A new partial identification system calculates causal bounds for modified treatment policies when exposure data contains severe gaps. Standard causal methods assume complete data coverage, pretending that every treatment combination exists across all patient backgrounds. The new approach isolates well-documented data regions and bounds unknown zones by shifting reference anchors inward from boundary edges.
Standard projections force outside data points directly against the sharp outer rim of observed evidence. This edge crowding acts like balancing weights on a thin fence line, which breaks the smooth mathematical curves needed for stable estimation. To repair this flaw, the new procedure steps anchors away from the perimeter into the interior space. By enforcing Lipschitz continuity, the system controls how far potential outcomes can drift outside the known territory.
The research team developed influence functions and evaluated their interior projection technique across synthetic simulations and chemical exposure datasets. The simulations measured confidence interval coverage rates across sparse continuous treatments. The team confirmed their intervals met nominal coverage where standard methods failed, and demonstrated that supposed protective pesticide effects collapsed under slight outcome variation.
Researchers can now construct asymptotically normal estimators and reliable confidence intervals for complex treatment policies without relying on parametric extrapolation. The framework handles continuous exposures and chemical mixtures even when specific treatment combinations are rare or missing.