Ordinal Confirmatory Factor Analysis
Also known as: CFA for ordinal data, polychoric CFA, WLSMV CFA, categorical CFA
Ordinal confirmatory factor analysis (Ordinal CFA) tests a pre-specified factor structure when the observed indicators are ordinal — typically Likert-type survey items. By using polychoric correlations and robust estimators such as WLSMV, it avoids the bias that arises from treating categorical responses as continuous.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
+3 more
When to use it
Use Ordinal CFA whenever you want to confirm a theory-driven factor structure and your indicators are ordinal items — Likert scales with seven or fewer categories are the prototypical case. It is appropriate when the number of categories is small enough that treating items as continuous would distort fit statistics or loadings. Ordinal CFA is the right choice after an Ordinal EFA has suggested a structure on a calibration sample, allowing confirmatory testing on a new sample. Do not use Ordinal CFA when the factor structure is unknown and needs to be discovered — use Ordinal EFA first. Do not apply it to binary items without switching to the dichotomous parameterisation (or IRT). Do not use standard (continuous) CFA with ML estimation on ordinal items with fewer than five categories, as this systematically underestimates loadings and inflates fit problems.
Strengths & limitations
- Correctly accounts for the categorical nature of Likert-type items, avoiding attenuation of loadings and distortion of fit indices that occur with ML on ordinal data.
- WLSMV is computationally efficient and handles missing data under MAR via pairwise approaches or FIML extensions.
- Thresholds provide additional diagnostic information about response scale functioning, revealing floor/ceiling effects.
- Supports full measurement invariance testing (configural, metric, scalar) on the polychoric scale across groups.
- Factor loadings and correlations estimated on the latent-response scale are directly comparable to IRT parameters, facilitating bridges between traditions.
- Requires larger samples than continuous CFA; WLSMV is sensitive to sparse cells when categories are extreme or items have many categories with few respondents.
- Model fit indices have different distributions from standard ML equivalents; software-specific corrections must be used, and published cutoffs (CFI > 0.95) are only approximate guidelines for WLSMV.
- Scalar invariance is defined on the polychoric (latent-response) scale; if thresholds are not invariant, observed-score comparisons across groups remain problematic.
- Polychoric correlations can become inadmissible (outside −1 to 1) when cell frequencies in the bivariate tables are very small, leading to non-positive-definite matrices.
Frequently asked
Can I use standard CFA with ML if my Likert scale has seven points?
With seven categories the continuous approximation is much better than with four or five, and ML CFA may give acceptable results if the distributions are not severely skewed. However, Ordinal CFA with WLSMV is still preferable when the data are clearly non-normal or when you plan scalar invariance testing, because it directly models the categorical response mechanism.
What is the minimum sample size for Ordinal CFA?
There is no universal minimum, but WLSMV requires a stable polychoric correlation matrix. Simulation studies suggest at least 200 cases for simple models with many indicators per factor. Models with few indicators per factor or skewed response distributions may need 400 or more. Very sparse bivariate tables (e.g., cells with fewer than five observations) cause estimation problems regardless of overall N.
How does Ordinal CFA relate to IRT?
The graded response model and the partial credit model in IRT are mathematically equivalent to single-factor ordinal CFA under the normal ogive or logistic link. Ordinal CFA generalises IRT by allowing multiple correlated factors and equality constraints. IRT software emphasises person-level scoring and item-level diagnostics, whereas ordinal CFA software emphasises structural fit and factor correlations.
What software runs Ordinal CFA?
Mplus provides the most comprehensive implementation via the WLSMV estimator with categorical indicators. In R, the lavaan package supports WLSMV (called DWLS there) for ordered indicators. OpenMx and the R package lsfa also support variants. LISREL introduced the approach historically through its PRELIS polychoric preprocessing step.
How do I test measurement invariance in Ordinal CFA?
The standard sequence runs configural (same pattern, all parameters free), metric (equal loadings across groups), and scalar (equal loadings and thresholds) models. In WLSMV, model comparisons use the Satorra-Bentler scaled difference chi-square test rather than a simple chi-square difference, because WLSMV fit functions are not chi-square distributed. Partial invariance — freeing non-invariant thresholds — is acceptable when substantively justified.
Sources
- Flora, D. B. & Curran, P. J. (2004). An empirical evaluation of alternative methods of estimation for confirmatory factor analysis with ordinal data. Psychological Methods, 9(4), 466–491. DOI: 10.1037/1082-989X.9.4.466 ↗
- Muthén, B. O. (1984). A general structural equation model with dichotomous, ordered categorical, and continuous latent variable indicators. Psychometrika, 49(1), 115–132. DOI: 10.1007/BF02294210 ↗
How to cite this page
ScholarGate. (2026, June 3). Ordinal Confirmatory Factor Analysis. ScholarGate. https://scholargate.app/en/psychometrics/ordinal-confirmatory-factor-analysis
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Confirmatory factor analysisPsychometrics↔ compare
- EFAStatistics↔ compare
- Item Response TheoryPsychometrics↔ compare
- Measurement InvariancePsychometrics↔ compare
- Ordinal EFAPsychometrics↔ compare