An assessment of the utility of a Bayesian framework to improve response propensity modes in longitudinal data

Response propensity (RP) models are widely used in survey research to analyse response processes. One application is to predict sample members who are likely to be survey nonrespondents. The potential nonrespondents can then be targeted using responsive and adaptive strategies with the aim of increasing response rates and reducing survey costs. Generally, however, RP models exhibit low predictive power, which limits their effective application in survey research to improve data collection. This paper explores whether the use of a Bayesian framework can improve the predictions of response propensity models in longitudinal data. In the Bayesian approach existing knowledge regarding model parameters is used to specify prior distributions. In this paper we apply this approach and analyse data from the UK household longitudinal study, Understanding Society (first five waves) and estimate informative priors from previous waves data. We use estimates from RP models fitted to response outcomes from earlier waves as our source for specifying prior distributions. Our findings indicate that conditioning on previous wave data leads to negligible improvement of the response propensity models’ predictive power and discriminative ability.