Statistics Glossary → Factor and component scores
Factor and component scoresconcept
Scores giving each observation's value on the latent factors found by a factor analysis, or on the components of a principal component analysis.
In principal component analysis the components are exact linear combinations of the observed variables, so component scores are computed exactly and, when standardised, have mean 0 and standard deviation 1. In factor analysis the factors cannot be derived exactly from the observed variables (factor indeterminacy), so the scores can only be estimated, and different methods give different results. The regression (Thurstone) method maximises the correlation with the true factor, but its scores are biased and may be correlated with one another. Bartlett's method gives unbiased scores. The Anderson-Rubin method produces uncorrelated scores with unit variance. The ten Berge method preserves the correlations between factors.
Scores can be used in later analyses (regression, group comparisons). But estimated scores contain measurement error, and because they do not carry this uncertainty they can weaken relationships or make standard errors look too small. Where possible, relationships should therefore be estimated directly within a structural equation model. Simple sums or means of items weight all items equally, whereas factor scores weight items according to their loadings. If a rotation was applied, the scores are computed from the rotated solution.
An everyday example
A company measures employee satisfaction with 12 questions, and factor analysis reveals the factors 'pay satisfaction' and 'relationship with manager'. Each employee gets a score on both factors. An employee with a pay score of −1.2 and a manager score of +0.8 is clearly less satisfied with pay than average and more satisfied than average with their manager.
In MerQur
In principal component analysis, 'Save component scores to the dataset' adds the standard scores as columns PCA_PC1 and so on. In exploratory factor analysis, the 'Factor scores' method (Regression, Bartlett, Anderson-Rubin, ten Berge) is chosen and the scores can be saved as columns such as FAC1_REG. In PLS-SEM, 'Extra output (save construct scores to the dataset)' adds the construct scores as new columns.
Analyses where this concept appears
Principal Component Analysis (PCA)Exploratory Factor Analysis (EFA)PLS-SEM (Partial Least Squares Path Modeling)