In the analysis of attrition within large and complex longitudinal social science data, there are significant and powerful computational difficulties that are routinely encountered. This paper will present new computational solutions to these complex issues in the context of an analysis of changing attitudes to gender roles over the last twenty years.

The example is taken from the British Household Panel Study (BHPS) which has collected data on attitudes to gender roles biennially since 1991. The paper will focus upon one Likert item which explores the question: ‘A husband’s job is to earn money; a wife’s job is to look after the home and family’. These items comprise the five response categories: ‘Strongly Agree’, ‘Agree’, ‘Neither Agree Nor Disagree’, ‘Disagree’ and ‘Strongly Disagree’.
First, the cumulative logit or proportional odds model will be used to model the ordinal responses to this item. Residual heterogeneity will be handled by incorporating a respondent-specific random effect into the response model. Second, a binary logistic random effects model will be used to analyse dropout. Finally, we will fit a joint response-dropout model with correlated random effects which allows us to test for the association between the response and dropout processes.
As part of the session, we will demonstrate the efficacy of the statistical software package SABRE (http://sabre.lancs.ac.uk ) (see Berridge and Crouchley, 2011) in fitting these complex statistical models to large datasets, and in incorporating a large number of explanatory variables.