Transporting Effects across Populations

An effect measured in one population need not equal the effect in another. The notebooks make the difference explicit by separating subgroup effects from the subgroup mix in the study and target populations.

The central relation

For discrete groups, the target effect can be written as

\tau^{T}=\sum_g \pi_g^{T}\tau^{S}(g).

$g$ labels groups that modify the effect, $\tau^{S}(g)$ is the effect learned in group $g$ of the study population, and $\pi_g^{T}$ is that group's share in the target population. The superscripts $S$ and $T$ refer to the study and target populations.

Reweighting effects for a target population

The recovery example shows how target population weights change the transported effect. The public-data experiment compares study and target summaries and makes poor support visible before reweighting.

Study and target population weights across effect-modifying groups. — Transport changes the weight assigned to each subgroup effect. Groups common in the target population receive more influence in the target estimate.

Study and transported effect estimates in a public-data example. — The public-data comparison shows the change produced by target weighting. It should be interpreted only where the study contains adequate examples of the target groups.

Transport requires both overlap and a defensible account of the variables that change the effect. Reweighting cannot recover a target group that is absent from the study.

Notebooks


1. Finite-Stratum Transport
2. Checking support before reweighting
3. Contrasting successful transport with failure