Unbiased judges can appear biased because other actors sort cases beforehand

Evidence

There are other important issues that complicate the investigation of judicial bias. First, there are many actors in the criminal justice system (CJS) other than judges, including prosecutors and defense attorneys, and the actions of these actors can impact the decisions of judges. Put differently, even if judges are completely unbiased with respect to race and gender, their decisions may resemble bias as an artifact of these other actors’ decisions and actions that took place before the case arrived at the docket of the judge. Furthermore, the decisions of these actors are interrelated. For example, if a particular judge is known to be racially biased against a particular group of defendants, defense attorneys may be more likely to plea bargain with prosecutors to avoid the case being adjudicated by the judge. In this scenario, the composition of the cases that are brought in front of this judge, as well as their case characteristics, would be different in comparison to counterfactual scenarios in which the judge was not racially unbiased. This type systematic “sorting” is produced by the defendant (or his/her defense attorney). Systematic sorting can also be generated by other non-judge actors. For example, prosecutors have substantial discretion regarding whether to press charges and what type of charges to press against a defendant. If this decision is racially biased, the case files that end up in front of judges will have gone through sorting based on the race or gender of the defendants. This would confound the analysis of judges’ decisions, if sorting is not adjusted for. — Joseph Avery, Jim Sidanius, Joel Cooper, Gregory Davis, Amanda N., Bias in the Law, 775

Categorization

• XREF: Connects to selection bias / collider bias discussions in causal inference and to the omitted-variable logic the chapter addresses in the next paragraph.
• QUESTION: How do current empirical studies actually adjust for this sorting/selection effect — which statistical techniques specifically?