Estimating effects on changes in health outcomes using scores – An alternative visualization

Abstract

Epidemiologists often try to estimate the effect of an exposure on change in health, a goal that applies to, at least partially, many situations. For example, they may study the effects: of education on deterioration of cognitive function in later life; of e-cigarette use cessation on change in weight; or of waist circumference on functional and physical decline. Here we focus on continuous outcomes and change between two time points, but our comments also are relevant to trajectories and to changes in dichotomous measures.Here we comment on results of Glymour et al. concerning studies of the effect of baseline exposure on change scores and in particular, the advisability of adjustment for the baseline value of the outcome. First, we identify relevant causal relationships and summarize them in a causal graph. Our graph includes more variables than, and differs from the graph used by Glymour et al. for their main results. These causal relationships and graph we use lead to the same main conclusions as those reached by Glymour et al., but the causal graph we use may often better describe certain aspects of the causal structure. Our alternative conceptualization includes additional, relevant variables and effects and may be easier for some to construct and interpret. Second, we suggest another alternative conceptualization and model that is compatible with the first one. Third, we discuss several additional considerations and issues, like the horse-racing effect and selection bias. We show that these effects can be included in the causal graphs.