Analysis of counts with two latent classes, with application to risk assessment using physician visit records

Motivated by the CAYACS program at BC Cancer Research Center, this thesis project introduces a latent class model to formulate event counts. In particular, we consider a pop- ulation with two latent classes, such as an at-risk group and a not-at-risk group of cancer survivors in the CAYACS program. Likelihood-based inference procedures are proposed for estimating the model parameters with or without one class fully specified. The EM algo- rithm is adapted to compute the MLE; a pseudo MLE of the model parameters is proposed to reduce computing intensity and improve inference efficiency using readily available sup- plementary information. The estimation procedures are studied via simulation regarding both efficiency and robustness. We illustrate the methodology with the physician claim data of the CAYACS cohort for risk assessment throughout the project. With the latent class model, we identify risk factors for cancer survivors to late and on-going problems and obtain an alternative, perhaps more desirable, comparison of the cohort with the general population.