Can You Improve Your IQ?
Yes, and the Question Is Which Ability

0 Quick Answer

Updated September 19, 2026 by Structural. Yes, measured ability can improve, and the strongest evidence for that is schooling rather than any brain game. The reason this question keeps getting a pessimistic answer is that most people ask it in a form that cannot be answered, and most studies test it with a tool that was never suited to the job.

Direct answer: the useful question is not "can IQ be raised" but "which ability, trained how, at what age, for how long". An IQ score is not a single substance. It is a summary of many narrow abilities, and those abilities do not all respond to the same input. Training reliably improves the tasks it resembles. How far those gains travel to untrained tasks, and whether they reach the latent ability a test estimates, is the open question, and the honest answer differs by intervention: years of schooling move test scores across the board, while short programs such as dual n-back move the trained task and little else.

The single strongest piece of evidence on this question is not a brain game trial. It is the meta-analysis of education effects by Ritchie and Tucker-Drob (2018), which pooled 142 effect sizes across more than 600,000 participants and found that each additional year of schooling raises measured intelligence by roughly 1 to 5 points, with effects that persist across the lifespan. Schooling is cognitive training. It is simply the version that runs for years instead of four weeks.

1 The Question Is Malformed, and That Is Why the Answers Disappoint

Ask whether IQ can be improved and you will get two confident camps. One says no, intelligence is fixed and largely heritable. The other says yes, and sells you an app. Both are answering a question that has no clean answer, because the question treats IQ as one thing that either moves or does not.

Consider the equivalent question in physical fitness. "Can you improve fitness?" is unanswerable in that form. Improve what? A powerlifter and a marathon runner are both extremely fit and would perform poorly at each other's sport. Training that builds maximal strength does very little for aerobic capacity, and the reverse is equally true. Nobody concludes from this that fitness is fixed. They conclude that you have to name the capacity before you can train it.

Cognitive ability works the same way, and the structure is better documented than most people realise. A modern intelligence battery does not measure one ability. It measures a hierarchy of them, and the top level summary that people call IQ is a weighted composite of the levels beneath it. When someone asks whether IQ can be raised, the honest response is to ask which part of that hierarchy they mean, because the parts behave differently.

This is not a dodge. It is the difference between a question that has an answer and one that does not. Once you name the ability, the literature becomes far less pessimistic than its reputation suggests.

2 What an IQ Score Actually Contains

The most widely accepted map of human cognitive ability is the Cattell-Horn-Carroll framework, usually shortened to CHC. It is not a theory one lab invented. It is the result of decades of factor analysis across hundreds of datasets, consolidated by McGrew (2009) into the structure that essentially every current test publisher now uses to organise its batteries.

CHC describes three levels. At the top sits general ability, usually written as g. Beneath it sit broad abilities, around eight to sixteen depending on the version. Beneath those sit roughly eighty narrow abilities, each one a specific skill that can be isolated and measured. You can read the full structure in The CHC Model, and how ACIS maps onto it in Cognitive Domains.

The broad abilities that appear in most batteries include the following.

This matters enormously for the training question. These are not synonyms. Vocabulary depth and mental rotation speed are different capacities with different developmental trajectories, different neural substrates, and different responsiveness to practice. Lumping them into one number is useful for prediction, but it is actively misleading when the question is whether training works, because it invites you to test a training program against a composite that most of the program never touched.

Here is the concrete consequence. Suppose a four week program genuinely improves working memory span by a meaningful amount. Working memory contributes perhaps one fifth of the variance in a full scale composite. Even a large, real improvement in one broad ability moves the composite by a fraction of its actual size, and that fraction is easily buried by measurement error. The training could work and the study would still report nothing.

3 The Intervention That Demonstrably Works

If cognitive ability were fixed, the single most studied intervention in the history of the field would show nothing. It shows a great deal.

Ritchie and Tucker-Drob (2018) conducted the definitive meta-analysis on education and intelligence. They pooled three independent research designs: studies exploiting policy changes that increased compulsory schooling, studies comparing students who differed in schooling for reasons unrelated to ability, and studies following the same individuals across school years. All three designs converged. Each additional year of education produced gains of approximately 1 to 5 IQ points, and those gains did not vanish. They were still detectable decades later.

This is a causal finding, not a correlation. The compulsory schooling designs are natural experiments: a law changes, an arbitrary cohort receives an extra year, and their measured ability rises relative to the cohort born a year earlier. There is no plausible route by which a birth date causes higher intelligence except through the extra schooling.

The work of Ceci (1991) had already assembled the converging evidence from a different direction: scores drop measurably over long school holidays, children who start school later score lower, and intermittent schooling produces cumulative deficits. Schooling is doing something to measured ability, continuously, for years.

Now consider what schooling actually is as an intervention. It is thousands of hours, distributed over years, in which a person is repeatedly required to acquire declarative knowledge, manipulate symbolic systems, reason about unfamiliar material, and produce language under evaluation. It is not one exercise. It is a broad, sustained, effortful curriculum that happens to align with a substantial share of what intelligence batteries sample.

That is the template for what works. Any honest discussion of cognitive training should start from the fact that the largest and best identified effect in the literature comes from an intervention lasting years, not weeks.

4 Why Dual N-Back Failed, and What That Failure Does Not Prove

In 2008, Jaeggi and colleagues published a study in PNAS reporting that training on a dual n-back task improved performance on a fluid reasoning measure, with gains scaling by training dose. The result was striking because it claimed transfer from a working memory task to reasoning, which is exactly the kind of far transfer the field had struggled to demonstrate.

The replication record that followed was poor. Redick and colleagues (2013) ran a version with an active control group and no-contact control group and found improvement on the trained task with no transfer to reasoning. Melby-Lervag and Hulme (2013) meta-analysed 23 studies and concluded that working memory training produced short-lived, task-specific effects without generalisation. Au and colleagues (2015) found a small positive effect on fluid reasoning, and the ensuing exchange over whether it survived correction for baseline differences and publication bias has never fully resolved.

So dual n-back does not reliably raise reasoning scores. That is a fair reading of the evidence. What is not fair is the conclusion that usually gets attached to it, namely that cognitive ability is therefore untrainable.

Look at what was actually tested. A dual n-back task requires holding two streams of stimuli in mind and reporting matches at a lag. It trains a narrow ability inside working memory: the maintenance and updating of items under interference. The outcome measure was typically a matrix reasoning test, which requires inducing an abstract rule from a visual pattern. Matrices are only one of the four families used to measure that ability, and what a fluid reasoning test is built to isolate is rule induction on material the person has never studied, which is the same property that makes it resistant to drilling. Induction and memory updating are different narrow abilities living under different broad abilities in the CHC hierarchy.

The study design asks whether practising one narrow ability improves a different narrow ability under a different broad factor. The answer turned out to be mostly no. That is a finding about transfer distance. It is not a finding about trainability, and treating it as one is a category error.

The typical trial compounds the problem. Most ran between two and five weeks, with sessions of twenty to thirty minutes, often three to five times a week. Total training time frequently landed under fifteen hours. Compare that to the intervention that demonstrably works, which is a school year: roughly a thousand hours of varied, effortful, instructor-guided cognitive work. Concluding that ability is fixed because fifteen hours of one narrow exercise did not move a composite is like concluding that exercise does not build muscle because a fortnight of wrist curls did not change someone's deadlift.

5 Transfer Distance Is the Variable That Decides Everything

The organising concept in this literature is transfer, meaning the extent to which practice on one task improves performance on another. It is conventionally split into near transfer, where the trained and untrained tasks share structure, and far transfer, where they do not.

Near transfer is reliably observed. Practise a span task and you get better at other span tasks. Practise mental rotation and other spatial tasks improve. The evidence for this is not seriously disputed.

Far transfer is where the field struggles. Sala and Gobet (2017) reviewed the domains where far transfer has been claimed, including chess instruction, music training, and working memory programs, and found that the apparent effects shrink toward zero as design quality improves, particularly once active control groups are used.

This is the finding that gets reported as "brain training does not work". A more precise statement is that distant transfer is weak and that most commercial programs promise exactly that, because promising near transfer is unimpressive. Nobody buys an app that makes you better at the app.

But notice the implication running the other way. If near transfer is real and far transfer is weak, then the correct strategy is not to abandon training. It is to shorten the distance. If you want to improve a narrow ability, train that narrow ability rather than hoping a distant one leaks across.

The left column is a set of plausible matches, not a set of proven treatments. Each entry shares cognitive demands with the narrow ability named beside it, and shared demands are the condition under which transfer has been observed at all: where they are absent, gains have not appeared. The strength of the evidence differs by row. Spatial training has a meta-analysis of 217 studies behind it, Uttal et al. (2013), with an average effect of about half a standard deviation and transfer to untrained spatial tasks. Working memory training has one pointing the other way, Melby-Lervåg, Redick and Hulme (2016): gains on the trained and closely related tasks, no reliable effect on measures of intelligence. The other rows rest on the general pattern rather than on trials of that row.

6 The Matching Principle

Stated carefully: training is most likely to move the abilities whose structure it shares, and the effects that have been measured grow with how much structure is shared and how long the training runs. That is a working principle, a summary of the transfer literature, not a law with a measured proportionality constant. Nobody has quantified structural overlap well enough to predict the size of a gain from it, and the honest expectation for any new program is the one the meta-analyses give: reliable near transfer, weak or absent far transfer.

The principle organises the apparently contradictory literature without dismissing any of it. Schooling works because it is long, effortful, and overlaps with a large share of what batteries sample, and even there the measured gain is of the order of one to five points per year. Dual n-back mostly does not transfer to reasoning: the trained task improves, measures of intelligence do not. Mental rotation training improves spatial tasks, including ones that were not trained, with a meta-analytic effect of about half a standard deviation. Chess makes you good at chess, because the overlap with anything else is thinner than enthusiasts assume.

The chess instruction evidence provides a concrete test of transfer claims, including why active comparison groups change the interpretation.

Three different things get called improvement, and the principle only makes sense once they are kept apart. The first is a practice effect: a higher score on the same or a very similar task, which every retest produces and which says nothing about ability. Scharfen, Peters and Holling (2018) put the average gain from a first to a second administration at about a third of a standard deviation, roughly five IQ points. The second is transfer: gains on untrained tasks that share structure with the trained one. This is real, modest, and what the table above is about. The third is a change in a latent ability, or in g: gains that appear across many different tasks at once. That is the claim brain training makes and does not deliver, and it is not fully settled even for education. In the one cohort where the question was tested directly, Ritchie, Bates and Deary (2015) found that the association between schooling and later test scores ran through specific skills rather than through g. The pessimistic reading of the field is right about the third kind of gain and wrong about the first two, and that pattern is what a hierarchical model of ability predicts: transfer decays with structural distance. The CHC structure is a map of which training is likely to reach which outcome, not a guarantee that it will.

It also explains why the composite is the wrong outcome measure for most training studies. If you train one narrow ability and measure a composite built from twenty subtests, you have deliberately diluted your own signal. The right measure is the subtest closest to what you trained, followed by progressively more distant ones, so you can see the gradient rather than a single blunt verdict.

7 Age Changes the Size of the Effect

Cognitive abilities do not develop on one schedule, and they are not equally plastic at all points on that schedule. This is one of the best established facts about the structure of ability, and it is routinely ignored in discussions of training. The same schedule governs measurement: fluid abilities peak in early adulthood while vocabulary keeps growing for decades, which is why taking an IQ test as an adult only makes sense against norms drawn from adults of a similar age.

Fluid abilities peak early. Cross-sectional and longitudinal work summarised by Salthouse (2004) shows reasoning and processing speed rising through adolescence, peaking in the twenties, and declining gradually thereafter. Crystallized abilities follow the opposite path, rising through middle age and holding late. You can see how this shapes score interpretation in Average IQ by Age.

The implication for training is direct. An intervention delivered while an ability is still on its ascending trajectory is working with development rather than against it. The same intervention delivered after the plateau is fighting a different battle. This is why the strongest intervention findings in the literature come from childhood and adolescence, and why adult training studies report smaller effects even when they are otherwise well designed.

Intervention research in early childhood makes this concrete. Programs delivering intensive, sustained, structured cognitive input to young children have produced substantial score gains, and the systematic review by Protzko, Aronson and Blair (2013) catalogued which manipulations moved measured intelligence in children and which did not. The pattern favours interventions that are early, sustained, and cognitively broad.

None of this means adults cannot improve. Crystallized abilities remain highly responsive throughout life, because acquiring knowledge does not depend on developmental plasticity in the way that reasoning speed does. An adult who reads widely and studies a quantitative discipline for three years will measurably improve on the subtests that sample those abilities. The realistic difference is which abilities respond most, and by how much. It is also worth separating two things that get merged here, because the Lothian Birth Cohort work distinguishes rank order stability from absolute change: standing relative to your own cohort is remarkably persistent across decades even while raw performance on the same tasks rises and then falls.

8 Absence of Evidence Is Not Evidence of Absence

The strongest form of the pessimistic argument is that no study has demonstrated large, durable gains in general ability from training. That is roughly accurate. It is also much weaker than it sounds, because it describes the state of the evidence rather than the state of the world.

Consider what a study capable of settling the question would require. It would need to train several narrow abilities simultaneously with structurally matched exercises. It would need to run for years rather than weeks, since the intervention we know works runs for years. It would need participants young enough to be inside the developmental window for fluid abilities. It would need an active control group doing something equally demanding. It would need to measure at the subtest level so that transfer gradients are visible rather than diluted into a composite. And it would need to follow participants long enough afterward to distinguish durable change from temporary boost.

That study is expensive, slow, and difficult to fund. It has no commercial sponsor, because no product can be sold on a five year timeline. It is unattractive to academic incentives, because a doctoral candidate cannot wait five years for a first result. And the null result that a shorter, cheaper version produces is publishable and career-safe, which means the literature fills up with the underpowered version of the question.

So the evidence base has a specific shape. It is dense with short, cheap, narrow studies and empty where the long, expensive, broad studies would be. Interpreting that emptiness as a finding about human potential mistakes a funding pattern for a fact of nature.

The correct epistemic position is that the question is open in the region where the evidence is thin, and that the evidence we do have from the one long intervention that has been studied properly, which is education, points toward substantial effects rather than away from them.

9 The Flynn Effect as Proof of Concept

There is one more piece of evidence that the pessimistic reading has to explain away, and it is large. Measured intelligence rose substantially across the twentieth century in essentially every country with data.

The meta-analysis by Trahan and colleagues (2014) pooled 285 studies and put the rate at roughly 2.31 points per decade. Over the span the data covers, that is a shift of well over a standard deviation. Full treatment in The Flynn Effect Explained.

The genome did not change on that timescale. Something environmental moved measured ability by an amount that dwarfs any training study, and the leading candidates are longer schooling, more cognitively demanding work, better nutrition, smaller families, and a visual environment saturated with abstract symbolic representation.

Notice that the gains were not uniform across abilities. They were concentrated in fluid and visual reasoning measures, with much smaller changes in vocabulary and arithmetic. That uneven pattern is consistent with the matching principle, though it does not test it: the abilities that moved most were the ones whose supporting environment changed most.

Whatever else the Flynn effect proves, it rules out the claim that population level cognitive performance is immovable. The mechanism is diffuse and operates over generations, which makes it a poor guide to what an individual should do next week. But as a proof that the ceiling is not where pessimists place it, it is decisive.

10 Heritability Does Not Mean Fixed

The most common objection to any claim about improving intelligence is that intelligence is highly heritable. Heritability estimates for adult IQ commonly land between 0.5 and 0.8, and the figure is often deployed as though it settles the question. It does not, and the reason is a persistent misunderstanding of what the statistic means.

Heritability is the proportion of variance in a trait, within a specific population at a specific time, that is associated with genetic variance in that population. It is a property of a population, not of a person, and it says nothing about whether the trait responds to intervention. Full treatment in Is IQ Genetic?.

Height is the standard illustration and it is a good one. Height is more heritable than intelligence by most estimates, and average height in many countries rose by several inches over the twentieth century through nutrition alone. High heritability coexisted with a large environmental shift, because heritability measures the distribution of differences under current conditions, not the range of outcomes possible under different ones.

There is also a subtlety that cuts against the fatalistic reading. Heritability of intelligence rises with age, from roughly 0.2 in early childhood to 0.6 or higher in adulthood. The usual explanation is gene-environment correlation: as people gain autonomy they increasingly select environments that match their propensities, so genetic differences get amplified through environmental channels. If environment is the amplifier, then environment is a lever, and the childhood window where heritability is lowest is precisely where intervention has the most room.

11 What Actually Moves the Needle

Pulling the threads together, the interventions with real support share a small number of features. They run long, they demand effort, they are structurally close to the ability targeted, and they are difficult enough to stay difficult.

The last card deserves emphasis because it is where most people have the largest available gain and the least interest in looking. Chronic sleep restriction degrades processing speed and working memory measurably. Somebody sleeping five hours a night is not measuring their capacity. They are measuring their capacity minus a deficit, and removing the deficit will look exactly like an improvement on a retest. Substance use belongs in the same category of removable deficit, and it is also where the popular version has drifted furthest from the evidence, because the widely quoted Dunedin result on persistent adolescent cannabis use lost most of its force once discordant twin and confounding analyses tried to reproduce it.

12 Dose, Frequency, and Duration

Every intervention question in physiology comes with a dose. Nobody would ask whether running improves endurance without specifying how far, how often, and for how long. Cognitive training is discussed as though dose does not exist, which is one reason the literature looks so discouraging.

The arithmetic is worth doing. A typical working memory training study delivers around twenty sessions of twenty five minutes, which is about eight hours of training. One school year delivers on the order of a thousand hours. The intervention that works delivers more than a hundred times the dose of the intervention that does not, and the field then reports the second as evidence about the first.

Frequency matters independently of total volume. Distributed practice outperforms massed practice across essentially every learning domain studied, which means the same total hours spread over months will generally beat the same hours crammed into weeks. Most training studies compress rather than distribute, because a study that runs for two years is a study nobody funds.

Difficulty is the third variable and the most commonly mishandled. A task that stops being hard stops training anything. Adaptive difficulty, where the task tracks performance and stays near the edge of capability, is a design feature of the better protocols and absent from most commercial products, which are engineered to feel rewarding rather than to stay effortful.

13 The Fadeout Problem, and What It Actually Tells Us

A pattern that recurs across early intervention research is fadeout: initial score gains shrink over the years following the end of the program. It is often cited as evidence that gains are illusory.

The pattern is real and the interpretation is too quick. Bailey and colleagues (2017) examined why early childhood intervention effects fade and identified a mechanism that has nothing to do with the gains being fake. Control groups catch up, because the environment eventually delivers to everyone much of what the program delivered early. If a preschool program teaches skills that school teaches anyway two years later, the gap closes without the treated children losing anything.

They also identified the conditions under which effects persist. Skills that are foundational rather than easily acquired elsewhere, and environments that continue to support what the intervention built, produce durable differences. That is a design specification, not a counsel of despair.

Applied to individual training, the lesson is that a gain sustained by ongoing practice behaves differently from a gain produced by a finite program and then abandoned. Nobody expects fitness gains to persist years after training stops. Expecting cognitive gains to is a double standard.

14 What Does Not Work, Stated Plainly

An honest article on this subject has to be as specific about failures as about successes, because the failures are where most of the money goes.

Commercial brain training apps promising general intelligence gains have a poor evidence record. The problem is structural rather than incidental: their business model requires short, pleasant sessions with visible progress, and the training features that appear to matter are long duration, sustained difficulty, and structural overlap with a named target ability. A product optimised for retention is optimised against the mechanism.

Listening to classical music does not raise intelligence. The original 1993 finding described a brief improvement on a spatial task after listening to Mozart, which was inflated by media coverage into a claim about intelligence that the authors never made, and subsequent work attributed the small effect to arousal and mood rather than anything specific to the music.

Nootropics and supplements have not demonstrated meaningful gains in healthy, well-nourished adults. Correcting a genuine deficiency helps, which is a different claim.

Practising IQ test items raises IQ test scores without necessarily raising ability. This is the most important failure mode to understand, and it gets its own section next.

15 Practice Effects Are Not Improvement

Take the same intelligence test twice and you will usually score higher the second time: about a third of a standard deviation on average, roughly five IQ points, and about half a standard deviation by the third sitting, after which the gains level off, according to the meta-analysis by Scharfen, Peters and Holling (2018). This is a well documented measurement phenomenon, not a change in ability, and it is the single largest source of self-deception in this area. That phenomenon is why what makes an IQ score accurate cannot be reduced to internal consistency: a test also has to show that a person gets a similar score on a second occasion, with the practice gain accounted for rather than mistaken for growth.

Retest gains come from several sources. You recognise item formats and lose no time to comprehension. You have already solved specific items and can recall answers. You are less anxious in a familiar procedure. None of these reflect a change in the underlying capacity the test was built to estimate.

This is why test publishers specify retest intervals, and why professional evaluations record whether a battery has been administered recently. It is also why a program that trains on items resembling test items will show impressive gains that do not generalise. The training has taught the test rather than the ability.

The distinction to hold is between improving your score and improving the thing your score estimates. Only the second is worth pursuing, and only the first is easy to sell. If you want to know how much of a change is real, the relevant concept is the standard error of measurement, covered in Reliability vs Validity, and the norming process described in How IQ Scores Are Normed.

16 How to Tell Whether You Actually Improved

If you decide to train, measurement is where most self-assessment goes wrong. A single before and after comparison on the same test cannot distinguish improvement from practice effect, normal fluctuation, or a better night of sleep. A usable baseline needs an instrument with real norms and a reported reliability figure, which is the minimum a real online IQ test has to publish before a before-and-after comparison can mean anything.

Four principles make the comparison more informative.

Measure at the subtest level rather than the composite, because a composite dilutes exactly the signal you are looking for. If you trained a spatial ability, look at the spatial subtests. A change there with no change elsewhere is evidence of specific transfer. A uniform rise across everything is more likely to be practice, effort, or condition.

Allow enough time between measurements that item memory has decayed, and understand that a shorter interval inflates the second score.

Expect the effect to be modest and specific rather than large and global, and treat any claim of a large global jump over a short period with suspicion, including your own.

Interpret changes against the test's measurement error rather than as exact values. A five point difference on an instrument whose standard error is three points is not clearly a real change. ACIS reports domain level results across six domains and twenty subtests, which is the granularity this kind of comparison requires.

17 A Realistic Plan

Everything above reduces to a sequence that is unglamorous and defensible.

Start by measuring a profile rather than a single number, because you cannot target an ability you have not identified. A profile shows which domains are already strong and which have room, and the room is where returns are largest.

Fix the floor before chasing the ceiling. Sleep, aerobic fitness, and correcting any nutritional deficiency are the cheapest available gains for anyone currently operating below their own baseline, and they are frequently larger than anything a training program will deliver.

Choose one or two narrow abilities rather than trying to raise everything at once, and pick training whose structure genuinely overlaps the target. Use the table in section 5 as the filter: if you cannot articulate what the exercise and the ability share, that is a reason to expect nothing.

Commit to a timescale measured in years rather than weeks, keep the difficulty at the edge of your capability, and distribute practice rather than cramming it. Then re-measure at the subtest level after a long enough interval, and read the result against measurement error rather than as a verdict.

None of this promises thirty points. What it promises is that the effort goes into channels where the mechanism is plausible and the evidence is favourable, instead of into a game that was never going to reach the ability you cared about.

18 FAQ: Improving Cognitive Ability

19 Best Next Step

You cannot target an ability you have not measured. A single composite number tells you where you sit overall and nothing about which of the underlying abilities has room to move, which is the only information a training decision actually needs.

ACIS reports across six cognitive domains and twenty subtests, so the result is a profile rather than a label. If you want to understand the structure before testing, start with The CHC Model. If you want to understand what the resulting number does and does not mean, read What IQ Measures and Reliability vs Validity.

20 Sources Behind This Page

This page argues against a common reading of the training literature, so every claim above is anchored to a primary source you can check. The meta-analyses and the replication failures are both included, because an honest case has to survive its own counter-evidence.

• Ritchie, S.J. & Tucker-Drob, E.M. (2018). How much does education improve intelligence? A meta-analysis. Psychological Science, 29(8), 1358-1369. Pooled 142 effect sizes across more than 600,000 participants using three independent designs, finding 1 to 5 IQ points per additional year of schooling with effects durable across the lifespan.
• McGrew, K.S. (2009). CHC theory and the human cognitive abilities project. Intelligence, 37(1), 1-10. The consolidation of the Cattell-Horn-Carroll framework into the three-stratum structure of broad and narrow abilities that current test publishers use.
• Jaeggi, S.M., Buschkuehl, M., Jonides, J. & Perrig, W.J. (2008). Improving fluid intelligence with training on working memory. PNAS, 105(19), 6829-6833. The original dual n-back transfer claim, including the dose-response pattern that later work struggled to replicate.
• Redick, T.S. et al. (2013). No evidence of intelligence improvement after working memory training: A randomized, placebo-controlled study. Journal of Experimental Psychology: General, 142(2), 359-379. Improvement on the trained task with no transfer to reasoning once an active control was included.
• Melby-Lervag, M. & Hulme, C. (2013). Is working memory training effective? A meta-analytic review. Developmental Psychology, 49(2), 270-291. Twenty-three studies pooled, finding short-lived task-specific effects without generalisation.
• Au, J. et al. (2015). Improving fluid intelligence with training on working memory: A meta-analysis. Psychonomic Bulletin & Review, 22(2), 366-377. A small positive transfer estimate, and the starting point of the unresolved exchange over baseline differences and publication bias.
• Sala, G. & Gobet, F. (2017). Does far transfer exist? Negative evidence from chess, music, and working memory training. Current Directions in Psychological Science, 26(6), 515-520. Far transfer effects shrink toward zero as design quality and control group activity improve.
• Ceci, S.J. (1991). How much does schooling influence general intelligence and its cognitive components? Developmental Psychology, 27(5), 703-722. The converging evidence from summer holidays, delayed school entry, and intermittent schooling.
• Protzko, J., Aronson, J. & Blair, C. (2013). How to make a young child smarter: Evidence from the database of raising intelligence. Perspectives on Psychological Science, 8(1), 25-40. A systematic catalogue of which manipulations move measured intelligence in children and which do not.
• Bailey, D., Duncan, G.J., Odgers, C.L. & Yu, W. (2017). Persistence and fadeout in the impacts of child and adolescent interventions. Journal of Research on Educational Effectiveness, 10(1), 7-39. Why early gains fade, and the conditions under which they persist instead.
• Trahan, L.H., Stuebing, K.K., Hiscock, M.K. & Fletcher, J.M. (2014). The Flynn effect: A meta-analysis. Psychological Bulletin, 140(5), 1332-1360. Open access. Scores rose about 2.31 points per decade across 285 studies, concentrated in fluid and visual measures.
• Salthouse, T.A. (2004). Localizing age-related individual differences in a hierarchical structure. Intelligence, 32(6), 541-561. The differing developmental trajectories of fluid and crystallized abilities that determine when an intervention is working with development or against it.
• Uttal, D.H., Meadow, N.G., Tipton, E., Hand, L.L., Alden, A.R., Warren, C. & Newcombe, N.S. (2013). The malleability of spatial skills: A meta-analysis of training studies. Psychological Bulletin, 139(2), 352-402. Across 217 studies, spatial training produced an average effect of 0.47 standard deviations, stable over delays and transferring to untrained spatial tasks.
• Melby-Lervåg, M., Redick, T.S. & Hulme, C. (2016). Working memory training does not improve performance on measures of intelligence or other measures of far transfer. Perspectives on Psychological Science, 11(4), 512-534. Gains on trained and closely related working memory tasks, with no reliable transfer to measures of intelligence.
• Ritchie, S.J., Bates, T.C. & Deary, I.J. (2015). Is education associated with improvements in general cognitive ability, or in specific skills? Developmental Psychology, 51(5), 573-582. In 1,091 people tested at 11 and 70, the association between education and later scores ran through specific skills rather than through g.
• Scharfen, J., Peters, J.M. & Holling, H. (2018). Retest effects in cognitive ability tests: A meta-analysis. Intelligence, 67, 44-66. Scores rise about a third of a standard deviation from a first to a second administration and about half by the third, then level off.