How IQ changes with age: two answers, both true

0 The Short Answer

Your IQ score, meaning your standing relative to people your own age, changes remarkably little across a lifetime. Your raw cognitive performance, meaning what you can actually do, changes a great deal. Those are two different measurements, and the question "does IQ change with age" gets a different answer depending on which one you meant.

The evidence for the first claim is unusually good. In 1932 the Scottish Council for Research in Education tested almost every eleven year old in Scotland, 87,498 children, on a single paper and pencil instrument called the Moray House Test Number 12. Decades later, Ian Deary's group in Edinburgh tracked survivors down and gave them the same test again. Reporting in Psychology and Aging in 2011, Gow and colleagues found stability coefficients of .67 from age 11 to age 70, .66 from 11 to 79, and .51 from 11 to 87. Nothing else in psychology has been measured over a span like that.

Point the same instruments at absolute performance and the picture reverses. Speeded and novel reasoning tasks peak early and decline for decades afterwards, while vocabulary and accumulated knowledge keep climbing into later life. A single composite number hides that split completely, which is why what a score actually reports matters more than the number itself.

The bridge between the two facts is age norming. An IQ score is computed against a reference group of people the same age as you. A 100 at 25 and a 100 at 70 both mean average for that age, and they sit on top of very different raw performance. Most of the confusion on this topic dissolves once that is clear.

1 Two Questions Hiding Inside One

"Does IQ change with age" is not one question, and the pages that rank for it almost never say so. One version asks whether your position in the distribution moves. The other asks whether your absolute capability moves. They have opposite answers, and a page that gives only one of them is not wrong so much as half finished.

Rank order stability is the technical name for the first. It asks: if you lined up a thousand people by test score at age 11 and lined the same thousand up again at age 70, how similar would the two lines be? The answer, from the Lothian cohorts, is very similar but not identical. A correlation of .67 means the ordering is largely preserved and also means plenty of individuals swapped places.

Absolute change is the second. It asks: on the same items, under the same conditions, does a person get more or fewer right at 70 than at 25? For anything that depends on speed or on reasoning with unfamiliar material, the answer is fewer. For anything that depends on accumulated knowledge, the answer is often more. Nothing about the first answer constrains the second, because a correlation is computed on positions, not on levels. Every person in a cohort can decline by ten raw points and the correlation will not budge.

This is also the boundary between this page and its companion. Why the average stays 100 at every age is a population question about how the scale is constructed and standardized. It is answered on a separate page and is not repeated here. What follows is the within person question: what happens to one individual, measured twice, decades apart. Keeping those separate is the single most useful move available to a reader of this literature, and it depends on knowing what an intelligence test measures in the first place.

2 The Longest Follow Up in Psychology

The Scottish Mental Surveys were an accident of policy that turned into the best dataset on this question anyone will ever have. In 1932 the Scottish Council for Research in Education administered the Moray House Test Number 12 to 87,498 children born in 1921, essentially every eleven year old in the country. The exercise was repeated in 1947 with 70,805 children born in 1936. Both sets of scores sat in an archive for half a century, which is why this evidence exists at all: nobody designed a seventy year follow up, they simply found one waiting.

The test itself matters, because a stability coefficient is only as meaningful as the instrument producing it. The Moray House Test Number 12 is a group administered paper and pencil test with a 45 minute limit and 71 numbered items, 75 in total, scored out of 76. The item mix is broad: 14 following directions items, 11 same opposites, 10 word classification, 8 analogies, 6 practical items, 5 reasoning, 4 proverbs, 4 arithmetic, 4 spatial, 3 mixed sentences, 2 cypher decoding and 4 others. That is a general ability test in the ordinary sense, closer to the item formats used in group tests than to a single puzzle task, and it loads heavily on the general factor those items share.

Recruitment began in 1999. The Lothian Birth Cohort 1921 tested 550 survivors, 234 men and 316 women, in Edinburgh between 1999 and 2001, then again in 2003 to 2005, then a third time in 2007 to 2008. The Lothian Birth Cohort 1936 recruited 1,091 people between 2004 and 2007, tested at a mean age of 69.5 years. Every one of them sat the identical test they had taken as a child.

Two notes on reading that table. The sample column gives how many people completed the test at the later occasion, and the correlation itself is computed on the smaller number who have a usable score at both ages, so the pairwise figure is never larger and is usually a little smaller. The corrected column is a modelled adjustment for the fact that survivors span a narrower range of ability than the 1932 population did, not a second measurement.

The last row comes from a separate paper. Deary, Pattie and Starr, writing in Psychological Science in 2013, extended the follow up to age 90 in 106 people and reported a raw coefficient of .54, rising to .67 once range restriction was accounted for. They also checked that the old instrument was still doing its job, reporting strong concurrent and predictive validity for the Moray House Test against standard cognitive tests at both ages. Gow and colleagues described their own 76 year span, from 11 to 87, as believed to be the longest period over which a stability coefficient for intelligence had been computed, and the 2013 paper stretched it by three more years.

3 Who Is Still Alive to Be Measured

Every figure in the table above was computed on survivors, and survivors are not a random sample of the people who took the test in 1932. This is the limitation that most write ups of the Lothian findings skip, and the original papers are explicit about it.

The 2011 paper reported the comparison directly. Among Lothian Birth Cohort 1921 members, those who returned for the third wave of testing had scored higher as children than those who did not: 47.7 against 45.5 on the Moray House Test at age 11, a difference significant at p equal to .014. At age 79 the gap was much wider, 61.9 against 57.2, significant at p below .001. People who score higher live longer and are more likely to keep turning up for a clinic appointment in their late eighties, so the surviving sample gets progressively more selected with every wave.

Selection of this kind narrows the spread of scores, and a narrowed spread mechanically shrinks a correlation. That is why the authors report range restriction corrected values alongside the raw ones, and why the corrected figure for the age 87 comparison rises from .51 to .61. Their own note is worth keeping: the participants tested at 87 were even less representative of the original 1932 population than those tested at 79. Correction is a modelled adjustment, not a measurement, so the honest reading is that true lifelong stability sits somewhere at or above the raw coefficients rather than exactly on them.

Selection does not make the finding fragile. It cuts in a specific direction, and the direction is towards understatement. But it does mean a reader should treat .51 as a floor for the surviving population rather than a precise constant of human nature, in the same way that any reliability figure has to be read alongside the sample that produced it. The general principle is covered under how measurement quality is judged, and the sampling half of it under sampling and fairness problems in testing.

4 What a Stability Coefficient Does Not Promise

A correlation of .67 is a statement about a group, and it makes no promise at all about any particular person in it. This is the point at which the Lothian findings get oversold, usually by writers who are impressed by the number and have not asked what it would look like from the inside.

Start with shared variance. Gow and colleagues put the minimum estimates at 26 to 44 percent of the variance in later life scores accounted for by childhood performance on the same test. They give a range rather than a single figure because there is a genuine methodological argument, going back to Ozer in 1985, about whether squaring a correlation is the right way to express shared variance. Take the more generous end and well over half the variation in old age scores is still unexplained by the childhood score.

Now translate that into people. With a coefficient near .5, a person in the top fifth as a child is much more likely than chance to be in the top fifth in old age, and is still perfectly capable of finishing in the middle. Movement across the distribution is not an anomaly in these data. It is a large fraction of what the data contain. Individual trajectories cross, and the correlation summarises the average tendency of a thousand of them rather than the fate of any one.

There is a second reason the coefficient understates true stability, and it pushes the opposite way from selection. No test is perfectly reliable, so part of the observed disagreement between two administrations is simply measurement error rather than real change. Gow and colleagues note this explicitly. Between selective survival on one side and unreliability on the other, the raw numbers are conservative, and the sensible summary is that lifelong stability is high and clearly incomplete.

Practically, this is why a single score should be read as a band rather than a point, and why the exact percentile for a score and how rare a given score is are more useful framings than a bare number. A childhood score is evidence about an adult, not a verdict on one.

5 Why 100 at 25 and 100 at 70 Are Different Performances

An IQ score is a rank within an age band, and once you understand the arithmetic that produces it, most of the confusion about ageing and intelligence disappears. The score is not a count of what you got right. It is a statement about where your raw performance sits inside the distribution of raw performances produced by people your age.

The Lothian analyses show the procedure in miniature. To model change within old age, Gow and colleagues took the raw Moray House score, regressed it on the participant's age in days at the moment of testing, kept the standardized residual, and converted that residual to a scale with a mean of 100 and a standard deviation of 15. Age was removed by construction. Whatever remained described how a person performed relative to others tested at the same age, which is exactly what a published IQ score does using a normative sample instead of a regression.

The consequence is worth stating plainly. If a 70 year old and a 25 year old both score 100, they have both landed at the midpoint of their own age group, and the 25 year old will almost certainly have answered more items correctly on any speeded task. If the same person scores 100 at 25 and 100 at 70, their raw performance on speeded and novel material has fallen while their position among peers has held. That is not a contradiction. It is the design working as intended.

It also explains a common and reasonable complaint: that age norming flatters older adults. It does not flatter anyone. It answers a specific question, which is how a person compares with their contemporaries, and it deliberately refuses to answer a different one, which is how a person compares with their younger self. If the second question is what you care about, the score is the wrong instrument and raw performance on matched tasks is the right one. The machinery is described in full under the norming process itself, and the historical alternative that it replaced is covered under the mental age idea it replaced.

6 The Design That Separates Ageing From Cohort

If you compare a group of 70 year olds with a group of 25 year olds today, every difference you find is a mixture of ageing and of everything else that separates people born in 1956 from people born in 2001. The Seattle Longitudinal Study exists because of that confound, and its results are the strongest demonstration available that the mixture matters.

The study began in 1956 as an ordinary cross sectional survey of adults aged 20 to 70. In 1963 Warner Schaie converted it into a multiple cohort longitudinal design, adding fresh random samples from the membership of a large health maintenance organization at seven year intervals and following participants until death or dropout. That structure lets the same six abilities be examined two ways: as differences between age groups measured once, and as changes within the same individuals measured repeatedly.

The two pictures do not agree, and the disagreement is systematic. Writing in Research in Human Development in 2005, Schaie reported that in the cross sectional data four of the six abilities show consistent negative age differences, reaching significance at age 46 for inductive reasoning, spatial orientation and perceptual speed, and at age 39 for verbal memory. Across the full age span the cross sectional gap averages roughly two standard deviations. In the longitudinal data, reliable within person decline appears decades later.

Numeric facility is the case that gives the game away. Cross sectionally it looks almost flat, with only minor differences between age groups. Longitudinally it shows early and steep decline. Schaie's explanation is cohort: over the 70 year range his data cover, inductive reasoning and perceptual speed rose by about a standard deviation across successive birth cohorts, while numeric facility drifted slightly downwards. A negative cohort trend makes an ability look stable in a snapshot even when individuals are losing ground fast. The broader version of that cohort story is the secular rise in test scores, and the site keeps a separate page on generational score comparisons.

7 Both Designs Are Biased, in Opposite Directions

The tidy conclusion would be that longitudinal data are correct and cross sectional data are misleading. The literature does not support the tidy conclusion, and the disagreement between two of the field's most serious researchers is the reason.

Timothy Salthouse has argued the opposite case for years. In Neurobiology of Aging in 2009 he concluded that some aspects of age related cognitive decline begin in healthy educated adults in their 20s and 30s, with cross sectional declines of about .02 to .03 standard deviation units per year for adults under 60, steepening to about .04 to .05 per year in a sample of roughly 800 adults aged 61 to 96. His argument is that longitudinal designs mask decline because people improve on a test simply by having taken it before, and that the improvement is large enough to swamp several years of genuine change. He is careful to exempt the crystallized side: measures built on accumulated knowledge, he reports, keep rising until at least age 60 in the same samples.

He put numbers on that in Current Directions in Psychological Science in 2014, using an episodic memory composite from the Swedish Betula project. Among adults born in 1949, tested at 40 in 1989 and again at 45 in 1994, the cross sectional comparisons showed the older group lower by 0.99 and 2.40 T score units on the two occasions, while the longitudinal comparison on the same people showed a gain of 2.08. Two separate methods for estimating change without prior test experience both put the real change at minus 1.26, a gap of 3.34 T score units between what the repeated testing showed and what the experience free estimate showed. Salthouse called the illustration crude, since it rests on two occasions and on group means, but the direction has been replicated in his Virginia Cognitive Aging Project, where more than 2,300 people had completed at least two occasions by December 2013.

Longitudinal designs also lose their weaker performers to attrition, exactly as the Lothian waves did. Putting both together, Salthouse concluded that within cohort comparisons resemble the cross sectional trends more closely than the longitudinal ones. Schaie's reply, published in Neurobiology of Aging in 2009, is that this reifies what he calls the cross sectional fallacy: inferring change within individuals from differences between groups measured once. Age changes and age differences can only coincide, he argues, in a perfectly stable environment with no cohort variation, and no such environment has existed in the last century.

The useful reading is that the truth is bracketed rather than pinned. Cross sectional estimates are inflated by cohort differences. Longitudinal estimates are deflated by practice effects and by the loss of weaker participants. Real decline in speeded and novel reasoning almost certainly begins earlier than the longitudinal curves suggest and later than the cross sectional ones do. Anyone selling a single confident age is selling one design's bias. This is also why preparation and practice before a session is treated carefully on this site rather than dismissed.

8 Fluid Reasoning Falls While Crystallized Knowledge Holds

The abilities inside a composite do not age together, and the gap between the earliest and the latest peak is measured in decades. In CHC terms, Gf, fluid reasoning, peaks early and declines from there. Gc, comprehension knowledge, holds or rises well into later life. Gs, processing speed, peaks earliest of all. A single number averages across all of them and shows you none of them.

The largest convergent test of this comes from Hartshorne and Germine, publishing in Psychological Science in 2015. They combined 48,537 online participants with the normative samples of the WAIS-III (2,450 people aged 16 to 89), the WMS-III (1,225 people) and 26,850 vocabulary responses from the General Social Survey collected between 1974 and 2012. Their finding was not that ability peaks at one age. It was that different abilities peak at wildly different ages, and that no age exists at which a person is at peak on everything.

That last pair is a cohort effect rather than a lifespan effect, and the authors tested it directly. Pooling the General Social Survey series with the Wechsler and web vocabulary data, they estimated that the age of peak vocabulary performance has been arriving about 0.96 years later with each passing year of testing, significant at p equal to .0003. The peak is not just late. It is getting later, for the same reason scores in general have drifted upward across generations.

Salthouse's own data agree on the crystallized side even while he disputes the timing on the fluid side. Measures built on accumulated knowledge, he reports, are consistently found to increase until at least age 60. That is the one point on which the cross sectional and longitudinal camps do not argue.

The consequence for reading your own result is direct. A 60 year old with strong verbal comprehension and slow processing speed is not showing a defect in one domain. That profile is what the lifespan literature predicts, and it is only visible if the abilities were measured separately. The definitions are set out under the Gf and Gc distinction in full, the taxonomy under the CHC structure behind the domains, and the speed domain on its own under processing speed as a measured domain.

9 Why One Number Hides the Pattern

If fluid and crystallized abilities move in opposite directions with age, then a composite is the worst possible summary of an ageing profile, because the two trends partially cancel. A person whose reasoning has slipped and whose vocabulary has grown can produce the same composite at 65 that they produced at 35, having changed substantially in both directions underneath it.

This is the honest case for measuring domains separately rather than reporting one figure. A battery that yields index scores across verbal comprehension, fluid reasoning, quantitative reasoning, visual spatial ability, working memory and processing speed shows the shape of the change. A composite alone shows only its net. Neither is more accurate than the other. They answer different questions, and only the profile answers the one an ageing reader is usually asking.

ACIS is built that way: 20 subtests across six primary cognitive domains, with subtest scaled scores on a mean of 10 and a standard deviation of 3, and composites on a mean of 100 and a standard deviation of 15. The Full Scale composite reaches an omega of .9886 with a g loading of .958 and a standard error of measurement of about 1.60 IQ points, figures drawn from a technical analysis set of 2,750 complete records. The higher order model fit was CFI .9761, TLI .9726, RMSEA .0406 and SRMR .0217, with a chi-square of 916.703 on 166 degrees of freedom.

What a profile buys an older adult specifically is interpretive: it separates a low index that reflects the ordinary trajectory of speed from a low index that does not fit the pattern at all. That distinction is invisible in a composite. The domain list is set out under the six primary domains, the argument for breadth under a broad multi domain battery, and the underlying numbers under the published reliability tables.

10 Is IQ Fixed? What the Longitudinal Data Say

"Fixed" is the wrong word in both directions, and the Lothian data show why with unusual clarity. Childhood ability turned out to be a powerful predictor of the level a person reached in old age, and no predictor at all of how fast they declined once they got there.

Gow and colleagues modelled Moray House performance between ages 79 and 87 with a growth curve that separated level from rate of change. Age 11 score predicted the level strongly, with a standardized coefficient of .57 at p below .001. It predicted the slope not at all, at minus .03 with a p value of .666. Years of education behaved the same way: it predicted level at .12, p equal to .001, and had no relationship with rate of change. The authors' summary is the part worth carrying away. Higher early life intelligence was protective of intelligence in old age because of stability across the lifespan, not because it slowed the decline experienced within old age.

That result cuts against two popular positions at once. It undermines the idea that a childhood score is destiny, because a large share of variance in late life ability is not accounted for by it. It also undermines the cognitive reserve story in its strongest form, at least for this measure and this age window, because the people who started higher fell at the same rate as everyone else. They simply fell from further up.

None of this is the same question as whether ability can be raised deliberately. That is a separate literature with its own evidence, and this site handles it separately under whether training moves the number. Nor is it a claim about genetics: stability across a lifespan is compatible with a range of causal accounts, which is the subject of what heritability does and does not mean. The mirror question, whether an exposure pushes ability down, has been attacked with the same logic of measuring first: the Dunedin birth cohort tested its members at 13, before anyone had used cannabis, and reported an eight point fall by 38 among persistent adolescent onset users, a figure resting on 23 people that largely disappeared once twins discordant for use were compared, a sequence set out under the replication record on cannabis and measured ability.

The defensible summary is that adult rank order is stable enough to be one of the most reliable individual differences psychology measures, and unstable enough that a substantial minority of people end up somewhere they would not have been predicted to end up. Both halves matter, and the prediction is to a distribution rather than to a point, which is the same caution that applies to school outcomes and measured ability.

11 Comparing Two Scores of Your Own

The practical version of this question is usually personal: someone has a score from school and a score from last week, and wants to know whether the difference means anything. Usually it does not, and there are four reasons that stack.

The first is measurement error. Every score is a point estimate with a band around it. On ACIS the standard error of measurement for the Full Scale composite is about 1.60 IQ points, and comparing two administrations means carrying the error from both, so the interval around a difference is wider than the interval around either score. Two numbers that differ by a few points are consistent with no real change at all.

The second is practice. In Salthouse's Betula illustration the gap between observed longitudinal change and change estimated without prior test experience was 3.34 T score units, which on a scale with a standard deviation of 10 is about a third of a standard deviation. That is not measured in IQ points and should not be converted into them, but it is large enough to reverse the sign of a five year change. If two administrations used the same or similar items, the later one is biased upward for reasons that have nothing to do with ability, and the shorter the gap between them, the worse the bias.

The third is that the tests were probably different instruments with different norms. A school screening from 1994 and an adult battery from today do not share a reference sample, an item pool, or a scaling decision. Comparing them is not comparing like with like, and the mismatch is often larger than any true change. The size of that gap is easy to underestimate, since the Stanford-Binet 5 still in clinical use is anchored to a reference sample collected in 2003 while the Woodcock-Johnson V was normed between 2022 and 2024, and the instrument by instrument inventory prints that norm year next to the age range for every battery, which is the quickest way to see whether two of your own scores were ever on the same footing. The site covers the practical version of this under free quizzes against validated instruments and how the main online batteries compare.

The fourth is age norming, discussed above: the reference group moved between the two occasions, so identical scores can conceal real raw change and different scores can conceal real raw stability.

What survives all four is a profile comparison rather than a composite comparison. If the same battery is administered twice with a meaningful gap, a shift in the shape of the profile, verbal holding while speed falls, is more interpretable than a shift in the single number. Even then, an unsupervised administration adds testing conditions to the list of things that differ between occasions, which is why the standard error of measurement is the first thing to read before either score is taken at face value.

12 What This Evidence Cannot Tell You

Nothing on this page is a way of finding out whether your own decline is normal. The literature describes averages and distributions in cohorts. It does not diagnose individuals, and an online score is not the instrument for that question even in principle.

The Lothian and Seattle findings describe healthy community dwelling adults. Pathological decline is a different phenomenon with a different trajectory, and it is identified by clinical assessment that includes history, informant report, medical investigation and supervised neuropsychological testing, not by comparing two numbers. A person worried about memory or reasoning needs a clinician, and the cost of an unnecessary reassurance from a self administered test is that they do not get one.

There are also limits inside the research itself that a careful reader should hold onto. The Lothian cohorts are Scottish, born in 1921 and 1936, educated in a specific system, and tested on an instrument designed in the 1930s. The Seattle sample was drawn from the membership of one health maintenance organization in one American city. Salthouse's participants were recruited by advertisement and scored roughly half a standard deviation to a full standard deviation above the Wechsler normative means. None of these is a representative sample of humanity, and the fact that they agree with one another on the broad pattern is more persuasive than any one of them alone.

Finally, the group statistics here are not individual predictions. A stability coefficient of .67 tells you about the average behaviour of a thousand trajectories. It tells you nothing about which of them is yours. That distinction matters as much here as it does in measured ability and later health outcomes, and it is the main reason a supervised assessment is a different product from an online one, as set out under supervised testing against unsupervised testing.

13 How to Read Your Own Score Across a Lifetime

The reader who came here asking whether their intelligence has changed is entitled to a usable answer, and it has three parts.

Your standing among people your own age has probably moved less than you think. That is the plainest implication of correlations of .67 and .66 measured across six and seven decades, and it holds even after allowing for the fact that survivors are a selected group. If you were a strong test taker at school, the base rate says you are still above average now, and the base rate also says a meaningful minority of people in your position are not. What you think is a poor instrument for settling it either way, since a self estimate pooled over 115 samples and nearly 37,000 people correlates about .30 with a measured score and predicts a range some 55 points wide, which is why the impression of having slipped carries far less information than a childhood coefficient does, as how poorly people place themselves on the scale makes clear.

Your raw performance has certainly moved, and it has moved in different directions in different domains. Speed and novel reasoning are lower than they were at 25. Vocabulary and accumulated knowledge are probably higher. Neither of those is visible in a composite, which is why a domain profile is the more informative output for anyone past their thirties and why the shape of the profile is worth more attention than the headline figure.

And the number itself is a comparison, not a quantity. It was computed against people your own age, which is what makes it interpretable and also what makes it useless for the specific question of how you compare with your younger self. If that is the question, matched raw performance is the measurement you want, and no standardized score will substitute for it. For interpretation of the number you do have, what counts as a good score and score interpretation and routing are the practical starting points, and the current Wechsler adult scale is the supervised reference the online instruments are measured against.

All of this sits inside a professional framework rather than outside one. The Standards for Educational and Psychological Testing, published in 2014 by the American Educational Research Association, the American Psychological Association and the National Council on Measurement in Education, require that a test be interpreted only for the population and purpose on which its evidence was gathered, that normative samples be described so users can judge their relevance, and that reliability evidence be reported alongside any score. APA testing standards make the same demand of anyone reporting a result. Applied to ageing, they say something specific and useful: an age normed score answers an age normed question, and a claim about what a person has lost or gained since their youth needs evidence the score was never designed to supply.

14 Frequently Asked Questions