Bayesian factor analysis (BFA) assumes the normal distribution of the current sample conditional on the parameters. Practical data in social and behavioral sciences typically have significant skewness and kurtosis. If the normality assumption is not attainable, the posterior analysis will be inaccurate, although the BFA depends less on the current data due to prior information. This article proposes to apply a robust procedure to the sample before performing a BFA. Examples show that this procedure leads to a more accurate evaluation of the factor structure when data contain outliers.