Bayesian Exploratory Factor Analysis (BEFA)
Bayesian Exploratory Factor Analysis · Also known as: Bayesian factor analysis, BEFA, Bayesian common factor model, probabilistic factor analysis
Bayesian exploratory factor analysis applies a full probabilistic framework to the common factor model. By placing prior distributions over factor loadings and unique variances, it yields posterior distributions rather than point estimates, quantifies uncertainty around every loading, and can treat the number of factors as an unknown to be inferred from data.
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.
When to use it
Use Bayesian EFA when you want genuine uncertainty quantification around factor loadings — for example, in small samples (n < 200) where classical standard errors are unreliable, or when you have substantive prior knowledge to encode. It is also well suited when the number of factors is unclear and you want the data to help decide. Do NOT use it as a drop-in replacement for classical EFA when a quick, reproducible output is needed: MCMC is computationally intensive, requires convergence diagnosis, and demands more statistical expertise to specify and interpret. Bayesian EFA is not appropriate if items are binary and a dichotomous IRT model is the correct measurement model, or if the goal is confirmatory — for that, use Bayesian CFA instead.
Strengths & limitations
- Produces full posterior distributions over loadings, providing credible intervals rather than point estimates alone.
- Uncertainty about the number of factors can be propagated through the analysis rather than treated as fixed.
- Shrinkage priors automatically regularise weak loadings toward zero, reducing spurious cross-loadings in small samples.
- Posterior predictive checks provide a principled way to evaluate and compare competing factor solutions.
- Naturally accommodates missing data and hierarchical extensions within the same model.
- MCMC estimation is computationally expensive and can require thousands of iterations and careful convergence monitoring.
- Prior specification is non-trivial: different priors on the loadings or unique variances can meaningfully influence the posterior in small samples.
- Rotational indeterminacy still applies and must be handled explicitly post-hoc or through identifiability constraints in the prior.
- Output interpretation requires familiarity with Bayesian concepts (posterior, credible interval, MCMC convergence) that many applied researchers lack.
- Software support is less mature than for classical EFA; implementations exist in R (befa, BayesFM) but require more setup than a standard EFA call.
Frequently asked
Is Bayesian EFA the same as classical EFA with uncertainty bounds?
Not quite. Classical EFA produces standard errors from asymptotic theory and is sensitive to distributional assumptions at small n. Bayesian EFA produces credible intervals from the actual posterior distribution, incorporates prior information, and treats the number of factors as a random variable. In large samples with well-behaved data the numerical results are often similar, but the inferential logic and small-sample behaviour differ substantially.
How do I choose a prior for the loading matrix?
A common choice is the multiplicative gamma process prior (Bhattacharya and Dunson, 2011), which shrinks columns of the loading matrix progressively so that factors with little explanatory power are automatically suppressed. For confirmatory-style constraints a lower-triangular prior structure is often used. If you have genuine prior knowledge from a pilot study, an informative normal prior on specific loadings is defensible. Avoid entirely diffuse priors: they do not penalise complexity and negate one of Bayesian EFA's main advantages.
How many factors does Bayesian EFA select?
With a shrinkage prior the effective number of factors is determined by how many columns of the posterior loading matrix have at least one loading clearly away from zero. You can compare models with different fixed k using DIC or WAIC, or use a reversible-jump scheme that moves between k values during sampling. In practice the posterior tends to concentrate on a narrow range of k values, making the model-selection uncertainty easier to communicate than choosing a single k via parallel analysis.
When should I use Bayesian CFA rather than Bayesian EFA?
Use Bayesian CFA when you have a pre-specified factor structure — for example, from prior literature or a previous EFA study — and want to test how well it fits while quantifying parameter uncertainty. Use Bayesian EFA when the structure is genuinely unknown and you want the data, regularised by priors, to reveal it.
Does Bayesian EFA work with ordinal items?
Yes, through an ordinal probit extension: each ordinal item is modelled as a discretisation of a continuous latent variable, and the factor model is placed on those latent variables. This is computationally heavier but appropriate for Likert-type items and avoids the assumption of interval-level measurement.
Sources
- Lopes, H. F. & West, M. (2004). Bayesian model assessment in factor analysis. Statistica Sinica, 14(1), 41–67. link ↗
- Ghosh, J. & Dunson, D. B. (2009). Default prior distributions and efficient posterior computation in Bayesian factor analysis. Journal of Computational and Graphical Statistics, 18(2), 306–320. DOI: 10.1198/jcgs.2009.07145 ↗
How to cite this page
ScholarGate. (2026, June 3). Bayesian Exploratory Factor Analysis. ScholarGate. https://scholargate.app/en/psychometrics/bayesian-exploratory-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.
- Bayesian Confirmatory Factor AnalysisPsychometrics↔ compare
- Confirmatory factor analysisPsychometrics↔ compare
- EFAStatistics↔ compare
- Item Response TheoryPsychometrics↔ compare