ABSTRACT
The likelihood for multistate processes under intermittent observation, unlike the likelihood for right-censored data, does not factor into functionally independent components where each involves a different parameter vector. This chapter discusses approaches to model assessment based on comparison of parametric and nonparametric estimates, model expansion, examination of residuals or influence measures, and predictive assessment. When the states occupied by individuals are observed only intermittently, all approaches but model expansion are problematic except in special cases; nonparametric estimation of transition intensities and construction of residuals are not feasible. Pure nonparametric estimation of cumulative transition intensity functions is intractable for most multistate models when processes are under intermittent observation. Two exceptions are competing risk models and illness-death models. In some cases, it can be obtained nonparametric estimates of state occupancy probabilities, or probability distributions for entry times that may be helpful provided observation times are ignorable.
