Predictive Validity in Intelligence Tests: What It Shows

What does predictive validity actually support?

Predictive validity is evidence that scores from a specified test are related to a later outcome for a defined group under stated conditions, in a way that supports a particular interpretation or use. The Standards for Educational and Psychological Testing (2014) frame validity around evidence and theory for proposed score interpretations and uses; validity is not a permanent quality conferred by calling an instrument an intelligence test. To judge a claim, the reader needs to know what score was studied, which outcome followed, who took the test, how it was administered, and what decision the evidence is meant to support.

A coefficient summarizes how scores and outcomes vary together across the studied observations. It does not say that a person with a given score has that same probability of succeeding, establish that the score caused the outcome, or authorize the same interpretation in a different population. In an explicitly illustrative case, a relationship between scores and later course grades could be relevant to a similar course and student group; it would not alone show what one applicant will earn, or establish usefulness for a workplace decision. Those questions require evidence about the relevant outcome and setting.

Reports may give an observed correlation and a corrected meta-analytic estimate. The observed value summarizes the associations available in the included samples; a corrected value estimates what the relationship might be under specified adjustments for features such as measurement error or restricted score range. The corrected figure is therefore not another direct observation, and its meaning depends on the assumptions and procedures used. Neither form turns a group relationship into an individual forecast. A stable, substantial relationship can still inform a well-matched decision; bounding the inference preserves that value while preventing it from traveling automatically across tests, populations, or outcomes.

The distinction matters when comparing school grades with work outcomes. Each is a different criterion, assembled under different conditions, and evidence about one does not silently validate a score use in the other. The next question is therefore not simply whether intelligence scores predict, but what the school-grade evidence actually pooled and whether its design establishes that the score came before every grade it relates to.

Sources: Standards for Educational and Psychological Testing (2014), full text

What does the school-grade estimate measure?

The meta-analysis *Intelligence and school grades: A meta-analysis* reports a sample-size-weighted observed correlation of r=.44 between standardized intelligence-test scores and school grades, alongside a corrected population estimate of ρ=.54. These are summaries of associations across the studies Roth and colleagues included, not a forecast that a particular student will attain a particular grade. The observed coefficient describes the pooled empirical relationships under the analysis’s weighting; the corrected estimate adjusts for sampling error, measurement error, and range restriction. That adjustment is an estimate under correction assumptions, not an additional set of directly observed scores and grades. The correction attempts to account for factors that can make observed correlations smaller than an underlying population relationship, but the resulting value remains tied to those analytic choices. It should not be read as the grade variance explained for every student or school.

The timing matters to the word predictive. The accessible publisher abstract for *Intelligence and school grades: A meta-analysis* reports the correlations and corrections, but does not establish that every included intelligence test was administered before the grades used as criteria. Accordingly, the pooled r=.44 and ρ=.54 support describing an average score-grade association in the assembled evidence; they do not, on the accessible record, establish uniform prospective prediction. This limitation narrows the claim rather than removing the result: an association across included samples can be informative even where the abstract does not document the temporal order study by study.

A separate paper, *The Predictive Validity of Four Intelligence Tests for School Grades: A Small Sample Longitudinal Study*, is explicitly identified as longitudinal and concerns later grades. That design label makes it a distinct example of temporal ordering as a research question. Its full text was not accessible in the recorded research, so this section does not report its sample details, coefficients, or findings. The contrast is methodological: a study designed around test scores and subsequent grades can address timing directly, whereas a broad meta-analytic association should not be assumed prospective unless the underlying studies’ sequence is verified. The paper’s title alone cannot supply results or establish how broadly they generalize.

Grades are meaningful outcomes, but they are not a pure readout of reasoning ability. They reflect performance on assessed coursework within particular instructional and institutional settings; the relationship can therefore include what students learned, what was taught, how assessment was conducted, and other context. This is an interpretation of what the criterion represents, not a causal decomposition measured by the meta-analysis. The pooled coefficient does not separate those contributions or show that changing a test score would change a grade.

Taken carefully, the estimate says that the included studies show a notable average association between standardized intelligence measures and school grades, with a higher corrected estimate under the analysis’s adjustments. It does not establish that each study tested future grades, that the association is identical across subjects or ages, or that it specifies an individual student’s outcome. Those boundaries leave the pooled finding useful: it describes a broad relationship in the assembled evidence while keeping temporal claims and personal forecasts dependent on evidence designed to support them.

Sources: Intelligence and school grades: A meta-analysis; The Predictive Validity of Four Intelligence Tests for School Grades: A Small Sample Longitudinal Study

Why do school-grade relationships vary by subject and age?

A school-grade coefficient is most useful as evidence about a criterion resembling the one a reader cares about. In *Intelligence and school grades: A meta-analysis*, the reported association varied across subject areas, grade levels, test types, and publication years. Those moderator patterns qualify the pooled estimate described above: it is a summary across unlike studies, not a single relationship that can be carried unchanged from one course or age group to another. The meta-analysis documents variation between study conditions; it does not establish that any one moderator caused the differences.

The reason is partly criterion alignment. A broad intelligence measure samples abilities used across many tasks, while a grade records performance in a particular subject, at a particular stage, under a particular curriculum and assessment system. They can share demands—such as learning, reasoning, or working with information—without measuring the same thing. A mathematics grade, for example, reflects performance in the assessed mathematics course; it is not interchangeable with a general ability score. The score may relate to that outcome, but the strength and meaning of the relationship depend on what the course actually asks students to do and how achievement is recorded.

Age and grade stage also change the setting in which a criterion is produced. Early schooling, later secondary study, and advanced coursework differ in content, instruction, assessment, and the opportunities students have had to learn the tested material. A standardized ability test and a grade may therefore overlap differently across stages. The Roth meta-analysis identifies grade-level and publication-year variation, but its reported moderator analysis does not isolate whether curriculum, selection into studies, test composition, or another feature explains each difference. These are plausible contextual considerations, not causes established by that synthesis.

For a reader comparing evidence, ask how closely the studied criterion matches the actual course, grade stage, or placement outcome at issue. Evidence drawn from a similar subject and educational level is more directly informative than a pooled result spanning distant criteria. This is a bounded comparison rule: similarity improves relevance, but it does not make a group association an individual forecast or show that ability determines grades. Nor does a moderator average tell a school what will happen locally. The synthesis reports patterns across studies; a local decision still depends on its students, course, assessment, and the evidence supporting that particular use. It is also useful to distinguish the label of a predictor from the content of the criterion: two studies may both call their measure an intelligence test while using different test formats, and two grade records may represent different subject demands. The moderator findings warn against treating those labels as sufficient evidence of equivalence. They do not supply a conversion from a test score to a course grade, or show that choosing a more closely matched test would itself improve educational outcomes.

Sources: Intelligence and school grades: A meta-analysis

Why did the older job estimate change?

The familiar .51, the .31 re-estimate, and the 2026 work estimates describe different evidence and analytic histories. They should not be arranged as three readings of one fixed quantity. The .51 is the frequently cited historical estimate from an earlier synthesis; the .31 figure is reported in the later discussion of a reanalysis of older meta-analyses; and the 2026 synthesis calculates contemporary estimates from a separate, newer evidence base. Differences between them reflect both the studies available and the assumptions used to correct their results.

*Revisiting meta-analytic estimates of validity in personnel selection: Addressing systematic overcorrection for restriction of range* (2022) reexamined prior personnel-selection meta-analyses. Its authors argued that common artifact-distribution procedures could overcorrect validity estimates, and reported revised estimates lower by roughly .10 to .20 validity points on average. The .31 comparison belongs to that older evidence and reanalysis. It is not a result computed from the 2026 contemporary dataset, and it should not be presented as a direct rival estimate based on the same studies, period, or procedure.

Range restriction explains one part of the dispute. In a hiring study, researchers may observe only people who were selected and hired. If selection favored applicants with higher scores on a predictor, the hired group can have a narrower spread of scores than the original applicant pool. Because correlations can be affected by how much predictor variation is present, meta-analysts may correct the observed relationship to estimate what it might have been across a broader applicant group. That corrected figure is an estimate under a model, not a newly observed correlation among all applicants.

The correction depends on information or assumptions about how much the range narrowed and how the restriction process relates to the predictor and outcome. If the assumed restriction is too strong or otherwise mismatched to the data, the correction can shift the estimated relationship too far upward. The 2022 paper’s central challenge was that some procedures used to model artifact distributions systematically produced overcorrection in the older selection literature. Revising those assumptions lowered several estimates. This is a methodological disagreement about how to recover a broader relationship from selected samples, as well as a disagreement about estimate size; it is not evidence that no relationship exists.

The same paper reports that much of the relative ordering of selection procedures remains similar after its revisions, even though several average validity estimates decline. That distinction matters: changing the estimated magnitude does not necessarily reverse every comparison among methods. But a preserved ordering cannot validate a particular instrument for a particular job or guarantee an individual outcome. The estimate and the ranking answer different questions, and both remain bounded by the historical studies and correction method under review.

The 2026 article, *Is general mental ability still the best predictor of job performance? Integrating contemporary meta-analytic evidence*, places the .31 older-data figure alongside its own newer analysis while identifying their distinct bases. Its contemporary sample is not a replication of the 2022 reanalysis, and the older .51 and .31 should not be collapsed into a trend line as if only the calendar changed. The studies, measures, and correction choices differ. Thus a careful comparison preserves labels: .51 as the historical synthesis figure, .31 as an older-data re-estimate discussed in the later literature, and contemporary values as outputs of a separate synthesis. None is a universal coefficient for all work settings.

Readers should therefore interpret disagreement by asking what data were reanalyzed and what range-restriction assumptions were applied, rather than choosing the largest or newest number by default. Sackett et al. (2022) did not erase the observed association between cognitive ability and job outcomes; they argued that some earlier corrections overstated its magnitude. The 2026 synthesis addresses contemporary evidence separately. Keeping these comparisons distinct prevents a methodological revision to older studies from being mistaken for a direct estimate of present-day job performance, while preserving the substantive point that the size of a validity estimate depends on both evidence base and analytic procedure.

Sources: Revisiting meta-analytic estimates of validity in personnel selection: Addressing systematic overcorrection for restriction of range; Is general mental ability still the best predictor of job performance? Integrating contemporary meta-analytic evidence; Validity and utility of selection methods in personnel psychology: Practical and theoretical implications of 85 years of research findings

What do contemporary work estimates actually say?

The 2026 synthesis, “Is general mental ability still the best predictor of job performance? Integrating contemporary meta-analytic evidence,” reports two different kinds of estimate. Across its contemporary set of 212 studies and 109,396 participants, the observed mean correlation was .14 when studies were weighted by sample size and .15 when each study received equal weight. After the authors applied corrections intended to estimate operational validity, results varied from .18 to .22 across four combinations of correction and weighting choices. Their preferred summary is .20: equal weighting under their preferred restriction-of-range correction. So .14 is the observed N-weighted mean, not the corrected estimate; .18–.22 is the corrected range, not an observed mean. Keeping those labels attached is essential, because each number answers a different synthesis question.

Weighting determines how much influence each study has. N-weighting gives larger samples greater influence, while equal weighting gives each included study the same influence regardless of size. Neither is a purely clerical choice: the resulting average can move when study sizes and estimates differ. The paper reports .14 under N-weighting and .15 under equal weighting before correction, and .18 versus .20 under its preferred correction restriction afterward. Those paired values make the choice visible instead of hiding it inside a single headline coefficient. Equal weighting is the authors’ preferred summary for this analysis; that preference should be reported as their analytic judgment, not mistaken for a result that follows automatically from the data.

Correction changes the target as well. An observed correlation summarizes the relationships available in the assembled studies. Operational-validity estimates adjust for artifacts such as imperfect measurement and restriction of range, under assumptions about how those artifacts arose. Different plausible correction and weighting combinations therefore yield nearby but non-identical values, .18–.22. These are not four separate outcomes experienced by workers; they are estimates produced by alternative analytic choices applied to study evidence. The preferred .20 is useful as the paper’s considered contemporary summary because the authors state the correction and weighting behind it. It is still a synthesis estimate with assumptions, not a constant that applies unchanged to every employer, role, test, or performance measure.

The older .31 figure belongs in a separate time-and-evidence comparison. In the 2026 paper, .31 is attributed to Sackett and colleagues’ 2022 reanalysis of older personnel-selection meta-analyses; it was not calculated from the newer 212-study set. The reanalysis questioned correction practices used with older datasets, while the contemporary synthesis assembles recent evidence and makes its own correction choices. Combining .31 with .20 as though the two were rival estimates from one common pool would erase differences in source studies, period, and analytic context. Reporting them separately lets readers see both the methodological revision to the historical literature and the newer synthesis without claiming that time alone caused the difference.

A coefficient of .20 is a correlation-scale summary of how scores and a defined work criterion covary across evidence; it does not mean that 20 percent of people succeed, that a particular applicant has a one-in-five chance of success, or that a score causes performance. The 2026 authors recommend .20 as a judgment-call summary after presenting the observed and corrected alternatives. That recommendation offers a clear contemporary reference point, while the range and choices show why it should not be quoted without its method. For a specific decision, the relevant question remains whether the studies’ tests, people, jobs, and performance criteria resemble the proposed use. This synthesis supports a bounded group-level relationship, not an individual forecast or a universal hiring rule.

The observed-to-corrected shift should also not be read as proof that actual workers’ performance correlations are secretly higher by a fixed amount. The operational estimate depends on how the synthesis models measurement error and range restriction, and those assumptions target a corrected relationship rather than a newly observed workforce. The 2026 paper’s presentation is useful precisely because it exposes both layers: readers can see the directly summarized observed association and the adjusted estimates authors consider relevant for operational validity. If a later account reports only .20, it should retain the fact that this is the preferred corrected, equally weighted estimate under a stated restriction correction. If it reports .14, it should call that the N-weighted observed mean. Accurate labels allow comparison without turning methodological adjustments into claims about individual outcomes.

Sources: Is general mental ability still the best predictor of job performance? Integrating contemporary meta-analytic evidence

Why is the work criterion harder to generalize?

The coefficient changes meaning with the outcome being predicted. The 2026 synthesis, “Is general mental ability still the best predictor of job performance? Integrating contemporary meta-analytic evidence,” focuses its contemporary work estimate on overall job performance. That criterion is not interchangeable with a training score or a measure of technical proficiency. A supervisor’s overall rating, demonstrated learning during training, and accurate completion of a technical task each represent a different outcome. Evidence linking a cognitive measure to one cannot simply be carried over to the others, even when all are described informally as doing well at work. The paper’s lower contemporary estimates for overall job performance therefore qualify the claim about that criterion; they do not establish that cognitive measures have no relationship with learning or technical skill.

The distinction matters because criteria capture different parts of work. Overall performance can aggregate multiple responsibilities and judgments, while training performance concerns learning or results during a training context, and technical proficiency concerns execution of particular tasks. Those labels identify different questions, not successive names for one latent result. A correlation with a broad rating cannot tell a reader how strongly a score relates to a specific course assessment, equipment operation, or accuracy standard unless that outcome was itself studied. Conversely, evidence about training does not automatically establish prediction of later overall performance. The 2026 synthesis makes the overall-performance estimate the relevant result for its stated criterion; it should not be stretched to answer every workplace question.

The same synthesis reports job-complexity moderator findings using O*NET-derived Data, People, and Things dimensions. It finds a weak relation between the Data dimension and validity (r = −.04), a positive relation for Things (r = .29), and a positive Data-by-Things interaction: the Data relationship is more positive where Things demands are higher. These are occupational-level patterns in the assembled study evidence, not person-level effects or coefficients for any named job. They suggest that a single overall estimate can conceal variation associated with how jobs are characterized, but they do not provide a ready-made prediction for an individual occupation or worker.

The paper interprets Things-oriented work as including demands involving equipment and information processing. That is the authors’ interpretation of the occupational dimension, not a demonstrated causal mechanism showing that equipment or information processing produces a stronger score-performance link. The interaction also cautions against reading the Data and Things results as isolated rules: the reported Data relationship differs with the level of Things demands. In practical terms, a job label alone does not identify the exact criterion or establish that this moderator pattern applies locally. The result is a reason to attend to job and outcome definitions when reading a pooled coefficient, not a basis for ranking occupations or deciding who will perform well.

A separate training-validity estimate reported elsewhere in the article should therefore be read as evidence about training performance, not substituted for the contemporary overall-performance result. The two findings can coexist because they address different criteria and evidence bases. Lower validity for the overall-performance criterion does not erase evidence concerning learning or technical proficiency; neither does a training relationship guarantee later job performance. A training result concerns performance in that learning context, while technical proficiency names a distinct task-related outcome; neither label should be silently converted into a broad supervisor judgment. For any interpretation, first name the outcome the coefficient represents. Only then can a reader judge whether it speaks to the question at hand. The 2026 synthesis supplies a contemporary summary and useful occupational moderators, while its criterion boundaries prevent that summary from becoming a universal measure of work success.

Sources: Is general mental ability still the best predictor of job performance? Integrating contemporary meta-analytic evidence

What makes a work criterion relevant?

A test-outcome link is pertinent when the evidence supports the interpretation being made for a sufficiently similar decision context. The *Standards for Educational and Psychological Testing (2014)* treats validity as evidence and theory supporting a proposed interpretation and use of scores. That framing makes the question more specific than whether a test has ever correlated with an outcome: what score meaning is claimed, for which people, under what conditions, and for what decision? Evidence for one interpretation cannot simply be carried to another because both involve a test score.

The criterion must represent the outcome the decision actually concerns. Imagine an institution saying that a reasoning score predicts success in a particular training program. That is an illustration, not a reported case. Evidence about completion, demonstrated learning, or a defined performance assessment in a comparable program would address the claim more directly than a vague label such as “success,” whose meaning could shift between institutions. If the criterion is a supervisor’s broad judgment, a course mark, or a later job outcome, each answers a different question. The institution’s chosen outcome should be defined well enough to interpret what a score-outcome association means.

Population and test comparability also limit the inference. A relationship documented among people entering one kind of program may not carry unchanged to a different age group, selection process, or level of prior preparation. Likewise, evidence from a test with different content, score construction, or administration does not automatically establish the relevance of another test that shares a broad label. The Standards’ interpretation-and-use framework directs attention to whether the evidence fits the intended population and use; it does not supply a universal coefficient or guarantee that a study sample represents every local applicant group.

Administration conditions and criterion quality matter because they shape what the recorded scores mean. If the test is administered under materially different conditions, or the outcome is recorded inconsistently, the observed link may not reflect the intended comparison. In the employment context, the *Uniform Guidelines on Employee Selection Procedures, 28 CFR 50.14* provide a U.S.-specific example of guidance that connects selection procedures with job-related criteria and validation evidence. They are not global law and do not prove that any particular intelligence test predicts a particular job outcome. Their relevance here is the bounded principle that a claimed employment use should be connected to the work and criterion at issue.

A score is also one source of information within a decision, not a complete account of a person or a guaranteed forecast. Evidence that a group-level relationship exists may support limited use when the test, people, conditions, and outcome are reasonably matched; it still leaves uncertainty about any individual. Other relevant evidence can inform a decision alongside a score without turning a statistical association into certainty. These questions are tools for understanding whether an institution’s claim is well matched. They are not a demand that an individual reader conduct a validation study; the organization relying on a consequential score is responsible for explaining the evidence and how it supports that use.

Sources: Standards for Educational and Psychological Testing (2014), full text; Uniform Guidelines on Employee Selection Procedures, 28 CFR 50.14

What should a reader do with a score?

For curiosity, use Test IQ Free’s free 50-question test for low-stakes reasoning practice and review its worked answers. Its displayed IQ range is a provisional estimate from an unvalidated conversion with no validated adult reference population; it is not a confidence interval or established population comparison. The optional $9 report explains the same conversion, so payment adds explanation, not validity. The form samples pattern reasoning, number patterns, and spatial reasoning, so its questions can give you material to inspect: which operation you recognized, where an answer went off track, and how the worked solution proceeds. Treat that as feedback about these items and your approach on this attempt, rather than a stable ranking of your general ability.

For a consequential education or employment decision, ask the institution relying on a score to identify the outcome, population, evidence, administration conditions, and decision rule behind its use. A browser result from this test should not guide high-stakes education, employment, diagnosis, or gifted-identification decisions. Seek a qualified assessment and context-specific evidence instead. Worked feedback can still help someone learn how they approached reasoning items; that useful practice is different from forecasting an individual’s consequential outcome.

Sources: Test IQ Free assessment

Questions readers ask