The use of a non-response study in retrospect

The authors report about a study on the use and possibilities of a non-response study in retrospect. A non-response study is meant to control for systematic differences between responders and non-responders in having some property. If so there is selection bias. The problem investigated is: Which kinds of non-response studies in retrospect are possible and useful and how important are time trends? Three scenarios for non-response studies are investigated. These scenarios lead to three estimators of the fraction of the population having the property. The first scenario is not to perform a non-response study. The first estimator uses only the original information from the responders and doesn't control for selection bias. The second scenario is to perform a study in retrospect on non-responders only. The second estimator controls for selection bias but not for time trends. The third scenario is to perform a study in retrospect on both responders and non-responders. The third estimator controls for selection bias and time trends. It is concluded that a non-response study being performed in the same period as the main study is prefered to a study in retrospect, and that control for selection bias isn't useful in case of an extreme ratio between the numbers of responders and those of re-examined non-responders. Because of this last conclusion it was decided not to perform a non-response study in case of the 1991 gastro-intestinal diseases population study in the Netherlands.VHIHIG