Propensity score weighting across counterfactual worlds: longitudinal effects under positivity violations

Abstract

When examining a contrast between two interventions, longitudinal causal inference studies frequently encounter positivity violations when one or both regimes are impossible to observe for some subjects. Existing weighting methods either assume positivity holds or produce effects that conflate interventions' impacts on ultimate outcomes with their effects on intermediate treatments and covariates. We propose a novel class of estimands -- cumulative cross-world weighted effects -- that weights potential outcome differences using propensity scores adapting to positivity violations cumulatively across timepoints and simultaneously across both counterfactual treatment histories. This new estimand isolates mechanistic differences between treatment regimes, is identifiable without positivity assumptions, and circumvents the limitations of existing longitudinal methods. Further, our analysis reveals two fundamental insights about longitudinal causal inference under positivity violations. First, while mechanistically meaningful, these effects correspond to non-implementable interventions, exposing a core interpretability-implementability tradeoff. Second, the identified effects faithfully capture mechanistic differences only under a partial common support assumption; violations cause the identified functional to collapse to zero, even when the causal effect is non-zero. We develop doubly robust-style estimators that achieve asymptotic normality and parametric convergence under nonparametric assumptions on the nuisance estimators. To this end, we reformulate challenging density ratio estimation as regression function estimation, which is achievable with standard machine learning methods. We illustrate our methods through analysis of union membership's effect on earnings.