Abstract

We investigate methods for estimating a rich set of sickness absence indicators based on the Netherlands Working Conditions Survey (NWCS). The main target population of interest is the population of employees in the Human Health and Social Work Activities branch. For each indicator, annual figures for a detailed set of subpopulations defined by several cross-classifications of health branches, regions, age classes and firm size classes are to be estimated. As the NWCS is not large enough to provide reliable direct estimates at such levels of detail, small area estimation methods are employed. In particular, we use time-series multilevel models to borrow strength from NWCS data over the last ten years as well as over the cross-sectional domains of interest. The models used are unit-level models for continuous, binary, count and categorical data, depending on the indicator. The models contain multiple random effect terms for the cross-classifications of interest, including interactions with independent or random walk temporal effects. The fixed effects include those for auxiliary variables used in the NWCS survey weighting. The models are estimated in a Bayesian simulation framework using data augmentation and Gibbs sampling. For each indicator the complete set of small area estimates is benchmarked to the national figure based on the NWCS weights. We discuss the efficiency of our method, both from a computational and statistical point of view. Results are also compared to estimates based on cross-sectional multilevel models applied separately to the annual datasets.