Skip to content
Playbook
AdvancedHRPeopleOpsManager

Selection science — Schmidt & Hunter's 100-year meta-analysis, updated

The single most cited paper in I/O psychology — Schmidt & Hunter's 1998 meta-analysis ranked every predictor of job performance.

10 min read
On this page
60-Second Summary
  • Top predictors (validity coefficients): structured interviews (~0.58), work sample tests (~0.54), GMA tests (~0.51-0.65), integrity tests (~0.46).
  • Bottom predictors: graphology, years of education, age, interest inventories — near zero.
  • Unstructured interviews are weak (~0.20) yet ubiquitous. Structure is the single biggest legal + scientific win you can make this quarter.
  • Combine 2-3 predictors with low overlap (e.g. work sample + structured interview + integrity) for compounding validity.

Hiring is the single most consequential HR decision. The science is unusually clear — and unusually ignored. Schmidt & Hunter ran the meta-analysis; Sackett et al. updated it in 2022. Use it.

What the meta-analysis says

PredictorValidity (r)Notes
Structured interview0.58Standardised questions, scored against a rubric
Work sample test0.54Realistic task from the job itself
GMA (cognitive ability)0.51–0.65Often legally constrained in some jurisdictions
Integrity test0.46Predicts counterproductive work behaviour
Job knowledge test0.48Strong for technical roles
Unstructured interview0.20Ubiquitous, weak, and bias-prone
Years of experience0.18Surprisingly weak above 1-2 yrs
Graphology0.02No predictive value

What changed in the update

Sackett, Zhang, Berry & Lievens (2022) re-estimated the validities after correcting for over-corrections in the original meta-analysis. GMA dropped from ~0.65 to ~0.31-0.42; structured interviews held up; work samples remained strong. The hierarchy mostly stands; the absolute numbers are more modest.

Building an evidence-backed process

  1. Define the job competencies (≤6).
  2. Pick 2-3 predictors that map to them and have low overlap.
  3. Score every candidate on the same rubric.
  4. Use compensatory or multiple-hurdle scoring — decide in advance.
  5. Audit hires + performance after 6 and 18 months. Recalibrate.
The legal layer

Validity science is necessary but not sufficient. Adverse impact testing (4/5ths rule in US) must run alongside. A predictor that works but discriminates is still illegal.

Written by Pawan Joshi.Sources cited inline.
First published 16 Jun 2026See site changelog →