fairness-and-bias-analysis

name	fairness-and-bias-analysis
description	Use when evaluating fairness and bias of a selection procedure — distinguishing the several meanings of "fairness," testing for predictive bias (differential prediction via moderated regression), and examining measurement bias (DIF, item sensitivity review). Covers what subgroup-mean differences do and don't imply, when bias analyses are warranted, and the statistical pitfalls. Triggers: "adverse impact vs bias", "differential prediction", "predictive bias", "measurement bias", "DIF analysis", "is the test fair / biased", "subgroup differences", "item sensitivity review".
version	1.0.0
author	OpenMatter-Network
license	MIT
category	research
tags	["Community","io-psychology","personnel-selection"]
permissions	[]

Fairness and bias analysis

Two distinct ideas that are routinely conflated. Keep them separate.

Fairness is a social concept, not a single statistic

"Fairness" has no single agreed definition (statistical, psychometric, or social). Recognized meanings include:

• Equal group outcomes (e.g., equal passing rates). The Standards reject this as the definition of fairness: group outcome differences alone do not indicate bias — though they should trigger heightened scrutiny for possible bias.
• Equitable treatment of all examinees (testing conditions, access to practice materials, feedback, retest opportunities, reasonable accommodation, mode of administration).
• Comparable access to the construct — accessible testing so all candidates can show their standing without being advantaged/disadvantaged by construct-irrelevant characteristics (age, race, ethnicity, gender, SES, cultural/linguistic background, disability).
• Lack of bias.

There is broad agreement that equitable treatment, access, bias, and scrutiny when subgroup differences appear are important — but no agreement that "fairness" can be uniquely defined in terms of any one of them.

Bias is a technical concept — two forms

Bias = systematic error that differentially affects the performance of different subgroups.

• Predictive bias — slope and/or intercept of the predictor→criterion regression differs across groups (a predictor–criterion relationship issue).
• Measurement bias — construct-irrelevant variance producing systematically higher/lower scores for a subgroup (a score issue, for predictors or criteria).

Crucial point on consequences

A subgroup-mean difference (adverse impact) is a negative consequence, but it is evidence against validity only if it traces to a measurement property of the procedure (i.e., bias). If the group difference on the procedure mirrors a real difference in the work-relevant outcome (i.e., no predictive bias), the consequence is a policy issue for the user, not a validity defect.

Predictive bias / differential prediction

Test via moderated multiple regression (MMR): regress the criterion on the predictor, subgroup membership, and their interaction. Slope and/or intercept differences signal predictive bias. MMR is preferred over comparing separate subgroup correlation coefficients.

• Frame the question as "is the subgroup's performance underpredicted?" — only underprediction signals bias against that group. Simply knowing slopes/intercepts differ doesn't answer it. (In U.S. cognitive-ability research, slope differences are rare; when intercept differences occur they typically take the form of overprediction of minority performance — Schmidt, Pearlman, & Hunter, 1980; and corrected analyses, e.g., Berry & Zhao, 2015, still find little underprediction.)
• Consider effect sizes as well as statistical significance (Nye & Sackett, 2017; Dahlke & Sackett, 2017).

Technical cautions (predictive bias)

1. Analyze predictors as operationally used (e.g., test the composite when selection uses a composite, not each test separately).
2. A confident, unbiased criterion is a prerequisite.
3. Statistical power is a chronic problem — small total/subgroup samples, unequal subgroup sizes, range restriction, and predictor unreliability all reduce power to detect slope/intercept differences.
4. Check the homogeneity-of-error-variance assumption; use alternative tests when it's violated.
5. Use an unbiased estimate of the intercept difference and operational validity parameters (not observed parameters).

Predictive bias and mean differences can exist independently; analyze predictive bias when there's compelling reason to question whether predictor and criterion relate comparably across subgroups and appropriate data exist. Where relevant research exists, generalized evidence can inform the question.

Measurement bias

Construct-irrelevant variance raising/lowering scores for a subgroup — hard to detect because it requires comparing an observed score to a true score. Approaches:

• Item sensitivity review — diverse reviewers examine items (and instructions to candidates and scorers) for language/content that could carry differing meaning across subgroups or be demeaning/offensive. Value depends on content; use is a matter of professional judgment.
• Differential item functioning (DIF) — identifies items on which members of different subgroups with the same total score (or same IRT true score) perform differently. Notes:
  • Needs large samples for stable results.
  • Domains where DIF is common have rarely shown sizable, replicable DIF (Sackett et al., 2001); for cognitive tests it's common to find roughly equal numbers of items favoring each subgroup, netting to little test-level bias.
  • DIF is not a routine/expected part of selection development; explore it when appropriate data exist. Especially useful in cross-cultural / linguistically different testing.

Pitfalls

• Equating adverse impact with bias, or "no bias" with "fair."
• Testing each component instead of the operational composite.
• Running underpowered bias analyses and reading a null as "no bias."
• Comparing subgroup correlations instead of using MMR.
• Treating any slope/intercept difference as bias without asking about underprediction direction.

Checklist

• "Fairness" meaning(s) at issue named explicitly
• Subgroup differences treated as a scrutiny trigger, not a verdict
• Predictive bias tested with MMR on the operational predictor/composite
• Underprediction direction (not mere difference) interpreted; effect sizes reported
• Power, range restriction, unreliability, error-variance homogeneity addressed
• Unbiased parameter estimates used
• Measurement bias considered (item sensitivity review and/or DIF) where data/justification exist
• Equitable treatment and access to the construct addressed (see accommodations skill)

See also

criterion-related-validation · selection-decisions-and-scoring (composites & subgroup tradeoffs) · candidate-accommodations (equitable treatment/access) · internal-structure-validation · technical-validation-report

Source: Principles (5th ed., 2018), "Fairness and Bias."