Estimating models for panel survey data under complex sampling

Complex designs are often used to select the sample which is followed over time in a panel survey. We consider some parametric models for panel data and discuss methods of estimating the model parameters which allow for complex schemes. We incorporate survey weights into alternative point estimation procedures. These procedures include pseudo maximum likelihood (PML) and various forms of generalized least squares (GLS). We also consider variance estimation using linearization methods to allow for complex sampling. The behaviour of the proposed inference procedures is assessed in a simulation study, based upon data from the British Household Panel Survey. The point estimators have broadly similar performances, with few significant gains from GLS estimation over PML estimation. The need to allow for clustering in variance estimation methods is demonstrated. Linearization variance estimation performs better, in terms of bias, for the PML estimator than for a GLS estimator.