Latent class analysis (LCA) posits a categorical latent variable that explains the associations among categorical manifest variables. Simultaneous LCA (Clogg-Goodman) relates an explanatory categorical grouping variable to the latent classes; van der Heijden, Dessens and Böckenholt generalize this to continuous explanatory variables via the concomitant-variable latent-class model, and work out an EM estimation procedure in detail. The method is illustrated on an example of juvenile delinquency.
"We discuss a tool called the concomitant-variable latent-class model, which generalizes this work to continuous explanatory variables. An EM estimation procedure to estimate the model is worked out in detail, and the model is applied to an example on juvenile delinquency."
A clean, load-bearing methods paper: it is where the "let class membership depend on covariates" idea gets a fully worked EM recipe for continuous predictors, which is exactly the machinery that later became routine in Latent GOLD and the Bandeen-Roche latent-regression tradition. The value is in the M-step decomposition — separating the measurement update from a weighted regression of latent class on covariates — because that modularity is what makes the concomitant model composable with the rest of the LCA toolkit (class-count selection, standard errors, the three-step workflow). It is unglamorous but foundational, and a good example of why the wiki wants the original estimation paper and not only the later syntheses that cite it.