Methods for analysing complex panel data using multilevel models with an application to the Brazilian labour force survey

Abstract

Data sets commonly used in the social sciences are often obtained by sample surveys with complex designs. These designs usually incorporate a multistage selection from a population with a natural hierarchical structure. In addition, these surveys can also be carried out in a repeated manner including a rotating panel design, which is a source of planned non-response. Unplanned non-response is also present in panel data in the form of panel attrition and intermittent nonresponse.

Methods are developed to handle this type of data complexity. These methods follow the Multilevel Model framework which is reviewed. Longitudinal growth curve models accounting for the complex data hierarchy are fitted. Recognizing the need to account for the complex correlation structure present in the data, multivariate multilevel models are then adopted. Alternative modified correlation structures accounting for the rotating sample design are proposed. Multivariate multilevel models are fitted utilizing these alternative correlation structures. In addition, models estimated using robust methods are compared with those estimated using standard methods.

A method for calculating a set of longitudinal sample weights that accounts for attrition is proposed. Models utilising the conditional sample weights and longitudinal weights are fitted using the Probability-weighted Iterative Generalized Least Squares (PWIGLS) estimation method. Furthermore, an extension to PWIGLS for multivariate multilevel models is developed. Models fitted through different estimation methods are compared and the best approaches are suggested.

Data from the Brazilian labour force survey, Pesquisa Mensal de Emprego (PME) are used. The PME has a complex sampling design that includes a multistage selection of the sample units and a rotating panel design characterised as 4-8-4. The methods developed are used to investigate the labour income dynamics of employed heads of households in the PME survey.

More information

Identifiers

ORCID for Peter W.F. Smith: orcid.org/0000-0003-4423-5410

Catalogue record

Export record