Model-based clustering and analysis of life history data

Methods and models for longitudinal data with categorical, multi-dimensional outcomes are quite limited, but they are essential to the study of life histories. For example, in the Swiss Household Panel, information on the co-residence and professional status of several thousand individuals is available through to age 45 years. Interest centres on the time and order of life course events such as having children and working full or part time and the duration of the phases that they delineate. With data of this type, optimal matching and clustering algorithms relying on a distance metric or parametric models of duration in a competing risks framework are used; the appropriateness of each derives from competing goals and orientation. We prefer model-based approaches when certain goals are paramount: simulation of individual trajectories; adjusting for time-dependent covariates; handling multistate trajectories and missing outcomes. Several of these goals are particularly challenging when the number of states is of moderate size, and many transitions are infrequent and/or time inhomogeneous. Using the Swiss Household Panel, we demonstrate the appropriateness of latent class growth curve models for analysing sequence data. In particular, models including heterogeneous dependence structure provide new techniques for assessing goodness of fit as well as yield insights into social processes.