IQ Metrics
IIF Certified Assessment Start IQ Test
IIF Certified Assessment
Data studyBrain and cognition · Version 2026.11

Can a brain scan show your IQ? Every MRI study, sample size against result

We collected every published MRI study of healthy people we could read that relates the brain to an intelligence score, 2008-2026, and set each result against its sample size, method and whether family background was taken into account.

About this study

Report type
Data study: our ledger of published MRI-intelligence results, sample size against result
Data
23 results from 18 papers, 2008-2026; healthy samples of 107 to 18,363 people (UK Biobank, ABCD, HCP, cohorts)
Sources
OpenAlex searches (2,208 records), 78 records read, Europe PMC open full texts, the 2026 Science paper
Years
Results 2008 to 2026; searches as of 11 November 2026
Licence
CC BY 4.0 (our dataset and charts)
Version
2026.11
Sources
30
Short answer

No: a brain scan cannot show one person’s IQ. In the 12 largest results we found (1,022 to 18,363 people), brain MRI and intelligence scores correlate at a median r = .35, so the scan accounts for about 12% of the differences in test scores. IQ Metrics Research ledger of 23 results, 2008-2026, as of 11 November 2026.

23 resultsmedian r .35 at 1,000+about 12% explainedCC BY 4.0
12%of score differences a scan accounts for in the biggest studies
4%for total brain volume alone
36% lowerbrain-IQ link (connectivity) after neighbourhood-SES adjustment (2026)

01 Key findings

  1. A scan is not an IQ test. In the 12 results with 1,000 or more people, the median r is .35: the brain measure accounts for about 12% of the differences in scores, and nothing about any one person.IQ Metrics calculation
  2. Brain size matters least. Total brain volume, the oldest measure, links at r = .18 to .28 in the three volume results with 1,000+ people: about 4% of score differences.Nave, Cox, Lee
  3. Bigger studies did not shrink the link in our ledger. The rank correlation between sample size and r is -.02; studies of 1,000+ people report a median of .35 against .42 for 100-999.IQ Metrics calculation
  4. Neighbourhood opportunity explains part of it. In ABCD children, adjusting for neighbourhood social and economic opportunity cut the brain-IQ link by 30 to 40% (connectivity r = .36 to .23).Marek et al. 2026
  5. Few studies check. 4 of 23 results adjust for family income, SES or the family; one prints the link both ways.IQ Metrics ledger
  6. Prediction got stricter. All 12 results since 2021 either test their model on unseen people or do not say; same-sample correlations are mostly pre-2020.IQ Metrics ledger

02 Statistics to quote

r = .35

median link between brain MRI and an intelligence score in the 12 results with 1,000 or more people (about 12% of score differences).

IQ Metrics calculation
4%

of the differences in intelligence scores accounted for by total brain volume in the three volume results with 1,000+ people (median r = .19).

IQ Metrics calculation
r = .28

total brain volume and a g factor in the largest result, 18,363 UK Biobank adults aged 44-81 (95% CI .25 to .30).

Cox et al. 2019
36% lower

brain-IQ prediction from resting connectivity after adjusting for neighbourhood social and economic opportunity: r .36 to .23 in ABCD children tested on 2,263 unseen children (Marek 2026).

Marek et al. 2026
4 of 23

results in our ledger take family income, an SES index or the family into account; only one study prints the link both before and after.

IQ Metrics ledger
-.02

rank correlation between sample size and reported r across all 23 results: bigger studies did not report smaller links on average (IQ Metrics calculation).

IQ Metrics calculation
r = .42 vs .21

median for 13 results tested on unseen people vs 7 same-sample correlations, which are mostly of brain volume.

IQ Metrics calculation
11,878

children enrolled in the ABCD Study sample behind the 2026 SES result; 4,579 were analysed after quality checks.

Marek et al. 2026
23

results from 18 papers, 2008-2026, with n from 107 to 18,363, in healthy samples (no patient groups).

IQ Metrics ledger
13,608

UK Biobank adults in the preregistered brain-size study: total brain volume and fluid intelligence r = .19 with income-related status in the model.

Nave et al.
r = .24

pooled brain volume-IQ link in the Pietschnig 2015 meta-analysis (88 studies, more than 8,000 people), which also found the published links overestimated.

Pietschnig et al. 2015
50%

the most any ledger result accounts for (Choi 2008, r = .71 from 50%%, 225 people, regions picked and tested within one study).

Choi et al. 2008

03 Can a brain scan show your IQ?

No: in the 12 largest published results we found, brain MRI accounts for about 12% of the differences in intelligence scores between people (median r = .35), far too little to read one person’s IQ from a scan. Our ledger holds 23 results from 18 papers, 2008-2026, each relating an MRI measure (brain volume, the cortex, white matter, connectivity at rest or during tasks) to a general intelligence score in a healthy sample, with the sample size, the method and whether family circumstances were taken into account.

A correlation of r = .35 describes a group. It says that, across thousands of people, those with a given brain pattern score a little higher on average; it does not say that a scan can place any individual on the IQ scale. The brain-imaging results most often quoted in the news were found in a few hundred people; the largest samples, from the UK Biobank, the Adolescent Brain Cognitive Development (ABCD) Study and the Human Connectome Project (HCP), report links of r = .19 to .54. An IQ score itself is measured with a test, as our guide to the IQ score range explains.

We built our own ledger because the nearest existing work is one systematic review: Vieira and colleagues (Intelligence, 2022) reviewed 30 machine-learning studies found by a Scopus search to 15 December 2020 and estimated r = .42 for predicting general intelligence from fMRI [28] [6]. That review covers out-of-sample prediction only, does not flag which studies took family income into account, and ends before the biobank-sized studies of 2021-2026. Our ledger adds same-sample correlations, results through November 2026, the SES flag, sample-size bands, the share of score differences in plain words and an open CSV.

Brain MRI
Magnetic resonance imaging: a scanner measures brain structure (volume, cortical thickness, white matter) or, with functional MRI (fMRI), changes in blood oxygen that track brain activity. It does not use radiation.
Correlation (r)
A number from -1 to +1 for how closely two measures move together across people. r = 0 means no link; r = .30 is a modest link; r = 1 would mean one predicts the other exactly.
Share of score differences (r squared)
r x r, as a percentage: how much of the spread in test scores between people a brain measure accounts for, in that sample. r = .30 gives 9%. It is a property of a group, not a person.
Out-of-sample prediction
A model is built on some people and tested on others it has never seen (cross-validation or a held-out sample). It guards against fitting noise, which same-sample correlations cannot do.
Resting-state functional connectivity
How closely the activity of brain regions rises and falls together while a person lies in the scanner doing nothing. Task fMRI records the same during a task such as remembering a sequence.
General intelligence (g)
What different mental tests share: people who do well on one tend to do well on others. Studies here use full-scale IQ, a g factor, a fluid reasoning score or a test battery total.
Socioeconomic status (SES)
A family’s social and economic position: income, parental education or the opportunities of the neighbourhood. It is linked both to test scores and to brain measures, so it can make a brain-IQ link look stronger.
Brain-wide association study (BWAS)
A study that links one trait, such as an IQ score, to measures across the whole brain. Marek and colleagues showed in 2022 that such studies need thousands of people to give stable numbers [21].

04 How many MRI studies link brain scans to intelligence?

We found 23 results from 18 papers that print both a sample size and a usable correlation, published 2008-2026. They came from 2,208 unique records in 12 OpenAlex searches, of which 564 named an MRI measure and an intelligence score, 409 passed a filter for patient groups and reviews and 51 printed a correlation in the abstract; with 27 named leads from the reviews and the news, we read 78 records by hand. The count is a lower bound: many studies print their result only in a figure or a table we could not open, and we list every record we read and left out.

Figure 1

From 2,208 search records to 13 kept papers, plus 5 from named leads

Stage 1: unique records from 12 OpenAlex searches. Stages 2-4: scripted word rules on titles and abstracts (written before screening). Stage 5: the 51 survivors read by hand; 13 kept. The 27 named leads read the same way added 5 papers: 18 papers and 23 results in all; 4 reviews went to a separate table. As of 11 November 2026.

Source: OpenAlex [1]; Europe PMC [2]; screening by IQ Metrics (data/screen_decisions.tsv)

Data table
stagecountpct_of_firstpct_of_previous
Unique OpenAlex records (12 searches)2208100100
Names an MRI measure and an intelligence score56425.525.5
No disorder, patient or animal words in the title; not a review40918.572.5
Abstract prints a correlation or variance explained512.312.5
Kept after reading the abstract or open text130.625.5
Embed this chart

Paste this HTML into your page: it carries the numbers and the credit link to this figure.

Most results come from three big shared datasets: 8 from the ABCD Study of American children aged 9-10, 7 from the Human Connectome Project of young adults and 2 from middle-aged and older UK Biobank adults; the other 6 come from cohorts and samples of a single laboratory. The search for intelligence in brain scans began before MRI: in 1988 Haier and colleagues measured glucose use in the cortex with positron emission tomography (PET) while people solved an abstract reasoning test [5]. We show that study as history, not as a row.

Figure 2

From one PET study in 1988 to biobank-sized MRI samples

Every ledger result by year and brain measure, plus the 1988 PET study that started the search for intelligence in brain scans (history only). Colour = same-sample correlation, test on unseen people, or not stated. 18 of the 23 results appeared in 2018 or later. As of 11 November 2026.

Source: Our ledger [1]; Haier et al. 1988 [5]

Data table
brain measuretypefromtolabeldetail
Historyevent1988Haier 1988Haier and colleagues: glucose use in the cortex during an abstract reasoning test, PET (history only, not a ledger row)
Brain volumeevent2009Shenkin 2009Shenkin 2009: r = .21, n = 107, Own sample (community volunteers, Scotland)
Brain volumeevent2018Nave 2018Nave 2018: r = .19, n = 13,608, UK Biobank
Brain volumeevent2019Cox 2019Cox 2019: r = .28, n = 18,363, UK Biobank
Brain volumeevent2019Lee 2019Lee 2019: r = .18, n = 1,022, Human Connectome Project (siblings)
Cortex thickness or areaevent2024Michel 2024 (grey)Michel 2024 (grey): r = .39, n = 11,876, ABCD Study
Cortex thickness or areaevent2026Marek 2026 (thickness)Marek 2026 (thickness): r = .23, n = 2,263, ABCD Study
White matterevent2011Ganjavi 2011Ganjavi 2011: r = -.15, n = 198, NIH MRI Study of Normal Brain Development
White matterevent2018Alnaes 2018Alnaes 2018: r = .39, n = 729, Philadelphia Neurodevelopmental Cohort
White matterevent2024Michel 2024 (white)Michel 2024 (white): r = .35, n = 11,876, ABCD Study
Resting connectivityevent2016Hearne 2016Hearne 2016: r = .38, n = 317, Human Connectome Project
Resting connectivityevent2018Dubois 2018Dubois 2018: r = .45, n = 884, Human Connectome Project
Resting connectivityevent2021Sripada 2021Sripada 2021: r = .42, n = 5,937, ABCD Study
Resting connectivityevent2022Marek 2022Marek 2022: r = .34, n = 3,604, ABCD Study
Resting connectivityevent2026Marek 2026 (RSFC)Marek 2026 (RSFC): r = .36, n = 2,263, ABCD Study
Task fMRIevent2020Sripada 2020Sripada 2020: r = .50, n = 967, Human Connectome Project
Task fMRIevent2022Tetereva 2022 (tasks)Tetereva 2022 (tasks): r = .56, n = 873, Human Connectome Project
Several measures togetherevent2008Choi 2008Choi 2008: r = .71, n = 225, Own sample (young adults)
Several measures togetherevent2015Ritchie 2015Ritchie 2015: r = .42, n = 672, Lothian Birth Cohort 1936
Several measures togetherevent2022Tetereva 2022Tetereva 2022: r = .57, n = 873, Human Connectome Project
Several measures togetherevent2024Michel 2024Michel 2024: r = .44, n = 11,876, ABCD Study
Several measures togetherevent2024Popp 2024Popp 2024: r = .25, n = 1,030, Human Connectome Project
Several measures togetherevent2024Popp 2024 (new sample)Popp 2024 (new sample): r = .19, n = 567, Independent sample (not named in the abstract)
Several measures togetherevent2025Wang 2025Wang 2025: r = .54, n = 10,754, ABCD Study
Embed this chart

Paste this HTML into your page: it carries the numbers and the credit link to this figure.

05 Do bigger studies find a weaker link between the brain and IQ?

Not in our ledger: results with 1,000 or more people report a median r = .35, against .42 for studies of 100-999 people, and the rank correlation between sample size and r is -.02, which is no relation. Marek and colleagues showed the inflation directly inside one dataset: when they drew small samples from the ABCD Study, “irreproducible, inflated effect sizes were ubiquitous”, and the links stabilised only with about 2,000 people or more [21]. Our ledger cannot show that shrinking because it holds no result with fewer than 100 people: the classic small studies (for example Andreasen 1993, 67 volunteers) print no r in the text we could read [8].

Figure 3

Sample size against result: 23 brain-scan results from 107 to 18,363 people

Each dot is one result: people scanned (across, log scale) and the reported link between the brain measure and an intelligence score (up, r). Results with 1,000 or more people have a median r of .35 against .42 for 100-999 people; the rank correlation between sample size and r is -.02 (IQ Metrics calculation). As of 11 November 2026.

Brain volumeCortex thickness or areaWhite matterResting connectivityTask fMRISeveral measures together

Source: Our ledger of 23 results [1]; r and n as printed in each paper

Data table
resultnrgroup
Ganjavi 2011198-0.1White matter
Shenkin 20091070.2Volume
Choi 20082250.7Multimodal
Ritchie 20156720.4Multimodal
Hearne 20163170.4Resting FC
Dubois 20188840.4Resting FC
Alnaes 20187290.4White matter
Nave 2018136080.2Volume
Cox 2019183630.3Volume
Lee 201910220.2Volume
Sripada 20209670.5Task fMRI
Sripada 202159370.4Resting FC
Marek 202236040.3Resting FC
Tetereva 20228730.6Multimodal
Tetereva 2022 (tasks)8730.6Task fMRI
Michel 2024118760.4Multimodal
Michel 2024 (grey)118760.4Cortex
Michel 2024 (white)118760.4White matter
Popp 202410300.2Multimodal
Popp 2024 (new sample)5670.2Multimodal
Wang 2025107540.5Multimodal
Marek 2026 (RSFC)22630.4Resting FC
Marek 2026 (thickness)22630.2Cortex
Embed this chart

Paste this HTML into your page: it carries the numbers and the credit link to this figure.

The flat pattern hides two opposite trends. Brain-volume studies are large and report modest links (r = .18 to .275 in samples of 1,000 or more), while prediction studies of connectivity and combined scans moved from a few hundred young adults in the HCP to thousands of ABCD children and kept reporting r = .34 to .54 [21] [20] [25]. The highest result, r = .71, comes from 225 young adults whose brain regions were picked and tested within the same study [11]; the second highest, r = .57, combines seven task scans, structure and resting connectivity in 873 HCP adults [22].

Figure 4

Every result with its 95% interval, largest sample first

Square = reported r (area = people scanned); line = 95% interval. Only Cox 2019 prints an interval; the other 22 are Fisher-z intervals from r and n (IQ Metrics calculation), which ignore model choices and so are too narrow for prediction studies. No pooled estimate: the rows mix measures and methods. As of 11 November 2026.

Reported link with an intelligence score (r)

Source: Our ledger; intervals by IQ Metrics except Cox 2019 [17]

Data table
resultstudydetailnestimateci_lowci_high
All 23 resultsCox 2019Volume, UK Biobank183630.30.30.3
All 23 resultsNave 2018Volume, UK Biobank136080.20.20.2
All 23 resultsMichel 2024Combined, ABCD118760.40.40.5
All 23 resultsMichel 2024 (grey)Cortex, ABCD118760.40.40.4
All 23 resultsMichel 2024 (white)White m., ABCD118760.40.30.4
All 23 resultsWang 2025Combined, ABCD107540.50.50.6
All 23 resultsSripada 2021Rest FC, ABCD59370.40.40.4
All 23 resultsMarek 2022Rest FC, ABCD36040.30.30.4
All 23 resultsMarek 2026 (RSFC)Rest FC, ABCD22630.40.30.4
All 23 resultsMarek 2026 (thickness)Cortex, ABCD22630.20.20.3
All 23 resultsPopp 2024Combined, HCP10300.20.20.3
All 23 resultsLee 2019Volume, HCP10220.20.10.2
All 23 resultsSripada 2020Task, HCP9670.50.50.5
All 23 resultsDubois 2018Rest FC, HCP8840.40.40.5
All 23 resultsTetereva 2022Combined, HCP8730.60.50.6
All 23 resultsTetereva 2022 (tasks)Task, HCP8730.60.50.6
All 23 resultsAlnaes 2018White m., PNC7290.40.30.5
All 23 resultsRitchie 2015Combined, LBC19366720.40.40.5
All 23 resultsPopp 2024 (new sample)Combined, new sample5670.20.10.3
All 23 resultsHearne 2016Rest FC, HCP3170.40.30.5
All 23 resultsChoi 2008Combined, own sample2250.70.60.8
All 23 resultsGanjavi 2011White m., NIH MRI study198-0.1-0.3-0
All 23 resultsShenkin 2009Volume, own sample1070.200.4
Embed this chart

Paste this HTML into your page: it carries the numbers and the credit link to this figure.

The intervals are narrow for the biobank-sized studies (Cox and colleagues print .25 to .30 for 18,363 people [17]) and wide for studies of a few hundred. A model tested on a new sample can do worse than in its own: Popp and colleagues report r = .25 in 1,030 HCP adults and r = .19 in a completely independent sample of 567 [24].

Figure 5

Samples grew from hundreds to thousands; the reported links did not fall to zero

Each bubble is one result by publication year (across) and reported r (up); bubble area = people scanned. Every result since 2021 in the ledger tests its model on people it was not trained on, or does not say; 12 of the 12 results with 1,000 or more people appeared from 2018 on. As of 11 November 2026.

Same-sample correlationTested on unseen peopleNot stated in the abstract
People scanned5005k15k

Source: Our ledger [1]

Data table
resultyearrgroupn
Ganjavi 20112011-0.1in-sample198
Shenkin 200920090.2in-sample107
Choi 200820080.7out-of-sample225
Ritchie 201520150.4in-sample672
Hearne 201620160.4in-sample317
Dubois 201820180.4out-of-sample884
Alnaes 201820180.4out-of-sample729
Nave 201820180.2in-sample13608
Cox 201920190.3in-sample18363
Lee 201920190.2in-sample1022
Sripada 202020200.5out-of-sample967
Sripada 202120210.4out-of-sample5937
Marek 202220220.3out-of-sample3604
Tetereva 202220220.6out-of-sample873
Tetereva 2022 (tasks)20220.6out-of-sample873
Michel 202420240.4not stated11876
Michel 2024 (grey)20240.4not stated11876
Michel 2024 (white)20240.4not stated11876
Popp 202420240.2out-of-sample1030
Popp 2024 (new sample)20240.2out-of-sample567
Wang 202520250.5out-of-sample10754
Marek 2026 (RSFC)20260.4out-of-sample2263
Marek 2026 (thickness)20260.2out-of-sample2263
Embed this chart

Paste this HTML into your page: it carries the numbers and the credit link to this figure.

Figure 6

Most results sit between r = .1 and r = .6

All 23 results in bins of 0.1. Median r = .38. One result is negative (corpus callosum area after correcting for brain size, [[ganjavi2011]]); the highest is .71 (Choi 2008, regions chosen and tested in one study of 225 people). As of 11 November 2026.

Source: Our ledger; bins by IQ Metrics

Data table
fromtoresults
-0.2-0.11
-0.100
00.10
0.10.23
0.20.34
0.30.46
0.40.54
0.50.64
0.60.70
0.70.81
Embed this chart

Paste this HTML into your page: it carries the numbers and the credit link to this figure.

06 Does brain size correlate to intelligence?

Yes, weakly: in the three largest brain-volume results (1,022 to 18,363 people) total brain volume correlates with intelligence scores at r = .18 to .275, about 4% of the differences in scores at the median. The largest, 18,363 UK Biobank adults aged 44-81, gives r = .275 between total brain volume and a g factor from four tests [17]; a preregistered UK Biobank study of 13,608 adults gives r = .19 for fluid intelligence with height and socioeconomic status in the model [16].

The link survives inside families, which rules out the simplest version of a family-background explanation: Lee and colleagues found that, among siblings in the HCP (1,022 people), the sibling with the larger brain tends to score higher, a correlation of .18 after correcting for measurement error [18]. The 2015 meta-analysis by Pietschnig and colleagues pooled 88 studies of more than 8,000 people, healthy and clinical, to r = .24 and reported that “strong and positive correlation coefficients have been reported frequently” while small ones were often left out [27]. Our news article on what a correlation of 0.24 means and our study of famous brains weighed after death show why a heavy brain is not a sign of genius; brain size has also changed over human evolution, as our study of whether human brains are shrinking shows.

One brain-structure result runs the other way: in 198 healthy American children and adolescents, a larger corpus callosum (the band of fibres between the hemispheres) went with slightly lower IQ once total brain size was taken into account (r = -.147) [9]. In 107 community-dwelling adults aged 75-81, whole brain volume correlated r = .21 with a g factor, but the link nearly vanished (.03) once intracranial area, an estimate of the brain’s largest size earlier in life, was taken into account [10].

07 Which kind of brain scan is most linked to IQ?

Task fMRI and combined models report the strongest links (median r = .53 and .44), brain volume the weakest (.20), with white matter (.35), the cortex (.31) and resting connectivity (.38) between them. The groups are small, two to seven results each, and they differ in method: almost every volume result is a same-sample correlation, while almost every connectivity and task result is a model tested on unseen people, so the comparison mixes brain measure and method.

Figure 7

Brain volume shows the weakest links; combined and task scans the strongest

One dot per result; box = middle half, whiskers = lowest and highest. Median r: brain volume .20 (4 results), cortex .31 (2), white matter .35 (3), resting connectivity .38 (5), task fMRI .53 (2), several measures together .44 (7). Groups of 2-7 results: read the dots, not the boxes. As of 11 November 2026.

Reported r.

Source: Our ledger; quartiles by IQ Metrics

Data table
brain measurenlowq1medianq3high
Brain volume40.20.20.20.20.3
Cortex thickness or area20.20.30.30.40.4
White matter3-0.10.10.40.40.4
Resting connectivity50.30.40.40.40.4
Task fMRI20.50.50.50.50.6
Several measures together70.20.30.40.60.7
Embed this chart

Paste this HTML into your page: it carries the numbers and the credit link to this figure.

Task scans do well partly because the tasks themselves are tests: Sripada and colleagues predicted general cognitive ability at r = .50 from the brain’s response to a working-memory task in 967 HCP adults [19], and Marek and colleagues note that task-fMRI links with cognitive ability shrink once performance on the scanner task is taken out [21]. Combining many kinds of scan helps: in 10,754 ABCD children, Wang and colleagues reached r = .54 with 45 kinds of MRI together, and the same children’s family and environment factors alone reached r = .49 [25]. In 11,876 ABCD children, grey matter and white matter together accounted for 19.0% of differences in cognitive performance, more than either alone [23].

Figure 8

Median r by brain measure and sample size

Median reported r per cell; empty = no result. Counts: brain volume 1 and 3; cortex thickness or area 0 and 2; white matter 2 and 1; resting connectivity 2 and 3; task fmri 2 and 0; several measures together 4 and 3. No result in the ledger has fewer than 100 people: the classic small studies print no r we could read. As of 11 November 2026.

Source: Our ledger; medians by IQ Metrics

Data table
brain measure100-999 people1,000 or more
Brain volume0.20.2
Cortex thickness or area0.3
White matter0.10.4
Resting connectivity0.40.4
Task fMRI0.5
Several measures together0.50.4
Embed this chart

Paste this HTML into your page: it carries the numbers and the credit link to this figure.

Whole-brain summaries do worst. In Molloy and colleagues’ study of 5,937 ABCD children (Cerebral Cortex, 2025), whole-brain graph measures such as global efficiency and small-worldness “fail to exhibit meaningful relationships” with general cognitive ability, while measures of single brain regions do, a result repeated in 847 HCP adults [4]. We did not make it a row because its abstract prints no r and we could not open its text. Our study of how much of the brain has been mapped explains what a connectome is, and our news article on what MRI scans of gifted children show covers the smaller studies of high scorers.

08 Can an MRI predict intelligence in people it has never seen?

Partly: 13 results test a brain model on people it was not built on, with a median r = .42, which is about 18% of the differences in scores; the 7 same-sample correlations have a median of .21, mostly for brain volume. Even the best-tested models miss individuals widely: Hilger and colleagues, predicting IQ from grey matter in 308 adults, found errors of more than ten IQ points and predictions that “varied closely around the sample mean” [7].

Figure 9

The field moved from same-sample correlations to tests on unseen people

Results in the ledger by publication period and analysis. 9 of the 12 results from 2021-2026 test the model on people it was not trained on; the other 3 do not say in the abstract. Testing on unseen people is the stricter check, but it is not a test on a new population unless the sample is new (2 results). As of 11 November 2026.

Source: Our ledger

Data table
periodSame-sample correlationTested on unseen peopleNot stated in the abstracttotal
2008-2015314
2016-2020437
2021-20269312
Embed this chart

Paste this HTML into your page: it carries the numbers and the credit link to this figure.

Testing on unseen people is not the same as testing on a new population. Most out-of-sample results split one dataset (HCP or ABCD) into training and test parts; only two use a separate sample: Popp’s model fell from r = .25 to .19 on a new sample of 567 people [24], and Choi’s explained 50% of IQ differences in an independent sample whose size the abstract does not give [11]. Vieira’s review found that “small sample sizes” and untreated confounders, including kinship, were common risks of bias in this literature [28].

09 Do family income and neighbourhood explain the link between brain scans and IQ?

Partly: in the one study that prints both numbers, adjusting for the social and economic opportunity of the child’s neighbourhood cut the link between children’s brain scans and IQ scores by 30 to 40%, the paper’s own figure (resting connectivity r = .36 to .23, cortical thickness r = .23 to .13). Family income and parental education, used the same way, weakened the brain-IQ models less than the neighbourhood measure did. Marek and colleagues (Science, June 2026) analysed 4,579 ABCD children aged 9-10 from 11,878 enrolled, building models on 2,316 and testing them on 2,263 [26].

Figure 10

Adjusting for neighbourhood opportunity cut the brain-IQ link by 30 to 40%

Marek and colleagues (Science, 2026), ABCD Study children aged 9-10: out-of-sample r between brain-based predictions and IQ scores (NIH Toolbox total), trained on 2,316 children and tested on 2,263, before and after adjusting for the social and economic domain of the Child Opportunity Index. Resting connectivity .36 to .23; cortical thickness .23 to .13; the paper reports a reduction of 30 to 40%. The only ledger study that prints both numbers. As of 11 November 2026.

After adjusting for SESBefore

Out-of-sample r with IQ scores

Source: Marek et al. 2026 [26]; drop in % by IQ Metrics

Data table
brain measureAfter adjusting for SESBefore
Connectivity0.20.4
Thickness0.10.2
Embed this chart

Paste this HTML into your page: it carries the numbers and the credit link to this figure.

Of 649 measures of children’s lives, the one most strongly linked to brain connectivity was the social and economic opportunity of the child’s neighbourhood (r = .24 for the strongest single connection, against .16 for IQ scores) [26]. Adjusting for it weakened 95% of the single brain-IQ links, and about 70% were no longer statistically significant. The authors also report that brain-IQ models trained only on children from better-off neighbourhoods failed, and conclude that the link is consistent with “confounding by socioeconomics”. The paper reports a reduction, not a disappearance: the adjusted link (r = .23) is still there.

Figure 11

4 of 23 results take family circumstances into account

One square per result, coloured by whether the model behind the printed r adjusts for family income, an SES index, parental education or neighbourhood, or uses a within-family design. 4 results from 3 papers do; only Marek 2026 prints the r both ways (2 results). As of 11 November 2026.

Source: Our ledger; flags read from each paper [16] [18] [26]

Data table
resultgrouplabeldetail
marek2026ctBefore and after printedMarek 2026 (thickness)Marek 2026 (thickness): Before and after printed
marek2026rsBefore and after printedMarek 2026 (RSFC)Marek 2026 (RSFC): Before and after printed
nave2019Adjusted onlyNave 2018Nave 2018: Adjusted only
lee2019Adjusted onlyLee 2019Lee 2019: Adjusted only
choi2008Not adjusted for family circumstancesChoi 2008Choi 2008: Not adjusted for family circumstances
shenkin2009Not adjusted for family circumstancesShenkin 2009Shenkin 2009: Not adjusted for family circumstances
ganjavi2011Not adjusted for family circumstancesGanjavi 2011Ganjavi 2011: Not adjusted for family circumstances
ritchie2015Not adjusted for family circumstancesRitchie 2015Ritchie 2015: Not adjusted for family circumstances
hearne2016Not adjusted for family circumstancesHearne 2016Hearne 2016: Not adjusted for family circumstances
alnaes2018Not adjusted for family circumstancesAlnaes 2018Alnaes 2018: Not adjusted for family circumstances
dubois2018Not adjusted for family circumstancesDubois 2018Dubois 2018: Not adjusted for family circumstances
cox2019Not adjusted for family circumstancesCox 2019Cox 2019: Not adjusted for family circumstances
sripada2020Not adjusted for family circumstancesSripada 2020Sripada 2020: Not adjusted for family circumstances
sripada2021Not adjusted for family circumstancesSripada 2021Sripada 2021: Not adjusted for family circumstances
marek2022Not adjusted for family circumstancesMarek 2022Marek 2022: Not adjusted for family circumstances
tetereva2022Not adjusted for family circumstancesTetereva 2022Tetereva 2022: Not adjusted for family circumstances
tetereva2022tNot adjusted for family circumstancesTetereva 2022 (tasks)Tetereva 2022 (tasks): Not adjusted for family circumstances
michel2024Not adjusted for family circumstancesMichel 2024Michel 2024: Not adjusted for family circumstances
michel2024gNot adjusted for family circumstancesMichel 2024 (grey)Michel 2024 (grey): Not adjusted for family circumstances
michel2024wNot adjusted for family circumstancesMichel 2024 (white)Michel 2024 (white): Not adjusted for family circumstances
popp2024Not adjusted for family circumstancesPopp 2024Popp 2024: Not adjusted for family circumstances
popp2024xNot adjusted for family circumstancesPopp 2024 (new sample)Popp 2024 (new sample): Not adjusted for family circumstances
wang2025Not adjusted for family circumstancesWang 2025Wang 2025: Not adjusted for family circumstances
Embed this chart

Paste this HTML into your page: it carries the numbers and the credit link to this figure.

Only 4 of 23 results (3 papers) in our ledger take family circumstances into account at all: the UK Biobank brain-size study, with socioeconomic status in its model [16], the sibling study, where a within-family design holds the family constant [18], and the 2026 ABCD study. Most prediction studies adjust for age, sex, head motion or brain size but not for income, so their r may include the part of the link that runs through family circumstances. Sripada and colleagues found in 5,937 ABCD children that brain connectivity patterns linked to cognitive ability “significantly statistically mediated” the effect of SES, an analysis that treats the brain as a path, not a confounder [20].

10 How much of the difference in IQ scores can a brain scan explain?

About 12% in the biggest studies (median r = .35 across 12 results of 1,000 or more people), from 4% for total brain volume to 29% for the best combined model (r = .54 in 10,754 ABCD children) [25]. Squaring r converts it into the share of the spread in test scores that the brain measure accounts for, in that sample: r = .275 in 18,363 UK Biobank adults is 7.6% [17], and the 2026 ABCD cortical-thickness result after SES adjustment, r = .13, is 1.7% [26].

Figure 12

In the largest study of each kind, a scan accounts for 8% to 25% of score differences

r squared x 100 for the result with the most people in each brain-measure class (IQ Metrics calculation): the share of the differences in test scores between people that the brain measure accounts for, in that sample. Total brain volume in 18,363 people: 7.6%. As of 11 November 2026.

Source: Our ledger; r squared by IQ Metrics [17] [23] [20] [19]

Data table
labelvalue
Task fMRI: Sripada 2020 (n = 967)25
Several measures together: Michel 2024 (n = 11,876)19
Resting connectivity: Sripada 2021 (n = 5,937)18
Cortex thickness or area: Michel 2024 (grey) (n = 11,876)15.4
White matter: Michel 2024 (white) (n = 11,876)12.4
Brain volume: Cox 2019 (n = 18,363)7.6
Embed this chart

Paste this HTML into your page: it carries the numbers and the credit link to this figure.

The rest, the large majority, is everything a scan does not capture: the parts of thinking that no current scanner resolves, the test’s own measurement error, and the person’s experience, schooling and circumstances on the day. A share of differences is a statement about a population; it does not mean that 12% of any one person’s IQ “comes from” the brain region measured. The same caution applies to the popular brain claims checked in our brain statistics fact-check, the myths in left brain, right brain and five other myths about intelligence and our news article on what brain scans actually show.

11 What does this mean, and what does it not?

Brain MRI and intelligence scores are linked, modestly and reliably, in large healthy samples; the link is a correlation, not a cause, and part of it runs through family circumstances. None of the 23 results shows that a scan can read one person’s intelligence, and none supports sorting people into kinds of brain by intelligence. Where a model was tried on a new sample, its r fell (.25 to .19) [24].

Replication is the test that matters here, as in our replication scorecard of famous psychology findings: the biobank-era studies are the replications of the small early ones, and they report smaller, steadier numbers for brain volume. For what intelligence covers beyond a test score, see whether intelligence is limited to IQ and human brain facts; to exercise working memory and attention, try our brain games.

A scan cannot give your IQ. A test is how IQ is measured.

Our IQ test measures pattern reasoning and compares your result with adults, with a percentile.

Take the IQ test

12 Every result in our ledger

Table 1

Every result in our ledger (23 rows from 18 papers)

Largest sample first. * = r from a printed R squared or share of variance (square root; IQ Metrics calculation). Share = r squared x 100 (IQ Metrics calculation). SES = whether the model behind the printed r adjusts for family circumstances. The CSV adds journal, DOI, intervals and licences. As of 11 November 2026.

Cox 2019 [17]2019UK Biobank44-8118,363Brain volumeSame-sample correlation.287.6%No–Total brain volume with a latent g factor, age- and sex-corrected
Nave 2018 [16]2018UK Biobankadults13,608Brain volumeSame-sample correlation.193.6%Adjusted–Preregistered; total brain volume with fluid intelligence; SES in the model, no unadjusted r printed in the abstract
Michel 2024 [23]2024ABCD Studyabout 1011,876Several measures togetherNot stated in the abstract.44 *19.0%No–Grey plus white matter structure, regularised structural equation models
Michel 2024 (grey) [23]2024ABCD Studyabout 1011,876Cortex thickness or areaNot stated in the abstract.39 *15.4%No–Grey matter volume, thickness and surface area together
Michel 2024 (white) [23]2024ABCD Studyabout 1011,876White matterNot stated in the abstract.35 *12.4%No–White matter volume, FA and MD
Wang 2025 [25]2025ABCD Study9-1010,754Several measures togetherTested on unseen people.5429.2%No–45 kinds of MRI stacked; environment (incl. family income) modelled separately (r = .49)
Sripada 2021 [20]2021ABCD Study9-105,937Resting connectivityTested on unseen people.42 *18.0%No–Leave-one-site-out; printed as cross-validated partial eta squared
Marek 2022 [21]2022ABCD Study9-103,604Resting connectivityTested on unseen people.3411.6%No–Largest replicable multivariate association in the paper (univariate maximum r = 0.14)
Marek 2026 (RSFC) [26]2026ABCD Study9-102,263Resting connectivityTested on unseen people.3613.0%Before and after.23Ridge regression trained on 2,316 children, tested on 2,263; SES = Child Opportunity Index social and economic domain
Marek 2026 (thickness) [26]2026ABCD Study9-102,263Cortex thickness or areaTested on unseen people.235.3%Before and after.13Cortical thickness, same design
Popp 2024 [24]2024Human Connectome Projectyoung adults1,030Several measures togetherTested on unseen people.256.2%No–Structural-functional coupling (diffusion plus resting fMRI), cross-validated
Lee 2019 [18]2019Human Connectome Project (siblings)young adults1,022Brain volumeSame-sample correlation.183.2%Adjusted–Within-family correlation (sibling differences), corrected for measurement error; a family design holds family income and upbringing constant
Sripada 2020 [19]2020Human Connectome Projectyoung adults967Task fMRITested on unseen people.5025.0%No–N-back 2-back vs 0-back contrast
Dubois 2018 [14]2018Human Connectome Projectyoung adults884Resting connectivityTested on unseen people.45 *20.0%No–Cross-validated; controlled for gender, age and brain volume
Tetereva 2022 [22]2022Human Connectome Projectyoung adults873Several measures togetherTested on unseen people.5732.5%No–Stacked model of task fMRI, structural MRI and resting connectivity
Tetereva 2022 (tasks) [22]2022Human Connectome Projectyoung adults873Task fMRITested on unseen people.5631.4%No–Seven task fMRI scans stacked
Alnaes 2018 [15]2018Philadelphia Neurodevelopmental Cohort8-22729White matterTested on unseen people.3915.2%No–Diffusion MRI, 10-fold cross-validation; general cognitive factor
Ritchie 2015 [12]2015Lothian Birth Cohort 1936about 73672Several measures togetherSame-sample correlation.42 *18.0%No–Six structural measures together; the paper prints 18-21%: the lower end is used
Popp 2024 (new sample) [24]2024Independent sample (not named in the abstract)adults567Several measures togetherTested on unseen people.193.6%No–Same model applied to a new sample
Hearne 2016 [13]2016Human Connectome Projectyoung adults317Resting connectivitySame-sample correlation.3814.4%No–Mean connectivity of the edges the same study picked; in-sample by design
Choi 2008 [11]2008Own sample (young adults)young adults225Several measures togetherTested on unseen people.71 *50.0%No–Cortical thickness plus task fMRI during reasoning; regions chosen in the same study
Ganjavi 2011 [9]2011NIH MRI Study of Normal Brain Development6-18198White matterSame-sample correlation-.152.2%No–Corpus callosum midsagittal area, after correcting for total brain volume and age; the only negative link in the ledger
Shenkin 2009 [10]2009Own sample (community volunteers, Scotland)75-81107Brain volumeSame-sample correlation.214.4%No–Whole brain volume with a g factor; falls to .03 after correcting for intracranial area

Source: Our ledger [1] [2] [3]

Table 2

Reviews and meta-analyses (kept apart, never pooled with the ledger)

These pool other studies, so they are not rows. Two well-known meta-analyses are listed without a number because their abstracts are withheld by the publisher in every index we can read and we could not open their texts.

ReviewYearWhat it poolsStudiesPeopleResultNote
Pietschnig et al. 2015 [27]2015Meta-analysis of brain volume and IQ88more than 8,000r = .24Pools healthy and clinical samples; reports that published effects declined over time and were overestimated
Vieira et al. 2022 [28]2022Systematic review of predicting intelligence from MRI with machine learning30–r = .42 (95% CI .35 to .50)Out-of-sample prediction studies only; search to 15 December 2020; fluid intelligence from a single test r = 0.15 (0.13-0.17)
McDaniel 2005 [29]2005Meta-analysis of in vivo brain volume and intelligence––Not read: no numberAbstract withheld by the publisher in OpenAlex and Semantic Scholar; full text not open: no number printed here
Gignac and Bates 2017 [30]2017Meta-analysis of brain volume and intelligence by test quality––Not read: no numberAbstract withheld by the publisher; the author copy refused our request (HTTP 403): no number printed here

Source: Each review as cited; read by IQ Metrics

Table 3

Records we read and left out (56) and why

Reasons follow data/INCLUSION-RULE.md, written before screening. Leaving a paper out says nothing about its quality: most were left out because the abstract prints no usable r, or because the sample is a patient group, which this ledger does not cover.

Haier 1988Cortical glucose metabolic rate correlates of abstract reasoning and attention studied with positron emission tomographyPET, 1988: history only (timeline)
Willerman 1991In vivo brain size and intelligenceClassic small MRI study: no abstract online, n and r could not be read
Andreasen 1993Intelligence and brain structure in normal individualsClassic small MRI study (n = 67): the abstract prints no r
Posthuma 2003Genetic Correlations Between Brain Volumes and the WAIS-III Dimensions of Verbal Comprehension, Working Memory, Perceptual Organization, andScore is one WAIS-III dimension (working memory), not a general score
Peng 2004Diffusion tensor images in children with early-treated, chronic, malignant phenylketonuric: correlation with intelligence assessmentPatient group (phenylketonuria)
Witelson 2005Intelligence and brain size in 100 postmortem brains: sex, lateralization and age factorsPost-mortem brain weight, not MRI
Garde 2005Decline in intelligence is associated with progression in white matter hyperintensity volumeOutcome is decline in intelligence over time in older adults, not a score level
Hasan 2008Quantitative diffusion tensor imaging and intellectual outcomes in spina bifidaPatient group (spina bifida)
Rhein 2014Neuroanatomical Correlates of Intelligence in Healthy Young Adults: The Role of Basal Ganglia VolumeOnly sex-split r for parts of a test; no r for a general score
Bódizs 2014Sleep spindling and fluid intelligence across adolescent development: sex mattersEEG (sleep spindles), not MRI
Ohtani 2014Medial Frontal White and Gray Matter Contributions to General Intelligencen = 26: below the n >= 100 cut and not a classic named by the volume meta-analyses
Hai-b 2014Preliminary study of high resolution MR structural imaging of cerebral cortex in infants and correlated with intelligence developmentInfants and young children; no general intelligence score in the abstract
Nestor 2015Attentional Control and Intelligence: MRI Orbital Frontal Gray Matter and Neuropsychological CorrelatesAttentional-control tasks and regional grey matter; no r for a general score with n printed
Hernández 2015Brain iron deposits and lifespan cognitive abilityBrain iron in older people with lifespan cognitive change; not a general score link in a healthy sample
Luciano 2015Structural Brain MRI Trait Polygenic Score Prediction of Cognitive AbilitiesPolygenic scores of brain traits, not a brain measurement
Finn 2015Functional connectome fingerprinting: identifying individuals using patterns of brain connectivityNo intelligence r in the abstract (fingerprinting paper; fluid intelligence result not read)
Basten 2015Where smart brains are different: A quantitative meta-analysis of functional and structural brain imaging studies on intelligenceCoordinate meta-analysis of brain regions; no r
Fu 2016Diffusion tensor imaging study in Duchenne muscular dystrophyPatient group (Duchenne muscular dystrophy)
Greene 2018Task-induced brain state manipulation improves prediction of individual traitsPrints only bounds (over 20% and under 6%) for variance explained, no point estimate
Dubois 2018Resting-State Functional Brain Connectivity Best Predicts the Personality Dimension of Openness to ExperienceOutcome is personality (openness), not intelligence
Shen 2018Resting-State Connectivity and Its Association With Cognitive Performance, Educational Attainment, and Household Income in the UK BiobankEffect printed as standardised betas of single connections, not a correlation
Parham 2019Prediction of General Intelligence Using DTI Datan not printed in the abstract; journal could not be checked
Lett 2019Cortical Surfaces Mediate the Relationship Between Polygenic Scores for Intelligence and General IntelligenceVariance explained is for polygenic scores, not for a brain measure
Schmitt 2019The genetics of cortical myelination in young adults and its relationships to cerebral surface area, cortical thickness, and intelligence: AGenetics of cortical myelin; no brain-intelligence r
Kremen 2019Influence of young adult cognitive ability and additional education on later-life cognitionEducation and later-life cognition; no brain-intelligence r
He 2019Deep neural networks and kernel regression achieve comparable accuracies for functional connectivity prediction of behavior and demographicsAbstract prints no r for fluid intelligence
Mihalik 2019ABCD Neurocognitive Prediction Challenge 2019: Predicting Individual Fluid Intelligence Scores from Structural MRI Using Probabilistic SegmeChallenge proceedings: no abstract we could read
Elliott 2019General functional connectivity: Shared features of resting-state and task fMRI drive reliable and heritable individual differences in functAbstract prints no r
Jiang 2019Gender Differences in Connectome-based Predictions of Individualized Intelligence Quotient and Sub-domain ScoresAbstract prints no r
Frith 2020Intelligence and creativity share a common cognitive and neural basisPrinted r is between intelligence and creativity, not brain and intelligence
Vanderwal 2020Stability and similarity of the pediatric connectome as developmental measuresConnectome stability as a developmental measure; no general intelligence r with n
Hilger 2020Predicting intelligence from brain gray matter volumeAbstract prints the error in IQ points, not r
Judd 2020Cognitive and brain development is independently influenced by socioeconomic status and polygenic scores for educational attainmentOutcome is working memory, not a general score
Serin 2021NBS-Predict: A prediction-based extension of the network-based statisticMethods paper (also a schizophrenia sample); n for the intelligence result not in the abstract
Rasero 2021Integrating across neuroimaging modalities boosts prediction accuracy of cognitive abilityAbstract prints only the gain in variance explained (1% to 4%), not the prediction r
Wallert 2021Cognitive dedifferentiation as a function of cognitive impairment in the ADNI and MemClin cohortsCognitive impairment cohorts (ADNI)
Martins 2021Gray matter correlates of reading fluency deficits: SES matters, IQ does notOutcome is reading fluency, not intelligence
Dadi 2021Population modeling with machine learning can enhance measures of mental healthAbstract prints no brain-only r
Feilong 2021The neural basis of intelligence in fine-grained cortical topographiesPrints a range across 360 brain regions, no single whole-brain r
Tian 2021Machine learning prediction of cognition from functional connectivity: Are feature weights reliable?Prints a range (r = 0.2-0.4) across models, no single r
Thiele 2021Multitask brain network reconfiguration is inversely associated with human intelligenceAbstract prints no r
Wu 2022A multicohort geometric deep learning study of age dependent cortical and subcortical morphologic interactions for fluid intelligence predicDeep-learning prediction across cohorts; the printed R is not for a general intelligence score with n
Muri 2022Cortical thickness and its relationship to cognitive performance and metabolic control in adults with phenylketonuriaPatient group (phenylketonuria)
Cecelja 2022Aortic Distensibility Measured by Automated Analysis of Magnetic Resonance Imaging Predicts Adverse Cardiovascular Events in UK BiobankAortic MRI, not the brain
Wei 2022Functional connectome of vicarious traumatization: detecting and helping individuals vulnerable to mental illnessOutcome is vicarious traumatization, not intelligence
Tong 2022Transdiagnostic connectome signatures from resting-state fMRI predict individual-level intellectual capacityTransdiagnostic psychiatric sample
Ooi 2022Comparison of individualized behavioral predictions across anatomical, diffusion and functional connectivity MRIAbstract prints no r for cognition and no open full text was found
Kristanto 2022What do neuroanatomical networks reveal about the ontology of human cognitive abilities?Abstract prints no r
Kirkegaard 2023A Multimodal MRI-based Predictor of Intelligence and Its Relation to Race/EthnicityRace and ethnicity comparison: outside this ledger (no group comparisons of intelligence)
Voldsbekk 2023Shared pattern of impaired social communication and cognitive ability in the youth brain across diagnostic boundariesSample mixes diagnostic groups; social communication focus
Thiele 2023Multimodal Brain Signal Complexity Predicts Human IntelligenceEEG, not MRI
Fürtjes 2023General dimensions of human brain morphometry inferred from genome‐wide association dataGenome-wide data on brain morphometry; no brain-intelligence r
Joseph 2023Greater maltreatment severity is associated with smaller brain volume with implication for intellectual ability in young childrenSample selected for maltreatment
Hofmann 2025The utility of explainable AI for MRI analysis: Relating model predictions to neuroimaging features of the aging brainBrain-age model; the printed r is not with intelligence
Molloy 2025Regional, but not brain-wide, graph theoretic measures are robustly and reproducibly linked to general cognitive abilityWhole-brain graph measures showed no meaningful link; abstract prints no r and no open full text found (text only)
Parham 2026Structural Connectivity of Brain Regions May Predict Human IntelligenceNo DOI and no journal in the record; could not be checked

Source: OpenAlex [1]; screening by IQ Metrics

13 How we built the dataset

  1. 01

    Rule first

    We wrote the inclusion rule (data/INCLUSION-RULE.md) on 11 November 2026 before screening: healthy samples, MRI, a general intelligence score, a printed r or R squared and a printed n; bands and brain-measure classes fixed in advance.

  2. 02

    Search

    12 OpenAlex title and abstract searches (2,208 unique records) and 27 named leads from the reviews, the brief and the news hooks. OpenAlex metadata is CC0 [1].

  3. 03

    Screen

    Scripted word rules kept 51 records that print a correlation; we read 78 records by hand (abstracts, plus the open full text in Europe PMC where the abstract printed no usable r [2]). Every decision and reason is in data/screen_decisions.tsv.

  4. 04

    Check

    Each typed value has a quote; a script found every quote in the cached abstract or text and every n, r, R squared and interval inside a quote (248 checks), and confirmed that no paper is flagged retracted in OpenAlex.

  5. 05

    Compute

    r from R squared (square root), bands, medians, quartiles, the share of score differences (r squared), Fisher-z intervals and rank correlations: all IQ Metrics calculations, scripted in data/analyse.py.

Read every number with these in mind

  • The ledger is a lower bound: 23 results whose abstract or open text prints both n and r. Many studies print results only in figures or paywalled tables; “we found none” means none in what we could read.
  • No result has fewer than 100 people, so the ledger cannot show the inflated small-study effects that Marek and colleagues demonstrate by resampling; those come from their paper, not from our rows.
  • Results are not independent: several come from the same HCP or ABCD participants, and three from one ABCD paper.
  • Methods differ (same-sample correlations, cross-validation, held-out sets, structural equation models), so medians mix kinds of evidence; we show every result rather than pool them.
  • 7 results are converted from a printed share of variance (r = square root); one uses the low end of a printed range (18-21%); one is a within-family correlation corrected for measurement error.
  • Intervals other than Cox 2019 are Fisher-z intervals from r and n; for prediction studies the true uncertainty is larger.
  • Patient groups are out by design. Nothing here is medical advice, and no result describes a group’s or a person’s intelligence.

Sources

  1. OurResearch. OpenAlex works API: 12 title and abstract searches (2,208 unique records) and 37 named-lead look-ups; metadata, abstracts, citation counts and retraction flags. 2026. openalex.orgCC0: "Every record is released under a CC0 public-domain license" (help.openalex.org/hc/en-us/articles/24397762024087-Pricing); abstracts read by script and cited as facts
  2. EMBL-EBI. Europe PMC REST API: open-access full texts of four papers and one abstract. 2026. europepmc.org/RestfulWebServiceEach text keeps its article licence; facts cited only
  3. Allen Institute for AI. Semantic Scholar Graph API: abstract look-ups for four papers whose abstracts OpenAlex does not carry. 2026. api.semanticscholar.orgAPI terms; abstracts cited as facts only
  4. Cerebral Cortex. Molloy and Sripada (2025). Regional, but not brain-wide, graph theoretic measures are robustly and reproducibly linked to general cognitive ability. Cerebral Cortex (ABCD n = 5,937; HCP n = 847). 2025. doi.org/10.1093/cercor/bhaf074Publisher copyright: facts cited only (n, r, ages, dataset)
  5. Intelligence. Haier et al. (1988). Cortical glucose metabolic rate correlates of abstract reasoning and attention studied with positron emission tomography. Intelligence (history only). 1988. doi.org/10.1016/0160-2896(88)90016-5Publisher copyright: facts cited only (n, r, ages, dataset)
  6. bioRxiv. Vieira et al. (2021). Preprint of the systematic review (bioRxiv, posted 28 October 2021): search date and methods read from the text. 2021. doi.org/10.1101/2021.10.19.462649CC BY-ND 4.0 (preprint): facts cited only
  7. Brain Structure and Function. Hilger et al. (2020). Predicting intelligence from brain gray matter volume. Brain Structure and Function (n = 308; abstract read, no r printed). 2020. doi.org/10.1007/s00429-020-02113-7CC BY (article licence as recorded by OpenAlex); facts cited
  8. American Journal of Psychiatry. Andreasen et al. (1993). Intelligence and brain structure in normal individuals. American Journal of Psychiatry (67 volunteers; abstract prints no r). 1993. doi.org/10.1176/ajp.150.1.130Publisher copyright: facts cited only (n, r, ages, dataset)
  9. PLoS ONE. Ganjavi et al. (2011). PLoS ONE (NIH MRI Study of Normal Brain Development, n = 198). 2011. doi.org/10.1371/journal.pone.0019698CC BY (article licence as recorded by OpenAlex); facts cited
  10. BMC Geriatrics. Shenkin et al. (2009). BMC Geriatrics (Own sample (community volunteers, Scotland), n = 107). 2009. doi.org/10.1186/1471-2318-9-12CC BY (article licence as recorded by OpenAlex); facts cited
  11. Journal of Neuroscience. Choi et al. (2008). Journal of Neuroscience (Own sample (young adults), n = 225). 2008. doi.org/10.1523/jneurosci.3259-08.2008CC BY-NC-SA article licence: facts cited only, nothing reproduced
  12. Intelligence. Ritchie et al. (2015). Intelligence (Lothian Birth Cohort 1936, n = 672). 2015. doi.org/10.1016/j.intell.2015.05.001CC BY (article licence as recorded by OpenAlex); facts cited
  13. Scientific Reports. Hearne et al. (2016). Scientific Reports (Human Connectome Project, n = 317). 2016. doi.org/10.1038/srep32328CC BY (article licence as recorded by OpenAlex); facts cited
  14. Philosophical Transactions of the Royal Society B Biological Sciences. Dubois et al. (2018). Philosophical Transactions of the Royal Society B Biological Sciences (Human Connectome Project, n = 884). 2018. doi.org/10.1098/rstb.2017.0284Publisher copyright: facts cited only (n, r, ages, dataset)
  15. JAMA Psychiatry. Alnæs et al. (2018). JAMA Psychiatry (Philadelphia Neurodevelopmental Cohort, n = 729). 2018. doi.org/10.1001/jamapsychiatry.2017.4277Publisher copyright: facts cited only (n, r, ages, dataset)
  16. Psychological Science. Nave et al. (2018). Psychological Science (UK Biobank, n = 13,608). 2018. doi.org/10.1177/0956797618808470Publisher open access (other licence): facts cited only
  17. Intelligence. Cox et al. (2019). Intelligence (UK Biobank, n = 18,363). 2019. doi.org/10.1016/j.intell.2019.101376CC BY (article licence as recorded by OpenAlex); facts cited
  18. Intelligence. Lee et al. (2019). Intelligence (Human Connectome Project (siblings), n = 1,022). 2019. doi.org/10.1016/j.intell.2019.01.011Publisher copyright: facts cited only (n, r, ages, dataset)
  19. Human Brain Mapping. Sripada et al. (2020). Human Brain Mapping (Human Connectome Project, n = 967). 2020. doi.org/10.1002/hbm.25007CC BY (article licence as recorded by OpenAlex); facts cited
  20. Translational Psychiatry. Sripada et al. (2021). Translational Psychiatry (ABCD Study, n = 5,937). 2021. doi.org/10.1038/s41398-021-01704-0CC BY (article licence as recorded by OpenAlex); facts cited
  21. Nature. Marek et al. (2022). Nature (ABCD Study, n = 3,604). 2022. doi.org/10.1038/s41586-022-04492-9CC BY (article licence as recorded by OpenAlex); facts cited
  22. NeuroImage. Tetereva et al. (2022). NeuroImage (Human Connectome Project, n = 873). 2022. doi.org/10.1016/j.neuroimage.2022.119588CC BY-NC-ND article licence: facts cited only, nothing reproduced
  23. Journal of Neuroscience. Michel et al. (2024). Journal of Neuroscience (ABCD Study, n = 11,876). 2024. doi.org/10.1523/jneurosci.0465-23.2023CC BY (article licence as recorded by OpenAlex); facts cited
  24. NeuroImage. Popp et al. (2024). NeuroImage (Human Connectome Project, n = 1,030). 2024. doi.org/10.1016/j.neuroimage.2024.120563CC BY-NC-ND article licence: facts cited only, nothing reproduced
  25. eLife. Wang et al. (2025). eLife (ABCD Study, n = 10,754). 2025. doi.org/10.7554/elife.105537.3CC BY (article licence as recorded by OpenAlex); facts cited
  26. Science. Marek et al. (2026). Science (ABCD Study, n = 2,263). 2026. doi.org/10.1126/science.aee6213CC BY-NC article licence: facts cited only, nothing reproduced
  27. Neuroscience & Biobehavioral Reviews. Pietschnig et al. 2015. Meta-analysis of brain volume and IQ. Neuroscience & Biobehavioral Reviews. 2015. doi.org/10.1016/j.neubiorev.2015.09.017Publisher copyright: facts cited only
  28. Intelligence. Vieira et al. 2022. Systematic review of predicting intelligence from MRI with machine learning. Intelligence. 2022. doi.org/10.1016/j.intell.2022.101654Publisher copyright (journal version); facts cited only
  29. Intelligence. McDaniel 2005. Meta-analysis of in vivo brain volume and intelligence. Intelligence. 2005. doi.org/10.1016/j.intell.2004.11.005Publisher copyright: facts cited only
  30. Intelligence. Gignac and Bates 2017. Meta-analysis of brain volume and intelligence by test quality. Intelligence. 2017. doi.org/10.1016/j.intell.2017.06.004Publisher copyright: facts cited only

Version history

  • 2026.1111 November 2026First release. Next update due May 2027 (searches re-run every six months; ABCD and UK Biobank papers appear monthly), sooner if a study in the ledger is corrected or retracted, or if a new study prints the brain-IQ link before and after family-income adjustment.
Leon Falkner

Leon Falkner, Data & Research Analyst. Leon owns the arithmetic on IQ Metrics: the percentile tables, the score converter and the figures in everyone else's drafts. On the research desk Leon pulls the data, builds each index and writes the method. More about Leon.

Questions

Can a brain scan show your IQ?

No. In the 12 largest published results (1,022 to 18,363 people), brain MRI links to intelligence scores at a median r = .35, about 12% of the differences between people. That describes groups; a scan cannot place one person on the IQ scale. IQ is measured with a test.

Does brain size correlate to intelligence?

Weakly. Total brain volume correlates with intelligence scores at r = .18 to .275 in the three largest volume studies (1,022 to 18,363 people), about 4% of score differences. A 2015 meta-analysis of 88 studies found r = .24 and evidence that published effects were inflated.

Can an MRI measure intelligence?

Not for an individual. MRI models tested on people they were not built on predict intelligence scores at a median r = .42 across 13 results, so errors of ten IQ points or more are normal. One study of 308 adults found predictions clustered around the average.

Does brain size affect IQ?

A larger brain goes with a slightly higher score on average, and the link holds between siblings in 1,022 young adults (r = .18 after correcting for measurement error). The authors read their genetic analyses as pointing to a causal effect; a correlation alone cannot show cause.

Does family income explain the link between the brain and IQ?

Partly. In a 2026 Science study of 4,579 ABCD children, adjusting for neighbourhood social and economic opportunity cut the brain-IQ link by 30 to 40%; family income and parental education weakened it less. A smaller link remained (r = .23 and .13).

What did the 2026 study of 12,000 children find?

Marek and colleagues started from 11,878 ABCD children and analysed 4,579 with brain scans. Socioeconomic measures had the strongest brain links of 649 measures, and adjusting for neighbourhood opportunity reduced the brain-IQ prediction from r = .36 to .23. It did not reduce it to zero.

Which part of the brain is linked to intelligence?

No single part. Studies find links spread across the frontal, parietal and temporal cortex, white matter tracts and connections between networks; a 2025 study of 5,937 children found whole-brain efficiency measures unrelated to ability while single-region measures were related. Task and combined scans report the strongest links (median r = .53 and .44).

Are bigger brains smarter?

On average, very slightly: total brain volume and fluid intelligence correlate at r = .19 in a preregistered study of 13,608 UK Biobank adults, about 4% of score differences. Many people with smaller brains score higher than people with larger ones.

How much of IQ can a brain scan explain?

About 12% of the differences in test scores in the biggest studies (median r = .35), from about 4% for brain volume to 29% for the best combined model. The rest reflects what scans miss, test error and life circumstances.

Why do small brain-imaging studies report stronger results?

Chance swings are larger in small samples, and studies with striking results are more likely to be published. Marek and colleagues showed in the ABCD Study that brain-behaviour links became stable only with about 2,000 people; our ledger has no result under 100 people.

Use the data

Share it, cite it, download it

You may reuse everything on this page under CC BY 4.0: in articles, reports, lessons and posts. The one condition is credit to IQ Metrics with a link to this page.

Cite this study

Version 2026.11. Citing the version lets readers find the exact numbers.

Archived copy with a DOI: 10.5281/zenodo.24126068 (report PDF and data on Zenodo).

Falkner, L. (2026). Can a brain scan show your IQ? Every MRI study, sample size against result (Version 2026.11) [Data set]. IQ Metrics Research. https://iqmetrics.org/research/can-a-brain-scan-show-your-iq/ (DOI: 10.5281/zenodo.24126068)

Embed the key numbers

Paste this into an article or post. It lists the main numbers and credits the source.

Get the data

One row per published result (23 rows, 18 papers): an MRI brain measure related to a general intelligence score in a healthy sample. Sample size, dataset, ages, brain-measure class, analysis type, r as printed or converted from R squared, the share of score differences (r squared), a 95% interval, whether family circumstances were adjusted for and the r after adjustment where printed. Empty = not printed in the text we could read.

Download CSV
Data dictionary
ColumnWhat it holds
row_id, study, first_author, year, journal, doiOur id, first author and year, publication year (OpenAlex), journal and DOI
dataset, ages, n, n_noteSample (UK Biobank, ABCD, HCP, cohort or own sample), ages as printed, people behind the printed r, and what that n counts
sample_size_band<100, 100-999 or 1,000+ (fixed before computing)
modalityVolume, Cortex (thickness, area), White matter, Resting FC, Task fMRI or Multimodal
analysisin-sample (same-sample correlation), out-of-sample (cross-validation or held-out sample) or not stated
r_printed, r_squared_printedThe effect as printed: a correlation, or a share of variance explained
r_used, r_from_r_squared_iqm_calcr used on the page; yes = square root of the printed share (IQ Metrics calculation)
share_of_score_differences_pct_iqm_calcr squared x 100 (IQ Metrics calculation)
ci_low, ci_high, ci_source95% interval: printed, or Fisher z from r and n (IQ Metrics calculation)
ses_controlled, r_after_ses_controlWhether the model adjusts for family income, SES, parental education, neighbourhood or uses a within-family design; r after adjustment where both are printed
note, paper_licence_openalexOur note; the article licence recorded by OpenAlex (facts only are cited)

Share this finding

A finished image and a ready message. The image carries the source line, so the credit travels with it.

Each network gets a version sized for it unless you edit the message.

Press kit

For journalists, researchers and teachers

Everything on this page may be quoted, charted and republished with credit. No embargo, no permission needed, no sign-up. If you need a number that is not here, ask.

Summary you can paste

Plain news style. Edit freely; keep the credit.

IQ Metrics Research built a ledger of 23 published results (18 papers, 2008-2026) relating brain MRI to a general intelligence score in healthy samples, with sample size, method and whether family circumstances were taken into account. In the 12 results with 1,000 or more people, the median correlation is r = .35: the scan accounts for about 12% of the differences in scores, total brain volume about 4%. Only 4 results adjust for family circumstances; in the one study that prints both, adjusting for neighbourhood opportunity cut the link by 30 to 40% (the paper's figure). Source: IQ Metrics Research, https://iqmetrics.org/research/can-a-brain-scan-show-your-iq/

Angles for a story

  • A scan accounts for about 12% of score differences in the biggest studies (Figure 12)
  • The 2026 ABCD study: neighbourhood opportunity cut the brain-IQ link by 30 to 40%, not to zero (Figure 10)
  • Brain size and intelligence: r = .18 to .275 in samples of 1,022 to 18,363 (the section)
  • Only 4 of 23 results adjust for family income or background (Figure 11)
  • Sample size against result, every study on one chart (Figure 3)

Quotes you can use

Attributable as written, with the name and title. For a different angle or a quote on a specific figure, ask through the contact page.

  • A correlation of about .35 between a scan and a test score is real and it replicates. It is also far too weak to read anyone’s IQ off a brain image, and nobody in these studies claims it is.

    Leon Falkner, Data & Research Analyst, IQ Metrics Research
  • Only four of the twenty-three results take family circumstances into account. The one study that shows the number both ways finds the link drops by a third or more.

    Leon Falkner, Data & Research Analyst, IQ Metrics Research
  • Brain size is the oldest idea in this field and it explains the least: about four per cent of the differences in scores in the largest samples.

    Leon Falkner, Data & Research Analyst, IQ Metrics Research

Files

  • CSV of every result: study, year, journal, DOI, dataset, ages, n, brain measure, analysis, r as printed, r used, share of score differences, interval, SES flag, r after SES, licence (CC BY 4.0)
  • Every figure as a print-ready PNG: use the Image button under each figure
  • Embed code and data table under every figure
  • Share images for social media in three sizes, with the credit line built in
  • Citations in APA, MLA, Chicago, BibTeX and plain text

Terms, contact and corrections

Credit: “IQ Metrics Research, Can a brain scan show your IQ? (2026)” with a link to this page. On air or in print, “IQ Metrics” and the page address are enough.

What it covers: published MRI studies of healthy samples that print a sample size and a correlation (or R squared) with a general intelligence score; reviews are listed apart; patient groups are out of scope.

Contact: the research team, through the contact page. Corrections are made on this page with a dated note in the version history.

Keep reading

Related research