Bias Goes Both Ways

Evidence of racial discrimination does not tell you which way it should go

I believe that the American justice system is biased against African-American defendants. There are several studies to this effect which I believe are credible. Take, for instance, Arnold, Dobbie, and Yang (2017), who are studying discrimination in whether people are granted bail or not. They use a judge IV, which is a method which comes up a lot and is worth going over. The problem with simply comparing defendants who have otherwise similar records is that there might be unobserved differences in likelihood to reoffend which might be apparent to a judge, but not to us, the econometrician. For example, the charge of assault conceals an enormous variation in the severity of the offense – everything from an embarrassing drunken shoving match to a premeditated assault on a rival gang member. It is reasonable for them to take into account such information when making a decision, and if that unobserved information is correctly correlated with race, then we might mistake justified and unjustified discrimination.

What we can use, though, is the severity of a judge across different cases, as some judges are harsher than others. So long as defendants are randomly assigned to the judges, that’ll allow you to suss out the real likelihood of misconduct, and thus whether there is bias. If the rates of observed misconduct after adjusting differ, then there must be inaccurate discrimination.

Arnold, Dobbie, and Hull (2022) is a descendant of this, and agrees with the results, although it makes some intriguing methodological points. (In brief: judge IVs reveal only the effect on people at the margin of being convicted or sentenced or what have you. If you trace out a line on a graph describing the effects, each judge gives you only a small point of the line. We assume that there is one true index of risk underlying all of this. If judges differ in their skills in detecting risk, or in what they regard as dangerous, we are no longer uncovering one index, and we will need quite a lot more work. The upside to this is that they separate out the judges discrimination into two components, an average shift in the likelihood of misconduct, and a noisiness of the signal by race. Since most defendants get bail, noisier signals actually leads to fewer people being held. Note, though, that they are concerned with all disparities, and not whether those disparities are justified.)

When we get into jury selection, we again find that whites-only juries are more likely to convict Black defendants, and imprison them for longer. Anwar, Bayer, and Hjalmarsson (2012) is the first of these studies. Who, exactly, is summoned to the jury pool varies from day to day. Since Sarasota and Lake Counties, the setting of their study, are less than five percent Black, you will very often end up with no African-Americans at all in the jury. When this happens, a given Black defendant is 16 percentage points more likely to be convicted, and a white defendant is 7 percentage points less likely to be convicted. These results are not driven by peremptory challenges to simply strike African-Americans from the pool, because conditional on being included among the potential jurors, Blacks are just as likely to be seated.

They use the same sort of analysis in Houston in Anwar, Bayer, and Hjalmarsson (2022) , although they lack information on the race of the jurors. Instead they possess only the neighborhood from which they come. What they can show is that places with more minorities are underrepresented in their appearances in the jury pool, and that when they are absent, minorities are much more likely to be convicted.

African-Americans are also more likely to be stopped by the police, and their car searched. Vinay Tummarakota surveyed the evidence here, and it’s extremely convincing. Very well-designed studies consistently show that, holding speed and quality of car constant, African-Americans are more likely to get pulled over. Nor are they more likely to have contraband, at least in Texas. Feigenbaum and Miller (2022) show that African-Americans are much more likely to be searched, but no more likely to be found with illegal drugs or guns – and notably, officers that search more are no more or less likely to find contraband.

We can also look at things which are related to the criminal justice system. The Child Protective Services have the ability to, in certain cases, remove the children from the custody of their parents. Using random variation in who is assigned to different cases or to screen calls, Baron, Doyle, Emanuel, Hull, and Ryan (2024) show that otherwise identically risky Black parents are 55% more likely to have their children placed in foster care.

We have established, I hope satisfactorily, that there is a gap between the treatment of whites and Blacks. What I would like to do is urge us to be very careful in how we interpret this – the existence of a gap does not imply that Blacks are sentenced or denied bail unjustly. On the contrary, it can indicate just as well that whites are sentenced to prison enough! Perhaps the system takes away Black children from their parents more often – but if being removed from the home is good for the child, on the margin, perhaps it is Black children who are better off for it.

Consider the Anwar, Bayer, and Hjalmarsson papers. Their results need not indicate that whites discriminate against Blacks – rather, it could indicate that Blacks are, on average, more violent than their peers, and that their co-racialists then discriminate in their favor! Both outcomes are possible given the difference we observe.

Or consider the study on CPS. As they themselves note in the paper, being taken away from your parents is good if they’re bad enough, and at the margin, the best evidence indicates that more children should be separated from their parents. The bias is real, just in practice it runs against the best interests of white Americans. When we consider the effect of incarceration on the children of those incarcerated, we find something similar. Norris, Pecenco, and Weaver (2021) use data from Ohio, and find that having your parents incarcerated reduces your own chance of being incarcerated by 4.9 percentage point, or by 40%. In Colombia, Carolina Arteaga (2023) showed that parental incarceration increases child educational attainment by .78 years, on average. When the actions have impacts on other people, it is not clear which way bias runs – bias against Black parents can be bias in favor of Black children, and harsher policing in minority neighborhoods will disproportionately benefit the law-abiding neighbors.

Similarly, it is not at all clear that we have too many traffic stops. The point is to pull over people who are going too fast and endangering others. It could be that we need to pull over more white drivers, and that balancing out disparities by simply reducing the number of Black drivers pulled over would be bad. The last time we had a wave of depolicing, in 2020, we also had an enormous spike in traffic fatalities. It would not do to repeat it.

I have not come to a definitive answer on all of the world’s problems, and what the optimal level of everything is. I would simply like us to be much more cautious about interpreting evidence of bias as evidence that whatever Black people are experiencing should be reduced. Our goal should be to seek what is optimal, not what is equal.