Cancer
Collider Bias in Estimating the Effects of Pre-Incident-Cancer Exposures on Post-Incident-Cancer Outcomes Matthew M Coates* Matthew Coates Danica Anukam Zuo-Feng Zhang Onyebuchi A. Arah
Studies estimating the effects of exposures before incident cancer on post-incident-cancer outcomes, such as cancer progression, cancer recurrence, or cancer mortality, are restricted to people with incident cancer or pre-cancer malignancies, potentially creating selection bias. This study identifies examples of this bias in literature, explains the structural causal mechanisms underlying this bias with causal diagrams, and quantifies the potential impact of this bias through simulations. In a causal diagram, incident cancer is (i) a mediator between pre-incident exposure and post-incident cancer consequences and (ii) a potential collider between pre-incident exposure and risk (e.g., genetic, infectious, or environmental) factors that lead to both higher cancer risks and outcomes of interest such as tumor aggressiveness. Estimating the direct effect of pre-incident exposures on post-incident outcomes (e.g., the effect of cancer screening on cancer survival that conditions on cancer diagnosis) requires adjustment for this mediator-outcome confounding, even under a randomized exposure. We simulated data using causal diagrams and parameters informed by published literature to quantify the degree of bias in estimates of the effects of pre-incident exposures on post-incident outcomes for a set of exposures of interest identified from the literature. In simulations based on results from published studies, a combination of pre-incident exposure causing cancer and strong uncontrolled confounding of cancer incidence and the post-incidence outcome could create substantial, varying bias in the estimated effects. Studies estimating the effect of pre-incident exposures on post-incident outcomes among people with cancer should adjust for common causes of cancer incidence and subsequent cancer outcomes, use bias analysis to assess potential collider bias, or, if relevant, target alternative estimands that do not depend on measurement of such common causes.