Time-varying confounders affected by prior exposure: when do they require causal inference methods?

摘要

Background: With the increasing popularity of causal directed acyclic graphs and "causal inference methods" (also called g-methods), epidemiologists are paying more attention to time-varying confounders, particularly those affected by prior exposure. Such variables may cause bias in a standard analysis, whether or not the model adjusts for them. Aims: To clarify the situations in which causal inference methods are needed. Methods: Directed acyclic graphs help to illustrate two different scenarios that depend on the causal question of interest. In one, the exposure of interest is a cumulative measure that takes into account prior exposure history. In the other, the exposure of interest is a current exposure (measured in the same time interval as, or just prior to, the outcome), while prior exposure history is merely a confounder. Results: When the causal question of interest involves a cumulative exposure measure, causal inference methods are required because results from a standard analysis will be biased. However, when the causal question concerns a current exposure only, both causal inference methods and standard analysis should be unbiased. Conclusion: Researchers should always adjust for time-varying confounders affected by prior exposure; the causal question of interest determines which methods can achieve a correct adjustment.