What the 2024 Review Changed About Adult Cognitive Score Stability

What exactly did the 2024 review establish?

The 2024 review strengthened the case that adults tend, on average, to retain similar relative standing on cognitive measures over time. It did not establish that an adult’s own score remains unchanged, or that every test has the same stability. The central result is a pooled, model-based estimate: at a modeled mean age of 20 and a five-year retest interval, rank-order stability was ρ = .76. Here, rank-order stability means the degree to which people’s relative positions in a group correspond across two testing occasions. The estimate describes an association across people; it is not a predicted score change for any one person. The scope helps explain what changed. Earlier syntheses described in “The Stability of Cognitive Abilities: A Meta-Analytic Review of Longitudinal Studies” covered narrower questions, including twin studies or particular intelligence-test scores. The 2024 review instead brought together longitudinal evidence across a wider age span, cognitive abilities, instruments, and populations, then examined whether age, time between tests, measured ability, instrument, and other study features moderated the associations. Its abstract reports 205 studies, 87,408 participants, and 1,288 test–retest correlations for observed scores. Those totals show the breadth of the assembled evidence, but they do not mean that every age, ability, or test was equally represented. The review’s adult estimate should therefore be read with its conditions attached. The value ρ = .76 is a model estimate for age 20 and a five-year interval, rather than a direct average from one study or a universal coefficient for adults. Its question was how strongly measured differences between people were ordered similarly at two occasions, and how that association varied across the lifespan and study conditions. It was not a test of whether the average score stayed constant, nor a diagnostic check of whether a particular person’s score changed beyond measurement error. Those questions require different evidence. The broader age coverage also matters. The authors report low rank-order stability in preschool children, a rapid rise during childhood, and consistently high average stability from late adolescence through late adulthood. They also report that stability declined as the interval between assessments grew, with that decline flattening at about five years. These patterns replace an unqualified statement such as “adult cognitive scores are stable” with a more useful one: relative standing is often persistent in adult samples, while the estimate depends on age, interval, ability, and measurement. “High” in this context describes a group-level correlation, not certainty about any person. The paper thus changes the level of precision in the claim more than its basic direction. It supports substantial average continuity in adults while making clear that stability is conditional and varies across the evidence base. Its modeled result does not establish an individual score’s shelf life, certify a particular instrument, or show that cognitive ability cannot change. For a reader comparing results, the relevant question is not simply whether adults are stable. It is whether the studies behind the estimate resemble the person, test, score, and interval at issue.

Sources: The Stability of Cognitive Abilities: A Meta-Analytic Review of Longitudinal Studies (author postprint); The Stability of Cognitive Abilities: A Meta-Analytic Review of Longitudinal Studies

Why can a group stay similarly ordered while scores move?

A correlation can remain high even when scores change because it tracks whether people keep similar relative positions, not whether their scores stay numerically identical. In The Stability of Cognitive Abilities: A Meta-Analytic Review of Longitudinal Studies, rank-order stability means the persistence of individual differences across two testing occasions. The authors describe it as typically assessed with a test–retest Pearson correlation: a summary of how scores on one occasion vary alongside scores on another. A high correlation therefore says that people who scored relatively higher at the first occasion tended, on average, to score relatively higher at the second. It does not report the direction or size of every person’s change.

Mean-level change asks a different question: did the group’s average score rise or fall? Individual-level change asks how a particular person’s score changed. Neither answer follows from rank-order stability alone. Consider an explicitly illustrative group of four people whose scores are 10, 20, 30, and 40 on one occasion. If every score increases by five points by the next occasion, the average rises by five, but the ordering is unchanged and the correlation is perfect. That example shows why a rising average can coexist with stable relative standing; it is not a description of participants or a result reported by the review.

The same distinction holds when changes are unequal. Suppose the four illustrative scores become 12, 19, 34, and 37. The average still rises, and one person’s score falls, but the broad ordering remains similar. A correlation summarizes the pattern across the group; it does not tell us that every person changed by the same amount or that each person stayed close to their original score. If some people move past others, that reordering can lower the correlation, especially when changes are large relative to the differences between people. Yet a high overall correlation can still coexist with meaningful changes for some individuals.

The 2024 review makes this distinction central to its subject. Its ρ = .76 estimate for a modeled age of 20 and five-year interval is an average rank-order association drawn from longitudinal studies, not an estimate of how many points one adult should gain or lose. The authors note that rank-order analyses can indirectly reflect whether many people undergo large changes relative to others, because widespread reordering affects the association. That indirect group-level signal is not the same as directly measuring each participant’s change. A person could change substantially while many peers change in parallel, leaving relative positions mostly intact; another person could change little while peers’ scores shift around them.

So “stable” needs its object named. The group may show stable relative ordering, while its average moves; a person’s observed score may change, while the group correlation remains high. These statements are compatible because they describe different quantities. A changed raw result—the count or value before interpretation against a norm group—does not by itself show a lasting change in ability. Nor does a high population-level correlation establish that an individual’s result was unchanged. To interpret a personal difference, a reader needs evidence about the exact measure, its precision, the time between administrations, and whether conditions and scoring were comparable. The meta-analysis answers a broad question about continuity in relative standing; it cannot settle those individual questions by itself.

Sources: The Stability of Cognitive Abilities: A Meta-Analytic Review of Longitudinal Studies (author postprint)

How do age and retest interval change the estimate?

The 2024 review found that cognitive scores were less stable in early childhood, became more stable through development, and showed high modeled average stability from late adolescence through late adulthood. Stability also declined as the gap between tests grew, with the decline flattening at about five years. These are patterns in rank-order correlations across studies, not a timetable for how long one person's score remains useful. Age and elapsed time matter, but neither supplies an individual forecast.

The age pattern comes from a model fitted across longitudinal research, rather than one group followed continuously from infancy to old age. In The Stability of Cognitive Abilities: A Meta-Analytic Review of Longitudinal Studies, the authors analyzed 205 studies, 87,408 participants, and 1,288 test-retest correlations. Their model described low stability in preschool children, a rapid increase during childhood, and consistently high average stability from late adolescence through late adulthood. The broad coverage addresses uncertainty about whether stability might decline again in very old age. The model found high average stability into late adulthood, while allowing variation among studies and measures. It does not mean every adult ability, instrument, or individual follows the same curve.

Retest interval changes the estimate because a longer span gives more time for differences between people to shift. The review found a declining association as intervals lengthened, but the decrease was not linear: it flattened from around five years onward. The model therefore does not support a simple rule such as “a score expires after five years.” Five years marks the approximate point where the average modeled decline became less steep, not a recommended retesting interval or a guarantee that a particular result remains current. Nor does a flatter curve mean that no meaningful changes occur in particular abilities or people.

For the model's age-20 and five-year reference point, the review estimated mean rank-order stability at ρ=.76. It also examined a modeled rₜₜ=.80 benchmark, described as the minimum stability sufficient for individual-level diagnostic decisions. The age and interval models indicated that adults met this benchmark over intervals exceeding five years. That is a modeled group-level pattern, not a clinical rule licensing an individual decision. It also does not establish that every instrument in the review has the precision or validity needed for that decision. A meta-analysis threshold cannot replace evidence about a specific test and its proposed use.

The headline estimate sits amid substantial between-study variation. For the age-20, five-year estimate, the paper reports I²=98.661, indicating that study estimates varied considerably beyond sampling error alone under the analysis. This describes differences among studies, not the probability that one person's score will change or a personal uncertainty range. The pooled curve describes an average pattern across the assembled evidence; the heterogeneity cautions against treating it as universal. Samples, instruments, abilities, and procedures can differ even when the broad age and interval trends remain visible.

For a personal comparison, elapsed time is context rather than a verdict. Two scores collected a short time apart may differ because of measurement variation or conditions; scores years apart may preserve relative standing on average while one person's observed performance changes. To judge a change, a reader needs evidence for the instrument, score type, and intended interpretation, including its precision and conditions of repeated use. The 2024 review narrows the claim that adult cognitive scores “stay stable”: relative standing is generally persistent in modeled adult samples, and stability depends on age and interval. The calendar alone cannot show whether an individual's score changed meaningfully.

Sources: The Stability of Cognitive Abilities: A Meta-Analytic Review of Longitudinal Studies (author postprint)

Why do ability and instrument matter?

The 2024 review did not find one stability value that applies to every cognitive measure. At its modeled reference point, age 20 and a five-year retest interval, estimates for general intelligence and knowledge-based abilities were about .77 to .80, while estimates for abilities relying more on effortful processing were about .65 to .75. These are pooled subgroup estimates across studies and measures, not coefficients for a particular instrument or a forecast for any individual. In “The Stability of Cognitive Abilities: A Meta-Analytic Review of Longitudinal Studies,” the authors found general intelligence most stable, with comprehension knowledge, quantitative knowledge, and reading and writing similarly stable. Fluid reasoning, learning efficiency, and working-memory capacity tended to show lower stability. The labels describe broad classes of measured ability: knowledge-based tasks draw on acquired information, while effortful-processing tasks make demands such as reasoning through unfamiliar material or maintaining and manipulating information. They do not divide every test neatly into two kinds. The meta-analysis pooled measures assigned to ability categories; its estimates summarize between-person retest correlations across the available studies, not changes in one person's capacity. Those differences may reflect both the abilities and the instruments used to measure them. The review reports that reliability adjustment reduced some gaps between ability categories. In particular, the difference between general intelligence and fluid reasoning became smaller after adjustment, suggesting that differences in test consistency accounted for part, but not all, of the observed contrast. The authors discuss test features that could matter: knowledge measures may be less speeded, while processing measures may impose strict time limits; test length and the spread of item difficulty can also affect reliability. That explanation is plausible within the paper's analysis, but the meta-analysis does not isolate a single feature as the cause of the subgroup pattern. The same caution applies to the word “instrument.” A test made from one narrow task samples a different range of performance than a battery that combines several subtests and reports both domain scores and a composite. The authors note that batteries differ in the number and type of tasks, how they are presented, testing setting, reliability, abilities represented, and how component scores are weighted. Shortening a battery or selecting only a few subtests can also change the score's content and precision. Consequently, a repeated score can reflect continuity on the sampled tasks without representing every cognitive ability a broader assessment would cover. Repeating the same instrument and switching instruments are also different comparisons. Reusing a test can preserve a more direct link between occasions, but familiarity may affect performance. Using a different test may reduce some retest effects, yet the relationship between scores then depends partly on how closely the two tests measure overlapping abilities. The review notes that a retest correlation across different instruments is limited by the concurrent association between them. Its broad average therefore cannot be transferred automatically to a new test, a shortened form, or a browser quiz simply because each is described as an intelligence or reasoning test. For a reader comparing two results, the practical first question is whether they sampled comparable content and used comparable scoring. Check whether both scores came from the same instrument and scale, whether a battery or selected subtests were used, and whether speed, response format, or testing conditions differed. If the measures changed, a difference may combine actual performance change with differences in what was asked and how the results were scored. The 2024 findings make the blanket phrase “adult cognitive scores stay stable” more conditional: average rank-order stability varies with the ability and the test. They do not tell a person that one specific score is stable, changed, or suitable for a consequential decision.

Sources: The Stability of Cognitive Abilities: A Meta-Analytic Review of Longitudinal Studies (author postprint)

What does reliability adjustment fix, and what does it leave?

Reliability adjustment estimates how measurement error may weaken an association between scores taken at different times. It helps researchers ask whether differences in stability across ages, abilities, or tests arise because measures differ in consistency. It does not remove genuine change, make unlike tests interchangeable, or calculate a corrected result for one person. The 2024 review used it as a robustness check, not a conversion rule for an individual's score.

In Test Reliability—Basic Concepts, reliability means consistency of test scores across specified occasions, editions, or raters. The conditions matter: consistency across administrations of one form does not establish consistency on a different task or occasion. Measurement error is variation in observed scores not attributable to the intended signal under the model being used. It can weaken the correlation between measurements. If error also differs across groups or measures, it may distort comparisons between correlations. The meta-analysis explains that correcting for estimated reliability can test this possibility; it cannot show that remaining variation is error-free or that a test measures the intended ability well.

The size and limits of the review's check matter. Reliability estimates were available for 250 of the 1,288 effect sizes, from 45 samples. The authors reran relevant moderator analyses in this subset, both before and after adjusting correlations and their variances for measurement error. Adjusted stability estimates were higher than unadjusted estimates by roughly .06 to .11 across the abilities discussed. Some differences between ability categories and instruments also changed in statistical significance after adjustment. This supports the view that measurement consistency can contribute to apparent differences in stability; it does not establish that reliability explains every difference or that adjusted estimates are the best values to apply elsewhere.

The authors caution that the reliability-data subset was much less comprehensive and might not represent all studies. They compared adjusted results with unadjusted results from the same subset and with the complete data. This makes the adjustment informative as a sensitivity analysis: it shows how conclusions shift under one treatment of measurement error. Missing reliability information limits which studies enter the corrected analysis, and the adjustment cannot recover information original studies did not report. The result remains a pooled estimate, not a personal score correction.

Reliability answers a different question from validity. Reliability concerns consistency under stated conditions; validity concerns the evidence and reasoning that support a particular interpretation and use of scores. The Standards for Educational and Psychological Testing says each intended score interpretation should have a rationale and summary of relevant evidence and theory. A consistent result alone cannot establish that a score supports a diagnosis, predicts an outcome, or compares a person with an appropriate reference group. Nor does limited reliability alone show that every interpretation is invalid; evidence must be judged for the use at hand.

Reliability adjustment helps researchers test whether differences in measurement consistency partly account for patterns in a broad review. It leaves unresolved whether a particular instrument is precise enough for a decision, whether its interpretation is supported for that population, and whether a changed personal result reflects lasting ability change. It cannot supply missing norms or validate an online conversion. Those questions need evidence tied to the exact test, score, population, and proposed use.

Sources: The Stability of Cognitive Abilities: A Meta-Analytic Review of Longitudinal Studies (author postprint); Test Reliability—Basic Concepts; Standards for Educational and Psychological Testing

What does a narrower repeat-test study add?

The UK Biobank repeat-assessment study adds a useful qualification: continuity depends on the task being repeated. It followed a selected group of 20,346 adults who completed cognitive measures again after an average of 4.33 years. That is a different question from the 2024 meta-analysis, which pooled 1,288 correlations across 205 studies and modeled average rank-order stability across many abilities, instruments, ages, and intervals. The meta-analysis estimates a broad pattern across research; UK Biobank shows how much repeat consistency can vary among a few specific tasks. These findings complement each other rather than conflict. The UK Biobank paper examined brief computerized tasks, not a comprehensive intelligence battery. Its verbal-numerical reasoning task contained 13 reasoning questions under a two-minute limit. The reaction-time measure used a timed symbol-matching task. The visual-memory measure asked participants to remember positions of card pairs and counted errors. Because the content and response demands differ, each score captures performance under a particular format. A repeat association from one task cannot simply be transferred to a test with different content, timing, scoring, or administration. For the repeated scores, the researchers reported intraclass correlations (ICCs). An ICC is a statistic describing how consistently repeated measurements distinguish people under a specified model. In this study, the ICC was 0.65 for verbal-numerical reasoning, 0.57 for reaction time, and 0.16 for visual-memory errors. The large gap is the point: the same cohort and broad retest period did not produce one stability value for every task. The low visual-memory estimate signals weak relative consistency for that brief measure in these circumstances. It is not a universal ranking of memory against reasoning, nor a coefficient for other tests. The study also examined average score changes, a separate question from whether people retained similar relative positions. Reasoning and visual-memory scores improved slightly on average, while reaction-time performance worsened. The reported mean changes on these measures did not exceed the study’s estimated smallest real difference. The authors suggested practice effects as one possible explanation for small improvements, but the pattern alone cannot establish why each score moved. Familiarity, measurement variation, and other changes in testing conditions are possible considerations; the study does not identify a single cause for an individual’s result. The repeat sample places important limits on the comparison. Participants lived within 35 kilometres of the UK Biobank coordinating centre in Stockport, and 21% of those invited responded. The follow-up interval ranged from two to seven years, with the reported 4.33 years as its mean. The measures were brief, bespoke tasks, and the number of participants with complete repeat data differed by task. In particular, verbal-numerical reasoning had 4,523 repeat observations, compared with 19,327 for reaction time and 19,017 for visual memory. The overall count of 20,346 therefore does not mean every task estimate came from all participants. This narrower study cannot replace the broader synthesis, and the synthesis cannot erase the task-level differences. The 2024 review answers how longitudinal associations look across a wide body of research and reports substantial variation among studies. UK Biobank answers how a handful of its own brief measures performed when repeated in a geographically selected cohort. Differences in task design, participants, and analytic scope mean that the estimates should be read side by side, not treated as competing measurements of one universal adult score. For a reader comparing personal results, the practical lesson is to inspect the task before interpreting a stability claim. Ask whether both scores came from the same instrument and scale, whether the form and conditions were comparable, and whether the instrument has evidence for the interpretation being made. A broad average across studies cannot tell you whether a change on a different test reflects a lasting change in ability. The UK Biobank example makes that boundary concrete: “cognitive test” covers measures whose repeat consistency can differ substantially, even within one research cohort.

Sources: Test-Retest Reliability of Cognitive Tests Used in the UK Biobank; The Stability of Cognitive Abilities: A Meta-Analytic Review of Longitudinal Studies (author postprint)

Why can retesting raise observed performance?

A score can rise on a second sitting because the person learned something about the task, received coaching, saw the same items again, or took the test under different conditions. The ability being sampled may also have changed. The score difference alone does not identify which explanation applies, and it does not establish that general intelligence increased. Retest performance is an observation to interpret, not a mechanism by itself. The 2007 meta-analysis “Retesting in Selection: A Meta-Analysis of Coaching and Practice Effects for Tests of Cognitive Ability” examined 50 studies, comprising 107 samples and 134,436 participants. Across them, the adjusted overall effect size was .26. Effects were larger when practice was accompanied by coaching and when people took identical forms. This is evidence that repeat exposure can improve cognitive-test performance in organizational and educational testing settings. It does not forecast one adult's result or establish equal effects across tests. The review said further research was needed on how retesting affects the validity inferences drawn from scores. The moderator findings point to plausible routes for an increase. Coaching may teach an approach to a task or familiarity with the test format. Repeating identical items can make the second attempt partly a memory task. Even with different items, a person may understand the response format better or use time more effectively. These routes can improve performance on the retest without showing equal gains on unfamiliar tasks. The meta-analysis documents larger effects under coaching and identical forms; this explanation of why exposure could matter is interpretation, not a separate causal result for every included test. Other explanations remain possible in an individual comparison. The sampled ability may genuinely have changed, A first score may also have been unusually low because of measurement error, so a later score moves closer to typical performance. Sleep, attention, stress, interruptions, timing, instructions, or the testing environment may differ between sittings and affect the result. Without evidence about the exact instrument and administration, a reader cannot apportion the change among these explanations. A score change after practice does not prove practice caused all of it; a change without practice does not prove lasting ability change. Practice-effect research answers a different question from the 2024 lifespan synthesis of rank-order stability. It asks whether scores tend to shift across occasions and which retest conditions accompany larger gains. Rank-order stability asks whether people retain similar relative positions across occasions. A group can retain much of its ordering while individual scores move, and practice-related gains can coexist with that pattern. Neither group-level result identifies the cause or meaning of one person's change. The “Standards for Educational and Psychological Testing,” published by the American Educational Research Association, American Psychological Association, and National Council on Measurement in Education, frame interpretation around intended use and supporting evidence. Applied to a retest, the same numerical scale is not enough: a reader needs to know what was measured, how scores are interpreted, and whether evidence supports the proposed comparison. Repeat consistency alone cannot establish that a particular interpretation is valid or useful. For a personal comparison, note the exact test and score type, time between attempts, whether items or forms repeated, any coaching or practice, and changes in testing conditions. Treat the difference as observed performance unless instrument-specific evidence supports a stronger interpretation. This record does not calculate a corrected score; it identifies what would need checking. A brief online reasoning result should remain a low-stakes practice observation, not a personal forecast. Ask: what changed between attempts, and what evidence lets this test support the conclusion I want to draw?

Sources: Retesting in Selection: A Meta-Analysis of Coaching and Practice Effects for Tests of Cognitive Ability; Standards for Educational and Psychological Testing

What can a pooled stability estimate say about your own result?

A pooled stability estimate says that adults in the studies tended, on average, to keep similar relative standing over time. It cannot tell whether your own score changed by more than measurement error, validate the test you took, or decide whether the score suits a consequential decision. The Standards for Educational and Psychological Testing tie interpretation to evidence for the proposed use; a broad finding about stability is not a substitute. First identify what each number represents. A raw score records performance on a form; a norm-referenced score interprets performance against a defined comparison group. Then check whether results came from the same instrument, scale, form, interval, and administration conditions. Ask what evidence describes the test’s precision for repeated scores and supports your intended interpretation. “Test Reliability—Basic Concepts” defines reliability as consistency across testing occasions, editions, or raters. Consistency matters, but does not by itself establish that an interpretation is valid. If an important education or clinical decision depends on the result, consult a qualified professional and consider other relevant evidence. For low-stakes curiosity, Test IQ Free offers a 50-question reasoning test and worked answers. Its provisional estimated IQ range uses an unvalidated conversion with no validated adult reference population; it is neither a confidence interval nor a norm comparison. The optional $9 report adds explanation, not validity. You can [try the free test](/test) for practice, keeping its estimate separate from professional assessment. Population-level continuity supplies context; the exact measure and intended use determine what you can responsibly conclude about your result.

Sources: Standards for Educational and Psychological Testing; Test Reliability—Basic Concepts