Incomplete data in event history analysis.
2015-11-19T08:55:42Z (GMT) by
Incomplete data present a serious problem in the modelling of event histories. Two particular forms of incompleteness are in evidence for data of this form. The first is due to recording of event times in interval-censored form. For single non-repeatable events this can be accommodated by using methods for modelling grouped survival times, such as those of Prentice and Gloeckler (1978) and Finkel- stein (1986). The other, more serious, problem relates to incomplete recording of follow-up measurements which would typically be included as time-dependent covariates in survival models. A number of methods exist for handling incomplete data. These include multiple imputation for variables subject to incompleteness and the application of iterative algorithms such as EM and the data augmentation algorithm. In this thesis, a method for handling both these types of incompleteness is derived based on multiple imputation combined with an adaptation of Finkelstein's method to handle time-varying covariates. This method is then investigated via Monte Carlo simulation and applied to data arising from the annual screening of those aged 75 years and over in the town of Melton Mowbray, as performed through the local general practice. Its performance is compared with that of more traditional approaches to modelling data collected in studies of this type. It is shown that parameter estimates can be considerably affected by the choice of approach to modelling. Whilst there are some problems with the implementation of this technique, particularly with reference to the model for the multiple imputation of the repeated risk factor values, it shows promise for application to studies of this form, particularly if combined with improved models for multiple imputations. The data from the annual screenings are assumed missing at random, but the techniques used could be extended to cover non-ignorable missing data mechanisms of known form.