A Boosted Multistate Model of Partnership Trajectories in Germany

Angela Carollo , Max Planck Institute for Demographic Research
Guillermo Brise\~no Sanchez, Karlsruhe Institute of Technology
Valeria Ferraretto, University of Bologna
Nicole Hiekel, Max Planck Institute for Demographic Research

Romantic partnerships trajectories over time can be analyzed by means of multistate models. Individuals move between states such as dating, cohabiting, marriage and union dissolution in non-random ways, usually determined by some observable and non-observable characteristics. To disentangle the heterogeneity in these transitions, researchers usually fit regression models, choosing sets of predictors based on theoretical considerations, previous findings, and available data. In order to exploit the full potential of rich survey data, such as the German Family Panel (pairfam), we suggest to combine theoretical considerations with data-driven approaches of variable selection, such as statistical boosting, to identify the best set of predictors for each transition in a multistate model. The result is an interpretable statistical model for each transition, in which the covariates’ effects where selected and estimated automatically by the boosting algorithm. In this study, we combine statistical boosting algorithms with multistate models to study partnership trajectories in Germany and to identify the best predictors of transitions between states in a relationship. We discuss advantages and disadvantages of the proposed approach, and we detail future directions of research.

See paper

 Presented in Session 39. Innovations in Life Table and Multistate Modeling