Propensity Score Estimation with Multi-Layer Neural Networks
The propensity score analysis is a statistical tool for deriving causal inferences in a broad class of settings when the randomized controlled experiment is not feasible in terms of design or research ethics. Observational studies provide enormous amounts of data statistical processing of which, nevertheless, can not explicitly answer the question of how large is the effect of the treatment on the outcome due to the effects of the confounders. The propensity score makes it possible for a researcher to reduce the confounders to one dimension to account for the confounding effects. In this light, it is of supreme importance to find the best tool for estimating the propensity score. A variety of machine learning tools have been investigated regarding their performance in estimating probabilities of receiving the treatment given the covariates as well as their computational efficiency. It has been shown that one-layer neural networks perform reasonably well (Setoguchi et al., 2008). However, not much research has been conducted on the ability of multi-layer neural networks to provide an unbiased estimation of propensity score. This research is focused on the assessment of the efficiency of multi-layer neural networks in estimating propensity scores. The study showed the utility of using multi-layered neural networks in estimating the propensity scores in the scenarios when both the outcome and propensity score models are non-linear. However, in the case of low exposure prevalence adding each hidden layer must be done with caution.
Migunov, Igor, "Propensity Score Estimation with Multi-Layer Neural Networks" (2022). ETD Collection for Fordham University. AAI29320828.