How much does a hiring test actually improve who you hire?

A validity coefficient is not an answer to any business question

Ask an assessment vendor how good their instrument is and you will be given a number between zero and one. It is a real number, it is usually honestly derived, and on its own it tells you nothing you can act on. It is a correlation. Nobody hires a correlation.

The same coefficient produces wildly different outcomes depending on two facts about your situation that the vendor does not know. Take a validity of 0.31, the current published estimate for a cognitive ability test, against a base rate of 50% — half your applicants would work out if you hired at random:

One instrument, one coefficient, four answers ranging from transformative to marginal. And note which rows are which. Vendor examples are habitually drawn at selection ratios of 0.05 and 0.10; Schmidt and Hunter observed in 1998 that actual selection ratios are typically in the 0.30 to 0.70 range, which is the bottom half of that table.

This is not an argument against testing. A ten-point lift on a thousand hires is a hundred people, and few interventions available to an HR function move a hundred people. It is an argument against quoting a coefficient as though it were a result, which is what almost every page in this market does.

Three numbers, and the one you can actually move

The base rate is the share of your applicants who would perform satisfactorily if you hired them at random. It sets the headroom. At a base rate of 80% there are at most twenty points for any instrument to capture, and a validity of 0.31 captures about thirteen of them at an aggressive selection ratio. If your base rate is already high, selection is not your problem and no assessment will make it your solution.

The selection ratio is how many you hire divided by how many apply. It is a multiplier on validity, because an instrument can only sort people you are in a position to reject. At a selection ratio of 1.0 the most valid test ever built changes nothing at all. This is also the number most organisations can move fastest — widening the top of the funnel usually beats improving the instrument, and it is rarely the option on the table because nobody sells it.

The validity coefficient is the one everybody argues about and the one with the least room in it. The published estimates for whole categories of instrument sit between 0.19 and 0.42. Choosing well inside that band matters; it matters less than either of the other two numbers.

A useful discipline: before evaluating any instrument, write down your base rate and your selection ratio and compute the ceiling. If the best possible predictor could only buy you four points, stop the procurement.

The validity numbers, and the two revisions the market has not caught up with

The presets in the calculator are the estimates from Sackett, Zhang, Berry and Lievens (2022), who re-analysed the evidence base and found that corrections for restriction of range had been applied systematically too aggressively for twenty-five years. The revision reordered the field.

The first thing the market has not absorbed is that cognitive ability is no longer top of the list. Structured interviewing predicts better than any test in the table. Any vendor page still quoting 0.51 for an ability test in 2026 is quoting a superseded figure, and several do.

The second is newer still and almost nobody has it. Sackett and colleagues published a follow-up in 2023 collecting estimates that had appeared since, and it revises cognitive ability again — downward, to 0.23, on the basis of a meta-analysis of 113 twenty-first-century studies, which drops it from fifth to twelfth among the predictors examined. Assessment centres move the other way, up to 0.33. Both figures are in the calculator’s preset list, marked. Run the model at 0.31 and then at 0.23: on a 1-in-10 selection ratio and a 50% base rate the success rate falls from 71.5% to 66.0%, which is the difference between five and a half extra good hires per hundred.

Third, and least discussed: read the SD column. These are means of distributions, not constants. Structured interviews have an SD of ρ of 0.19 and integrity tests 0.20, which means the honest way to use this calculator is to run it twice — once at the mean and once a standard deviation below — and to make the decision that survives both.

The money model, and the experiment that says be careful who you show it to

The second panel implements Brogden–Cronbach–Gleser, the standard utility model. The gain per person hired per year is the validity coefficient multiplied by SDy — the standard deviation of the value of an employee’s output — multiplied by the mean standardised predictor score of the people you actually hired. Costs are subtracted across everyone tested, not everyone hired, which is why the selection ratio drives the answer harder than the price per test does.

Two things about that model deserve to be on the page rather than in a footnote.

SDy is a guess wearing a decimal point. The conventional estimate is 40% of salary, from Schmidt and Hunter, and it is worth reading what they actually wrote: that figure is described as a minimum— “at minimum 40% of the mean salary of the job” — and “a lower bound value”, with actual values “typically considerably higher”. It is not a typical value and it is not measured. Halve it and every number in the panel halves.

The output is usually too large to be believed, and that is an empirical finding rather than an opinion. Latham and Whyte gave 143 experienced managers either a validity-based case for adopting a selection procedure, or the same case plus a utility analysis showing substantial net financial benefits. The utility analysis reduced managerial support for adopting the procedure. Whyte and Latham ran it again in 1997 with an internationally recognised expert explaining the logic on video and then answering questions live, and the effect survived that too.

The expert in question, Steven Cronshaw, published a reply in the same issue that is the most useful thing anyone has written about how to use these numbers. He accepts the effect is real and substantial, and argues the moderator is perceived persuasive intent: a utility analysis read as an attempt to sell an intervention backfires, where the same analysis used at arm’s length to inform an investment decision may not. That is the reason this page ships a calculator you drive with your own numbers rather than a return-on-investment claim with ours in it — and the reason the honest thing to take to a decision-maker is the success-rate figure from the first panel, not the currency figure from the second.

What the model assumes, and where it breaks

• The criterion is a line drawn through a continuum.Taylor–Russell needs performance sorted into satisfactory and unsatisfactory, which real performance is not. Where you draw that line changes the base rate, and the base rate changes the answer. Two people modelling the same job with different definitions of “worked out” will get different results and both will be right.
• Predictor and criterion are assumed jointly normal and linearly related. Reference implementations of the model, including Niels Waller’s TaylorRussell package for R, integrate a multivariate normal distribution to produce the tables. Where the relationship is not linear — a threshold effect, say — the tables mislead.
• You hire strictly top-down, and everyone you offer accepts.The second half of that is almost never stated and is the assumption most likely to fail. Scullen and Meyer argued in 2014 that the standard models rest on the “inherent, but dubious, assumption that all job seekers in a given applicant pool are pursuing that particular job and no other jobs”, and that once candidates apply to several employers at once, small pools lose their strongest people to competing offers — producing significant negative effects on expected new-hire quality that the selection-ratio models do not capture. Treat the first panel’s output as a ceiling.
• The two panels measure the gain differently and do not agree. The first reports the share of the maximum possible improvement in the success rate; the second reports the gain on a continuous criterion, where — this is Brogden’s 1946 result — the share of the maximum possible gain is exactly the validity coefficient itself. At a 10% selection ratio and a 50% base rate those are 43% and 31% of maximum respectively. Neither is wrong. Dichotomising a criterion changes what “the gain” means, and any page that reports one of these numbers without saying which is being imprecise.
• None of this is a fairness analysis.A procedure can be valid, utility-positive, and still produce a selection rate for one group below four-fifths of the highest group’s. The two questions are independent and both have to be answered. Run the adverse impact ratio on the same cut you modelled here.

What AssessAll can and cannot tell you here

The validity presets in this calculator are published meta-analytic estimates for categories of instrument. None of them is an AssessAll figure, and there is no AssessAll figure to offer. A criterion-validity coefficient requires scores collected before hire and performance measured afterwards, on a sample large enough and spread across enough employers to mean anything. AssessAll has not published such a study on any instrument, so no coefficient for an AssessAll assessment appears on this page or anywhere else on this site.

What the platform does publish is the specification, which is the part you can check before you have outcome data: the item count and time limit on every form, the competencies each one reports, and what the report does at the boundary. Those are the inputs to a sensible decision when a validity coefficient is not available — and asking a vendor for them is a fair test of whether they have them.

One practical consequence of the arithmetic above, and it applies to buying from us as much as from anyone: because cost falls on everyone tested and gain only on those hired, a single expensive instrument administered to a whole applicant pool is the worst shape a screen can have. A short, cheap stage that removes most of the pool before an expensive one runs changes the economics far more than a better instrument does. That is what a staged assessment blueprint is for, and it is the argument the calculator makes on its own if you halve the number of applicants tested at full price.

Frequently asked questions

Sources

Every item below was verified at a DOI record, a publisher-deposited abstract or the paper’s own text on 13 September 2026.

• Taylor, H. C., & Russell, J. T. (1939). The relationship of validity coefficients to the practical effectiveness of tests in selection: discussion and tables. Journal of Applied Psychology 23(5), 565–578. doi:10.1037/h0057079
• Brogden, H. E. (1946). On the interpretation of the correlation coefficient as a measure of predictive efficiency. Journal of Educational Psychology 37(2), 65–76 — the result that the validity coefficient is the share of the maximum possible gain on a continuous criterion.
• Brogden, H. E. (1949). When testing pays off. Personnel Psychology 2(2), 171–183 — the linear utility model.
• Cronbach, L. J., & Gleser, G. C. (1965). Psychological Tests and Personnel Decisions (2nd ed.). University of Illinois Press, Urbana.
• Schmidt, F. L., Hunter, J. E., McKenzie, R. C., & Muldrow, T. W. (1979). Impact of valid selection procedures on work-force productivity. Journal of Applied Psychology 64(6), 609–626.
• Schmidt, F. L., & Hunter, J. E. (1998). The validity and utility of selection methods in personnel psychology. Psychological Bulletin 124(2), 262–274 — the 40%-of-salary lower bound on SDy and the 0.30–0.70 selection-ratio range, both p. 263.
• 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:10.1037/apl0000994
• Sackett, P. R., Zhang, C., Berry, C. M., & Lievens, F. (2023). Revisiting the design of selection systems in light of new findings regarding the validity of widely used predictors. Industrial and Organizational Psychology 16, 283–300 (open access) — the source of the 0.23 and 0.33 updates. doi:10.1017/iop.2023.24
• Latham, G. P., & Whyte, G. (1994). The futility of utility analysis. Personnel Psychology 47(1), 31–46.
• Whyte, G., & Latham, G. (1997). The futility of utility analysis revisited: when even an expert fails. Personnel Psychology 50(3), 601–610.
• Cronshaw, S. F. (1997). Lo! The stimulus speaks: the insider’s view on Whyte and Latham’s “The futility of utility analysis”. Personnel Psychology 50(3), 611–615.
• Scullen, S. E., & Meyer, R. D. (2014). Pitfalls in modelling the effects of recruitment on selection outcomes. Journal of Management 40(6), 1675–1699 — on the offer-acceptance assumption.
• Waller, N. G. TaylorRussell(R package, University of Minnesota) — a reference implementation, used here to corroborate the model’s normality assumption and the 1939 pagination.