Stepwise Latent Class Analysis in the Presence of Missing Values on the Class Indicators

While latent class (LC) modeling using bias-adjusted stepwise approaches has become widely popular, little is known on how these methods are affected by missing values. Using synthetic data sets, we illustrate under which conditions missing values introduce biases in the estimates of the relationship between class membership and auxiliary variables. We apply three-step LC analysis with both modal and proportional class assignments, as well as the recently proposed two-step LC analysis method. Our results show that stepwise LC analysis yields unbiased parameter values as long as the MAR assumption holds in the step-one model. When this assumption does not hold because covariates are omitted from the step-one model, each of the stepwise approaches yields some bias, but bias is much larger with modal class assignments. The amount of bias is affected by the amount of deviation from MAR, the proportion of missing values, and the separation between the classes.

Automated summary

This study investigates how missing values affect latent class modeling using bias-adjusted stepwise approaches and proposes a new analysis method. Using synthetic data sets, the results show that unbiased parameter estimates can be obtained in stepwise LC analysis when the MAR assumption holds in the first step model. However, bias is introduced when this assumption does not hold, especially with modal class assignments. The amount of bias is influenced by the deviation from MAR, the proportion of missing values, and the separation between classes.