Causal Inference
Characterizing the Causal Structure of Lead Time Bias in Screening Studies with Causal Diagrams and Causal Estimands Matthew Coates* Matthew Coates Coates Department of Epidemiology, Fielding School of Public Health, University of California, Los Angeles (UCLA); Practical Causal Inference Lab, UCLA
Lead-time bias refers to the phenomenon that comparing post-diagnosis survival time between screen-detected and symptom-detected cancers can create the appearance of longer survival in the screened group, even without a true survival benefit, because screening reduces the lead time between onset and diagnosis. In prior literature, lead-time bias has been discussed as confounding, as information bias due to measurement error, and as selection bias. Despite being discussed in great detail, particularly in the cancer literature, lead-time bias has not been commonly described using tools from formal causal inference frameworks, such as counterfactual language and causal diagrams. Using causal diagrams and causal estimands that specify the definition, timing, and measurement of the target population, exposure, and outcome, we describe the mechanisms underlying lead-time bias. Randomized trials can be used to avoid lead-time bias through a combination of (1) creating an often-implicit target population that includes people without the disease, (2) balancing time since disease onset between treatment and control arms, and (3) aligning the start of follow-up in the arms. Some analyses susceptible to lead-time bias, including trial-based analyses, restrict the study population to individuals who are diagnosed, perhaps implying a target population of those who are diagnosed who may benefit from screening. Identifying and estimating the effect of screening in this target population, defined after intervention, is challenging and requires data beyond those diagnosed. Explicitly defining survival since disease onset as the outcome clarifies the measurement error that arises when using time since diagnosis and which generates lead-time bias. We use causal diagrams to demonstrate that multiple other biasing mechanisms, including confounding and selection bias, often co-occur in studies subject to lead-time bias.