An application of a machine learning technique in microeconomics: using a neural network to enhance prediction in the service of estimation in the context of the South African child support grant
Date
2022
Authors
Wootton, Kyle
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
In this study, my aim is to show how machine learning can be used for prediction in the service of estimation in the context of a microeconomic research question: namely, whether the South African Child Support Grant (‘the grant’) improves the nutrition of children who receive it. Specifically, I show how a fully connected artificial neural network can be used as a novel, and potentially superior, approach to constructing an input variable in microeconomic research where the input variable is the result of high-dimension prediction. The hypothesis is that, if a neural network can be used to improve predictive performance when constructing the input variable, the estimation step that relies on this input as a covariate should also be more accurate. The input variable in question is caregiver motivation which is a covariate used as part of the identification strategy when determining the impact of the grant on nutrition. Caregiver motivation is constructed as the standardised difference between predicted application delay and actual application delay. Actual application delay is the number of days between the child becoming eligible for the grant and receiving the grant. Predicted application delay is the expected number of days between a child becoming eligible for the grant and receiving the grant given a set of observable characteristics. When comparing eligible children who receive the grant to eligible children who do not, I find the motivation variable constructed using a neural network results in the grant having a statistically significant impact on child nutrition – a result consistent with theoretical expectations and qualitative empirical evidence. In contrast, the motivation variable constructed using ordinary least squares (OLS) regression finds no statistically significant improvement in child nutrition from the grant – a result contrary to theoretical expectations and qualitative empirical evidence. As a result, I argue the neural network is a better predictor of application delay for the grant than the OLS regression.
Description
A research report submitted in partial fulfilment of the requirement for a Degree of Master of Commerce (Economics) to the Faculty of Commerce, Law and Management, School of Economics and Finance, University of the Witwatersrand, 2022