# What actually predicts job performance? What the 2022 re-analysis changed

> The hiring research most people quote was overstated. A 2022 re-analysis re-ranked the methods, put structured interviews first, and found years of experience barely predicts anything. Here is what that means for a small team.

David at SortTheStack, October 9, 2026. Original: https://sortthestack.com/blog/what-predicts-job-performance

The short answer: **a well-run structured interview, a test of the specific
knowledge the job needs, and a realistic sample of the work** are the strongest
predictors of job performance we have. **Years of experience is one of the
weakest.** And every method predicts less than the figures you have probably
seen quoted.

That ranking comes from a 2022 re-analysis by Paul Sackett, Charlene Zhang,
Christopher Berry and Filip Lievens in the _Journal of Applied Psychology_. It
revisited the meta-analyses behind one of the most widely cited summaries in
hiring research and concluded that the old numbers had been systematically inflated.

This article explains what changed, what the numbers do and do not mean, and
how a team without an HR department can use them.

## The numbers most people quote are from 1998

If you have read about evidence-based hiring, you have probably met
[Schmidt and Hunter (1998)](https://doi.org/10.1037/0033-2909.124.2.262). Their
summary of 85 years of research put work sample tests, cognitive ability tests
and structured interviews at the top, each with an estimated validity above .50.
It has shaped hiring advice for a quarter of a century.

[Sackett and colleagues (2022)](https://doi.org/10.1037/apl0000994) looked at
how those estimates were produced. Raw correlations between a selection method
and later job performance are usually corrected upward, partly because studies
only measure the people who were hired, and that narrower group understates how well
the method separates the whole applicant pool. The correction is legitimate.
The authors' argument is that the way it was usually applied overcorrected,
often by a lot, because it assumed far more restriction than most studies
actually had.

Re-doing the corrections, they found that most methods kept their rank but lost
.10 to .20 of their estimated validity. In their words, selection procedures
"remain useful, but selection predictor–criterion relationships are
considerably lower than previously thought."

## The revised ranking

These are the paper's estimates for the methods a small business is most likely
to use, from its Table 3. "Validity" here is the estimated correlation between
the method and later job performance.

| Method                      | 1998 estimate | 2022 estimate | Practical for a small team?                       |
| --------------------------- | ------------- | ------------- | ------------------------------------------------- |
| Structured interview        | .51           | .42           | Yes. Costs preparation time, not money.           |
| Job knowledge test          | .48           | .40           | Yes, where the job has a body of knowledge.       |
| Empirically keyed biodata   | .35           | .38           | Rarely. Needs data from many past hires to build. |
| Work sample test            | .54           | .33           | Yes, if it is short and resembles the real work.  |
| Cognitive ability test      | .51           | .31           | Possible, with care. See subgroup differences.    |
| Integrity test              | .41           | .31           | Usually bought from a test publisher.             |
| Unstructured interview      | .38           | .19           | The default, and only half as predictive.         |
| Conscientiousness (general) | .31           | .19           | Weak on its own.                                  |
| Job experience (years)      | .18           | .07           | Barely predictive.                                |

Two methods often used in small-business hiring do not appear in the revised
list at all. Reference checks and years of education were excluded because
there was not enough information to re-estimate them, so the paper offers no
updated figure for either.

## Three things the table changes

### 1. The interview you run matters more than the fact that you interview

The gap between structured and unstructured interviews is the most useful
number in the paper for a small business: .42 against .19. A structured
interview asks every candidate the same job-related questions and rates the
answers against anchors decided in advance. An unstructured interview is a
conversation that goes wherever it goes.

Both can fit in the same hour. On these estimates, one is more than twice as
predictive as the other.

The structured interview's estimate also varies a lot from study to study
(a standard deviation of .19, with a lower 80% credibility value of .18). That
spread is the warning: "structured" covers a range of practice, and a
loosely structured interview will not get you the average result.

### 2. Years of experience is a weak screen

A validity of .07 means that, across the studies reviewed, knowing how many
years someone had worked told you very little about how well they would do
the job. An earlier meta-analysis dedicated to the question,
[Van Iddekinge and colleagues (2019)](https://doi.org/10.1111/peps.12335),
reached a similar conclusion: pre-hire work experience correlated about .06
with job performance across 44 samples, and not at all with turnover.

That does not make experience irrelevant. It means the _count of years_ is a
poor stand-in for what the years were supposed to produce. "Five or more years
of accounting experience" is easy to check and weakly predictive. "Has run a
month-end close independently" is harder to check and much closer to the
thing you need.

This is the pattern behind most weak hiring criteria. **They are chosen because
they are easy to evaluate, not because they predict much.** A year count, a
degree title and a recognisable employer name can all be read in seconds. The
evidence that someone can do the work takes longer to find.

### 3. Everything predicts less than you think

A correlation of .42 is strong by the standards of this research and modest by
the standards of everyday intuition. Candidates who do better in a structured
interview tend to perform better, with many exceptions in both directions.

The practical consequence is that no single method should decide a hire.
Combining two or three good ones beats leaning on any one, and the strongest
single predictor still leaves plenty of room for surprises.

## What the numbers do not tell you

The paper is careful about its limits, and they matter for how you use it.

- **These are averages across many jobs and studies.** A method's validity in
  your firm, for your role, can be higher or lower.
- **"Job performance" is largely supervisor ratings**, mixed with some
  objective measures. Ratings are themselves an imperfect measure of how good
  someone is at a job.
- **Many of the underlying studies tested current employees, not applicants.**
  The work sample estimate rests on 54 studies, 53 of them with incumbents, and
  the job knowledge estimate comes entirely from such samples because no
  applicant meta-analysis exists. Applicants are a more varied group, and the
  correction for that difference is exactly what the paper argues was overdone.
- **It ranks methods, not resumes.** Nothing in the table measures how well
  someone reads a resume. Resume screening sits upstream of all of it: it
  decides who gets the chance to be evaluated by the stronger methods at all.

## Validity is not the only consideration

The paper pairs each method with its average Black–White subgroup difference,
because a method can predict well and still screen out one group of applicants
far more than another. Cognitive ability tests show a large difference
(d = .79 in the paper's table); structured interviews a much smaller one
(d = .23). The paper notes that some of these subgroup figures come from
samples other than job applicants, so treat them as approximate.

The authors call this the validity–diversity tradeoff. For a small employer
the practical reading is simple: the structured interview combines the highest
validity in the table with a comparatively small subgroup difference, which is
one more reason to start there. Whether a particular
test creates legal risk depends on where you operate and how you use it, which
is a question for an employment lawyer, not a research table.

## A hiring process built on the evidence

Here is one way to turn the ranking into a process a small team can run. It is
our interpretation of the research, not a procedure the paper prescribes.

1. **Write the requirements as evidence, not proxies.** For each thing the job
   needs, write what you would expect to see from someone who has it.
   "Has prepared year-end working papers for an external audit" rather than
   "3+ years audit experience".
2. **Screen resumes for that evidence.** Read for whether the application shows
   the thing, says nothing about it, or shows the opposite. Treat "says nothing"
   as unknown rather than as a no. Our guide to
   [screening a large pile of resumes](https://sortthestack.com/blog/how-to-screen-a-pile-of-resumes)
   walks through this step in detail.
3. **Add one short work sample or knowledge check** for the shortlist, built
   from the real job. Keep it short enough that a strong candidate with other
   options will still do it.
4. **Run a structured interview.** Same questions, in the same order, for every
   finalist. Write down what a strong, acceptable and weak answer looks like
   before the first interview. Our guide to
   [running a structured interview](https://sortthestack.com/blog/structured-interviews) has a worked
   example.
5. **Rate each piece separately, then combine by a rule you set in advance.**
   A meta-analysis by
   [Kuncel and colleagues (2013)](https://doi.org/10.1037/a0034156) found that
   combining information mechanically, by a formula or fixed rule, predicted job
   performance substantially better than holistic judgement, by more than 50%,
   even when the judges were experts in the jobs concerned. Discuss the result
   afterwards; just do not let the discussion replace the ratings. A free
   [candidate scorecard](https://sortthestack.com/blog/candidate-scorecard-template) does this on one
   page.

None of this needs software. It needs the requirements written down before
anyone reads an application, and the discipline to evaluate everyone against
them.

## Where SortTheStack fits

SortTheStack covers step 2. It drafts requirements from your job description
for you to review, then reads every application against them and quotes the
evidence it found, marks what it could not find as unclear rather than missing,
and leaves the decision to you. It does not run interviews or tests, and it
does not predict job performance. If step 2 is where your time goes, you can
[see how the screening works](https://sortthestack.com/product).

## 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.
  [doi.org/10.1037/apl0000994](https://doi.org/10.1037/apl0000994). Figures in
  this article are from its Table 3.
- Schmidt, F. L., & Hunter, J. E. (1998). The validity and utility of selection
  methods in personnel psychology. _Psychological Bulletin, 124_(2), 262–274.
  [doi.org/10.1037/0033-2909.124.2.262](https://doi.org/10.1037/0033-2909.124.2.262)
- Van Iddekinge, C. H., Arnold, J. D., Frieder, R. E., & Roth, P. L. (2019). A
  meta-analysis of the criterion-related validity of prehire work experience.
  _Personnel Psychology, 72_(4), 571–598.
  [doi.org/10.1111/peps.12335](https://doi.org/10.1111/peps.12335)
- Kuncel, N. R., Klieger, D. M., Connelly, B. S., & Ones, D. S. (2013).
  Mechanical versus clinical data combination in selection and admissions
  decisions: A meta-analysis. _Journal of Applied Psychology, 98_(6),
  1060–1072. [doi.org/10.1037/a0034156](https://doi.org/10.1037/a0034156)