Teaching Neural Networks to Estimate Propensity Scores

• In Event: 25.018 - Advanced Analysis and Inference Issues

Abstract

Implementing neural network algorithms into propensity score analysis as an alternative to logistic regression requires less statistical know-how, searches all possible interactions between covariates, identifies nonlinear relationships between outcomes and predictor variables, and provides supervised training of algorithms to data. Nevertheless, the problematic implications due to use of neural networks is well reported across statistical, epidemiological, and methodological journals. This study reciprocates the counsels of previous literature by providing a comparison of logistic regression, neural networks with a Bayesian framework, and neural networks with maximum likelihood in propensity score estimation. A simulation study of a wide range of sample sizes and diverse sets of predicting variables and step-by-step procedures for implementing a neural network in propensity score analysis are demonstrated.

Authors

• Zachary Kendall Collier, University of Florida

• Walter L. Leite, University of Florida