Description
Estimates each respondent's score on factors from a factor solution, checks how determinate those scores are, and extends the solution to new variables without refactoring.
Factor scores cannot be computed exactly, only estimated. The robust simple option is to sum the variables that load highly on a factor with equal weights. The most common method, Thurstone's regression approach, treats each factor as a criterion to predict: the weights are the inverse of the item correlation matrix times the factor structure matrix, applied to standardised item scores. Estimated scores correlate more strongly with each other than the factors do (for example .67 against .50), which matters if they feed later analyses. The deeper issue is indeterminacy: a given solution is consistent with a range of possible factors, nearly identical when communalities are high but potentially even uncorrelated when they are low. Extension analysis correlates estimated factor scores (or uses a shortcut with the weights) with variables that were not in the analysis, giving their loadings on the existing factors without disturbing the original solution.
Attributes
- Discipline
- Survey & Psychometrics
- Loss type
- Siloed
- Stage
- Evaluation, Discovery
- Data
- Quantitative
- Effort
- Days
Steps
- Obtain a factor solution with pattern matrix, structure matrix and factor correlations from EFA or CFA.
- Compute regression weights as the inverse item correlation matrix times the structure matrix (or choose unit weights on salient items, or Bartlett scores if unbiasedness matters).
- Apply the weights to standardised item responses to obtain each respondent's estimated factor scores.
- Check determinacy: compare intercorrelations of the estimated scores with the factor correlations, and inspect communalities or determinacy coefficients; treat scores from low-communality factors with caution.
- For extension, multiply the correlations of new variables with the original items by the rescaled weights to get their structure loadings, then convert to pattern loadings.
- Use the scores in segmentation or regression only with their estimation error acknowledged, or move to SEM to keep the factors latent.
Sources
- Latent Variable Models (Loehlin & Beaujean, 2017)
- Structural Equation Modeling: A Bayesian Approach (Lee, 2007)
Also known as
- Regression factor scores
- Thurstone factor score weights
- Factor score indeterminacy check
- Extension analysis
Categories
- Analytics & Behaviour
- Survey & Psychometrics