| Original Full Text | UC DavisUC Davis Electronic Theses and DissertationsTitleThe Economics of University Peer Effects and Employee Training in Microenterprises in UgandaPermalinkhttps://escholarship.org/uc/item/3vm4k7p4AuthorAhimbisibwe, IsaacPublication Date2024 Peer reviewed|Thesis/dissertationeScholarship.org Powered by the California Digital LibraryUniversity of CaliforniaThe Economics of University Peer Effects and Employee Training in Microenterprises in UgandaByISAAC AHIMBISIBWEDISSERTATIONSubmitted in partial satisfaction of the requirements for the degree ofDOCTOR OF PHILOSOPHYinAgricultural and Resource Economicsin theOFFICE OF GRADUATE STUDIESof theUNIVERSITY OF CALIFORNIADAVISApproved:Travis Lybbert, ChairScott CarrellStephen BoucherCommittee in Charge2024iAbstractHuman capital development is crucial for economic growth. Policymakers must understand thefactors influencing human capital accumulation since they may vary by setting and sector. Forexample, micro-enterprise employees enhance their skills through on-the-job training or vocationalprograms. In educational institutions, family and school investments and peer interactions canaffect human capital growth. Studies show that characteristics of peers, such as the gender of aclassmate, can impact academic performance in elementary school. Other research has highlightedthe importance of homophily in social network formation. Other research has highlighted the im-portance of homophily in social network formation. The latter is especially important in highereducation settings, where individuals are likely to engage in assortative matching based on charac-teristics such as shared identity. Thus, the ethnicity of peers may also play a crucial role in humancapital development in settings of ethnic diversity.The first essay in this dissertation quantifies high-ability and coethnic peer effects in highereducation located in an ethnically diverse setting. While empirical research has documented the neg-ative impact ethnic diversity has on several political and economic outcomes in Sub-Saharan Africa,including economic growth, political engagement, conflict, and contributions to public goods, weknow relatively little about educational peer effects in such settings, which are generally character-ized by high ethnic diversity and cross-ethnic mixing. This chapter studies the effect of coethnic andhigh-ability peers in student groups on academic outcomes at a large public university in Uganda, acountry with pronounced ethnic heterogeneity and segregation. I link data on student-level univer-sity admissions with subsequent grades. Upon admission, dorm assignments are random conditionalon gender, providing exogenous variation in peer group formation. On average, high-ability peers(irrespective of ethnicity) and coethnic peers (irrespective of ability) positively affect a student’s per-formance. Whereas the coethnic peer effect disappears by the year of graduation, the high-abilitypeer effect persists and even increases in magnitude over time. Lastly, I find that the effect ofhigh-ability coethnic peers on performance is statistically indistinguishable from that of high-abilitynoncoethnic peers.The second essay uses a causal forest algorithm to analyze heterogeneity in coethnic peereffects by estimating a grade dose-response function and treatment effects resulting from interethniciirelations. Specifically, I train the causal model to optimize heterogeneity in students’ characteris-tics, including ethnic groups. This model predicts each student’s conditional average treatment ofcoethnic share while doing data-driven sample splits to estimate heterogeneity. I find that coeth-nic peer effects are strongest for the largest ethnic group. This is the group that portrays moreethnic attachment than other ethnic groups in this setting. I also find the lowest effects for thesecond-largest group, which controls the central government and is thus likely to identify more withnational identity than tribal identity.The last essay uses a field experiment to examine how employers select employees for trainingand the demand for training from employees. Along with collaborators, I elicit employers’ beliefsabout which of their employees it would be socially optimal to train and their preferences overwhich employees they choose to train. I then investigate whether employers’ selection of workers isindividually rational. Finally, I measure employees’ self-selection into training and their alignmentwith employers’ selections. To ensure incentive compatibility of employer and employee choices,I provide employees from a sample of metalworking SMEs with free, high-quality skills training.Additionally, I conduct practical skills tests to measure employee metalworking skills before andafter training. My analysis shows that owners perceive that training improves the quality of atrained worker. Yet, when offered the opportunity to choose an employee for training, they do notselect workers whose quality would improve the most from training. Instead, they choose workerswith ties to the firm, as those workers are perceived to be most profitable post-training but wouldnot gain the most from training.iiiAcknowledgementsThis achievement would not have been possible without the support of several individuals, duringmy intellectual and personal growth. First, I want to thank my committee, Travis Lybbert (chair),Scott Carroll, and Stephen Boucher. They have shown belief in me as a researcher and fosteredan environment that has led to tremendous growth during my time at UC Davis. Additionally,Travis and Steve have provided a ‘community’ outside of school, helping me miss my family a lotless while at Davis, especially during the holidays. I have learned an incredible amount from eachof my committee members, both professionally and personally. I am also grateful for their supportduring the job market process.Outside my committee, I have obtained support from several Economics and ARE professorsduring my job market and thesis journey at UC Davis, such as Arman Razae, Dalia Ghanem, andRachael Goodhue, as well as from development group folks, especially Ashish Shenoy. I also wantto acknowledge my co-authors, Andy Brownback, Arman Razae, and Sarojini Hirshleifer. I havelearned an incredible amount from them, especially how to apply economic theory in developmenteconomics and how to implement and manage field experiments.My time at UC Davis would not have been possible without a supportive student community.I especially want to thank Scott Sommerville for his friendship; Saloni Chopra for her researchcollaboration; Jeffrey Hadacheck and James Keeler for their friendship and constant help withcoding in R and editing my earlier manuscripts. I did not have any R coding experience at thestart of my PhD, but James and Jeff helped me hone my skills. I also want to thank Kavin Dihnfor helping me prepare for my micro preliminary exam and Alex Machinda for collaborating andhelping me with data scraping.Lastly, I want to thank my family who have consistently reminded me of the meaning of lifeand helped me feel loved. First, I want to thank my mom and dad, who, despite only completingGrade Four and barely speaking English, appreciated education’s role in breaking the cycle ofpoverty that has plagued my ancestral home. They instilled in me a strong foundation growing upin southwestern Uganda, where less than one percent attain post-secondary education. My oldersiblings, Sarah and Henry, took care of me and my younger siblings while my parents were workingand taught us the value of hard work. My sister was the first to come to the US for graduate school,ivinspiring me that obtaining a PhD from a first-class university was possible, and encouraging meto apply. To my younger brother Peter, for his constant friendship and generous heart, and to mybrother-in-law Joe, for his financial support when I first came to the US for graduate studies, andto my other siblings also, thank you.vContents1 Introduction to the Essays 12 Peer Effects and Ethnicity in Uganda 62.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122.2.1 Ethnicity in Uganda . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122.2.2 Ugandan Higher Education and Makerere University Kampala . . . . . . . . . 142.3 Empirical Setting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152.3.1 Applications, Admissions, and Sample Definition . . . . . . . . . . . . . . . . 152.3.2 Dorm Assignment and Defining Peer Groups . . . . . . . . . . . . . . . . . . 172.3.3 Identifying Peer Effects at MUK . . . . . . . . . . . . . . . . . . . . . . . . . 182.4 The Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202.4.1 Academic and Demographic Characteristics . . . . . . . . . . . . . . . . . . . 202.4.2 Ethnicity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222.4.3 Descriptive Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252.5 Empirical Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262.5.1 Mean Effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262.5.2 Heterogeneous Peer Effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . 322.6 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 362.6.1 Mean Effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 362.6.2 Persistence of Mean Effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . 392.6.3 Heterogeneous Peer Effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40vi2.6.4 Robustness checks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 462.6.5 Discussion and Contextualizing Results . . . . . . . . . . . . . . . . . . . . . 482.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 532.8 Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 602.8.1 Data Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 602.8.2 Ethnic and Geographic Boundaries . . . . . . . . . . . . . . . . . . . . . . . . 612.8.3 Deriving the Reduced-Form Peer Effect . . . . . . . . . . . . . . . . . . . . . 662.8.4 Additional Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 682.8.5 List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 713 Heterogeneity in Coethnic Peer Effects 723.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 723.2 Empirical estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 743.2.1 Estimating the CATE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 753.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 773.3.1 Dose Response to Coethnic Share . . . . . . . . . . . . . . . . . . . . . . . . . 773.3.2 Heterogeneity by Ethnicity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 793.3.3 Heterogeneity by Ethnicity and Gender . . . . . . . . . . . . . . . . . . . . . 803.3.4 Interethnic Effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 813.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 823.5 Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 853.5.1 Deriving the Identifying Assumption . . . . . . . . . . . . . . . . . . . . . . . 853.5.2 Additional Figures and Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . 854 Beliefs and the Demand for Employee Training 884.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 884.2 Context and Conceptual Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . 924.2.1 Metal Fabrication in Uganda . . . . . . . . . . . . . . . . . . . . . . . . . . . 924.2.2 Conceptual Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 934.3 Research Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 964.3.1 Firm Evaluation Sample . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97vii4.3.2 Sample of Potential Trainees . . . . . . . . . . . . . . . . . . . . . . . . . . . 974.3.3 Training Treatment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 984.4 Data and Outcomes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 994.4.1 Baseline Survey–Owners . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 994.4.2 Baseline Survey–Workers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1004.4.3 Practical Skills Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1014.5 Preliminary Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1014.5.1 Descriptive Statisitics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1024.5.2 Beliefs about Worker Quality . . . . . . . . . . . . . . . . . . . . . . . . . . . 1034.5.3 Do Owners Select the Most Teachable Worker for Training? . . . . . . . . . . 1044.5.4 Do Owners Select the Most Profitable Worker for Training? . . . . . . . . . . 1044.5.5 Employee Firm Ties and Selection for Training . . . . . . . . . . . . . . . . . 1054.5.6 Worker Demand . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1074.5.7 Worker Demand Vs Owner Selection . . . . . . . . . . . . . . . . . . . . . . . 1074.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1084.7 List of Figures and Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1124.7.1 Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1124.7.2 Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1174.8 Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1214.8.1 Time Line and Sample Construction . . . . . . . . . . . . . . . . . . . . . . . 121viiiList of Tables2.1 Descriptive Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262.2 Evidence against Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 302.3 Mean Effects in Year One: Coethnic vs High-ability Share . . . . . . . . . . . . . . . 362.4 The Effect of High-ability Coethnic and High-ability Noncoethnic peers on AcademicPerformance in Year One. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 382.5 Persistence of Mean Effects in Follow-up Years . . . . . . . . . . . . . . . . . . . . . 412.6 Differential Effect by Gender: Coethnic vs High-ability Share . . . . . . . . . . . . . 422.7 Differential Effect by Ability Type: Coethnic vs High-ability Share . . . . . . . . . . 442.8 Differential Effect by Degree Type: Coethnic vs High-ability Share . . . . . . . . . . 452.9 Differential Effect by Ethnic Salience: Coethnic vs High-ability Share . . . . . . . . . 472.10 Coethnic vs High-ability Share: Outcome as GPA . . . . . . . . . . . . . . . . . . . . 49A1 Ethnic fractionalization in a district . . . . . . . . . . . . . . . . . . . . . . . . . . . 63A2 Ethnicity/Language Group Composition . . . . . . . . . . . . . . . . . . . . . . . . . 65A3 Differential effect by Ethnic Salience (Nonmajority) Coethnic vs High-ability Share . 69A4 More Evidence against Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70B1 Gender Differences CATE by Ethnicity . . . . . . . . . . . . . . . . . . . . . . . . . . 87C1 Owner/Firm-level Descriptive Statistics . . . . . . . . . . . . . . . . . . . . . . . . . 117C2 Worker-level Descriptive Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118C3 Factors Affecting Selection for Training . . . . . . . . . . . . . . . . . . . . . . . . . . 119C4 Worker Demand for training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120C5 Project Timeline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123ixList of Figures2.1 Geographic Segregation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132.2 Distribution by ethnicity: student sample vs general population . . . . . . . . . . . . 152.3 Peer Group Construction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182.4 Distribution of Coethnic Share across Peer Groups. . . . . . . . . . . . . . . . . . . . 272.5 Comparison to past Papers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 502.6 Distribution of Peer Group Sizes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 713.1 Dose-Response Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 773.2 Difference by Ethnicity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 783.3 Differences by Ethnicity and Gender . . . . . . . . . . . . . . . . . . . . . . . . . . . 793.4 Interethnic Effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81B5 Coefficient plot by Ethnicity: CATE . . . . . . . . . . . . . . . . . . . . . . . . . . . 86B1 Owners Overestimate Worker Quality. . . . . . . . . . . . . . . . . . . . . . . . . . . 113B2 Teachability: Owner Selection vs Other Workers . . . . . . . . . . . . . . . . . . . . 114B3 Perceived Profitability: owner Selection vs Most Teachable . . . . . . . . . . . . . . . 115B4 Perceived profitability: Owner selection vs worker demand most teachable . . . . . . 116xChapter 1Introduction to the EssaysHuman capital accumulation, including education and training, is essential for economic growth.Becker (1964) highlights that investments in education, training, and health boost individual pro-ductivity and lifetime earnings. Vocational training programs, supported by various governments,significantly enhance worker skills (King and Palmer 2010). Peer effects also impact human capitaldevelopment, especially in educational settings. During early education, exogenous factors such asgender (Gong, Lu, and Song 2019) and home environment, such as domestic violence (Carrell andHoekstra 2010), can influence grades in elementary school. In higher education, students may bemore likely to sort themselves despite exogenous peer groups (Carrell, Sacerdote, and West 2013).While research often focuses on Western colleges, where race (Black or White) is a regression con-trol, high ethnic diversity and segregation in Sub-Saharan Africa may complicate and add newdimensions of peer effects.Uganda is characterized by high ethnic heterogeneity and segregation (Uganda Bureau ofStatistics 2016; Alesina and Zhuravskaya 2011). I use a setting at Makerere University KampalaUganda (MUK), which is centrally located and one of the most prestigious universities in Uganda,to study peer effects and ethnicity in chapters One and Two. By its virtue of location and prestige,MUK enrolls students from disparate ethnic regions of the country. It is, thus, at the MUK campusthat most students interact with peers of different ethnicities. Additionally, dorm assignments arerandom upon admission and within each cohort, providing exogenous assignments to peer groups.Specifically, I define a peer group as students admitted to the same school, such as the School ofMedicine, and randomly assigned to the same dorm. My identifying assumption is that, conditional1on gender, school, and cohort, a student’s peer quality measures (share of coethnic peers and shareof high-ability peers) are uncorrelated with unobservables and student characteristics. Specifically,this is the variation that I exploit. Consider two students admitted to majors in the same schooland year but randomly assigned to different dorms. One student may be exposed to x coethnicpeers in their dorm, and the other student may be exposed to y coethnic peers in another dorm asa result of randomization, where x ̸= y.My analysis in Essay One shows that both coethnic peers and high-ability peers are essentialfor academic performance when students first arrive on campus. Initially, the presence of coethnicpeers significantly boosts performance, but this effect disappears by the time students graduate.In contrast, the impact of high-ability peers persists and even increases in magnitude over time.Additionally, I find that the entire effect of peer influence is driven by students with limited priorexposure to different ethnic groups. These are students who have graduated high school in theirdistrict of origin and whose ethnic identity is likely more salient as a result of migrating to thecity for university in a different ethnic region (Okunogbe 2024). My findings might be explainedthrough peer-to-peer learning channels and psychological channels, such as the contact hypothesis inWilliam (1947). This setting makes it easy for students to learn which of their peers are high-abilityand seek study help from them if needed. Also, student-teacher interactions are limited becauseuniversities in this setting do not offer office hours, making peer-to-peer learning essential for studentsuccess. Lastly, the results showing that coethnic peer effects disappear by graduation imply thatas students progress, they navigate diverse environments, experience cross-ethnic interactions, andform cross-ethnic networks, making the psychological boost of coethnic peers less important.Essay Two builds on the first essay and studies heterogeneity related to coethnic peers.Specifically, I use causal forest estimation methods developed in Athey and Imbens (2016) andAthey and Wager (2021) to analyze the nonlinearities and potential interethnic composition effects.Causal forest uses data-driven sample splits, reducing researcher bias in selecting the relevant hetero-geneity dimensions. Additionally, the causal forest method enables the capture of high-dimensionalnonlinearities while avoiding overfitting by employing training sampling differently from an esti-mation sample. Essentially, I estimate Conditional Average Treatment Effects (CATE) for eachindividual by feeding the causal forest algorithm an estimation regression similar to the primaryestimation regression of Chapter One.2I find substantial differences in the estimated CATE by ethnicity. For example, the dis-tribution of the largest ethnic group, Baganda, stochastically dominates the distribution of otherethnicities and the average CATE of this group is larger than the average treatment effect (ATE). Incontrast, the distribution of the second largest group, Banyankore, is centered to the left of the ATE.I also find that heterogeneity by gender within each ethnicity is not unidirectional. For example,coethnic peer effects seem to matter more for women in the Basoga, the third largest ethnic group,whereas, for Baganda, coethnic peer effects seem to matter more for men. The patterns of resultsof this heterogeneity suggest that homophily (as a result of ethnic attachment), not inter-ethnicrelations, may be driving coethnic peer effects in this setting.The last essay studies human capital development in terms of employee training. Motivatedby Becker (1962) who predicts an under-provision of general skills training in perfectly competitivemarkets because owners would have to pay a worker the post-training marginal product. Otherwise,a firm faces the friction of separation by the employee. Because of such market failures, governmentsin developed countries step in to provide subsidized training. However, the consequences of trainingunder private provision might be more extreme in the developing world, as such settings often sufferfrom low productivity and imperfectly competitive markets (Hsieh and Klenow 2009).Together with collaborators, we provide a free training program to employees of small (4-14workers) metalworking firms in Uganda and study how owners select workers for training. Addi-tionally, we elicit from owners about the perceived gain from training each of their workers andoffer objective measures of quality. Since our training is free and carefully designed to limit non-monetary costs, the owners in this evaluation sample should be able to pay a worker their post-training marginal product if they anticipate any separation by the worker after training and shouldselect a worker who would improve the most from training.Although I find that owners believe our training would improve workers’ skills or humancapital through our baseline questions and incentive-compatible elicitation for training selection, ourdata show that owners do not select the workers who would improve most from training. Instead,they select a worker with the strongest ties to the firm. My analysis suggests that owners operaterationally and try to maximize the gap between the marginal product and the wage they can paytheir workers without causing that worker to leave. In doing so, they select the worker who is leastlikely to separate from the firm, often a relative, rather than the worker who would gain the most3from training. Lastly, the analysis shows that owners’ beliefs do not align with the beliefs of theworkers.4ReferencesAlesina, A., and E. Zhuravskaya. 2011. “Segregation and the Quality of Government in a CrossSection of Countries.” American Economic Review 101:1872–1911.Athey, S., and G. Imbens. 2016. “Recursive partitioning for heterogeneous causal effects.”Proceedings of the National Academy of Sciences 113:7353–7360.Athey, S., and S. Wager. 2021. “Policy Learning With Observational Data.” Econometrica 89:133–161.Becker, G.S. 1964. Human Capital: A Theoretical and Empirical Analysis, with Special Referenceto Education. Chicago: University of Chicago Press.—. 1962. “Investment in Human Capital: A Theoretical Analysis.” Journal of Political Economy70:9–49.Carrell, S.E., and M.L. Hoekstra. 2010. “Externalities in the Classroom: How Children Exposedto Domestic Violence Affect Everyone’s Kids.” American Economic Journal: Applied Economics2:211–28.Carrell, S.E., B.I. Sacerdote, and J.E. West. 2013. “From Natural Variation to Optimal Policy? TheImportance of Endogenous Peer Group Formation.” Econometrica 81:855–882.Gong, J., Y. Lu, and H. Song. 2019. “Gender Peer Effects on Students’ Academic and NoncognitiveOutcomes: Evidence and Mechanisms.” Journal of Human Resources, pp. .Hsieh, C.T., and P.J. Klenow. 2009. “Misallocation and Manufacturing TFP in China and India*.”The Quarterly Journal of Economics 124:1403–1448.King, K., and R. Palmer. 2010. Planning for Technical and Vocational Skills Development. Paris:UNESCO.Okunogbe, O. 2024. “Does Exposure to Other Ethnic Regions Promote National Integration? Evi-dence from Nigeria.” American Economic Journal: Applied Economics 16:157–92.Uganda Bureau of Statistics. 2016. “The National Population and Housing Census 2014 – MainReport.” Working paper, Kampala, Uganda.William, J., Robin M. 1947. The Reduction of Intergroup Tensions: A Survey of Research onProblems of Ethnic, Racial, and Religious Group Relations. New York: Social Science ResearchCouncil.5Chapter 2Peer Effects and Ethnicity in Uganda2.1 IntroductionUnderstanding the determinants of academic and other outcomes for students in higher educationcontinues to be a priority for university administrators and policy-makers. While significant progresshas been made in understanding the role of peer effects on academic performance (Sacerdote 2011;Foster 2006; Zimmerman 2003) and other outcomes, such as major choice De Giorgi, Pellizzari,and Redaelli (2010), and cheating (Carrell, Malmstrom, and West 2008). Most of this research hasbeen conducted in the West. Whether or not these results translate to developing countries, suchas those of Sub-Saharan Africa (SSA), is unclear. Indeed, since peer effects reflect social dynamicsthat can change dramatically across cultural contexts, it seems likely that these effects could operatequite differently in non-Western settings. One specific reason to doubt the external validity of theexisting peer effects literature on SSA is the degree and nature of ethnic diversity that characterizesmuch of the region. Such heterogeneity combined with ethno-linguistic differences may, for example,complicate student collaboration, thereby muting the positive effects of high-ability peers on studentperformance.Uganda, the setting for this study, consists of over 50 ethnicities (Uganda Bureau of Statistics2016). These ethnicities are geographically segregated, although there is considerable ethnic mixingin the capital of Kampala. Several studies link high ethnic heterogeneity in SSA to several pooreconomic outcomes, such as public goods provision, economic growth, and firm productivity, and6to negative effects on social indicators, especially social trust.1 A prevalent bias in favor of coethnicinteraction partly explains these documented costs associated with high ethnic diversity in SSA.Coupling this diversity with strong ethnic segregation, as is the case in Uganda, further exacerbatesmistrust (Alesina and Zhuravskaya 2011). This added friction to social interaction, cooperation,and collaboration is costly in general but may be especially apparent in student performance atuniversities that draw from disparate ethnic regions and, hence, where many students first interactintensively with ethnicities other than their own.This paper leverages the higher education context in ethnically diverse and segregatedUganda to explore the effects of coethnic and high-ability peers on academic outcomes. This uniqueempirical setting raises a number of questions that this paper studies. Does the share of coeth-nic peers within a student’s peer group affect academic performance more or less than the shareof high-ability peers? Do high-ability coethnic peers matter more than high-ability noncoethnicpeers? Does the context of Ugandan higher education translate into coethnic peer effects strongerfor some students than others? The contribution of this paper to the peer effects literature hingeson providing credible answers to these questions in this distinctive setting.In the empirical stage for this analysis, I link administrative records of student applications,admissions, and post-admission academic performance from a large public university in Uganda.These records include students enrolled in most of the STEM, social sciences, and business degrees inthe years 2009-2017 at this prestigious national university that is centrally located and, by admittingstudents from across the country, creates a microcosm of Uganda’s rich ethnic heterogeneity. Forthe purposes of this research, this feature is particularly interesting given the strong geographicsegregation of ethnicities in Uganda, which means that many students arrive at the university withlittle prior exposure to other ethnicities but are suddenly surrounded by the full diversity thatconstitutes the country as a whole. In the analysis that follows, I classify students who graduatedhigh school from their districts of origin as those with less prior exposure to other ethnicities andfor whom the ethnic diversity on campus is most salient.Being surrounded by coethnic peers at this large university might provide a sense of belonging1For example, cross country quality of government (Alesina and Zhuravskaya 2011); cross country public policies(Easterly and Levine 1997); productivity of a firm in Kenya (Hjort 2014); public goods provision in Uganda (Habya-rimana et al. 2007) Additionally, regarding public goods, Gisselquist, Leiderer, and Niño-Zarazúa (2016) show thathigh ethnic diversity may lead to welfare gains. For ethnicity and social trust, see Alesina and La Ferrara (2000)7and stability, thereby enhancing academic performance. Interacting with high-ability students cansimilarly improve performance in this setting because contact with instructors is limited (e.g., officehours are not offered), so learning from peers is important. In addition to testing the direct effectsof coethnic and high-ability peers, I also estimate the interaction effect of these two peer types sincehomophilous coethnic sorting could hamper or help learning from peers depending on the academicability of these coethnic peers. In this analysis, I rely on exogenous variation in the share of coethnicand high-ability peers in a given student’s peer group to test for these direct and interaction effects.The administrative data I use in this paper provide students’ demographic and academiccharacteristics, including whether they were admitted on merit scholarships, which I take as anindicator of high ability. These records do not, however, report student ethnicity. I overcomethis limitation by exploiting linguistic and cultural characteristics common to Uganda and SSA,where surnames reflect one’s native languages and, thus, ethnicity. To do so, I apply a machinelearning algorithm common in computational linguistics introduced in Cavnar and Trenkle (1994)and recently adapted by ? to the Ugandan context to a national administrative dataset of 2016voter registrations that includes over 14 million Ugandans. This external data set provides trainingdata I use to build a classification model that predicts ethnicities using student surnames.This paper’s causal identification of peer effects hinges on the random assignment of incomingstudents into dorms, which provides exogenous variation in peer groups. Specifically, a peer groupin this analysis consists of students admitted to majors in the same school and assigned to live inthe same dorm. Upon admission and conditional on gender, dorm assignment is random. Sincethere is excess demand for dorm beds, actual residence in dorms is not guaranteed. Some end upliving off-campus, but dorm assignments shape campus life for some of these students, as they mayengage in extracurricular activities within their assigned dorm. In addition to exogenous assignmentto peer groups, the econometric strategy exploits idiosyncratic year-to-year variation in coethniccomposition. Moreover, each student’s course list is predetermined at the time of admission, andstudents do not meet their classmates and dormmates until orientation week. Thus, the results inthis paper are not driven by selection into peer groups. I control for dorm, classroom (course-by-year), and major fixed effects to account for correlated shocks and differences that might confoundmy estimates.The results of this analysis indicate that the coethnic peers are as important as high-ability8peers in this setting, especially in the first year. That is, I find that coethnic peers (irrespective ofability) and high-ability peers (irrespective of ethnicity) increase a student’s performance in the firstyear. Specifically, adding five coethnic peers to a peer group of size 25, which would increase thenumber of coethnic peers in a group from the twenty-fifth to the seventy-fifth percentile, increasesa student’s performance by 0.19 percentage points. Additionally, the same change of adding high-ability peers to a group of 25 increases a student’s performance by 0.15 percentage points. Botheffects are significant at the 5% level and are about 0.02 standard deviation change in a student’sperformance in the first year.Nevertheless, the effect of coethnic peers disappears by the time a student graduates butthat of high-ability peers persists and even increases. Specifically, the effect of coethnic peers in thethird year, which is the final year for almost all the majors in this setting, is half of that observed inthe first year. Yet the effect of high-ability peers in the third year is 1.5 times that of the first yearin magnitude. Lastly, although I find that suggestive evidence shows that coethnic peers mattermore than high-ability noncoethnic peers as a student advances during university education, theeffects of both types of peers are statistically indistinguishable.Beyond the average effects, heterogeneous impacts indicate that the effect of coethnic shareis mostly driven by students of assumed high ethnic salience. For example, adding five coethnicpeers to a group of 25 increases the academic performance of a high ethnic salience student by 0.05standard deviations, which is 2.5 times the average effect. Nevertheless, like the mean effect ofcoethnic peers, this effect on students with high ethnic salience fades as a student progresses. Onthe contrary, coethnic peers have a positive and significant effect on high-ability, not low-ability,students that persist into the third year, suggesting that the benefits of coethnic peers throughouta student’s university career can be reaped by those posed to succeed when they enter university.Qualitative insights from the specific university context of this study align with potentialunderlying explanations for these results, including peer-to-peer learning and cultural and psycho-logical factors. In this setting, both coethnicity and academic ability are readily and generallyobservable. Incoming freshmen can easily identify coethnic peers through physical features and cul-tural characteristics, including names and language. As the academic year unfolds, they also learnwho among their peers are high-ability because publicly posted scores and grades reveal academicmerit scholarship status or through frequent interactions. Given the prevalence of ethnic student9organizations and activities on campus, which suggests a degree of homophily that shapes studentlife, it is natural for incoming students to seek out coethnic connections and support. Such connec-tions can be critical to a student’s successful transition to a novel setting of high ethnic diversityand, possibly, latent inter-ethnic tensions that may prevail on campus.This university setting is also characterized by classical lecture-style instruction with fewopportunities to interact with faculty or consult with teaching assistants, which makes informalpeer-to-peer learning especially important. For both incoming and continuing students havinghigh-ability peers in one’s peer group can therefore provide an advantage. If anything, the benefitof such informal peer tutoring increases as students progress to more advanced courses in theirdegree programs.The finding that a higher coethnic share matters on average and especially so for students ofhigher ethnic salience is suggestive of other psychological mechanisms. Enrolling at a large, centrallylocated national university may increase ethnic identity salience and attachment, as social identitytheory in Tajfel (1982) predicts. The presence of coethnic peers in one’s university environmentmay thus be beneficial for such students. This is similar to the finding reported in Okunogbe (2018),showing that the ethnic pride of Nigerian youth increases when they do national service in a regionwhere they are not part of the ethnic majority. Also, since several students are forced to navigate aspace that is diverse compared to their pre-university schools, having coethnic peers precludes inter-ethnic barriers. Moreover, I find the differential effect of the share of coethnic peers on studentsof high ethnic salience observed in the first year disappears as a student progresses. This indicatesthat through frequent cross-ethnic interactions at the university, these types of students make cross-ethnic networks and factors other than the shared ethnic identity of peers begin to matter more foracademic performance. This phenomenon can be interpreted by the contact hypothesis in William(1947). This might also explain why the effect of high-ability peers increases over time.This paper contributes to several strands of literature on peer effects in college. Althoughmixed, prior evidence largely indicates that post-secondary peer effects meaningfully impact educa-tion outcomes, such as major choice and academic performance. For example, Zimmerman (2003)and Sacerdote (2001) exploit random roommate assignments at US colleges to study roommatepeer effects. Zimmerman (2003) finds significant but small peer effects when using pre-treatmentacademic characteristics to measure peer quality and also detects nonlinear effects that are condi-10tional on the student’s SAT scores. Sacerdote (2001) finds null effects using the ability of a peerbut significant nonlinear effects at Dartmouth on academic outcomes. Additionally, he finds strongeffects on some social outcomes (e.g., fraternity membership). Carrell, Fullerton, and West (2009)argues that roommates are a small part of one’s college life, which might explain why Zimmerman(2003) and Sacerdote (2001) find no strong dorm or roommate peer effects. Exploiting exogenousassignments at the United States Air Force Academy, where students are assigned to peers withwhom they spend a majority of time together, Carrell, Fullerton, and West (2009) find strongeracademic peer effects than roommate peer effects. More recently, Mehta, Stinebrickner, and Stine-brickner (2018) use a panel data set that tracks students’ time allocation and friendships at BereaCollege and found that peers have an effect on study efforts.These peer effects studies primarily focus on college peers in the West. Their main econo-metric specifications include the average quality of peers measured by pre-treatment academic char-acteristics on the right-hand side variables. Given the setting, these papers also control for race,usually a binary indicator for white or black. In the SSA region, however, high ethnic diversityintroduces new complexity and nuance to peer effects. For high-ability noncoethnic might have anegative or null effect on academic performance if high ethnic diversity leads to inter-ethnic rivalriesand discrimination that spill into classrooms. I find the opposite: the identity of a high-ability peerdoes not matter. High-ability peers (irrespective of ethnicity) affect a student’s academic perfor-mance, suggesting that peer effects observed in studies in the West also exist in this setting. I findthat coethnic peers are also important in the first and second years.This paper also contributes to the literature exploring the role of ethnic diversity on economicand social outcomes in SSA more broadly (Easterly and Levine 1997; Habyarimana et al. 2007;Alesina and Zhuravskaya 2011; Miguel 2004; Gisselquist, Leiderer, and Niño-Zarazúa 2016; Alesinaand La Ferrara 2000; Håkansson and Sjöholm 2007; Hooghe 2007). In contrast to these moregeneral studies, this analysis focuses on a different question, albeit one with clear importance andpolicy relevance. Understanding peer effects from social networks play out in higher educationinstitutions with high ethnic diversity may enable more informed admissions and other academicprocesses, which often feature explicitly or implicitly in facilitating (or potentially undermining)cross-ethnic cooperation among young adults. High-ability peers affecting academic outcomes morethan coethnic peers as students progress may suggest that Ugandan youth are less ethnically biased11or able to adapt to ethnic diversity. However, it is important to note that coethnic peers mighthave lasting impacts on social networks outside school or other outcomes that are unavailable in mydata.Peer effects in higher education in SSA have been understudied for several reasons, includingdata constraints. A few studies that have explored college peer effects in the region use data froma South African university (Garlick 2018; Corno, Ferrara, and Burns 2019). Nevertheless, Garlick(2018) focuses on peer effects under two different assignment rules (random and residential tracking),while Corno, Ferrara, and Burns (2019) focuses on how exposure to roommates of another racechanges one’s stereotypes. Race (white vs black) is salient in South Africa for historical reasons andgeneral population composition, unlike other African countries. Therefore, I add to the literatureby studying higher education peer effects at a university in a context about which we know verylittle. I find that in this setting, coethnic and high-ability peer effects exist, especially in the firstyear.2.2 Background2.2.1 Ethnicity in UgandaUganda has over 50 ethnic groups that belong to three broader Bantu-speaking tribes (UBOS 2006).The largest nine ethnic groups constituted 71% of the population according to the 2002 UgandaPopulation and Housing Census. Groups may differ by traditions (e.g., dressing), language, food,economic activities, and sometimes by physical characteristics (e.g., skin tone). This pronouncedethnic diversity is also characterized by distinct geographic segregation as shown in Figure 2.1.Indeed, Alesina and Zhuravskaya (2011) rank Uganda the 4th most segregated countries in theworld based on a spatial segregation index.Historic migration and ethnic kingdoms drive these segregated settlement patterns. Bantu-speaking groups are clustered in the country’s South, Central, and Western parts, while Nilotic andNilo Hamites peoples are clustered in the Northern and Eastern parts. For purposes of the analysisthat follows, I retrace current ethnic borders to historic kingdoms (see Appendix Section 2.8.2).Inter-region migration is limited except for rural-to-urban migration into the capital, Kampala, foreconomic opportunities. By contrast, rural-to-rural migration across ethnic clusters rarely opens12Figure 2.1: Geographic Segregation.Notes: Ethnicity by district is the proportion of each ethnicity within a district. Data source: 2014 census.District shape files can be downloaded from https://data2.unhcr.org/en/documents/details/83043.economic opportunities and is limited due to cultural reasons.Although ethnic divisions existed in pre-colonial Uganda, some were exacerbated duringBritish colonialism (Tornberg 2013). The first post-independence government made efforts to reducethe importance of ethnic identities by abolishing historic kingdoms and preaching national unity,13an effort that met with resistance from some kingdoms, especially those with economic or politicalpower. The current government allowed ethnic groups to reinstate their historical kingdom; someethnic groups did. While current inter-ethnic competition and recent historical conflicts can betraced to political and sometimes historical factors (Mamdani 2001), inter-ethnic competition oroutright conflict is generally not as intense as in neighboring countries.Although English, the official language of Uganda, is spoken in public offices and taughtin schools, native linguistic diversity is high.2 Differences between native languages are correlatedwith physical distance, implying that one may partially comprehend the language of a neighboringtribe. Luganda is the most familiar native language because it is native to the Kampala region. Iexploit this language diversity to predict ethnicity in Section 4.2.2.2 Ugandan Higher Education and Makerere University KampalaAlthough Uganda has one of the youngest populations in the world, post-secondary school educa-tion is low: the post-secondary enrollment rate for college-age Ugandans was only 6.85% duringthe 2017/18 academic year (NCHE 2018). Nine public and 44 private universities offered degreeprograms during the 2018/19 academic year (NCHE 2018), of which Makerere University Kampala(MUK) ranks first in quality and size.MUK is well-known in the SSA as it is one of the oldest universities in the region. Itwas established in 1922 as a technical school to facilitate training workers for the British colonialgovernment. It is centrally located in Kampala and admits students from across the country. Forsome students, it is at this university that they meet and interact with people of different ethnicitiesfor the first time. With the exception of Baganda, the diversity of the MUK student populationmirrors that of the country as a whole (see Figure 2.2).2WorldAtlas reports Uganda’s language diversity index of 0.929, which indicates that most Ugandans speak atleast one native language.14Figure 2.2: Distribution by ethnicity: student sample vs general populationSource: MUK admissions, 2009-2017 and 2014 Census (UBOS)2.3 Empirical Setting2.3.1 Applications, Admissions, and Sample DefinitionThe Ugandan public university and pre-collegiate nationwide system offer a unique setting thatI exploit to identify coethnic peer effects. First, national pre-collegiate exit exams and publicuniversity merit scholarships are centrally administered. Second, the Uganda Examination Board,an organization separate from MUK, runs an algorithm for all MUK admissions. Thus, there is noroom to manipulate the composition of its student population.Students are admitted under two schemes: (I) National merit scholarship and (II) self-funding scheme. A student lists up to six majors in order of preference during application. Admissionto a major (cutoffs) is a function of the student’s preference set, admission in national exams, and15the university’s capacity. A student’s major (and course list) are predetermined during admission,3-4 months before enrolling. Each major non-extension major is housed within a school, which is asmaller unit within a college. A school is locally termed as faculty or department, but I will adoptthe ‘school’ term for simplicity.Students in the same majors take almost all their first year classes together since courses havepredetermined sequencing. Still, they interact with students from other majors within classrooms,usually within the same school, who share the same course requirements on a daily basis. In addition,students within a school usually share common spaces, such as computer labs, food canteens, studyrooms, and libraries.Students cannot select into different sections within the same major, as sections do not existin this setting. Because of this, most student’s course sequence is also pre-determined before astudent reports to campus. The performance data show that over 98% of classes in each year arenon-elective. Moreover, The university offers evening and day class options as ‘different’ majorswhen a degree, such as business administration, is in demand. Still, the day class is a ‘different’major from the evening class, and students must apply and get admitted to either the day orevening class cohort separately. For example, students who intend to obtain a Bachelor of BusinessAdministration degree can apply and be admitted to either the day cohort or the evening cohort.Students admitted to the day cohort cannot take classes and sit for their exams with studentsadmitted to the evening cohort. I restrict the sample to day cohorts as evening majors do notqualify for the national merit scholarship. This is important because the merit scholarship is mymeasure of high-ability as I define in the coming sections.Students stick with the majors offered during admissions but can apply to change withinthe first two weeks of their freshman year. Approvals depend on the capacity of the intended majorand student performance and are thus rare. I find major change cases are less than 2.75% in theten-year period of my sample.3 Non-STEM majors, especially business and social sciences, tend tohave relatively large class sizes.3I compare the major student’s enrollment and the major at the admissions and find a mismatch of 2.75%. Thisnumber includes students whose major switch applications were approved and possibly some data entry errors whenentering admissions data.162.3.2 Dorm Assignment and Defining Peer GroupsConditional on gender and upon admission, dorm assignment is random. MUK has nine single-sexlarge dorms: three are female and six are male dorms. There are more incoming students assignedto dorms than there are beds to accommodate them. I observe dorm assignments but not thesubsequent residence status and room assignments. Each student’s admission letter indicates theassigned hall, which determined by the administration by simple random assignment. Studentsmust formally apply to their assigned dorm for residence, at which point a dorm administrator andcommittee allocates beds according to a university-wide priority list that favors students on nationalmerit scholarships in majors and schools perceived to be especially rigorous, such as medicine andengineering.The remaining beds are then assigned to students according to the order of their dormapplication. While students who are not allocated a bed in their assigned dorm must arrangefor their own housing off-campus, their initial dorm assignment continues to shape campus life asassigned students have access to shared spaces with entertainment and dining facilities in thesedorms. Extracurricular activities such as student government elections are also organized by dormassignment irrespective of residence.In general, a peer group consists of individuals with shared or similar characteristics whointeract in social or other settings. Specific definitions of peer groups are context-specific. Carrell,Fullerton, and West (2009) define a peer group as a squadron at the US Air Force Academy, whileFoster (2006) consider a peer group to be students living on the same dorm floor and Pre-collegiatestudies, such as Carrell and Hoekstra (2010), use a cohort definition. In the MUK context, I definea peer group, as illustrated in Figure 2.3, as students within the same school who are assigned tothe same dorm. Although most of the prior studies on college peer effects observe roommates, I donot observe room allocations and residence status, so I restrict my definition to dorm assignment.By focusing on the cohort-residential peer groups, I use a “strict" definition of a peer group,but it also allows me to study peers with whom a student spends most of their time. For instance,students within the same school and dorm may spend a lot of time together, such as walking to andfrom classes, attending classes together, and sharing common spaces at school and within the dorm.One may argue that a major hall year is a better peer group since students in the same major take17Figure 2.3: Peer Group ConstructionNotes: This diagram illustrates a peer group definition. Students in these defined peer groups are much morelikely to interact regularly with each other, including those of the same or different ethnicity. High-ability studentsare defined as those on merit scholarships, a status that is widely known among all students.100% of their classes. However, the focus of this paper is coethnic and high-ability shares, and usinga way smaller peer group definition reduces variation in the coethnic share as most of the coethnicshare of the smallest ethnicities will be zero in a lot of peer groups.2.3.3 Identifying Peer Effects at MUKEstimating peer effects may be econometrically challenging for three reasons: self-selection (Hoxby2002), endogeneity (Manski 1993), and correlated common shocks (Bramoullé, Djebbari, and Fortin2009). This section highlights the characteristics of this setting that provide solutions to these issuesrelated to measuring peer effects.Self-selection arises when people choose to join a group based on some pre-treatment char-acteristics. As stated in Hoxby (2002) “.. if everyone in a group is high achieving, many observersassume that achievement is an effect of belonging to the group instead of a reason for belonging toit." In the case of colleges, self-selection exists because students can select into classrooms, majors,and sometimes, dormitories. Peer effects literature typically employs two strategies to deal with18selection in peer effects papers. First, conditional on some pre-treatment characteristics, such asgender and ability, peers result from random assignment (Sacerdote 2001; Zimmerman 2003; Car-rell, Fullerton, and West 2009; Foster 2006). However, random assignments into classrooms in USstudies are difficult. So, with the exception of Carrell, Fullerton, and West (2009), these studies usea setting where roommates at some universities are randomly assigned.The second approach involves exploiting natural variation in a cohort or group composition.The idea behind this approach is that year-to-year variation (e.g., gender, race, and class size) ob-served at the group level is a reflection of a natural variation in a general population (idiosyncratic).This approach has been used in pre-collegiate peer effects studies (Carrell, Hoekstra, and Kuka2018; Hoxby 2000)My approach leverages characteristics of this setting described in Section 2.3.1. Peers areclassmates who potentially live together. As aforementioned, conditional gender dorm assignmentat MUK is random. Since I do not observe roommate assignments and residence status, this paperestimates the intent to treat (ITT) of the peer qualities defined later. Unlike most US universities,students do not select courses or majors post-admission, which has the convenient feature thatstudents do not sort into classes or classrooms based on characteristics or exposure (or not) todifferent types of peers.The reflection problem is the endogeneity problem challenge, which arises from a feedbackloop of peers. This is a challenge because a student’s and their peers’ outcomes are simultaneouslydetermined. One of the approaches in the literature is to use preexisting characteristics that areexogenous to the dependent variable, such as race and gender. For example, Carrell and Hoekstra(2010) uses the presence of family problems when studying peer effects of children linked to domesticviolence on academic outcomes. I use pre-collegiate characteristics, as most of the literature, toexploit exogenous variation in treatment variables, which are coethnic share and high-ability sharewithin a student’s peer group.In Uganda, students’ ethnic identities are determined at birth. An argument may be madethat ethnicity is part of multifaceted identities, a function of collective cultural traits, and thatan individual’s ethnicity may change through self-identification (Sen and Wasow 2016). I am notconcerned that this exists in Uganda to the extent that it would confound my estimates. First,I follow the official categorizations of ethnic groups in UBOS (2006), and admissions do not have19ethnic quotas or any form of affirmative action based on ethnicity. Thus, there is no incentive tochange one’s ethnic identity during university applications. Second, I use linguistic characteristicsto predict ethnicity instead of self-identification. I describe these variables in Section 2.4.2 below.The last main challenge is contemporaneous common shocks, especially if they are correlatedwith academic performance. My setting uses random assignments at the same university, whichreduces the possibility of such shocks. Nevertheless, there may be shocks that affect some peergroups differently. Thus, the main regressions include all group fixed effects, such as dorm andclassroom, to account for observed characteristics that might confound the main effect.2.4 The Data2.4.1 Academic and Demographic CharacteristicsPre-university CharacteristicsThe analysis in this paper uses several data sources: MUK’s administrative records on academic anddemographic characteristics observed from applications, admissions, and post-admission academicperformance for students entering the university during 2009-2017, and ethnicity is predicted bystudent surnames.I observe students’ application data from 2009 to 2017. The student applications includethe student’s name, type of application, admission scheme (merit scholarship or private scheme),and offered majors, as well as age and religious identity. All student records are de-identified pre-analysis, although most student admission data, such as major, are publicized on university noticeboards and in newspapers.Measure of High-abilityEvery year, 4,000 students are admitted to public universities on a government merit sponsorshipbasis of performance in high school national exams, most of which enroll at MUK relative to otherpublic universities (HESFB Uganda, 2012). These scholarships are awarded to the top studentswithin a major, and the number of spots per major is relatively constant across years. Meritscholarship application forms are submitted at the time of national exam registration before students20take their exams. Therefore, almost all A-level graduates are automatically considered for thegovernment merit scholarship, as the sponsorship does not require a separate application. Studentsare ranked based on their high school GPA within their preference set, and the top students areoffered a scholarship until each major’s scholarship spots are filled up. That is, the scholarship isdetermined by high school GPA.4High school GPA is a proxy for ability as it may pick up a student’s innate ability, effortduring high school, and success in an academic context. I therefore use this as an imperfect butinformative proxy of “academic potential," which accounts for both a student’s subject combinationand performance in this selected subject combination. Each major has a high school subject com-bination required for a student’s successful college career from a university’s perspective. I define“high-ability" students within each major as those enrolled with the national merit scholarship.Lastly, it is usually public knowledge which of a student’s peers are admitted through merit as uni-versity registration numbers differ by merit status. Moreover, admission lists are usually publishedin newspapers and university notice boards.University Academic performanceI observe student transcripts from 14 departments belonging to six colleges. Each student’s tran-script lists all courses taken, credit units, and performance in percentages by semester year of studyduring which the course was taken. Therefore, I can observe these students’ classmates and howthey have progressed from matriculation to completion of coursework. Unlike schools in the West,letter-grade ranges assignment is the same across all majors, and professors do curve grades. Pro-fessors at MUK do not assign letter grades. They submit each student’s performance on a 0-100%scale, and the central system assigns the letter grades. Also, most majors take three years to com-plete, and thus students take a lot of courses per semester (a min of six, and some majors requirestudents to take up to ten courses in some semesters).4There are a few variations. For example, Ugandan public universities have a gender affirmative action policy thatawards a ‘free’ 1.5 additional points to every girl during admission. This 1.5 free point is also awarded to girls whenthey are being considered for non-merit admission schemes. In addition, a small proportion of the merit scholarshipis awarded through the district quota to the top four students graduating from their district of origin who did notobtain the merit scholarship through the direct route. Therefore, the number of district quota spots is proportionalto ethnicity size. District quota applications are made at the same time as the national merit applications.212.4.2 EthnicityUniversity applications and admissions do not capture the ethnic identity of students, althoughethnicity is one of the most salient identities among Ugandans. I overcome this by exploitinglinguistic differences reflected in surnames. Ugandans’ surnames are in their native languages.5This naming pattern is not random or unique to Uganda. Historically African parents chose namesintentionally. However, with the arrival of colonists, first names are now in foreign languages, suchas English (in Anglophone countries) or French (Francophone countries). The meanings of mostUgandan surnames can be traced to the father’s tribal clan and religiosity or prevailing conditions atthe time of birth, among others. These are linguistic characteristics I use to predict one’s ethnicity.Data Appendix 2.8.2 describes how I trace ethnic boundaries from current administrative units tohistorical kingdoms.Using surnames to trace one’s identity is not new in economics and other fields. For example,surnames have been used in mobility studies to trace wealth across generations within a family inthe West (Barone and Mocetti 2016; Clark and Cummins 2015). Some studies have also usedsurnames to predict ethnic identity across several countries. For instance, Bhusal et al. (2020) usesurname frequency in the Nepalese historical censuses to predict one’s caste in their paper studyinghow revolutions may have altered political representations and inclusion in Nepal. Using fuzzymatching and naïve Bayes machine learning techniques on historical records, Monasterio (2017)studies surnames and ancestry in Brazil.Predicting Ethnicity and Constructing Coethnic ShareMore related, ? exploits rural-urban linkages in Uganda and applies machine learning on represen-tative Uganda surnames to predict the rural origin of Uber drivers in Kampala. His study exploreshow Uber drivers adjust their online hours when their probable ancestral homes experience a nega-tive weather shock. Therefore, agroecological zones form a basis for his predictions. His procedure,like the machine learning section in Monasterio (2017), follows a text categorization procedure devel-oped by Cavnar and Trenkle (1994). This process has been widely used in computational linguisticsand involves breaking down a name into N-grams.5Trevor Noah mentions the same pattern in South Africa in his book “Born a Crime" (PP.). Also, see thishttps://www.bbc.com/news/world-africa-37912748 for another example22Following the literature, there are two common approaches: use frequencies to predict prob-abilities as in Bhusal et al. (2020) and train a machine learning algorithm on some training dataset by applying tools, such as gradient boosting. Method (I) computes simple probability usingthe frequency of each surname. Suppose {E1, E2, ..., En} is a set of all ethnicities in a population.Also, suppose Ns∈Ei is the number of times a surname, s, belongs to an ethnicity, Ei. Then theprobability of belonging to a particular ethnicity is computed as:(2.1)Ns∈Ei∑∀nNs∈EiTo illustrate, consider the surname “AHIMBISIBWE": it appears 17,559 times in the name trainingdata, of which 13,904 occurrences in the Ankole region/ethnicity. Therefore, there is a 79.2%probability that a student with the surname “AHIMBISIBWE" is of Ankole ethnicity.Method (II), which is my preferred, follows ? and computational linguistics and begins bybreaking a surname into N-grams. Taking “AHIMBISIBWE" as an example, Method (II) breaksthis surname into 1-grams (“A", “H", “ M", “B", “I", etc); 2-grams (“AH”, “HM”, “MB”, “BI", etc);3-grams (“AHI", “HIM", “IMB", etc..) and so forth. The algorithm can now count the number offrequencies each n-gram appears in a surname and in each region. The most common weightingapproach used in linguistics is the term frequency-inverse document frequency (tf-idf) that combinesapproaches developed by Luhn (1957) and Jones (2004). I then apply gradient boosting on N-gramsand tf-idf features on an external data set described in Section 2.4.2, producing a classificationmodel that I apply to students’ surnames.The second method is preferred to the first in this paper for two reasons. First, by fol-lowing the tf-idf weighting procedure, the algorithm picks each surname’s most unique linguisticcharacteristics. Second, it does not require an exact match in the name database.This algorithm predicts N probabilities if we have N ethnicities. Given that languages are notexclusively unique, some probabilities are non-zero or 1. We can, therefore, interpret the predictedprobabilities as a measure of ‘confidence’ that a student belongs to a particular ethnicity. Oneapproach is to use ethnicity corresponding to the top predicted ethnicity (the ethnicity a predictionis most confident about), as is common in the literature.23I can then compute the share of coethnic peers using two approaches: (A) and (B). Firstly, bysingle ethnic identity assignment (A), I assign an individual a single ethnicity category correspondingto the group the algorithm is most confident about. This is common in studies employing machinelearning algorithms to predict ethnicity, religion, or area of origin in the literature. This methodassumes that individuals’ top predicted ethnicity corresponds to the ‘true’ ethnicity and treatsethnicity as a categorical variable without considering potential measurement errors. Using toppredicted is common in the literature. The average probability corresponding to ethnicity in thealgorithm is most confident equals .792 (median=.861), which is high.Secondly, by joint probability estimation, I consider all the probabilities that a given surnamebelongs to different ethnicities. This method acknowledges potential measurement errors associatedwith using categorical variables for ethnicity. It estimates the probable fraction of coethnic peers ina peer group by considering the joint probabilities of two individuals belonging to the same group.That is, student i’s share of coethnic peers in a peer group G, SEiG is computed as:(2.2) Using category assignment (A) : SEiG =∑k∀̸=iNumber of coethnicsNG − 1(2.3) Using joint probability estimation (B): SEiG =16∑e=1NG−1∑∀k ̸=iΠeiΠe′iNG − 1,where NG is the peer group size, Πei is the predicted probability that an individual i belongsto ethnicity group e. Lastly, I collapse ethnicities to 16 ethnic/language groups as described inAppendix 2.8.2. The main analysis uses the probable coethnic share in a group in equation (2.3),but the results remain unchanged when I use the share of coethnic peers computed using equation(2.2). Throughout this paper, I use the share of coethnic peers and the probable share of coethnicpeers synonymously for simplicity and ethnicity to mean the most probable ethnicity in the empiricaland results sections.24Training DataI use nationwide voter registration to train the machine learning model (gradient boosted). Thesedata contain names, voter ID numbers, date of birth, sex, polling station, and area of residence. Thearea of residence is given for all units of administration parish, sub-county, county, and district. Ilink these voter data to spatial administrative and public data containing ethnic boundaries tracedfrom historic kingdoms described in Appendix 2.8.2.People register to vote from a polling station within their parish of residence. Moreover,in many cases, people who live in cities outside their areas of origin often register to vote in theirancestral homes. Because voter registration is manual, it only takes place once between elections.A few Ugandans own cars to travel, so most walk, as long-distance public transportation is costly.Thus, the cost of registering to vote in a village different from their residence is high.2.4.3 Descriptive StatisticsTable 2.1 provides summary statistics for demographic and academic characteristics in Panel A andpeer group averages in Panel B. About half of the student population is female, and about 31%are high-ability (enrolled through national merit scholarships). Most of the students have declaredreligion, and as expected in this context, most students are either Catholic or Anglican. The averageage of incoming students is 20. Lastly, about 36% of the students in the sample graduated from ahigh school with their home district (these are the type of students I assume to have higher ethnicsalience). Because of Uganda’s high ethnic segregation (Figure 2.1), this group of students may nothave interacted with peers of different ethnic groups.Given my peer group definition, the average group size is 25, which is small, albeit the SDfor peer group and cohort sizes are large. Close to 75% of the peer groups are of size 50 and below,as the Appendix Figure 2.6 portrays. STEM majors, which comprise most of my sample majors,usually admit a few students relative to other majors.The average coethnic share is 25%, and the average co-ethnic share on merit is 7%. To explorehow treatment intensity (shared ethnicity) may vary by ethnicity and if MUK is representative ofUganda’s ethnicity distribution, I plot the distribution of ethnicities in Figure 2.4. The CDFs inthis figure show that 80% of peer groups have a coethnic share of 0.2 or less.25Table 2.1: Descriptive StatisticsN Mean SDPanel A: Student characteristicsHigh-ability 25,487 0.31 0.47Age 25,487 20.16 1.43Female 25,487 0.48 0.50High ethnic salience 25,487 0.36 0.48Anglican 25,487 0.37 0.47Catholic 25,487 0.31 0.47Muslim 25,487 0.09 0.29Pentecostal 25,487 0.06 0.24Seventh Adventist 25,487 0.02 0.13Unspecified Religiosity 25,487 0.05 0.22Other Religions 25,487 0.01 0.09Panel B: Peer group variablesPeer group Size 996 25.64 21.64High-ability share 996 0.35 0.24Coethnic share 996 0.24 0.19High-ability coethnic share 996 0.08 0.12Low-ability coethnic share 996 0.15 0.15Panel C: Course levelAll year grades (%) 1,061,905 67.83 9.61Year One grades (%) 321,538 66.90 9.65Year Two grades (%) 343,950 67.50 9.80Year Three grades (%) 330,626 68.52 9.26Notes: Data are from MUK and are restricted to students admitted to non-extension day majors at sixcolleges for 2009-2017, excluding 2010. A peer group comprises of students admitted to majors within aschool major in the same year and assigned to the same dorm. Unspecified religion indicates wheneverreligious identities are not provided or entered as “Christian". Christianity is usually a correction of severalor nondenominational religions in this context. Religion “Other" includes the smallest religions (where thecount is less than 100 in the sample), such as Bahai, Jehovah’s Witness, traditional religions, and Intambiro.Apart from age, Panel A variables are constructed to be binary.2.5 Empirical Strategy2.5.1 Mean EffectsDirect Effects of Coethnic and High-ability PeersAs aforementioned, a peer group refers to students admitted to majors in the same school f andassigned to dorm d in year t. For simplicity, I will index the peer group fdt with G in the estimationequations in this section. To estimate the direct effects of coethnic or high-ability peers on academicoutcomes, I use a model that exploits variation in coethnic composition across peer groups within26Figure 2.4: Distribution of Coethnic Share across Peer Groups.Notes: Data used to produce these distributions are from MUK and are restricted to students admitted tonon-extension day majors from six colleges for 2009-2017, excluding 2010. Coethnic share is computed as theleave-me-out proportion of coethnic peers in a group. This figure plots the 10 largest ethnic groups by the totalnumber of MUK students (out of the 16 total ethnic groups).a year and year-to-year variation:(2.4) yijcG = β0 + ϕ1SEiG + ϕ2SHiG + β2XiG + β3X̄G + δj + αc + λd + θm + γs + εijcG,where yijcG is the first year percent grade that student i of ethnicity j and belonging to group Gobtained in course c. SEiG is the probable coethnic share of in i’s peer group defined in Section 2.4.2in equation(2.3), and SHiG is the share of high-ability peers (coethnic and noncoethnics). The mainestimation controls for δj , which is i’s most probable ethnic group, XiG is a vector of i’s backgroundcharacteristics and includes i’s own ability, and X̄G =∑∀k ̸=iXiGNG−1 is the vector of exogenous variables(the average background characteristics of i’s peers, except high-ability). Additionally, αc, λd, θm,and γs represent classroom, dorm, major, and high school subject combination fixed effects (FE).Lastly, εijcG is the error term. I cluster standard errors at the peer group to account for the potentialerror correlation across individuals in a group.27The coefficients of interest are ϕ1, which captures the effect of attending lectures and poten-tially living with coethnic peers in this setting, and ϕ2, which captures the effect of attending classand potentially living with high-ability peers irrespective of their ethnicity. I take several steps toensure that ϕ1 and ϕ2 are unbiased. I control for several FE to deal with bias arising from correlatedshocks.First, correlated shock in this setting may arise from differences across classrooms. There-fore, I include classroom FE to control for unobserved differences in courses, such as performance,instructor effects, and classroom diversity. In addition, classroom FE should control for differencesin major by year since the student’s major and course list are determined at the time of admis-sion. Nevertheless, students may take courses with peers admitted to majors outside their schoolsif cohort sizes are small and major course requirements are related. This implies nonrandom ex-posure to coethnic peers because of the systematic differences in the share of some ethnicities inthe MUK sample and general population. Thus, αct also controls for this systematic difference inethnic exposure across students in addition to controlling ethnicity FE. Relatedly, I include majorFE, θm to control for differences between majors. Each regression will control for ethnicity, major,and classroom FE at the minimum in the results section.Second, I control for dorm FE to control for factors, such as renovation and dorm conditions,that might affect academic performance. In addition, cultures are different across dormitories. Forexample, Ricart-Huguet and Paluck (2023) show that cultures, such as outgoing and academicmindedness, are different across MUK dorms to the extent that they affect interpersonal outcomes.Third, although evaluated at the same cutoffs, students entering the same major may oc-casionally take different subject combinations during upper high school. Therefore, I include highschool subject group FE, γs, to capture the differences in types of incoming students. When comput-ing high school weighted GPA, each major has different requirements to capture incoming students’academic preparedness. Take Bachelor of Commerce, for example, the required HS subjects aremath and economics, but students who take one of the two and those who take both can qualifyif they perform above the cutoffs. Students graduating with math and economics have a higherperceived potential for success in Bachelor of Commerce classes than those graduating with oneof the two subjects. Therefore, controlling γs captures the unobserved differences in academicpreparedness across students in the same major.28Concerning self-selection, dorm assignment is random, and each student’s course list andclassmates are predetermined before entry at the time of admission, as mentioned in Section 2.3.1.Two lines of concern can be made for potential sources of self-selection. First, although dormassignment is random, the on-campus residence may be biased to STEM students admitted throughthe merit scholarship. This is only statistically meaningful if merit scholarships are correlated withethnicity and if dorm assignment was not random.Correlation between ethnicity and obtaining a scholarship in a STEM major is possible ifthe top secondary schools are concentrated in one ethnic region, where students from that regiongraduate with the highest A-level scores to qualify for the merit scholarship. As Panel B of Figure2.2 shows, the student population is biased towards the two largest ethnic groups. Coincidentally,the most elite secondary schools in the country belong to these regions because of historical reasons.Nevertheless, this is not an issue, as dorms and majors do not have ethnic quotas and equation (A1)controls for i’s probable ethnic group, which controls for differences in the levels of stratification.Also, when I regress merit ethnicity fixed effects, I find the explained variation is less than 1%.Additionally, students may select into majors by manipulating the rank of their choices. Thismight cause selection into majors even though an organization separate from the university handlesadmissions and even though obtaining admission is quasi-experimental. This is possible since theranking of program cutoffs does not change from one year to another, although actual cutoffs maychange. This is not a concern as a peer group of classmates who potentially live together, and dormassignment is random.As a test, I provide balance tests in Table 2.2, which presents evidence against selection.Each column is an independent estimation similar to specification (A1). I run these regressions atthe aggregated to the student level (not course level). Each pre-university characteristic is regressedagainst the coethnic share in Panel A, while in Panel B, each pre-university characteristic is regressedagainst the high-ability coethnic share. Panel B also controls for a student’s ability. Additionally,I regress the share of coethnic or high-ability peers on all the pre-university characteristics andreport the estimates and the F stat in Appendix Table A4. The correlation between pre-universitycharacteristics and the primary variable of interest would be high and significant if nonrandomsorting into peer groups existed.From Table 2.2, the correlation between each student’s characteristics and the share of29Table 2.2: Evidence against Selection(1) (2) (3) (4) (5) (6) (7) (8) (9)Age Anglican Catholic Muslim Pentecostal SDA High EthnicSalienceOtherReligion High-abilityPanel A: Coethnic share as the independent variableCoethnic share -0.00 -0.01 0.02 0.01 -0.02 -0.01 0.04 0.00 0.02(0.11) (0.04) (0.03) (0.02) (0.02) (0.01) (0.04) (0.01) (0.03)R-squared 0.13 0.07 0.05 0.27 0.03 0.03 0.11 0.06 0.17N 25,487 25,487 25,487 25,487 25,487 25,487 25,487 25,487 25,487Panel B: High-ability share as the independent variableHigh-ability share 0.06 -0.00 -0.02 -0.01 0.02 -0.01 0.03 0.01(0.08) (0.03) (0.03) (0.01) (0.01) (0.01) (0.03) (0.01)R-squared 0.13 0.07 0.05 0.27 0.03 0.03 0.13 0.06N 25,487 25,487 25,487 25,487 25,487 25,487 25,487 25,487 25,487Notes: Data are from MUK and are restricted to students admitted to non-extension day majors fromsix colleges for 2009-2017, excluding 2010. Each column is an independent regression that regresses apre-university characteristic against the share of coethnics. All regressions include school-by-year (notclassroom), ethnicity, and hall FEs. Also, all regressions in panel B control for own ability. Standarderrors are in parentheses and clustered at the peer group level.*p<0.1, **p<0.5, ***p<0.01coethnic peers (Panel A) or the share of high-ability peers (Panel B) is practically zero and notsignificant in all regressions, providing evidence that students are not selecting into peer groups. Inaddition, the F stats from Appendix Table A4 are small, 0.84 and 2.15 when the share of coethnic(column one) or share of high-ability (column two) peers are regressed against all the pre-universitycharacteristics, respectively. This indicates that the results presented in this paper are unlikely tobe biased because of nonrandom sorting.Interpreting the Magnitude of ϕ1 and ϕ2As Section 2.3.2 describes, not all students assigned to a dorm end up residing in their assigneddorm due to capacity constraints. Living off campus does not, however, exclude a student fromdorm-based peer groups; it just alters the nature and frequency of interactions. Given my peergroup definition, consider two types of students based on the extent of interactions with others ina given peer group: fully compliant and partially compliant. Fully compliant students live in theirassigned dorms and can thus interact as dorm residents with other fully compliant peers and, inother ways, with their partially compliant peers. Partially compliant students live off-campus and,therefore, do not interact as dorm residents with others in the peer group I construct for them.Both types of students are likely to interact daily (within and across each type) in classes and study30groups.If I observed residents, I could estimate the local average treatment effect of coethnic andhigh-ability peers using dorm assignment as an instrument for dorm residence to account for endoge-nous dorm residency. Since I do not observe residence, ϕ1 and ϕ2 in equation (A1) are effectivelyreduced-form peer effects estimates based on dorm assignment. These reduced-form effects are adata-weighted average of the peer effects for fully compliant peers and partially compliant peers.6I expect the reduced-form peer effects in Equation (A1) to be less than the local averagetreatment effect of coethnic and high-ability peers. Also, the existence of partially compliant peerswill likely attenuate peer effects. To illustrate the logic behind this claim, consider a parallelwith Carrell, Fullerton, and West (2009), who use dorm floors to reconstruct peer groups at theAir Force. Their empirical setting allows them to construct pseudo-peer groups that span therelevant peer group (squadrons). Interactions are expected to still exist within these pseudo-peergroups but at a reduced rate than the true squadron-based peer groups. Although these pseudo-peer groups comprised 66.6% of peers from squadrons, the presence of peers with whom studentsinteract less frequently attenuates estimated peer effects. Analogously, in the MUK context, I expectthat reduced-form peer effects based on dorm assignment and, therefore, including both residentand non-resident students in peer groups will underestimate the true peer effects operating in thissetting.Coethnic and High-ability Interaction EffectsAlthough evidence in the literature is mixed, the effect of the high-ability peers (ϕ2) is expectedto be positive. Classrooms in the setting are relatively large, and services, such as office hours, donot exist, making peer-to-peer learning important. The sign of the coethnic share coefficient (ϕ1)is largely ambiguous. The coefficient ϕ1 could be zero on average, although it could be positive ornegative for several reasons.Bayer et al. (2020) show that minority students in introductory economics classes report alower sense of belonging than non-minority students, and some studies (e.g., Walton and Cohen6Even knowing the overall proportion of each cohort residing in dorms (i.e., residence compliance) does notnecessarily solve this issue as I cannot use this compliance rate as the first stage to inversely weight reduced-form inthe equation proposed in Bloom (1984) and as equation (A4b) in Appendix Section 2.8.3 without additional (strong)assumptions.312011) show that interventions to increase a sense of belonging improve academic outcomes forstudents. Additionally, some students might suffer from imposter syndrome exacerbating theirsense of belonging. Thus, having coethnic peers might be significant for some student as it mayincrease a sense of belonging during student interactions.On the other hand, students of shared ethnicity may gravitate toward one another for culturalreasons, such as language, traditions, and beliefs. These homophilous tendencies and coethnic biasduring interactions in diverse societies might lead to ethnic-based sorting into study and friendshipgroups. In this case, the effect of coethnic peers on academic performance will be indirectly throughhigh-ability coethnic peers. For example, it might be detrimental for coethnic peers to isolate ifthey are, on average, low-ability compared to noncoethnic peers. A low-ability student might benefitfrom a higher share of high-ability than a higher coethnic share in a peer group.To capture the effect of the pre-university academic quality of coethnics, I use an equationsimilar to equation (A1).(2.5) yijcG = β0 + ϕ1SELiG + ϕ2SEHiG + ϕ3SE′HiG + β2XiG + β3X̄G + δj + αct + λd + θm + γs + εijcG,where SEHiG , SELiG and SE′HiG are the probable shares of high-ability coethnic, low-ability coethnic,and high-ability noncoethnic peers, respectively. All other terms are the same as those in equation(A1).Therefore, the setting provides four sources of variation of interest in the share of peers thatis: (A) high-ability and coethnic; (B) low-ability and coethnic; (C) high-ability and noncoethnic;and (D) low-ability and noncoethnic. Coefficient ϕ1 in equation (A1) captures the average effect of(A) and (B), while ϕ2 captures the average effect of (A) and (C). If students are sorting on ethnicitywhen forming study groups, (A) should matter than (C). In such cases, we can think of the coethnicpeers operating indirectly through high-ability ethnic peers.2.5.2 Heterogeneous Peer EffectsTo estimate heterogeneous effects, I estimate equation (A1), including the interaction of the twotreatments with the dummy variable that captures the heterogeneous dimension listed below.32(2.6)yijcG = β0+ϕ1SEiG+ϕ2SHiG+φ1SEiG×di+φ2SHiG×di+β2XiG+β3X̄G+δj+αc+λd+θm+γs+εijcG,where di can be gender, ability, or assumed level of ethnicity salience, while φ1 and φ2 are thedifferential impacts on di of coethnic and high-ability share, respectively. All other terms are thesame as those in equation (A1).If coethnic and high-ability peers matter for academic performance, the average effects maybe conceptually different depending on dimensions, such as the size of each ethnicity at the universityand level of prior exposure to noncoethnic Ugandans. If increasing a sense of belonging is a channelthrough which coethnic peers might work, then the effect of coethnic peers could be zero for largeethnic groups who are less likely to suffer a low sense of belonging. However, coethnic peers maymatter positively for small ethnic groups with limited exposure to different ethnicities prior toUniversity.Also, coethnic peer effects in the presence of high ethnic heterogeneity may also matter dueto inter-ethnic impacts. For instance, there are ethnic groups that share values with other groupsor portray less in-group bias. In such cases, fewer co-ethnic peers may not matter as those studentswould easily integrate with other ethnicities. More broadly, if inter-ethnic uncongenial relationshipsexist in Ugandan societies, they could spill over into schools, creating ‘bad’ peers. Nevertheless, thisis unlikely in Uganda, as inter-ethnic tensions are not that common. The analysis will, therefore,explore heterogeneity in other dimensions.Differential Effects by GenderSeveral studies report differential peer effects on academic and non-academic outcomes by genderin several settings. For example, Carrell and Hoekstra (2010) study peer effects of kids exposed todomestic violence on test scores and disciplinary incidents in a classroom and find that peer effectsare significant and stronger for boys, not girls. Additionally, Stinebrickner and Stinebrickner (2006)use HS GPA to study peer effects on sudy habbits and academic perforamcne at Berea College andfind that HS GPA captures the effect on study habits and significant peer effects on girls. Morerecently, using a field experiment at a public school in Peru, Zárate (2023) finds low peer effects33on academic outcomes but stronger on social skills, such as network connectivity, and psychologicalmeasures of social skills, such as altruism, which vary by gender.Given the coethnic bias reported in the literature, it is likely that friendships are formedalong ethnic lines. For example, using a setting in SSA similar to Uganda, Salmon-Letelier (2022)report that friendship networks form along ethnicity or religion lines in Nigeria’s state and federalunity schools, respectively. If such homophily exists, it may create differential impacts by gendersince friendship groups overlap with study groups.There is also long-standing anthropological literature, such as de la Cadena (1995) exploringethnicity and gender that finds women are more ethnic than men in the community of Cusco.Studying how information affects homophily, Gallen and Wasserman (2023), finds that womenportray homophile tendencies more than men in an online college mentoring platform. Also, Jacksonet al. (2022) track university students’ friendships and study partnerships in their Caltech Cohortstudy and find assortative homophily by gender and ethnicity exists and persists substantially overtime among friendship and study groups.Differential Effects by Ethnic SalienceHaving a coethnic in a peer group might be useful for students with high ethnic salience due tomigrating from their home regions to attend university and experiencing a “diversity shock" whenthey arrive at the campus. Migrating from one’s ethnic region to attend a university located in adifferent ethnic region with different cultures could cause immigrant students to be aware of theirown ethnic identity, leading to greater attachment to their own ethnicities. This is the phenomenonin Okunogbe (2018), who finds greater ethnic pride among Nigerian youth randomly assigned toserve in a region where the ethnic majority is different from their own ethnicity. These hypothesesalso align with the psychology literature on social identity, which suggests that the salience of one’sethnic identity increases when one migrates away from one’s native region.In addition, such students could face social isolation as they encounter cultural barriers,which may increase their stress levels and contribute to a lack of sense of belonging. Moreover,students from certain ethnicities may experience discrimination from other groups, leading themto isolate themselves.7 These students are forced to navigate a new learning environment where7Another potential reason for the isolation of certain ethnicities is inter-ethnic conflicts and competition spilling34classrooms are more diverse than their high schools. Yet, several studies report generally lowtrust levels in addition to high in-group bias in highly ethnically diverse societies. Having a highproportion of coethnic peers in their peer group can be beneficial for students with high ethnicsalience, especially if they belong to small ethnic groups.Differential Effects by Degree TypeStudies on post-secondary education have reported differences by subject type. For example, Carrell,Fullerton, and West (2009) find peer effects are stronger in math and science courses, smaller insocial sciences, and absent in foreign languages and physical education at the Air Force Academy.Studying peer effects from the field of study at an Italian university, Brunello, De Paola, andScoppa (2010) find that peer effects are stronger in the ‘hard’ sciences (engineering, math, andnatural sciences) but absent in social sciences and humanities. I do not observe course names, butI observe the course code (e.g., STA101) and the degree type. MUK offers degrees in either arts orsciences. Arts degree comprises a wide range of degrees, such as business-related, social sciences,and humanities, and so do science degrees.There are other reasons in this context to anticipate why peer effects might differ by degreetype. For example, classrooms in arts degrees may differ from those in science classes, as they are,on average, larger. Additionally, the proportion of high-ability peers in arts degrees is smaller dueto the design of the national merit scholarship scheme. Generally, larger class sizes would reduceinteraction with the professor by increasing the student-to-teacher ratio. Since student-teacherinteractions outside the classroom are limited in this setting, class size effects are more likely tomanifest through peer effects. Also, large classrooms might increase the need for a sense of belongingand may reduce intimate cross-cultural interactions among students. It is easier to sort based onethnicity as the probability of the coethnic presence of ethnicity is high.Differential Effects by AbilityHeterogeneous peer effects by a student’s own ability and peers’ average ability have been shown toexist in the literature. For example, Zimmerman (2003) finds students in the middle of the verbalSAT distribution have negative peer effects from low-ability roommates. Also, Carrell, Fullerton,into classrooms, although ethnic conflicts are not common in this setting.35Table 2.3: Mean Effects in Year One: Coethnic vs High-ability Share(1) (2) (3) (4)Coethnic share 1.054** 1.064** 1.046** 0.936**(0.47) (0.47) (0.47) (0.47)High-ability share 0.799*** 0.833*** 0.848*** 0.735**(0.29) (0.29) (0.29) (0.28)High-ability 3.790*** 3.792*** 3.791*** 3.787***(0.09) (0.09) (0.09) (0.09)R-squared 0.326 0.326 0.328 0.328N 321,452 321,452 321,452 321,452Dorm FE No Yes Yes YesIndividual Controls No No Yes YesGroup Controls No No No YesNotes: Data are from MUK and are restricted to students admitted to non-extension day majors from sixcolleges for 2009-2017, excluding 2010. A peer group comprises students admitted to majors within the sameschool assigned to the same dorm. Each column is an independent regression, but the outcome is the coursegrades in all the specifications. The differences between each specification are indicated at the bottom andcome from the controls. All regressions control for own ethnicity, gender, own ability, major FE, and HSsubject combination FE, but gender drops out (2)-(4) since dorms are single-sex. Individual controls includeage, religious indicators, and graduating from the district of origin. Group controls include the leave-me-outaverages of individual controls in addition to peer group size. SEs are parentheses and are clustered at thepeer group level.*p<0.1, **p<0.5, ***p<0.01and West (2009) find suggestive evidence of non-linearity peer effects. Verbal SAT peer effectsare strong for students in the bottom third of the distribution. Given reduced student-teacherinteraction in this setting, peer effects may exist through study partnership channels, especially forlow-ability students.2.6 Results2.6.1 Mean EffectsTable 2.3 estimates various specifications of equation (A1). All specifications control for ethnicity asdescribed in section 2.5, ability, gender, student’s major, and classroom and HS subject combinationfixed effects. The difference between specifications is shown at the bottom of each column. It comesfrom controls in each regression, as I begin with a simple regression and progressively add morecontrols. Since I do not know the residence status of the students, the coefficients reported in this36should be interpreted as intent to treat effects.Given no evidence of selection, as reported in Section 2.5, we do not expect the coefficientsto change significantly as we move from column (1) to (4). The table shows that the share ofcoethnic and high-ability matters significantly for academic performance. The effect of coethnicshare is stable at around one percentage point (pp) , while that of high-ability peers is around0.8pp. Adding dorm fixed effects and individual and group controls does not alter the effects.The results in specification From specification (4) imply that adding five coethnic peersto a typical peer group of size 25 (corresponding to the average group size) increases a student’sperformance by 0.19pp (5/25 × 0.936). This effect is equivalent to 0.02 standard deviations in astudent’s performance in the first year. The effect of high-ability share is 0.735, which implies thatadding two more high-ability peers to a typical group of size 25 increases a student’s performance by0.15pp (5/25 ×.735), which is also about 0.02 standard deviations change in a student’s performance.For context, adding five coethnic peers to a typical peer group of size 25 (correspondingaverage size as Table 2.1 shows) is equivalent to moving from the group where the number ofcoethnic peers corresponds to the twenty-fifth percentile to a group where the number of coethnicpeers corresponds to the seventy-fifth percentile.8 For simplicity, I will interpret the results as theeffect of adding either five coethnic or high-ability peers to a group of 25.Table 2.3 also shows the effect of own ability is much larger than the effect of coethnic andhigh-ability share. The table shows that high-ability students perform about four percentage pointshigher than low-ability peers. This difference is large as it corresponds to a change in grade thatwould move a student whose first-year grade is equal to the average from the second-class lower(Fairly Good) performance range to a second-upper (Good) range. The average of grades reportedin Table 2.1 is equivalent to second class lower degree type in this setting.Results show positive and direct peer effects from a higher share of high-ability (irrespectiveof ethnicity) and coethnic peers. However, it is likely that the high ability of coethnic peers mightmatter, while high-ability noncoethnic peers do not. To test this, I break down the treatmentvariables into four: (A) high-ability coethnic peers, (B) low-ability coethnic peers, (C) high-abilitynoncoethnic peers, and (D) low-ability noncoethnic peers and compute the share of each as described8It is important to note that distribution of high-ability peers may be different from that of coethnic peers in agroup. I use a marginal change of 5 peers in a typical group size for simplicity of interpretation.37Table 2.4: The Effect of High-ability Coethnic and High-ability Noncoethnic peers on AcademicPerformance in Year One.(1) (2) (3) (4)(A) High-ability coethnic share 0.928* 0.980* 1.053** 0.865*(0.52) (0.52) (0.51) (0.50)(B) Low-ability coethnic share 0.686* 0.692* 0.696* 0.648(0.41) (0.41) (0.41) (0.41)(C) High-ability noncoethnic share 0.963*** 0.992*** 0.991*** 0.875***(0.32) (0.32) (0.32) (0.32)R-squared 0.326 0.326 0.328 0.328N 321,375 321,375 321,375 321,375Dorm FE No Yes Yes YesIndividual controls No No Yes YesGroup controls No No No YesNotes: Data are from MUK and are restricted to students admitted to non-extension day majors from sixcolleges for 2009-2017, excluding 2010. A peer group comprises students admitted to majors within the sameschool, and assigned to the same dorm. Each column is an independent regression, but the outcome is thecourse grades in all the specifications. The differences between each specification are indicated at the bottomand come from the controls. All regressions control for own ethnicity, gender, own ability, major FE, and HSsubject combination FE, but gender drops out (2)-(4) since dorms are single-sex. Individual controls includeage, religious indicators, and graduating from the district of origin. Group controls include the leave-me-outaverages of individual controls in addition to peer group size. SEs are parentheses and are clustered at thepeer group level.* p<0.10, ** p<0.05, *** p<0.01in Section 2.5. If high-ability coethnic peers matter while high-ability noncoethnic peers do not,(A) should be positive and significant while (C) should not, or at least (A) should be larger than(C), indicating that coethnic peers matter indirectly through high-ability peers.These results in Table 2.4 largely follow the pattern observed in Table 2.3. When runningthese regressions, I exclude category (D). Therefore, the reported coefficients should be interpretedrelative to that reference group. From the preferred specification (4), adding five high-ability coeth-nic or noncoethnic peers to a group of size 25 increases a student’s course grade by 0.18pp relativeto low-ability noncoethnic peers, which is equivalent to 0.02 standard deviations. The same tablealso shows that low-ability coethnic peers have a large and positive effect on academic performance.However, it is imprecise.Lastly, estimates in both Table 2.4 and Table 2.3 don’t change or change very little when Iadd dorm FE, individual controls, and group controls. This is consistent with exogenous assignmentinto peer groups and that correlated shocks are less likely to drive results reported in this paper. The38results do not suggest that high-ability coethnic peers matter more than high-ability noncoethnicpeers in the first year, which is contrary to my prior. Taken together, these results show that, onaverage, high-ability peers directly and significantly impact every student’s grades regardless of theirethnicity. Also, these results show that coethnic peers have a positive effect on grades, althoughthey suggest high-ability coethnic matter more than low-ability coethnic peers.2.6.2 Persistence of Mean EffectsAll the results presented thus far focus on student performance during their first year at MUK. Byextending the analysis to the subsequent years of their education, I test for the persistence of thesepeer effects. If peer effects from social networks persist as a student advances throughout theircollege career, then the effects of high-ability and coethnic peers observed in Section 2.6.1 shouldalso be evident in the follow-on years. Since selection into courses is limited, this setting allowsme to explore the persistence of peer effects. I estimate specification (A2) separately for each ofthe three academic years of undergraduate education at MUK and present the results in Table 2.5.Panel A compares the effect of coethnic share to that of high-ability share, while Panel B compareshigh-ability coethnic to noncoethnic peers.Column (1) restates the results in Section 2.6.1 to facilitate comparison. The effect ofcoethnic share persists into the second year, but it is almost half of the magnitude of the first year’seffect in the third year. That is, the effect of adding five coethnic peers into a peer group of size25 is 0.10pp in the third year, which is not statistically different from zero and is almost half of theeffect observed in the first year, as reported in Table 2.3. In comparison, the effect of high-abilityshare persists and even increases in the third year. From Panel A, adding two high-ability peers toa group of 25 increases a student’s performance by 0.22pp in the third year Yet, the same changeincreases student performance by 0.15pp in the first year. Thus, the effect of high-ability peers inthe third year is 1.5 times the effect observed in the first year, and that of coethnic peers is one-halfof what is observed in the first year.Panel B shows results similar to those in Panel A. Relative to low-ability noncoethnic peers,the effect of high-ability coethnic and noncoethnic in the third year is positive and significant. Incontrast, that of low-ability coethnic peers in the third year is not significant and is about 25% lowerthan the effect observed in the first year. Additionally, the effects of high-ability peers (coethnic39and noncoethnic) relative to low-ability peers increase from the first to the third year.The results in Table 2.5 show that the role of shared identity, if not coupled with ability,falls as time goes on. However, the effect of high-ability peers rises as time goes on, although theeffect of high-ability coethnic peers increases more than that of high-ability noncoethnic peers. Theresults are suggestive of evolving study groups or social networks.9 For example, students mightform stronger study bonds with high-ability peers as time goes on.2.6.3 Heterogeneous Peer EffectsResults in Section 2.6.1 show that, on average, going to school and potentially living with coethnicand high ability affects academic performance, but the effect of the coethnic share falls as time goeson. I now turn to see whether there are differential impacts on different dimensions mentioned inSection 2.5.2, as such differential effects might shed light on some mechanisms.Differential Impacts by GenderTable 2.6 presents the differential effects by gender. As dorms are single-sex, I control for genderinstead of dorm FE. Controlling for dorm FE does not change the results, but gender drops out. Thetable also represents results across the years and p-values corresponding to testing the significanceof the treatment variables for girls: coethnic share (ϕ1 + φ1) and high-ability share (ϕ2 + φ2) inequation (2.6). These results reveal several patterns.First, girls perform lower than boys by 0.743pp, significant in the first year, but they performbetter than boys by 1.061pp in the third year. The difference in performance between boys andgirls is not significant in the second year.Second, the coethnic share is positive but not significant for both boys in the first andthird years, but it is marginally significant in the second year. Although economically meaningful,the differential impact by gender is not significant across years. These results suggest that, unlikeboys, girls might be benefiting from a higher coethnic share. The p-value testing the sum of the9I will explore this mechanism when I run surveys later. It is possible that students hang out with coethnic peersat the start of their university career because it is more organic. However, as time goes on, networks may evolve asthey learn which of their peers are high-ability (coethnic or noncoethnic). They might form stronger networks withhigh-ability coethnic peers, weaker networks with low-ability coethnic peers, and somewhat strong networks withhigh-ability noncoethnic peers, as high-ability peers may be perceived as more beneficial for academic performance.The survey will ask students about their study and friendship groups throughout their undergrad career to see if theyare constant or changing overtime.40Table 2.5: Persistence of Mean Effects in Follow-up YearsYear One Year Two Year ThreePanel A: Coethnic vs High-abilityCoethnic share 0.936** 1.136** 0.491(0.47) (0.48) (0.45)High-ability share 0.735** 0.839*** 1.101***(0.28) (0.28) (0.28)R-squared 0.328 0.384 0.380N 321,375 343,761 330,158Panel B: High-ability (Coethnic vs Noncoethnic)High-ability coethnic share 0.867* 1.201** 1.347***(0.50) (0.52) (0.50)Low-ability coethnic share 0.639 0.775* 0.476(0.41) (0.41) (0.39)High-ability noncoethnic share 0.879*** 0.951*** 1.166***(0.32) (0.32) (0.32)R-squared 0.326 0.383 0.378N 310,867 333,187 320,752Dorm FE Yes Yes YesIndividual controls Yes Yes YesGroup controls Yes Yes YesNotes: Data are from MUK and are restricted to students admitted to non-extension day majors from sixcolleges for 2009-2017, excluding 2010. A peer group comprises students admitted to majors within thesame school and assigned to the same dorm. Each column is an independent regression, but the outcome iscourse grades in all regressions. All regressions control for own ethnicity, own ability and major, HS subjectcombination, and classroom FE. Individual controls include age, religious indicators, and graduating fromthe district of origin. Group controls include the leave-me-out averages of individual controls in addition topeer group size. SEs are parentheses and are clustered at the peer group level.* p<0.10, ** p<0.05, *** p<0.01significance of coethnic share and the interaction of the coethnic share and female dummy is largelysignificant in the first and second years and marginally significant in the third year. These resultsimply that adding five coethnic peers in a group of 25 increases academic performance for boys by0.13pp, which is equivalent to a 0.01 standard deviation change in academic performance in thefirst year. On the contrary, the same in coethnic peers would increase girl’s performance by 0.25PP,equivalent to 0.03 standard deviation in the first year. Thus, the effect of coethnic share on boys isabout 30% lower than the average effect in the first year, yet the effect on girls is about 30% largerthan the average effect observed in Table 2.3.Third, the differential effect of high ability by gender is small, insignificant, and sometimes41Table 2.6: Differential Effect by Gender: Coethnic vs High-ability ShareYear One Year Two Year ThreeCoethnic share 0.630 0.928* 0.136(0.55) (0.56) (0.53)High-ability share 0.661** 0.821*** 1.101***(0.31) (0.31) (0.31)Female -0.743*** 0.243 1.061***(0.21) (0.21) (0.21)Coethnic share × Female 0.605 0.398 0.668(0.48) (0.49) (0.46)High-ability share × Female 0.136 -0.021 -0.074(0.42) (0.42) (0.41)p-val Coethnic share: Female 0.015 0.001 0.097p-val High-ability share: Female 0.059 0.053 0.013R-squared 0.328 0.384 0.380N 321,375 343,761 330,158Dorm FE Yes Yes YesIndividual controls Yes Yes YesGroup controls Yes Yes YesNotes: Data are from MUK and are restricted to students admitted to non-extension day majors from sixcolleges for 2009-2017, excluding 2010. A peer group comprises students admitted to majors within thesame school and assigned to the same dorm. Each column is an independent regression, but the outcome iscourse grades in all regressions. All regressions control for own ethnicity, major, HS subject combination,and classroom FE. Individual controls include age, religious indicators, and graduating from the district oforigin. Group controls include the leave-me-out averages of individual controls in addition to peer groupsize. SEs are parentheses and are clustered at the peer group level. The table also reports p-values forcoethnic share and high-ability share of female students. These tests correspond to ϕ1 + φ1 and ϕ2 + φ2 inequation (2.6).* p<0.10, ** p<0.05, *** p<0.01negative. Still, the share of high-ability peers has a positive and significant effect on boys and girls.As the average effect reported in Table 2.3, the share of high-ability persists and increases into thethird year for both boys and girls.Differential Impacts by AbilityTable 2.7 shows the differential effects by own ability across the years. The table also reports resultsacross the years and p-values corresponding to testing the significance of the treatment variables forhigh-ability: coethnic share (ϕ1+φ1) and high-ability share (ϕ2+φ2) in equation (2.6). High-ability42students perform higher than low-ability peers, as Section 2.6.1 already reported. The results revealseveral other patterns.First, the effect of coethnic share on the academic performance of low-ability students is notsignificant across all the years and is negative in the third year. On the other hand, the effect ofhigh-ability share on low-ability students is positive and significant across all the years and evenhigher in the third year than in the first year. For example, adding five high-ability peers to agroup of 25 increases a slow student’s performance by 0.14pp and 0.25pp in the first and third year,respectively. Although the differential impact by ability is not significant across all the years, itis negative and economically meaningful in the third year, implying that high-ability peers have alarger effect on low-ability students than high-ability students.Second, the differential impact of coethnic share is large and significant in Year One andlargely persists into Year Three. Table 2.7 shows adding five coethnic peers to a group of 25 increasesthe performance of high-ability students by 0.51pp, which is about 3.5 times the average effectreported in Table 2.3. This effect is about 0.05 standard deviation of the first year’s performance.This differential impact of coethnic share on high-ability students persists significantly into the thirdyear, albeit at a reduced magnitude.Differential Impacts by Degree TypeI estimate the differential effect by degree type and report it in Table 2.8. Since I control for theMajor FE, the arts major dummy drops out of the regressions. Although not shown, the resultschange significantly when I control for the degree type dummy instead of major FE changes. Thetable also reports results across the years and p-values corresponding to testing the significance ofthe treatment variables for arts majors: coethnic share (ϕ1 + φ1) and high-ability share (ϕ2 + φ2)in equation (2.6).This table shows the differential impact of high-ability share by degree type is large andsignificant across all the years and more than doubles from the first to third year. The resultsshow that adding five high-ability peers to a group of 25 increases a student in the art’s majorperformance by 0.33pp in the first year, which is almost 1.5 times the average effect reported inTable 2.3. Moreover, this effect increases to 0.362pp in the third year, which is 3.3 times the effectreported 2.5. The effect of high-ability share on a student who is a degree major in the third year43Table 2.7: Differential Effect by Ability Type: Coethnic vs High-ability ShareYear One Year Two Year ThreeCoethnic share 0.064 0.274 -0.126(0.48) (0.48) (0.46)High-ability share 0.716** 0.911*** 1.234***(0.31) (0.32) (0.30)High-ability 3.259*** 3.013*** 2.816***(0.20) (0.20) (0.19)Coethnic share × High-ability 2.467*** 2.513*** 1.846***(0.51) (0.51) (0.48)High-ability share × High-ability -0.015 -0.269 -0.429(0.42) (0.44) (0.41)p-val Coethnic share: High-ability 0.000 0.000 0.003p-val High-ability share: High-ability 0.082 0.115 0.046R-squared 0.328 0.384 0.380N 321,452 343,840 330,236Dorm FE Yes Yes YesIndividual controls Yes Yes YesGroup controls Yes Yes YesNotes: Data are from MUK and are restricted to students admitted to non-extension day majors from sixcolleges for 2009-2017, excluding 2010. A peer group comprises students admitted to majors within thesame school and assigned to the same dorm. Each column is an independent regression, but the outcome iscourse grades in all regressions. All regressions control for own ethnicity, own ability, and major, HS subjectcombination, and classroom FE in addition to individual and group controls. Individual controls includeage, religious indicators, and graduating from the district of origin. Group controls include the leave-me-outaverages of individual controls in addition to peer group size. SEs are parentheses and are clustered at thepeer group level. The table also reports p-values for coethnic share and high-ability share of high-ability.These tests correspond to ϕ1 + φ1 and ϕ2 + φ2 in equation (2.6).* p<0.10, ** p<0.05, *** p<0.01is very large, as it corresponds to 0.08 standard of academic year in the third year.Lastly, the table also shows the differential impact of coethnic share by degree type issignificant in the first year but not in the second and third years. Adding five coethnic peers to agroup of 25 increases the academic performance of a student in the arts degree by 0.22pp more thanfor a student in the science majors.44Table 2.8: Differential Effect by Degree Type: Coethnic vs High-ability ShareYear One Year Two Year ThreeCoethnic share 0.586 1.101** 0.650(0.51) (0.52) (0.49)High-ability share 0.455 0.384 0.034(0.36) (0.37) (0.36)Coethnic share × Arts degree 1.080** 0.188 -0.074(0.49) (0.49) (0.46)High-ability share × Arts degree 0.975* 1.321** 2.979***(0.59) (0.60) (0.57)p-val Coethnic share: Arts degree 0.006 0.017 0.283p-val High-ability share: Arts degree 0.000 0.000 0.000R-squared 0.328 0.384 0.380N 321,375 343,761 330,158Dorm FE Yes Yes YesIndividual controls Yes Yes YesGroup controls Yes Yes YesNotes: Data are from MUK and are restricted to students admitted to non-extension day majors from sixcolleges for 2009-2017, excluding 2010. A peer group comprises students admitted to majors within thesame school and assigned to the same dorm. Each column is an independent regression, but the outcomeis course grades in all regressions. All regressions control for own ethnicity, ability, major and HS subjectcombination, and classroom FE. Individual controls include age, religious indicators, and graduating fromthe district of origin. Group controls include the leave-me-out averages of individual controls in additionto peer group size. SEs are parentheses and are clustered at the peer group level. The table also reportsp-values for coethnic share and high-ability share of arts degree. These tests correspond to ϕ1 + φ1 andϕ2 + φ2 in equation (2.6).* p<0.10, ** p<0.05, *** p<0.01Differential Impacts by Ethnic SalienceI proxy high ‘ethnic salience’ using a dummy variable equal to one if a student graduated high schoolfrom a district of origin and zero otherwise.10 Since most Ugandan districts are ethnically segregated,these students have generally had much less exposure to other ethnicities prior to enrolling at MUKthan their peers of the same ethnicity who graduated high school outside their district of origin (e.g.,as a boarding student in or near Kampala). I estimate the differential impacts by ethnic salienceand present these results in Table 2.9. The table also presents these results across the years, whichshow interesting patterns and p-values corresponding to testing the significance of the treatment10I treat Kampala Metropolitan area, which includes Kampala and Wakiso as the one district as these two sharethe cities and there are clear boarders between these.45variables for arts majors: coethnic share (ϕ1+φ1) and high-ability share (ϕ2+φ2) in equation (2.6).First, students with assumed high ethnic salience perform significantly lower than thosewith low assumed ethnic salience. However, this negative difference reduces over time and is nolonger significant in the third year. That is, students of assumed high ethnic salience perform0.63pp, significant at 1% lower than those of assumed low-ethnic salience in the first year, but thecoefficient of this dummy increases to -0.17 and is no longer significant in the third year.Second, the effect of coethnic share on students with low assumed ethnic salience is notsignificant across the years. However, Table 2.9 shows that students of high ethnic salience typebenefit from a high share of coethnic peers in the first and second year. The differential effects ofcoethnic in the first, second, and third years are 2.088pp (significant), 1.323pp (significant), and0.515 (insignificant), respectively. The table also reports the p-values of treatment variables atthe bottom, which show that coethnic peers are important for students with high ethnic saliencein the first year and second year. These results imply that adding five coethnic peers to a groupof 25 increases academic performance by 0.42pp more for students assumed to be of high-ethnicsalience type than those assumed to be of low ethnic salience in the first year. That is, adding fivecoethnic peers to a group of 25 leaders leads to a 0.44pp increase in academic performance, which isequivalent to a 0.05 standard deviation in the first year and is 2.5 times the average effect reportedin Table 2.3.Third, although positive, the differential effect of high-ability share by ethnic salience is smalland insignificant. The share of high-ability peers has a positive and significant effect on studentsassumed to be of low ethnic and high ethnic salience in the first year, which persists and increasesfor both types of students in the third year.These results indicate that students who might suffer from cultural and diversity shock whenthey arrive at MUK to study benefit more from coethnic peers than high-ability peers. Nevertheless,the effect of coethnic share decreases from the first to the second and disappears by the time thestudent graduates.2.6.4 Robustness checksOne area of concern revolves around the potential impact of measurement error on estimates of thecoethnic share in a peer group. As aforementioned, I use the probable coethnic share in a peer46Table 2.9: Differential Effect by Ethnic Salience: Coethnic vs High-ability ShareYear One Year Two Year ThreeCoethnic share 0.132 0.613 0.285(0.51) (0.52) (0.49)High-ability share 0.716** 0.737** 1.044***(0.30) (0.31) (0.31)High ethnic salience -0.626*** -0.463*** -0.173(0.17) (0.17) (0.17)Coethnics Share × High ethnic salience 2.088*** 1.323*** 0.515(0.48) (0.48) (0.46)High-abillity Share × High ethnic salience 0.070 0.297 0.162(0.30) (0.33) (0.31)p-val Coethnic share: High ethnic salience 0.000 0.000 0.130p-val High-ability share: High ethnic salience 0.024 0.002 0.000R-squared 0.328 0.384 0.380N 321,375 343,761 330,158Dorm FE Yes Yes YesIndividual controls Yes Yes YesGroup controls Yes Yes YesNotes: Data are from MUK and are restricted to students admitted to non-extension day majors from sixcolleges for 2009-2017, excluding 2010. A peer group comprises students admitted to majors within thesame school and assigned to the same dorm. Each column is an independent regression, but the outcomeis course grades in all regressions. All regressions control for own ethnicity, ability, and major, HS subjectcombination, and classroom FE. Individual controls include age, religious indicators, and graduating fromthe district of origin. Group controls include the leave-me-out averages of individual controls in additionto peer group size. SEs are parentheses and are clustered at the peer group level. The table also reportsp-values for coethnic share and high-ability share of students assumed to be of high-ethnic salience. Thesetests correspond to ϕ1 + φ1 and ϕ2 + φ2 in equation (2.6).* p<0.10, ** p<0.05, *** p<0.01group to account for this. Nevertheless, I get practically similar results when I re-estimate theresults using a single ethnicity corresponding to the category the algorithm is most confident about.I discuss robustness in relation to using student-level aggregated data and, thus, a different set offixed effects in this section.GPA as the Dependent VariableI re-estimate the average effects at the student level, using GPA as the outcome (not course-levelgrades), and report these results in Table 2.10. Panel A compares the ethnic and high-ability shares47within a student’s peer group. In contrast, Panel B compares the effect of higher-ability coethnicpeers to high-ability noncoethnic peers relative to low-ability noncoethnics. In addition, (1) isthe same as the main effects regressions reported in Section 2.6.1 and is included for comparisonpurposes. The results obtained using GPA as the outcome are similar to those reported in Section2.6.1. Naturally, the magnitudes of coefficients are different since the outcome variables are different.The results are consistent when I use school-by-year in place of classroom FE.From Panel A, the effect of high-ability and coethnic share is positive and significant whenthe outcome is GPA Similarly, from Panel B, high-ability coethnic and noncoethnic peers positivelyand significantly affect academic performance. However, panel B shows the effect of low-abilitycoethnic is precisely estimated when I use GPA. As in Table 2.4 of the main effects, using GPAas an outcome also suggests that high-ability coethnic matter as much as high-ability noncoethnicpeers although. Although the effect of high-ability coethnic peers is larger than that of high-abilitynoncoethnic peers in panel B column (2), the difference of 0.02 is not significantly different fromzero.2.6.5 Discussion and Contextualizing ResultsI find that the share of high-ability and coethnic peers positively and directly affects academicperformance, although the effect of the latter does not persist. The results reported in Table 2.3suggest a mean peer effect size of 0.02 SD for both peer types. These are reduced-form effects basedon dorm assignment, which is likely to be an underestimate of the true peer effect (treatment effecton the treated). This effect is comparable to what zimmer2003 finds at Williams College as Figure2.5 shows. The effect Garlick (2018) finds at the University of Cape Town (UCT) using randomlyassigned residential peers assignment is larger than what find, although his reported confidenceintervals are large. The estimate is Garlick (2018) also reduced form effect because the authorobserves dorm assignments but not roommates. However, compliance is high in Garlick (2018) andis probably characterized by students who enroll at UCT from out of the city.11Interestingly, I find strong coethnic reduced-form peer effects—equal to 0.05 standard devi-ations, especially for students of assumed high ethnic salience, which is comparable to the average11The author mentions that people who do not live on campus in private residences, most likely with families (page348).48Table 2.10: Coethnic vs High-ability Share: Outcome as GPAOutcome variable: %course gradesOutcome variable:GPAPanel A: Coethnic vs High-abilityCoethnic share 0.936** 0.112**(0.47) (0.05)High-ability share 0.735*** 0.066**(0.28) (0.03)R-squared 0.328 0.277N 321,452 25,298Panel B: High-ability (Coethnic vs Noncoethnic)High-ability coethnic share 0.867*** 0.091*(0.51) (0.05)Low-ability coethnic share 0.639 0.076*(0.39) (0.04)High-ability Noncoethnic Share 0.879*** 0.081**(0.32) (0.03)R-squared 0.328 0.277N 321,375 25,298Classroom FE Yes NoSchool-by-year FE No YesDorm FE Yes YesIndividual controls Yes YesGroup controls Yes YesNotes: Data are from MUK and are restricted to students admitted to non-extension day majors from sixcolleges for 2009-2017, excluding 2010. A peer group comprises students admitted to majors within thesame school and assigned to the same dorm. Each column is an independent regression, but the outcome iscourse grades in all regressions. All regressions control for own ethnicity and dorm, major, and HS subjectcombination FE in addition to individual and group controls. Individual controls include age, religiousindicators, and graduating from the district of origin. Group controls include the leave-me-out averages ofindividual controls in addition to peer group size. SEs are parentheses and are clustered at the peer grouplevel.* p<0.10, ** p<0.05, *** p<0.01effect in Carrell, Fullerton, and West (2009) at the Air Force Academy. In short, in this setting withhigh ethnic diversity, I still find both ability and coethnic peer effects where high ethnic diversityis expected to dampen peer effects of higher ability. Ethnic diversity effects are unlikely to play arole at MUK.While channels behind peer effects literature, in general, are unclear, I hypothesize on themechanisms behind these results by discussing explanations for these results in this section based49Figure 2.5: Comparison to past PapersThis figure compares my average estimate and estimate of students of assumed high ethnic salience to past paperswith randomly assigned peers and significant average effects. Carrell, Fullerton, and West (2009) Table 3 column(6) page 452 reports a coefficient of 0.382 on peer SAT verbal, equivalent to a 0.05 increase in GPA. Additionally,Carrell, Fullerton, and West (2009) estimate is about 2.5 times that reported in Zimmerman (2003) Table 3 column(“First semester”) page 17. Garlick (2018) Table 4 column (1) reports a coefficient of 0.216 on the dormitory meanhigh school GPA, equivalent to a 0.04 SD. Lastly, the effects of coethnic and high-ability share in Table 2.3 areabout 0.02 SD increase in academic performance. Additionally, I report the peer effects on students of high ethnicsalience type, showing that coethnic peer effects are about 2.5 times the average effect for these types of students.on the mean and heterogeneous effects reported in the results section and the characteristics of thiscontext.12 The suggestive channels at play in this context that I discuss in this section includepeer-to-peer learning, friendships, and psychological and cultural reasons.Peer-to-Peer Learning and Study BehaviorTable 2.3 shows that one’s own ability positively and significantly affects academic performance, aneffect that persists into the final year of most majors. High-ability peers may influence the academic12I do not test mechanisms directly. I am yet to start collecting primary data through surveys, for which I havealready obtained IRB approval, including local IRB.50performance of their peers by facilitating peer-to-peer learning, such as leading discussion groups.This is especially important as office hours (professors or TAs) do not exist and because of theclassical style of lectures in this setting. Students in need of extra help might rely on high-abilitypeers for additional assistance.Students can identify their high-ability classmates through several methods, especially duringthe academic year advances. Firstly, student registration numbers differ by the enrollment scheme,such as merit scholarship status (high-ability). Secondly, it is common for newspapers to publish thenames of the top students in the country (those with a high chance of obtaining a merit scholarship)once the national exams are out. However, this usually occurs several months before students enrollat university, and newspapers are not delivered outside the largest cities. Lastly, it is typicalfor students’ course grades, especially in midterm marks, to be publicly posted on departmentnoticeboards. Consequently, it becomes easy to identify and seek assistance from high-ability peersin ways that can impact academic outcomes.Another potential explanation through which high-ability peers can influence others is byaffecting study efforts. Several studies utilizing time-use data have examined how a student’s studybehavior is influenced by the study behaviors of their peers (e.g., Mehta, Stinebrickner, and Stine-brickner 2019; Frijters, Islam, and Pakrashi 2019). For instance, Mehta, Stinebrickner, and Stine-brickner (2019) show that students exhibit studious behavior if their peers, assigned randomly orconnected through organic friendships, invest a lot of time studying at college or did during highschool. It is conceivable that high-ability peers who have earned merit scholarships might haveachieved it due to investing a significant amount of time into studying during high school or are do-ing so while at MUK. This intensified study behavior among high-ability peers could have a positiveimpact on the study behaviors of their peers.Coethnic FriendshipsIncoming freshmen can easily identify coethnic peers through physical features and cultural char-acteristics, including names and language. Shared ethnicity friendships are likely more organic dueto shared identity since literature shows that coethnic bias exists in ethnically diverse societies. Forexample, Salmon-Letelier (2022) finds that ethnicity is important during friendship formation inNigeria’s state schools. Even studies outside the SSA report homophilous assortativity in student51interactions in study groups and friendships based on gender and ethnicity (Jackson et al. 2022).Therefore, students might form ethnic-based friendships within randomly assigned groups explain-ing the suggestive evidence on why high-ability coethnic students might matter more for academicsuccess as Table 2.4. This aligns with the Berea college freshman time-use (Mehta, Stinebrickner,and Stinebrickner 2019), which finds that using friends as peers is a stronger predictor of a student’spropensity to study.Nevertheless, the same table reports that high-ability noncoethnic peers also positively andsignificantly affect academic performance. Students may likely seek high-ability peers for academichelp, irrespective of ethnicity. Thus, having high-ability coethnic peers is an added advantagebecause students may sort into friendships or study groups based on ethnicity when unaware ofwhich of their peers are high-ability.Cultural and Psychological ReasonsLastly, these results also suggest cultural and psychological explanations at play. Many studentsmigrating from rural districts might feel isolated as they navigate a diverse environment as they nolonger belong to an ethnic majority. This could hamstring a sense of belonging for such students,which could have a negative effect on academic performance. As Table 2.9 shows, students of highethnic salience perform lower than those of low ethnic salience in the first year.In this case, a higher share of coethnic peers might be perceived equally or even more impor-tant compared to the share of high-ability peers by students experiencing a diversity shock. Thismechanism might explain why I find coethnic peer effects in Table 2.3 are positive and significant,and even stronger in column one of Table 2.9 where the interaction of coethnic share and graduatingHS from the district of origin is positive and significant.If this mechanism is at play, this interaction should be even stronger for small groups (ex-cluding Banyankore/Kiga and Baganda groups, which are the largest two groups that makeup 65%of the student population) as the smallest groups tend even to be more segregated as Figure 2.1shows. Appendix Table A3 shows that the interaction is larger for smaller ethnic groups.Nevertheless, this interaction could capture the influence of cultural shocks stemming fromdifferences in diversity in the learning environment and between life in the city and rural areas.Beyond navigating diverse classrooms, migrated students encounter an urban lifestyle distinct from52their rural upbringing. Additionally, it’s conceivable that students who graduate from a high schoolwithin their district of origin might have predominantly resided at home, even though most Ugandanhigh schools offer boarding facilities. Such students might struggle to build a support network withpeers, especially noncoethnic ones.Table 2.9 also shows that the interaction’s magnitude reduces from Year One to Year Three,and so does the mean effect of the coethnic share in Table 2.3. This pattern in the coefficientsindicates that the importance of coethnic peers to students of assumed high ethnic salience goesaway by the time they graduate. It is possible that cross-ethnic friendships emerge as these typesof students acquaint themselves with peers through frequent interactions, making the coethnicshare less important. The contact hypothesis, first introduced by William (1947), can explain thisphenomenon.Also, as students learn more about peers, ethnicity-based networks become less importantcompared to assortative matching based on attributes such as ability and study behaviors thatmatter more for academic success at college. This evolution of networks and information gainmight explain why the effect of ability share increases with time.2.7 ConclusionEthnic diversity has widespread and measurable impacts on a host of social, political and economicoutcomes. In Sub-Saharan Africa, latent ethnic tension can deteriorate social trust and reinforcehigh coethnic favoritism. In the context of higher education, which brings students into closecontact with ethnic diversity – often for the first time – ethnic heterogeneity may hamstring studentcollaboration and undermine academic performance with long-run implications. This paper providescausal estimates of peer effects on performance in the unique setting of higher education in Uganda,one of the region’s most ethnically diverse and segregated countries.I define a student’s peer group as students admitted to majors within the same school inthe same year who are assigned to the same dorm. This allows me to study the effects of peers withwhom a student is likely to interact during school and non-school activities. Dorm assignments arerandom conditional on gender after a student is admitted, and courses are pre-determined at thetime of admission before a student enrolls, providing an exogenous variation across peer groups.53I find that coethnic peers (irrespective of ability) and high-ability peers (irrespective of ethnicity)have a positive and significant effect on grades in the first year. However, the mean effect of coethnicpeers does not persist until a student graduates.These mean results mask significant heterogeneity in coethnic peer effects. First, I findstrong and positive coethnic peer effects for students of high ethnic salience that do not persist untila student graduates. These are students who graduated from secondary schools in their districts ofbirth and have relatively limited exposure to ethnicities different from their own prior to arrival atcampus. I also find a strong positive coethnic peer effect for high-ability students, not low-abilitystudents, that persists. This suggests that the benefits of coethnic peers can be reaped by thosewho have the capacity to succeed academically. The results also suggest coethnic peers have a largerpositive impact on girls than boys.These results have a number of implications for higher education policy and administrationin Uganda and, perhaps, in comparable settings with high ethnic diversity. First, the positiveimpact of high-ability peers on academic performance underscores the importance of fostering anenvironment that encourages peer-to-peer learning. For example, universities could implementoptimal peer group assignments where low-ability students are mentored by high-ability students.Second, the positive effect of coethnic peers in the initial years on students assumed to be of highethnic salience suggests that there could be benefit of implementing programs that facilitate cross-cultural awareness, shared cultural events, and increase a sense of belonging. Given the existenceof ethnic student organizations in this setting, which suggests a degree of homophily that shapesstudent life, it is natural for incoming students of high ethnic salience to benefit from coethnicconnections and support.These results also suggest there might be a short-term cost to ethnic integration policies.For example, if a university peer group assignment algorithm breaks any homophily on ethnicityand enforces cross-ethnic mixing, it might have a negative effect on students who benefit from ahigher share of coethnic peers, especially those assumed to be of high ethnic salience.This paper points to several promising questions for future research. I find that a higher shareof high-ability coethnic and noncoethnic peers increases a student’s academic performance. At firstglance, these findings suggest that college students at MUK may portray less coethnic bias duringclassroom interactions, such as study group formations that have an effect on economic outcomes.54In such cases, the peer effects in this setting work through channels, such as study effort, as somestudies using colleges in the West (e.g., Stinebrickner and Stinebrickner 2006) report. However, thesefindings do not preclude other channels, such as coethnic cooperation and inter-ethnic competition.For example, high-ability coethnic peers might affect academic performance through cooperationwith peers of shared ethnicity, while noncoethnic high-ability might increase competition wherestudents of different ethnicities compete to the extent that increases academic performance.Additionally, I study the first order of ethnic diversity on academic performance by focusingon coethnicity within a peer group. This paper does not study higher-order effects, such as theethnic composition of noncoethnic peers, which is open for future research. For example, theremight be an optimal pairing, tripling, quadrupling, etc., of ethnicities that could be beneficial ordetrimental to academic performance. This kind of question requires going beyond studying theeffect of ethnic diversity that would regress a Herfindahl index computed from ethnic shares withina student peer group on academic performance.Lastly, this paper investigates short-term high-ability and coethnic peer effects by focusingon academic outcomes and finds that high-ability peers (irrespective of ethnicity) affect academicperformance. However, it is unclear if a similar pattern of findings exists in the long term. Studentsmay strategically engage during classroom interactions in a way that does not extend beyond theclassroom. For instance, students might strategically select into study groups with higher-abilitypeers irrespective of ethnicity when doing homework but select into coethnic friend groups whenforming non-education social networks. Cross-ethnic mixing at university may not change intereth-nic attitudes or social networks post-graduation if this happens. I focus these questions on theadditional work I have initiated using the same setting of this paper.55ReferencesAlesina, A., A. Devleeschauwer, W. Easterly, S. Kurlat, and R. Wacziarg. 2003. “Fractionalization.”Journal of Economic growth 8:155–194.Alesina, A., and E. La Ferrara. 2000. “Participation in Heterogeneous Communities*.” The QuarterlyJournal of Economics 115:847–904.Alesina, A., and E. Zhuravskaya. 2011. “Segregation and the Quality of Government in a CrossSection of Countries.” American Economic Review 101:1872–1911.Barone, G., and S. Mocetti. 2016. “Intergenerational mobility in the very long run: Florence 1427-2011.” Temi di discussione (Economic working papers) No. 1060, Bank of Italy, Economic Researchand International Relations Area, Apr.Bayer, A., S.P. Bhanot, E.T. Bronchetti, and S.A. O’Connell. 2020. “Diagnosing the Learning Envi-ronment for Diverse Students in Introductory Economics: An Analysis of Relevance, Belonging,and Growth Mindsets.” AEA Papers and Proceedings 110:294–98.Bhusal, B., M. Callen, R. Gulzar, Saad;Pande, S.A. Prillaman, and D. Singhania. 2020. “Does Rev-olution Work? Evidence from Nepal’s People’s War.” EGA Working Paper Series No. WPS-116.Center for Effective Global Action. Unversity Of California, Berkeley, pp. .Bloom, H.S. 1984. “Accounting for No-Shows in Experimental Evaluation Designs.” EvaluationReview 8:225–246.Bramoullé, Y., H. Djebbari, and B. Fortin. 2009. “Identification of peer effects through social net-works.” Journal of Econometrics 150:41–55.Brunello, G., M. De Paola, and V. Scoppa. 2010. “PEER EFFECTS IN HIGHER EDUCATION:DOES THE FIELD OF STUDY MATTER?” Economic Inquiry 48:621–634.Carrell, S., R. Fullerton, and J. West. 2009. “Does Your Cohort Matter? Measuring Peer Effects inCollege Achievement.” Journal of Labor Economics 27:439–464.Carrell, S.E., M. Hoekstra, and E. Kuka. 2018. “The Long-Run Effects of Disruptive Peers.”American Economic Review 108:3377–3415.Carrell, S.E., and M.L. Hoekstra. 2010. “Externalities in the Classroom: How Children Exposedto Domestic Violence Affect Everyone’s Kids.” American Economic Journal: Applied Economics2:211–28.Carrell, S.E., F.V. Malmstrom, and J.E. West. 2008. “Peer Effects in Academic Cheating.” Journalof Human Resources 43:173–207.Cavnar, W.B., and J.M. Trenkle. 1994. “N-gram-based text categorization.”Clark, G., and N. Cummins. 2015. “Intergenerational Wealth Mobility in England, 1858–2012: Sur-names and Social Mobility.” The Economic Journal 125:61–85.Corno, L., E.L. Ferrara, and J. Burns. 2019. “Interaction, stereotypes and performance. Evidencefrom South Africa.” IFS Working Papers No. W19/03, Institute for Fiscal Studies, Jan.56De Giorgi, G., M. Pellizzari, and S. Redaelli. 2010. “Identification of Social Interactions throughPartially Overlapping Peer Groups.” American Economic Journal: Applied Economics 2:241–75.de la Cadena, M. 1995. Women Are More Indian: Ethnicity and Gender in a Community nearCuzco. Duke University Press.Depetris-Chauvin, E., and R. Durante. 2017. “One Team, One Nation: Football, Ethnic Identity,and Conflict in Africa.” Documentos de Trabajo No. 492, Instituto de Economia. Pontificia Uni-versidad Católica de Chile.Easterly, W., and R. Levine. 1997. “Africa’s Growth Tragedy: Policies and Ethnic Divisions.” TheQuarterly Journal of Economics 112:1203–1250.Eifert, B., E. Miguel, and D.N. Posner. 2010. “Political Competition and Ethnic Identification inAfrica.” American Journal of Political Science 54:494–510.Foster, G. 2006. “It’s not your peers, and it’s not your friends: Some progress toward understandingthe educational peer effect mechanism.” Journal of Public Economics 90:1455–1475.Frijters, P., A. Islam, and D. Pakrashi. 2019. “Heterogeneity in peer effects in random dormitoryassignment in a developing country.” Journal of Economic Behavior & Organization 163:117–134.Gallen, Y., and M. Wasserman. 2023. “Does information affect homophily?” Journal of PublicEconomics 222:104876.Garlick, R. 2018. “Academic Peer Effects with Different Group Assignment Policies: ResidentialTracking versus Random Assignment.” American Economic Journal: Applied Economics 10:345–69.Gisselquist, R.M., S. Leiderer, and M. Niño-Zarazúa. 2016. “Ethnic Heterogeneity and Public GoodsProvision in Zambia: Evidence of a Subnational “Diversity Dividend”.” World Development78:308–323.Green, E. 2008. “District Creation and Decentralisation in Uganda.”Habyarimana, J., M. Humphreys, D.N. Posner, and J.M. Weinstein. 2007. “Why does ethnic diversityundermine public goods provision?” American Political Science Review 101:709–725.Hjort, J. 2014. “ Ethnic Divisions and Production in Firms.” The Quarterly Journal of Economics129:1899–1946.Hooghe, M. 2007. “Social Capital and Diversity Generalized Trust, Social Cohesion and Regimes ofDiversity.” Canadian Journal of Political Science / Revue canadienne de science politique 40:709–732.Hoxby, C.M. 2002. “The Power of Peers: How Does the Makeup of a Classroom Influence Achieve-ment? (Research).” Education Next 2:57.—. 2000. “The Effects of Class Size on Student Achievement: New Evidence from PopulationVariation*.” The Quarterly Journal of Economics 115:1239–1285.Hudson, M.C., and C.L. Taylor. 1972. World handbook of political and social indicators. Yale57University Press New Haven, Conn.Håkansson, P., and F. Sjöholm. 2007. “Who Do You Trust? Ethnicity and Trust in Bosnia andHerzegovina.” Europe-Asia Studies 59:961–976.Jackson, M.O., S.M. Nei, E. Snowberg, and L. Yariv. 2022. “The Dynamics of Networks and Ho-mophily.” Working Paper No. 30815, National Bureau of Economic Research, December.Jones, K.S. 2004. “A statistical interpretation of term specificity and its application in retrieval.” J.Documentation 60:493–502.Luhn, H.P. 1957. “A Statistical Approach to Mechanized Encoding and Searching of Literary Infor-mation.” IBM Journal of Research and Development 1:309–317.Mamdani, M. 2001. When Victims Become Killers: Colonialism, Nativism, and the Genocide inRwanda. Princeton University Press.Manski, C.F. 1993. “Identification of Endogenous Social Effects: The Reflection Problem.” TheReview of Economic Studies 60:531–542.Mehta, N., R. Stinebrickner, and T. Stinebrickner. 2018. “Time-Use and Academic Peer Effects inCollege.” Working Paper No. 25168, National Bureau of Economic Research, October.—. 2019. “TIME-USE AND ACADEMIC PEER EFFECTS IN COLLEGE.” Economic Inquiry57:162–171.Miguel, E. 2004. “Tribe or Nation? Nation Building and Public Goods in Kenya versus Tanzania.”World Politics 56:327–362.Monasterio, L. 2017. “Surnames and ancestry in Brazil.” PLoS ONE 12.Morris, H.F. 1966. “The Uganda Constitution, April 1966.” Journal of African Law 10:112–117.NCHE. 2018. “THE STATE OF HIGHER EDUCATION AND TRAINING IN UGANDA 2018/19.”A report on higher education delivery and institutions, The National Council for Higher Educa-tion.Okunogbe, O. 2018. Does Exposure to Other Ethnic Regions Promote National Integration?:Evidence from Nigeria. The World Bank.Ricart-Huguet, J., and E.L. Paluck. 2023. “When the Sorting Hat Sorts Randomly: A NaturalExperiment on Culture.” Quarterly Journal of Political Science 18:39–73.Sacerdote, B. 2011. “Chapter 4 - Peer Effects in Education: How Might They Work, How Big AreThey and How Much Do We Know Thus Far?” Elsevier, vol. 3 of Handbook of the Economics ofEducation, pp. 249–277.—. 2001. “Peer Effects with Random Assignment: Results for Dartmouth Roommates*.” TheQuarterly Journal of Economics 116:681–704.Salmon-Letelier, M. 2022. “Friendship Patterns in Diverse Nigerian Unity Schools.” ComparativeEducation Review 66:709–732.58Sen, M., and O. Wasow. 2016. “Race as a Bundle of Sticks: Designs that Estimate Effects ofSeemingly Immutable Characteristics.” Annual Review of Political Science 19:499–522.Ssentongo, J.S. 2016. “’THE DISTRICT BELONGS TO THE SONS OF THE SOIL’: DECEN-TRALISATION AND THE ENTRENCHMENT OF ETHNIC EXCLUSION IN UGANDA.”Identity, Culture and Politics: An Afro-Asian Dialogue 17:60–96.Stinebrickner, R., and T.R. Stinebrickner. 2006. “What can be learned about peer effects using col-lege roommates? Evidence from new survey data and students from disadvantaged backgrounds.”Journal of Public Economics 90:1435–1454.Tajfel, H. 1982. “Social Psychology of Intergroup Relations.” Annual Review of Psychology 33:1–39.Tornberg, H. 2013. “Ethnic fragmentation and political Instability in post-colonial Uganda: un-derstanding the Contribution of colonial rule to the plights of the Acholi people in NorthernUganda.”Turyahikayo, B. 1976. Transafrican Journal of History 5:194–200.UBOS. 2006. “2002 UGANDA POPULATION AND HOUSING CENSUS.” Analytical report,Uganda Bureau of Statistics, October.Uganda Bureau of Statistics. 2016. “The National Population and Housing Census 2014 – MainReport.” Working paper, Kampala, Uganda.Walton, G.M., and G.L. Cohen. 2011. “A Brief Social-Belonging Intervention Improves Academicand Health Outcomes of Minority Students.” Science 331:1447–1451.William, J., Robin M. 1947. The Reduction of Intergroup Tensions: A Survey of Research onProblems of Ethnic, Racial, and Religious Group Relations. New York: Social Science ResearchCouncil.Zimmerman, D.J. 2003. “Peer Effects in Academic Outcomes: Evidence from a Natural Experiment.”The Review of Economics and Statistics 85:9–23.Zárate, R.A. 2023. “Uncovering Peer Effects in Social and Academic Skills.” American EconomicJournal: Applied Economics 15:35–79.592.8 Appendix2.8.1 Data AppendixThis section provides details that are not highlighted in the main data Section 4 of this paper.Linking Student DataMUK stores data on students’ applications, admissions, and results in separate databases and offices.There is no unique identifier that can link databases in some cases.STEP I: Computing GPA. My data cleaning process starts from the results database. Thesedata list courses and course units (for some), and exam scores in percentages by program, depart-ment, semester, and year of study. They also list the calendar year when the exam was taken.These data cover 2008-2017. However, to match the admissions sample, I restrict the results sampleto 2009-2017 years. I convert the exam scores from the percentage scale to letter grades using theinformation on the back of the transcripts and available in the code book. I then compute GPAsby semester and year.STEP II: Merging with admissions. Each admitted student has two unique identifiers: studentnumber and registration number. I use the latter to merge results and admissions data. There is a93.9% merge rate at this stage.Step III: Determining cohorts. The admissions and graduation programs are coded differentlyin many cases. The undergrad (graduation) program may admit students through different cohorts(e.g., evening and day classes). Take the graduation program “Bachelor of Science in Computer Sci-ence", for example; it is coded as “BCSCS". However, BCSCS students may be admitted to throughtwo cohorts: day classes(“CSC") or evening classes (“CSE"). This distinction was necessary becausethe cohort forms one’s peer. I use the university codebook to ensure the admission, enrollment, andgraduation programs are consistent. Since I restrict the sample to day majors, CSC appears in myfinal sample, while CSE does not.Step IV: Merging with the name data. After correcting obvious misspellings in the names, Imerged these data with data that predicted ethnicities. Merging on names in the training data givesa merge rate of 98.6%. Merging features produced by ML classification is irrelevant since ethnic60predictions can be made for every surname.Lastly, I deleted all the 2010 observations because the hall assignment is unavailable formany students admitted through a private scheme. The university officials in the admissions officementioned that there was a problem/data bleach with the information system in 2010, where theuniversity lost a lot of records.2.8.2 Ethnic and Geographic BoundariesThe Ugandan parliament’s gate has engravings of symbols and names of 15 administrative units atthe time of independence from Britain. The administrative units were federal states, districts, orTerritories (The Constitution of Uganda, 1964). The federal states were historical kingdoms, whichincluded Ankole, Buganda, Bunyoro, and Toro, and the territory of Busoga. The districts includedAcholi, Bugisu, Bukedi, Karamoja, Kigezi, Lango, Madi, Sebei, Teso, and West Nile. Coincidentally,these kingdoms and districts’ boundaries followed ethnic/tribe boundaries that existed before theBritish colonial government but were exacerbated by British colonists.However, the colonial government introduced a notion of a district as an administrative unit,which initially was a way to group similar ethnicities in geographical proximity. Kingdoms werehistorically centralized and ethnically segregated, with a traditional king as a ruler. However, thiswas different for districts. Some districts, such as Sebei and Bugisu, were ethnically segregatedbut followed a different system of local political leadership, such as clans or chiefdoms. There werealso districts (e.g., West Nile and Bukedi) that were a cluster of several, and sometimes unrelated,relatively small ethnic groups. For example, the West Nile comprised mostly Lugbara people butincluded smaller ethnic groups, such as the Alur and Kakwa.President Obote abolished kingdoms in 1966 for political reasons and changed the status offederal states to districts, and split the formerly powerful federal state (kingdom) of Buganda intofour districts (Morris 1966).13 Since then, the number of districts has increased to 135 over theyears, with the highest increase happening under the current government for reasons such as servicedelivery and ethnolinguistic conflict management, among others. Some studies report politicalreasons as the most prominent explanations for new district creation (Green 2008).13District is the second-largest unit of administration after the federal government. The districts divide intocounties. Counties divide into sub-counties. Sub-counties divide into parishes/villages, which divide further intocells/villages.61Most importantly, new districts are curved out of existing districts at the time of creation.It has been rare to create a district by carving out counties that initially belonged to two separatedistricts over the years. Interestingly, albeit unsurprising, new districts tend to be more segregatedby ethnicity (Ssentongo 2016). For example, the population of Nebbi is 96.2% of Alur ethnicity,although it was carved out of the West Nile district in 1974, which mainly comprised the Lugbarapeople. The creation of new districts sometimes begins with smaller ethnicities wanting to breakaway from the majority ethnicity in the original bigger district for reasons such as autonomy andbringing resources closer to them. But also, the government will offer a county a district status forpolitical support.I can trace current administrative units to historical kingdoms using publicly available dataon administrative units from the Ugandan Ministry of Local Government. I complement the publicdata with data from the 2014 census from UBOS. The Census data contain the population break-down by ethnicity for each district, confirming ethnicity within each district. That is, the censusreports the number of each 66 ethnicities that reside in each district (i.e., 136 X 66 observations). Icompute the proportion of each ethnicity in a district and rank these proportions from the highestto the lowest.The top-ranked ethnicity informs the ethnic region that the district belongs to. The averageproportion of the top ethnicity by population is 0.737 (the median is 0.813), indicating high ethnicsegregation within each district. These UBOS data help me confirm the historic ethnic regions andgive the final ethnic and geographic boundaries. I then create ethnic clusters by combining both thecurrent and historical administrative units to give final ethnic geographical borders. Using just the1962 districts and kingdoms that were created by the British colonial government would give wrongborders as the colonial sometimes bundled together ethnic groups that did not have centralizedgovernments, such as those found in the eastern parts of the countrySpecifically, when retracing the ethnic borders, the ethnicity with the highest proportionin a district based on UBOS data combined with historical settlement patterns supersedes thesegeographic boundaries established by the British colonial government. Additionally, This studyignores the smallest ethnicities within each district. Take Abim district, for example, the populationof Abim is 87% of Karimojong ethnicity and geographically belongs to the Karimojong subregion.Using both UBOS data and historic settlement, this study identifies Abim within Karimojong62borders when running the ML algorithm. However, Abim comprises other small minority groups,such as Gimara (0.033%). By ignoring ethnic groups that make up 13% of Abim’s population, Iam implicitly assuming that the smallest ethnicities are forced to assimilate with the largest ethnicgroups within that district, or they are immigrant groups.I use two formulas when allocating each district to the ethnic border (I) proportion of thehighest ethnicity in the district and (II) ethnic fractionalization index. The two methods shouldgive very similar borders. I use both for consistency. The ethnic fractionalization index introducedin Hudson and Taylor (1972) gives the probability that two randomly from a region (a district inthis setting) belong to two different ethnic groups. I.e.,(A1) FRACj =E∑e=1πje (1− πje) ,where j indexes a district, πje is the proportion of ethnic group e in district j. Using UBOS ethnicitybreakdown data by district, county, and sub-county, (I) and (II) are highly correlated (-0.981).Table A1: Ethnic fractionalization in a districtN mean sdEthnic fractionalization index 135 0.388 0.26Max proportion in a district 135 0.727 0.22From Table A1, the average proportion of the largest ethnicity in a district is 0.727, andthe median is even higher (median=0.802). This implies that it is rare to find districts with equalshares of ethnicities. The average probability that two individuals are randomly selected from adistrict is low, and the median is also lower (0.345).14 However, I compute this probability for thewhole country, and I get 0.933. This is the same value reported in Alesina et al. (2003). Therefore,although Uganda is ethnically diverse as a whole, its subnational units are not. When constructingthe training sample, I restrict districts where the ethnic fractionalization index is low (< 0.5), andthe max proportion in a district is 0.7 and above.Even though UBOS reports that Uganda has over 50 ethnic groups, 45 (68.2%) of the 6614This is based on 66 ethnic groups in the census data. When I use ethnic clusters/language groups from TableA2, this index falls to 0.23563ethnic groups reported in 2014 census data contribute to less than 1% of the population each, and 22ethnicities (33%) contribute a combined total of less than 1% of Uganda’s population. The smallestethnic groups are either non-Ugandan immigrant groups or indigenous groups. The immigrantethnicities may be scattered across the country or segregated in the refugee resettlement areas.15The indigenous groups are tiny in that even though they are segregated, they only make up a smallpart of the district population. This leaves 32 unique ethnicities (out of 66) based on district andethnicity clusters.Students do not report ethnicity or places of origin during the application stage but theirhome districts. Although I observe home districts for most students, using reported districts wouldignore cases of internal migration, especially rural-urban migration. Instead, I use students’ sur-names to predict their ethnicity as Ugandans’ last names are almost usually in their native language,as Section 2.4.2 highlights. I combine ethnicities whose languages have high lexical similarity andmutual intelligibility to create a language group to proxy ethnicity.Using language groups to proxy ethnicity has been used in several African studies to proxyethnicity (e.g., Eifert, Miguel, and Posner 2010; Depetris-Chauvin and Durante 2017) as lan-guage and ethnicity usually overlap. The similarity in languages implies similarities in cultures,facilitating the ease of interaction in ethnically heterogeneous societies. Moreover, although notalways, local languages in different follow a dialect continuum, which further informs my languagegroups/ethnicity. For example, historical and current Ankole and Kigezi people living in the SWpart speak the same language but with different accents and are therefore combined to form the“Banyankore/kiga" ethnic group. Another basis for combining two or more ethnicities is historical.For example, the Tooro kingdom (Batoro) was historically part of the Bunyoro kingdom (Banyoro)until the early 19th century (Turyahikayo 1976). Therefore, Batoro and Banyoro form one eth-nicity (language group). Combining groups that are mutually intelligible and similar reduces theethnic groups to 16 groups. Another concern for the performance of the classification algorithm ishow segregated ethnicities are. As Figure2.1 portrays, ethnicities within Uganda are geographicallysegregated.15UNCHR ranks Uganda as the fifth largest refugee host nation. See this link: accessed 4/14/2364Table A2: Ethnicity/Language Group CompositionEthnicity/language group Composition NumberAlur_Jonam Alur, Jonam 2SWBanyankore, Bakiga, Bafumbira,Banyaruguru, Banyarwanda, Batagwenda,Barundi, Bahororo8Ganda Baganda 1Gisu Bagisu and Babukus 2Iteso Iteso 1Jopadhola Jopadhola 1Kakwa Kakwa 1Kelenjin Pokot and Sabiny 2Karimojong Karamoja, Jie, Dodoth, Napore,Nyagia 5Madi Madi 1Northern Luo Acholi, Lango, Kumam, and Ethur 4Nyoro Batuku, Bunyoro, Batoro, Bagungu,Babwisi 5Rwenzori Bakonzo, Baamba 1Samia_nyole_gwe Banyole, Basmia, Bagwe 3Soga Basoga, Bagwere, Bakenyi 3West Nile Lugbara, Aringa 2Extremely smallVonoma(.008%), SoTopeth(.007%),Shana(.003%), Reli (.025%),Chope(.102%), Nube(.086%),Ngikutio(.017%), Mvuba(.009%),Mening(.008%), Lendu(.056%),Kuku(.140%), Kebuokebu(.161%),Bahehe(.012%), Gimar(.03%),Ikteuso(.041%), Batwa(.018%),Baruli(.565%), Banyabutumbi(.03%),Banyabindi(.049%), Aliba(.006%),Banyara(.142%), Nyangia(0.028%),Non-Ugandan(1.4%)24All 66Notes: Source is the Uganda population and housing census of 2014. Groupings were informed usingseveral sources as this section mentions.652.8.3 Deriving the Reduced-Form Peer EffectAs described in the main text, this paper estimate the reduced-form peer effect based on randomdorm assignment. In this section, I derive and discuss the relationship between this reduced-formestimate and the true underlying peer effect. Starting with equation (A1) and simplifying subscripts,we can write the individual specific effect of ‘actual’ high-ability share, S̃i on student i’s grade as(A2) Yi = ρXi + ϕS̃iG + eiwhere ϕ is the effect of the share of high-ability peers in a student’s peer group on her academicperformance. If I observed both random dorm assignment and actual (endogenous) dorm residence,it would be natural to use an IV approach to estimate the local average treatment effect of peerson academic performance, using dorm assignment to instrument for dorm residence as follows:S̃iG = κ10Xi + κ11SHiG + e1i(A3a)Yi = κ20Xi + κ21SHiG + e2i(A3b)where SHiG is the share of high-ability peers computed from peer groups as the result of the dormassignment as in equation (A1) that may not be equal to S̃iG because some students do not live indorms. Equation (A3a) as the first stage capturing the effect SHiG on S̃iG, while κ211 captures thereduced form of the high-ability share due to dorm assignment. Substituting equation (A3a) intoequations A2 will give:κ20 ≡ ρ+ κ10(A4a)κ21 ≡ ϕκ11(A4b)e2i ≡ ϕe1i + ei(A4c)Thus, the true high-ability peer effect (ϕ) is equal to κ21κ11 . That is, the IV estimate weightsthe reduced-form effect by the inverse of the first stage. Since I only observe dorm assignment, notresidence, I am unable to recover this structural peer effects coefficient, so estimates captured in66equation (A1) are reduced-form estimates of peer effects based on dorm assignment.672.8.4 Additional ResultsDifferential Impacts by Ethnic SalienceThe results presented in this section should be interpreted in conjunction with the effects in Section2.6.3. I proxy high ethnic salience as graduating high school from one’s district of origin. Asillustrated in Figure 2.1, non-majority groups are even more segregated and might consequentlyencounter greater diversity shock when they relocate to the capital for university education. This isespecially true since they are also the most underrepresented group at MUK. I present the differentialeffect by diversity shock in Table A3.More on Robustness checks68Table A3: Differential effect by Ethnic Salience (Nonmajority) Coethnic vs High-ability ShareYear One Year Two Year ThreeCoethnic share 0.813* 1.037** 0.422(0.47) (0.48) (0.46)High-ability share 0.605** 0.735** 1.000***(0.28) (0.29) (0.29)High ethnic salience (nonmajority) −1.394*** −1.043*** −0.644**(0.27) (0.27) (0.26)Coethnic share × high ethnic salience (nonmajority) 3.254* 2.243 1.286(1.71) (1.70) (1.66)High-ability share × high ethnic salience (nonmajority) 1.355** 1.095* 0.972*(0.56) (0.59) (0.52)p-val Coethnic share (nonmajority): high ethnic salience 0.016 0.054 0.298p-val High-ability share (nonmajority): high ethnic salience 0.026 0.084 0.173R-squared 0.328 0.384 0.380N 321,375 343,761 330,158Dorm FE Yes Yes YesIndividual controls Yes Yes YesGroup controls Yes Yes YesNotes: Data are from MUK and are restricted to students admitted to non-extension day majors from sixcolleges for 2009-2017, excluding 2010. A peer group comprises students admitted to majors within thesame school and assigned to the same dorm. Each column is an independent regression, but the outcome iscourse grades in all regressions. All regressions control for own ethnicity, own ability, and major, HS subjectcombination, and classroom FE. Individual controls include age, religious indicators, and graduating fromthe district of origin. Group controls include the leave-me-out averages of individual controls in additionto peer group size. SEs are parentheses and are clustered at the peer group level. Nonmajority ethnicitiesexclude the largest two groups (Banyankore/Kiga and Baganda). The table also reports p-values for coethnicshare and high-ability share of arts degree. These tests correspond to ϕ1+φ1 and ϕ2+φ2 in equation (2.6).* p<0.10, ** p<0.05, *** p<0.0169Table A4: More Evidence against SelectionCoethnic share High-ability shareAge 0.000 0.000(0.00) (0.00)Anglican -0.000 -0.002(0.00) (0.00)Catholic 0.001 -0.002(0.00) (0.00)Muslim 0.001 -0.003(0.00) (0.00)Seventh Day Adventist -0.003 -0.004(0.00) (0.01)Pentecostal -0.002 0.002(0.00) (0.00)High ethnic salience 0.001 0.002(0.00) (0.00)Other Religions 0.001 0.007(0.01) (0.01)High-ability 0.000 0.011***(0.00) (0.00)Peer group Size 0.000** -0.000(0.00) (0.00)R-squared 0.420 0.263N 25,323 25,323Joint Fstat 0.84 2.15Notes: Data are from MUK and are restricted to students admitted to non-extension day majors fromsix colleges for 2009-2017, excluding 2010. Each column is an independent regression that regresses eitherthe coethnic share or high-ability share on all pre-university characteristics. All regressions include school-by-year FE (not classroom), ethnicity, and dorm FE. SEs clustered at the peer group level.*p<0.1, **p<0.5, ***p<0.001702.8.5 List of FiguresFigure 2.6: Distribution of Peer Group Sizes.Notes: Data are from MUK and are restricted to students admitted to non-extension day majors from six collegesfor 2009-2017, excluding 2010. A peer group includes students admitted to majors within the same school andassigned to the same dorm.71Chapter 3Heterogeneity in Coethnic Peer Effects3.1 IntroductionUniversity administrators do not have control over endogenous social networks among students,although they can implement policies that directly influence students’ peers through dorm as-signments. For example, Makerere University’s dean changed the dorm assignment system fromalphabetical to random in the 1970s to avoid ethnic clustering within dorms (Ricart-Huguet andPaluck 2023). However, I find that a student’s grades increase with a higher share of coethnic peersin Chapter One. Anecdotal evidence also shows that room assignments within each dorm are doneto encourage interactions for academic success among students within the same major and school,although dorm assignments are random.1Thus, an optimization problem for university administrators should seek to maximize theexpected academic success and help promote national identity, not tribal identity, by the timestudents graduate and enter the job market. The latter is crucial in a setting like Uganda, whichis characterized by high ethnic segregation, and university campuses are the only opportunity forstudents migrating from disparate regions to interact with peers of different ethnicities before theyenter the workforce. Obtaining the optimal policy involves estimating the potential outcomes fordifferent treatments and using these estimates to inform policy decisions (Athey and Wager 2021).The average effect of coethnic peers reported in Essay One obscures important variations in1My discussion with one of the dorm custodians at Makerere revealed that they intentionally assign rooms afterreceiving a list of students randomly allocated to each dorm. One criterion they consider is the proximity of students’majors and departments, both in courses and physical location.72interethnic interactions and other heterogeneity that may be crucial for policy. For example, theaverage effect does not explore any higher-order composition peer effects that may be detrimentalor beneficial for student success. The Baganda and Banyankore communities, for instance, maybe unfriendly toward each other, as these two ethnic groups have long competed politically.2 It isnot surprising that Baganda’s presidential voting against the current regime has overwhelminglyincreased in the last elections. Baganda may feel victimized because of persistent electoral losses.On the other hand, the Baganda may feel superior to northern minority groups. Thus, interethnicdynamics may differ for the Baganda if they are in a peer group with a large number of students ofBanyankore descent compared to a group with a larger number of northern minorities.Additionally, Essay One shows that coethnic peers are beneficial for academic performance,but a policy planner might be interested in the optimal number of coethnic peers that is mostbeneficial without hurting non-coethnic peers. Such a planner might be interested in learning aboutthe non-linearities that might exist in the form of a dose-response function regarding the share ornumber of coethnic peers within a student’s peer group.To analyze the non-linearities and potential interethnic composition effects, I use casualforest estimation methods in Athey and Imbens (2016) and Athey and Wager (2019). Casualforest uses data-driven sample splits, reducing researcher bias in selecting the relevant heterogeneitydimensions. Additionally, the causal enable the capture of high-dimensional nonlinearities whileavoiding overfitting by employing both training and estimation samples (the “honest approach").3Essentially, I estimate Conditional Average Treatment Effects (CATE) for each individual under thismethod by feeding the causal forest algorithm an estimation formula similar to my main estimationregression (equation 4) of Essay One.My analysis reveals several findings. The predicted treatment effects are slightly nonlinear,with an increase in the coethnic share. The effect is negative and not significant when the coethnicshare is small (less than 0.3) but positive and significant when the ethnic share is above 0.35.Additionally, the predicted treatment effects differ substantially by ethnic group. Interestingly, the2In a recent mobilization tour in Buganda, the presidential candidate, Bobi Wine, seemed to subtly galvanize theBaganda to resist the Banyankore’s occupation of the Baganda’s ancestral land. The Banyankore have controlledUganda’s government since 1986, and Uganda’s capital is geographically located in the Buganda kingdom. Conse-quently, the Banyankore have acquired extensive property on land belonging to the ancestors of most Baganda andbuilt institutions to prolong their stay in political power, which could scar intergroup relations between these twogroups.3Moreover, treatment effect estimates using honesty fitting are asymptotically normal (Athey and Imbens 2016).73effect is largest for the largest ethnic group (Baganda), where 100% of their predicted individualtreatment effects are to the right of the average treatment effect (ATE). Although positive, thepredicted treatment effects are smallest for the second largest group, Banyankole, for which I maybe powered to detect the treatment effects.Additionally, the predicted treatment effects reveal significant gender differences across eth-nicities. These differences are minimal among the Banyankore (the second-largest ethnic group) andmost pronounced among the Basoga (the third-largest group). Notably, gender effects are not uni-directional: female Basoga students benefit more from the treatment than their male counterparts,while the reverse is true for the Baganda. When analyzing interethnic effects using a causal forest,the results do not show much variation in the predicted treatment or support the hypothesizedinterethnic effects, possibly due to the model’s limited power in capturing heterogeneity related tointerethnic composition.In addition to the psychological channels discussed in Essay One, the heterogeneity analysisin this chapter reveals a high level of homophily. This effect is most pronounced among the largestethnic group, which likely has the strongest ethnic ties, and is lowest among the ethnic group thatcontrols the central government. Controlling the national government may cause the Banyankoreto identify more with being Ugandan than with their ethnic group.The rest of this chapter is organized as follows. Section 3.2 provides the empirical strategy,including explaining how I implement the causal forest algorithm. Section 3.3 gives the results,while I provide a discussion and conclusion in Section 3.4.3.2 Empirical estimationSince I am interested in analyzing the heterogeneity related to coethnic peer effects, I ignore high-ability peer effects for this section to reduce the dimensionality when estimating the CATE describedin Section 2.1 below. when I run the specification (4) in Essay One without controlling for high-ability peer effects, coethnic peer effects do not change since ethnic and high-ability share are notcorrelated because of random assignment. That is, my CATE estimation hinges on the followingequation.yijG = β0 + ϕ1SEiG + β2XiG + δj + ωf + λd + γs + εijG,74where yijcG is the GPA that student i of ethnicity j and belonging to peer group G obtained n thefirst year. Controls are similar to equation (4) of Essay One. SEiG is the probable coethnic share ofi’s peer group. The main estimation controls for δj , which is i’s most probable ethnic group, XiG isa vector of i’s background characteristics and includes i’s own ability. Additionally, ωf , λd, and γsrepresent school-by-year, dorm, and high school subject combination fixed effects (FE). Lastly, εijGis the error term. School refers to department or faculty as defined in Section 3.3 of Essay One.The coefficients of interest are ϕ1, which captures the effect of attending lectures and po-tentially living with coethnic peers in this setting. The identifying assumption is that conditionalon ethnicity, school, gender, and cohort, the coethnic share is independent of unobservable and astudent’s characteristics. That is,SEiG ⊥ (U,X) | Ethnicity, School, Gender, CohortI show why this identifying assumption is true in Appendix 3.5.1. For identifying SEiG, I only δj , ωfand λd (since dorms are single-sex). I control for subject HS FEs to improve precision.3.2.1 Estimating the CATEAs supervised machine learning techniques, casual forests predict heterogeneity in causal treatmenteffects to estimate CATE defined as, τ̂i = E[Y1i − Y0i | Xi = x], where Y1 and Y0 represent thepotential outcomes for the i-th individual when treated and untreated, respectively, and X is avector of observable characteristics. Causal forests as in the case of Athey and Imbens (2016)do not assume a specific functional form for the relationship between the outcome, treatment,and covariates, allowing for complex interactions and non-linearities. They can naturally handleheterogeneity in treatment effects, providing individual-specific estimates of the treatment effect,hence the subscript i.The splitting criteria in causal trees are designed to maximize the difference in treatmenteffects across the resulting subgroups (leaves) while ensuring accurate estimation within each leaf.This involves choosing splits that lead to significant differences in treatment effects rather than justimproving the fit of the outcome model, as in the case of random forests. Additionally, the algorithmestimates the individual treatment effects “honestly” and accurately. That is, in an “honest" causal75forest, the data is split into two sets: one used for constructing the tree (finding splits) and the otherfor estimating the treatment effects within the leaves. This separation helps to provide unbiasedestimates of treatment effects, as the estimation data was not used to determine the splits. As anadvanced technique, honest causal forests ensure multiple trees are built, each time using a differentsubset of data for splitting and estimation, as mentioned. The results are then averaged to producestable estimates. Honesty also eliminates any biasedness that would have resulted from overfitting.I use the GRF R package by Athey, Tibshirani, and Wager (2019) to implement causalforests. I do not begin by fitting regression forests to estimate the nuisance functions for theconditional mean outcome and the treatment propensity score as in Athey and Wager (2019).4 Iam essentially estimating equation (1), which is a version of my main estimation equation in EssayOne.In this approach, I provide the outcome variable (Year One GPA), treatment variable (coeth-nic share), and covariates, such as gender, and include the FEs required for identification as SectionAppendix 3.5.1 shows to the “causal forest" function, which internally handles the estimation ofnecessary nuisance parameters. The function then grows the causal forest by optimizing splits tomaximize treatment effect heterogeneity while maintaining accurate predictions, as aforementioned.Additionally, I exploit GRF’s tuning parameters, such as setting the minimum node size to optimizethe performance.Lastly, since my treatment variable is continuous, the causal forest will provide a partialeffect of coethnic share as in the case of Wooldridge (2010).5 In my case, GRF non-linearly andnon-parametrically uses a splitting criterion to maximize τ̂i = E[Cov(SE ,Y |X)Var(SE |X)], where X is a vectorof characteristics whose heterogeneity I am interested in and FEs.76Figure 3.1: Dose-Response FunctionCATE is estimated using causal forest algorithms in the GRF package. The plot also includes the CI of thepredicted CATE. Let τ̂i be individual i’s predicted CATE. Each τ̂i is predicted with a variance (σ2). Thus, the90% CI is given by τ̂i +/-, qnorm (0.9) * sqrt(σ2).3.3 Results3.3.1 Dose Response to Coethnic ShareFrom Figure (4) Essay One, we observe substantial variation in the treatment (coethnic share)by ethnic group. To visualize the relation between treatment and the predicted effects, I plotthe individual predicted CATE against the treatment in Figure 3.1, including the 90% confidenceinterval (CI) using the variance predicted for each CATE using the casual forest algorithms.I note several things. First, the CIs are large, which is expected—the authors of the causalforest report that the CIs tend to only converge in extensive samples. Second, most of the predicted4The authors estimate those functions because the treatment units (schools) in their data exhibit selection. EssayOne Table 2 provides evidence against selection. Although I use observational data, random dorm assignment ensuresno selection.5Wooldridge (2010), define the partial effect of a variable, w, E[y | w;X] as a derivative of E[y | w] with respectto w keeping X fixed77Figure 3.2: Difference by EthnicityCATE is estimated using causal forest (cf) algorithms in the GRF package. The plot includes two verticallines at zero and 0.0724. The latter corresponds to the average treatment effect from the causal forest: “aver-age_treatment_effect(cf)"treatment effects are positive. Third, the predicted CATE are increasing in the coethnic share.CATE is negative when the coethnic share is below 0.3, and positive and statistically significantcoethnic share is around 0.35 or above. Lastly, the treatment effects seem to peak at a coethnicshare of about 0.7, which corresponds to 18 coethnic peers in average-sized peer groups but theeffect at 0.7 is not different from the effect at 0.6 coethnic share. When I break peer groups intosmall (size below the average) and large (size above the average), the partners are qualitativelysimilar to those portrayed in Figure 3.1.678Figure 3.3: Differences by Ethnicity and GenderCATE is estimated using causal forest (cf) algorithms in the GRF package. The plot includes two verticallines at zero and 0.0724. The latter corresponds to the average treatment effect from the causal forest: “aver-age_treatment_effect(cf)". Within each panel, blue, green, purple, and black correspond to the first, second,third, and fourth ranking in terms of group sizes in my data. Additionally, the plot includes two vertical lines:one at zero and one at 0.0724, where the latter corresponds to the predicted average treatment effect.3.3.2 Heterogeneity by EthnicityFigure 3.2 plots of the CDFs of the individual predicted treatment effects (τ̂i) by ethnicity. I breakthe 16 groups into four panels: the largest four, the second largest four, and so forth. The causalforest also provides functionality for estimating the overall ATE, which is not just an average of theindividual treatment effects. A corresponding coefficient plot is shown in Appendix Figure B5.This figure shows substantial variation in the distribution of the predicted treatment effectsespecially. The top left panel shows that the Baganda have the largest treatment effect as the CDFof this group lies to the right of the ATE. Appendix Figure B5 shows that CATE for Bagandais significant. The second largest group, which we are part of, lies to the left of the ATE. The6It is worth noting that the maximum coethnic share is .54 for the large groups, which is low as expected. Thepatterns are similar over this range of coethnic peers.79remaining two groups, Acholi and Basoga, show that 50% of the CATEs are below the ATE.From the second panel, only 15% of the predicted effects are below the ATE for the Banyoro,Bagisu, and Iteso. In contrast, for Lugbara Baganda, 80% of the predicted average treatment effectslie to the left of the ATE, although positive for most individuals. We observe some variation in thethird panel. However, the pattern is unclear, as the distributions of the smallest group, Rwenzori,and the largest group, Japadhola, overlap and stochastically dominate the other two groups.Lastly, we do not observe variation across the smallest four groups, as the last panel portrays.However, this panel also shows that the predicted τ̂i’s groups are large as the CDFs lie to the rightof the ATE. We should also note that these groups are very underpowered to estimate the treatmenteffects accurately. For example, for the smallest group, Kagwa, we only have 50 instances of thisethnicity in the data, making it difficult to predict the treatment effects accurately.3.3.3 Heterogeneity by Ethnicity and GenderGender norms may differ by ethnicity. For example, women in the central part of the country areculturally expected to be subservient to men, and some ethnicities in the eastern part of Ugandahave cultural norms, such as genital mutilation, that differ by gender. I plot the CDF of theestimated treatment effects by gender and ethnicity in Figure 3.3. The top panel plots the largestfour groups, while the bottom panel plots the third most prominent groups. I include only thesegroups for ease of comparison. Additionally, the ethnic groups in the bottom panel are very differentfrom the ethnic groups in the top panel in terms of language, culture, and sometimes geography. Ireplicate a similar analysis in Appendix Table B1 for 14 out of sixteen ethnicities.The top figure shows considerable differences in the estimated effects by gender across allethnic groups. There is no overlap in the predicted effects by gender for each ethnicity. Also,the differences are not consistently larger for one gender across ethnic groups. For example, thepredicted estimated effects for males are larger than the estimated effects for female students amongthe Baganda and Acholi. The opposite is true for Basoga and Banyankole. Lastly, the mostprominent differences among Basoga are the smallest among the Banyankole. From the bottompanel, the treatment effects do seem to differ by gender across ethnic groups except Alur/Jonam.The bi-directional nature of differences by gender is the direction that requires further investigationI am unable to do right now due do data constraints as it may require qualitative data.80Figure 3.4: Interethnic EffectsCATE is estimated using causal forest (cf) algorithms in the GRF package. This figure plots the mean predictedConditional Average Treatment Effects (CATE) over the indicated pairs. The proportion of Banyankole and theproportion of Baganda in a peer group on the left and The proportion of Banyankole and the proportion of otherethnic groups (excluding Baganda) on the right. I estimate the correlations are equal to -0.44 and -0.258 in (I)and (II), respectively. Although computed correlations are somewhat low, they are significant. Thus, trees mightsplit on one of these variables, especially in pair (I) when growing the casual forest.3.3.4 Interethnic EffectsFigure 3.4 plots the predicted CATEs of two pairs share of Banyankore and Baganda and share ofBanyankore and other ethnic groups. I hypothesize that if heterogeneity in the estimated treatmenteffects exists, it should show up in the left panel, not in the right panel. As mentioned, Bagandaand Banyankole have a long history of political competition and may probably behave in a hostileway towards each other. I thus expected the predicted treatment effect should be high when theproportion of the other group is high, especially among the Baganda, as Section 3.3.2 showed thatthey benefit most from a higher share of coethnic pairs.7. Additionally, heterogeneity should not7If the mechanism through with noncoethnic peers works is competition the expected predicted treatment couldbe high when the proportion of the other group is high81show up in the right panel since Banyankore and ethnic groups do not have a long history ofanimosity or political competition. Nevertheless, suppose other groups (mostly small groups) feelmarginalized by the central government led by the Banyankore. In that case, a higher proportionof Banyankore might lead to a higher predicted treatment effect. However, the figure shows nosignificant heterogeneity in the predicted treatment effects. It is worth noting that the range inthe predicted CATEs is low, as shown by the grid values in both panels. It is also worth notingthat, given my sample size, I may be underpowered to grow to detect effects from trees grown forthis prediction. When a parallel OLS equation with estimation, I also do not obtain significantdifferences.3.4 ConclusionUniversity administrators aim to optimize assignment policies to boost student academic perfor-mance and help governments foster national identity over tribal identity, especially in ethnicallysegregated contexts like Uganda. This involves estimating outcomes for various treatments to in-form policy decisions. My focus is on addressing coethnic peer effects and interethnic dynamics,which can significantly impact student success. For instance, historical political tensions betweenthe Baganda and Banyankore groups may affect how students from these ethnic groups interact andtheir perceptions when they arrive on campus. Also, policymakers might be interested in nonlineari-ties in coethnic peers in a peer group and the potential discrimination or isolation faced by minoritystudents. Ensuring the optimal assignment rule avoids harm and promotes positive interactionsamong diverse student groups is crucial.I use causal forest to estimate expected treatment effects. This utilizes methods that leveragedata-driven sample splits to minimize researcher bias in selecting relevant heterogeneity dimensionsAthey and Imbens (2016); Athey and Wager (2019). This approach avoids overfitting by employ-ing separate training and estimation samples. I estimate Conditional Average Treatment Effects(CATE) for each individual based on an individual’s covariates.My results reveal slightly nonlinear predicted treatment effects based on ethnic share, withinsignificant effects when the coethnic share is below 0.3 but significant positive effects above 0.35.The largest effects are seen among the Baganda, while the Banyankole show the smallest positive82effects, which portray high homophily.The Buganda Kingdom is the strongest historical institution in Uganda, and it has surviveduntil today. People from this group often show allegiance to their traditional king more than thecentral government. On the other hand, members of the Banyankore (the second largest group)might likely identify more as Ugandans than identifying primarily with their ethnic group due totheir political status in the country. The people of Banyakore descent have controlled Uganda’spolitical government for the last 38 years. Moreover, the Banyankore did not reinstate their historickingdom as the Baganda did when the central government offered an opportunity to do so. It mightexplain why a large group, such as the Baganda, cares more about having coethnic peers than theBanyankore.Lastly, the analysis also reveals Gender differences in CATE that vary by ethnicity, withnotable disparities among the Basoga and minimal differences among the Banyankore. Interethniceffects do not show significant variations, possibly due to model limitations. The analysis highlightsstrong homophily, particularly among the Baganda, who maintain allegiance to their traditionalkingdom, unlike the politically dominant Banyankore, who may identify more as Ugandan.These results suggest that it might be feasible to optimize student performance by reallocat-ing students across dorms, at least initially. This could involve leveraging nonlinearities in studentcharacteristics to enhance academic outcomes. For coethnicity, segregating dorms might initiallyimprove performance but could exacerbate long-term negative interethnic attitudes, contrary to thecontact hypothesis highlighted in the first essay, which suggests that intergroup contact reducesprejudice. A thorough cost-benefit analysis, informed by the contact hypothesis literature, is es-sential to balance short-term academic gains with potential long-term social costs, ensuring thatpolicies foster both academic success and social cohesion. Considering this cost-benefit analysis ortrade-offs between academic performance and social cohesion is an area of future research.ReferencesAthey, S., and G. Imbens. 2016. “Recursive partitioning for heterogeneous causal effects.”Proceedings of the National Academy of Sciences 113:7353–7360.Athey, S., J. Tibshirani, and S. Wager. 2019. “Generalized random forests.” The Annals of Statistics47:1148 – 1178.83Athey, S., and S. Wager. 2019. “Estimating Treatment Effects With Causal Forests: An Application.”Observational Studies 5:37–51.—. 2021. “Policy Learning With Observational Data.” Econometrica 89:133–161.Ricart-Huguet, J., and E.L. Paluck. 2023. “When the Sorting Hat Sorts Randomly: A NaturalExperiment on Culture.” Quarterly Journal of Political Science 18:39–73.Wooldridge, J.M. 2010. Econometric Analysis of Cross Section and Panel Data, 2nd ed. MIT Press,Accessed: 2024-06-24.843.5 Appendix3.5.1 Deriving the Identifying AssumptionLet Y = f(SE , X, U) be the GPA production function, whereSE denotes coethnic share, X denotescovariates (including ability A), U denotes unobserved factors. Also, let F denote school/faculty.Conditional on gender and cohort:1. Dorm assignment: D ⊥ (U,X, F,E)⇒ D ⊥ (U,X,E) | FThis is true because unconditional randomization implies conditional randomization. That is,dorm assignment is independent of unobservable, individual covariates, school, and ethnicity,which implies that conditional on school, dorm assignment is independent of U , X, and E2. Peer group definition: G = f(F,D)⇒ G ⊥ (U,X,E) | FIn Essay One, I define a peer group comprised of students admitted to the same school andrandomly assigned to the same dorm within each cohort and gender. Thus, given a randomdorm assignment, my peer group is independent of U , X, and E conditional on school.3. Peer measures:SE = f(G,E)⇒ SE ⊥ (U,X) | F,EThe peer measure, the coethnic share is a function of peer group and the number of coethnicpeers. Combining (1.) and (2.) gives the identifying assumption in (3).3.5.2 Additional Figures and Tables85Figure B5: Coefficient plot by Ethnicity: CATECATE is estimated using causal forest (cf) algorithms in the GRF package. This figure gives a coefficient plot byethnicity along the 90% CI, computed as τ̂i +/- qnorm*sd(τi). Each τ̂i is predicted with its variance. Ethnicitiesare arranged in terms of the mean CATE from the lowest to the largest. The numbers in parentheses correspondrank of the ethnicity in terms of the sample size. These ranks should mirror the relative sizes in the country asFigure 2 of Chapter One. As aforementioned the CI are large, possibly driven by the high dimensional data anda not-so-large sample.86Table B1: Gender Differences CATE by EthnicityEthnicity Female Male DifferenceBasoga (3) 0.163 0.054 0.109Banyankore/kiga (2) 0.051 0.024 0.027Alur/Jonam (10) 0.083 0.050 0.032Jopadhola (9) 0.088 0.093 -0.005Rwenzori (12) 0.085 0.092 -0.007Karimajong (14) 0.079 0.091 -0.012Lugbara (8) 0.054 0.067 -0.013Bagisu (6) 0.077 0.093 -0.016Sabiny (11) 0.064 0.085 -0.021Basamia/Bagwe (13) 0.064 0.085 -0.021Iteso (7) 0.082 0.112 -0.029Acholi/Langi (4) 0.047 0.080 -0.032Banyoro (5) 0.082 0.127 -0.046Baganda (1) 0.100 0.173 -0.073Notes: Table presents the AVERAGE CATE by ethnicity excluding the smallest two groups with a samplesize below 100. The numbers in the parentheses are ranks of the ethnic group size. For example, Baganda(1) refers to Baganda as the largest ethnic group. I do not include the P-values testing the significance ofthe difference. Given the range in the predicted CATE is tight as shown by the range in Figure 3.2.87Chapter 4Beliefs and the Demand for EmployeeTraining4.1 IntroductionThe classic Becker (1962) model of human capital predicts that firms will not invest optimally,leading to the underprovision of general skills training. This suboptimal equilibrium is the resultof the fact that, in perfectly competitive labor markets, firms will pay employees less than theirmarginal product in order to recoup the costs of training, which will induce employees to leave forother firms. Thus, government programs that subsidize training are common, particularly in richand middle-income countries. Underprovision of human capital in firms, however, is particularlyrelevant in developing countries since firms in those settings are characterized by low productivity,which is a likely constraint to growth (Hsieh and Klenow 2009). Information frictions in suchsettings, however, may create unique distortions in labor markets. Thus, even if firm training issubsidized, firms may not choose the optimal employee to train.In this study, we examine the sub-optimal provision of employee training using a novelexperiment. We offer free training to firms belonging to one of the most critical skill-intensivemanufacturing subsectors in Uganda, metal fabrication, and study how owners select workers fortraining. Specifically, we study if owners choose the socially optimal worker that would push themetal fabrication subsector’s production possibility frontier outward the most or if they behave88individually rationally by selecting a worker to maximize the firm’s profits but is not necessarily theworker whose quality would not improve the most from training. Given that our training is free anddesigned to also minimize non-monetary costs, firm owners in our study sample should afford topay a marginal product to workers post-training, ameliorating the anticipated friction of separationby trained workers.We carry out data collection with metal fabrication firm owners, including incentive-compatibleselection for training. We also elicit incentive-compatible owners’ perceived quality of their workersat baseline and endline with or without training. Additionally, we collect perceived profitabilityfrom training each of their worker and data that proxies worker ties with the firm, such as if theworker is a relative, the owner’s trust in the worker, and the perceived likelihood of separation.From the workers’ side, we elicit incentive-compatible demand for training. Specifically, we askworkers to request an amount of money they are willing to accept (WTA) to attend training. Wepay winners their WTA if randomly assigned to receive training when they attend training. Lastly,we test workers’ quality using an objective measure (practical tests) scored by assessors from oneof the prominent government vocational institutes.We randomize small metal fabrication firms (4-14 employees) in our evaluation sample intotreatment and control and offer a training program to the treatment group using a curriculumcarefully designed through consultation with metal fabrication experts and lecturers from one ofthe prominent vocational training institutions in Uganda. Like most manufacturing firms in thedeveloping world, metal fabrication firms in Uganda are small. Small firms are predominant inthe Ugandan economy, where SMEs account for 90 percent of private-sector production accordingto Uganda’s Bureau of Statistics (UBOS 2014). This is typical of most developing economieswhere SMEs contribute up to 40 percent of GDP, according to the World Bank. The productivitydifferential between SMEs and large firms – especially in industries with economies of scale – isparticularly important in developing countries where small firms are also less likely to transforminto large firms (Van Biesebroeck 2005; Olafsen and Cook 2016).Our preliminary analysis reveals that, on average, owners believe that our training programhas a positive benefit (measured by a perceived improvement in quality) on their workers. However,they are less likely to select a worker who would increase the productivity of the metal fabricationsubsector the most. This is a worker whose skills would increase most at the firm. Instead, we find89that owners are more likely to select the worker with the highest perceived profitability post-trainingbut would not improve most from training.That is, we find that owners select workers who they trust and who have strong ties to thefirm. Specifically, owners rank a family member 10.1 percentiles higher for training relative to anon-family member. We also find that a one-point increase in perceived trust is associated withan 8.1 percentile higher rank and a similar increase in perceived risk of separation is associatedwith a 1.1 percentile lower rank in terms of worker selection for training. All these coefficients aresignificant at the one percent level. Third, we find that strong ties to the firm and trustworthinessare also significantly and positively correlated with perceived profitability from training a worker.Yet, there is a negative association between perceived profitability and perceived teachability.Put together, our results reveal that even though our intervention provides a free trainingprogram, eliminating credit constraints on the side of the owner to provide training, and thus, ownerscould afford to pay the workers their post-training marginal profit, owners are not socially optimal.Instead, they are individually rational in that they select workers with the largest gap betweenpost-training marginal product and the wage they can get away with paying without forcing thatworker to leave the firm. Such employees typically have strong ties to the firm, such as relativesor workers who are highly perceived to be reliable by the owner in our setting. On the contrary,workers have a different objective function in that they demand training to maximize their marginalproduct and, consequently, their lifetime earnings. Our results show that the workers’ demand fortraining does not strongly align with the owners’ selection for training.Our study contributes to several strands of literature. First, we contribute to studies docu-menting the effectiveness of several training programs in the developing world. Most of these papersstudy vocational programs.1 For example, Alfonsi et al. (2020) compares firm-provided training tovocational training and finds both types of training improve employment and earnings outcomes forUganda’s “disadvantaged" youth starting in the labor market, but the impact of vocational trainingshows almost twice that of apprenticeships. We contribute to this literature by studying currentfirm employees and how anticipated frictions affect worker selection for training, and thus effective-1For instance, McKenzie (2017) reviews different training programs across different settings, Card et al. (2011)studies a youth training program provided by the government in the Dominican Republic, Cho et al. (2013) studyvocational training programs in Malawi, Hirshleifer et al. (2016) study effects of vocational training targeted tounemployed youth in Turkey on labor market outcomes. Most of these papers report modest treatment of thetraining programs they study.90ness of training programs in settings. We offer training to firms in one of the crucial subsectors inUganda, and we find that owners do not choose the workers that would improve most from trainingand potentially have larger labor market outcomes. Instead, they select workers with low perceivedgains who have strong ties to the firm.We also contribute to the literature on the under-provision of training in skills that aretransferable between firms (Acemoglu and Pischke 1998; Becker 1962; Prendergast 1993; Acemoglu1997). We provide two possible explanations for these settings. First, our descriptive analysisshows that most employees entered the firm through apprenticeship, suggesting that owners haveconfidence in their ability to train low-quality hires. However, firm apprentices may not be effectivefor the industry because they lack standardization. Also, recent literature shows that training insimilar settings increases the separation of trained workers (Brown et al. 2024; Frazer 2006). Weprovide experimental evidence that owners do not select workers with the highest gains from trainingeven though they could potentially afford the marginal product in anticipation of separation.Perhaps the more related, Cefala et al. (2023) study the under-provision of training in agri-cultural markets using two experiments. In the first experiment, both the control and treatmentfarmers receive incentives to train workers on their farms, but the treatment payouts are conditionalon the farmer attending training. They find that tying the incentives to actually providing trainingreduces under-provision of training, but trained workers and non-training firms appropriate thereturns to training. In the second experiment, they tie trained workers’ incentive payouts to theworker working for a farmer who provides training. They find that this reduces under-provision oftraining even though owners do not receive financial incentives to train. The results can be inter-preted that reducing the risk of separation and “poaching externality" increases training provision.Similarly, we find that owners prefer to train workers with stronger ties to the firm. Although theworker trained in their second experiment is chosen by the farmer, our experiment differs in that weobserve the choice set over which the owner is deciding on who to send to training. For example,we observe if owners are deciding between a worker who is a relative and one who is not. Thatis, we can explain why training programs may not be effective because we observe the opportunityof training one worker over the other (we can compute the unrealized gains from owner choices).Lastly, our experiment goes beyond owners and studies worker decisions.914.2 Context and Conceptual Framework4.2.1 Metal Fabrication in UgandaMicro, Small, and Medium Enterprises (MSMEs) are a strategic part of Uganda’s policy for economicdevelopment. The National Development Plan II of Uganda’s government (GoU) underscores therole of MSMEs in wealth and job creation and outlines government objectives to provide institutionalsupport for MSME growth (GoU 2015a). According to the Uganda Bureau of Statistics (UBOS),MSMEs constitute over 90 percent of the private sector in Uganda and contribute 18 percent of theGDP. Additionally, manufacturing-related firms, including welding and steel, made up 8 percent ofall MSMEs in 2011 and employed 15.4 percent of the workforce in 2012 (GoU 2018). Using statisticsfrom UBOS, metal products (steel and fabrication) are among the top seven manufacturing sectors,according to the Index of Production. The same report shows that metal and steel fabrication grewsubstantially, over 10 percent, from 2011 to 2017, reflecting the dynamic nature and potential ofthe manufacturing industry in Uganda.Despite the economic importance and potential for the economy, Ugandan MSMEs face amyriad of challenges. GoU (2015b) summarises Uganda’s MSME policy and lists access to credit,lack of adequate technical skills, and informality as the top three challenges MSMEs face hinderingtheir growth. More recently, and specifically for metal fabrication in Uganda, Bassi et al. (2023)show that another challenge that limits growth is the customer-tailored nature of Uganda’s weldingbusinesses that also stifles specialization.During our listing exercise, we visited around 2,053 firms in Kampala and neighboring sub-urbs and found that less than 40 percent of firms employed at least four workers. In a differentsample, Bassi et al. (2023) reports an average number of six employees in welding firms in Kam-pala even after oversampling the largest firms. Small welding firms in our setting are likely to faceinternal constraints long reported in the literature in other settings that hinder growth, such as sub-optimal management practices in small businesses across several developing countries (McKenzieand Woodruff 2016).2The metal fabrication industry is particularly relevant to our intervention since it is anexample of knowledge-intensive (and capital-intensive) manufacturing that is concentrated only in2Management quality also matters even among large firms in the developing world (Bloom et al. 2012).92small firms in developing countries. As mentioned, there are thousands of small metal fab fabricationfirms in and around Kampala alone, and such firms are typical across the developing world.4.2.2 Conceptual FrameworkWhy would firms that stand to increase their profits through technical training not seek suchopportunities, or why would workers who could increase their wages not make a profitable investmentin the future? Answers to such questions highlight possible frictions that may overwhelm anypossible benefits in the absence of our intervention in this context. Anticipation of such frictionsmay prevent firm owners in our context from demanding training, and thus, training programs maynot arise naturally. We summarize these frictions in three categories:(I) Owners do not have an accurate perception of the value of training.(II) Owners’ selection of workers for training is inconsistent with their beliefs.(III) Worker’s beliefs about the training’s value do not align with those of owners.While these frictions are uniquely tailored to our context, suboptimal investment in humancapital is not. Across both the developed and developing world, firms regularly under-appreciatethe value of training. We document why these frictions arise and why they could lead to under-investment in training by observing them directly in the context of an exogenously-provided (ratherthan endogenously-demanded) training program below.Owners do not have accurate perceptions about the value of trainingOne hurdle to overcome in establishing a training program could be the existing perceptions aboutsuch training programs. In particular, the take-up of a training program will be necessarily hamperedby the perception that such programs are ineffective or poorly suited to the needs of the existingmarketplace. Moreover, firm owners fail to understand how potential benefits from training may bedistributed between the firm and their workers.Incorrect perceptions about training may lead to under-appreciation of a training programand investment. Education literature shows that lower educated parents may expect lower returnsto education, which in turn may affect investment in a child’s education (Brown 2006) as beliefs93about returns to education can affect investment choices (Attanasio, Boneva, and Rauh 2022). Inour setting, firm owners have low levels of formal education and low levels of vocational training,which may affect the effectiveness of our training program. Failure to perceive potential returns totraining may be present even among large firms in developing economies, foregoing potential returnsto even simple low-cost training (Adhvaryu, Kala, and Nyshadham 2018).Additionally, owners in our setting may be overconfident in their ability to hire and traintheir workers, which is strikingly similar to a long history of research on over-confidence amongentrepreneurs (Malmendier and Tate 2005). Entrepreneurs may be overly optimistic about theirlikelihood of success in identifying talented workers or may overstate the abilities of their workforceor their ability to train their workers adequately. This is likely prevalent in our setting as workerroles are flexible, and responsibilities overlap. Thus, overconfident owners may believe that theiremployees already possess the necessary skills or can learn on the job without formal training. Thisover-optimism would diminish the demand for training and distort the assignment to training whenpresented with a free training program.3Owners’ selection of workers for training are inconsistent with their beliefsEven with the correct perceptions about the value of training in terms of worker benefit, firmowners may not select the optimal worker–the most teachable (who would benefit from training themost). Misaligned incentives between owners and workers, behavioral biases, risk aversion, and non-standard production functions could all lead to inconsistencies between beliefs about an individualworker’s returns to training and the selection of workers for training, leading owners not to selectthe most teachable worker. These inconsistencies could manifest themselves in lower average returnsto the training program.Incomplete contracts may characterize this setting. Incomplete contracts naturally leadowners to underinvest in training because of the relationship-specific nature of the investment. Thisis a classical prediction of the hold-up problem in Hart and Moore (1988) and is elaborated on inAcemoglu and Pischke (1998).4 Firm owners will seek to assign workers to training from whom3This could also partly explain the prevalence of apprenticeship in Sub-Saharan Africa reported in Filmer and Fox(2014). Also, in addition to the desire to accrue rents from paying less than market wages to low-skilled employees,firm owners may be confident in their ability to train, so they regularly hire low-skilled workers whom they can trainand pay less than market wages, at least during apprenticeships.4Moreover, solutions to hold-up problems, such as designing wage contracts to offer different wages for different94they can capture the largest returns. This may involve sacrificing a worker’s returns to training fora higher likelihood that the worker remains at the firm.These firms likely practice relational contracting, implying that owners will rely on otherdimensions, such as trust, to select workers for training. This may undermine potential trainingbenefits as this setting is generally characterized by low social trust (Falk et al. 2018). Moreover,social psychology literature shows that non-Western societies emphasize values rooted in loyaltyand morals, underscoring relational contracting. Thus, to avoid separation after training, ownersmay select workers who they believe would stay with the firm based on how much they trust them.Alternatively, firm owners might employ family members or select family members for trainingbecause they trust them more.5Alternatively, a profit-maximizing owner might be individually rational, not socially optimal.A socially optimal owner will select the most teachable worker even though only some of the returnsfrom training may be extracted by the owner. An individually rational owner will select the mostprofitable worker—the worker they perceive to increase the firm’s profit the most after training.The most profitable worker may not be the most teachable.Also, impatience may inhibit the owner from maximizing the returns to training. A present-biased owner may outweigh the loss of production they see from assigning one of their betterworkers to training and decide to send a lower-quality worker who may not benefit as much. Anowner who assigns a worker for training pays an immediate cost in the form of lost productivityfrom that worker while attending training. This conflicts with the long-run benefit of having a moreproductive worker after training. A present-biased owner may under-select high-quality workerswho may benefit more from training because of the large, immediate loss in productivity. 6Lastly, firms with atypical production functions where the complementarity between workerskills is particularly high may rationally select workers for training who will not see the highestindividual returns. For instance, if a firm’s productivity is limited by the lowest skilled worker,tasks proposed in Prendergast (1993) and later discussed in Leuven (2005), is difficult. The model in Prendergast(1993) predicts that if firms offer higher wages for difficult tasks, it may induce workers to invest in training to obtainthe required skill level. The lack of institutional structures that enforce contracts makes this difficult.5This is consistent with literature (e.g., Bertrand et al. 2008; Ilias 2006) across different settings in the developingworld that report high family involvements in micro firms for reasons, such as avoidance of urgency costs.6The analysis presented in the version does not include results on time inconsistency biases of owners. To evaluatethe impact of this intertemporal conflict, we will identify whether or not owners who display present-biased preferencesover cash payments are more likely to express present-biased preferences in their selection of workers for training.95then the firm owner may maximize their returns by sending the lowest skilled worker for training,regardless of the individual returns. Optimizing under the constraint of such a production functionis complicated and requires both correctly perceiving the individual returns to training as well asidentifying the complementarities between the skills of workers.Workers and owners’ beliefs about training do not alignOwners and employers may have different objective functions. Individually rational owners seekto maximize profit, while their employees seek to maximize their lifetime earnings (post-trainingmarginal product). The lack of contractual terms that tie an employee’s marginal product totraining may lead to misaligned incentives for a training program. Just as the owner has theincentive to select workers for training based on their ability to extract the returns from training,workers similarly have the incentive to demand training to maximize private gains from training.Thus, workers who perceive themselves to be more mobile, more connected to other em-ployers, or more likely to transition to self-employment would have a higher demand for training.7Lastly, worker demand for training may be determined by the perceived average improvement fromtraining.4.3 Research DesignOur study relies on the random assignment of firms to treatment and control groups to study theselection preferences of employees into training according to employer and employee preferences.Appendix 4.8.1 Figure 4.8.1 gives a heuristic process flow of sample construction. Within firms, weassign two workers to training: one according to owner preferences and one according to workerpreferences.Our study relies on two main types of data collection: surveys and practical tests. Inaddition to detailed information on the firms and workers, the surveys include incentive-compatibleelicitation of preferences for training. We also rely on applied practical metal fabrication tests to7Higher demand for training could even show in other forms of training, such as apprenticeships. For example,Frazer (2006) develops a model where even among apprenticeships that are firm-specific, apprentices learn the firm’stechnology, which they replicate in future self-employment. He confirms the model predictions with the GhanaianManufacturing Enterprise Survey and finds that 77 percent of the apprentices express strong preference for self-employment.96measure worker productivity.4.3.1 Firm Evaluation SampleWe first identify our evaluation sample by relying on a brief census survey of all metal fabricationfirms located in Kampala subburbs.8 We used this screening survey to physically map firm locationsand identify firms with the targeted number of employees. We limit eligibility to firms with at leastfour employees since choosing employees for training is a decision that is relevant to experiencedmanagers who have more than one or two employees. In addition, it would be impractical for ownersto lose more than 50% of their workforce during training. We excluded firms where the owner oremployees had previously trained with our expert trainer as we did not want this prior interactionto affect take-up. We then approached all the qualifying firms for a marketing exercise and invitedthe owners for orientation.We organized marketing to facilitate adequate take-up. During these visits, we talked withboth owners and employees since both were to engage in the training either through selection totraining or attendance. The marketing visit included an invitation to an orientation event, whichwas another opportunity for employees and owners to learn more about the training.Thus, firms are part of the evaluation sample based on these three criteria. First, they passedthe number of employees requirement and did not have prior interaction with our expert. Second,they must have attended one of the orientation events. That is, they had concrete informationabout training to make informed decisions during the incentive selection exercises we discuss below.Lastly, they were located close to one of three training centers.4.3.2 Sample of Potential TraineesDuring the baseline survey, we interviewed the owners and two workers whom we identified topotentially receive training based on the elicitation exercises. Since this elicitation takes placebefore treatment is assigned, we identify a sample of potential trainees during the baseline survey.We selected the first worker at each firm using workers’ collective preferences. Workersfrom the participating firm cast bids in an auction to be the first potential trainee based on their‘willingness to accept’ participation in training. We asked each worker to demand an amount of8We do not list in our pilot area.97between 0 and 120,000 Shillings ($34) they would be willing to accept to attend training. Thisallowed us to collect information on each worker’s value on assignment to training. The workerwith the highest value for training will report the lowest willingness to accept for assignment (thatis, they require the least subsidy to attend the training), and is thus, selected as the first potentialfor training. Workers cast bids simultaneously to limit information sharing between workers thatwould influence bidding behavior, implying we elicit demand for training only if most (not all) of theworkers were present at some firms.9 Since we paid out the demand made by employees wheneverthey were in the treatment, our demand elicitation was, therefore, incentive-compatible. We callthis elicitation “worker demand elicitation".After worker demand elicitation, we carried out the owner elicitation. The owner’s preferencedetermined the second worker of the potential trainees at each firm. The owners revealed theirpreference for training by ranking their workers from the most preferred worker to the least preferredworker to receive training. Since this ranking determined one of the firm’s two potential trainees,it is, therefore, incentive-compatible.We only revealed the winner of worker demand elicitation simultaneously with the owner’selection after the owner elicitation. We did this to ensure that the owners’ true ranking was notinfluenced by their knowledge of the winner of the worker demand elicitation exercise at each firm.In case of a tie between owner selection and demand elicitation, we pick the second preferred workerby the owner and the winner of the demand elicitation. The ties happened about 20 percent of thetime during our data collection.4.3.3 Training TreatmentWe randomly assigned firms in our evaluation sample where the baseline surveys (owner and em-ployee) and practical tests had been completed into treatment and control groups. The main trainingintervention for this study focuses on technical training for employees of small firms. We investedsubstantially in developing this technical training program for this study with the intention thatit could be implemented more widely once the study is complete. We identified a metal fabrica-9There were cases where not all firm employees were present during this elicitation exercise. In such cases, weonly carried out elicitation if a sufficient percent of the workers were present at the firm during the survey date,otherwise, we would reschedule the survey for the following day. Specifically, we required at least three out of fouror five-employee firms, 4 out of 6-8 employee firms, and at least five employees for larger firms to be present beforeeliciting demand.98tion expert who, before designing the training, conducted semi-structured interviews with firms todetermine the gaps in their technical knowledge. After the training was initially piloted, we hireda curriculum expert to work with our implementing expert to develop a detailed and potentiallyscalable lesson-by-lesson curriculum.10We designed the training to engage and be accessible to employees of small manufacturingfirms. This population has limited formal education and has low levels of literacy, as Table C2shows. This is in contrast to publicly available metal fabrication training in Kampala, which oftenexpects trainees to have had a high school education. Thus, the training focuses on demonstrationsand hands-on practice and does not expect trainees to engage with detailed written materials.The design of the training itself also accounts for a target population that is very low-incomeand has a high opportunity cost of time since they are already working. Thus, the training tookplace in temporary training facilities that are close to the employees’ workplaces to facilitate regularattendance. Thus, to ensure workers are not away from their firms for more than a third of theirworking days per week, we conducted no charge training sessions just two days a week for six toeight hours a day, for a total of 23 days (totaling 168 hours) for each training cohort. Additionally,all trainees received meals and transport reimbursements on training days.4.4 Data and OutcomesWe collected data at three points in time. At baseline, we surveyed owners and workers. Thesesurveys include pre-specified incentive-compatible preference elicitation and the selection of workersin the potential trainee sample in Section 4.3.2 described above. Also, at baseline, we implementedpractical (i.e. applied metal fabrication) tests for the potential trainee sample. Then, immediatelyafter completing training, we conducted a follow-up practical test.4.4.1 Baseline Survey–OwnersOnce we identified the evaluation sample, we conducted a baseline survey of the owners of all partic-ipating firms. We collected data on the owner, the firm, and the firm’s employees. The information10The intention is that this curriculum could be implemented by skilled metal fabricators with limited prior teachingexperience. Scalable curriculums are increasingly common in development, especially for business skills training. Toour knowledge, this is the first technical skills training that is designed to be scaled. Moreover, this training curriculumis at par with the expectations of Uganda’s Level One Directorate of Industrial Training (DIT) for metal fabrication.99on the firm includes data on profitability, assets, access to credit, some measures of productivity,and personnel practices. The survey also included a worker roster that collected information ona worker’s history with the firm as well as their reliability and productivity (as perceived by theowner). This will be our primary source of data and includes all of the metal fabrication workersat the firm. The information from owners includes incentive-compatible elicitation of the owner’stime preference and a measure of risk attitudes.In addition to the incentive-compatible selection for training, the essential information fromthe baseline survey is the incentive-compatible elicitation to capture the owner’s beliefs about theirworker’s quality and gains of training for each of their workers. In these elicitations, owners provideda guess of scores each of the workers would obtain in an objective applied metal fabrication test–practical tests as described in Section 4.4.3 below. Specifically, we ask about beliefs about theirworkers’ scores on this practical test that would be scored on a 0-100 scale.11 Specifically, we askowners about their beliefs about each of their workers’ competency: (I) at baseline, (II)) at follow-up, assuming they are not trained, and (III) assuming that they are trained. The difference between(III) and (II) is our pre-specified measure of teachability. We pay owners whenever their beliefs arewithin the 10 percent range of the correct score.In addition, we ask the owners about the perceived short-run opportunity cost of sendingworkers to training and the perceived post-training profitability of training each worker. That is,each owner ranks each of their workers in terms of perceived profitability gain. These data fromelicitations and perceived profitability form the central part of the analysis, together with practicalskills test scores.4.4.2 Baseline Survey–WorkersAs part of the baseline survey, we interview the two potential trainees from each firm once they areidentified. The worker roster in the owner survey collects basic information on each worker, such astheir tasks and their work history at the firm. The worker survey, however, collects a more detailedwork history as well as complete wage information, particularly from any work done at other firms.The survey also collects more in-depth data on worker characteristics, mainly Raven’s matrices, to11When doing this, we explain to the owners that a score of fifty would correspond to an average metal fabricator inKampala, a score of zero would correspond to a person without any metal skills, and a score of 100 would correspondto a metal fabrication expert.100capture cognitive ability. In addition, we will conduct incentive-compatible elicitations of risk andtime preferences.We asked workers about their beliefs about the impact of training. The question askingowners about their beliefs on the perceived impact of our training on their competency was notincentivized to prevent hedging behavior. We incentivized, however, the question about their beliefabout the average score of the trained worker. Lastly, we asked workers about how they think gainsfrom training would be shared between the firm and themselves (i.e., if their wage will be increasedafter training).4.4.3 Practical Skills TestsAfter we had determined the two potential trainees from each firm, we invited all potential traineesto participate in a baseline practical skills test. The skills test involved the fabrication of a standardmetalworking product in such a way that evaluators could test several relevant skills (e.g., measuring,cutting, welding, grinding). The evaluators assessed each worker based on quality, speed, safety,and the efficient use of materials. Scores range from 0 to 100 percent, with the highest scorecorresponding to the output of a master metal fabricator with decades of experience. The testswere scored by accredited assessors (different from the trainer) who teach at one of the prestigiousand accredited vocational institutes in Kampala.Soon after concluding the training, potential trainees from the treatment and control groupsrepeated the baseline practical skills test. With repeated observations, we can calculate the improve-ment of each worker and evaluate the effect of the training on technical skills. We can then comparetrue improvement in skills to the perceived improvement collected in our incentive-compatible elic-itations.4.5 Preliminary ResultsOur experiment is still ongoing at the endline practical test stage. The results I present in thissection are preliminary and correlational. We provide regression equations for the results in eachrespective section.1212We shall revise this to include the empirical strategy section when updating the analysis with endline data1014.5.1 Descriptive StatisiticsTable C1 reports firm-level summary statistics, while Table C2 reports summary statistics of theworkers in the potential trainee sample. Our sample comprises dynamic firms and active en-trepreneurs in terms of employee and product turnover who have been in business for an averageof ten years. Additionally, an average of one employee quit, while an average of two new employeeswere hired in the firm’s previous year. Despite the artisanal business nature of the firms in oursetting, close to 50 percent of the entrepreneurs reported that they introduced new products in thelast year. The average profit was approximately 1,551,100 Ugandan shillings ( $430).The average number of employees is five, which is comparable to the 5.88 sample averageof welding firms in the same setting reported in Bassi et al. (2023). Almost all the entrepreneursin our sample are committed, not survival entrepreneurs, as only six percent mentioned they wouldleave metal fabrication if offered a job with a wage equivalent to their current profits. Also, theseentrepreneurs are optimistic about the future. More than 80 percent expect their business to survivefor at least five years, and the average number of employees is expected to more than double in thenext five years. Additionally, these entrepreneurs expect close to 70 percent of current businessesto survive for at least five years, portraying a general positive outlook on the economy.These entrepreneurs have low levels of formal education. The average years of formal educa-tion is 10.4, equivalent to completing Year Three of O-level (Ordinary Level) education in secondaryschool. Moreover, less than 20 percent have completed formal metal fabrication training. Thus,most of these entrepreneurs likely learned their skills on the job, as most have been entrepreneursfor more than ten years. It is possible these entrepreneurs may not correctly perceive returns totraining.Credit constraints may be a factor affecting the provision of training in our setting, as lessthan five percent have applied for loans to train workers. When we asked why they had not appliedfor training, most of these entrepreneurs believed it was unnecessary, while others mentioned theydid not know how or where to apply for loans. It is likely that entrepreneurs in our sample rely ontheir own ability to train their staff. It is thus unsurprising that more than 80 percent of employeesmentioned that they had been apprentices at some point at the current firm, as Table C2 shows.Turning to Table C2, the average age of workers is 24 years old, and all the workers employed102in this subsector are male. Additionally, it appears most of the workers in our sample started at theircurrent firm, as the ratio of years in metal fabrication and years at their current firm is 1.01. As withthe owners, their education level is low. The average education is barely one year of post-primaryeducation, which is two years less than the average of owners reported in Table C1. Moreover,only 5 percent of the employees have completed vocational training, indicating a need for trainingprograms Yet, Alfonsi et al. (2020) show vocational training in this setting is more beneficial thanon-the-job training.The average hourly wage is 4,140 shs ($1.14), and our sample comprises full-time employees,defined as working at least four days per week at the firm, as per the sample restriction. Althoughnot reported, most of the employees are paid based on completed orders. Additionally, a quarterof the employees have worked for more than one firm, and close to half discuss job opportunitieswith someone from a different firm. Lastly, more than 80 percent stated that they had ever beenapprentices at the firm, implying that most workers use apprenticeships to obtain employment inthe sector and that most employees are engaged in some informal training. The average in thissector is way larger than the proportion of apprentices in other sectors Filmer and Fox (2014).134.5.2 Beliefs about Worker QualityFigure B1 plots the owner baseline perception of quality against the workers’ practical test scores,which is our objective measure of quality. Additionally, the plot includes three lines of fit: the45-degree line, the best-fit for owner-selected workers, and the best-fit for other workers. We cannote several findings from this figure. First, owners are sophisticated in the sense that they candistinguish a high-quality worker from a low-quality worker among their metal fabrication staff.The correlation between perceived quality and objective score is high (close to 70 percent).Nevertheless, owners seem to overestimate the quality of their workers on average. From theY-axis, there are several firms where the perceived quality is above 80 percent or even equal to 100,but none of the workers scored above 80 percent in the baseline practical test. Interestingly, ownersseem to perceive their selected workers to be of higher quality than the other workers. Yet, the13Filmer and Fox (2014) Figure 3.18 reports that an average of 20 percent of SSA youth have ever been anapprentice, where the proportion is close to eight percent in Uganda. They report statistics of apprenticeships acrossall sectors in six countries using a standardized survey. This implies Uganda’s metal fabrication subsector may bean outlier, as apprenticeships are a more prevalent way of finding employment and acquiring metal fabrication thanother sectors.103correlation of their perceptions with practical tests is lower than that of worker-selected workers.That is, the slope of the line of best-fit for owner-selected workers is lower than the slope of the lineof best-fit for all other workers.4.5.3 Do Owners Select the Most Teachable Worker for Training?Figure B2 plots the perceived benefit to workers from training each worker on the Y-axis. Each spikerepresents a firm, and each dot represents the perceived teachability of each worker. The longer thespike, the larger the dispersion in perceived teachability within a firm. From the graph, perceivedteachability is positive on average. That is, most owners believe that their workers’ quality willimprove with training. Despite their beliefs, a lot of owners do not select the worker that wouldimprove most from training, implying that owners are not socially optimal.4.5.4 Do Owners Select the Most Profitable Worker for Training?Figure B3 plots owner selection for training on the left panel and most teachable on the right panelby rank of perceived profitability post-training on the right panel. Suppose i represents a worker,and k represents the ranking of perceived profitability. Each bar on the left panel can be read as“What proportion of the owners that selected worker i ranked k in terms of perception of how muchi would each increase a firm’s profit after training?” Using our teachability measure, we created adummy that is equal to one when a worker is the most teachable (right panel) at a firm and zerootherwise. Thus, we interpret each bar on the right panel as “What is the proportion of firms wherethe most teachable worker ranked k in terms of perception of how much profit after training?”From both panels, we see that owners are sophisticated in that only a small proportioncannot rank their workers in terms of profitability. From the left panel, we observe that more than70 percent of owners selected workers that ranked first or second in terms of profitability for training,and close to 55 percent selected the worker they perceived to bring in the most profit to the firmfor training.In the right panel, there seems to be no correlation between perceived profitability andperceived teachability. That is, most owners do not perceive the most profitable as the mostteachable, as only about 10 percent perceive the most teachable worker as the most profitable.Additionally, a significant number of firms perceive the least profitable workers as the most teachable,104with close to 40 percent ranking the most teachable fourth or greater.Taken together, these results show that owners are individually rational in selecting workersto maximize profits. However, they do not perceive this increase in profitability to come from workerswho would improve most from training. This implies that owners consider other dimensions in theirdecision-making, such as considering which workers might separate after training. We report theseresults in Section 4.5.5 below.4.5.5 Employee Firm Ties and Selection for TrainingTo study what variables predict owner selection and perceived profitability, we use the followingregression:(A1) Yij = β0 + ϕ1Familyij + ϕ2Trustij + ϕ3Separationij + ϕ4Tenureij + δj + εij ,where Yij is the outcome of interest, such as owner j’s selection ranking of worker i’s fortraining. Familyij is an indicator that is equal to one if the worker is related to the owner,Separationij is the perceived likelihood that the worker will leave the firm in the one year, andTenureij is how long the firm has employed a worker, and δj are firm fixed effects.Table C3 presents several versions of equation (A1) where the selection ranking of workersfor training is the outcome in (1) and (4), perceived profitability in (2) and (5), and teachabilityin column (3). Additionally, columns (4) and (5) control for the percentile of a worker’s perceivedteachability. All regressions control for fixed effects, and standard errors are clustered at the firmlevel. These regressions can be interpreted as correlational rather than causal, as the experiment isongoing.The findings reveal that the owner consistently selects a worker who is a relative for training.This coefficient is positive and significant at the one percent level in (1) and does not change whenwe control for perceived teachability in (4). Additionally, a family member is perceived to be moreprofitable. The coefficient is also positive and significant in both columns (2) and (5). A coefficientof 0.101 from (1) indicates a family member is ranked 10.1 percentiles higher compared to a non-family member. Also, an owner perceives a family member to increase a firm’s profitability by 10.3percentiles more than non-family members when selected for training. Lastly, being a relative is105not associated with perceived improvement from training.Trust, measured by worker perceived reliability, is also positively and highly significant whenthe outcome is owner selection for training or perceived profitability. From column (1), a one-unitincrease in trust, measured on a Likert scale, is associated with an increase of about 8.1 percentilesin the worker’s ranking for selection for training, while the same change in trust is associated witha 10.4 percentile increase in profitability in column (2). On the contrary, trust and perceivedteachability are negatively correlated. A one-unit increase in trust is associated with a decrease ofseven percentiles in teachability.Conversely, the likelihood of leaving is negatively associated with training selection or per-ceived profit and is not significant when the outcome is teachability. From (1), a one-unit increasein the perceived likelihood of worker separation in one year is associated with a decrease of approx-imately one percentile in their training selection ranking or perceived profitability.Tenure shows mixed results. The coefficient is negative and significant in (3) when teacha-bility is an outcome, barely significant when the outcome is selection for training, and not significantwhen the outcome is perceived profitability. From (3), one more year of tenure at the firm is as-sociated with a 1.4 percentile decrease in teachability and a 0.6 percentile less likely selection fortraining. This result implies that owners may have confidence in their ability to train. They mayperceive workers who have been at the firm for a long time to have already gained skills such thatadditional training would not be beneficial.Lastly, we add teachability as a control in models (4) and (5), which shows a negative and sig-nificant impact in model (5) when perceived profitability is an outcome but not significant in model(4) when selection for training is an outcome. The results from (5) indicate that a one-percentileincrease in teachability is associated with a 0.1 percentile decrease in perceived profitability. Thiscounterintuitive result suggests that workers who are perceived as more teachable are also perceivedto be less profitable, implying factors that proxy relational contracting, such as family, trust, andperceived separation, matter for owner selection for training and this perceived profitability.1064.5.6 Worker DemandTo study what variables that predict worker demand for training, we use the following regression:(A2)Dij = β0+ϕ1Qualityij+ϕ2WGainij+ϕ3OGainij+ϕ4WageChangeij+ϕ5ExpectedTenureij+δj+εij ,where Dij is worker i’s at firm j demand for training. Qualityij is the perceived quality ofthe trained staff at the endline, WGainij the worker’s own perceived gain in quality from training,OGainij is the owner’s perceived gain in quality from worker training, WageChangeij is the per-ceived wage change after training, ExpectedTenureij is he expected tenure at the current firm, andδj are firm fixed effects.The perceived average score of trainees reflects the incentivized average score of trained staffon the endline practical test. Own perceived benefit is the measure of how much a worker believeshe would gain in skill if selected for training, which may be different from the owner’s perceptionof the benefit.The outcome variable is the percentile of demand for training. All regressions control forfixed effects. The difference between columns one and two is from the variables we control for. Wedo not have data on perceived wage change after training for all the workers, and some workers didnot know how long they expected to stay at the firm. Therefore, the number of observations incolumn two is reduced compared to column one. The regressions from columns one and two showthat the owner’s perceived teachability training is the strongest predictor of workers’ demand fortraining. Additionally, the perceived gain for training is associated with an increase in demand fortraining. However, the coefficient of own perceived benefit is qualitatively different between columnsone and two.4.5.7 Worker Demand Vs Owner SelectionFigure B4 is a scatter plot of the owner’s percentile ranking for training on the Y-axis and thepercentile of worker demand for training on the X-axis. As expected, we observe bunching aroundone on both axes. The worker selected for training is consistently ranked number one, hence thebunching around one on the Y-axis. Similarly, the worker with the lowest demand for training107corresponds to the highest rank/percentile, leading to bunching around one on the X-axis. Thegraph indicates weak alignment between worker demand for training and the owner’s selection fortraining, implying that the workers selected by the owner are not necessarily those who demandtraining the most.4.6 ConclusionThe classic Becker (1962) model of human capital predicts that firms will underinvest in general skillstraining because, in competitive labor markets, firms will pay employees less than their marginalproduct to recoup training costs, prompting employees to leave for other firms. This underprovisionis especially pertinent in developing countries, where low productivity constrains growth. Govern-ment subsidies for training are common in richer countries to address this issue, but in developingcountries, information frictions may lead firms to select suboptimal employees for training even ifit is subsidized. This study examines the suboptimal provision of employee training through anexperiment in Uganda’s metal fabrication sector, offering free training and analyzing whether own-ers choose the socially optimal worker who would maximize sectoral productivity or the one whomaximizes firm profits but would not improve most from training.We conducted data collection with metal fabrication firm owners, including incentive-compatibleselection for training. We also elicited owners’ perceived quality of their workers at baseline andendline, both with and without training, as well as perceived profitability from training each workerand data on worker ties to the firm, such as kinship, trust, and perceived likelihood of separation.From the workers’ side, we gathered incentive-compatible demand for training by asking them tospecify a willingness-to-accept (WTA) amount to attend training, with winners paid their WTAif selected. Workers’ quality was objectively assessed through practical tests scored by vocationalinstitute assessors. We randomized small metal fabrication firms (4-14 employees) into treatmentand control groups, offering a training program to the treatment group, using a curriculum designedwith input from metal fabrication experts and vocational lecturers.Our preliminary analysis reveals that, on average, firm owners believe that our trainingprogram positively impacts their workers’ quality. However, they tend to select workers based onperceived post-training profitability rather than those who would maximize productivity gains for108the metal fabrication sector. Owners prefer workers they trust and those with strong ties to thefirm, such as family members, ranking them significantly higher for training. Despite free trainingeliminating credit constraints and mitigating the anticipated risk of separation since owners canafford to pay workers their post-training marginal product, owners remain individually rational,choosing workers with the largest gap between post-training marginal product and wage, oftenrelatives or highly trusted workers.Our findings suggest that interventions to improve the effectiveness of training programsshould consider frictions, such as separation, and aim to align owner-worker incentives better. Lastly,while we find that owners have positive perceptions of our training program and training in general,the analysis at this stage does not test whether they are underestimating or overestimating theeffectiveness of training. We shall test this once our endline is complete. We find, however, that theowners are overestimating their worker quality on average.109ReferencesAcemoglu, D. 1997. “Training and Innovation in an Imperfect Labour Market.” The Review ofEconomic Studies 64:445–464.Acemoglu, D., and J.S. Pischke. 1998. “Why Do Firms Train? Theory and Evidence.” The QuarterlyJournal of Economics 113:79–119.Adhvaryu, A., N. Kala, and A. Nyshadham. 2018. “The Skills to Pay the Bills: Returns to On-the-job Soft Skills Training.” Working Paper No. 24313, National Bureau of Economic Research,February.Alfonsi, L., O. Bandiera, V. Bassi, R. Burgess, I. Rasul, M. Sulaiman, and A. Vitali. 2020. “TacklingYouth Unemployment: Evidence From a Labor Market Experiment in Uganda.” Econometrica88:2369–2414.Attanasio, O., T. Boneva, and C. Rauh. 2022. “Parental Beliefs about Returns to Different Typesof Investments in School Children.” Journal of Human Resources 57:1789–1825.Bassi, V., J.H. Lee, A. Peter, T. Porzio, R. Sen, and E. Tugume. 2023. “Self-Employment Withinthe Firm.” Working Paper No. 31740, National Bureau of Economic Research, September.Becker, G.S. 1962. “Investment in Human Capital: A Theoretical Analysis.” Journal of PoliticalEconomy 70:9–49.Bertrand, M., S. Johnson, K. Samphantharak, and A. Schoar. 2008. “Mixing family with business:A study of Thai business groups and the families behind them.” Journal of Financial Economics88:466–498, Darden - JFE Conference Volume: Capital Raising in Emerging Economies.Bloom, N., B. Eifert, A. Mahajan, D. McKenzie, and J. Roberts. 2012. “ Does Management Matter?Evidence from India *.” The Quarterly Journal of Economics 128:1–51.Brown, G., M. Hardy, I. Mbiti, J. McCasland, and I. Salcher. 2024. “Can Financial Incentivesto Firms Improve Apprenticeship Training? Experimental Evidence from Ghana.” AmericanEconomic Review: Insights 6:120–36.Brown, P. 2006. “Parental Education and Investment in Children’s Human Capital in Rural China.”Economic Development and Cultural Change 54:759–789.Card, D., P. Ibarrarán, F. Regalia, D. Rosas-Shady, and Y. Soares. 2011. “The Labor Market Impactsof Youth Training in the Dominican Republic.” Journal of Labor Economics 29:267–300.Cefala, L., P. Naso, M. Ndayikeza, and N. Swanson. 2023. “Under-training by Employers in SpotLabor Markets: Evidence from Burundi.” Unpublished, Mimeograph.Cho, Y., D. Kalomba, A.M. Mobarak, and V. Orozco. 2013. Gender Differences in the Effects ofVocational Training: Constraints on Women and Drop-Out Behavior. The World Bank.Falk, A., A. Becker, T. Dohmen, B. Enke, D. Huffman, and U. Sunde. 2018. “Global Evidence onEconomic Preferences*.” The Quarterly Journal of Economics 133:1645–1692.Filmer, D., and L. Fox. 2014. Youth Employment in Sub-Saharan Africa. Africa Development Series,110Washington, DC: World Bank.Frazer, G. 2006. “Learning the master’s trade: Apprenticeship and human capital in Ghana.” Journalof Development Economics 81:259–298.GoU. 2015a. “Second National Development Plan (NDPII) 2015/16 – 2019/20.” Working paper,Government of Uganda, National Planning Authority.—. 2018. “Trade and Industry Sector Statistical Abstract.” Working paper, Ministry of Trade,Industry and Cooperatives (MTIC), Kampala, Uganda.—. 2015b. “Uganda Micro, Small and Medium Enterprise (MSME) Policy: Sustainable MSMEsfor Wealth Creation and Socio-Economic Transformation.” Working paper, Ministry of Trade,Industry and Cooperatives (MTIC), Kampala, Uganda, June.Hart, O., and J. Moore. 1988. “Incomplete Contracts and Renegotiation.” Econometrica 56:755–785.Hirshleifer, S., D. McKenzie, R. Almeida, and C. Ridao-Cano. 2016. “The Impact of VocationalTraining for the Unemployed: Experimental Evidence from Turkey.” The Economic Journal126:2115–2146.Hsieh, C.T., and P.J. Klenow. 2009. “Misallocation and Manufacturing TFP in China and India*.”The Quarterly Journal of Economics 124:1403–1448.Ilias, N. 2006. “Families and firms: Agency costs and labor market imperfections in Sialkot’s surgicalindustry.” Journal of Development Economics 80:329–349.Leuven, E. 2005. “The Economics of Private Sector Training: A Survey of the Literature.” Journalof Economic Surveys 19:91–111.Malmendier, U., and G. Tate. 2005. “CEO Overconfidence and Corporate Investment.” The Journalof Finance 60:2661–2700.McKenzie, D. 2017. “How Effective Are Active Labor Market Policies in Developing Countries? ACritical Review of Recent Evidence.” The World Bank Research Observer 32:127–154.McKenzie, D., and C. Woodruff. 2016. “Business Practices in Small Firms in Developing Countries.”Management Science 63:2967–2981.Olafsen, E., and P.A. Cook. 2016. Growth Entrepreneurship in Developing Countries: A PreliminaryLiterature Review. Washington, DC: The World Bank Group, License: Creative Commons At-tribution CC BY 3.0.Prendergast, C. 1993. “The Role of Promotion in Inducing Specific Human Capital Acquisition.”The Quarterly Journal of Economics 108:523–534.UBOS. 2014. “Statistical Abstract. Uganda Bureau of Statistics.”Van Biesebroeck, J. 2005. “Firm size matters: Growth and productivity growth in African manu-facturing.” Economic Development and cultural change 53:545–583.1114.7 List of Figures and Tables4.7.1 Figures112Figure B1: Owners Overestimate Worker Quality.020406080100Owner's Perceptions of Worker Quality (%)0 20 40 60 80 100Practical test score (%)Owner-Selected Worker selectedLine of Best-Fit (Owner-Selected) Line of Best-Fit (Worker selected)The X-axis (practical test score) is the objective measure of baseline quality, while the Y-axis is the perceived measureof quality. As mentioned in the data section, the perceived baseline was incentivized. We surveyed each owner aboutwhat they thought each of their workers would obtain if they were invited to take part in the test, and we gave anowner 10,000 shillings whenever their guess was within the correct range.113Figure B2: Teachability: Owner Selection vs Other WorkersNotes. Each line/spike represents a firm. The firms are sorted first on the perceived gains of the selected worker andsecond on the percentile rank of workers selected for training relative to other workers within the firm. We excludecases when the within-firm standard deviation in the perceived teachability is zero.114Figure B3: Perceived Profitability: owner Selection vs Most Teachable0%20%40%60%PercentN/A 1st2nd 3rd 4th>=5thTraining Selection0%20%40%60%N/A 1st2nd 3rd 4th>=5thMost TeachableThe X-axis is the rank in terms of perceived profitability. We surveyed each owner to determine which of theirworkers would lead to a profit increase after they had been trained. After this, each owner ranked the workers thatwould lead to profit from the most profitable to the least profitable. “Ranked None” refers to owners who believethat training does not increase profits or who cannot differentiate the profitability of their workers.115Figure B4: Perceived profitability: Owner selection vs worker demand most teachable0.2.4.6.81Owner ranking of worker (percentile)0 .2 .4 .6 .8 1Worker demand for training (percentile)Owner's Selection All Other Workers Line of Best-FitBunching on one is expected as the selected workers will rank on top of the distribution.1164.7.2 TablesTable C1: Owner/Firm-level Descriptive StatisticsN Mean SD Min MaxAge 336 36.50 8.85 19 73Years in metal fabrication 337 14.25 7.86 0 49Completed formal metal fabrication training 275 0.19 0.40 0 1Education (years) 336 10.35 4.18 0 22Would shut down business a job offer(wage=profits) 337 0.06 0.23 0 1Firm age 330 10.35 7.95 0 51Number of employees 336 5.35 1.82 4 14New employees hired last 12 months 336 2.26 3.40 0 30Number of Employees who quit last 12 months 337 1.02 1.72 0 12Introduced new products last 12 months 337 0.45 0.50 0 1Revenue in the last 30 days (10,000) 337 803.35 1884.09 0 30,000Profits in the last 30 days (10,000) 337 155.11 224.44 0 1,500Employees specialize on tasks 337 0.29 0.46 0 1Very Likely still in business in 5 years 337 0.84 0.36 0 1Expected number of employees in 5 years 333 11.41 8.26 2 70Expected number of similar firms (out of 10) in 5 years 333 6.88 2.61 1 10Ever received a loan 337 0.25 0.43 0 1Ever applied for a loan to train employees 337 0.03 0.16 0 1Ever received for a loan to train employees 337 0.01 0.11 0 1Change wages after working 337 0.92 0.28 0 1Notes: Data are from the baseline survey. Revenue and profit are measured in Ugandan Shillings (shs). Toconvert shs to dollars, divide the shs amount by 3,600. Unless specified, the variable is binary whenever themin is zero and the max is one, such as “Completed formal metal fabrication training”. Ever received loan refersto any other type of loan irrespective of training loan. We interviewed businesses if they have ever receivedany loan from any commercial bank, microfinance institution, savings cooperative, or other certified institutionsseparately from if they have ever applied or received a loan to train their workers specifically.117Table C2: Worker-level Descriptive StatisticsN Mean SD Min MaxAge 674 24.07 5.19 18 53Female 674 0.01 0.10 0 1Years in metal fabrication 673 4.53 3.76 0 23Years at firm 573 4.49 5.04 0 50Education 673 8.34 3.14 0 22Completed vocational training 674 0.04 0.20 0 1Ever under apprentice at firm 674 0.82 0.39 0 1Hourly wage 660 4139.50 3300.13 0 12,373Days per week 674 5.79 1.26 0 7Days last month 673 22.58 7.36 0 30Worked for other firm last month 604 0.24 0.43 0 1Discusses opportunities with different firm workers 674 0.49 0.50 0 1Notes: Data are from the baseline survey. Revenue and profit are measured in Ugandan Shillings. An averageeducation of 8.34 years is equivalent to completing one and a half years of secondary education. Divide by 3,600to convert the hourly wage to dollars. A few of the workers did not remember when they started at the firm.Ever under apprentice at the firm includes both employees who were currently at the firm under apprenticeshipor started and completed an apprenticeship at the firm. Opportunities refer to job opportunities. We askedworkers if they share/discuss job opportunities with workers from different firms.118Table C3: Factors Affecting Selection for Training(1) (2) (3) (4) (5)OwnerrankingProfitabilityrankingTeachabilityPctileOwnerrankingProfitabilityrankingFamily 0.101*** 0.103*** 0.007 0.100*** 0.104***(0.02) (0.02) (0.02) (0.02) (0.02)Trust 0.083*** 0.104*** -0.070*** 0.081*** 0.097***(0.01) (0.01) (0.01) (0.01) (0.01)Likelihood of leaving -0.011*** -0.010** -0.002 -0.011** -0.010**(0.00) (0.00) (0.00) (0.00) (0.00)Tenure -0.006* 0.004 -0.014*** -0.006* 0.003(0.00) (0.00) (0.00) (0.00) (0.00)Teachability percentile -0.033 -0.093***(0.03) (0.03)R-squared 0.085 0.148 0.096 0.086 0.155N 1,782 1,722 1,776 1,776 1,714The outcomes in columns are 0-1 percentile range of the indicated variable. Family is an indicatorof whether a worker is related to the owner. Trust and likelihood of leaving are captured usinga 0-10 Likert scale. Trust measures the extent to which the owner can depend (all the time) onthe worker to do assigned work reliably, while the likelihood of leaving measures the perceivedseparation of a worker in on year.All regressions control for control for firm FEs.Standard errors in parentheses* p<0.10, ** p<0.05, *** p<0.01119Table C4: Worker Demand for training(1) (2)Perceived average score of the trainees -0.161 -0.166(0.17) (0.20)Own perceived benefit from training 0.051** 0.496*(0.02) (0.28)Owner perceived benefit from training 0.874*** 0.736***(0.19) (0.27)Perceived wage change 0.003(0.01)Expected tenure -0.003(0.00)R-squared 0.053 0.063N 636 495The outcome is a percentile of worker demand for training, which we measured using theincentive-compatible amount of money a worker was willing to accept to attend training. Ex-pected Tenure is measured in years, while all other controls are measured on a 0-1 scale.All regressions control for control for firm FEs.Standard errors in parentheses* p<0.10, ** p<0.05, *** p<0.011204.8 Appendix4.8.1 Time Line and Sample ConstructionOur research and data collection process began with an in-person census of mental fabrication firmsin the Kampala metropolitan areas coupled with marketing. Our two-person team (enumerator andmarketer) visited all firms in our study area, identified those that meet the criteria, and introducedthe program. Our marketers were students from nearby vocational schools with the ability toclearly explain the nature of the training, including potential skills and benefits, with the purposeof recruiting businesses in the study area into our training program. We gave business owners anopportunity to opt in by applying.One month later, business owners who applied and their workers were invited for a 3-4hour orientation where they interacted with our implementing partner and accredited trainers. Thetrainers provided a high-level introduction to the curriculum and other training details, such asfacilitation on training days. Both listing and orientation were crucial in identifying our evaluationsample by identifying firms that met the number of employees requirement or were located at thetraining centers.Figure 4.8.1 highlights the workflow of the sample construction. Every firm that attendedthe orientation had more information about our training program, making the owner likely to buyinto our training program. We also considered the firms located close to the training centers eventhough they did not attend orientation, as these would incur a small cost (in terms of travel time).As mentioned, our training program covered the financial cost eliminating credit constraints thatwould affect employee training. We randomized at the firm level after conducting the post-baselinesurvey. As the figure shows, two employees are selected at each firm through incentive-compatibledemand elicitation with the workers or owner elicitation for worker selection. We identify spilloverworkers as any other workers at the treatment firm whom the owner did not select for training ordid not have the highest value for training.Figure 4.8.1 illustrates our sample selection procedure:121Census of firms intarget neighborhood4-10 workers?Prior Sebbtraining?OrientationAttendorientation?Located neartrainingcenter?Baseline surveysHighest valuefor training?Selected byowner?Non-TraineePotential TraineeTreatmentFirm?TreatmentFirm?Spillover WorkerControl: poten-tial trained orspillover workerTrained Worker Out of Samplenoyesyesnono noyesyesyesnoyesnoyesnoyes122Table C5: Project TimelineOct 2022 - Jan 2023 • Listing and marketing• Introduced our program to the owners and allworkers that present at the firm.• Identified firms that would qualify to meet thenumber of employees collected, location, andcontact information.• Allowed firms an opportunity to apply toparticipate in our training program.Feb 2023 • Orientaion• Our expert trainer introduced the curriculumto owners.Jul - Nov 2023 • Baseline survey• Collected demographics from both owner andworkers.• Incentive-compatible exercises with the ownerabout perceived worker quality.• Collected Owner selection for training.• Collected worker demand for training, andthus, worker selection for training.Aug - Nov 2023 • Baseline practical tests• Objective measure for worker quality atbaseline for all workers in our evaluationsample.Jan 2024 • Treatment assignmentFeb-May 2024 • Training• Every cohort trains for two days a week and6-8 hours per day.May -June 2024 • Endline practical tests• Objective measure for worker quality at endlinefor all workers in our evaluation sample.We invited workers from each location whenever we completed the baselinesurvey from that location as enumerators moved to another location duringthe baseline survey, which explains the overlap between baseline surveys andbaseline practical tests.123 |
|---|