Non-response bias in studies of survey data with an application to residential mobility

Paul Clarke, University of Bristol
Fiona Steele, University of Bristol
Elizabeth Washbrook, University of Bristol

We consider the issue of non-response or attrition bias in binary outcome models when the response mechanism takes a particular form: the probability an outcome is observed depends causally on the realized outcome. Our analysis is motivated by the study of residential mobility, in which movers are more likely to drop out of the study because of loss of contact, but is also relevant to applications such as the effects of a treatment of health outcomes. A simple modelling framework incorporates the response mechanism of interest and is contrasted with the more standard two-equation selection model approach. We use simulations to characterize the nature of the biases and the detection of non-ignorable non-response under different estimation approaches. We explore the sensitivity of the estimates to the normality assumption and the imposition of exclusion restrictions. The model is applied to an analysis of residential mobility in the British Household Panel Study.