A sensitivity analysis of secular trends in risk factors and mortality based on cohort studies

Because of a "healthy responder effect," secular trends in mortality based on cohort studies may be biased if based on responders only. Because responders are selected on the basis of their health at study entry, subjects just entering a study are not comparable with subjects who have been in the study for several years. The result may be an artificial increase in mortality, which impedes analyzing the effect of secular trends in risk factors on mortality. The objective of this paper is to suggest a solution by using data on nonresponders and applying a sensitivity analysis. We illustrate this solution with data on trends in smoking prevalence and all-cause mortality based on a large Danish cohort study with 19 years of complete follow-up on responders and nonresponders. Secular trends in mortality based on the whole sample vs responders only illustrated that results based on responders were biased. In a sensitivity analysis, the observed person-years of nonresponders were distributed among six categories of persons with respect to smoking behavior (never-smokers; ex-smokers; noninhaling current smokers; and current smokers of 1-14, 15-24, and > or =25 gm tobacco per day) according to preset assumptions regarding smoking habits. The observed deaths among nonresponders were then distributed on the six smoking categories according to relative risks derived from a Poisson regression analysis among responders. This procedure allowed us to study the effect of adjustment for smoking on the unbiased secular trend in mortality. By applying different assumptions regarding smoking habits among nonresponders, we explored the effect of the assumptions on the adjusted secular trend in mortality. We conclude that secular trends in mortality based on responders in a cohort study are likely to be biased. If complete follow-up on nonresponders is available, this method could prove useful in other cohort studies.