How to investigate and adjust for selection bias in cohort studies

That conditioning on a common effect of exposure and outcome may cause selection, or collider-stratification, bias is not intuitive. We provide two hypothetical examples to convey concepts underlying bias due to conditioning on a collider. In the first example, fever is a common effect of influenza and consumption of a tainted egg-salad sandwich. In the second example, case-status is a common effect of a genotype and an environmental factor. In both examples, conditioning on the common effect imparts an association between two otherwise independent variables; we call this selection bias.

In order to investigate potential selection bias in population-based cohort studies, participants (n = 28098) and non-participants (n = 40807) in the Malmö Diet and Cancer Study (MDCS) were compared with regard to cancer incidence and mortality. MDCS participants were also compared with participants in a mailed health survey with regard to subjective health, socio-demographic characteristics and lifestyle. Cancer incidence prior to recruitment was lower in non-participants, Cox proportional hazards analysis yielded a relative risk (RR) with a 95% confidence interval of 0.95 (0.90-1.00), compared with participants. During recruitment, cancer incidence was higher in non-participants, RR: 1.08 (1.01-1.17). Mortality was higher in non-participants both during, 3.55 (3.13-4.03), and following the recruitment period, 2.21 (2.03-2.41). The proportion reporting good health was higher in the MDCS than in the mailed health survey (where 74.6% participated), but the socio-demographic structure was similar. We conclude that mortality is higher in non-participants than in participants during recruitment and follow-up. It is also suggested that non-participants may have a lower cancer incidence prior to recruitment but a higher incidence during the recruitment period.