In this article, we present a two-stage estimation approach applied to multilevel latent class analysis (LCA) with covariates. We separate the estimation of the measurement and structural model. This makes the extension of the structural model computationally efficient. We investigate the robustness against misspecifications of the proposed two-stage and the classical one-stage approach for models where a direct effect exists between indicators of the LC model and covariate, and the direct effect is ignored.