The Big Five: Five Dimensions Recovered from Language, and the Boundary Nobody Has Settled

The idea underneath the Big Five is older than psychology's use of it, and it begins with a count. Francis Galton, in an essay called Measurement of Character published in the Fortnightly Review in 1884, reasoned that the number of conspicuous aspects of human character could be gauged by counting the words a language keeps for describing them. He went through an appropriate dictionary, Roget's Thesaurus, sampled many pages of its index, and estimated that it contained fully one thousand words expressive of character, each with a separate shade of meaning. That casual count is the origin of the lexical tradition, and the lexical tradition is what eventually produces five factors. Galton also founded eugenics and coined the word, which is not a detail that can be dropped from a sentence putting him at the head of a story about measuring people. Neither the essay nor Galton appears in our own research file; both are added here.

The first systematic version of Galton's count is Allport and Odbert's, in 1936. They worked through an unabridged English dictionary, sorted the person-describing words into columns, and the column they judged to hold stable traits held about 4,500 terms. That number is not a discovery about people. It is a discovery about English, and about two men's judgment of which words name something durable. Everything downstream in this article inherits both of those facts.

Cattell took the lexical list and compressed it. Working across roughly 1943 to 1957 he applied cluster analysis and arrived at 16 primary factors, measured by the questionnaire that still carries the name, the 16PF. Our own file states the outcome in the same sentence as the achievement, and the outcome is the whole reason the Big Five exists: the sixteen-factor structure proved difficult to replicate across laboratories. Goldberg's 1990 paper reviews that failure and the convergence on five from inside the tradition; Block's 1995 critique covers the same history from the other side. Five factors are not five because five is the natural number of things a person is. Five is what survived when different laboratories, different samples and different item sets were asked to produce the same answer.

The five were first recovered in a place academic psychology was not reading. Tupes and Christal reported five recurrent factors in a United States Air Force technical report in 1961, Recurrent Personality Factors Based on Trait Ratings. It sat in the Air Force technical literature, and the Journal of Personality reprinted it in full only in 1992, by which point the structure it described had been rediscovered several times over. Our own file's account of the discovery starts with Norman in 1963 and does not mention them, which hands Norman a priority that is not his: Norman was replicating Tupes and Christal. The five dimensions were first set out in a military technical report, filed, and left largely unread for three decades.

Norman, in 1963, ran factor analyses of trait-adjective ratings and reported a replicated five-factor structure in peer-nomination ratings. Goldberg, in 1990, arrived at the same five from a different direction and gave the set the name that stuck. Both papers are in our file and both are correctly cited there. The correction above is about what precedes them, not about them.

The instrument our file calls the gold standard is the Revised NEO Personality Inventory, developed by Costa and McCrae and published in 1992: 240 items resolving to 30 facets, six of them under each of the five domains. The facet layer is where the model stops being abstract, and it is the answer to the objection most readers reach for first, that five dimensions are far too coarse to describe anybody. They are. The facets are what the coarse dimensions are made of, and two people with identical Extraversion scores can be assembled from quite different facets underneath. The inventory is a copyrighted commercial instrument, so no item, scoring key or profile sheet from it appears anywhere on this page.

This is where an article about the Big Five is likeliest to overstate, and where our own file does. T_1_08 reports that the five-factor structure replicated in more than fifty cultures across six continents using translated versions of the inventory, with factor congruence coefficients above 0.90 for all five factors in most of them, and attaches the whole sentence to McCrae and Costa's 1997 paper in American Psychologist. That paper analysed six translations of the Revised NEO Personality Inventory: German, Portuguese, Hebrew, Chinese, Korean and Japanese, drawn from five distinct language families, with 7,134 participants in total, and it reported that the American five-factor structure was closely reproduced in each. Six translations is a real and interesting result. It is not a census of fifty cultures, and the paper's own move is an inference toward universality from six diverse cases rather than a count of anything. The paper is titled Personality Trait Structure as a Human Universal, and a title is an argument being made, not a finding being reported.

A fifty-culture study does exist and it is not the one our file cites. McCrae, Terracciano and the members of the Personality Profiles of Cultures Project published Universal Features of Personality Traits From the Observer's Perspective: Data From 50 Cultures in the Journal of Personality and Social Psychology in 2005, eight years after the paper our file credits. Swapping it in changes two things at once, and both matter. The scale goes up. And the method changes from what people say about themselves to what other people say about them, which is a different kind of evidence, not a bigger helping of the same kind.

Our file also contains the strongest thing standing against its own universality claim, tucked into a subordinate clause. Saucier and Goldberg reviewed lexical studies in German, Dutch, Hungarian, Italian, Czech, Filipino, Hebrew, Turkish, Korean and Chinese in the Journal of Personality in 2001, and found that they generally recover factors resembling the Big Five, though some languages yield six or seven factors rather than five. Eight years later Saucier went further, concluding from inclusive lexical studies that six recurrent dimensions fit better than five. If the number of factors a language yields depends on the language, then five is not a fact about people. It is a fact about a particular family of samples analysed a particular way. That is not a debunking, it is the correct reading of what a factor solution is, and section 08 makes it explicit.

Two published tests ran outside those samples and both came back negative, and neither is in our file. The first is about the model, the second about the instruments. Gurven, von Rueden, Massenkoff, Kaplan and Lero Vie administered a Big Five instrument to the Tsimane, a forager-farmer population of the Bolivian Amazon, and reported in the Journal of Personality and Social Psychology in 2013 that the five-factor model did not hold on tests of internal consistency, response stability or external validity. Replication did not improve in a second Tsimane sample in which adults rated their spouses. What the data supported instead was a Tsimane Big Two: prosociality and industriousness. Traits the Big Five assigns to conscientiousness, being efficient, persevering and thorough, bundled together with being energetic, relaxed and helpful, cutting straight across the boundaries the model expects.

That result is itself contested, and the contest is unusually clean. Van der Linden, Dunkel, Figueredo, Gurven, von Rueden and Woodley of Menie re-analysed the same Tsimane data in the Journal of Cross-Cultural Psychology in 2018 and reached a different structural conclusion. Two of the 2013 paper's authors, including its first author, are on the re-analysis. One dataset, overlapping teams, two answers, both in print. That is what a live scientific question looks like from the outside, and carrying both is more honest than crowning either.

The larger test points the same way and states its conclusion very carefully. Laajaj and colleagues pooled 29 face-to-face surveys covering 94,751 respondents across 23 low- and middle-income countries and reported in Science Advances in 2019 that the commonly used Big Five question sets generally failed to measure the traits they are intended to measure, and did not pass standard validity tests in those samples. Read the conclusion as its authors wrote it. It is about the instruments failing, not about personality having a different shape in poorer countries. Collapsing that distinction produces a falsehood in either direction, and both directions are tempting.

One further cross-cultural finding belongs here before the summing up, and it is robust and runs against the obvious prediction. Costa, Terracciano and McCrae reported in 2001 that women score slightly higher than men on Agreeableness and Neuroticism across nearly all the cultures studied, and that the differences are larger in more gender-egalitarian societies rather than smaller. Schmitt, Realo, Voracek and Allik found the same paradox across 55 cultures in 2008, and Mac Giolla and Kajonius replicated it in 2018. Our own file then draws a conclusion from it: that this contradicts social role predictions and suggests biological contributions. The finding is solid. That reading is one of at least two. Heine, Lehman, Peng and Greenholtz described the reference-group effect in 2002: a person answering a Likert scale rates themselves against an implicit local comparison group, so a difference between national means partly measures who the respondents were comparing themselves to. If the implicit comparison group differs between a more and a less egalitarian society, a measured difference can widen without any underlying trait difference changing at all. This page carries the pattern as robust and the biological reading as one interpretation among at least two, which is less than our file claims and considerably more than nothing.

Refused: that the Big Five is a human universal, that the same five dimensions have been found in every culture on Earth. The structure replicates well and repeatedly in large, literate, industrialised, questionnaire-taking populations, which is a genuine and non-trivial achievement, and the refusal does not diminish it. What it has not been shown to do is hold cleanly everywhere. The Tsimane data did not support it. Twenty-three low- and middle-income countries, 29 surveys and 94,751 respondents did not pass standard validity tests on the usual instruments. Several lexical studies recover six or seven factors rather than five. And our own file's more-than-fifty-cultures figure is attached to a paper that studied six translations. Whether the failures reflect a limit on the structure or a limit on the questionnaire is a live and unsettled question, and this page leaves it unsettled.

Before any number: heritability is a population statistic, not a personal one. A heritability of 50 percent does not mean that half of any one person's conscientiousness is genetic. It means that, in the particular population sampled at the particular time it was sampled, about half of the observed variance between people is statistically associated with genetic variance between them. It implies nothing about whether a trait can change, and the estimate itself moves with the environment: our own file on psychometrics, T_1_10, records that heritability estimates run lower in deprived environments and higher in enriched ones. T_1_08 makes heritability a headline claim and carries no version of this caveat anywhere. It is supplied here because the single most common misreading of this entire subject is precisely the one it blocks.

Twin studies put the heritability of all five factors in a moderate band, roughly 40 to 60 percent. Our file then prints five specific per-trait percentages and attributes them to Vukasovic and Bratko's 2015 meta-analysis in Psychological Bulletin. That meta-analysis does not report them. What it reports, across 62 independent effect sizes and more than 100,000 participants, is an overall heritability estimate of about 40 percent and a significant moderator effect of study design: twin designs give estimates slightly below .50, while family and adoption designs give estimates slightly above .20. The design split is the paper's actual headline, it is missing from our file entirely, and it is far more interesting than the five numbers it replaces, because it says the answer more than doubles depending on which relatives you compare. The origin of those five percentages could not be identified for this page. A figure whose source cannot be named does not get set in type here.

The rest of the variance is not where most readers would put it. Shared environment, meaning everything siblings raised in the same household have in common, contributes relatively little to adult Big Five variance, roughly 0 to 11 percent. What remains beyond heritability is non-shared environment plus measurement error, and that last phrase is doing quiet work. A slice of what looks like environmental influence is simply the unreliability of the questionnaire: the same person answering the same instrument twice does not give the same answers, and that noise has to land somewhere in the accounting. Our own file's counter-argument table makes the identical point from the other side, setting about 50 percent non-shared environment plus error against the claim that personality is primarily genetic.

Molecular genetics is where the story gets strange. Personality traits are highly polygenic: thousands of variants, each of very small effect. Lo and colleagues, in Nature Genetics, published online in December 2016 and in print in 2017, identified six genomic loci reaching genome-wide significance for personality traits and showed correlations with psychiatric disorders. Our file states that the total variance explained by common variants remains small, at roughly 5 to 15 percent. That band is wrong in both directions at once, and the paper that shows why is not in our file.

Power and Pluess estimated Big Five heritability directly from common genetic variants in Translational Psychiatry in 2015, rather than inferring it from family resemblance. Neuroticism came out at 15 percent, with a standard error of 0.08 and P = 0.04, and Openness at 21 percent, P below 0.01; both reached statistical significance. Extraversion, Agreeableness and Conscientiousness did not, with point estimates of 0.08, below 0.001, and 0.01 respectively. A point estimate that fails to reach significance is not a small measured effect. It is a failure to distinguish the effect from zero. The authors' own comparison is that twin studies give around 40 to 60 percent of the variance in the Big Five while these direct estimates generally tend to be half of twin-study estimates, and they conclude that common variants account for about a quarter of the causal genetic variation. So twin designs say roughly half, direct measurement of common variants recovers a fraction of that, and for three of the five traits it recovers nothing that reaches significance at all. Neither figure is a lie. The gap between them is the finding, and nobody can fully explain it. What the field did in response was scale up: Nagel and colleagues published a meta-analysis of genome-wide association studies for neuroticism in 449,484 individuals in Nature Genetics in 2018, identifying novel loci and pathways. Six loci to nearly half a million people is what hunting variants of tiny effect actually looks like.

The biological story our file tells is Eysenck's. In The Biological Basis of Personality, published in 1967 by Charles C. Thomas of Springfield, Illinois, Eysenck proposed that extraversion is linked to cortical arousal, with introverts chronically more aroused and therefore seeking less stimulation, and neuroticism to the reactivity of the limbic system. Our file's wording for the modern evidence is careful and worth preserving exactly: partially supported by functional imaging studies showing that amygdala reactivity correlates with neuroticism. Two things need saying that our file does not say. The first is about the man. In May 2019 an enquiry at King's College London into publications Eysenck co-authored with Grossarth-Maticek concluded that 26 of them were unsafe and recommended retraction; King's made the finding public later that year, and journals subsequently retracted a set of papers and issued expressions of concern on dozens more, some of them sixty years old. That enquiry was scoped to the Grossarth-Maticek collaboration on personality and disease and deliberately excluded Eysenck's sole-authored work, and the 1967 book is not among the retracted or flagged publications. So the arousal theory is not retracted, and Eysenck's record carries the largest research-integrity finding in the history of British psychology. Both of those statements are true, and this page carries both rather than choosing one.

The second thing is about the brain, and it is a caution rather than a refutation. Avinun, Israel, Knodt and Hariri looked for associations between the Big Five traits and variability in brain grey or white matter, in NeuroImage in 2020, in a sample far larger than the earlier personality-neuroscience studies that had reported such associations, and found little evidence for them. Note the scope carefully. That is brain structure, not brain function. It does not test Eysenck's cortical-arousal account, which is a functional claim, and it does not test the task-evoked amygdala finding our file cites. What it does is put down a marker: this is one of several places on this page where a small, striking early result shrank when a large sample looked again. A well-powered null is evidence, and it is labelled as evidence here.

Refused: that personality is about 50 percent genetic, so roughly half of who you are is fixed at birth. Every clause of that misreads the statistic. A heritability estimate is a population variance figure, not a share of any individual's traits. It changes by more than a factor of two depending on whether the design compares twins or adoptive families. It moves with the environment being sampled. Direct measurement of common genetic variants recovers only a fraction of it and reaches statistical significance for two of the five traits. And heritability says nothing whatever about malleability: mean levels of the traits shift systematically with age, and life events move them further, both of which are the subject of the next section.

Personality is stable, and the stability has been measured rather than assumed. Roberts and DelVecchio pooled 3,217 test-retest correlations from 152 longitudinal studies in Psychological Bulletin in 2000 and traced rank-order consistency across the lifespan: about .31 in childhood, .54 through the college years, .64 at about age 30, then a plateau near .74 between the ages of 50 and 70, with the test-retest interval held constant at 6.7 years across the whole curve. Two things make those numbers mean something rather than merely be numbers. Consistency rises with age and never reaches unity, so our file is right that the plaster metaphor, the idea that personality sets hard in early adulthood, overstates it. And that interval is not a technicality. A test-retest correlation without it is not a fact, because .74 over about seven years is a completely different statement from .74 over a lifetime. No restatement of the figure anywhere on this page travels without its interval.

Rank-order stability and mean-level change are different things and both can be high at once. Roberts, Walton and Viechtbauer pooled longitudinal studies in the same journal in 2006 and found that across the lifespan Agreeableness and Conscientiousness increase while Neuroticism decreases, with most of the movement concentrated in young adulthood, roughly ages 20 to 40; Openness increases in adolescence and declines slightly after midlife. Our file names the pattern the maturity principle: people become more socially mature with age. Everybody in a cohort can become more conscientious while keeping their position relative to everybody else, which is why these two findings do not contradict each other. Confusing them is the commonest error in popular writing about personality change, and getting it right is a real service to a reader. The mean-level finding also did not arrive uncontested: Costa and McCrae disputed it at the time and the authors' reply sits in the same volume, so this page does not present it as settled by acclamation.

Life events move personality too, and the size word is the whole claim. Bleidorn, Hopwood and Lucas reviewed life events and personality trait change in the Journal of Personality, online in 2016 and in print in 2018: a first job, marriage, divorce, widowhood and unemployment can each produce change beyond normative maturation, though the effects are typically small. Typically small is our own file's hedge and not a caution invented here, and it survives into every restatement on this page, because the compressed version of that sentence is divorce changes your personality, and that is not what the literature says.

Trait psychology's confidence was broken in 1968 by one book. Walter Mischel's Personality and Assessment reported that the cross-situational consistency of behaviour is typically around r = .30, a figure that became known as the personality coefficient, and argued that situations therefore matter more than traits. Two things about that number. It is not a trivially small correlation in behavioural science. And Mischel was not claiming that traits do not exist: he was claiming that a single behaviour in a single situation is poorly predicted by a trait score, which is true, and which improves sharply as soon as behaviour is aggregated across many situations.

The resolution most researchers now accept is interactionist: behaviour is jointly determined by personality traits and situational demands, and traits predict behaviour aggregated across many occasions rather than any single one. The better half of the story is that the man who broke trait psychology's confidence supplied the repair. Mischel and Shoda's Cognitive-Affective Processing System model, in Psychological Review in 1995, showed that individuals are consistent in their if-then situation-behaviour profiles: consistent in which situations produce which behaviour, even when the raw behaviour looks inconsistent. Consistency lives in the pattern, not in the average. That is the best single idea in this section and it is worth more than the coefficient it succeeded.

Funder stated the settlement plainly in the Journal of Research in Personality in 2006: the trait approach and the social-cognitive approach are complementary rather than competing. Traits describe what people typically do. Cognitive-affective units explain how they do it. Neither answers the other's question, and treating the pair as rivals was the error.

Our file's summary asserts significant predictive validity for life outcomes including health, longevity, occupational performance, relationship quality and psychopathology, and then supplies not one number for any of them anywhere in the document. Here is the shape of that literature. Barrick and Mount meta-analysed 117 studies in Personnel Psychology in 1991, across five occupational groups (professionals, police, managers, sales, and skilled and semi-skilled workers) and three criteria (job proficiency, training proficiency and personnel data). Conscientiousness was the only Big Five dimension showing consistent relations with all job performance criteria across all occupational groups. Extraversion was a valid predictor for the two occupations involving heavy social interaction, management and sales. For the remaining dimensions the estimated true-score correlations varied by occupational group and by criterion type. No coefficient from that paper appears on this page, and the reason is worth stating: three different figures for the same finding circulate in secondary sources, the difference between them is which corrections were applied, and the primary text was not read for this article.

Then the correction, which is recent and large. Sackett, Zhang, Berry and Lievens revisited meta-analytic estimates of validity in personnel selection in the Journal of Applied Psychology in 2022 and argued that all five standard approaches to correcting for range restriction produce substantial overcorrection. Their conclusion, in the published abstract's own words: most of the same selection procedures that ranked high in prior summaries remain high in rank, but with mean validity estimates reduced by .10 to .20 points; structured interviews emerged as the top-ranked selection procedure; and selection predictor-criterion relationships are considerably lower than previously thought. This is not a correction aimed at personality measures specifically. It hits cognitive ability tests the same way. What it means for this article is that a sentence of the form conscientiousness predicts job performance either carries a size word or explains why it does not, as the 1991 meta-analysis above does, and where a 2022 estimate exists, the size word is modest.

The best single answer to the replication question in this field points both ways at once, which is why it is the honest headline. The Life Outcomes of Personality Replication Project ran preregistered, high-powered replications of 78 previously published trait-outcome associations, with a median sample of 1,504, and reported in Psychological Science in 2019 that 87 percent of the attempts came out statistically significant in the expected direction, with the replication effects typically 77 percent as strong as the corresponding originals. Read both halves. Against the reflex that all of psychology collapsed in the replication crisis: these findings mostly held, at a rate most areas of the field would envy. Against overstatement: they held smaller, by about a quarter, which is the standard signature of publication bias in the original literature. One result, two disciplines.

Roberts, Kuncel, Shiner, Caspi and Goldberg compared personality traits against socioeconomic status and cognitive ability as predictors of mortality, divorce and occupational attainment in Perspectives on Psychological Science in 2007, and argued that the predictive power of personality traits is comparable in magnitude to that of the other two for several consequential life outcomes. The paper is called The Power of Personality, and the title will do the overstating for a reader who is not careful. What it says is that personality predicts about as well as socioeconomic status and cognitive ability do. The corollary, which belongs in the same breath, is that none of the three predicts these outcomes very strongly. It is a statement about relative standing among weak predictors.

One domain beyond work, because a single literature is a poor calibration. Poropat meta-analysed the five-factor model against academic performance in Psychological Bulletin in 2009 and found conscientiousness the strongest and most consistent Big Five predictor of academic achievement, with a relationship large enough to rival intelligence at secondary and tertiary level in some of the pooled data. The same size discipline applies. And this meta-analysis predates the correction argument above, which concerned personnel selection specifically but raises the same class of question about correction practice generally.

The most useful thing a good taxonomy does is catch a repackaged construct, and our own file's counter-argument table records the Big Five doing exactly that. Crede, Tynan and Harms meta-analysed the grit literature in the Journal of Personality and Social Psychology in 2017 and found grit overlapping with the perseverance-of-effort component of conscientiousness at rho = .84. That rho denotes a disattenuated correlation, corrected for the unreliability of both measures; an uncorrected correlation between two scales could not approach .84 given their reliabilities. Grit, sold as a distinct predictor of success, is very largely conscientiousness under a newer name. A framework that can establish that is earning its keep, and it is worth noticing that our file puts the result in its own counter-arguments table rather than leaving it out.

There is a serious rival and it is not a fringe one. Ashton and Lee proposed HEXACO in Personality and Social Psychology Review in 2007: a sixth factor, Honesty-Humility, covering sincerity, fairness, greed avoidance and modesty, derived from lexical studies in multiple languages where six factors emerged more naturally than five. HEXACO is not the Big Five plus one. Its Agreeableness and its Emotionality differ from Big Five Agreeableness and Neuroticism in how the content is allocated between them, and Ashton, Lee and de Vries set that reallocation out themselves in 2014, which is the part our file states without explaining.

Our file states that Honesty-Humility is the strongest predictor of workplace counterproductive behaviour, at about r = -.40, and that it correlates negatively with Dark Triad traits. That number belongs to a study rather than to the world: Lee, Ashton and de Vries reported it in Human Performance in 2005, predicting workplace delinquency and integrity from the HEXACO and five-factor models. Three cautions travel with it. One correlation, for one predictor, against one criterion, in one literature, is not a settled ranking. The HEXACO group also built the scale. And self-report predictors of self-reported misconduct share method variance, which inflates the association between them for reasons that have nothing to do with either construct. Set that beside section 07, where the whole personnel-selection validity literature was revised downward for correction artefacts.

The debate our file states is whether HEXACO represents a genuine improvement over the Big Five or merely a rotation of the same factor space, with Big Five proponents arguing that Honesty-Humility content is partially captured by low Agreeableness. That word rotation is the most useful thing in this article. A factor solution is not a discovery of parts. It is a set of axes drawn through a cloud of correlations, and the axes can be turned. Different placements describe the same cloud, and the reason five became standard is that five kept replicating across samples, methods and laboratories, not that anybody ever found five objects. On that reading HEXACO is not a heresy but a different placement of the axes with a case behind it, and Saucier's independent finding of six recurrent dimensions in inclusive lexical studies strengthens that case from outside the HEXACO group. Our own file lists HEXACO as the counter to the claim that the Big Five is the optimal personality taxonomy, which is exactly the right place for it.

Paulhus and Williams named the Dark Triad in the Journal of Research in Personality in 2002: narcissism, meaning grandiosity, entitlement and a need for admiration; Machiavellianism, meaning strategic manipulation, cynicism and a pragmatic morality; and psychopathy, meaning callousness, impulsivity and antisocial behaviour. The three are conceptually distinct and empirically correlated with each other at roughly r = .25 to .50, which is the first hint that they may not be three separate things.

Our file records that Dark Triad traits predict workplace deviance, unethical behaviour, short-term mating strategies and interpersonal aggression, and gives no effect size for any of the four. Read predicts at the size everything else on this page has established: a modest correlation in a self-report literature, not a determination of anything about anybody. Section 07 is the calibration for that word, and it applies here without amendment.

Our file's own stated debate is the interesting part: whether the three Dark Triad traits are truly distinct from each other and from the Big Five at all, given that low Agreeableness combined with low Conscientiousness captures substantial variance in all three. The file also records the Dark Tetrad, which adds everyday sadism as a fourth dimension following Buckels, Jones and Paulhus in Psychological Science in 2013, and it flags measurement concerns with the brief self-report Dark Triad scales. So the concession is already on the record in our own corpus: the Dark Triad may be a rebranding of one corner of the five-factor space rather than a discovery beyond it. That matters, because this is the part of the subject most likely to reach a reader through popular media in the shape of a diagnosis. Nothing here is a diagnosis, and clinical psychopathy is a different subject with a different literature.

One contrast first, because it separates two objects that share a name. The Minnesota Multiphasic Personality Inventory, begun by Hathaway and McKinley in the 1940s, and its successors are the most widely used clinical personality measures. Our file gives the MMPI-2 at 567 items and the MMPI-3, updated in 2020, at 335 items; the publisher's own record confirms 335 true-false items across 52 scales, with a normative sample of 810 men and 810 women aged 18 and over matched to United States Census Bureau demographic projections, the first update since the mid-1980s. Its validity scales are empirically keyed to detect response distortion, which is to say faking good or faking bad. Our file records strong evidence for clinical use alongside criticism for item-content overlap and changing norms. The MMPI is not a Big Five instrument and this page does not present it as one. Its role here is the contrast in design philosophy: a clinical inventory built by empirically keying items against diagnosed groups, against a trait model built by factor-analysing the structure of ordinary language. Two completely different objects, both called a personality test.

Now the instrument most readers have actually taken. The Myers-Briggs Type Indicator has poor psychometric properties despite its popularity, and our file names three specific problems: low test-retest reliability for the type classifications, the information thrown away by cutting continuous dimensions into dichotomies, and a four-factor structure that does not replicate in factor analyses as cleanly as Big Five models do. The retest figure needs its range and its source, and our file supplies neither. Pittenger's 1993 review in Review of Educational Research cites a study in which 50 percent of subjects were reclassified on one or more of the four scales across a five-week interval. His 2005 review reports the wider literature as giving somewhere between 39 percent and 76 percent changing at least one letter on retest, depending on the interval and the sample. So the honest statement is a range with a most-cited point inside it, not a bare one in two.

The reason for that instability is the design, and stating it precisely is the clearest way to see what a continuous dimension is. The Big Five gives every person a position on each of five continua, and the distributions are unimodal, which means most people sit somewhere near the middle rather than clustering at two ends. The Myers-Briggs takes four continua and cuts each into two boxes. A person near the middle of a scale therefore flips category on a trivial change of mood or item wording, comes out with a different four-letter type, and nothing about them has changed. That is not carelessness by the people who use it. It is what necessarily happens when a cut point is imposed on a distribution that has no natural gap in it. And our own file's own cited paper makes the point without hostility: McCrae and Costa, in the Journal of Personality in 1989, reinterpreted the Myers-Briggs from the perspective of the five-factor model, mapped its four scales onto four of the five factors, and found them meaningful as continuous measures while rejecting the type dichotomies. The four scales are measuring something. The sixteen boxes are the problem.

Our file records the instrument's popularity in business and career counselling at roughly two million administrations a year. That figure is a publisher's estimate. It originates with the company that acquired the rights in 1975, is repeated through the manual and from there into the academic and popular literature, and no independent audit of it was located for this page. So it is carried here as what it is, an estimate by the party that sells the instrument, rather than as a measured statistic. In an article whose subject includes the difference between commercial personality typing and psychometrics, that distinction matters more than it usually would.

Refused: that the Big Five is basically a better Myers-Briggs, five personality types instead of sixteen. The Big Five has no types at all. It has five continuous dimensions on which every person sits somewhere, usually not far from the middle, and there is no defensible cut point that turns a score into a category. Which cuts against this article too: a page that describes somebody as high in openness, full stop, as though that named a kind of person, has quietly reintroduced the error it is criticising. Nor is a Big Five profile a fixed portrait. Rank-order consistency plateaus near .74 over a 6.7-year interval and never reaches unity, and mean levels shift systematically with age.

Our file places digital personality assessment at its speculative tier and the placement is right. Youyou, Kosinski and Stillwell reported in the Proceedings of the National Academy of Sciences in 2015 that a model trained on Facebook Likes predicted a person's Big Five scores at r = 0.56, against r = 0.49 for that person's own Facebook friends filling in a personality questionnaire about them. The thresholds are the arresting part: 10 Likes to beat a work colleague, 70 to beat a cohabitant or friend, 150 to beat a family member, 300 to beat a spouse. Participants averaged 227 Likes. Now the sentence that almost never travels with that study. Both the computer and the human judges were scored against the target's own self-report. The self-report is the yardstick, so what the study shows is that the model matches what you say about yourself better than your friends do. It is not a demonstration that an algorithm knows you better than you know yourself, and our file's flags on generalisability, privacy and misuse are the right ones to keep.

Politics is the other entry at that tier, and the numbers need replacing. Our file reports Openness correlating with liberal political orientation at about r = .30 and Conscientiousness with conservatism at about r = .20, citing Carney and colleagues in Political Psychology in 2008, while noting that causal direction, mediating mechanisms, and whether these reflect genuine traits or politically tinged self-presentation all remain debated. Sibley, Osborne and Duckitt meta-analysed 73 studies covering 71,895 people in the Journal of Research in Personality in 2012 and report Openness at r = -.18 with political conservatism and Conscientiousness at r = .10 with conservatism, both of which their own framing calls significant but weak. That is roughly half our file's figures in both cases, which is the classic shape of a single-study headline that shrank when the studies were pooled. And the more important result in that meta-analysis is neither point estimate. The Openness relationship is strongly moderated by systemic threat and uncertainty, indexed by national homicide and unemployment rates: it reaches r = -.42 where systemic threat is low and falls to a negligible r = -.07 at only moderate levels of threat. A relationship that swings that far with national circumstance is not a constant of human nature.

Two taxonomies have been tested properly and failed, and they belong in the body of this article rather than in a footnote, because they show what failure looks like against a background where the five-factor structure largely held. Blood type personality theory, widely believed in Japan and South Korea, holds that ABO blood type determines personality: Type A earnest, Type B creative, Type O confident, Type AB rational. Rogers and Glendon tested it in Personality and Individual Differences in 2003, and Nawata tested it again with large-scale surveys in Japan and the United States in 2014. Studies with more than 10,000 participants find no meaningful relationship between blood type and personality. This is not a subject for mockery. It is a live belief system in two large countries with documented consequences in employment and dating, and it is a clean demonstration of what it looks like when a personality taxonomy is tested and does not survive.

The same for astrology. There is no credible evidence that birth date or zodiac sign predicts personality traits, and Hartmann, Reuter and Nyborg examined the relationship between date of birth and individual differences in personality and general intelligence in a large-scale study in Personality and Individual Differences in 2006. Our file records studies with more than 15,000 participants finding no significant differences in Big Five scores by sun sign, and attributes any self-reported effects to confirmation bias and the Barnum or Forer effect: a description vague enough to fit everybody reads as uncannily personal. That mechanism is worth naming for an uncomfortable reason as well as a comfortable one. It is a risk the Big Five runs too, in popular presentation, and an article unwilling to say so is not being straight with its reader.

The deepest objection to the Big Five is not about any of its findings. It is about the premise section 01 stated, and our own file does not mention it. Jack Block's A Contrarian View of the Five-Factor Approach to Personality Description, in Psychological Bulletin in 1995, makes four objections and every one of them survives being stated plainly. The lexical hypothesis assumes that what matters about personality is what lay language happened to encode, which is a strong and unargued premise. The number of factors recovered is not discovered but is a consequence of the variables selected, the sample used and the rotation chosen. The resulting five are broad to the point of being uninformative about any individual. And factor-analysing adjectives can recover the structure of raters' implicit personality theories rather than the structure of personality itself. Block returned to the argument six years later, and the exchange is on the record on both sides: Costa, McCrae, Goldberg and Saucier all replied in print in the same 1995 volume, and Block published a rejoinder to them.

Refused: that personality testing is pseudoscience, the Big Five included, and none of it predicts anything. This is the mirror error and it gets refused as firmly as the first one. The five-factor structure is recovered repeatedly from independent methods: self-report, peer rating, lexical analysis and language use. Rank-order consistency rises to about .74 by later middle age, over a 6.7-year interval. A preregistered, high-powered replication project found 87 percent of 78 published trait-outcome associations replicating in the expected direction, at about 77 percent of the original effect size, which is a record most areas of psychology would envy. And the model does real discriminative work: it is what showed that grit is very largely conscientiousness renamed. The correct verdict is neither proven universal nor junk. It is a well-replicated descriptive taxonomy with modest, real predictive validity, an unresolved cross-cultural boundary, and an explanatory story it does not yet have.