Cohort studies with low baseline response may not be generalisable to populations with different exposure distributions.

Several advocates of mega-cohorts in epidemiology recently stated that high baseline response needs not be a driving scientific factor. We illustrate how cohort studies with low baseline response can become less generalisable than cohort studies with reasonable high baseline response. Assuming exposure prevalences of 32%, and 39% for former, and current smoking in the general population, we postulate that 10% in each group have a specific genotype G that makes them susceptible for the effect of smoking on the outcome (synergism between genotype and smoking). The existence and the mechanism of this synergism is unknown to the investigator. The parameter of interest is the average effect of smoking on the outcome expressed as relative risk (RR). We consider three different scenarios. In scenario 1, the RR for former and present smokers is 50 and 100 for subjects with G, translating into RR of 5.9 and 10.9 at population level. In scenario 2, the according figures are 10 and 20 for subjects with G and 1.9 and 2.9 at the population level, while in scenario 3, the according numbers are 2 and 5, and 1.1 and 1.4, respectively. We assume a differential baseline response by genotype G with overall baseline responses of 10% (low) and 50% (high). In all scenarios, we observe generalisability for a high baseline response. In contrast, for low baseline response the RRs for former and current smokers lack generalisability with loss of dose response relationship leading to 2.58 and 2.09 in scenario 1, to 1.29 and 1.21 in scenario 2, and in scenario 3 with RRs very close to unity (1.03 and 1.04 respectively). RR estimates may lack generalisability and dose response relationships even may be inverted if baseline response is low. Thus, even mega-cohorts should strive for a reasonable high baseline response.