Alan Griffith

Working Papers

Random Assignment with Non-Random Peers: A Structural Approach to Counterfactual Treatment Assessment
(Job Market Paper)

Recent efforts by economists to exploit peer effects by creative peer assignment have come up short due in part to endogenous peer selection. That is, even conditional on random assignment, agents choose their peers, and failure to account for this selection may crucially bias predictions of the effects of alternative policies. To address this shortcoming of the literature, I build a two-part model in which (1) agents form a network; (2) conditional on the realized network, outcomes are determined by a process that allows for non-linear peer effects. To overcome difficulties in identification and estimation of network-formation games, agents in my model make continuous linking decisions subject to a budget constraint. I show that, under certain conditions, this model has a unique strictly positive equilibrium, which can then be used for identification and estimation. In modeling peer effects, I explicitly model network endogeneity as an omitted variable problem, and further propose a method to recover these omitted variables in estimating the network-formation game. I estimate the parameters of the two-part model using innovative data on networks and outcomes from a randomized study in Rajasthan, India, then show that the model performs well in matching predictions to realized out-of-sample outcomes. This paper makes important contributions to the methodology of peer effects estimation as well as the theory and econometrics of network formation, while providing an important link between structural and experimental approaches to policy evaluation.

Available here.

Network Partitioning and Social Exclusion under Different Selection Regimeswith Clara Delavallade and Rebecca Thornton

While most social programs are based on some form of exclusion of sub-populations, we know little about how being excluded, and the selection process, affect social inclusion. This paper compares peer effects of an after-school program, under three different (randomly assigned) network-formation regimes: endogenously formed, popularity vote, and randomly assigned. We find substantial evidence of homophily within endogenously-formed and elected networks. When participation was randomly assigned, we find segregation of friendships due to the program. We do not find this among elected networks, mainly because they were already highly partitioned. Lastly, we find that social exclusion – not being elected in a school with popular voting – reduced education aspirations and self-confidence.

Available here.

Participation, Learning, and Equity in Education -- Can We Have It All?with Clara Delavallade and Rebecca Thornton (submitted)

The Sustainable Development Goals set a triple educational objective: improving access to, quality of, and gender equity in education. This study is the first to document the effectiveness of a multifaceted educational program—delivered to 230 primary schools in rural India—on students’ participation and academic performance, while also examining heterogeneous impacts and sustainability. We find that the program reduced gender gaps in school retention and improved learning during the first year of implementation, but did not yield sustained effects on school attendance or learning, nor did it bridge gender inequalities in school performance over the two-year period.

Available here.

How Many Friends Do You Have? An Empirical Investigation into Censoring-Induced Bias in Social Network Data

In analyzing peer effects in a linear-in-means framework, identifying who interacts with whom is crucial. This suggests the need to collect detailed network data. However, taking a cue from AddHealth, many data-collection efforts only permit resondents to list up to a maximum number of links, leading to censoring and mismeasurement of peer groups. Within a linear-in-means framework, I document the extent of bias due to censoring analytically and by simulation. I then demonstrate that censoring-induced bias is present in empirical applications using data from AddHealth and an experiment in rural Nepal. After documenting the bias, I provide strategies to recover consistent estimates and discuss limitations of these strategies. This paper provides important contributions to the literature on design of network surveys as well as estimation of peer effects in the presence of data limitations.

Available here.

Selected Works in Progress

Equilibrium in Concave Network Formation Games

When Interventions Affect the Network: A Decomposition of Treatment Effects in a Partial Treatment Setting, with Clara Delavallade and Rebecca Thornton