Personnel selection and validity generalization: Schmidt-Hunter meta-analysis and the science of hiring

Intuition Beginner

How should a company hire people? For most of the twentieth century the standard answer inside personnel psychology was that "it depends" — on the job, the firm, the local situation. Each job, it was believed, needed its own bespoke selection study, because what predicted success as a clerk supposedly differed from what predicted success as a salesperson or a manager.

Frank Schmidt and John Hunter, working at the University of Iowa from 1977 onward, spent forty years testing that assumption across hundreds of studies and millions of workers. Their finding inverted the field's default. The "it depends" variance between jobs and firms shrank dramatically once they corrected for four statistical artifacts — small samples, imperfect performance ratings, restricted score ranges among incumbents, and pass-or-fail scoring. After correction, the best predictor of job performance was the same everywhere.

That predictor is a general cognitive-ability test — an IQ-like instrument measuring reasoning, verbal, and numerical skill. Across jobs from clerk to manager, the correlation with performance is about $0.5$ to $0.65$. Structured interviews, where every candidate gets the same questions scored systematically, come second. Unstructured interviews, reference letters, and graphology are weak or useless. The finding has been replicated in dozens of countries and industries and reshaped employment law, military practice, and corporate hiring.

Visual Beginner

The picture plots the operational validity of the major personnel-selection methods as horizontal bars, ordered from strongest to weakest. The lengths are the Schmidt-Hunter 1998 meta-analytic estimates after correction for statistical artifacts.

The drop from cognitive ability and structured interviews (above $0.5$) down to graphology and age (near zero) is the load-bearing visual fact. Methods at the top of the chart should be the load-bearing components of any hiring system; methods at the bottom should not.

Worked example Beginner

Setting. In the early 1980s the United States Army faced a selection problem at industrial scale. Every year it had to absorb roughly $200,000$ new recruits, sort them across about $150$ military specialties, and absorb the cost of training failures when a recruit washed out of a specialty they were not suited for. The Army commissioned Project A, a multi-year validation study led by John Campbell and David Knapp, to redesign selection and assignment around evidence rather than tradition.

Step 1. Instruments. Project A built on the Armed Services Vocational Aptitude Battery (ASVAB), a ten-subtest cognitive battery administered to every recruit. It added the TAPAS — the Tailored Adaptive Personality Assessment System — to measure conscientiousness and other traits. The ASVAB yields a general factor that closely mirrors standard IQ tests, plus specific aptitude composites for clerical, mechanical, electronics, and combat skills.

Step 2. Validation. The Project A team tracked roughly $50,000$ recruits through training and into their first tour of duty, collecting hands-on job-performance measures (work samples, supervisor ratings, training grades) rather than relying on supervisors' global impressions alone. The correlations were computed by specialty and aggregated.

Step 3. Results. The general cognitive factor extracted from the ASVAB predicted training success at $r\approx 0.6$ to $0.8$ and hands-on job performance at $r\approx 0.4$ to $0.6$, varying by specialty complexity. Conscientiousness from the TAPAS added incremental validity on top of cognitive ability, especially for disciplined-performance criteria. The Army's selection and assignment system has been refined on the basis of these data ever since.

What this tells us: a large-scale, criterion-validated selection study using a cognitive-ability battery reproduces the Schmidt-Hunter pattern — general mental ability is the dominant predictor, and a personality measure adds incremental validity — at the scale of hundreds of thousands of workers and hundreds of specialties.

Check your understanding Beginner

Formal definition Intermediate+

The Schmidt-Hunter program rests on four formal objects: the validity coefficient, the operational validity, the four statistical artifacts that attenuate it, and the meta-analytic combination of corrected validities across studies. The definitions below fix notation.

Definition (Validity coefficient). Let $X$ be a predictor score (for example, a GMA composite) and $Y$ a criterion score (for example, a supervisor rating of job performance, a work-sample score, or a training grade) in a population of incumbents. The validity coefficient is the Pearson correlation

$${\rho }_{XY}=\frac{\mathrm{Cov}(X,Y)}{{\sigma }_X{\sigma }_Y},$$

interpreted as the standardised slope of the linear regression of $Y$ on $X$. In personnel selection the criterion-referenced quantity of interest is the operational validity — the correlation between predictor and an external measure of job performance, not between predictor and another predictor.

Definition (Attenuation by criterion unreliability; Spearman 1904). Let $X^{\ast }$ be the observed predictor score and $Y^{\ast }$ the observed criterion score, where $X^{\ast }=X+{\epsilon }_X$ and $Y^{\ast }=Y+{\epsilon }_Y$ are classical-test-theory decompositions with $\mathrm{Var}({\epsilon }_X)>0$ and $\mathrm{Var}({\epsilon }_Y)>0$ independent of true scores. Define predictor reliability $r_{XX}=\mathrm{Var}(X)/\mathrm{Var}(X^{\ast })$ and criterion reliability $r_{YY}=\mathrm{Var}(Y)/\mathrm{Var}(Y^{\ast })$. The Spearman attenuation formula ^{[Spearman1904]} gives

$${\rho }_{X^{\ast }{Y}^{\ast }}={\rho }_{XY}\sqrt{r_{XX}}\sqrt{r_{YY}},$$

so the observed correlation is always smaller in magnitude than the true correlation. Solving for the true correlation yields the disattenuation ${\rho }_{XY}={\rho }_{X^{\ast }{Y}^{\ast }}/(\sqrt{r_{XX}}\sqrt{r_{YY}})$.

Definition (Range restriction; Thorndike 1949 Case II). Let $(X,Y)$ be jointly normal in the unrestricted applicant population with correlation $\rho$. Suppose selection is made on $X$ alone, retaining only those applicants with $X\ge c$ (or, more generally, restricting $X$ to a sub-range). Write $u={\sigma }_{X,\mathrm{restricted}}/{\sigma }_{X,\mathrm{unrestricted}}$ for the ratio of restricted to unrestricted predictor standard deviations; $u<1$ under direct range restriction. The observed correlation in the restricted incumbent sample is

$${\rho }_{\mathrm{obs}}=\frac{\rho u}{\sqrt{1-{\rho }^2+{\rho }^2{u}^2}},$$

the Thorndike Case II formula. The inverse, $\rho ={\rho }_{\mathrm{obs}}u/\sqrt{1-{\rho }_{\mathrm{obs}}^2+{\rho }_{\mathrm{obs}}^2{u}^2}$, is the range-restriction correction applied in validity generalization.

Definition (Validity generalization; Schmidt-Hunter 1977). Let $\{{\hat{\rho }}_i{\}}_{i=1}^k$ be observed validity coefficients from $k$ independent primary studies sharing a common predictor construct and criterion construct, each with sample size $n_i$ and artifact measurements. The validity-generalization estimate of the operational validity is

$${\hat{\rho }}_{\mathrm{VG}}=\frac{{\sum }_i{w}_i{\hat{\rho }}_i^{(c)}}{{\sum }_i{w}_i},w_i=n_i,$$

where ${\hat{\rho }}_i^{(c)}$ is the $i$-th observed validity individually corrected for sampling error, criterion unreliability, range restriction, and dichotomisation. The percent of variance in $\{{\hat{\rho }}_i\}$ accounted for by the four artifacts is

$$\%{\mathrm{Var}}_{\mathrm{art}}=\frac{{\mathrm{Var}}_{\mathrm{art}}(\hat{\rho })}{{\mathrm{Var}}_{\mathrm{obs}}(\hat{\rho })}\times 100\%.$$

The situation-specific-validity hypothesis asserts $\%{\mathrm{Var}}_{\mathrm{art}}\ll 100\%$ (residual variance is genuine situational moderation); validity generalization asserts $\%{\mathrm{Var}}_{\mathrm{art}}\approx 100\%$ (residual variance is artifactual). The Schmidt-Hunter 1977 finding was $\%{\mathrm{Var}}_{\mathrm{art}}\approx 80\%$ for GMA-criterion validity.

Counterexamples to common slips

• Validity generalization means any cognitive test works anywhere. No. What generalizes is the construct of general mental ability and its operational validity for performance criteria. A specific test must still be locally administered and shown to load on $g$; an off-the-shelf test with poor psychometric properties does not inherit the $0.51$ validity merely by being labelled a cognitive test.

• The $4/5$ rule is a scientific significance test. No. The four-fifths rule — selection rate for a protected group below $80\%$ of the majority-group rate — is a regulatory guideline from the 1978 Uniform Guidelines on Employee Selection Procedures ^[EEOC1978], not a statistical hypothesis test. It flags disparate impact for investigation but says nothing about its cause or its remedy.

• Meta-analysis is just averaging correlations. No. Schmidt-Hunter correct each individual study for four statistical artifacts before inverse-variance-weighted pooling, and they quantify the residual between-study variance after correction. Naïve averaging of uncorrected correlations would underestimate operational validity and overestimate situational moderation.

• Each job needs a unique selection system. Partly false. The validity of GMA generalizes across jobs; what is job-specific is the appropriate job-knowledge test, the work-sample content, and the personality profile. The Campbell-Knapp Project A data show GMA validity varying with job complexity but positive everywhere.

• AI selection tools outperform traditional tests. Contested. Machine-learning selection algorithms can match GMA validity when trained on enough performance data, but they introduce new bias risks (training-set labels encode historical inequalities) and require the same local validation that any selection procedure requires. There is no meta-analytic evidence that AI selection strictly dominates GMA plus structured interviews.

• Personality matters more than intelligence for hiring. No. Conscientiousness is the strongest personality predictor of performance ($r\approx 0.31$ corrected) and adds incremental validity on top of GMA, but GMA is primary. The best two-predictor combination is GMA plus a measure of conscientiousness or integrity, with combined operational validity around $0.6$.

• You can fake a personality test, so they are useless. Overstated. Faking is a real concern, especially on applicant samples versus incumbent samples. But personality tests corrected for faking retain useful incremental validity, and forced-choice formats (like the TAPAS) reduce fakability.

Key result: the Schmidt-Hunter 1998 meta-analytic findings Intermediate+

Result (Schmidt-Hunter 1998). Aggregating $85$ years of personnel-selection research, general mental ability (GMA) is the single best predictor of overall job performance across all jobs, with operational validity $\hat{\rho }\approx 0.51$ uncorrected and $\hat{\rho }\approx 0.65$ after correction for sampling error, criterion unreliability, range restriction, and dichotomisation. The rank order of operational validities of the major selection methods, drawn from the same meta-analysis, is:

Method	Operational validity (corrected)
GMA + integrity test	$\approx 0.65$
GMA + structured interview	$\approx 0.63$
Work-sample test	$\approx 0.54$
GMA alone	$\approx 0.51$
Structured interview	$\approx 0.51$
Integrity test	$\approx 0.41$
Conscientiousness	$\approx 0.31$
Reference check	$\approx 0.26$
Unstructured interview	$\approx 0.20$
Years of job experience	$\approx 0.18$
Graphology	$\approx 0.02$

The combination GMA + integrity (or GMA + conscientiousness) is the most efficient two-predictor battery for overall job performance. GMA predicts training success even more strongly than job performance (about $0.6$ corrected). GMA validity increases with job complexity (from about $0.23$ for the simplest jobs to about $0.58$ for the most complex), but is positive for every job complexity level studied.

Defence. Schmidt-Hunter 1998 ^{[SchmidtHunter1998]} aggregated every available meta-analysis of personnel-selection validity through the mid-1990s — building on Hunter-Hunter 1984 ^{[HunterHunter1984]}, Schmidt-Hunter-Pearlman 1985 ^{[SchmidtHunterPearlman1985]}, and the validity-generalization program launched in Schmidt-Hunter 1977 ^{[SchmidtHunter1977]}. The artifact-correction framework, applied uniformly across the primary-study pool, yielded the table above. The Sackett-Lievens 2008 review ^{[SackettLievens2008]} and the Sackett et al. 2017+ updates ^{[Sackett2017]} reproduced the rank order with larger databases; the GMA validity estimate moved within a few hundredths but the rank order of methods was stable.

The $80\%$-plus artifactual-variance finding from Schmidt-Hunter 1977 ^{[SchmidtHunter1977]} was the load-bearing methodological result. Before 1977 the field believed that validity was situation-specific: a cognitive test that predicted clerical performance in firm A need not predict it in firm B. Schmidt and Hunter showed that once the four artifacts were corrected, the residual between-study variance was small. The situational-specificity hypothesis was an artefact of under-powered primary studies with noisy criteria and restricted ranges. Validity generalization follows.

The Project A field validation (Campbell-Knapp 2001 ^{[CampbellKnapp2001]}) provided an independent check at the scale of the United States Army. The ASVAB-extracted GMA factor predicted training success at $r\approx 0.6$ to $0.8$ and hands-on job performance at $r\approx 0.4$ to $0.6$ across roughly $50,000$ recruits — the same rank order and magnitude as the meta-analytic estimate.

Caveat. The Schmidt-Hunter validity estimates describe the average operational validity across the jobs and criteria in the primary-study pool. Situational moderators exist: GMA validity is higher for complex jobs and for training criteria, lower for simple jobs. Local validation remains advisable, especially when a test has disparate impact on a protected group. The artifact corrections recover the operational validity of a properly administered selection procedure; they do not license the use of a poorly constructed test.

Bridge. The Schmidt-Hunter finding is the foundational reason that the construct of general mental ability — extracted as the leading principal component of any diverse cognitive battery in 29.13.04 — becomes a personnel decision variable rather than merely a psychometric construct: it identifies the operational validity of $g$ with the most-studied number in applied psychology, and the bridge is between the factor-analytic tradition of Spearman 1904 and the employment-utility framework of Taylor-Russell 1939. This is exactly the I/O application of the within-versus-between logic defended in 29.13.04: situational moderation is a within-population residual, not a between-population refutation of validity generalization. The pattern appears again in 29.08.04 as the person-situation debate, where Mischel's situationist critique of trait psychology dissolves once aggregated — the central insight of the density-distribution resolution mirrors the artifact-correction logic here. The result generalises to the personnel-selection survey 29.15.01 as the empirical spine that justifies the I/O field's preference for cognitive-ability and structured-interview batteries, and builds toward 29.13.01 as the psychometric-measurement foundation that the attenuation formula presupposes.

Exercises Intermediate+

Interpretive debates and developments Master

Result 1 (Spearman 1904: the attenuation formula). Charles Spearman's 1904 paper in the American Journal of Psychology ^{[Spearman1904]} introduced both the rank correlation coefficient and the correction for attenuation. Spearman showed that the observed correlation between two noisy measures is bounded above by the geometric mean of their reliabilities times the true correlation: ${\rho }_{X^{\ast }{Y}^{\ast }}={\rho }_{XY}\sqrt{r_{XX}}\sqrt{r_{YY}}$. The attenuation formula is the foundation of every subsequent artifact correction in personnel-selection meta-analysis. Spearman's paper is also where the general factor $g$ first appeared (see 29.13.04); the same paper that delivered $g$ delivered the methodological machinery Schmidt and Hunter would later use to defend $g$'s operational validity.

Result 2 (Ghiselli 1966: the first compendium). Edwin Ghiselli's 1966 compendium The Validity of Occupational Aptitude Tests ^{[Ghiselli1966]} assembled every available validity coefficient across occupational groups. Ghiselli's data were the empirical substrate on which the situation-specificity hypothesis was built: validity coefficients varied widely across jobs and firms. Ghiselli himself did not apply artifact corrections, so the spread in his data reflected sampling error and other artifacts in addition to any genuine situational moderation. The Ghiselli era established the empirical puzzle; Schmidt and Hunter supplied the resolution.

Result 3 (Schmidt-Hunter 1977: validity generalization). Schmidt and Hunter's 1977 Journal of Applied Psychology paper ^{[SchmidtHunter1977]}, "Development of a General Solution to the Problem of Validity Generalization," applied the four-artifact correction to a sample of GMA-criterion validity studies and showed that approximately $80\%$ of the observed between-study variance vanished after correction. The residual $20\%$ was attributed to genuine situational moderators (job complexity, criterion type). The paper overturned the situation-specificity doctrine that had dominated personnel psychology since Ghiselli. The artifact-correction framework was expanded and refined in Schmidt-Hunter-Pearlman 1985 ^{[SchmidtHunterPearlman1985]}.

Result 4 (Hunter-Hunter 1984: the 100-year meta-analysis). Hunter and Hunter's 1984 Psychological Bulletin paper ^{[HunterHunter1984]}, "Validity and Utility of Alternative Predictors of Job Performance," synthesised a century of selection research and reported the canonical GMA validity estimate of $r\approx 0.53$ corrected for the four artifacts. The paper also introduced the utility calculus — translating validity coefficients into dollar-value productivity gains via the Brogden-Cronbach-Gleser formula — that made validity generalization an operational rather than purely academic result. The Hunter-Hunter paper is the bridge between the validity-generalization program and the personnel-selection survey that opens this chapter 29.15.01.

Result 5 (Schmidt-Hunter 1998: the canonical summary). Schmidt and Hunter's 1998 Psychological Bulletin paper ^{[SchmidtHunter1998]}, "The Validity and Utility of Selection Methods in Personnel Psychology: Practical and Theoretical Implications of 85 Years of Research Findings" (124:262), is the canonical summary of the program. The headline table — GMA $0.51$ uncorrected, $0.65$ corrected; work samples $0.54$; structured interviews $0.51$; integrity tests $0.41$; conscientiousness $0.31$; unstructured interviews $0.20$; references $0.26$; experience $0.18$; graphology $0.02$ — is the most-cited single result in personnel-selection research. The paper also ranked the combinations: GMA + integrity test ($0.65$) and GMA + structured interview ($0.63$) are the optimal two-predictor batteries for overall job performance.

Result 6 (Project A; Campbell-Knapp 2001: independent field validation). The United States Army's Project A, led by John Campbell and David Knapp from the early 1980s onward and summarised in their 2001 volume Exploring the Limits in Personnel Selection and Classification ^{[CampbellKnapp2001]}, was the largest single-site criterion-related validation study ever conducted. Roughly $50,000$ recruits were tracked through training and into their first tours of duty across about $150$ military specialties. The ASVAB-derived GMA factor predicted training success at $r\approx 0.6$ to $0.8$ and hands-on performance at $r\approx 0.4$ to $0.6$. The companion TAPAS personality inventory added incremental validity via conscientiousness. Project A is the independent field validation that the Schmidt-Hunter meta-analytic numbers describe reality at scale.

Result 7 (Sackett-Lievens 2008; Sackett et al. 2017+: the modern updates). Sackett and Lievens's 2008 review in Annual Review of Psychology ^{[SackettLievens2008]} updated the Schmidt-Hunter meta-analyses with the larger primary-study databases available by the late 2000s. The rank order of selection-method validities was stable; the GMA estimate moved within a few hundredths. Sackett et al.'s 2017+ updates ^{[Sackett2017]} extended the analysis to situational-judgment tests, assessment centres, and modern structured-interview formats. The overall picture is robust: GMA is the dominant predictor; structured interviews and work samples are competitive alternatives; combinations of GMA with a personality or integrity measure remain optimal.

Result 8 (legal framework: Taylor-Russell 1939; Griggs 1971; Uniform Guidelines 1978). The Schmidt-Hunter program has a parallel legal history. Taylor and Russell's 1939 Journal of Applied Psychology paper ^{[TaylorRussell1939]} introduced the expectancy tables linking validity, base rate, and selection ratio to the expected success rate among hires — the utility framework that makes validity coefficients operationally meaningful. The Supreme Court's 1971 decision in Griggs v. Duke Power Co. ^[Griggs1971] established the disparate-impact doctrine: a selection test with disproportionate impact on a protected group must be shown to be job-related. The 1978 Uniform Guidelines on Employee Selection Procedures ^[EEOC1978] codified the four-fifths rule and the three validation strategies (criterion-related, content, construct). The Schmidt-Hunter evidence is the scientific input; the legal framework is the regulatory constraint; the personnel-selection practitioner operates at their intersection.

Result 9 (the GMA-complexity interaction). GMA validity is not literally constant across jobs. The Attention and Processing Capacity framework (Hunter 1986) and subsequent reconstructions show GMA validity rising from about $r=0.23$ for the least complex jobs (unskilled manual labour) to about $r=0.58$ for the most complex (professional, managerial, technical). The increase is monotone in job complexity as measured by the Dictionary of Occupational Titles' data-people-things ratings. The interaction does not overturn validity generalization — GMA validity is positive at every complexity level — but it refines the practical recommendation: for the most complex jobs, GMA is even more dominant; for the simplest jobs, alternative procedures (work samples) come closer.

Result 10 (machine learning and AI selection). The post-2015 emergence of ML/AI selection tools — résumé screeners, video-interview scorers, gamified assessments — is the most contested frontier. The published evidence (Sackett et al. 2017+; McDaniel et al. 2023 on algorithmic interviewing) shows that well-validated ML tools can match GMA validity when trained on large, criterion-relevant datasets, but introduce new bias risks: training labels encode historical inequalities; black-box models obscure the construct being measured; local validation is harder. The Schmidt-Hunter framework provides the validation standard against which any new selection procedure — including an AI tool — must be evaluated.

Synthesis. The foundational reason that the Schmidt-Hunter program succeeded where situation-specificity failed is that the four statistical artifacts — sampling error, criterion unreliability, range restriction, dichotomisation — are systematic attenuators whose variance contribution can be removed by analytic correction, and the central insight of validity generalization is that the residual between-study variance after correction is small relative to the artifactual variance, typically below $20\%$ of the observed total. Putting these together with the Hunter-Hunter utility calculus and the Taylor-Russell expectancy tables, this is exactly the picture that identifies the operational validity of $g$ — extracted as the leading principal component of any cognitive battery in 29.13.04 — with the most-studied number in applied psychology, and the bridge is between Spearman's 1904 attenuation formula and the modern regulatory framework of Griggs and the Uniform Guidelines.

The pattern appears again in 29.08.04 as Mischel's person-situation debate, where the density-distribution resolution mirrors the artifact-correction logic: cross-situational variance in behaviour, like cross-study variance in validity, dissolves once aggregated over enough observations. The result generalises to the personnel-selection survey 29.15.01 as the empirical spine of evidence-based hiring, builds toward the psychometric foundation of 29.13.01 as the measurement-theoretic substrate that the attenuation formula presupposes, and the pattern recurs in every domain in which a within-population regularity has been mistaken for a between-population refutation — from the validity-generalization debate of the 1970s to the AI-bias debates of the 2020s. The central methodological lesson — that situational moderation is a residual to be measured after artifact correction, not a default to be assumed — is the load-bearing interpretive content of the modern synthesis.

Full argument set Master

Proposition 1 (Validity generalization: artifact correction recovers a constant operational validity). Let $\{(X_i^{\ast },Y_i^{\ast }){\}}_{i=1}^k$ be $k$ independent validity-study datasets, each drawn from a population in which $(X_i,Y_i)$ are jointly normal with common true correlation $\rho$. Suppose each study $i$ observes predictor scores $X_i^{\ast }$ with reliability $r_{XX,i}$, criterion scores $Y_i^{\ast }$ with reliability $r_{YY,i}$, range-restriction ratio $u_i={\sigma }_{X,\mathrm{restricted},i}/{\sigma }_{X,\mathrm{unrestricted}}$, and (optionally) dichotomisation of $X_i^{\ast }$. Then under classical test theory the observed validities ${\hat{\rho }}_i$ converge in probability to a common corrected value $\rho$ after artifact correction, and the percent of between-study variance attributable to the artifacts tends to $100\%$ as the per-study sample sizes grow.

Proof. Under classical test theory and joint normality, the expected observed correlation in study $i$ is given by composing the attenuation, range-restriction, and dichotomisation effects. Suppressing the dichotomisation term (which contributes an analogous multiplicative attenuation), the expected observed correlation is

$$\mathbb{E}[{\hat{\rho }}_i]=\rho \cdot \sqrt{r_{XX,i}}\cdot \sqrt{r_{YY,i}}\cdot \frac{u_i}{\sqrt{1-{\rho }^2+{\rho }^2{u}_i^2}}=:\rho \cdot a_i,$$

where $a_i\in (0,1]$ is the attenuation factor for study $i$, the product of the four artifact ratios. The corrected validity in study $i$ is ${\hat{\rho }}_i^{(c)}={\hat{\rho }}_i/a_i$, which is an asymptotically unbiased estimator of $\rho$ with variance going to zero as $n_i\to \infty$.

In the meta-analytic pool, the observed between-study variance is

$${\mathrm{Var}}_{\mathrm{obs}}(\hat{\rho })=\underset{\text{artifact variance}}{\underbrace{{\rho }^2{\mathrm{Var}}_i(a_i)}}+\underset{\text{sampling variance}}{\underbrace{\frac{1}{k}\sum _i{\mathrm{SE}}^2({\hat{\rho }}_i)}}+\underset{\text{genuine moderation}}{\underbrace{{\mathrm{Var}}_{\mathrm{mod}}}}.$$

The first term — variance in the attenuation factors $a_i$ across studies — accounts for differences in reliability, range restriction, and dichotomisation across the primary-study pool. The second term is the average sampling-error variance. The third term is the genuine situational moderation that survives artifact correction. The variance after correction removes the first term and most of the second:

$${\mathrm{Var}}_c({\hat{\rho }}^{(c)})={\mathrm{Var}}_{\mathrm{mod}}+o(1/k).$$

Consequently,

$$\%{\mathrm{Var}}_{\mathrm{art}}=\frac{{\mathrm{Var}}_{\mathrm{obs}}-{\mathrm{Var}}_c}{{\mathrm{Var}}_{\mathrm{obs}}}\times 100\%=\frac{{\rho }^2{\mathrm{Var}}_i(a_i)+\text{sampling variance}}{{\rho }^2{\mathrm{Var}}_i(a_i)+\text{sampling variance}+{\mathrm{Var}}_{\mathrm{mod}}}\times 100\%.$$

When the artifact variance and sampling variance dominate the genuine-moderation variance — the empirical Schmidt-Hunter finding — this ratio is close to $100\%$. The empirical Schmidt-Hunter estimates of $\%{\mathrm{Var}}_{\mathrm{art}}\approx 80\%$ across diverse GMA-criterion literatures imply ${\mathrm{Var}}_{\mathrm{mod}}$ is roughly one quarter of the artifactual plus sampling variance, establishing validity generalization: the corrected validity $\rho$ is the constant operational validity of the predictor-criterion pair across the situations sampled. $\square$

Proposition 2 (Disattenuation can inflate sampling variance; the correction-then-pool ordering matters). Under the same setup as Proposition 1, disattenuating each study individually by $a_i^{-1}$ and then pooling yields a more efficient estimator than pooling the observed validities and then disattenuating by the mean attenuation, provided the per-study attenuation factors $a_i$ are heterogeneous. Equivalently, the standard Schmidt-Hunter procedure of correcting each primary validity before inverse-variance pooling is statistically efficient; the alternative of pooling-then-correcting is not.

Proof. Let ${\hat{\rho }}_i$ have asymptotic variance ${\mathrm{SE}}^2({\hat{\rho }}_i)=(1-{\rho }_i^2{)}^2/(n_i-1)$. After disattenuation, ${\hat{\rho }}_i^{(c)}={\hat{\rho }}_i/a_i$ has asymptotic variance ${\mathrm{SE}}^2({\hat{\rho }}_i)/a_i^2$ by the delta method. The inverse-variance-weighted pooled estimator is

$${\hat{\rho }}_{\mathrm{pool}}^{(c)}=\frac{{\sum }_i{w}_i{\hat{\rho }}_i^{(c)}}{{\sum }_i{w}_i},w_i=\frac{1}{{\mathrm{SE}}^2({\hat{\rho }}_i^{(c)})}=\frac{a_i^2}{{\mathrm{SE}}^2({\hat{\rho }}_i)}.$$

Studies with less attenuation (larger $a_i$, i.e. better reliability, less range restriction) receive more weight because their corrected estimates are more precise. The pooled estimator has asymptotic variance $1/{\sum }_i{w}_i$.

By contrast, the pool-then-correct estimator computes ${\bar{\rho }}_{\mathrm{obs}}={\sum }_i{n}_i{\hat{\rho }}_i/{\sum }_i{n}_i$ (sample-size weighting on observed validities) and then applies the mean attenuation ${\bar{a}}^{-1}$. This estimator ignores the heterogeneity in $a_i$: it weights a tightly-attenuated study (small $a_i$, large corrected variance) the same as a loosely-attenuated study (large $a_i$, small corrected variance) of the same sample size, which is statistically inefficient. The asymptotic variance of the pool-then-correct estimator exceeds that of the correct-then-pool estimator by a factor that scales with ${\mathrm{Var}}_i(a_i)/{\bar{a}}^2$.

Consequently, the Schmidt-Hunter procedure — correct each study for the four artifacts and then pool the corrected validities by inverse corrected variance — is the statistically efficient estimator. The alternative procedure, while sometimes used in informal syntheses, loses efficiency proportional to the heterogeneity of artifact levels across studies. This is the formal grounding for the meta-analytic protocol that the Schmidt-Hunter program standardised. $\square$

Connections Master

• Industrial and organizational psychology survey 29.15.01 is the chapter anchor for the present unit and supplies the broader I/O framework — selection, appraisal, motivation, leadership, job design — within which personnel selection sits as the most-studied application. The present unit deepens the survey's brief treatment of validity generalization into the full Schmidt-Hunter program, with the artifact-correction framework, the meta-analytic findings, and the Project A field validation all presupposing the personnel-selection apparatus that the survey introduces. The Schmidt-Hunter evidence is the empirical spine of the survey's recommendation that hiring systems be built around cognitive-ability and structured-interview batteries.

• Intelligence testing history 29.13.04 supplies the construct of general mental ability ($g$) on which the entire Schmidt-Hunter program rests. Spearman's 1904 attenuation formula — the same paper that introduced $g$ — is the methodological foundation of validity generalization; the present unit presupposes 29.13.04's account of the positive manifold, the factor-analytic extraction of $g$, and the deviation-IQ metric. The two units converge on the operational validity of $g$ as the most-robust individual-differences construct in psychology: 29.13.04 establishes what $g$ is, the present unit establishes what $g$ predicts in the employment context.

• Psychometrics survey 29.13.01 provides the measurement-theoretic substrate that the artifact-correction framework presupposes. The classical-test-theory decomposition $X^{\ast }=X+\epsilon$, the definition of reliability as true-score variance over observed-score variance, and the Spearman attenuation formula are all developed in 29.13.01 before being applied here to the personnel-selection context. The present unit's intermediate definitions of reliability, range restriction, and validity generalization are the I/O face of the psychometric apparatus that 29.13.01 develops in full.

• Person-situation debate 29.08.04 is the personality-psychology counterpart to the validity-generalization debate. Mischel's 1968 situationist critique of trait psychology argued that cross-situational consistency of behaviour was too low to support trait constructs; the density-distribution resolution (Fleeson; Mischel-Shoda) showed that the apparent inconsistency was largely a sampling artifact, and that aggregated behaviour reveals strong trait coherence. The structural parallel to Schmidt-Hunter is exact: in both literatures, an apparent situational specificity dissolved once the cross-observation variance was decomposed into artifact versus genuine moderation. The two units together establish a general methodological principle for the individual-differences sciences.

Historical & philosophical context Master

Frank Schmidt and John Hunter, working at the University of Iowa from the mid-1970s onward, launched the validity-generalization program as a methodological response to the situation-specificity doctrine that had dominated personnel psychology since Edwin Ghiselli's 1966 compendium ^{[Ghiselli1966]} of occupational aptitude-test validities. The situation-specificity position held that the validity of a cognitive test for clerical performance in one firm need not transfer to another firm or another job, because each validity study was a locally-bound measurement of a locally-bound construct. Ghiselli's data — wide variability in reported validities across jobs and firms — appeared to support this position.

The Schmidt-Hunter 1977 paper in the Journal of Applied Psychology ^{[SchmidtHunter1977]} showed that approximately $80\%$ of the observed variance in the Ghiselli-era validity coefficients was attributable to four statistical artifacts: sampling error (most primary studies had $n<100$), criterion unreliability (supervisor ratings are noisy), range restriction (incumbents are a selected subsample), and dichotomisation (many studies scored the predictor pass-fail rather than continuously). Once corrected, the residual between-study variance was small, and the operational validity of GMA for job performance was approximately constant across jobs and firms. The situation-specificity hypothesis was an artifact of under-powered primary studies with noisy criteria.

The 1984 Hunter-Hunter paper in Psychological Bulletin ^{[HunterHunter1984]}, titled "Validity and Utility of Alternative Predictors of Job Performance," extended the synthesis to a century of research and reported the canonical GMA validity of $0.53$ corrected. The 1998 Schmidt-Hunter paper in Psychological Bulletin (124:262) ^{[SchmidtHunter1998]} is the canonical summary: $85$ years of research distilled into a single table of selection-method validities, with GMA at the top and graphology at the bottom. The Schmidt-Hunter-Pearlman 1985 Personnel Psychology paper ^{[SchmidtHunterPearlman1985]} provided the full methodological statement of the artifact-correction framework.

The program has a parallel legal history. Griggs v. Duke Power Co. (1971) ^[Griggs1971] established the disparate-impact doctrine under Title VII of the Civil Rights Act of 1964: a selection test with disproportionate impact on a protected group must be shown to be job-related. The 1978 Uniform Guidelines on Employee Selection Procedures ^[EEOC1978] codified the four-fifths rule and the three validation strategies. The Schmidt-Hunter evidence is the scientific input to the legal framework; the personnel-selection practitioner operates at the intersection of validity generalization and adverse-impact regulation. The Sackett-Lievens 2008 review ^{[SackettLievens2008]} and the Sackett et al. 2017+ updates ^{[Sackett2017]} reproduced the Schmidt-Hunter rank order with larger databases and extended the analysis to modern structured-interview and situational-judgment formats. The Campbell-Knapp Project A study ^{[CampbellKnapp2001]}, conducted at the scale of the United States Army with roughly $50,000$ recruits, provided the independent field validation. Frank Schmidt's death in 2014 and John Hunter's in 2022 closed the program's primary contributors; the modern updates by Sackett, Lievens, Berry, and others continue the line.

Bibliography Master

• Berry, Christopher M., Paul R. Sackett, and Shonna D. Waters. "Faking on Personnel Selection Tests." Psychological Bulletin 138, no. 6 (2012): 1083–1121.

• Brogden, Hubert E. "On the Interpretation of the Correlation Coefficient as a Measure of Predictive Efficiency." Journal of Educational Psychology 37, no. 2 (1946): 65–76.

• Campbell, John P., and David J. Knapp, eds. Exploring the Limits in Personnel Selection and Classification. Mahwah, NJ: Lawrence Erlbaum Associates, 2001.

• Cronbach, Lee J., and Goldine C. Gleser. Psychological Tests and Personnel Decisions. 2nd ed. Urbana: University of Illinois Press, 1965.

• Equal Employment Opportunity Commission, Civil Service Commission, Department of Labor, and Department of Justice. Uniform Guidelines on Employee Selection Procedures. 29 C.F.R. §1607. Washington, DC: U.S. Government Printing Office, 1978.

• Ghiselli, Edwin E. "The Validity of Occupational Aptitude Tests in Personnel Selection." Personnel Psychology 18, no. 2 (1966): 135–163.

• Griggs v. Duke Power Co., 401 U.S. 424 (1971).

• Hunter, John E. "Cognitive Ability, Cognitive Aptitudes, Job Knowledge, and Job Performance." Journal of Vocational Behavior 29, no. 3 (1986): 340–362.

• Hunter, John E., and Ronda F. Hunter. "Validity and Utility of Alternative Predictors of Job Performance." Psychological Bulletin 96, no. 1 (1984): 72–98.

• McDaniel, Michael A., Deborah L. Whetzel, Frank L. Schmidt, and Steven D. Maurer. "The Validity of Employment Interviews: A Comprehensive Review and Meta-Analysis." Journal of Applied Psychology 79, no. 4 (1994): 599–616.

• Roth, Philip L., Allen I. Huffcutt, and Philip Bobko. "Ethnic Group Differences in Measures of Job Performance: A New Meta-Analysis." Journal of Applied Psychology 88, no. 4 (2003): 694–706.

• Sackett, Paul R., and Filip Lievens. "Personnel Selection." Annual Review of Psychology 59 (2008): 419–450.

• Sackett, Paul R., Chu Zhang, Christopher M. Berry, and Filip Lievens. "Doing What We Do Now, But Better: A Commentary on Macan and Merritt." Personnel Psychology 70 (2017): 393–407.

• Schmidt, Frank L., and John E. Hunter. "Development of a General Solution to the Problem of Validity Generalization." Journal of Applied Psychology 62, no. 5 (1977): 529–540.

• Schmidt, Frank L., and John E. Hunter. "The Validity and Utility of Selection Methods in Personnel Psychology: Practical and Theoretical Implications of 85 Years of Research Findings." Psychological Bulletin 124, no. 2 (1998): 262–274.

• Schmidt, Frank L., John E. Hunter, and Kenneth Pearlman. "Progress in Validity Generalization: 1975–1985." Personnel Psychology 38, no. 2 (1985): 265–285.

• Spearman, Charles. "The Proof and Measurement of Association between Two Things." American Journal of Psychology 15, no. 1 (1904): 72–101.

• Taylor, Hazel C., and John T. Russell. "The Relationship of Validity Coefficients to the Practical Effectiveness of Tests in Selection: Discussion and Tables." Journal of Applied Psychology 23, no. 5 (1939): 565–578.

• Thorndike, Robert L. Personnel Selection: Test and Measurement Techniques. New York: John Wiley & Sons, 1949.