Methods/Statistics
Using Transportability Methods for Subgroup Analyses Quoa Her* Michael Webster-Clark Anthony Matthews Robert Platt
Subgroup analyses are a key part of epidemiologic research. Unfortunately, they often yield imprecise treatment effect estimates (or false positives for heterogeneity) because the analytic populations are restricted to members of the subgroup; those not in the subgroup do not contribute information.
While this works when A) there is no difference in the true treatment effect between members and non-members or B) the true treatment effect differs between non-members and members conditional on measured covariates, these two cases are not exhaustive. The true treatment effect may differ marginally between members and non-members but be equal if we account for measured covariates that modify the treatment effect of interest. In such cases, methods developed to transport estimates to external target populations can reduce variance in subgroup estimates.
We explored this using data from the PRIME trial of panitumumab on one-year progression free survival (PFS) in patients with metastatic colon cancer for the subgroup of Hispanic individuals (a small group with a differing prevalence of a key tumor gene). We 1) weighted non-Hispanic White participants to resemble Hispanic participants in terms of key effect measure modifiers, 2) combined unweighted Hispanic participants and weighted non-Hispanic White participants in one data set, and 3) analyzed the weighted cohort, obtaining 95% confidence limits from percentiles of 1,000 bootstrap iterations. While a Hispanic-only analysis estimated a decrease in PFS at one-year of 17% (95% C.I. -8.8%, 45%), an analysis including reweighted non-Hispanic Whites was much more precise (8.7%, 95% CI -5.3%, 22%) while still differing from the full population estimate (-1.0%, 95% CI: – 7.5%, 5.9%).
While this method is not suitable for every subgroup (e.g., it would create bias if used to study the tumor gene that directly modified the treatment effect), it can improve precision without ignoring other measured effect measure modifiers.