Joshua Shea

I develop a framework to detect and measure bias amid sample selection and statistical discrimination, and apply this framework to study racial bias in police traffic searches. I model the search decision with a threshold model and allow the threshold to be random. This allows the direction and intensity of bias to depend on the officer’s belief of the probability that a driver carries contraband. This framework also allows me to evaluate each officer separately, thereby allowing for unrestricted heterogeneity in officer search preferences and beliefs. Sharp bounds on various measures of bias can be derived using bilinear programs. I use this framework to evaluate 50 officers in the Metropolitan Nashville Police Department and find 6 officers to be biased. For each of these officers, I construct sharp bounds on how search rates for minority drivers would change if they were treated as white drivers, and vice versa. The estimates suggest the intensity of bias depends on the officer’s belief of the probability that a driver carries contraband.

Instrumental variable (IV) strategies are widely used to estimate causal effects in economics, political science, epidemiology, psychology, and other fields. When there is unobserved heterogeneity in causal effects, standard linear IV estimators only represent effects for complier subpopulations (Imbens and Angrist, 1994). Marginal treatment effect (MTE) methods (Heckman and Vytlacil, 1999, 2005) allow researchers to use additional assumptions to extrapolate beyond complier subpopulations. We discuss a flexible framework for MTE methods based on linear regression and the generalized method of moments. We show how to implement the framework using the ivmte package for R.

We provide large sample distribution theory for support vector regression (SVR) with l1-norm, along with error bars for the SVR regression coefficients. Although a classical Wald confidence interval obtains from our theory, its implementation inherently depends on the choice of a tuning parameter which scales the variance estimate and thus the width of the error bars. We address this shortcoming by further proposing an alternative large sample inference method based on the inversion of a novel test statistic which displays competitive power properties and does not depend on the choice of a tuning parameter.