Causal analysis of continuous treatments with soft constraints
N. T. Longford

Abstract

Approaches based on balancing weights are firmly established in causal analysis of binary treatments in observational studies. Their extension to treatments defined on ordinal scales, which arise in dose- and exposure-response studies, remains a methodological challenge. We develop a method of balancing that naturally extends from binary treatment to a continuum of treatment doses. The solution, the estimated average dose-response function in a given sample or population, depends on a set of tuning parameters that quantify the relative priorities of balancing the background variables. The closed form of this function facilitates a detailed analysis of its finite-sample properties, including a sensitivity analysis for hidden bias. All the constraints considered in other approaches are converted in this method to penalties, and so are satisfied only approximately but according to the priorities specified by the analyst. The method is applied to a study of the effect of campaign advertising on donations in the 2008 U.S. presidential elections.

Submitted.

August 2025.