Working Paper: NBER ID: w27802
Authors: Ivan A. Canay; Magne Mogstad; Jack Mountjoy
Abstract: The decisions of judges, lenders, journal editors, and other gatekeepers often lead to significant disparities across affected groups. An important question is whether, and to what extent, these group-level disparities are driven by relevant differences in underlying individual characteristics, or by biased decision makers. Becker (1957, 1993) proposed an outcome test of bias based on differences in post-decision outcomes across groups, inspiring a large and growing empirical literature. The goal of our paper is to offer a methodological blueprint for empirical work that seeks to use outcome tests to detect bias. We show that models of decision making underpinning outcome tests can be usefully recast as Roy models, since heterogeneous potential outcomes enter directly into the decision maker's choice equation. Different members of the Roy model family, however, are distinguished by the tightness of the link between potential outcomes and decisions. We show that these distinctions have important implications for defining bias, deriving logically valid outcome tests of such bias, and identifying the marginal outcomes that the test requires.
Keywords: No keywords provided
JEL Codes: C21; C26; C51; J15; J16; J71; K14; K42
Edges that are evidenced by causal inference methods are in orange, and the rest are in light blue.
| Cause | Effect |
|---|---|
| decision-maker bias (D91) | outcomes experienced by different groups (I24) |
| biased decision-makers (D91) | favoritism (J71) |
| favoritism (J71) | disparities in outcomes across groups (I14) |
| bank's loans to approved black applicants yield higher profits than those to approved white applicants (G21) | racially biased lending practices (J15) |
| outcome tests (C52) | reveal whether a decision-maker is biased (D91) |
| generalized Roy model vs extended Roy model (C52) | understanding how potential outcomes influence decisions (D91) |
| extended Roy model (C59) | more accurate identification of bias (D91) |
| judge's marginal white defendants show a higher mean effect of release on misconduct than their marginal black defendants (K40) | racial bias against black defendants (J15) |
| mean effects are equal (C33) | judge is unbiased (K40) |
| flexibility in modeling (C52) | logical invalidity in outcome tests (C52) |