Semiparametric regression analysis of panel count data allowing for within-subject correlation

In this paper, a maximum likelihood approach is proposed for analyzing panel count data under the gamma frailty non-homogeneous Poisson process model. The approach allows one to estimate the baseline mean function and the regression parameters jointly while taking the within-subject correlation into account. The within-subject correlation is quantified explicitly by Pearson's correlation coefficient. Monotone splines are adopted to approximate the unspecified nondecreasing baseline mean function in the model. An expectation-maximization (EM) algorithm is derived to facilitate the computation by exploiting a data augmentation based on Poisson latent variables. The EM algorithm is robust to initial values, easy to implement, converges fast, and provides closed-form variance estimates. It can be also applied to the non-homogeneous Poisson model without frailty. The proposed approach is evaluated through simulations and illustrated by two real-life examples coming from a skin cancer study and a bladder tumor study. A companion R package PCDSpline has been developed and is available on R CRAN for public use.