Becker Outcome Test

What it is

The outcome test comes from Gary Becker’s economics of discrimination. Its core move is to shift attention away from how often a group is subjected to an action (stops, searches, loan reviews) and toward the results of that action (hit rates, repayment rates). Becker’s point is that a disparity in treatment rates is ambiguous on its own, but a disparity in outcomes at the margin can reveal whether the decision-maker is acting on animus rather than on accurate prediction.

Becker built the framework in The Economics of Discrimination (1957) and stated the outcome-test version explicitly for lending in a 1993 column: the way to test banks for bias was not rejection rates but whether loans granted to minorities repaid at the same rate as loans granted to whites. If minority loans repaid at a higher rate, the bank was holding minority applicants to a stricter standard.

The underlying logic

Two different things travel under the single word “discrimination”:

• Taste-based discrimination — the decision-maker has a preference against a group and will act against them even when it is costly or inaccurate to do so.
• Statistical discrimination — the decision-maker uses group membership as a proxy for a predicted trait, to be more accurate, with no animus involved.

The outcome test is built to detect the first kind. An unprejudiced officer trying to maximize successful searches keeps searching a group until the marginal search barely pays off; in equilibrium, that marginal success rate is equalized across groups. A prejudiced officer is willing to search the disfavored group even when the payoff is low — applying a lower threshold of suspicion to them. That extra low-value searching drags their success rate down.

The signature of bias

If a group is searched more but its search success (hit) rate is lower, that is the fingerprint of taste-based discrimination. If the more-searched group’s hit rate is equal or higher, the disparity in search rates is consistent with accuracy-driven enforcement rather than animus.

How it should be used

This is the practical payoff, and the part most often botched in public argument.

People routinely point to a raw disparity in stop or search rates — “group X is stopped at N times the rate of group Y” — and treat that gap, by itself, as proof of discrimination. The outcome test says that inference does not follow. Stop and search rates only tell you where enforcement is directed; they cannot separate “focusing on the right signals to get results” from “acting on prejudice.” Only the outcomes can do that.

The correct procedure:

1. Do not stop at the stop/search rate. A higher rate for a group is a question, not an answer.
2. Compare hit rates (or the relevant success measure) across groups.
3. Interpret: roughly equal hit rates point away from taste-based bias; a systematically lower hit rate for the more-scrutinized group points toward it.

Why raw rates can mislead

Suppose police concentrate searches on a corridor where trafficking is heavily concentrated. The searched population there is not demographically average, so search rates look “disproportionate.” But if searches in that corridor find contraband at the same rate across groups, the disparity reflects targeting the signal, not bias against a group. The outcome test recovers this; a rates-only analysis does not.

What the test does not settle

Using the tool well means marking its edges. The test is a discipline, not a verdict — and it can exonerate or indict depending on the numbers.

• Infra-marginality. The theory is about the marginal decision; the data usually give you the average hit rate. If the distributions of guilt or creditworthiness differ across groups, average hit rates can diverge (or coincide) without implying anything about the margin. This is the central methodological objection (Ayres 2002), and it motivated the later “threshold test” (Simoiu, Corbett-Davies, and Goel 2017), which estimates the actual decision thresholds instead of comparing averages.
• It targets animus, not proxying. Equal hit rates rule out taste-based bias at the margin; they do not show that no one is using race as a predictive proxy. Whether statistical discrimination is itself acceptable is a separate normative question the test cannot answer.
• Upstream base rates. The test takes differential offending or creditworthiness rates as given; it is silent on whether those base rates are themselves downstream of other inequities.
• Data quality. Hit rates require honest, complete recording of searches and outcomes; missing consent-search data or measurement gaming can corrupt the result.

Honest note on the empirical record

The outcome test does not presume the answer, and applied to real policing data it has often cut against the “just efficient enforcement” reading. The largest study to date (Pierson et al., roughly 100 million stops) found the bar for searching Black and Hispanic drivers was lower than for white drivers — lower hit rates, which the test reads as bias against those groups. Other work shows the standard test is fragile enough to flag bias against white drivers in a large share of agencies. So the correct use is genuinely two-directional: run it, compare hit rates, and follow the numbers wherever they land, rather than assuming they will vindicate a prior in either direction.

Bottom line

The outcome test is a discipline against a common error: mistaking unequal attention for unequal treatment. Its correct use is to force the argument down to results — compare hit rates, not stop rates — and to reserve the charge of discrimination for cases where the disfavored group is being acted on at a lower evidentiary bar. Used carefully, it can show that scrutiny was tracking a real signal rather than animus; used carelessly (average hit rates read as marginal ones), it misleads in either direction, which is why the infra-marginality caveat always travels with it.

References

• Gary S. Becker, The Economics of Discrimination, 2nd ed. (University of Chicago Press, 1971 [1957]).
• Gary S. Becker, “The Evidence Against Banks Doesn’t Prove Bias,” Business Week, April 19, 1993.
• John Knowles, Nicola Persico, and Petra Todd, “Racial Bias in Motor Vehicle Searches: Theory and Evidence,” Journal of Political Economy 109, no. 1 (2001): 203–229.
• Ian Ayres, “Outcome Tests of Racial Disparities in Police Practices,” Justice Research and Policy 4, nos. 1–2 (2002): 131–142.
• Camelia Simoiu, Sam Corbett-Davies, and Sharad Goel, “The Problem of Infra-marginality in Outcome Tests for Discrimination,” Annals of Applied Statistics 11, no. 3 (2017): 1193–1216.
• Emma Pierson et al., “A Large-Scale Analysis of Racial Disparities in Police Stops Across the United States,” Nature Human Behaviour 4 (2020): 736–745.