Estimating transition rates for multistate models from panel data and repeated cross-sections

`Incomplete' data sources, such as panel data and repeated cross-sectional data, are often used to estimate continuous time multistate models. However, relatively little is known about how accurate the estimates are.

In this thesis, methods are reviewed for deriving parameters of continuous-time Markov models from panel surveys and repeated cross-sections. The performance and accuracy of these methods are obtained in a simulation study.

The results of the simulation study for the panel data indicate that the method built on the EM (expectation-maximization) algorithm and the maximum likelihood estimates (MLE) constructed for `complete' data shows a bias in the estimates. The bias depends on the model specification and on the sampling scheme. Other methods based on MLE from panel data do not show a difference between the expected values of the transition rates and the true values. The results for the repeated cross-sections reveal that the chosen method can be used with age-varying parameters for models with two or three states.

The two best methods for panel data are applied to the US Health and Retirement Study to describe disability dynamics at old age. The two methods produce similar results. Smoothing the transition rates generates better results than the piecewise constant approximation. The estimation results confirm findings from the literature.