Postestimation with latent class analysis accounting for class uncertainty

Latent class analysis (LCA) is a statistical model with categorical latent variables in which the measured categorical outcomes have proportions of the outcome categories that differ between classes. In official Stata, the model is fit using the gsem, lclass() command. Applied researchers often need to follow up the LCA modeling with other statistical analyses that involve the classes from the model, from simple descriptive statistics of variables not in the model, to multivariate models. A simplified shortcut procedure is to assign the class with the highest predicted probability, but doing so results in treating the classes as fixed and perfectly observed, rather than latent and estimated, leading to underaccounting of uncertainty and downward bias in standard errors. We demonstrate how to utilize the existing official Stata multiple imputation (MI) capacity to impute classes based on the LCA postestimation results and present the resulting dataset to Stata mi procedures as valid MI data. The standard MI diagnostics that can be applied to the mi estimate results show that variances are noticeably underestimated when only the modal class is imputed. In the application that motivated this development, the variances were biased down by 25% to 40%.