Generalizing JTPA job-training effects to a national target: TRACE narrows the confidence interval by 60% and says whether baseline mismatch or follow-up loss is to blame

A generalizability diagnostic for survival outcomes now handles two forms of underrepresentation at once: a subgroup's covariate profile being rare in the trial relative to the target, and differential censoring eroding that subgroup over follow-up. Earlier diagnostics either conflated these two mechanisms or applied only to the baseline case, so a policymaker using JTPA findings could not tell which target subgroups the trial generalizes to reliably, why generalization fails elsewhere, or when during follow-up precision is lost. In a case study generalizing the JTPA Study to U.S. JTPA-eligible adults, restricting to reliably generalizable subgroups narrows the 95% confidence interval for the 24-month employment effect by 60% relative to the original target-population average treatment effect, with the two intervals covering different estimands.

The method localizes variance inflation to specific subgroups and specific follow-up intervals, and decides whether each subgroup's problem is baseline mismatch or censoring. It does so by decomposing the asymptotic variance of the efficient generalizability estimator into two terms, of which only a sampling-weight term and a censoring-weight term can diverge as representation weakens; the larger term identifies the responsible channel. On synthetic benchmarks, TRACE recovers the planted underrepresented regions, attributes each to its mechanism, and locates the follow-up windows driving time-varying underrepresentation; in the JTPA case, the censoring intervention activates on only a small number of trial person-intervals, indicating baseline mismatch as the main source.

The search solves the inner censoring-intervention choice separately from the outer baseline trim, making the joint optimization tractable over a tree ensemble. For each individual, the inner problem reduces to an affine lower envelope over an intervention budget, precomputed once by a parametric dynamic program so that each later query is a walk along the envelope. The outer level scores splits against the global variance objective using a Rashomon ensemble of near-optimal trees, and the output is an auditable four-region partition: reliably generalizable, censoring-depleted, baseline-underrepresented, and doubly underrepresented.

The demonstrated scope is generalization of survival outcomes over discrete follow-up intervals, covering synthetic benchmarks and one case study generalizing the JTPA Study to U.S. JTPA-eligible adults represented by the CPS; the method outputs a diagnostic partition and does not adjust the point estimator.

TRACE is diagnostic rather than corrective and does not adjust the point estimator; the trimmed estimand changes the inferential question rather than recovering the original target-population average treatment effect, and its point estimate and confidence interval condition on the selected trim without accounting for selection uncertainty. When both mechanisms are large on the same units, how credit is split depends on the objective.