Case-control studies are used in epidemiology to try to determine variables associated with a disease, by comparing those with the disease (cases) against those without (controls). Participation rates in epidemiology studies have declined over recent years, particularly in the control group where there is less motivation to participate. Non-participation can lead to bias and this can result in the findings differing from the truth.
A literature review of the last nine years shows that non-participation occurred in published studies as recently as 2015, and an assessment of articles from three high impact factor epidemiology journals concludes that participation bias is a possibility which is not always controlled for. Methods to reduce bias resulting from non-participation are provided, which suit different data structures and purposes. A guidance tool is subsequently developed to aid the selection of a suitable approach. Many of these methods rely on the assumption that the data are missing at random. Therefore, a new solution is developed which utilises population data in place of the control data, which recovers the true odds ratio even when data are missing not at random.
Chain event graphs are a graphical representation of a statistical model which are used for the first time to draw conclusions about the missingness mechanisms resulting from non-participation in case-control data. These graphs are also adapted specifically to further investigate non-participation in case-control studies.
Throughout, in addition to hypothetical examples and simulated data, a diabetes dataset is used to demonstrate the methods. Critical comparisons are drawn between existing methods and the new methods developed here, and discussion provided for when each method is suitable. Identification of factors associated with a disease are crucial for improved patient care, and accurate analyses of case-control data, with minimal biases, are one way in which this can be achieved.
