Structuring Coding for "State-" and "District-Level" Controls

While this discussion on the StataList (click here) does not specifically speak to the legal context per se, the logic of what is being discussed carries over directly to law-related research that seeks to include, e.g., district and circuit court control variables in a single regression specification. The relevant issues discussed span from how (right-side) control variable ordering in the coding can matter to which level of clustering makes the most sense.

Defendant/Prosecutor Cross-Racial Pairing Effects on Convictions

As racial disparities in the criminal context persist, studies of factors that might plausibly contribute likewise persist. One such inquiry, arising most notably in the death penalty context, considers the potential impact of cross-race defendant and victim effects on prosecutors' decisions to seek the death penalty. Relatively under-studied, however, includes potential cross-race defendant and prosecutor effects on conviction rates.

In a recent paper, Do Prosecutor and Defendant Race Pairings Matter? Evidence from Random Assignment, CarlyWill Sloan (USMA--econ) exploits quasi-random case assignments to prosecutors and misdemeanor case data (N=75,666) drawn from New York County's Early Case Assessment Bureau. Key results indicate variation across race and crime types and note "significant cross-race effects on conviction outcomes for property crimes, but not for drug, person, or other offenses. Specifically, Black defendants charged with property crimes are convicted at a rate 5 percentage points higher when assigned to a white prosecutor rather than a Black one (65 percent vs. 61 percent). White defendants, by contrast, show similar conviction rates regardless of prosecutor race."

Calculating "Win" Probabilities

A user-written Stata command available on SSC, [winprob], may interest legal scholars examining outcomes from two separate groups (e.g., treatment, control) in randomized controlled trials (RCTs). As the technical documentation explains, "winprob performs a non-parametric test for the win probability (also called the Wilcoxon-Mann-Whitney test probability, c-statistic, AUC, probabilistic index). The command returns the point estimate and associated confidence interval, standard error and hypothesis test for the win probability." In less technical language, "[winprob] calculates the non-parametric win probability of one group over another for a single outcome. The outcome can be ordinal or continuous. This is the probability of one group scores being at least as high or higher than those in the other group." More information, including on how to download, is found here.

Panel Data & Heteroskedasticity, Autocorrelation, and Cross-Sectional Dependence Concerns

Despite some obvious comparative advantages, panel data sets invite traditional complications as well. These complications include (but are not limited to) heteroskedasticity, autocorrelation, and cross-sectional dependence issues. How to manage such complications can be more "art" rather than "science," and is often specification- (and data-) specific. For a few recent as well as slightly older discussions of and resources on various approaches for navigating through panel data set challenges, click here, here, and here.