year as in addition to the introduction of the Grey- hounds, there was a signicant transformation of the Andhra Pradesh State Police counter-insurgency capabilities that occurred only after 1989. We now test for the pos- sibility of an announcement eect in the traditional dierence-in-dierence framework by adding a lead of the treatment variable. The results of this check are presented in table 2.14. Columns (1) and (2) reproduce the estimated treatment eects from the baseline specications reported in tables 2.3, 2.6 and 2.7. In order to capture the announcement eect we estimate the specications reported in tables 2.3, 2.6 and 2.7, together with a lead of the treatment variable. Columns (3) and (4) report the coecients on the treatment variable and its lead. Looking at the results for pcNSDP, we see that although the coecient on the treatment variable remains large, the sig- nicant coecients on the lead indicate an announcement eect. However, we do not nd any indication of an announcement eect for the industrial, manufacturing, registered manufacturing and unregistered manufacturing sectors. In addition to the coecient on the lead treatment variable being insignicant, on comparing column (1) with (3) and column (2) with (4) we see that the coecient on the treatment variable is unaected and still signicant. The results conrm the synthetic control methodology results seen graphically in gures 2.2 and 2.6. 60 Table 2.14: Greyhounds and Announcement Effects Baseline Announcement 1970-2000 1970-1997 1970-2000 1970-1997 (1) (2) (3) (4) NSDP Greyhounds 1251.9 2362.2 857.2 1617.7 (484.5) (669.6) (584.9) (534.6) Greyhounds(t + 1) 622.9 1019.6 (203.7) (354.1) Industry Greyhounds 294.4 734.8 375.3 705.3 (124.4) (307.8) (130.0) (265.7) Greyhounds(t + 1) -92.41 40.40 (91.78) (133.2) Manufacturing Greyhounds 238.6 530.3 318.7 538.0 (92.33) (230.8) (109.4) (202.2) Greyhounds(t + 1) -101.0 -10.62 (82.07) (122.5) Reg. Manufacturing Greyhounds 137.4 310.1 198.6 337.6 (72.03) (164.2) (82.39) (154.1) Greyhounds(t + 1) -74.30 -37.63 (80.64) (100.6) Unreg. Manufacturing Greyhounds 101.6 227.0 111.0 202.7 (20.57) (76.39) (35.41) (65.69) Greyhounds(t + 1) -12.77 33.22 (25.05) (53.65) 61 2.6 Conclusion Currently in its fth decade, the Naxalite movement has largely escaped the attention of both policy makers and academics. We exploit an explicit change in the govern- ment's counterinsurgecy policy to estimate the economic benets associated with a security based response. We nd that the introduction of the Greyhounds results in large signicant gains in pcNSDP for Andhra Pradesh. Additionally, we nd that these gains come through the non-agricultural sector (industry, manufacturing, ser- vices) rather than the agricultural sector. Based on anecdotal evidence there have been numerous calls on other aected states to follow the \Andhra Pradesh model". The rigorous evaluation of the Greyhounds presented in this paper should help inform policy makers. Since the synthetic control methodology is a case-study approach, we would like to stress that the ndings do not imply that other Naxalite aected states raising security forces similar to the Greyhounds would experience eects of similar mag- nitude. Neither does it imply that a security response is more eective in terms of raising pcNSDP compared to non-security based responses. It is possible that a \carrot" counterinsurgency policy could have a similar eect. However, the ndings presented here are still interesting as they show that output and investments respond to counterinsurgency policies. 62 Appendix 2.A Data Description Data on pcNSDP (and its various components) at current prices were downloaded from EPWRFITS, 30 based on data from the Central Statistical Oce of the Min- istry of Statistics and Programme Implementation, Government of India. The data were available in four separate series, each corresponding to a dierent base year i.e. 1970-71, 1980-81, 1993-94, and 1999-00. The data was rst de ated using the Nation- master 31 GDP de ator series for India using 1999 as the base year, thus converting the data into pcNSDP at constant 1999 prices, although corresponding to dierent base years. The pcNSDP data for each base year series were calculated using slightly dierent methodologies, thus rendering them non-comparable across base years un- less suitably linked. The Directorates of Economics and Statistics of the respective state governments are responsible for transforming back series of pcNSDP data so that they are compatible with the latest base year series. However, in the absence of data from the Directorates, we were compelled to link the various base year series using the following ad hoc method. Data on overlapping years for the dierent base year series were used to link the series, using dierent \linking coecients" for each state. For example, if data on a number of years for a state was available for both the 1993-94 series and the 1999-00 series, then the linking coecient for that state was calculated as the ratio of the average value of pcNSDP at constant 1999 prices for that state for the overlapping years in the 1999-00 series to that of the 1993-94 series. Then, data for the 1993-94 series was converted into data for the 1999-00 series by multiplying all observations 30 Economic and Political Weekly Research Foundation India Time Series (EPWRFITS) is an online statistical information portal aimed at aiding and promoting research on Indian economy. It is partly funded by the University Grants Commission and is available at http://www.epwrts.in 31 Nationmaster is a statistics information portal, available at www.nationmaster.com 63 for that state for the 1993-94 series by the linking coecient. This procedure was performed for each state, thus converting all 1993-94 series data into 1999-00 data. A similar procedure was then performed for each state in order to convert 1980-81 series data into 1999-00 data by linking with the converted 1993-94 series. In this manner, all data were suitably linked so that they were expressed in terms of their 1999-00 equivalents both for de ation as well as for linking purposes. The following varibales were used as predictors of pcNSDP and its various com- ponents. Demographic- Population density in 1971 and 1981 was obtained from the Census of India 1991. The 1981 values of the Human Development Index (HDI) were ob- tained from the third volume of the 10th Five-Year Plan, published by the Planning Commission, Government of India. 32 Data on the percentage of rural/urban/total population under the poverty line in 1973, 1977 and 1983 were also obtained from the Planning Commission. Adult literacy rate (literacy of population aged 15 and over) for the years 1971 and 1981 was obtained from the Ministry of Human Resource Development. Life expectancy at birth over ve year spans 1971-75, 1976-80, 1981-85 and 1986-90 are based on data collected by the Ministry of Rural Development and were obtained from INDIASTAT. 33 Infrastructure- Road density, measured in kilometers of road per 1000 square kilometers of area, in 1971 and 1981 was obtained from the Planning Commission. Similarly, the percentage of urban/rural/total households with access to safe drink- ing water in 1981 was obtained from the Planning Commission. Yearly data on per capita total/industrial/agricultural electricity consumption (measured in kilo- watt hours) and the percentage of villages electried for the period 1970-2000, based 32 Planning Commission hereafter. 33 INDIASTAT is another online statistical information portal, available at http://www.indiastat.com. 64 on various publications of the Central Electricity Authority, were obtained from EP- WRFITS. Credit- The yearly amount of credit utilized and sanctioned for the period 1972- 1988 were published in various reports of the Reserve Bank of India (RBI). These were then de ated and converted to per capita terms. Other variables- The data on total area under irrigation, net area sown and food- grain yields (kilograms per hectare) in 1970 and 1980 were obtained from the website of the Central Statistical Oce of the Ministry of Statistics and Programme Imple- mentation, Government of India. From the EOPP data 34 we obtained the following variables - Labor Reform Index constructed by Besley and Burgess (2004); the number of mandays lost due to industrial disputes; the membership of labor unions submitting returns, log per capita development expenditure and average monthly rainfall. For the industry level regressions reported in section 2.4.1 we use the data set constructed by Aghion et al. (2008). The data from the Annual Survey of Industries is used to construct a state-industry panel on the registered manufacturing sector for the period 1980-1997. Industries are observed at the three-digit level and they only consider industries that are in the data for at least ten years and active in at least ve of the 16 major states over this period. This was then combined with their measures of delicense and FDI reform. For the purpose of this paper we aggregate their data to the two-digit industry level. Political parties are grouped into the following categories and expressed as a share of the total number of seats in the state legislature: Congress parties (Indian National Congress, Indian National Congress Urs, Indian National Congress Socialist); Janta parties (Janata Dal, Janta Party, Lok Pal); Left parties (Communist Party of India, Communist Party of India Marxist); Hindu party (Bharatiya Janata Party); Regional parties (Telugu Desam Party, Shiv 34 These data and further details are available at the EOPP website http://sticerd.lse.ac.uk/eopp/ 65 Sena, Utkal Congress, Progressive Democratic Front). All the other parties and independents form the base category. Further details are available in Aghion et al. (2008). 66 Appendix 2.B Map Figure 2.22: Map of India: Naxalite Affected States 67 Appendix 2.C Yearly Treatment Eects Table 2.15: Yearly Treatment Effects Year NSDP Industry Manufacturing Manufacturing Services Agriculture Registered Unregistered 1989 814.63 64.52 65.18 62.98 2.22 209.97 397.05 1990 805.09 335.72 249.01 166.53 28.95 221.61 96.27 1991 1286.01 319.56 347.38 309.66 97.44 366.29 163.80 1992 1156.69 137.08 211.64 129.02 53.84 333.39 256.47 1993 1464.43 287.84 301.30 163.71 122.55 597.19 331.16 1994 2180.61 529.94 382.05 202.93 90.09 860.04 432.54 1995 2720.09 645.48 536.72 262.62 159.10 852.00 723.33 1996 2630.21 610.06 603.02 327.50 132.46 685.87 695.09 1997 2124.53 596.48 476.53 248.45 145.87 730.87 134.86 1998 2781.29 474.83 509.77 221.28 159.69 806.22 836.49 1999 2685.15 583.99 425.72 185.47 153.08 859.42 443.57 2000 4133.71 669.89 421.90 154.50 209.62 1310.66 1269.22 68 Chapter 3 Tiered housing allocation with pre-announced rankings: an experimental analysis 1 This chapter studies a version of the house allocation with existing tenants problem. In this problem, a set of indivisible goods must be allocated to a number of agents. Money exchanges are not feasible and some agents may be endowed with some of the goods. Examples include the assignment of oces to faculty members, on-campus housing to students, parking spaces to employees, schedules to crew in the trans- portation industry (airlines, railroads, truck companies), and many other. Following the existing literature, in the rest of the paper we refer to the indivisible goods as `houses'. We investigate the class of one-sided matching problems where agents are parti- tioned into tiers with dierent privileges. Indeed, in the vast majority of these cases individuals belong to groups with privileges that are identical within tiers and dif- ferent across them. For example, in university departments, oces are sequentially allocated to full professors, associate professors, assistant professors, lecturers, and graduate students. Undergraduate housing is sequentially allocated to students in their senior, junior, sophomore and freshman year. Senior crew members have rst right to pick, etc. Similar and sometimes more rigid tiered structures are present in other situations such as rms, fraternities or the armed forces. In those environ- ments, individuals within a stratum are on a level playing eld, but have rights which 1 This chapter is co-authored with Juan Carrillo. 69 supersede the rights of agents in the strata below. A house allocation mechanism is a systematic procedure that assigns houses to prospective tenants, allotting at most one house to each agent. At the outset, some agents are endowed with a house (existing tenants) and some others are not (new- comers). Similarly, some houses are occupied by an agent while others are vacant. An allocation mechanism can be evaluated on four desirable properties: (a) Pareto e- ciency (the houses should be optimally allocated given the preferences of the agents); (b) fairness (the assignment should respect the priority order); (c) individual ratio- nality (an agent should be no worse-o by participating in the mechanism); and (d) strategy-proofness (agents should not benet from misrepresenting their preferences). Three leading allocation mechanisms have been proposed in the literature, each with dierent advantages. The most commonly used mechanism in real life applica- tions, the Random Serial Dictatorship with squatting rights (RSD), satises properties (a,b,d) but not (c), thereby discouraging participation which can imply substantial losses. To get around this problem Abdulkadiro glu and S onmez (1999) propose the Top Trading Cycles (TTC), a mechanism that satises properties (a,c,d) but not (b). Under this procedure, an agent may end up with a worse allocation than someone below in the priority queue, a feature that may create tensions between agents due to issues related to envy or fairness. Finally, the well-known Gale-Shapley mechanism (GS) of two-sided matching theory (Gale and Shapley, 1962) has a natural counterpart in the housing allocation problem. Guillen and Kesten (2012) show that a mechanism used at one of MIT's undergraduate dormitories (the NH4) is theoretically equivalent to GS in the context of the housing allocation with existing tenants problem. This rule satises properties (b,c,d) but not (a). Naturally, Pareto ineciencies create ex-post incentives for swaps. With the imposition of a tiered structure, an individual in a higher tier can always 70 expropriate the house of an individual in a lower one. As a result of this new hier- archical structure, two of the four properties above mentioned need to be adjusted. First, fairness must apply only to agents in the same tier. We call it tiered fairness. Second, since property rights across tiers are compromised, individual rationality can no longer be globally guaranteed. Following Title (1998), we consider the weaker notion of tiered individual rationality, which incorporates the fact that an agent may be forced to switch to another house if her endowment is preferred by another agent who belongs to a higher tier. The rst step of our analysis consists in extending the three mechanisms discussed above to a multi-tiered structure, which we denote tSD, tTTC and tGS respectively. 2 Under the proper modication of fairness and individ- ual rationality, it is straightforward to show that the multi-tiered mechanisms keep the same theoretical properties as their single tier counterparts: tSD satises all but tiered individual rationality, tTTC satises all but tiered fairness and tGS satises all but Pareto eciency (Proposition 1). Therefore, in tSD some agents should rationally opt out. 3 As for tTTC and tGS, we should observe dierences in the nal allocations between the two but not in the rates of participation or truthful revelation. Whether these predictions match the empirical behavior of agents in a controlled laboratory environment is the main subject of this research. Surprisingly, there exist only two main experimental tests on housing markets with existing tenants. The seminal paper by Chen and S onmez (2002), from now on [CS], compares RSD and TTC in an experiment with 12 agents (9 tenants and 3 newcomers) and 12 houses under asymmetric information. 4 In that paper, TTC dominates RSD 2 Since we inform the subjects of the priority queue before they make any decision, we choose to call it tiered Serial Dictatorship (tSD) instead of tRSD. 3 Following the literature, Pareto eciency is dened from the point of view of the participating agents only. This means that tSD may result in Pareto ineciencies over the set of all agents. 4 As a robustness check, in a follow up paper Chen and S onmez (2004) compare RSD and TTC under complete information (where subjects know the payos of everyone). The qualitative results are the same. 71 both in terms of eciency (88% v. 74% of full eciency) and participation rates (79% v. 47%) whereas no signicant dierences are found in truth-telling rates conditional on participation (71% v. 74%). More recently, Guillen and Kesten (2012), from now on [GK], compares TTC and GS (or, more precisely, the equivalent NH4 version) in an identical setting. Using an ordinal eciency test, they nd that GS is more likely to Pareto dominate TTC than the other way around. GS also yields more participation than TTC (78% v. 48%) and comparable truth-telling rates (80% v. 69%). 5 Neither of these papers discusses the empirical fairness of the mechanisms. Our paper introduces two main changes in the experimental designs of [CS] and [GK]. First, we consider a hierarchical structure with 12 agents divided into 4 tiers, with 2 tenants and 1 newcomer per tier. As mentioned above, a stratied population is empirically relevant for a large class of applications. It is therefore interesting to determine the impact of such structure. Second and perhaps most importantly, we communicate to the agents not only their tier but also their positions in the priority queue before eliciting their participation decision and ranking of alternatives. From a theory standpoint, this should have no eect on either tTTC or tGS, where partic- ipation and truthful revelation are dominant strategies. It should aect participation (but not ranking) under tSD since agents have extra information about the likelihood of losing their endowed allocation. In the mechanisms used in the real world, the ranking is sometimes pre-announced (e.g., NH4 dormitory at MIT) and sometimes not (e.g., dormitory allocation at Carnegie Mellon). However, the experimental liter- ature has not studied the behavioral dierences between these methods. Instead, it has always assumed no pre-announcement. Whether the announcement has practical eects on choices is an empirical question that our experiment can address. 5 The paper also compares GS and NH4 and nds no dierence in the laboratory between these two (theoretically equivalent) mechanisms. 72 There are three other methodological changes with respect to the previous papers. First, we compare all three mechanisms (tSD, tTTC, tGS) within the same framework. Indeed, the results in [CS] and [GK] suggest that behavior in the laboratory may vary even under identical protocols. It is therefore desirable to make comparisons between mechanisms in a framework that is as unied as possible. Second, we conduct 6 rounds of the game with an identical payo matrix and set of players but with random reassignment of preferences over houses and positions in the queue at the end of each round. In practice, these mechanisms are typically played only one or a few times. However, it is important to understand the eect of experience in order to provide policy prescriptions. Repetition also allow us to perform tests of individual behavior. Third, given that fairness is emphasized in the theoretical literature as a desirable property, we study which mechanism dominates the others in that dimension. Final allocations are similar in tTTC and tGS even though payos are constructed in a way that they should be dierent for two-thirds of the subjects (all the tenants) if they played according to the theory. Allocations dier more in tSD. Performance, on the other hand, is similar under all three mechanisms. In all mechanisms, most allocations are fair (more than 90% of the time), participation rates are low (around 40%) and truthful revelation rates conditional on participation are high (around 90%) resulting in signicant losses in eciency. However, we also nd interesting dierences relative to the previous literature. Contrary to [CS], tSD outperforms tTTC in terms of cardinal eciency and truthful revelation rates. Contrary to [GK], dierences between tTTC and tGS are not statistically signicant. Overall, tSD{ a theoretically inferior mechanism{ has the highest eciency because participation rates under tTTC and tGS are as low as under tSD, and truthful revelation rates under tTTC and tGS are lower than under tSD. On the other hand and as predicted by theory, tTTC has the greatest proportion of unfair allocations. 73 The most striking result of our analysis is the highly signicant impact of the position in the queue on the decision to opt out. It seems that agents with an unfavorable draw in the priority queue suspect that they are unlikely to improve their endowment and choose to `play safe,' that is, to keep their initial allocation. By denition, this is suboptimal under tTTC and tGS. 6 Such conditioning could not be done by subjects if rankings were not pre-announced, so we suspect that this variation accounts for most of the dierences with the previous literature. In any case, the rank-dependency of choices has important policy implications that deserve to be explored in further detail. In particular, it raises the issue of which of the two theoretically equivalent designs, with known or unknown priorities, performs better in practice. As for the other variants introduced, we do not nd any systematic tier eect on the behavior of subjects after controlling for potential gains of participation and opting out payos. This may not be be surprising to some readers although it did not seem a priori obvious to us. We do not nd any change in behavior over the course of the experiment under any mechanism, implying either that ineciencies are unlikely to be reduced through experience or that subjects need substantially more than six matches to learn. Still, the combination of multiple observations and no learning permits an analysis at the individual level. We nd that many subjects never opt in and few always do, suggesting that almost no subject realizes that participating is a dominant strategy in tTTC and tGS despite the transparency of the instructions. The pattern is the opposite when it comes to preference revelation: more than 80% of subjects always report truthfully and less than 10% misrepresent their preferences more than once, implying that the majority of subjects do realize the benets (or the 6 It may be rationalized with an (ad-hoc) xed cost of opting in, evaluating the alternatives or thinking through the game, although we did not explore behavioral theories of this sort. 74 simplicity) of truthful revelation. Finally, we obtain an interesting result from the interaction of multiple matches and known priority queues. Indeed, the participation behavior of half of the subjects can be rationalized by a monotone rule of the type \I participate if and only if my position in the queue isx th or above", withx2f1; 2; 3g. Other related literature. The theoretical literature on housing allocation includes the `Hierarchical Ex- change' mechanism (P apai, 2000) which characterizes the class of Pareto-ecient, reallocation-proof and group strategy-proof mechanisms. Pycia and Unver (2007) in- troduce a new class, the `Trading Cycles with Brokers and Owners', and show that all group incentive compatible and ecient mechanisms belong to that class. Un- der the tiered environment, Title (1998) develops the `Tiered Exchange' mechanism that characterizes the class of group strategy-proof, tiered individually rational, tiered envy-free and Pareto ecient mechanisms. Finally, the house allocation problem with existing tenants is related to other matching problems like the school choice problem (Abdulkadiro glu and S onmez, 2003) and the kidney exchange problem (Roth, S onmez and Unver, 2004). 7 There exists also a rapidly expanding experimental literature on matching. Studies on one-sided matching problems include Olson and Porter (1994), Chen and S onmez (2006), Pais and Pint er (2008), Calsamiglia, Haeringer and Klijn (2010) and Feath- erstone and Niederle (2011). The only study with a multi-tiered environment is the eld experiment by Baccara et al. (2012) which analyzes the eect of network exter- nalities in the housing allocation problem. A set of vacant oces are oered to faculty members partitioned into three groups. The priority queue of agents within each tier is known to everyone at the outset and the allocation is implemented using the RSD 7 For a comprehensive survey of the literature see S onmez and Unver (2011). 75 mechanism. 8 The paper nds that externalities, such as institutional, co-authorship and friendship networks, strongly aect the subjects' choices. Finally, there is an ex- perimental literature on two-sided matching (see e.g. Nalbantian and Schotter (1995), Kagel and Roth (2000), Unver (2005), Haruvy, Roth and Unver (2006) and Pais et al. (2011) and Roth and Sotomayor (1990) for a survey of the theory). 3.1 The model 3.1.1 Basic denitions Since we use a multi-tier extension of a model that is standard in the literature, we present it brie y (we refer the reader to Abdulkadiro glu and S onmez (1999) for a formal and comprehensive exposition of the single-tier version). A tiered housing allocation problem with existing tenants consists of a nite set of agents, exogenously partitioned into a nite number L of ordered sets or `tiers', indexed by l. In each tier, agents are classied as either existing tenants (already occupying a house) or newcomers (not occupying a house). There is a nite set of houses, some of which are occupied by existing tenants while the others are vacant. We assume that the total number of agents is equal to the total number of houses. Also, there is a list of (strict) preference relations for each agent over all houses. Agents require at most one house and strictly prefer occupying a house rather than not. In a tiered environment, we model the privileges associated with belonging to a higher tier in two ways. First, members of higher tiers choose houses before those in lower tiers. Second, a member of a higher tier can expropriate the house occupied by an agent in a lower tier. The outcome of a tiered allocation problem is a `matching', that is, an assignment of houses to agents such that each agent is assigned at most 8 There are no existing tenants so the concept of tiered individual rationality does not apply. 76 one house and no house is assigned to more than one agent. The typical proper- ties that a matching mechanism should promote are eciency, fairness, participation and truthful revelation. However, due to the existence of hierarchical privileges, the standard denitions of fairness and participation need to be adapted to our stratied structure. We describe how these basic properties are dened in a tiered matching mechanism M. (a) M is Pareto ecient if there is no other matching which assigns each agent a weakly preferred house and at least one agent a strictly preferred house. (b) M is tiered fair if, when an agent prefers another agent's assignment, then (i) the other agent belongs to a higher tier, or (ii) the other agent is in the same tier and is ranked higher in the priority queue, or (iii) the other agent is in the same tier and is assigned her own house. 9 (c) M is tiered individually rational if no agent gets a house that is worse than her endowment at the beginning of the allocation process for her tier. 10 (d) M is strategy proof or incentive compatible in dominant strategies if no agent can benet by unilaterally misrepresenting her preferences independently of the preferences and announcements of the other agents. Notice that Pareto eciency ensures that agents do not want to ex-post exchange their allocations. Tiered fairness corresponds to the well-known (within tier) pairwise 9 The analogue of this property in single tier is called justied envy-freeness. Title (1998) in- troduces tiered envy-freeness to describe a matching where agents prefer their allocation to that of subjects in strictly lower tiers. By denition, all the mechanisms discussed in this paper are tiered envy-free, as members of higher tiers receive their allocation before those in lower tiers and can expropriate them. 10 This corresponds to her initial endowment only if she does not lose it to someone in a higher tier. If she does lose it, then it corresponds to a randomly selected house that was either vacant or previously occupied by someone in a higher tier and vacated. In that sense and following Title (1998), an agent who loses her house to someone in a higher tier may be worse-o with her new allocation, but there is nothing she can do about it. 77 stability condition in the college admission problem. Tiered individual rationality guarantees participation conditional on the endowment inherited after the choice of agents in higher tiers. Finally, strategy proofness is dened in dominant strategies, which ensures that truthful revelation should be aected neither by the (rational or irrational) participation and announcement decision of other players nor by the knowledge of the position in the priority queue. 3.1.2 The tiered allocation mechanisms We describe the modied versions of the Top Trading Cycles (tTTC), Gale-Shapley (tGS) and Serial Dictatorship with Squatting Rights (tSD) mechanisms, which incor- porate a tier structure and a known position in the priority queue. All mechanisms M2ftTTC; tGS; tSDg have a common structure. The allocation of houses is done sequentially by tiers. We start with tier 1 while participants in tiers 2 to L wait. When the allocation for tier 1 is completed, we proceed to tier 2 while participants in tiers 3 to L wait, and so on. Starting with the top tier, the allocation for each tierl takes place in the following way. First, all the houses that have been assigned to members of tiers 1 tol1 are not available anymore. If an existing tenant's house has not been allocated to someone at a higher tier, then the agent continues to occupy her house. If, however, that agent's house has been taken by a member of a higher tier, she is compensated with a randomly selected house that was either vacant or occupied by someone in a higher tier and vacated. Second, an ordering of agents in tier l, also called a priority queue, is chosen deterministically or stochastically. The ordering is communicated to all the members of the tier, who also observe the house occupied by each agent in the tier. Third, each existing tenant in the tier decides whether to participate in the allocation 78 mechanism or opt out. Those who opt out are assigned their houses and removed from the process. Fourth, the participating agents, i.e., the existing tenants who opt in and the newcomers, report their preferences over all the remaining available houses. Fifth, using the priority queue and the submitted preferences of the agents, the allocation in tier l takes place according to the mechanism M being used. The agents and their allocations are removed from the process, and the same steps are repeated in tier l + 1. Notice that setting the priority queue before the participation decision and preference revelation of agents makes the comparison between tTTC and tGS sharper by ensuring that the equilibrium allocations under these two mechanisms are unique and dierent from each other. The procedure always follows these general principles. The next sections describe the modications included in each of the three mechanisms in order to account for the tiered structure of our problem. Tiered Top Trading Cycle (tTTC) Building on Gale's top trading cycle idea (Shapley and Scarf, 1974), Abdulkadiro glu and S onmez (1999) proposed the top trading cycle (TTC) mechanism. 11 Consider the agents in tierl who decide to opt in along with the newcomers. Using the preferences submitted and the exogenous priority queue, the allocation procedure in our tiered version works as follows. Starting from the top of the queue, assign each agent in turn her top choice from among the available houses until someone requests the house of an existing tenant in the same tier. If at that point the existing tenant whose house is demanded is already assigned a house, then do not disturb the procedure. Otherwise, 11 This version is called \you request my house - I get your turn" in Abdulkadiro glu and S onmez (1999) and shown to be theoretically equivalent to Gale's idea of top trading cycles in the context of housing allocation with existing tenants. This version of TTC has been used in the experiments of [CS] and [GK]. 79 modify the remainder of the ordering by inserting the existing tenant to the top of the queue and proceed. Similarly, insert any existing tenant in tier l who has not yet been assigned a house at the top of the queue once her house is demanded. If at any point, a loop forms, it can only be formed by existing tenants in the same tier where each tenant demands the house of the tenant next in the loop. In such cases remove all agents in the loop by assigning them the houses they demand and proceed. Once all the agents in tier l have received their allocation, move on to tier l + 1. When the number of tiers is 1, this procedure reduces to the TTC mechanism proposed by Abdulkadiro glu and S onmez (1999, p.251). Tiered Gale-Shapley (tGS) [GK] have adapted the GS mechanism to the house allocation with existing tenants problem. We simply extend that mechanism to our tiered environment. For tierl and using the exogenous priority queue, construct a priority ordering for each available house as follows. If the house is vacant or occupied by someone in a lower tier, then the corresponding queue for this house is the same as the priority queue of the agents. If the house is currently occupied by someone in the same tier, then assign the highest priority for this house to the corresponding existing agent, and assign the remaining priorities without changing the relative ordering of the remaining agents. Then, using this priority ordering for each house and the submitted preferences of the agents, apply the following version of the deferred acceptance algorithm due originally to Gale and Shapley (1962): Step 1. Each agent applies to her top choice house. For each house, look at its pool of applicants and tentatively assign the house to the agent with the highest priority according to the priority ordering for that house and reject the rest of the applicants. 80 In general, Step k. Each rejected agent applies to her next top choice house. For each house, consider its applicants at this step together with the agent (if any) who is currently tentatively placed to it. Among these, assign the house to the agent with the highest priority according to the priority ordering for that house and reject the rest. The process for tier l ends when no agent in that tier is rejected. At that point, all tentative assignments are nalized, and we move on to tier l + 1. Tiered Serial Dictatorship with Squatting Rights (tSD) The serial dictator algorithm is the simplest and most commonly employed one. In each tier l, the tSD determines the allocation by simply working its way down the priority queue, assigning each participating agent, in turn, her top choice from among the remaining available houses. The unassigned houses are then passed on to the next tier. Under tSD, tenants may not want to participate since they are not guaranteed a house that is at least as good as the one they are assigned at the beginning of the process for their tier, resulting in potential losses. However, since members of a higher tier choose before those in lower tiers, it is tiered fair. Note that once an existing agent decides to participate, truthful preference revelation is a dominant strategy. 3.1.3 Properties Allocation mechanisms are evaluated (both theoretically and empirically) according to the properties they satisfy. It is well known that for the case of a single tier, no (one-sided or two-sided) matching mechanism satises simultaneously the four properties described in section 3.1.1 (see e.g. [GK] and the references therein). For the 81 same reason, no mechanism can satisfy the four properties in our tiered environment either. The proposition below describes which properties hold under each of the tiered mechanisms. Proposition 1 For any ordering of agents in each tier, the mechanisms satisfy the following properties: - tTTC is Pareto ecient, tiered individually rational and strategy-proof; - tGS is tiered fair, tiered individually rational and strategy-proof; - tSD is Pareto ecient, tiered fair and strategy-proof. Proof. The notions of fairness and individual rationality have been adequately mod- ied to be applied to each tier sequentially. The allocation algorithm of each mecha- nism has also been modied to be applied to each tier sequentially. Since the proper- ties hold for any given tier, the sequential application by tiers implies that they must necessarily hold for the tier modied mechanism. The idea is simple. In a stratied population, the notions of fairness and partic- ipation must be modied in order to be applied to each tier sequentially, that is, in order to ensure identical privileges within tiers and hierarchical privileges between tiers. The key issue is then to nd the proper denition of fairness and participation. Once this is achieved, it is straightforward to extend the logic of the single-tier mecha- nisms to a multi-tiered environments. Not surprisingly, the same properties that have been shown for the single-tier setting also hold in the multi-tier framework. Natu- rally, whether such theoretical properties hold in practice is an empirical question. [CS] and [GK] document interesting deviations from the theoretical predictions in the single-tier case. The experimental analysis conducted in the next sections report other deviations in our multi-tier environment with known priorities and repetition. 82 3.2 Experimental Design Our experiment is designed to compare the tTTC, tGS and tSD mechanisms in terms of participation of existing tenants, truthful preference revelation by participating agents, fairness of the nal allocation and aggregate eciency of the outcome. To this purpose we conduct three treatments which dier exclusively in M, the allocation mechanism employed. For each treatment we ran three sessions with twelve subjects per session for a total of 108 subjects. The subjects were undergraduate students at the University of California, Los Angeles who were recruited by email solicitation. All sessions were conducted at the California Social Science Experimental Laboratory (CASSEL). The interaction between subjects was fully computerized using an extension of the open source software package Multistage Games. 12 No subject participated in more than one session. In each session, participants played a total of 6 matches. In the rst match, the twelve participants were randomly assigned a role, labeled 1 to 12, and divided into 4 tiers of 3 participants each. Subjects in roles 1-2-3 were in tier 1, subjects in roles 4-5-6 were in tier 2, and so on. Subjects in roles 1-2-4-5-7-8-10-11 were existing tenants while subjects in roles 3-6-9-12 were newcomers. Hence, each tier consisted of two existing tenants and one newcomer. There were twelve dierent houses to be allocated, labeled A to L. Table 3.1 shows the payo to each participant (in dollars) as a function of the house she holds at the end of the allocation process. The square bracket, [ ], indicates that 12 This contrasts with [CS] and [GK] who ran their experiment fully and partly by hand, re- spectively. We do not think that this minor modication had an impact on the subjects' behavior. On the other hand, having the game fully computerized allowed us to play multiple matches in a relatively short period of time (6 matches in 1 hour). 83 the participant is occupying that particular house at the beginning of the match. Note that due to the existence of hierarchical privileges, existing tenants in tiers 2 to 4 might not be occupying these particular houses when the allocation mechanism reaches their tier, although they will be occupying some house. The payos dier from previous experiments and the priority queues are also pre-specied and announced to the participants. Payos and queues are carefully chosen with several objectives in mind. First, to facilitate comparisons, we want all the Pareto ecient house allocations to have the same aggregate payo. In the experiment, there are four Pareto ecient allocations within each tier, giving a total of 4 4 = 256 Pareto ecient house allocations, all with an aggregate payo of $235. Second, the initial payos are such that only one tenant (role 10) is occupying her most preferred house (again, remember that endowments may change for agents in tiers 2 to 4). Third, payos range from $4 to $25 providing signicant variation. Fourth and quite importantly, with these payos and queues the equilibrium allocations under tTTC and tGS are unique and dierent from each other. Again this facilitates comparisons of behavior across mechanisms. Fifth, the least interesting participation decision corresponds to a tenant who is rst in the queue. In order to minimize that case, we never put the newcomer last in the queue. As we will develop below, satisfying all these properties has two drawbacks: the payo from not participating is higher than we would like it to be and the equilibrium outcome under tTTC is unfair for some newcomers only (about 30% of the population). Unfortunately and despite our best eorts, it was not possible to avoid those shortcomings and still satisfy the ve other objectives. All treatments are implemented as games of incomplete information. At the be- ginning of each match, each subject knows her own tier, her position in the priority queue, and her payo of holding each of the twelve houses. Subjects also know the number of tiers, the number of existing tenants and newcomers in each tier and that 84 Table 3.1: Payoff matrix Houses Tier Role A B C D E F G H I J K L 1 1 [19] 22 11 5 13 10 7 4 6 17 15 8 2 25 [22] 5 11 10 13 4 7 8 17 6 15 3 25 22 13 7 8 15 10 11 5 17 4 6 2 4 5 25 [17] 19 4 6 15 13 10 22 8 11 5 22 17 25 [19] 6 8 15 5 11 4 10 13 6 22 17 25 19 6 11 15 4 8 5 13 10 3 7 8 6 11 22 [17] 19 5 10 13 25 15 4 8 17 4 6 25 22 [19] 11 13 10 8 15 5 9 17 4 25 10 22 19 5 8 13 6 15 11 4 10 11 5 8 17 10 4 [25] 22 15 13 6 19 11 13 11 4 10 5 25 8 [17] 15 6 22 19 12 10 17 5 13 25 4 6 22 15 8 11 19 the payo tables of other participants may dier. Finally, when the allocation process reaches their tier, they know which houses are not available anymore, which house is occupied by someone in the same tier with a higher position in the priority queue (if any), which house is occupied by someone in the same tier with a lower position in the priority queue (if any), which houses are occupied by members of lower tiers, and which houses are vacant. Figure 3.1 shows a sample screenshot for an existing tenant in tier 2 who has chosen to participate. 13 In each session, the allocation of houses is done sequentially, by tiers. We start with tier 1 while participants in lower tiers wait. The existing tenants rst have the option of keeping their current house (by opting `out') or participating (by opting `in'). If they opt out, they keep their own house and the process ends for them. The existing tenants opting `in' and the newcomers are simultaneously asked to submit 13 As explained in the instructions, the status of houses are V (white) for \vacant", L (orange) for \occupied by someone in a lower tier", NA (blue) for \not available", O (green) for \occupied by oneself" and OS for \occupied by someone else in the same tier" (red if in higher position and dark green if in lower position). Subjects had an instructions sheet during the experiment to remind them the meaning of the acronyms. 85 Figure 3.1: Sample Screenshot their preferences over the remaining houses. 14 The participants are assigned houses according to the allocation mechanism M (2ftTTC; tGS; tSDg) employed in that session, and these houses become unavailable for the lower tiers. Once the allocation for tier 1 is over, we move on to tier 2. If an existing tenant has lost her house to a member of a higher tier, then she is compensated with a randomly selected house which was either vacant at the beginning of the match or previously occupied by a member of a higher tier and vacated in the process. The allocation process in tier 2 then follows the same steps as in tier 1. The process continues with tiers 3 and 4, at which point the match ends. At the end of a match, subjects are randomly reassigned a role (1 to 12). Roles 1-2-3 are always associated with tier 1, 4-5-6 with tier 2, and so on. The payos associated with each role are always the same as described in Table 3.1. Subjects in roles 1-2-4-5-7-8-10-11 are always existing tenants and subjects in roles 3-6-9-12 are always newcomers. The only dierence is that the position in the priority queue for 14 When submitting their rankings, the participants know the number of existing tenants and newcomers who have decided to opt in. 86 a participant in a given role changes from match to match. 15 Once the new roles and priorities are reassigned, subjects play the same game under the same rules and using the same allocation mechanism. In each session, subjects make decisions over a total of 6 matches. At the beginning of each session, instructions were read by the experimenter stand- ing on a stage in the front of the experiment room. The experimenter fully explained the rules with special emphasis on the details of the allocation procedure and an- swered all questions. Next, the subjects went through a practice match in order to familiarize themselves with the computer interface and procedures. Subjects had to complete an interactive computerized comprehension quiz before they could proceed to the paid matches. Subjects were then asked to make their decisions over 6 matches. After each match, subjects were randomly reassigned a role. At the end of the ses- sion, one of the matches was randomly selected and subjects were paid privately and in cash their earnings in that match. Each session lasted for an average of 1 hour and the average earnings in each session were $18 plus a show up fee of $5. Table 3.2 summarizes the details of each session. Appendix 3.B provides the instructions used in sessions 1-3 for the tTTC mechanism (instructions for the other sessions are similar and available upon request). 3.3 Results We begin with a descriptive analysis of the nal allocations and then evaluate the performance of the three mechanisms on four fronts: aggregate eciency, allocative fairness, participation rates and truthful preference revelation. 15 We changed the ordering of the queue between matches to make sure that even if an individual happens to draw the same role as in a previous match, she still faces a dierent decision problem. However, the six orderings we selected for the six matches all satisfy the properties described above. 87 Table 3.2: Session details mechanism session date # of subjects # of matches tTTC 1 7/20/10 12 6 2 7/20/10 12 6 3 7/20/10 12 6 tGS 1 7/21/10 12 6 2 7/21/10 12 6 3 7/21/10 12 6 tSD 1 7/22/10 12 6 2 7/22/10 12 6 3 7/22/10 12 6 3.3.1 Final allocations Tables 3.3, 3.4 and 3.5 show the distribution of the nal allocation of houses by roles under the three mechanisms as well as the participation rates under each role (IN). The square bracket, [.], indicates the fraction of nal allocations that coincide with initial endowments. For the tTTC and tGS mechanisms, the cells shaded in grey indicate the theoretical prediction. This rst look at the data highlights four major ndings, which we will further investigate later. First, participation rates across mechanisms are low. This is surpris- ing, especially for tTTC and tGS where participation is a dominant strategy. Second, the nal allocations under tTTC and tGS are similar despite the fact that predicted outcomes (shaded boxes) are dierent for all eight tenants. In fact, outcomes are very much driven by the subjects' initial endowments, which are closer to the predictions in tGS than in tTTC. 16 Third, there is little dispersion in the choices under tTTC 16 Remember that endowment only coincides with rst-best choice for role 10. So participants who play at equilibrium in tGS end up with their initial endowment (except for role 10!) but only after order other houses rst. 88 Table 3.3: Allocations under tTTC Houses Tier Role A B C D E F G H I J K L IN 1 1 [.94] .06 0 0 0 0 0 0 0 0 0 0 .22 2 .06 [.89] .06 0 0 0 0 0 0 0 0 0 .22 3 0 .06 0 0 .06 0 0 0 0 .83 .06 0 - 2 4 0 0 [.89] 0 0 0 0 0 0 .11 0 0 .28 5 0 0 0 [1] 0 0 0 0 0 0 0 0 .39 6 0 0 .06 0 0 0 .89 0 0 0 .06 0 - 3 7 0 0 0 0 [.83] .06 0 0 .06 .06 0 0 .50 8 0 0 0 0 .06 [.94] 0 0 0 0 0 0 .33 9 0 0 0 0 .06 0 0 .06 .11 0 .72 .06 - 4 10 0 0 0 0 0 0 [.11] .22 .17 0 0 .50 .72 11 0 0 0 0 0 0 0 [.72] 0 0 .06 .22 .28 12 0 0 0 0 0 0 0 0 .67 0 .11 .22 - and tGS. For example, we can predict the choice of subjects in tiers 1-2-3 with a 72% to 100% accuracy. Dispersion in choices is much higher for subjects in tier 4 and for all subjects under tSD. Fourth, although participation rates are also low under tSD, the distribution of nal allocations is quite dierent from that of tTTC or tGS. This is due to higher truth-telling rates. 3.3.2 Eciency We compare the eciency of the three mechanisms using a cardinal measure. To this end, we rst dene two natural benchmarks: the earnings in any of the Pareto ecient equilibria and the earnings if no existing tenant participates and all the newcomers reveal their preference truthfully (since they are alone in their tier they can trivially pick their preferred house among the remaining ones). Aggregate earnings in these two scenarios are $235 and $211 respectively. For comparison, we also report the aggregate earnings if subjects play according to theory. These are $235 under tTTC 89 Table 3.4: Allocations under tGS Houses Tier Role A B C D E F G H I J K L IN 1 1 [.94] .06 0 0 0 0 0 0 0 0 0 0 .28 2 0 [.94] 0 0 0 .06 0 0 0 0 0 0 .22 3 .06 0 0 0 0 .06 0 0 0 .83 0 .06 - 2 4 0 0 [.78] .06 0 0 0 0 .06 .11 0 0 .39 5 0 0 .11 [.83] .06 0 0 0 0 0 0 0 .39 6 0 0 .11 .11 0 0 .72 0 0 0 .06 0 - 3 7 0 0 0 0 [.94] 0 0 0 .06 0 0 0 .33 8 0 0 0 0 0 [.89] 0 .06 0 0 .06 0 .39 9 0 0 0 0 0 0 0 .06 .11 0 .83 0 - 4 10 0 0 0 0 0 0 [.28] 0 .28 0 0 .44 .67 11 0 0 0 0 0 0 0 [.67] 0 0 0 .33 .50 12 0 0 0 0 0 0 0 .22 .50 .06 .06 .17 - (by denition, since this mechanism is Pareto ecient) and $211 under tGS. For tSD, the eciency if all existing tenants were forced to participate is $235 (again by denition), and the eciency is $211 if all tenants knew the preferences of the other agents and therefore could evaluate the optimality of opting in. With these premises, we can calculate overall empirical eciency as the ratio of the sum of actual earnings to the earnings in the Pareto ecient equilibria. In order to take into account the eect of endowments, we normalize this ratio by subtracting from both the numerator and the denominator, the sum of earnings in the match if no existing tenant participates and all the newcomers reveal their preference truthfully (in this case $211). Table 3.6 reports overall normalized empirical eciency for each match in each session under each mechanism. Eciency is low. In fact, in 50% of the matches the empirical eciency is no greater than the eciency when no tenant participates. T-tests 17 show that the aver- 17 The standard errors are clustered at the session level for all the t-tests reported in this section. Additionally, all the results hold if we do not normalize the eciency. 90 Table 3.5: Allocations under tSD Houses Tier Role A B C D E F G H I J K L IN 1 1 [.61] .06 .06 0 0 0 0 0 0 .28 0 0 .56 2 0 [.83] 0 0 0 0 0 0 0 .17 0 0 .17 3 .39 .11 .06 0 0 0 0 0 0 .44 0 0 - 2 4 0 0 [.67] .06 0 0 .17 0 0 .06 .06 0 .22 5 0 0 0 [.78] 0 0 .22 0 0 0 0 0 .39 6 0 0 .22 .17 0 0 .56 0 0 0 .06 0 - 3 7 0 0 0 0 [.56] .17 0 0 .06 .06 .17 0 .61 8 0 0 0 0 0 [.78] 0 0 0 0 .22 0 .33 9 0 0 0 0 .44 .06 0 .06 0 0 .44 0 - 4 10 0 0 0 0 0 0 [.06] .11 .06 0 0 .78 .78 11 0 0 0 0 0 0 0 [.78] .17 0 .06 0 .22 12 0 0 0 0 0 0 0 .06 .72 0 0 .22 - age empirical eciency is signicantly lower than theoretical under tTTC (p< 0:01) while there is no statistical dierence under tGS (p = 0:18). Empirical eciency under tSD is signicantly lower than theoretical conditional on all agents partici- pating (p < 0:01) but signicantly higher than theoretical under full information (p = 0:04). Comparing the empirical eciency between the mechanisms, Wilcoxon- Mann-Whitney tests shows that while empirical eciencies are not statistically dier- ent between tTTC and tGS (p = 0:71), the empirical eciency of tSD is signicantly higher than that of tTTC (p< 0:01) and tGS (p< 0:05). This comes as no surprise since the analysis of the distribution of nal allocations revealed strikingly similar patterns for tTTC and tGS. A measure of greater interest, however, is that of conditional empirical eciency. The conditional empirical eciency for a match is calculated as the ratio of the sum of actual earnings to the conditional Pareto ecient earnings. The latter is dened as the sum of the earnings under the Pareto ecient allocation for each tier given 91 Table 3.6: Normalized Empirical Efficiency Eciency Session Match tTTC tGS tSD 1 1 -0.08 0.46 0.00 2 0.00 0.42 0.25 3 0.25 -0.54 0.42 4 0.00 0.00 0.79 5 0.00 0.42 0.46 6 0.46 0.21 0.00 2 1 0.08 0.00 0.25 2 0.21 0.71 0.21 3 0.00 0.00 0.21 4 0.21 0.00 -0.08 5 0.21 -0.54 0.25 6 0.00 0.75 0.25 3 1 0.00 0.00 0.54 2 -0.29 0.00 0.33 3 -0.04 0.00 0.00 4 0.00 0.00 0.58 5 0.50 0.00 0.46 6 -0.50 0.00 0.46 Overall 0.06 0.10 0.30 * Theoretical = 1.00; ** Theoretical = 0.00 *** Theoretical if all participate = 1.00, Theoretical under full information = 0.00 the allocations observed in the previous tiers. So, for example, if subjects in tier l deviate from the Pareto eciency, it aects the Pareto ecient allocation and the corresponding earnings of subjects in tier l + 1. Once again, to take into account the eect of endowments, we normalize this ratio by subtracting from both the numerator and the denominator the (conditional) sum of earnings in the match if no existing tenant participates and all the newcomers reveal their preference truthfully. Table 3.7 reports the normalized conditional empirical eciency for each match in each session under each mechanism. It also reports the normalized conditional theoretical eciency for tTTC and tGS, which is calculated as the ratio of the sum of earnings 92 under the allocation subjects would have received had they behaved according to the theory conditional on the observed allocation of the previous tiers in that match to the conditional Pareto ecient earnings and normalized in the same way as before. For tSD, we compute the same two theoretical benchmarks as before but conditional on previous behavior. Table 3.7: Normalized Conditional Efficiency tTTC tGS tSD Session Match Th. Emp. Th. Emp. Th.1 Th.2 Emp. 1 1 1.0 0.13 0.21 0.13 1.0 0.00 0.00 2 1.0 0.00 0.50 0.46 1.0 0.00 0.25 3 1.0 -0.08 0.87 0.22 1.0 0.00 0.42 4 1.0 0.00 0.00 0.00 1.0 0.00 0.79 5 1.0 0.00 0.58 0.15 1.0 0.00 0.46 6 1.0 0.46 0.00 0.21 1.0 0.00 0.00 2 1 1.0 0.09 0.00 0.00 1.0 0.00 0.25 2 1.0 0.21 0.77 0.50 1.0 0.00 0.21 3 1.0 0.00 0.00 0.00 1.0 0.00 0.21 4 1.0 0.21 0.00 0.00 1.0 0.00 -0.37 5 1.0 0.21 0.46 -0.06 1.0 0.74 0.40 6 1.0 0.00 0.21 0.33 1.0 0.00 0.06 3 1 1.0 0.00 0.00 0.00 1.0 0.00 0.54 2 1.0 0.00 0.00 0.00 1.0 0.00 0.33 3 1.0 0.20 0.00 0.00 1.0 0.00 0.00 4 1.0 0.00 0.00 0.00 1.0 0.00 0.58 5 1.0 0.43 0.00 0.00 1.0 0.21 0.13 6 1.0 -0.11 0.00 0.00 1.0 0.00 0.46 Overall 1.0 0.10 0.20 0.11 1.0 0.05 0.26 Empirical and theoretical eciency can be compared using a simple t-test. Empiri- cal eciency is signicantly lower than theoretical under tTTC (p< 0:001) while there is no statistical dierence under tGS (p = 0:20). Empirical eciency under tSD is sig- nicantly lower than theoretical conditional on all agents participating (p< 0:01) but 93 not signicantly dierent than theoretical under full information (p = 0:18). Finally, we can compare the empirical eciency between the mechanisms. Wilcoxon-Mann- Whitney tests shows that while empirical eciencies are not statistically dierent between tTTC and tGS (p = 0:85), the empirical eciency of tSD is signicantly dierent than that of tTTC (p = 0:02) and tGS (p = 0:03). Overall, the results on eciency and conditional eciency are similar and summarized as follows. Result 1 (Eciency) Normalized eciency and conditional eciency is low in all three mechanisms: 0.06-0.10 in tTTC, 0.10 in tGS and 0.26-0.30 in tSD. The empirical eciency of tSD is higher than the empirical eciency of both tTTC and tGS. The result highlighting the eciency dominance of tSD is surprising in light of the previous experimental literature. As we will discuss below, it re ects the higher tendency to reveal truthfully under tSD than under either tTTC or tGS. Finally, in appendix 3.A.1 we evaluate the relative eciency of the dierent allocation mecha- nisms using the ordinal eciency test (OET) introduced by [GK]. With this criterion, we nd that no mechanism is more likely to Pareto dominate the others. 3.3.3 Fairness The second criterion to compare mechanisms is allocative fairness. Recall from Propo- sition 1 that tTTC is not tiered fair but tGS and tSD are. Surprisingly, the previous experiments have not studied the empirical fairness of the dierent mechanisms de- spite its importance. Table 3.8 summarizes the theoretical and empirical fairness for each observation of each subject broken down by mechanism and house endowment (tenant vs. newcomer). Recall that for each mechanism we have 3 sessions with 6 matches each and 12 subjects for a total of 216 individual observations, of which 94 two-thirds are existing tenants and one-third are newcomers. Just like we did for e- ciency, we construct the theoretical fairness prediction not from an ex-ante viewpoint but, instead, conditional on the allocations chosen by subjects in the tiers above. By construction, under tGS and tSD the equilibrium is always tiered fair, independently of the choices by previous subjects. Under tTTC, choices are predicted to be fair for all the 144 observations of existing tenants but only for 10 out of 72 observations of newcomers. Table 3.8: Tiered fairness by mechanism and endowment tTTC tGS tSD cond. theor. empirical cond. theor. empirical cond. theor. empirical tenant 144/144 139/144* 144/144 136/144*** 144/144 137/144*** (1.0) (.97) (1.0) (.94) (1.0) (.95) newcomer 10/72 60/72*** 72/72 67/72** 72/72 69/72 (.14) (.83) (1.0) (.93) (1.0) (.96) total 154/216 199/216*** 216/216 203/216*** 216/216 206/216*** (.71) (.92) (1.0) (.94) (1.0) (.95) (percentages displayed in parenthesis) *, **, ***: Signicantly dierent from the theoretical prediction at 90%, 95% and 99% condence level (sign test) When the conditional theory prescribes full fairness (tGS, tSD and existing tenants under tTTC), the empirical proportion of fair allocations is very high (.93 or above) though, in all but one case, they are statistically lower than the theoretical prediction. Fairness drops but remains high (.83) when the theory predicts low levels of fairness (newcomers under tTTC). The comparison between mechanisms is perhaps more revealing. T-tests of proportions show that for newcomers the proportion of tiered fair allocations is signicantly lower under tTTC than under tGS (p = :04, one- sided) or tSD (p = :01, one-sided) but not signicantly dierent between tGS and tSD (p = :47, two-sided), just like predicted by theory. The same t-test for existing 95 tenants shows no signicant dierences of fairness across the three mechanisms, again as predicted by theory. Overall, the inferiority of tTTC relative to tGS and tSD in tiered-fairness highlighted in the literature is not as strong in our experiment as the theory predicts but it is still signicant. The results on fairness can be summarized as follows. Result 2 (Fairness) As predicted by theory, the proportion of unfair allocations is very small under tGS and tSD and signicantly larger for newcomers under tTTC. 3.3.4 Participation Rates Next we compare the mechanisms on the basis of participation rates. Table 3.9 reports the participation rates under the dierent mechanisms broken down by session, match, tier and position in the priority queue. In general, the participation rate of existing tenants is low: 36.8% for tTTC, 39:6% for tGS and 41.0% for tSD. T- test of proportions shows that the overall participation rates are not statistically dierent between the three mechanisms. Furthermore, dierences in participation rates between mechanisms are not statistically signicant for any given tier, any match, any position in the queue or for existing tenants who have or have not lost their original house. Notice that no-participation may be optimal for some subjects under tSD because participants are not guaranteed their endowments. Under tGS, most tenants are indierent not ex-ante but at least in equilibrium between participating and not since in both cases they end up with their endowment. This could possibly explain the dierences in participation rates between our experiment and [GK] under tGS. However, we nd no compelling reason for non-participation under tTTC, where subjects could do substantially better by opting in. 96 Table 3.9: Existing tenants' participation decision tTTC tGS tSD Session 1 21/48 (.44) 25/48 (.52) 21/48 (.44) 2 21/48 (.44) 15/48 (.31) 18/48 (.38) 3 11/48 (.23) 17/48 (.35) 20/48 (.42) Tier 1 8/36 (.22) 9/36 (.25) 13/36 (.36) 2 12/36 (.33) 14/36 (.39) 11/36 (.31) 3 15/36 (.42) 13/36 (.36) 17/36 (.47) 4 18/36 (.50) 21/36 (.58) 18/36 (.50) Match 1 6/24 (.25) 9/24 (.38) 7/24 (.29) 2 11/24 (.46) 8/24 (.33) 10/24 (.42) 3 9/24 (.38) 7/24 (.29) 8/24 (.33) 4 8/24 (.33) 9/24 (.38) 11/24 (.46) 5 9/24 (.38) 10/24 (.42) 11/24 (.46) 6 10/24 (.42) 14/24 (.58) 12/24 (.50) Priority 1 29/45 (.64) 24/45 (.53) 29/45 (.64) 2 9/27 (.33) 11/27 (.41) 14/27 (.52) 3 15/72 (.21) 22/72 (.31) 16/72 (.22) Original House Lost 16/19 (.84) 13/18 (.72) 14/20 (.70) Not Lost 37/125 (.30) 44/126 (.35) 45/124 (.36) Overall 0.37 0.40 0.41 (percentages displayed in parenthesis) Result 3 (Overall Participation) Participation rates are low in all three mecha- nisms: 36.8% for tTTC, 39.6% for tGS and 41.0% for tSD. Dierences across mech- anisms are not statistically signicant. Next we look at the determinants of participation rates within mechanisms. First, standard t-tests of proportions performed on the data in Table 3.9 reveals that tenants in tiers 3 and 4 are more likely to participate than tenants in tiers 1 and 2 under all three mechanisms (p = 0:025 for tTTC, p = 0:061 for tGS, p = 0:062 for tSD). Participation is particularly low in tier 1. The test, however, does not control for other factors, such as the payo of opting out (which depends on her endowment at the beginning of the allocation process for her tier) or the maximum possible 97 gain. Second and again without controlling for other factors, we nd that under all mechanisms an existing tenant is more likely to participate when she is higher in the priority queue. On aggregate, participation rates in priority positions 1, 2 and 3 are 61%, 42% and 24% respectively. Also, t-tests of proportions show that participation rates are signicantly higher in position 1 than in 2 or 3 for tTTC (p = 0:01 and p< 0:001), in position 1 than in 3 for tGS (p = 0:014) and in positions 1 and 2 than in 3 for tSD (p < 0:001 and p = 0:004). The result is in accordance to theory for tSD, where participation is a dominant strategy for the rst agent in the queue and becomes probabilistically less interesting as the likelihood of losing the endowment to someone higher in the queue increases. Under tTTC and tGS participation is (weakly) dominant independent of the position. Overall, it suggests that participants are reluctant to opt in if they rationally (for tSD) or irrationally (for tTTC or tGS) believe that participation may result in a loss. Third, we analyze the evolution of participation over the course of the experiment. T-tests of proportions reveal no signicant dierences in participation rates between match 1 and match 6, between matches 1-2 and matches 5-6 or between matches 1-2-3 and matches 4-5-6, under any mechanism. One could a priori think that agents would learn through repetition that participation is always optimal under tTTC and tGS. This does not seem to be the case in our experiment. A possible reason is that we provide limited feedback to our subjects. Indeed, we only inform them of the nal vector of allocations and the subject's own payo. Our study being the rst to introduce repetitions, we think it is interesting and surprising to notice the lack of change in behavior between matches. However, future research should study in more detail the determinants of learning. We now turn to examine other factors that may be aecting the decision of agents to participate. Table 3.10 shows the number of agents who occupy a house (# tenants) and their participation rate (in) as a function of two variables: (i) the payo of opting 98 out (own house), that is, the value of the house they occupy when deciding whether to participate or not, and (ii) the potential gain of opting in (max. gain in) dened as the maximum possible net gain from participation. Table 3.10: Participation as a function of maximum gain in and own house payoff tTTC tGS tSD # tenants in # tenants in # tenants in Own house 25 2 .00 5 .20 1 1.00 22 18 .22 18 .22 20 .20 19 64 .42 58 .38 64 .47 17 51 .31 51 .41 51 .35 15 6 .50 9 .67 7 .86 10 1 1.00 - - - - 8 - - - - 1 .00 6 1 1.00 1 1.00 - - 5 - - 1 1.00 - - 4 1 1.00 1 1.00 - - Max. gain in 9 3 1.00 3 1.00 1 .00 7-8 6 .50 9 .56 9 .67 5-6 21 .43 21 .43 16 .44 3-4 64 .39 57 .35 65 .42 0-2 50 .26 54 .37 55 .35 Outside option and potential gains have the expected eect: participation by existing tenants increases as the payo of keeping their own house decreases and as the maximum attainable gain from participation increases. This suggests that even if subjects in tTTC and tGS do not always participate contrary to their best interest, the comparative statics on participation retain a rational avor. Indeed, 99 their behavior could be rationalized if we assumed that subjects have a xed cost of evaluating options and choosing an ordering: they will be more likely to spend this eort as the potential net gain of opting in increases. In order to investigate in more detail the determinants of participation, we perform two logit regressions where the dependent variable is a discrete choice variable that takes value 1 if the existing tenant opts in and 0 otherwise. The rst specication uses the maximum possible payo gain as an explanatory variable (columns 2-4) whereas the second specication uses the payo of keeping her own house (columns 5-7). The common set of independent variables include dummies for tiers, position in the queue, matches and whether the existing tenant has lost her original house. Table 3.11 reports the results. The rst specication conrms to a large extent the previous results on mean comparisons and highlights one of the major results of our analysis: after controlling for maximum possible gain, participation is negatively aected by being last in the priority queue. Maximum gain has a positive eect on participation (although only under tTTC). The second specication is also informative. Position in the prior- ity queue has still a signicant eect on participation and this eect appears to be stronger under tGS when the subject is in a lower tier. Payo in own house has a signicant negative impact on participation under tTTC. Finally, under either spec- ication, there is no eect on participation over time, no systematic eect of being in a lower tier and no impact of losing the original house (including session dummies do not change the results). The determinants of participation are summarized in the following result. Result 4 (Determinants of Participation) Participation increases when agents are higher in the priority queue. Participation also increases when the maximum 100 Table 3.11: Logit models of participation decisions tTTC tGS tSD tTTC tGS tSD Tiers 3&4 0.054 0.687 1.332 -0.296 0.320 1.190 (0.532) (0.355) (0.822) (0.528) (0.508) (0.659) Position 2 -0.908 -0.114 -0.138 -0.852 -0.161 -0.122 (0.732) (0.544) (0.388) (0.688) (0.547) (0.400) Position 3 -2.156* -0.631* -1.131* -2.284* -0.838** -1.152** (0.973) (0.248) (0.512) (0.962) (0.259) (0.405) Position 3*Tiers 3&4 1.259 -0.139 -1.227 1.403 -0.182** -1.233 (0.944) (0.181) (0.859) (0.932) (0.035) (0.794) Last 3 matches 0.056 0.645 0.667 0.048 0.651 0.657 (0.377) (0.450) (0.362) (0.359) (0.439) (0.382) Lost house 1.360 0.555 -0.007 1.643 0.526 0.115 (1.365) (0.566) (0.820) (1.313) (0.836) (0.688) Max. possible gain 0.153* 0.159 0.081 (0.078) (0.101) (0.095) Own house payo -0.080* -0.195 -0.050 (0.034) (0.107) (0.114) Constant -0.453 -1.356** -0.780 1.701** 3.007 0.479 (0.467) (0.354) (0.746) (0.131) (2.537) (1.981) N 144 144 144 144 144 144 Pseudo R 2 0.191 0.100 0.166 0.186 0.111 0.164 Standard errors clustered at session level in parenthesis. *, **: Signicant at 95% and 99% condence level, respectively. possible gain is high and the payo of opting out is low. It does not change over time or across tiers. One advantage of having subjects play multiple matches is that we can study behavior at the individual level. In particular, it is instructive to determine whether participation is bimodal (some individuals always participating and others never do) or spread. Table 3.12 presents the frequency of participation as a function of the number of times a subject was assigned the role of existing tenant. There are more subjects who never participate than subjects who always do (26 vs. 10). The lower average participation rate under tTTC reported in Result 3 may be driven by the 101 important number of subjects who always opt out (13). Such behavior may be due to a conservative strategy by individuals with problems to understand the rules governing the allocation mechanism, although such interpretation is speculative. Table 3.12: Frequency of participation Participation Frequency Participation Frequency Participation Frequency tTTC tGS tSD 0 1 2 3 4 5 6 0 1 2 3 4 5 6 0 1 2 3 4 5 6 Tenant Frequency 1 - - - - - - - - - - - - - - 1 0 - - - - - 2 3 2 0 - - - - 0 0 1 - - - - 1 1 1 - - - - 3 3 1 2 1 - - - 2 5 2 2 - - - 2 3 1 1 - - - 4 3 1 2 5 0 - - 3 4 3 2 2 - - 0 6 1 3 1 - - 5 3 1 3 0 2 0 - 0 4 2 1 0 0 - 3 2 1 4 1 1 - 6 1 1 0 1 1 0 0 1 1 0 0 1 0 0 0 2 0 0 0 0 0 Further analysis of the individual participation decisions reveals some interesting patterns. As developed above, priority in the queue is a crucial explanatory vari- able. In Table 3.13, we focus on monotone strategies based on queue position. We categorize subjects into those who: (i) always participate; (ii) participate only when rst or second in the queue; (iii) participate only when rst in the queue; (iv) never participate. 18 This classication explains the behavior of 44% to 64% of the subjects depending on the mechanism. As discussed above, these types are all characterized by some level of rationality, since they realize that the likelihood of improving their allocation by opting in is increasing in their position in the queue. 18 Some subjects never play in a certain position. So, for example, an agent who chooses IN when rst, OUT when second and never plays in third position would be classied as a type (iii). Note that agents who chooses IN when rst, OUT when third and who never play in second position could be classied as either (ii) or (iii). We put them in (ii). 102 Table 3.13: Classification of individual participation behavior tTTC tGS tSD (i) Always in 1 6* 4 (ii) in if 1st or 2nd 7* 3 6* (iii) in only if 1st 2 1 2 (iv) Always out 13 6 7 Total 23 (64%) 16 (44%) 19 (53%) * Includes one individual who did not participate when occupying her most preferred house Result 5 (Individual Participation) Only 9% of subjects always participate in the mechanism whereas 24% never do. We can classify 54% of subjects according to a monotone participation rule based exclusively on their ranking in the priority queue. 3.3.5 Truthful Preference Revelation Truthful preference revelation is a dominant strategy for agents participating in any of the three mechanisms under consideration. Table 3.14 presents for each mechanism the proportion of truthful announcements by session, tier, match, position in the queue, and house endowment. Even though participants submit rankings over all available houses, we only consider the relevant rankings. Indeed, ifn agents choose to participate in a given tier, they can only end up in a house ranked 1 ton, so we restrict attention to preference revelation over that set. 19 The overall proportion of truthful revelation is high: 86.4% under tTTC, 87.6% under tGS and 94.7% under tSD. T- tests of equality of proportions show that truthful revelation is higher under tSD than under tTTC (p = 0:023) or tGS (p = 0:045) whereas no signicant dierences 19 The case where all rankings are considered is discussed in appendix 3.A.2. Naturally, as we enlarge the set of houses ranked, the likelihood of non-truthful revelation (on purpose, due to inatten- tion or just due to a typographical mistake) increases. However, we show that the same qualitative properties hold under the full ranking specication. 103 are found between tTTC and tGS (p = 0:777). Higher truth-telling rates are partly responsible for the dierences in nal allocations between tSD (Table 3.5) and tTTC or tGS (Tables 3.3 and 3.4) and also for the dierences in eciency (Result 1). When we compare truth-telling rates across mechanisms for subsets of the data, the most signicant dierence is that newcomers are more likely to tell the truth under tSD than tTTC (p = 0:071) while participating existing tenants are more likely to tell the truth under tSD than tGS (p = 0:041). Table 3.14: Proportion of truthful preference revelation (relevant ranking) tTTC tGS tSD Session 1 38/45 (.84) 40/49 (.82) 45/45 (1.0) 2 41/45 (.91) 33/39 (.85) 39/42 (.93) 3 29/35 (.82) 40/41 (.98) 40/44 (.91) Tier 1 21/26 (.81) 24/27 (.89) 29/31 (.94) 2 29/30 (.97) 25/32 (.78) 27/29 (.93) 3 27/33 (.82) 29/31 (.94) 34/35 (.97) 4 31/36 (.86) 35/39 (.90) 34/36 (.94) Match 1 14/18 (.78) 19/21 (.91) 19/19 (1.0) 2 20/23 (.87) 15/20 (.75) 21/22 (.96) 3 18/21 (.86) 18/19 (.95) 20/20 (1.0) 4 17/20 (.85) 19/21 (.91) 22/23 (.96) 5 20/21 (.95) 19/22 (.86) 19/23 (.83) 6 19/22 (.86) 23/26 (.89) 23/24 (.96) Priority 1 50/56 (.89) 45/51 (.88) 55/56 (.98) 2 46/54 (.85) 52/56 (.93) 54/59 (.92) 3 12/15 (.80) 16/22 (.73) 15/16 (.94) Endowment newcomer 60/72 (.83) 64/72 (.89) 67/72 (.93) tenant 48/53 (.91) 49/57 (.86) 57/59 (.97) Overall 0.86 0.88 0.95 (percentages displayed in parenthesis) Next, we look at the determinants of truthful preference revelation within mecha- 104 nisms. Given the high rates of truth-telling, dierences in behavior when we partition the data are likely not to be signicant or to be signicant but small in magnitude. For example, under tGS truth-telling rates are higher in positions 1 and 2 than in 3 (p = 0:101 and p = 0:017 respectively). The position in the priority queue has no eect under tTTC or tSD although, in the latter case, it is mainly because agents almost always reveal truthfully. Also, in general we do not nd statistically signi- cant dierences under any mechanism when we look at truth-telling rates by matches or when we compare the behavior of newcomers and existing tenants. As in section 3.3.4, we also perform logit regressions where the dependent variable is a discrete choice variable that takes value 1 if the agent reveals his ranking truthfully and 0 otherwise. The regressions are not very informative due to the little variation in the truth-telling rates (results omitted for brevity). Result 6 (Overall Truth-telling) Truth-telling is high in all three mechanisms: 86.4% for tTTC, 87.6% for tGS and 94.7% for tSD. Truth-telling is higher under tSD than under tTTC or tGS and it is not aected by position in the queue, tier or match. As argued before, multiple matches allow us to perform an analysis at the individ- ual level. Table 3.15 presents the frequency of truth-telling by subjects as a function of the number of times the subject submitted preferences over houses. Less than 10% misreport their preferences more than once whereas 80% always report truthfully. It suggests that the observed aggregate dierences in truth-telling across mechanisms are mostly driven by the behavior of a few subjects. We conjecture that the dierences may re ect the greater diculty for some individuals to understand the subtleties of the mechanisms. In our nal individual analysis, reported in Table 3.16, we classify heuristically 105 Table 3.15: Frequency of truth-telling (relevant ranking) Truth-telling Frequency Truth-telling Frequency Truth-telling Frequency tTTC tGS tSD 0 1 2 3 4 5 6 0 1 2 3 4 5 6 0 1 2 3 4 5 6 Submission Frequency 1 0 4 - - - - - 0 1 - - - - - 1 4 - - - - - 2 1 2 1 - - - - 0 0 7 - - - - 0 0 2 - - - - 3 0 2 1 5 - - - 0 0 2 6 - - - 0 0 0 9 - - - 4 0 0 0 1 6 - - 0 1 0 1 8 - - 0 1 0 1 7 - - 5 0 0 1 0 1 9 - 0 1 0 0 0 3 - 0 0 0 0 2 5 - 6 0 0 0 1 0 0 0 1 0 0 0 0 0 4 0 0 0 0 0 0 4 into categories the 40 instances of misrepresentation. We nd 5 categories with 5 to 10 observations each. In (i), subjects rank one of the vacant house rst and then their true preference order. In (ii), subjects switch their top two choices. In (iii), subjects list their own house rst and the others in no discernible order. In (iv), subjects follow some pattern in the rankings but there is not enough data to put them in a separate category (for example, a subject second in the priority queue submits the rst two rankings truthfully and the third one randomly). In (v), subjects with no discernible pattern are lumped together. 20 One could argue that (i), (iii) and possibly (iv) follow some reasonable logic, and (v) corresponds to individuals who did not pay attention or did not understand the game. The most puzzling behavior is (ii), which would vaguely correspond to a subject who follows a `psychological win' (but admittedly very strange) behavioral strategy of the type `either I win by getting the rst ranked choice or I win by getting my truly preferred house.' Result 7 (Individual Truth-telling) 9% of subjects misreport their preferences more than once and 80% always report them truthfully. Less than 4% of the observa- tions follow an indiscernible pattern. 20 These categories are similar (but not identical) to those described in [CS]. 106 Table 3.16: Classification of Misrepresentations tTTC tGS tSD Total (i) Vacant house 5 2 2 9 (ii) Switch-top-two 4 3 1 8 (iii) Random after own 2 4 2 8 (iv) Unclassied 2 3 0 5 (v) Random 4 4 2 10 Total 17 16 7 40 3.4 Conclusion In this paper we have studied in the laboratory the tiered housing allocation prob- lem with existing tenants. We have evaluated the performance of three well-known mechanisms {Top Trading Cycles, Gale-Shapley and Random Serial Dictatorship{ on four fronts: eciency, fairness, participation and truthful revelation. Contrary to some of the literature, performance is similar across mechanisms with an eciency advantage for tSD due to higher truthful revelation rates and a fairness disadvantage for newcomers under tTTC. We have also introduced three novelties in our analysis: tiered structure, multiple matches and known priority queues. Three results have important policy implications for the future design of matching mechanisms in practical settings. First and foremost, announcing the priority order- ing before eliciting choices has a signicant eect on behavior, even in mechanisms where participation and truthful revelation are dominant strategies. Second, playing the same game a few times is unlikely to improve behavior. Probably, it would be necessary to provide substantial feedback to observe changes, although more research is needed to test this hypothesis. Finally, the individual analysis suggests that ex- tremely few subjects realize the dominance of participation. At the same time, the 107 majority of individuals who play out-of-equilibrium still follow some \reasonable" strategies. 108 Appendix 3.A Additional analyses 3.A.1 Ordinal Eciency Test [GK] suggest an ordinal eciency test (OET) to evaluate the relative eciency of dierent allocation mechanisms. To compare the allocations under two mechanism, say tTTC and tGS, the test works as follows. We pick each outcome under tTTC and Pareto compare it to each outcome under tGS. We then count the number of times tTTC dominates tGS and the number of times tGS dominates tTTC, and use a sign rank Wilcoxon test for equality of the matched pairs of dominations (comparisons between tTTC and tSD and between tGS and tSD are done is a similar manner). [GK] argue that the OET is the most appropriate test since we are concerned with eliciting ordinal (not cardinal) preferences from agents. While we concur with the argument, it is also true that (non-equilibrium) deviations from truthful revelation are likely to be aected by the cardinality of payos. Hence, empirical performance should take into account such an eect. More importantly, in [GK] the ordering is announced after the agents have made their decision. The eciency comparison can then be performed with respect to the 10,000 possible priority orderings. In contrast, we announce the ordering before eliciting preferences, so it only makes sense to consider that priority ordering. This results in a dramatic decrease in the number of observations which reduces the statistical power of the OET test. Indeed, since we have 18 matches under each mechanism, we get a total of 18 18 = 324 paired comparisons. Performing this test, we are unable to reject the null hypothesis of equality in the number of times that one mechanism dominates the others. Out of 324 comparisons, tTTC dominated tGS 37 times and tGS dominates tTTC 14 times (p = 0:448); tGS never dominates tSD and tSD dominates tGS 1 time (p = 0:317); 109 nally, tTTC never dominates tSD and tSD never dominates tTTC. 3.A.2 Preference Revelation: Full Ranking Analysis In this section we check the robustness of our preference revelation analysis by con- sidering all the rankings submitted by subjects. Naturally, the overall proportion of truthful preference revelation decreases. It becomes 81% under tTTC, 78% under tGS and 82% under tSD, which for tTTC and tSD are still above those found in the previous literature. Table 3.17 shows the analogue of Table 3.14 when all rankings are considered. We nd that most of the drop in truth-telling happens in tiers 1 and 2. This is not surprising: in higher tiers subjects have to submit rankings over a larger set of available houses, so they are more liable to make \mistakes" after the rst few (relevant) rankings. Overall, there are 385 observations where individuals submit preferences. In 345 cases, subjects reveal truthfully the relevant rankings (Table 3.14) and in 310 cases they reveal truthfully the entire ranking (Table 3.17). Therefore, there are 35 obser- vations where truthful rankings are submitted over the relevant ranking but not over the entire one. In Table 3.18 we take a closer look at these observations. As we can see, in almost half of these observations subjects make a mistake either in the last two rankings or in the ranking immediately after the relevant ones. 110 Table 3.17: Proportion of truthful preference revelation (full ranking) tTTC tGS tSD Session 1 36/45 (.80) 38/49 (.78) 40/45 (.89) 2 37/45 (.82) 29/39 (.74) 30/42 (.71) 3 28/35 (.80) 34/41 (.83) 38/44 (.86) Tier 1 16/26 (.62) 17/27 (.63) 22/31 (.71) 2 28/30 (.93) 21/32 (.66) 18/29 (.62) 3 26/33 (.79) 28/31 (.90) 34/35 (.97) 4 31/36 (.86) 35/39 (.90) 34/36 (.94) Match 1 14/18 (.78) 17/21 (.81) 17/19 (.90) 2 19/23 (.83) 14/20 (.70) 17/22 (.77) 3 17/21 (.81) 16/19 (.84) 17/20 (.85) 4 16/20 (.80) 18/21 (.86) 21/23 (.91) 5 18/21 (.86) 17/22 (.77) 17/23 (.74) 6 17/22 (.77) 19/26 (.73) 19/24 (.79) Priority 1 46/56 (.82) 40/51 (.78) 44/56 (.79) 2 44/54 (.82) 46/56 (.82) 51/59 (.86) 3 11/15 (.73) 15/22 (.68) 13/16 (.81) Endowment newcomer 57/72 (.79) 55/72 (.76) 58/72 (.81) tenant 44/53 (.83) 46/57 (.81) 50/59 (.85) Overall 0.81 0.78 0.82 (percentages displayed in parenthesis) Table 3.18: Misrepresentations after relevant ranking tTTC tGS tSD Total Mistake in last two ranks only 0 4 2 6 Mistake immediately after relevant ranks 3 3 5 11 Other mistakes 4 5 9 18 Total 7 12 16 35 111 Appendix 3.B Sample experimental instructions 3.B.1 Instructions for tTTC mechanism This is an experiment in the economics of decision making and you will be paid for your participation in cash at the end of the experiment. The entire experiment will take place through computer terminals, and all interaction between participants will take place through the computers. You will remain anonymous to me and to all the other participants during the entire experiment; the only person who will know your identity is the Lab Manager who is responsible for paying you at the end. It is important that you not talk or in any way try to communicate with other participants during the experiment. Remember that you are not being deceived and you will not be deceived: everything I tell you is true. In this experiment we are going to simulate a house allocation process. The procedure and payment rules will be described in detail below. We will start with a brief instruction period. During the instruction period, you will be given a complete description of the experiment and will be shown how to use the computers. You must take a quiz after the instruction period, so it is important that you listen carefully. If you have any questions during the instruction period, raise your hand and your question will be answered so everyone can hear. If any diculties arise after the experiment has begun, raise your hand, and an experimenter will come and assist you. Dierent participants may earn dierent amounts. What you earn depends partly on your decisions, partly on the decisions of others, and partly on chance. At the end of the session, you will be paid the sum of what you have earned in the experiment and a show-up fee of $5. Everyone will be paid in private and you are under no 112 obligation to tell others how much you earned. The experiment consists of 6 matches. The procedure in each match is exactly the same and is as follows: The Procedure is as follows: There are 12 participants divided into 4 tiers. Your participation ID and TIER is mentioned on your screen. [SLIDE #2] In each tier there are 3 participants. 2 of them are EXISTING tenants, that is, they currently occupy a house. 1 of them is a NEWCOMER who does not have a house yet. In all there are 8 existing tenants and 4 newcomers. Your ROLE of existing tenant or newcomer is also mentioned on your screen. [SLIDE #2] There are 12 houses labeled A-L to allocate. [SLIDE #2] Each house must be allocated to one and only one participant. Your payo for the match, denominated in dollars, depends on the house you hold at the end of the match and it is given in the payo table like this [SLIDE#2]. For example, if you hold house K at the end of the match, then your payo is $17. Should you be the current tenant of a house, then this fact is also indicated on your computer screen [SLIDE #2]. Note that dierent participants might have dierent payo tables and these payos are privately known. In the experiment, the allocation of houses is done sequentially, by tiers. We start with tier 1 while participants in lower tiers wait. When the allocation is done for tier 1 we proceed to tier 2 while participants in lower tiers wait and so on. 113 The match ends when the allocation process for all the tiers is over and we move on to the next match. For the next match the computer randomly reassigns the tiers and the roles of existing tenants and newcomers. The new assignments do not depend in any way on the past decisions of any participant including you and are done completely randomly by the computer. The second match then follows the same rules as the rst match. This continues for 6 matches after which the experiment ends. At the end of the experiment the computer randomly selects with equal proba- bility one of the 6 matches and your payo in the experiment is equal to your payo in the match selected by the computer. Classication of Houses: The houses are classied into the following categories indicated by their STATUS [SLIDE #2]. There are 2 main categories: Not Available and Available. NOT AVAILABLE houses are houses which have already been assigned and are no longer available for allocation. These are indicated by the color BLUE and labeled as NA. The AVAILABLE houses can be further classied into: If the house is occupied by you, then it is indicated by the color LIGHT GREEN and labeled (O). If it is occupied by someone else in your tier then it is labeled as (OS) and colored RED if the tenant is ranked higher than you in the priority queue (we will explain in a minute what this is) and DARK GREEN if the tenant is ranked lower than you in the priority queue. 114 If the house is occupied by someone in a tier lower than yours then it is indicated by the color ORANGE and labeled (L). If the house is vacant, that is, not occupied by anyone, it is indicated by the color WHITE and labeled (V). House Allocation is as follows: Tier 1: The house allocation process starts with tier 1 while the other tiers wait. Within tier1 the house allocation takes place in the following way. All participants in the tier are lined in a pre-determined priority queue. Your position in the queue is indicated on your screen. [SLIDE #2] Note that your position in the queue does not depend on any of your or anyone else's past decisions. Existing tenants rst, simultaneously, choose between taking part in the allo- cation process by choosing IN and not taking part by choosing OUT. [SLIDE #2] { If you are an existing tenant and choose OUT, you will keep your current house and the allocation process is over for you. { If you are an existing tenant and choose IN, you will then need to rank the available houses. [SLIDE #3] If you are a newcomer you cannot choose OUT. You need to rank the available houses. 115 Note that the participating existing tenants and the newcomers simultaneously submit their rankings and they need to rank all the available houses. No two houses should be given the same rank. Once the participating existing tenants and the newcomers have submitted their ranking of the available houses, the house allocation takes place in the following way: We proceed from the top of the priority queue. Based on her chosen ranking of the houses, for the rst participant in the queue, we look at the status of her top ranked house from among the remaining houses. If the Status of the house is L or V,i.e., if the house is occupied by someone in a lower tier or not occupied by anyone, then the participant at the top of the queue is assigned to it. Note that during the allocation process houses occupied by lower tiers are treated in the same manner as those that are vacant. Note that an existing tenant vacates her current house, once she is assigned another house. If the requested house is not V or L, it means the requested house is the current house of an existing tenant in the tier. In this case, the existing tenant is moved to the top of the priority queue, directly in front of the requester. This way the existing tenant is always guaranteed a house which is at least as good as the house she is living in, based on her chosen ranking of the houses. The process continues afterwards with the modied queue. If a cycle of requests are formed (e.g., I want John's house, John wants your house and you want my house), all members of the cycle are given what they want, and their new houses are removed from the system. 116 The process continues until all participants in tier1 are assigned a house. Tier 2: All participants are lined in a pre-determined priority queue. Your position in the queue is indicated on your screen. Note that your position in the queue does not depend on any of your or anyone else's past decisions. The houses assigned to members of tier 1 are not available (NA). In tier 2, if an existing tenant's house has not been allocated to someone in tier 1, then the agent continues to occupy her house. If however, the agent's house has been taken by a member of a higher tier, she is compensated with a randomly selected house which was either previously occupied by a member of tier 1 or not occupied by anyone (V). Existing tenants rst, simultaneously, choose between taking part in the allo- cation process by choosing IN and not taking part by choosing OUT. { If you are an existing tenant and choose OUT, you will keep your current house and the allocation process is over for you. { If you are an existing tenant and choose IN, you will then need to rank the available houses. If you are a newcomer you cannot choose OUT. You need to rank the available houses. The rest of the steps are exactly as those for tier 1. When the allocation for TIER 2 is over, we move on to TIER 3. The steps for TIER 3 are exactly the same as those for TIER 2. When the allocation for TIER 3 is over, we move on to TIER 4, where once again, the steps are exactly the same. 117 EXAMPLE: We will now go through a simple example to illustrate how the allocation method works. [Slide #4] Suppose that there are six participants in two tiers. Participants 1, 2 and 3 belong to tier 1 while 4, 5 and 6 belong to tier 2. In tier 1, participants 1 and 2 are existing tenants occupying houses W and X respectively, while 3 is a newcomer. In tier 2, participants 4 and 5 are existing tenants occupying houses Y and Z respectively, while 6 is a newcomer. In addition, houses U and V are vacant. We start with tier 1. Suppose the pre-determined priority queue is: 3-1-2. Suppose player 1 chooses OUT and player 2 chooses IN. Player 1 is allocated her current house W and the allocation process is over for her. Then suppose that players 2 who chose IN and player 3 who is automatically in because she is a newcomer, given their payos, enter the following ranking of houses: [SLIDE #5] Participant 2 Participant 3 Rank 1 V Z Rank 2 Y U Rank 3 U Y Rank 4 X V Rank 5 Z X Then the allocation for Tier 1 takes place in the following manner: Top choice among remaining Actions taken at the Priority Queue Remaining Houses houses and its status the end of the step Step 1 3 2 X;Y;Z;U;V Z; status-L 3 gets Z Step 2 2 X;Y;U;V V; status-V 2 gets V Step1: We start with participant 3. His top choice is house Z which has status L. Participant 3 is assigned house Z. 118 Step 2: Only participant 2 remains. His top choice among the remaining houses is V which is vacant. So participant 2 is assigned house V. [SLIDE #6] Now we move on to tier 2. Suppose the priority queue is 6-5-4. Since player 5 has lost her house to tier 1 she is compensated by a house randomly chosen from the set of houses that were previously occupied by tier 1 or vacant. Suppose she is compensated by house X. Now suppose the players 4, 5 choose IN and then players 4, 5 and 6 submit the following ranking of houses. Participant 4 Participant 5 Participant 6 Rank 1 X Y X Rank 2 Y U Y Rank 3 U X U Then the allocation for Tier 2 takes place in the following manner: Top choice among remaining Actions taken at the Priority Queue Remaining Houses houses and its status the end of the step Step 1 6 5 4 X;Y;U X; occupied by 5 6-5-4 becomes 5-6-4 Step 2 5 6 4 X;Y;U Y; occupied by 4 5-6-4 becomes 4-5-6 Step 3 4 5 6 X;Y;U X; occupied by 5 4 gets X, 5 gets Y Step 1: The priority queue is 6-5-4. Participant 6 has ranked house X as his top choice which is currently occupied by participant 5. Participant 5 is moved to the top of the queue. Step 2: The modied priority queue is 5-6-4. Participant 5 has ranked house Y as his top choice which is currently occupied by participant 4. Participant 4 is moved to the top of the queue. Step 3: The modied priority queue is 4-5-6. Participant 4 has ranked house X as his top choice which is currently occupied by participant 5. Now a cycle is created 119 where Participant 4 wants the house of participant 5 and participant 5 wants the house of participant 4. So participant 4 is given house X and participant 5 is given house Y. Step 4: Now only participant 6 is left. He gets house U. The following slide summarizes the rules of the experiment: [Read summary slides #7 and #8] *** PRACTICE SESSION *** We will now begin the Practice session and go through a practice match to fa- miliarize you with the computer interface and the procedures. During the practice match, please do not hit any keys until you are asked to do so, and when you enter information, please do exactly as asked. Remember, you are not paid for this prac- tice match. At the end of the practice match you will have to answer some review questions. Are there any questions before we begin? [AUTHENTICATE CLIENTS] You have just received your rst match. Notice your Role and tier. The existing tenant occupying house A in Tier 1 will see a screen like this [SLIDE #9]. Notice that you are asked to choose between taking part in the allocation process by choosing IN and not taking part by choosing OUT. In this the Existing tenant is occupying house A which is colored LIGHT GREEN. House B is colored DARK GREEN as it is occupied by someone else in the tier but with a lower position in the priority queue. The houses occupied by participants in tiers 2, 3 and 4 are colored Orange while houses C,F,K and L are colored white as they are not occupied by anyone. Newcomer in Tier 1 will see a screen like this [SLIDE #10]. Notice that you are asked to wait for your turn while the existing tenants in the tier choose between 120 taking part in the allocation process or not. Notice that houses A and B are colored RED as they are occupied by participants with higher position in the priority queue. The houses occupied by participants in lower tiers are colored Orange those that are not occupied by anyone are colored white. Those not in tier 1 will see a screen like this [SLIDE #11]. Notice that you are asked to wait for your turn. If you are an existing tenant this is indicated on your screen and the house you occupy is colored LIGHT GREEN. Existing Tenants in Tier 1 please click "OUT". Notice that you have been allocated the house that you were occupying. For example, the Existing tenant who was occupying house A will see a screen like this [SLIDE #12]. Newcomer in Tier 1 will see a screen like this [SLIDE #13]. Please give rank 1 to house C, rank 2 to house D, rank 3 to house E... and so on till rank 10 to house L. The Newcomer is allocated house C. Now we have moved to tier 2 The existing tenant occupying house E in Tier 2 will see a screen like this [SLIDE #14]. Notice that now houses A, B and C have become Not Available as they have already been assigned to someone in tier 1. In this the Existing tenant is occupying house E which is colored LIGHT GREEN. House D is colored RED as it is occupied by someone else in the tier but with a higher position in the priority queue. The houses occupied by participants in tiers 3 and 4 are colored Orange while houses not occupied by anyone are colored white. Notice that all participants in tiers 3-4 are asked to wait while we nish the allocation of tier 2. [SLIDE #15] You will also see that houses A, B and C 121 have become Not Available as they have already been assigned to someone in a higher tier. If you are an existing tenant this is indicated on your screen and the house you occupy is colored LIGHT GREEN. Existing Tenants in Tier 2 please click IN. [SLIDE #16] Now Existing Tenants and Newcomer in Tier 2 please give rank 1 to house D, rank 2 to house E, rank 3 to house F,... and so on till rank 9 to house L. The participant rst in the priority queue is allocated house D, the participant second in the priority queue is allocate house E and the participant third in the priority queue is allocated house F. Now we have moved to tier 3 The existing tenant occupying house G in Tier 3 will see a screen like this [SLIDE #17]. Notice that now houses D, E and F have also become Not Available as they have already been assigned to someone in tier 2. As the Existing tenant is occupying house G it is colored LIGHT GREEN. House H is colored DARK GREEN as it is occupied by someone else in the tier 3 but with a lower position in the priority queue. The houses occupied by participants in tier 4 are colored Orange while houses not occupied by anyone are colored white. Notice that all participants in tiers 4 are asked to wait while we nish the allocation of tier 3. You will also see that houses D, E and F have become Not Available as they have already been assigned to someone in a higher tier. If you are an existing tenant this is indicated on your screen and the house you occupy is colored LIGHT GREEN. Existing Tenants in Tier 3 please click OUT. Notice that you have been allocated 122 the house that you were occupying. For example, the Existing tenant was occupying house G will see a screen like this [SLIDE #18]. Newcomer in Tier 3 will see a screen like this [SLIDE #19]. Newcomer in Tier 3 please give rank 1 to house I, rank 2 to house J, rank 3 to house K and rank 4 to house L. The Newcomer is allocated house I. Now we have moved to tier 4 All the participants in tier 4 will see that houses G,H and I have also become Not Available. The existing tenant who was earlier occupying house I has lost his house but has now been compensated by another house, chosen randomly from the set of vacant houses. For example, if he is compensated with house L then he will see a screen like this [SLIDE #20]. Existing Tenants in Tier 4 please click IN. [SLIDE #21] Now Existing Tenants and Newcomer in Tier 4 please give rank 1 to house J, rank 2 to house K, rank 3 to house L. The participant rst in the priority queue is allocated house J, the participant second in the priority queue is allocate house K and the participant third in the priority queue is allocated house L. *** END OF PRACTICE SESSION *** The practice match is over. Please complete the quiz. It has 6 questions. If there are any problems or questions from this point on, raise your hand and an experimenter will come and assist you. 123 [START QUIZ] [WAIT for everyone to nish the Quiz] Are there any questions before we begin with the paid session? We will now begin with the 6 paid matches. If there are any problems or questions from this point on, raise your hand and an experimenter will come and assist you. [START MATCH 1] [After MATCH 6 read:] This was the last match of the experiment. Your payo is displayed on your screen. Please record this payo in your record sheet. [CLICK ON WRITE OUTPUT] Your Total Payo is this amount plus the show-up fee of $5. We will pay each of you in private in the next room in the order of your Subject ID number. Remember you are under no obligation to reveal your earnings to the other participants. 124 Chapter 4 Conclusion These essays presented empirical evidence on two issues in political economy and mechanism design. Chapter 2 presents results from an investigation into the eco- nomic eects of a counterinsurgency policy implemented in India. Chapter 3 presents results from an experimental examination of the tiered housing allocation with ex- isting tenants problem. This nal chapter now concludes with a brief review of the ndings, their implications and some suggestions for future research. 4.1 Economic benets of counterinsurgency The rst essay (Chapter 2) estimates the eects of counterinsurgency policies on economic activity, in the context of the Naxalite insurgency in the Indian state of Andhra Pradesh over the period 1970-2000. In addition to the destruction of infras- tructure, the environment of political uncertainty created during a con ict further dampens economic activity by reducing investments. This research nds evidence that a strong signal from the government that it is willing to invest in tackling the insurgency could mitigate this eect of the insurgency. We nd that the introduction of the Greyhounds results in large signicant gains in pcNSDP for Andhra Pradesh. Additionally, we nd that these gains came through the non-agricultural sector (industry, manufacturing, services) rather than the agri- cultural sector. In particular, the output and investments in the manufacturing sector, 125 responded strongly to the counterinsurgency policy, suggesting a dynamic relationship between insurgency, counterinsurgency and economic outcomes. The eld of the microeconomic eects of counterinsurgency policies remains largely underexplored. As more data becomes available, future line of work could measure the eects of Naxalite insurgency and the state's counterinsurgency response at a more disaggregated level. The linkages between con ict, security and economic outcomes is only starting to be understood and though, this research is based on an example from India, we believe it is relevant to a larger audience of policy makers and academics. 4.2 Tiered housing allocation: an experimental analysis The second essay (Chapter 3) examines the tiered housing allocation with existing tenants problem in a controlled laboratory environment. The performance of the modied versions of three mechanisms {Top Trading Cycles, Gale-Shapley and the Random Serial Dictatorship{ is evaluated in terms of eciency, fairness, participation and truthful revelation. We make three important additions to the existing literature: (i) we adapt and study three leading mechanisms in this tiered framework, (ii) we inform the agents of their position in the queue before they make any decisions and (iii) we play multiple matches. We nd that contrary to some of the existing literature, all three mechanisms exhibit similar behavior in terms of low rates of participation (around 40%), high rates of truth-telling conditional on participation (around 90%), high proportions of fair allocations (above 90%) and signicant eciency losses. Overall, the Random Serial Dictatorship has an eciency advantage due to higher truth-telling rates. In terms of policy implications for the future design of matching mechanisms, the most important nding is that announcing the priority ordering before eliciting 126 choices has a signicant eect on behavior, even in mechanisms where participa- tion and truthful revelation are dominant strategies. While the priority ranking is pre-announced in some real life mechanisms and not in others, the behavioral im- plications of doing so have not been studied so far. This study indicates that the rank-dependency of choices has important policy implications that deserve to be ex- plored in further detail. Further, individual level analysis reveals that extremely few subjects realize the dominance of participation. However, as shown in table 3.16, the majority of indi- viduals who do not follow the theoretical predictions still follow some \reasonable" strategies. Put together, these results suggest an investigation into the feedback re- quired by individuals in order to understand the properties of these mechanisms as a future avenue of research. 127 Comprehensive bibliography [1] Abadie, A. and Gardeazabal, J., (2003), \The Economic Costs of Con ict: A case Study of the Basque Country", American Economic Review, 93(1), 113-132. [2] Abadie, A., Diamond, A., and Hainmueller, J., (2010), \Synthetic Control Meth- ods for Comparative Case Studies: Estimating the Eect of California's Tobacco Control Program", Journal of the American Statistical Association, 105(490), 493-505. [3] Abdulkadiro glu, A. and T. S onmez (1999) \House Allocation with Existing Ten- ants." Journal of Economic Theory, 88, 233-260. [4] Abdulkadiro glu, A. and T. S onmez (2003) \School Choice: A Mechanism Design Approach." American Economic Review, 93, 729-747. [5] Achuthan, J. K., (2010), \Tackling Maoists: The Andhra Paradigm", Indian Defence Review, 25(2). [6] Aghion, P., Burgess R., Redding, S. J., and Zilibotti, F., (2008), \The Unequal Eects of Liberalization: Evidence from Dismantling the License Raj in India", American Economic Review, 98(4), 1397-1412. [7] Baccara, M., Imrohoroglu, A., Wilson, A., and L. Yariv (2012) \A Field Study on Matching with Network Externalities." American Economic Review, 102(5), 1773-1804. [8] Balagopal, K., (1988), \Herald the Hunting Dogs that are Grey in Colour", Economic and Political Weekly, Vol. 23, No. 28, 1413-1414. 128 [9] Beath, A., Christia, F., and Enikolopov, R., (2011), \Winning Hearts and Minds through Development: Evidence from a Field Experiment in Afghanistan", MIT Political Science Department Research Paper No. 2011-14. [10] Becker, G. S., (1968), \Crime and Punishment: An Economic Approach", Jour- nal of Political Economy, 76(2), 169-217. [11] Berman, E., Callen, M., Shapiro, J. N., and Felter, J. H., (2011), \Do Work- ing Men Rebel? Insurgency and Unemployment in Afghanistan, Iraq, and the Philippines", Journal of Con ict Resolution, 55(4), 496-528. [12] Berman, E., Shapiro, J. N., and Felter, J. H., (2011), \Can Hearts and Minds Be Bought? The Economics of Counterinsurgency in Iraq", Journal of Political Economy, 119(4), 766-819. [13] Besely, T. and Burgess, R. (2004), \Can Labor Regulation Hinder Economic Performance? Evidence from India", Quarterly Journal of Economics, 19(1), 91-134. [14] Borooah, V. K. (2008), \Deprivation, Violence, and Con ict: An Analysis of Naxalite Activity in the Districts of India", International Journal of Con ict and Violence, 2(2), 317-333. [15] Calsamiglia, C., Haeringer, G. and F. Klijn (2010) \Constrained School Choice: An Experimental Study." American Economic Review, 100(4), 1860-74. [16] Chadha, V. (2005). Low intensity con icts in India: an analysis. New Delhi: Sage Publications. [17] Chaney, E. (2008), \Assessing Pacication Policy in Iraq: Evidence from Iraqi Financial Markets", Journal of Comparative Economics, 36(1), 1-16. 129 [18] Chen, Y. and T. S onmez (2002) \Improving Eciency of On-Campus Housing: An Experimental Study." American Economic Review, 92, 1669-1686. [19] Chen, Y. and T. S onmez (2004) \An Experimental Study of House Allocation Mechanisms." Economics Letters, 83, 137-140. [20] Chen, Y. and T. S onmez (2006) \School Choice: An Experimental Study." Jour- nal of Economic Theory, 127, 202-231. [21] Crost, B., Felter, J., and Johnston, P. B., (2012), \Aid Under Fire: Development Projects and Civil Conct", Working Paper. [22] Eynde, O. V., (2011), \Targets of Violence: Evidence from India's Naxalite Con ict", Working Paper. [23] Featherstone, C. and M. Niederle (2011) \School Choice Mechanisms under In- complete Information: An Experimental Investigation." Working Paper. [24] Gale, D. and L. Shapley (1962) \College Admissions and the Stability of Mar- riage." American Mathematical Monthly, 69, 9-15. [25] Ganguly, S., and Fidler, D. P. (eds.) (2009). India and counterinsurgency: lessons learned. London: Routledge. [26] Gawande, K., Kapur, D. and Satyanath, S. (2012), \Natural Resource Shocks and Con ict in India's Red Belt", Working Paper. [27] Greenstone, M., (2007), \Is the \Surge" Working? Some New Facts", NBER Working Paper 13458. [28] Government of India (2005), Ministry of Home Aairs Annual Report 2004-05. 130 [29] Government of India (2006), Ministry of Home Aairs Annual Report 2005-06. [30] Guillen, P. and O. Kesten (2012) \Matching Markets with Mixed Ownership: The Case for A Real-life Assignment Mechanism." International Economic Re- view, 53(3), 1027-1046. [31] Gupta, D. K. (2007), \The Naxalites and the Maoist Movement in India: Birth, Demise, and Reincarnation", Democracy and Security, 3, 157-188. [32] Haruvy, E., Roth, A. and U. Unver (2006) \The dynamics of law clerk matching: An experimental and computational investigation of proposals for reform of the market." Journal of Economic Dynamics and Control, 30(3), 457-486. [33] Hindustan Times (2009), \Rajdhani held hostage in Naxal land; secured after ve hours", retrieved from http://www.hindustantimes.com/Rajdhani-held-hostage- in-Naxal-land-secured-after-ve-hours/Article1-469941.aspx as on June 14, 2011. [34] Hinrichs, P. (2012), \The Eects of Armative Action Bans on College Enroll- ment, Educational Attainment, and the Demographic Composition of Universi- ties", Review of Economics and Statistics, 94(3), 712-722. [35] Hoelscher, K., Miklian, J., and Vadlamannati, K.C., (2012), \Hearts and Mines: A District-Level Analysis of the Maoist Con ict in India", International Area Studies Review, 15(2), 141-160. [36] Imbens, G. W., and Wooldridge, J. M., (2009), \Recent Developments in the Econometrics of Program Evaluation", Journal of Economic Literature, 47(1), 5-86. [37] Iyengar, R., Monten, J., and Hanson, M., (2011), \Building Peace: The Impact 131 of Aid on the Labor Market for Insurgents", NBER Working Paper Series No. 17297. [38] Joshi, S. (2010), \How Naxal Violence hits India Inc. Hard", IBN Live News Re- port available at http://ibnlive.in.com/news/how-naxal-violence-hits-india-inc- hard/112950-7.html?from=tn as on June 14, 2011. [39] Kagel, J. and A. Roth (2000) \The Dynamics of Reorganization in Matching Markets: A Laboratory Experiment Motivated by A Natural Experiments." Quarterly Journal of Economics, 115(1), 201-235. [40] Kapstein, E. B., (2012), \Measuring Progress in Modern Warfare", Survival: Global Politics and Strategy, 54(1), 137-158. [41] Kocher, M. A., Pepinsky, T. and Kalyvas, S., (2011), \Aerial Bombing and Counterinsurgency in Vietnam War", American Journal of Political Science, 55(2), 201-218. [42] Kujur, R. (2008) \The Naxal Movement in India: A Prole", Institute of Peace and Con ict Studies Research Papers. [43] Lee, W. (2011) \Comparative case studies of the eects of in ation targeting in emerging economies", Oxford Economic Papers, 63(2), 375-397. [44] Lusome, R., and Bhagat, R. B., (2006), \Trends and Patterns of Internal Migra- tion in India, 1971-2001", Working Paper. [45] Lyall, J., (2009), \Does Indiscriminate Violence Incite Insurgent Attacks? Evi- dence from Chechnya", Journal of Con ict Resolution, 53(3), 331-362. 132 [46] Mukherjee, A. (2010) \India's experience with insurgency and counterinsur- gency", in S. Ganguly, A. Scobell and J. C. Liow (eds.), The Routledge Handbook of Asian Security Studies, London: Routledge, 139-157. [47] Montalvo, J. G., (2011) \Voting after the Bombings: A Natural Experiment on the Eect of Terrorist Attacks on Democratic Elections", Review of Economics and Statistics, 93(4), 1146-1154. [48] Nalbantian, H.R. and A. Schotter (1995) \Matching and Eciency in the Base- ball Free-Agent System: An Experimental Examination." Journal of Labor Eco- nomics, 13(1), 1-31. [49] Naxal Terror Watch (2010), blog available at http://naxalwatch.blogspot.com/ as on June 14, 2011. [50] Oetken, J. L. (2009), \Counterinsurgency against the Naxalites", in S. Ganguly and D. P. Fidler (eds.), India and counterinsurgency: lessons learned, London: Routledge, 127-151. [51] Olson, M. and D. Porter (1994) \An experimental examination into the design of decentralized methods to solve the assignment problem with and without money." Economic Theory, 4(1), 11-40. [52] Pais, J. and A. Pint er (2008) \School choice and information: An experimental study on matching mechanisms." Games and Economic Behavior, 64, 303-328. [53] Pais, J., Pint er, A. and R. F. Veszteg (2011) \College admissions and the role of information: an experimental study."International Economic Review, 52(3), 713-737. 133 [54] P apai, S. (2000) \Strategyproof Assignment by Hierarchical Exchange." Econo- metrica, 68(6), 1403-1433. [55] Pinotti, P. (2012), \The Economic Consequences of Organized Crime: Evidence from Southern Italy", Working Paper. [56] Planning Commission, Government of India, \Tenth Five-Year Plan: 2002-07, Vol. III, State Plans, Trends, Concerns and Strategies". [57] Priyadershi, V. (2009), \Fighting Left-Wing-Extremism through Special Task Forces", Centre for Land Warfare Studies, Article 1362. [58] Pycia, M. and U. Unver (2007) \A Theory of House Allocation and Exchange Mechanisms." Working paper. [59] Rajagopalan, R. (2008), Fighting like a Guerrilla: The Indian Army and Coun- terinsurgency. New Delhi: Routledge. [60] Raju, R. V. (2010), \Countering the Naxal Threat - I: An Analysis of Earlier Eorts", Institute of Peace and Con ict Studies, Article 3149. [61] Ramana, P.V. (2009), \A Critical Evaluation of the Union Government's Re- sponse to the Maoist Challenge", Strategic Analysis, 33(5), 745-759. [62] Roth, A.E., S onmez, T. and U. Unver (2004) \Kidney Exchange." Quarterly Journal of Economics, 119(2), 457-488. [63] Roth, A.E. and M. Sotomayor (1990) Two-Sided Matching: A Study on Game - Theoretic Modelling. Cambridge University Press. 134 [64] Sahni, A., (2007), \Andhra Pradesh: The State Advances, the Maoists Retreat", South Asia Intelligence Review Weekly Assessments & Briengs, 6(10), available at http://www.satp.org/satporgtp/sair/Archives/6 10.htm . [65] Sahni, A., (2008), \Fighting the Maoists with Mantras", South Asia Intelligence Review Weekly Assessments & Briengs, 7(2), available at http://www.satp.org/satporgtp/sair/Archives/7 2.htm . [66] Shapley, L. and H. Scarf (1974) \On Cores and Indivisibility." Journal of Math- ematical Economics, 1, 23-28. [67] Shatrugna, M. (1989), \NTR and the Naxalites", Economic and Political Weekly, Vol.24, No.28, p. 1570. [68] Singh, A. K. and Diwan, S. B. (2010), \Red Money", South Asia In- telligence Review Weekly Assessments & Briengs, 8(39), available at http://www.satp.org/satporgtp/sair/Archives/sair8/8 39.htm . [69] Solberg, K. E. (2008), \Health Crisis Amid the Maoist Insurgency in India", The Lancet, 371(9621), 1323-1324. [70] S onmez, T. and U. Unver (2011) \Matching, Allocation and Exchange of Discrete Resources", in J. Benhabib, A. Bisin, and M. Jackson (eds.) Handbook of Social Economics Vol. 1A, Elsevier. [71] Swami, P. (2010, April 13), \For a review of counter- insurgency doctrine", The Hindu. Retrieved May 7, 2012, from http://beta.thehindu.com/opinion/lead/article395529.ece. [72] Tata, M. (2010, April 19), \Lessons from Andhra", Outlook India. Retrieved May 7, 2012, from http://www.outlookindia.com/article.aspx?265014. 135 [73] Title, L.R. (1998) \Platonic Mechanism Design." Working Paper. [74] Tiwari, A. (2009, February 7), \Naxal-hit states to form Greyhound- like force", The Times of India. Retrieved May 7, 2012, from http://articles.timesondia.indiatimes.com/2009-02-07/india. [75] Unver, U. (2005) \On the survival of some unstable two-sided matching mecha- nisms." International Journal of Game Theory, 33(2), 239-254. [76] U.S. Army (2007), The U.S. Army/Marine Corps Counterinsurgency Field Man- ual (Field Manual No. 3-24), University of Chicago Press. [77] Zussman, A. and Zussman, N. (2006), \Assassinations: Evaluating the Eec- tiveness of an Israeli Counterterrorism Policy Using Stock Market Data", The Journal of Economic Perspectives, 20(2), 193-206. 136

Abstract (if available)

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Essays on political economy of privatization

PDF

Three essays on cooperation, social interactions, and religion

PDF

Three essays on economics of early life health in developing countries

PDF

Essays on education: from Indonesia to Los Angeles, CA

PDF

The political economy of R&D collaboration: micro- and macro-level implications

PDF

Essays on development and health economics

PDF

Essays on development and health economics: social media and education policy

PDF

Essays in development and experimental economics

PDF

Essays on political economy and corruption

PDF

Essays on effects of reputation and judicial transparency

PDF

Essays on education and institutions in developing countries

PDF

Essays in political economics

PDF

Essays on health economics

PDF

Three essays on health & aging

PDF

Comparative study on Asian financial capitals' competitiveness: focused on strengths and weaknesss of the city of Seoul

PDF

Three essays on the political economy of federal spending

PDF

Perspectives on state capacity and the political geography of conflict

PDF

Essays on economics of education and private tutoring

PDF

Political economy of private commercial arbitration in Colombia

PDF

Essays on education programs in Costa Rica

Asset Metadata

Creator Singhal, Saurabh (author)

Core Title Essays in political economy and mechanism design

School College of Letters, Arts and Sciences

Degree Doctor of Philosophy

Degree Program Economics

Publication Date 06/25/2013

Defense Date 05/07/2013

Publisher University of Southern California (original), University of Southern California. Libraries (digital)

Tag conflict,experimental economics,matching,OAI-PMH Harvest,political economy

Format application/pdf (imt)

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Carrillo, Juan D. (committee chair), Strauss, John A. (committee chair), Banerjee, Tridib K. (committee member), Nugent, Jeffrey B. (committee member)

Creator Email singhal.saurabhs@gmail.com,singhals@usc.edu

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c3-280064

Unique identifier UC11293728

Identifier etd-SinghalSau-1706.pdf (filename),usctheses-c3-280064 (legacy record id)

Legacy Identifier etd-SinghalSau-1706.pdf

Dmrecord 280064

Document Type Dissertation

Format application/pdf (imt)

Rights Singhal, Saurabh

Type texts

Source University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Access Conditions The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the a...

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA