UC Berkeley

This dissertation studies examples of applied econometrics for causal inference in labor and development economics. One of the fundamental problems in applied fields of economics is causal inference. Merely observing that event B occurred after event A is not enough to claim that A caused B. The field of economics, and the social sciences in general, are limited by ethics and practicality in their ability to conduct randomized field experiments, the gold standard for causality in other fields. Several statistical methods have been devised to obtain causal estimates from "natural" or "quasi" experimental settings--settings where plausibly exogenous variation in a treatment effect of interest can be found and exploited to produce an unbiased estimate of causal effects. Some of these methods include panel data with fixed effects, nearest-neighbor matching, and regression discontinuity. This dissertation explores applications of these econometric methods, as well as an actual randomized controlled trial, in issues of labor and development economics.

The first chapter uses panel data, and causal estimates are identified using a series of fixed effects to control for unmeasurable characteristics that could be correlated with both dependent and independent variables. The subject matter is the recruiting task of the United States military, which is the largest employer in the nation and spends over $4 billion each year to recruit roughly 200,000 new soldiers to maintain its troop levels. This recruiting task has become more expensive since the beginning of the wars in Iraq and Afghanistan. I use a detailed new dataset of all US military applicants over several recent years and find that deaths in Iraq of US soldiers had a significant deterrent effect on recruiting in the home county of the soldiers who were killed. The deterrent effect of local deaths is significantly larger than the deterrent from a death from outside the county. The deterrent exhibits significant heterogeneity across characteristics of deaths, recruits, and locations. Deaths from Iraq decrease recruiting, while deaths from Afghanistan actually increase recruiting. Recruits with higher test scores are more deterred by deaths, and the deterrent is larger and more negative in less populous and more racially diverse counties, but is significantly smaller and in many cases even positive in counties that voted for George W. Bush in the 2004 presidential election. The findings provide strong evidence that recruits are over-emphasizing local

information and have war-specific tastes and preferences that makes enlistment decisions more complicated than a full-information utility-maximization model of risk and monetary compensation would predict.

The second chapter uses nearest-neighbor matching techniques to look at performance of Major League Baseball players after they win awards in order to shed light on the more general question of how rational agents perform after they have been rewarded for good behavior up to that point. Comparing individual player's performance after winning major awards to their performance before winning shows that although players do perform significantly better in the year in which they win the award, performance after the award is generally indistinguishable from pre-award performance. Matching methods based on both baseball writer voting and performance statistics also indicate the likely absence of any sort of "curse" from winning awards for the winners themselves, their teams, and their teammates.

The third chapter, which is co-authored work with Michael Kremer and Edward Miguel, uses data from a randomized controlled trial, the Girls Scholarship Program (GSP), as well as the Kenya Life Panel Survey (KLPS) to conduct three types of analysis of bursary programs. We evaluate the effect of different targeting rules for secondary school scholarships, we estimate the impact of attending a primary school that took part in a scholarship program, and we estimate the effect of winning a scholarship from the program. Giving scholarships based on KCPE alone would lead to under representation of children whose parents have no secondary education and girls relative to their proportion of the population. Distributing the scholarships to the top students in each school as opposed to each district does little to alleviate this discrepancy. Analysis of the medium-run impacts of the Girls Scholarship Program, gave largely inconclusive but suggestive evidence that there were moderate benefits from attending a scholarship program school on the order of one half of the benefits observed in the original study held immediately after the scholarship program. The evidence indicates that scholarship winners did not benefit greatly from the award itself.