Background: Estimation of non-response bias by modelling prevalence as a function of the number of mailings required to achieve a response, or of the cumulative response, has been advocated, but the models have not incorporated age and sex, differential response rates by age and sex, or season of response.
Methods: The effect on age-sex standardized prevalence of estimating non-response bias using a variety of models was investigated using data on nine symptom and medication questions from 13,007 subjects in the three English centres of the European Community Respiratory Health Survey. Comparison was made of goodness of fit and the prediction of responses in a 25% follow-up sample with the observed values.
Results: Despite low response rates in Cambridge and significant decreases in prevalence with additional mailings or increasing cumulative response in Norwich, there were only small effects on estimated age-sex standardized prevalences. No model was consistently better for any centre or question.
Conclusions: The models are useful for exploring the sensitivity of estimated prevalence to non-response bias, but should be used with caution to adjust estimates. Ideally first mailings should be staggered over the whole year so that mailing and season are not confounded, and sufficient mailings or other contacts carried out for the whole sample to ensure a high response rate.