A multiple imputation-based method to correct non-response bias in prevalence surveys

Non-response introduces uncertainty in the results of most epidemiological surveys and it can bias the final estimates of prevalence. In the present paper a new method is proposed to deal with the non-response problem, when the survey is performed through subsequent stages, taking non-responders at one stage as target population for the next stage. While the existing methods are based on the a priori assumption of a specific prevalence trend across following stages, the new approach starts by analysing the non-response generating process and by choosing the model which best fits to the data available from responders. Using the maximum likelihood three hierarchic models are tested: 1) the no-bias model; 2) the two parameter model, which assumes two prevalence values, one for responders to the first stage and one for the rest of the sample; 3) the regression model, which hypothesizes a monotonous trend in prevalence across stages. In order to reflect the degree of uncertainty inherent to non-response on interval estimate width, missing data variability was mimicked by random multiple imputations. An example is also included, to compare point and interval estimates under the new approach with the estimates obtained with two other well-known methods.