Interpreting time-related trends in effect estimates.

Abstract

This paper reviews the sources of apparent time trends in effect. Apparent changes in effect may arise from changes in covariate distributions, background rates, exposure distribution, measurement quality, or selection factors. As with time trends in rates, time trends in effect must have at least one of these sources, since time itself has no effect. If background incidence is changing, however, time trends in effect become dependent on choice of effect measure, and interpretation must take this into account. Evidence that a trend arises from age-, cohort-, or period-related phenomena can indicate the relative plausibility of different explanations of the trend. Conversely, the relative plausibility of each explanation may indicate whether the trend is most appropriately viewed over the axis of age, birth cohort, or calendar time. Nevertheless, studies of short duration relative to an apparent trend (such as most case-control studies) must invoke strong assumptions to justify focusing on a particular axis. Illustrations are given from studies of electronic fetal monitoring and of smoking and lung cancer.