Statistical modelling of selective non-participation in health examination surveys

Abstract

Health examination surveys aim to collect reliable information on the health and risk factors of a population of interest. Missing data occur when some invitees do not participate the survey. If non-participation is associated with the variables to be studied, then the estimates based only on the participants cannot be generalised to the population of interest. In this case, the estimates have selection bias, which misleads the decision-makers. The purpose of this thesis is to develop statistical methods to reduce the selection bias in the cross-sectional data using additional data sources. The data, which we use, comes from the National FINRISK Study, and we aim to estimate the prevalences of self-reported daily smoking and self-reported heavy alcohol consumption. The sources of additional information are follow-up data consisting of hospitalisations and causes of deaths, and questionnaire data collected from the non-participants of health examination by contacting them again, called re-contact data. Follow-up data give indirect information after the follow-up period about the health behaviour of non-participants during the health examination while the re-contact data give information similar to the health examination survey. This thesis presents methods for utilising these sources of additional information. Multiple imputation has been applied for the use of re-contact data, and Bayesian statistical modelling has been implemented for the use of follow-up data. The thesis demonstrates that the use of additional data sources and these statistical methods leads to prevalence estimates for daily smoking and heavy alcohol consumption that are higher than those obtained from the participants only. Multiple imputation can be utilised for prevalence estimation if the re-contact data are available. Bayesian modelling is appropriate for the situation where re-contact data are not available but the follow-up data are and have follow-up period long enough to indicate about the differences between the participants and non-participants. This thesis presents means for reducing the selection bias caused by non-participation. It is important to reduce the magnitude of the bias for obtaining more reliable information for example to support decision making. The statistical methods used in this thesis can also be applied to other fields of research than in the health studies.