Examining Agent Behavior in the Criminal Justice System

Abstract

This dissertation studies the role of the criminal justice system in shaping socioeconomic outcomes and inequality, examining both how personnel policies and environments influence the behavior of judicial agents and how social policy interacts with the justice system to shape intergenerational economic well-being.

In Chapter I, I study the consequences of incomplete contracts in the high-stakes, multitasking setting of policing. In my context, highway patrol troopers face salient traffic enforcement targets, but must balance their effort on enforcement production with completing other non-enforcement responsibilities. While the enforcement target induces trooper effort, it simultaneously distorts trooper behavior and generates a range of socially suboptimal outcomes, including lower quality enforcement, disparities unwarranted on the basis of collision risk, and delayed completion of non-enforcement responsibilities. Given significant trooper-specific heterogeneity, I develop an approach to optimally assign troopers to locations that reduces the negative externalities produced by the existing incentive scheme. In contrast, alternative personnel policies which ignore this heterogeneity improve only a subset of outcomes.

In Chapter II, co-authored with Nikhil Rao, we show how to use natural experiments and a binary instrumental variable strategy to measure discrimination, adjusted for group differences in unobserved potential outcomes. Our approach does not require random assignment to decision-makers, a prerequisite for existing techniques. We study discrimination in two settings, using distinct natural experiments. First, we measure racial discrimination in misdemeanor prosecution in Washington using a difference-in-difference design, generated by a county budget cut. Before the cut, which reduced prosecution rates, we find no evidence of discrimination in prosecution conditional on unobserved potential recidivism. Afterwards, white defendants were more likely to be prosecuted than minority defendants. The gap is driven by prosecutors dropping low quality cases, which were more common among minority defendants. These patterns suggest prosecutors attenuated disparities created in prior stages of the criminal legal system. Second, we study socio-economic discrimination in student grade promotion in Michigan public schools using a regression discontinuity design. Economically disadvantaged students near a test score cut-off were promoted at lower rates than non-disadvantaged students, even conditional on unobserved academic ability.

In Chapter III, co-authored with Michael Mueller-Smith, Kevin Schnepel, and Caroline Walker, we study the lifetime banning, as introduced by United States Public Law 104-193, of individuals convicted of felony drug offenses after August 22, 1996 from ever receiving future SNAP benefits. Using a regression discontinuity design that leverages CJARS criminal history records with federal administrative and survey data, we estimate the causal impact of safety net assistance bans, finding significant reductions in SNAP benefit take-up, which creates unintentional spillovers to spouses and children and persist long after ban revocations occurred. While we observe limited changes to other adult outcomes, children's short- and long-run outcomes worsen, especially those impacted at young ages.

Finally, in Chapter IV, I study the efficiency and equity considerations of expanding judicial latitude using administrative court data from Oregon and a 1997 reform which permitted judges to consider new crime risk in addition to flight risk when making pretrial release decisions. Leveraging a regression discontinuity design, I show that marginal defendants are less likely to be released post-reform, with no corresponding reduction in pretrial rearrest rates. Extending existing tools to estimate disparate impact, I show that the reform widened the cross-judge unwarranted racial release gap as a result of differential changes in judicial decision-making errors.