Semiparametric survival models for routine register data
Routine registers offer researchers opportunities to carry out studies of covariate effects on lifetimes of rare diseases otherwise infeasible because of the large cohorts required. Familial relationships necessary for analysis of environmental or genetic factors can be identified by record linking. The vast amount of data and clustering of related individuals pose statistical challenges. As most statistical information is associated with the cases, an estimator based on a sample where cases are overrepresented can drastically reduce the sample size with only a minor loss of efficiency.
This thesis concerns regression of clustered cohort sampled survival data within the broad class of semiparametric transformation models. This class includes the proportional hazards and proportional odds models as special cases. Correlations within clusters are modeled by random effects.
We derive consistency and asymptotic normality of a weighted maximum likelihood estimator and provide a consistent estimator of its asymptotic variance. A likelihood ratio test for regression coefficients is also proposed. The method is shown to perform well on simulated data and is illustrated by application to a study on cardiovascular diseases among Swedish men.