Recent Publications

In 2017, the Trump Administration restored local law enforcement agencies’ access to military weapons and some other types of surplus military equipment (SME) that had been prohibited by the Obama Administration. The Justice Department background paper used to justify this decision cited two papers published by the American Economic Association. These papers used SME data collected with a 2014 Freedom of Information Act request and concluded that SME, supplied to local law enforcement by the federal government via the 1033 Program, reduces crime. Here we show that the findings of these studies are not credible due to problems with the data. Using more detailed audit data on 1033 SME, we show that the 2014 data are flawed and that the more recent data provide no evidence that 1033 SME reduces crime.

Police use of force bears on central matters of political science, including equality of citizen treatment by government. In light of recent high-profile officer-involved shootings (OIS) that resulted in civilian deaths, we assess whether, conditional on a shooting, a civilian’s race predicts fatality during police-civilian interactions. We combine Los Angeles data on OIS with a novel research design to estimate the causal effects of fatal shootings on citizen-initiated contact with government. Specifically, we examine whether fatal OIS affect citizen contact with the municipal government via use of the emergency 911 and nonemergency 311 call systems in Los Angeles. We find no average effect of OIS on patterns of 911 and 311 call behavior across a wide range of empirical specifications. Our results suggest, contrary to existing evidence, that OIS, in and of themselves, do not substantively change civic behavior, at least not citizen-initiated contact with local government.

Working Papers

We present a theoretical model predicting racially biased policing produces 1) more use of potentially lethal force by firearms against Black civilians than against White civilians and 2) lower fatality rates for Black civilians than White civilians. We empirically evaluate this second prediction with original officer-involved shooting data from nine local police jurisdictions from 2005 to 2017, finding that Black fatality rates are significantly lower than White fatality rates, conditional upon civilians being shot by the police. Using outcome test methodology, we estimate that at least 30% of Black civilians shot by the police would not have been shot had they been White. We also show that an omitted covariate three times stronger than our strongest included covariate would only reduce this estimate to 18%. Additionally, such an omitted covariate would have to affect Black fatality rates and not Hispanic fatality rates in order to be consistent with the data.

Outcome tests are a method for detecting bias in selection procedures using data on selected units. We use a principal stratification approach to establish lower bounds on this bias for general outcomes. We show that the Knox, Lowe, and Mummolo [2020] bound is sharp for a binary outcome, and weaken the assumptions required for that approach. We also show that the analogous bound for non-binary outcomes is not sharp and provide sharper lower bounds. We illustrate these methods with a re-analysis of the data in Anzia and Berry [2011] on the delivery of federal spending by male and female members of congress. Using Anzia and Berry [2011] data and assumptions, we find that at least 19.9% of men elected, would not have been elected, had they been women.

Outcome tests are a method comparing rates of observed outcomes across selected groups to evaluate bias in decision making processes. Building on the lower bound estimand from Knox et al. (2020), I derive a lower bound in terms of relative risks and develop a sensitivity analysis to weaken the selection-on-observables assumption. Additionally I develop a covariate adjusted sensitivity analysis to assess sensitivity to unmeasured covariates. I am able to estimate a bias adjusted outcome test robust to both measured and unmeasured confounders. Applying this outcome test and sensitivity analysis to settlement and roster data from the Chicago Police Department (1985-2016), I find evidence for gender bias in hiring. I estimate at least 7.9% men would not have been hired had they been women.

Courses

September 16, 2019 Finishing up Chapter 3 of R4DS Short review: options for visualization in ggplot2: color, fill, shape, size, …

Installation Install R software from CRAN Install RStudio Andrew Heiss has on his blog install instructions for both Windows and OS …

Recent Posts

We’ll follow along with the example used in Chapter 3 of Mostly Harmless. You can download the 1980 census from …

Using birthweight data from the MASS library we will work through evaluating different modeling specifications: library(MASS) # birthwt …

Nonparametric Regression Given the usual conditional expectation function \[E[y_i | \mathbf{x_i} = \mathbf{x}] = m(x)\] we can estimate …

Matrices in R! We can start by creating a 3X3 matrix and filling it with the numbers 1 to 9: A <- matrix(1:9, nrow = 3, ncol = 3, …