A Tutorial on Artificial Neural Networks in Propensity Score Analysis

Abstract

Artificial neural networks (NN) can help researchers estimate propensity scores for quasi-experimental estimation of treatment effects because they can automatically detect complex interactions involving many covariates. However, NN is difficult to implement due to the complexity of choosing an algorithm for various treatment levels and monitoring model performance. This research aims to develop a tutorial to facilitate the use of NN to derive causal inferences. The tutorial provides social scientists with a gentle overview of machine learning terminology and best practices for training, validating, and testing NN to estimate propensity scores. The veracity of NN is demonstrated in this study using data on 5,770 teachers from the Beginner Teacher Longitudinal Study. Propensity score analysis was used to estimate the effects of assigning mentors to new teachers on the probability of continuing in the teaching profession. The results show that NN provided a better covariate balance between treatment versions than multinomial logistic regression and generalized boosted modeling. The study's findings align with previous research showing NN's advantages over conventional propensity score estimation methods.

Keywords:

• Propensity score
• data mining
• categorical treatments
• potential outcomes framework
• quasi-experimental designs
• inverse probability of treatment weighting