Confounders and intermediaries in case-control study designs: a strategy for distinguishing between the two when measured using the same variable

Abstract

Purpose: An intermediary falls within the exposure-outcome pathway and is distinct from a confounder. In case-control studies, it may be difficult to discern between the two when both are measured by the same variable. Using data from a study on the effects of antipsychotic initiation on risk of death among older adults, where hospital use is both a confounder and intermediary, we illustrate the bias introduced when this distinction is overlooked and propose a modified exposure classification strategy to mitigate this.

Methods: We identified 5391 cases and 25,937 controls. Three analyses were completed: traditional analytic adjustment including hospital use (full), traditional analytic adjustment excluding hospital use (reduced) and exposure classification incorporating hospital use prior to antipsychotic initiation (extended).

Results: The unadjusted odds ratio (OR) was 2.8 (95% confidence interval (CI) 2.1-3.8). Full and reduced analytic adjustment resulted in ORs of 0.8 (95% CI 0.6-1.2) and 1.4 (95% CI 1.0-1.9), respectively. The extended exposure classification strategy produced an OR of 1.4 (95% CI 0.9-2.1) among those without hospital use prior to antipsychotic initiation.

Conclusions: Full analytic adjustment resulted in a biased estimate of effect. The extended exposure analysis differentiated between hospital use that occurred prior (confounder) and subsequent (intermediary) to antipsychotic initiation. This strategy may overcome the limitations of analytic adjustment alone.