Multilevel latent class analysis with gsem

Since version 15, Stata offers the possibility to estimate latent class models for categorical observed indicators being dichotomous, ordinal, or nominal. It also integrates manifest covariates to predict the class membership of the observations. Both parts of the model, the measurement one and the prediction one, are estimated simultaneously so that changes in the measurement part influence the estimates of the structural part and vice versa. According to Hayduk (1996) and Bakk and Kuha (2020), I propose a three-step approach. First, I estimate a sequence of latent class models identifying the most appropriate solution by the entropy criteria. Second, I analyze a profile plot of the item probabilities to attach meaningful labels to the latent classes. The assignment of observations to the discrete latent classes follows the highest probability rule. Third, I estimate a multinomial logit regression model to predict the discrete latent class membership by exogenous level 1 and level 2 variables. To estimate a logistic intercept-as-outcome model, I use the Stata xtmlogit command introduced by version 17. I demonstrate the usefulness of this approach presenting a latent class analysis of attitudes toward vaccination at the eve of the COVID-19 pandemic using the special eurobarometer 488 dataset. Being a vaccination supporter, a conspirator, or a naif is predicted within 28 European countries by the personal characteristics of the respondents and between countries by their fixed-effect dummy variables. To enlighten these black boxes, I estimate a random-effect intercept-as-outcome multinomial logit model with the xtmlogit command using exogenous level 2 variables like collective level of trust in government, GDP, poverty rate, and the global health security index. Finally, I discuss the main results and give some methodological considerations.