Multi-state models for clustered duration data: an application to workplace effects on individual sickness absenteeism

In this paper we specify and estimate three state duration models of work, sickness and exit from the job to explain individual absenteeism behaviour of primary school teachers. There is a large variation of sickness absenteeism records across schools and absenteeism records of workers within a school appear to be related. This clustering of individual absenteeism data may to a large extend be caused by workplace effects. Since it will be difficult to fully capture workplace effects with observed characteristics of the workplace, we also account for unobserved workplace effects in the models. The most flexible specification allows for non-parametric baseline hazards that differ per exit rate and workplace. A stratified partial likelihood approach is used to estimate the regression coefficients of this model. Conditional on these estimates we recover fixed unobserved workplace effects and semi-parametric baseline hazards in order to detect the causes for the observed variation and clustering in the data.