Local Longitudinal Modified Treatment Policies

Longitudinal Modified Treatment Policies (LMTPs) provide a framework for defining a broad class of causal target parameters for continuous and categorical exposures. We propose Local LMTPs, a generalization of LMTPs to settings where the target parameter is conditional on subsets of units defined by the treatment or exposure. Such parameters have wide scientific relevance, with well-known parameters such as the Average Treatment Effect on the Treated (ATT) falling within the class. We provide a formal causal identification result that expresses the Local LMTP parameter in terms of sequential regressions, and derive the efficient influence function of the parameter which defines its semi-parametric and local asymptotic minimax efficiency bound. Efficient semi-parametric inference of Local LMTP parameters requires estimating the ratios of functions of complex conditional probabilities (or densities). We propose an estimator for Local LMTP parameters that directly estimates these required ratios via empirical loss minimization, drawing on the theory of Riesz representers. The estimator is implemented using a combination of ensemble machine learning algorithms and deep neural networks, and evaluated via simulation studies. We illustrate in simulation that estimation of the density ratios using Riesz representation might provide more stable estimators in finite samples in the presence of empirical violations of the overlap/positivity assumption.