What predicts job performance? The research, in plain words

How researchers measure "predicts performance"

A selection method is useful if people who score higher on it also do better in the job later. Researchers express this as a correlation, usually reported as a number between 0 and 1. A 0 means the method tells you nothing about later performance; a 1 would mean perfect prediction, which no method reaches. In this field, about .30 is generally treated as useful, and .40 as strong.

Because any single study is small and noisy, the field relies on meta-analysis: pooling hundreds of studies into one estimate per method. The best-known pooled estimates were published by Schmidt and Hunter in 1998 and were quoted for twenty years. In 2022, Sackett, Zhang, Berry and Lievens re-examined the main meta-analyses behind those figures. The older estimates applied a large correction for the fact that studies only see the people who were hired, a narrower group than the applicants; in many studies that correction was too large or did not apply, which inflated the numbers. Most 2022 figures are lower, by about .10 to .20, while the order of the methods stayed broadly similar. They are now widely used as the reference point.

The 2022 ranking

Rounded figures for predicting overall job performance, from the 2022 re-analysis:

• Structured interviews: about .42. Everyone gets the same job-related questions, and answers are scored against a written guide.
• Job-knowledge tests: about .40. Tests of what the person knows about the work itself.
• Empirically keyed biodata: about .38. Application-form facts scored with a key built from past hires; rarely practical for a small employer.
• Work samples: about .33. The candidate does a piece of the real work and is scored on the result.
• General mental ability tests: about .31, lower than the .51 reported in 1998, and with the largest Black–White score difference of the methods in the 2022 table (US data), which is why many employers avoid relying on them.
• Integrity tests: about .31.
• Situational judgment tests: about .26. Written work dilemmas with set answer options.
• Unstructured interviews: about .19. A free-form conversation, scored by overall impression.
• Conscientiousness (personality questionnaire): about .19.
• Years of job experience: about .07, close to nothing.

Two things stand out. First, the gap between a structured and an unstructured interview is large: adding structure roughly doubles the predictive value of the same hour. Second, most of the methods near the top look at the work itself, either by asking about it in a disciplined way or by watching the person do it.

Why structure matters so much

An unstructured interview feels informative. Research says that feeling is mostly an illusion: in a 2013 experiment, students predicted classmates' grades from their past record, and for some classmates they also held an unstructured interview. The interview made their predictions worse, and they formed impressions just as confidently when, without their knowing, the answers they heard were random. The interview added noise, and the interviewer did not notice.

Structure fixes three problems at once. Every candidate faces the same questions, so differences in answers are differences between people, not between interviews. Answers are scored against written descriptions of good and poor answers, so the score is less about the interviewer's mood. And several interviewers can score the same answer and compare, which makes the process checkable afterwards. A 2014 review of twenty years of structured-interview research covers how to build one: what counts as structure, how to write the rating scales, and how far to allow follow-up questions.

Why work samples and simulations help

A work sample asks the candidate to do the job, or a slice of it, and scores what they produced. Its strength is obvious: it measures the behaviour you care about directly, instead of a report about it. Its weakness is cost. A real work sample for a support role means a real customer, a real system and an afternoon of a manager's time, so most teams skip it.

Simulations sit between a work sample and a situational judgment test. The candidate is placed in a realistic situation and must act, but the situation is staged, so it can be given to every candidate in the same form at low cost. A 2011 study of 196 applicants in high-stakes medical selection found that both a written situational judgment test and a higher-fidelity simulation (an assessment centre) predicted later job performance and added information beyond a knowledge test, and the assessment centre added information beyond the written test. This fits the idea that the closer a simulation is to real behaviour (the person must actually say or do something, not pick A, B or C), the more it behaves like a work sample.

What the numbers do not tell you

• No method is a verdict. A correlation of .40 leaves most of the variation in performance unexplained. The research supports using these methods as inputs to a human decision, not as the decision.
• Combining methods helps, but less than you would hope. Methods overlap. A structured interview plus a work sample is better than either alone, but not twice as good.
• The numbers are averages. A badly built structured interview can be worse than a good unstructured one. The method only earns its number when it is done properly: job-related content, the same for everyone, scored against a written guide.
• Fairness is separate from validity. A method can predict well and still produce large differences between groups. Check both.

How a small team can use this

1. Decide what the job needs before you look at anyone. Three to five skills, in plain words, each with a sentence describing a strong and a weak answer.
2. Give every candidate the same job-like task early: a short work sample or simulation that shows the skills in action. Score it against your guide.
3. Interview with structure. Same questions for everyone, scored with the same guide; use the task results to decide what to probe.
4. Keep the human decision. Use the scores to decide whom to interview and what to ask, not to rank people automatically.
5. Write down why. A short note per candidate makes the process explainable to the candidate and to anyone who asks later.