Structured interviews have the highest validity of any method studied (.42) and among the smallest Black–White subgroup differences (d = .23). On the evidence, they are not a trade-off at all. Cognitive ability tests show the largest subgroup difference by a wide margin (d = .79), and their revised validity is .31. The method that defined the trade-off has a weaker case than it did.
- Structured interviews have the highest validity of any method studied (.42) and among the smallest Black–White subgroup differences (d = .23). On the evidence, they are not a trade-off at all.
- Cognitive ability tests show the largest subgroup difference by a wide margin (d = .79), and their revised validity is .31. The method that defined the trade-off has a weaker case than it did.
- Years of job experience is the worst deal available: almost no predictive value (.07) and a substantial subgroup difference (d = .49).
- The trade-off has not vanished. It has moved. Sackett and colleagues conclude the conversation should shift from cognitive ability to the other methods that still carry adverse impact.
For decades, hiring teams were told they faced a trade-off. The methods that predicted performance best also produced the largest differences between demographic groups, so you could have accuracy or you could have a diverse shortlist, but not both. That framing shaped a great deal of practice, and a 2022 re-analysis of the evidence has substantially undermined it.
What the numbers mean
Two figures matter here, and they are easy to confuse.
Validity (r) is how strongly a method's results track later job performance, from 0 to 1.
Subgroup difference (d) is the standardized gap between average scores of two groups. By convention, .20 is small, .50 is moderate and .80 is large. A larger d means a method is more likely to produce adverse impact, meaning selection rates that differ substantially by group.
Neither number tells you a method is fair or unfair on its own. A d of zero on a method that predicts nothing is not an achievement. What matters is the pairing.
The pairing
| Method | Validity (r) | Subgroup difference (d) |
|---|---|---|
| Structured interviews | .42 | .23 |
| Job knowledge tests | .40 | .54 |
| Empirically keyed biodata | .38 | .33 |
| Work sample tests | .33 | .67 |
| Cognitive ability tests | .31 | .79 |
| Integrity tests | .31 | .10 |
| Assessment centres | .29 | .52 |
| Situational judgment (knowledge) | .26 | .39 |
| Conscientiousness, contextualized | .25 | −.07 |
| Interests | .24 | .33 |
| Conscientiousness, overall | .19 | −.07 |
| Unstructured interviews | .19 | .32 |
| Years of job experience | .07 | .49 |
Sources: validity estimates from Sackett et al. (2022); subgroup differences as reported in the same paper, drawn largely from Dahlke and Sackett (2017), with cognitive ability from Roth et al. (2011) and job knowledge from Roth et al. (2003).
Reading the table
Three patterns stand out.
The top-left quadrant is not empty. Structured interviews sit at the highest validity in the set with a subgroup difference less than a third the size of cognitive ability tests. Integrity tests pair .31 validity with a d of .10. Contextualized conscientiousness measures show a slightly negative d. The idea that predictive power requires accepting large group differences does not survive contact with this table.
Some widely used methods are bad on both counts. Years of experience predicts almost nothing and produces a moderate subgroup difference. It remains one of the most common screening criteria in use, usually as a hard filter applied before a human reads anything.
Unstructured interviews are worse than structured ones on both dimensions. Same activity, same hour, lower validity (.19 versus .42) and a larger subgroup difference (.32 versus .23). This is the clearest case in the literature where improving fairness and improving accuracy are the same intervention.
Why structure appears to help
The mechanism is not mysterious. An unstructured interview gives each interviewer latitude over what to ask, what to weight and what to conclude, and that latitude is where similarity bias operates. Research on demographic similarity effects has found that highly structured interviews are substantially more resistant to them, because there is less room for an interviewer's impression to do the work.
Structure does four things that each narrow that room: the same questions for everyone, questions drawn from the job rather than the interviewer's preference, criteria written before the candidate appears, and ratings recorded before discussion. Each one removes a point where an unexamined preference can enter.
It is worth being precise about the claim. Structure does not eliminate bias, and no credible reading of this evidence says it does. It reduces the variance that bias travels in.
What this does not say
That the trade-off is over. Sackett and colleagues are explicit. Reaching an adverse impact ratio of .80 in their modelling still required accepting lower validity, because methods other than cognitive ability also contribute adverse impact. What changed is that cognitive ability is no longer the centre of the problem.
That a low d makes a method safe to use. Adverse impact is measured on actual selection rates in your organization, not on meta-analytic averages. A method with a small average d can still produce adverse impact in a particular process, and a legally defensible process is tested on its own numbers.
That any of this replaces legal advice. Selection methods in the United States sit under the Uniform Guidelines, and comparable frameworks apply elsewhere. These figures inform a decision about method. They do not settle compliance.
What to do with it
If you are choosing between methods, the evidence points one way: structure the interview you are already running. It is the highest-validity method available, it has one of the smaller subgroup differences, it requires no vendor, and the organizations that need it most are the ones already spending the hour.
If you are reviewing an existing process, look first at your hard filters. Years of experience, used as a screen, costs you predictive accuracy and adds adverse impact at the same time. It is unusual to find a practice that is weak on both dimensions and almost universal in use.
Method
This is a narrative evidence review, not a new analysis. It draws on the published Sackett et al. (2022) estimates and the subgroup-difference figures reported alongside them. The authors note that some d values come from non-applicant or mixed samples, which is a real limitation on precision, and the ranking is more reliable than any individual figure.
Sources
- Sackett, P. R., Zhang, C., Berry, C. M., & Lievens, F. (2022). Revisiting meta-analytic estimates of validity in personnel selection: Addressing systematic overcorrection for restriction of range. Journal of Applied Psychology, 107(11), 2040–2068. https://doi.org/10.1037/apl0000994
- Dahlke, J. A., & Sackett, P. R. (2017). The relationship between cognitive-ability saturation and subgroup mean differences across predictors of job performance. Journal of Applied Psychology.
- Roth, P. L., Bevier, C. A., Bobko, P., Switzer, F. S., & Tyler, P. (2011). Ethnic group differences in cognitive ability in employment and educational settings: A meta-analysis.
- Roth, P. L., Huffcutt, A. I., & Bobko, P. (2003). Ethnic group differences in measures of job performance and job knowledge: A new meta-analysis.
How to cite this page
Behavior Hiring. (2026). Does structure reduce bias? Validity and adverse impact by hiring method. https://behaviorhiring.com/research/structure-bias-adverse-impact-hiring-methods/