Primary Submission Category: Machine Learning and Causal Inference
Investigating neural networks to learn an intervention response function on a continuous exposure
Authors: Mauricio Tec, Oladimeji Mudele, Kevin Josey, Falco Bargagli Stoffi, Francesca Dominici,
Presenting Author: Mauricio Tec*
This work investigates neural network (NN) techniques for estimating the expected change in an outcome of interest resulting from a hypothetical change in a continuous exposure. For example, a policymaker may want to know how many hospitalizations could be avoided by an intervention reducing air pollution from its current values in each location. We formulate this goal as learning an intervention response function (IRF) using the stochastic interventions (SI) framework (Hubbard & Van Der Laan, 2005). We illustrate how IRFs differ from exposure-response functions and can better support reasoning about certain intervention policies of interest, particularly under overlap violations.
We consider three mechanisms in which NNs can improve IRF estimation:
By learning a flexible model of the conditional density ratios required by IPTW and TMLE estimators for SIs (Diaz-Munoz & Van Der Laan, 2012);
By using architectures and priors promoting smoothness in the individual potential outcome curves as a function of the exposure;
By extending recent work on doubly robust estimation via targeted regularization for NNs (Shi et al., 2019; Nie et al., 2021) to the case of SIs.
These mechanisms are evaluated using previously proposed synthetic benchmark datasets for continuous treatments. Finally, we illustrate them in a real application to estimate the impact of a reduction in airborne exposure to toxic metals (Kodros et al., 2022) on hospitalizations among Medicare enrollees in the U.S.