Primary Submission Category: Sensitivity Analysis
Quantifying and interpreting reverse causation bias in epidemiological studies with sensitivity analysis
Authors: Jeremy Brown,
Presenting Author: Jeremy Brown*
In epidemiological studies reverse causation occurs when we want to estimate the effect of the exposure on the outcome, but the outcome affects exposure. This is a particular concern in cross-sectional and ecological studies, where temporality of exposure and outcome is not clear, therefore, these designs are rarely used for causal inference. In study designs more commonly used for causal inference, such as cohort studies, the temporal order of exposure and outcome can typically be ascertained and therefore reverse causation prevented. However, there may still be reverse causation bias if the outcome is misclassified, for example if it was present but undiagnosed at baseline. Using directed acyclic graphs we describe the structure of this bias, which is structurally analogous to bias due to a misclassified confounder. Methods to mitigate this bias are available, but are not always feasible. Bias formulas have not been commonly used, but have been derived and under certain assumptions are equivalent to unmeasured confounding bias formulas. We describe these formulas and, as an example, apply them to the results of a published case-control study examining the effect of oral contraceptives on uterine cancer. Applying plausible bias parameters to the observed odds ratio 2.35 (95% CI 1.29-4.26) led to bias-adjusted estimates ranging from 1.96 (95% CI 1.08-3.55) to 1.24 (95% CI 0.68-2.24) indicating the potential for reverse causation amongst other biases to alter findings.