When ‘less is more’: How statistical discrimination can decrease predictive accuracy

Felipe A. Csaszar; Diana Jue-Rajasingh; Michael Jensen

doi:10.1287/orsc.2022.1626

Citation: Csaszar, F. A., Jue-Rajasingh, D., and Jensen, M. (2023). When ‘less is more’: How statistical discrimination can decrease predictive accuracy. Organization Science 34(4) 1383–1399. doi:10.1287/orsc.2022.1626

Cite

BibTeX file

Paper highlights

Statistical discrimination theory argues that a decision-maker seeking accuracy should use group membership when it predicts an outcome, even if group membership does not cause that outcome. This paper shows how that argument breaks once the decision-maker is fallible. Under many realistic conditions, ignoring race, gender, or another noncausal group cue produces more accurate predictions.

The strongest mechanism is inconsistency. A theoretically optimal predictor can assign exactly the right weight to each cue every time. Human decision-makers cannot. Even small deviations in how they use group information can erase its limited predictive benefit and make the simpler, nondiscriminatory rule more accurate.

How the model works

The formal model compares predictions made with a causal cue, a correlated but noncausal group cue, both cues, or neither. It varies whether cues are observable, how strongly they correlate, environmental uncertainty, decision-maker inconsistency, and bias in interpreting the causal cue.

Main results

Using group membership improves accuracy only under a narrow set of conditions.
Across much of the modeled parameter space, the most accurate rule omits the group cue.
Inconsistency makes the less-is-more result especially strong because extra information creates another weight the decision-maker can misapply.
Even where statistical discrimination improves prediction, the gain is usually small in the model.

Why it matters

Group membership adds a cue that a fallible decision-maker can estimate badly or apply inconsistently; its theoretical information value may be smaller than the error it introduces.
A nondiscriminatory rule can therefore be more accurate. The legal, social, and moral harms of discrimination do not depend on whether group information improves prediction.
Better measurement of causal individual attributes further narrows the conditions under which group averages aid prediction.

How to use this paper

Cite this for

A formal model showing how statistical discrimination can decrease predictive accuracy.
The role of decision-maker inconsistency and diagnostic bias when a noncausal group cue is available.
A less-is-more result in which omitting information can improve prediction for fallible decision-makers.

Useful for teaching

Why more data does not automatically mean better prediction.
The difference between a cue being correlated with an outcome and a person being able to use that cue accurately.
How accuracy, legal standards, and moral objections to discrimination are related but distinct questions.

Careful claim

Under the modeled conditions, omitting a noncausal group cue can increase predictive accuracy for fallible decision-makers; the normative case against discrimination does not depend on this accuracy result.

Abstract

Discrimination is a pervasive aspect of modern society and human relations. Statistical discrimination theory suggests that profit-maximizing employers should use all the information about job candidates, including information about group membership (e.g., race or gender), to make accurate predictions. In contrast, research on heuristics in psychology suggests that using less information can be better.

Drawing on research on heuristics, we show that even small amounts of inconsistency can make predictions using group membership less accurate than predictions that do not use this information. That is, whereas statistical discrimination theory implies that better predictions can be achieved by using all available information about an individual (including group characteristics that may be correlated with but do not cause performance), our model shows that using all available information only improves predictive accuracy under a very specific set of conditions, thus suggesting that statistical discrimination often results in worse predictions.

By understanding when statistical discrimination improves or worsens predictions, our work cautions decision makers and uncovers paths toward reducing the occurrence of situations in which statistical discrimination benefits predictive accuracy, thus reducing its pervasiveness in society.