This dissertation develops and applies econometric methods to understand key issues in labor economics. In particular, this dissertation investigates methods to transparently measure quantities of central importance to understanding equity and efficiency in labor markets, like the productivity of workers and firms, the nature of worker's preferences and firm's strategic behavior, and the size of wage markdowns.
The first chapter studies the nature and implications of firm wage-setting conduct on a large online job board for full-time U.S. tech workers. Utilizing granular data on the choice sets and decisions of firms and job seekers, I first develop and implement a novel estimator of worker preferences that accounts for both the vertical and horizontal differentiation of firms. The average worker is willing to pay 14% of their salary for a standard deviation increase in firm amenities. However, at the average firm, the standard deviation of valuations of that firm’s amenities across coworkers is also equivalent to 14% of their salaries, indicating that preferences are not well mdescribed by a single ranking of firms. Following the modern Industrial Organization literature, I use the labor supply estimates to compute the wage markdowns implied by a series of models of firm conduct that vary in the degree to which worker preference heterogeneity gives rise to market power. I then formulate a testing procedure that can discriminate between these models. Oligopsonistic models of wage setting are rejected in favor of monopsonistic models exhibiting near uniform markdowns of roughly 18%. Relative to a competitive benchmark, imperfect competition substantially exacerbates gender gaps in both wages and welfare. However, blinding employers to the gender of job candidates would have negligible effects on wage inequality.
The second chapter proposes a novel framework for conducting causal inference when researchers wish to compare a large number of treatments, as in studies of value-added that aim to quantify heterogeneity in skill, productivity, or preferences across workers, decisionmakers or service providers. Rather than apply parametric assumptions about the data-generating process, the framework I propose leverages only the common assumption that assignment of observations to treatments is unconfounded, and as such leads to "design-based" inferences of causal effects (in contrast to "model-based" approaches). I first illustrate identification of the causal effects of interest when the mechanism governing assignment -- the propensity score -- is known. I then propose a method for estimating the features of the assignment mechanism when the true propensity scores are unknown and must be estimated. In settings with a large number of treatments (e.g. teachers, judges, or firms), the standard overlap assumption that all observations face a strictly positive probability of assignment to every treatment is likely to fail. I therefore propose a propensity score estimator that allows for structural failures of overlap, provide a computational guarantee for the estimation algorithm, and develop a finite-sample bound on the error of the estimator that holds with high probability. Finally, I provide an algorithm for using the estimated propensity scores to optimally trim the sample, such that a traditional notion of overlap is likely to hold on the resultant subsample and treatments can be reliably compared.
The third chapter applies the econometric framework of the second chapter to understand the distribution of productivity in a particular setting: legal defense for indigent individuals. This chapter quantifies the extent to which variation in case outcomes across indigent criminal defendants can be attributed to variation in the quality of their assigned counsel. Applying my estimation framework to data on case outcomes from three Texas counties that assign cases through conditionally randomized “wheel” systems, I find that attorney quality is highly variable. For defendants in felony cases, a one-standard-deviation decrease in attorney quality is associated with a 5.6 percentage-point increase in the probability of incarceration. These findings suggest that outcomes in criminal cases are driven in a nontrivial way by the luck of the draw, undermining the extent to which the criminal legal system can achieve traditional notions of fairness and efficiency. Using estimates of attorney quality, I evaluate the effects of a program that allowed defendants to choose attorneys. Perhaps because attorney quality is difficult to predict using observable characteristics, the program had essentially no effects on aggregate case outcomes, although it did significantly shift the burden of caseloads across attorneys.
