Multiple spells experienced by the same unit are often analysed as single spells under a number of restrictive assumptions. If these assumptions do not hold, the estimates of model parameters may be biased and the estimates of their standard errors are smaller than they should be.
We investigate some modifications to handling multiple spells observed in a longitudinal survey. One modification is a design-based approach which treats an individual as a cluster of spells, bringing an extra stage of clustering into the inferential process. Another modification intervenes into the model by allowing a flexible specification of the baseline hazard using dummy variables for each occurrence of the spell to account for possibly differential dependence of hazard on the duration and order of successive spells; it also allows for differential effects of the covariates on the spells of different orders. The estimation of parameters and their standard errors is done using a weighted likelihood (Chambless and Boyle, 1985) and a weighted partial likelihood (Binder, 1992). The new approaches are compared using data from the Canadian Survey on Labour and Income Dynamics.
|