Personality Traits Exist as Measurements Before They Exist as Essences

Gordon Allport and Henry Odbert begin with a dictionary, which is a rather grand place to look for the soul and a very good place to find adjectives behaving badly.

They worked through the whole of Webster’s New International and extracted nearly eighteen thousand words describing personality or behavior. About a quarter of them were judged to name stable traits rather than temporary states or evaluations.

The lexical hypothesis supplied the premise. If a distinction matters enough in social life, language will probably have made a word for it.

The dictionary was not a personality test. It was raw material with delusions of grandeur.

Raymond Cattell spent a decade reducing those terms with factor analysis, the method that finds variables correlating more with each other than with the rest.

He settled on sixteen primary factors. Early computer access let him run larger analyses, and it did not compel the universe to be five things wide. Cattell’s model was sixteen, and he disputed the reduction.

Then the reduction kept returning anyway.

Fiske, then Tupes and Christal with Air Force officer ratings, then Norman each found five broad clusters. Goldberg named them the Big Five.

Costa and McCrae reached overlapping terrain from questionnaire construction, fixing the familiar vocabulary of openness, conscientiousness, extraversion, agreeableness, and neuroticism.

This is a respectable origin story. Air Force ratings, dictionary analysis, and clinical questionnaire work arrive by separate doors and find five chairs already set out.

The chairs need not therefore be ghosts, organs, or tiny bureaucrats living behind the forehead with clipboards.

A factor is first a pattern in measurements. It is what the correlations make visible.

Which is exactly where the pattern begins to look like a possession.

If several methods recover the same broad dimensions, if the scores endure, and if they predict things that matter, then perhaps the inventory has located natural kinds. An actual quantum of conscientiousness, tucked behind the ribs, waiting for its questionnaire to announce it.

The evidence does not carry that far.

It supports a model of patterned differences. It supports measurement with error, longitudinal regularity, and limited prediction.

That a factor extracted from one instrument names a discrete causal entity inside a person is speculation. It may be true. Factor analysis itself cannot deliver it, because it begins with correlations among answers, ratings, or behaviors, not with a tour of the alleged inner furniture.

Spearman had already supplied the small warning label.

He developed machinery for treating a test score as a noisy signal, then asked what a correlation might be without measurement error.

Classical test theory condenses the point. An observed score is a hypothetical true score plus random error.

The true score is hypothetical. It is not a little engraved nameplate found under the skin.

The distinction is not a technicality.

A scale may be reliable, meaning a substantial portion of its observed-score variance is not random error, and still have a live validity problem.

Cronbach and Meehl’s paper on construct validity exists precisely because an attribute like anxiety has no clean external object to compare with its score.

Their hypothetical anxiety test is exemplary because it refuses a cheap finish.

A respectable correlation with palmar sweating is not enough. Peer ratings of tenseness, responses to electric shock, another anxiety scale, and differences between diagnosed and non-diagnosed groups would make the interpretation stronger.

Each result can also reveal that the test was measuring something else in a cheap anxiety moustache.

The Big Five have more to show than a neat factor structure.

Meta-analytic heritability estimates for the five dimensions cluster somewhere near half. Those are proportions of variation within the studied populations, not percentages of anyone’s personality delivered by genetics like a parcel.

Stability has a similarly disciplined meaning.

Pooling more than three thousand test-retest correlations, rank-order consistency rises from about .3 in childhood to roughly .5 in the college years, .6 around thirty, and near .75 in later middle age.

Rank-order stability means remaining in a similar place relative to peers. It coexists perfectly well with mean-level change. Conscientiousness and agreeableness tend to rise with age. Neuroticism tends to fall.

Prediction does not rescue essence either.

Across a large meta-analysis of occupational groups, conscientiousness correlates with job performance at around .22 after correction.

It is real, replicated, and modest. A trait score is one input into a working life, not a fortune cookie in a necktie.

Popular type systems make the temptation especially theatrical.

Jung’s Psychological Types supplied the vocabulary. The Myers-Briggs indicator turned it into an instrument with four dichotomies and sixteen boxes.

When Costa and McCrae mapped those dichotomies onto continuous five-factor dimensions, they found no support for the discontinuous, bimodal type structure the categories imply.

The continuum has not, alas, provided the dramatic satisfaction of being told that half the office is one species and the other half its natural enemy.

Even a serious trait scale has another border to cross. Groups.

Measurement invariance asks whether an instrument measures the same construct, in the same way, across specified groups.

Multi-group confirmatory factor analysis tests nested levels. Configural invariance for the same factor structure. Metric invariance for equal loadings. Scalar invariance for equal intercepts. Strict invariance for equal residual variances.

Scalar invariance is the necessary level before means can honestly be compared.

Kato’s depression-scale study gives the requirement a paper trail.

Across White American, White Australian, Indian, Filipino, and Singaporean samples, the twenty-item instrument achieved configural and metric invariance.

Full scalar invariance failed. Only after nine items were released from equality constraints did partial scalar invariance appear.

Identical factor shapes, then, do not guarantee that the same number says the same thing.

That is the whole lesson in one procedure. A construct is a claim about what a pattern of answers is, and the claim has to be argued, not asserted by a coefficient.

What survives is not that personality has finally found the hidden furniture of the self, nor that psychometrics has exposed personality as a crossword puzzle with a confidence interval.

It is that personality becomes empirical by being measured, and becomes philosophically treacherous when the measurement is mistaken for the thing.

The score gives a pattern a handle. It does not make the handle a creature.