Causal Inference
The transportability of the causal effect of an infectious disease intervention to a target population Christopher Boyer* Christopher Boyer Juan Gago
Infectious disease outbreaks are dynamic events characterized by repeated and often rapid shifts in infection patterns, population immunity, host behaviors, and antigenic evolution. This changing landscape makes it challenging to monitor the effectiveness of interventions to reduce or modify the course of disease, as new circumstances require continual re-estimation of their impact. However, direct re-estimation, through additional trials or observational studies, may be impractical as these studies may be hard to justify (in the case of controlled trials) or they may take too long and incur additional risk of bias (in the case of observational studies). In this case, a possible research goal is to “transport” the causal effect estimated in an initial trial to the new setting.
Recent methodological developments for analyses that extend –generalize or transport– inferences from a trial to a target population aim to accomplish this goal through techniques that standardize the data distribution observed in the trial to the covariate distribution in the target population, assuming that all relevant effect modifiers are measured. However, infectious disease outbreaks introduce additional complexities that often violate the core identifiability assumptions that motivate these techniques, even when all effect modifiers have been accounted for. Specifically, the transmission of an infectious disease between individuals introduces the possibility of interference, whereby an intervention affecting one individual’s risk of infection also affects the risk of their contacts. Furthermore, due to the dynamic nature of an outbreak, the structure and degree of post-baseline exposure are likely to vary between the trial and the target population.
In this study, we examine the transportability of infectious disease interventions from a source to target population and discuss the conditions under which relevant causal effect estimands are identifiable using existing approaches. We also propose alternative strategies based on transporting so-called per exposure effects, i.e. effects under hypothetical interventions on exposure. We derive estimators and illustrate our methods via an agent-based simulation of an outbreak in a source and target population.