Factor Loading

Factor Loading: Concept Architecture

Connect an observed measure to a latent factor

A factor loading is an estimated coefficient linking an observed variable to an unobserved factor in a specified measurement model. This page shows what the coefficient means, how scaling and rotation affect it, and why loadings help evaluate health measurement without becoming preference weights.

Read the coefficient within its model

In a one-factor measurement model, an observed item $X_i$ can be represented as an intercept, a loading on a latent factor $F$, and item-specific error $\varepsilon_i$:

$$X_i=\nu_i+\lambda_iF+\varepsilon_i.$$

The loading $\lambda_i$ expresses the expected change in the item for a one-unit change in the factor under the model. Its magnitude depends on how the item and factor are scaled; a standardized loading is easier to compare across items. The sign indicates direction under the chosen orientation of the factor, whose sign can be reversed without changing the substantive fit.

For questionnaires about physical function or mental well-being, factor analysis asks whether observed responses reflect one or more intended constructs. An exploratory analysis investigates plausible patterns; a confirmatory analysis tests a prespecified structure against observed covariation. A health-scale validation tutorial explains the role of factor analysis in assessing structural validity.

Work through a single-factor example

Suppose a standardized mobility item has a standardized loading of 0.80 on a standardized physical-function factor, with uncorrelated residual error in a one-factor model. Under those assumptions, a one-standard-deviation difference in the factor corresponds to a 0.80-standard-deviation difference in the item's model-predicted value. The factor accounts for $0.80^2=0.64$, or 64%, of the item's variance; the residual accounts for 36%.

This squared-loading interpretation relies on the single-factor, standardized, uncorrelated-error setup. With multiple correlated factors, communalities and factor–item correlations require the factor covariance structure; simply squaring one coefficient is not a general answer. A UCLA statistical-methods guide distinguishes factor pattern coefficients, factor–item structure correlations, and correlations among factors in oblique solutions.

Check the structure, not only the largest number

Inspect the loading matrix as a whole, the intended number of factors, cross-loadings, residual correlations, sample adequacy, and the fit of the chosen model. Rotation can make a pattern easier to interpret, but it changes the displayed coefficients and does not rescue an inappropriate construct or poor data. Confirmatory models also require attention to identification and global as well as local fit.

Compare measurement properties across the populations, languages, or disease groups in which an instrument will be used. A loading estimated in one sample need not be stable elsewhere; measurement invariance must be evaluated before interpreting group differences as differences in the underlying construct. An applied health-related quality-of-life study examines whether factor loadings are sufficiently comparable across stroke-history groups.

Keep measurement separate from valuation

Loadings can help assess whether questionnaire items support an intended health construct and whether a summary score has a defensible structure. That matters when patient-reported outcomes inform health economic evidence, subgroup comparisons, or mapped health utilities. A loading is nevertheless a statistical measurement coefficient, not a treatment effect, a probability, or a social preference weight for calculating quality-adjusted life years.

Report the extraction or estimation method, item and factor scaling, factor number, rotation, loadings with uncertainty where available, sample, fit, and any invariance checks. Do not interpret a large loading as proof that the item is unbiased, clinically important, or responsive to treatment; those properties require their own evidence. Distinguish common-factor analysis from principal components analysis, which constructs components to summarize observed variance rather than positing the same latent-error measurement model.