Heterogeneity assessment in causal data fusion problems
Previous works have formalized the conditions under which findings from a source population could be reasonably extrapolated to another target population, the so-called "transportability" problem. While most of these works focus on a setting with two populations, many recent works have also provided the identifiability of a causal parameter when multiple data sources are available, under certain homogeneity assumptions. However, we know of little work examining transportability when data sources are possibly heterogeneous, e.g. in the distribution of mediators of the exposure-outcome relation. The presence of such heterogeneity generally invalidates the transportability assumption required in most of the literature. In this paper, we will propose a general approach for heterogeneity assessment when estimating the average exposure effect in a target population, with mediator and outcome data obtained from multiple external sources. To account for heterogeneity, we define different effect estimands when the mediator and outcome information is transported from different sources. We discuss the causal assumptions to identify these estimands, then propose efficient semi-parametric estimation strategies that allow the use of flexible data-adaptive machine learning methods to estimate the nuisance parameters. We also propose two new methods to investigate sources of heterogeneity in the transported estimates. These methods will inform users about how much of the observed statistical heterogeneity in the transported effects is due to the differences across data sources in: 1) conditional distribution of mediator variables, and/or 2) conditional distribution of the outcome. We illustrate the proposed methods using four sites that were part of the Moving to Opportunity Study, which was an experiment that randomized housing voucher receipt to participating families living in public housing.
READ FULL TEXT