A1458
Title: Causal optimal transport of treatment effect to a target population with limited individual-level data
Authors: Tat Thang Vo - University Paris Est Creteil (France) [presenting]
Antoine Chambaz - Universite Paris 5 Rene Descartes (France)
Abstract: The transportability of empirical findings to new environments, settings, or populations is essential in most scientific investigations. One practical challenge of standard methods for transportability, however, is that they require the individual-level data on outcome, treatments, and case-mix characteristics to be fully accessible in the source study, along with individual-level data on case-mix characteristics for a random sample from the target population. In practice, data sharing is often subject to administrative barriers and privacy concerns, e.g., a pharmaceutical company may have access to individual-level data from its own study but only aggregate-level data for the target population, such as from a competitor's trial. In such a scenario, state-of-the-art methods generally rely on parametric G-computation or inverse weighting to adjust for the case-mix difference between the study and the target population. Unsurprisingly, the resulting effect estimates from these approaches can be severely biased if the modeling assumption imposed on the nuisance outcome and/or weight model is violated. Novel methods are developed for transportability that fully circumvent the need for strong parametric assumptions when there is restricted access to individual-level data, using computational optimal transport. The new methods allow the use of flexible data-driven methods to estimate nuisance parameters and rely on semi-parametric theory for valid asymptotic inference.
