What is Factor Analysis?

Factor analysis is a technique for identifying the smaller number of unobserved dimensions (factors) that account for the correlations among a larger set of observed variables. The premise: when many measures correlate in structured ways, something underneath is driving them, a latent variable that no single question captures but several together reflect. The analysis estimates how strongly each observed item loads on each factor, how many factors the data supports, and how much of the pattern they explain, turning a correlation tangle into an interpretable structure: these six items reflect "perceived ease", those four reflect "trust", and so on.

A Brief History

The method was born with a substantive claim: Charles Spearman's 1904 work observed that children's scores across diverse school subjects all correlated positively, and proposed a single underlying general ability (his famous g) to explain the pattern, inventing early factor-analytic machinery to formalise it. The technique outgrew the intelligence debates that spawned it (and that still contest g's interpretation) to become psychometrics' general-purpose engine: wherever multi-item questionnaires exist, factor analysis is how their internal structure was discovered, tested, or should have been.

Exploratory and Confirmatory

Exploratory factor analysis (EFA) asks the open question: how many dimensions live in these items, and which items belong to which? It is the instrument-development mode: draft a broad item pool, field it, and let the loading pattern reveal the constructs, pruning items that load nowhere or everywhere.

Confirmatory factor analysis (CFA) tests a stated structure: the theory says these ten items measure two specific factors this way; how well does the data fit that blueprint? It is the validation mode, the formal half of the construct-validity argument for any scale claiming to measure something in particular. (A neighbouring method, principal component analysis, performs similar compression with a different philosophy: PCA summarises variance without positing latent causes, and the two get conflated constantly.)

Factor Analysis in Research Practice

Its product-research jobs: building instruments (does the new satisfaction battery measure the three things it claims, or one blur?), auditing borrowed ones (do the standardised scales behave in your population as published?), structuring driver batteries before driver analysis (factors as predictors beat twenty collinear items), and reducing survey length honestly (if six items measure one factor, three well-chosen ones may suffice, a gift to respondent attention). Its requirements are real: comfortable sample sizes relative to item counts, correlations worth analysing, and judgment at the famous discretion points (how many factors to keep, which rotation, what loading threshold), each a place where two analysts can defensibly differ, which is why documented decisions are part of the method.

The Cautions

1. Factors are constructs, not discoveries of nature.
The analysis proposes structure; naming a factor "trust" is an interpretive act that theory and further validation must earn. Reifying factor labels is the field's oldest bad habit, dating to g itself.

2. Structure depends on the items you fed it.
Factors can only be made of what was asked; a dimension missing from the item pool is invisible, however important.

3. Sample and stability matter.
Small-sample factor solutions wobble; replication on fresh data is the difference between an instrument and an anecdote about one dataset.

The Takeaway

Factor analysis is the microscope for what questionnaires really measure: loadings revealing which items travel together, factors naming the latent drivers, EFA to discover and CFA to confirm. Use it to build shorter, cleaner, construct-honest instruments, document the judgment calls, replicate before believing, and hold the labels lightly: the mathematics finds the clusters; meaning remains the researcher's job.