Are pre-employment tests and skills assessments actually worth it, and do they predict job performance?

Yes, though by less than the figures still quoted in most textbooks and vendor decks. The reference point for three decades was Schmidt and Hunter's 1998 summary, which placed general mental ability at .51 and work sample tests at .54 [2]. In 2022, Paul Sackett, Charlene Zhang, Christopher Berry, and Filip Lievens found a methodological problem behind those coefficients: meta-analyses had applied range restriction corrections derived from predictive studies to concurrent studies, where restriction is usually small, and roughly 80% of these studies are concurrent [1]. Correcting heavily where little restriction exists produces overcorrection, so the published validities came out too high [1].

What the revised numbers look like

Mean operational validity fell by about .10 to .20 points across most predictors, and further for several: .21 lower for work sample tests, .20 lower for cognitive ability, .19 lower for unstructured interviews [2]. The rank order shifted as well. Structured interviews came out on top at .42, then job knowledge tests at .40, empirically keyed biodata at .38, and work sample tests at .33, with general mental ability at .31 [2]. A later meta-analysis limited to 21st century studies put cognitive ability lower still at .23, dropping it from fifth to twelfth [2].

Averages hide a large amount of variation

Mean validity is a poor guide on its own because the spread across studies is wide. Structured interviews average .42 with a standard deviation of .19, giving an 80% credibility interval of .18 to .66 [2]. A carefully built procedure lands near the top of that band and a rushed one near the bottom, so execution quality matters about as much as method choice [2].

Structured interviews lead the ranking because of standardization rather than anything special about interviewing. Every candidate answers the same predetermined questions and every answer is scored against defined anchors, which removes the noise that leaves an unstructured conversation predicting at only .19 [2]. The cost sits with the employer, who writes job-relevant questions, builds scoring guides, and trains the panel.

Job knowledge tests

Job knowledge tests measure what a candidate already knows and rank second at .40, with a tight standard deviation of .13 [2]. They are cheap to run and straightforward to defend on content validity grounds when items map to a documented job analysis [6]. They apply only where the knowledge is required at hire rather than trained afterward, and they carry a sizable subgroup difference of .54 [2].

Empirically keyed biodata

Empirically keyed biodata scores background responses against items statistically linked to later job outcomes, and it reaches .38 [2]. Its practical edge is the low end of its credibility interval, .26 against .18 for structured interviews, so a risk-averse employer has less downside [2]. Keying an instrument empirically needs enough hires with outcome data, which is why it suits volume programs.

Work samples and assessment centers

Work samples ask a candidate to perform a slice of the job and land at .33, well below the .54 previously reported [2]. Assessment centers, effectively multi-exercise work samples for managerial roles, land at the same .33 after correction, and both consume heavy candidate and assessor time while carrying large subgroup differences of .67 and .52 [2].

Cognitive ability tests

General mental ability tests fell furthest, from .51 to .31, and to .23 in 21st century samples [2]. They stay cheap and fast and work with candidates who have no relevant experience, but they carry the largest Black-White standardized mean difference in the set, .79 [2].

Integrity tests

Integrity tests predict at .31, revised down from .41 [2]. They keep appearing in optimized composites because they pair usable validity with a subgroup difference of only .10, adding predictive power without adding adverse impact [2][7].

Situational judgment tests

Situational judgment tests predict at .26 whether scored on knowledge or on behavioral tendency, and they work with candidates who have no job experience [2]. In multi-predictor models they often receive little weight once structured interviews and biodata are in the mix [7].

Personality questionnaires

Conscientiousness measures predict at .21 overall and .25 when items are worded to a work context, and they carry essentially no subgroup difference, so they help a composite on the diversity side [2]. Faking is the standing concern, since research on distortion in high-stakes settings shows candidates with higher cognitive ability are better at shifting responses toward an ideal profile [9].

Assessments earn their cost where hiring volume is high enough for a small gain in hire quality to compound, where the job has requirements you can measure before hire, and where the realistic alternative is an unstructured interview predicting at .19 [2].

Incremental validity is the number to ask for

What a test adds beyond your existing process matters more than its standalone coefficient. Across all composites of one to five predictors, mean validity comes to .51 using the pre-2022 figures and .47 using the revised ones, so multi-measure processes lost far less than any single measure did [2]. Removing cognitive ability entirely from a six-predictor composite cost .20 of validity under the old estimates and only .05 under the revised ones [2][7].

Where assessments tend not to pay

Very low volume is the first case, because a job analysis, a validation study, and ongoing monitoring carry a fixed cost that a handful of hires per year cannot absorb. The second is a mismatch between what the test measures and what the job requires, which fails on prediction and on the job-relatedness standard the EEOC applies [3]. The third is friction that removes strong candidates before anyone evaluates them, invisible in any vendor validity report. Where any of those holds, the same money spent building a real structured interview generally returns more, since it is the highest-validity method available and needs no vendor [2].

Adverse impact decides more assessment programs than validity does, because a procedure that disproportionately screens out a protected group is treated as discriminatory unless it has been validated [5]. Federal enforcement agencies generally treat a selection rate below four-fifths of the highest group's rate as evidence of adverse impact [4].

The evidence side of the trade-off

Subgroup differences vary enormously across methods and do not track validity. Cognitive ability tests carry the largest Black-White standardized mean difference at .79, then work samples at .67, job knowledge tests at .54, and assessment centers at .52 [2]. At the other end sit integrity tests at .10, structured interviews at .23, and conscientiousness at -.07 [2]. The older assumption was that reducing the weight on cognitive ability cost a great deal of validity, and the revised figures ease that dilemma. In a Pareto-optimized composite of six methods, the validity-maximizing solution reaches .61 at an adverse impact ratio of .42, and zeroing out cognitive ability drops validity only to .56 while raising the ratio to .67, so excluding it barely moves validity yet substantially reduces adverse impact [7][8].

What the trade-off still costs

Pushing that same model past the .80 ratio threshold requires validity to fall further, to .48, so the tension has narrowed rather than disappeared [7]. Where two procedures are substantially equally valid, the guidelines direct employers to the one with lesser adverse impact [5], and the EEOC lists searching for a less discriminatory alternative among its best practices [3]. An employer relying on a test must show it is job-related and consistent with business necessity, supported by a criterion-related, content, or construct validity study, and a vendor technical manual does not transfer that responsibility [3][6]. This is general information about federal standards, not legal advice, and state and local rules add obligations these sources do not cover.

Published assessment pricing is generally a subscription plus per-candidate credits rather than a flat per-test fee. TestGorilla's own pricing page lists a free tier with 10 credits per month, a Core plan at $142 per month billed annually, a Plus plan from $400 per month billed annually, and custom enterprise terms, with a skills test consuming one credit per candidate [11]. Those are vendor-published figures as of July 2026 and are not independently audited.

Candidate drop-off is the hidden cost line

The cost that rarely reaches a business case is the candidate you never evaluate. Criteria Corp reports, from an analysis of roughly 500,000 of its own tests, that employers can expect a completion rate near 80% for assessments up to 40 minutes, falling toward 60% once total testing runs beyond an hour [10]. That is a vendor claim about a vendor's own platform, and no independent benchmark of comparable scale was available to corroborate it. A 40% loss at the assessment stage outweighs the difference between most competing instruments, so calculate cost per completed assessment rather than cost per license, then weigh it against the validity the instrument adds on top of a structured interview [2].

Ask for the technical report before you ask for a demo. A credible vendor supplies validation evidence, subgroup statistics, and the samples behind them, and one that cannot is asking you to carry a legal obligation on trust [3].

The checks worth running, in order:

• Start from a documented job analysis, since the EEOC's standard is whether the procedure is job-related for the position and appropriate to your purpose [3].
• Ask which validation strategy supports the instrument; the Uniform Guidelines accept criterion-related, content, or construct validity studies [6].
• Ask whether the underlying samples were concurrent or predictive and how range restriction was handled, the exact issue that inflated older estimates [1].
• Ask for the credibility interval, not the mean validity alone, because .42 ranging from .18 to .66 is a different purchase than .40 with a narrow band [2].
• Request subgroup data, then calculate your own selection rates by group against the four-fifths standard once the tool is live [4].
• Compare at least one lower-impact alternative of comparable validity, since the guidelines point you toward the equally valid procedure with less adverse impact [5].
• Measure completion time and completion rate during a pilot, then decide whether the drop-off is acceptable at your volume [10].
• Confirm the accommodation path for candidates with disabilities before launch, because Americans with Disabilities Act obligations apply regardless of who built the tool [3].

It is not a lie detector

Integrity tests are written questionnaires and sit outside polygraph law, though the two get conflated. The Employee Polygraph Protection Act makes it unlawful for most private employers to "directly or indirectly, to require, request, suggest, or cause any employee or prospective employee to take or submit to any lie detector test" [12].

It is not a medical examination

Medical inquiries and physical examinations fall under separate rules with timing restrictions, distinct from cognitive, personality, and skills testing [3]. A personality inventory that strays into questions likely to reveal a mental health condition can cross that line.

It is not a personality type sorter

Instruments that assign candidates to types are built for development conversations, and none of the validity evidence here applies to them. The measured predictor in the research is trait conscientiousness, not a type label [2].

It is not a substitute for a job analysis

Buying a validated instrument does not make it valid for your job, since validation is position-specific and the employer holds that burden [3][6].