Monitoring Racial Bias in Traffic Enforcement with Noisy Proxies of Driver Behavior
Kai R. D. Cooper, Gregory Lanzalotto, Haosen Ge, Jacob Kaplan, Scott Desposato, Dean Knox, and Jonathan Mummolo
IN PREPARATION
Abstract
Racial bias has been documented in a variety of policing settings. Traffic stops represent the most common type of encounter in which civilians interact with police, which makes this an important setting for investigation of police behavior. Evidence-based debates in this area frequently rely on case-specific benchmarks to evaluate the racial distribution of police stops. Often, benchmarks such as per-capita demographics or not-at-fault drivers in vehicle collisions are used. However, these measures only approximate the intended comparison because they fail to (i) precisely define the causal contrast, (ii) accurately describe the nature of drivers visible to officers, (iii) explain disparities due to potentially biased selection decisions by officers. In this work we formally define racial bias as a violation of an individual fairness criterion: race should be uninformative for officer decision-making, conditional on the behavior of the civilian, i.e. whether or not they are driving dangerously. This latter, however, is unmeasurable. To address this issue, we build on the growing proximal causal inference literature to make use of cameras, crashes and checkpoints as negative outcome controls for identification, under certain assumptions which may be probed via a sensitivity analysis. We will extend the method's viability to a host of imperfect data settings which plague policing, e.g. when race is mismeasured or unavailable. We finish with applications of the approach to a variety of U.S. police jurisdictions.
Learning From Imperfect Research Designs: Automating Causal Inference When Classic Assumptions Fail
Kai R. D. Cooper, Guilherme Duarte, Luke Keele, Dean Knox, and Jonathan Mummolo
SUBMITTED TO THE AMERICAN JOURNAL OF POLITICAL SCIENCE (2026)
Abstract
Social science has developed an expansive design-based toolkit for causal inference, but key assumptions often fail in real-world settings. Partial identification offers an alternative: researchers can learn as much as possible through sharp bounds while transparently acknowledging limitations of data and design. We propose methodological improvements to automated partial identification that make it viable for applied social-science research, including new approaches for quantifying uncertainty, adjusting for covariates, handling continuous variables, assessing the consequences of relaxing or falsifying assumptions, and interpreting why bounds are narrow or wide. We then replicate and extend published studies spanning several causal designs to show how these advances deepen our understanding of empirical robustness. In some applications, implausible assumptions directly drive key causal claims; in another, they are inconsistent with observed data. Among other results, we present updated findings on counterinsurgency violence, voter habit formation, and racial bias in policing.
Evaluating the Validity and Robustness of Instrumental Variable Analyses
Kai R. D. Cooper, Guilherme Duarte, Luke Keele, Dean Knox, and Jonathan Mummolo
IN PREPARATION
Abstract
Instrumental-variable (IV) designs are widely used across numerous fields to estimate causal effects when the relationship between treatment and outcome is confounded, exploiting as-if randomized encouragements that nudge units into treatment. The validity of these designs rests on several assumptions that are often regarded as untestable, including monotonicity, the assumption that no units defy the encouragement, and exclusion, the assumption that the instrument does not directly affect the outcome. Drawing on extant but overlooked analytic results along with recent advances in automated partial identification, we present falsification tests and new sensitivity analyses that empirically evaluate the validity of these assumptions and the robustness of inferences to violations.Replicating and extending published examples, we show our techniques are flexible to the idiosyncrasies of applied settings by sharply bounding causal estimands in situations where key IV assumptions are believed to fail.
Did the Elizabeth Line Improve Air Quality in London?
Kai R. D. Cooper, Liang Ma, and Daniel J. Graham
IN PREPARATION
Abstract
Assessing the causal effects of major engineering interventions is difficult because implementation is nonrandom, spillovers are unavoidable, and suitable controls are often unavailable. We develop a temporal regression discontinuity (TRD) design for time-series settings in which treatment begins at a known date and may induce effects that evolve gradually. Within a potential-outcomes framework, we define short-run and finite-horizon effects, and introduce an approach for long-run effect estimation based on the derivative of the estimated effect curve. To accomplish this, we make use of Gaussian process regression and show how it offers advantages over similar methods in the literature. We apply the method to hourly air-pollution data from 77 London monitoring sites to study the opening of the Elizabeth Line on 24 May 2022. Our findings suggest that transport investments can yield gradual, uneven environmental benefits that are missed by discontinuity designs focused only on immediate effects.