Primary Submission Category: Generalizability/Transportability
Transportability sharp bounds
Authors: Guilherme Duarte,
Presenting Author: Guilherme Duarte*
Transportability is one of the biggest challenges in causal inference. Researchers are often familiar with techniques to estimate quantities such as the ATE, from experimental/observational data, but they still struggle with generalizing estimates from one environment to another. For example, they might run an experiment in Los Angeles and calculate an effect, but they still hesitate over which assumptions allow them to know what the effect would be in New York City without having to rerun the experiment there. Current papers emphasize solutions that indicate if transportation is permitted given particular structural assumptions. Nonetheless, limitations to this strategy are known as it puzzlingly says cases are not transportable when they in fact are, and even when a precise solution does not exist, it fails to provide informative bounds. Here I propose a general and complete algorithm to always provide upper and lower sharp bounds to transportable estimates. Among the inputs, one states two different environments (for instance, Los Angeles and New York City), introduces existing data to both (e.g., baseline mortality rates in each city), invariance assumptions, and an original quantity they want to transport. The algorithm, then, outputs sharp bounds for that estimate in the second environment. If both bounds converge, then one would say that the quantity is precisely transportable. To show the applicability of this method, two examples of the literature are analyzed.