Causal Inference
Framing and Extending Immortal Time Bias with Augmented Causal Diagrams and Annotated Causal Estimands: Immune Time versus Immune Study Population Matthew M Coates* Matthew Coates Onyebuchi A. Arah
As the increasing availability of large-scale real-world datasets has enabled many studies of causal effects in observational settings that require careful study design, immortal time bias has garnered increased attention. Immortal time is time experienced by study participants during which they are immune to or are not at risk for the outcome. Target trial emulation is a practical strategy that can help avoid inducing immortal time bias by aligning eligibility, treatment assignment, and the start of follow-up for outcome ascertainment. However, this alignment of eligibility, treatment assignment, and the start of follow-up can be more generally described in terms of appropriate definition, timing, and measurement of the target population, treatment, and outcome. Using causal diagrams augmented with selection nodes, treatment misclassification, or both, we show that immortal time bias can occur in a study when the study population sample (so-called eligibility) definition, treatment classification (assignment), or both are based on the study sample, treatment, or both being defined after the start of the study (time zero). We then show how causal estimands can be annotated to see when immortal time bias may be a concern, such as when the study sample differs from the target population based on the timing of sample selection. Finally, we generalize the first type of immortal time bias as an immune study sample bias arising from sample selection based on post-time-zero status that differs from the target population at time zero. The second type of immortal time bias is generalized as immune-to-treatment-effect bias, which arises from treatment misclassification based on post-time-zero exposure assignment. We also investigate whether and how these two biasing mechanisms could co-occur. We conclude by examining the augmented causal structures and estimands that lead to doomed study sample bias versus doomed-to-treatment-effect bias.