Investigating Grocery Shopping Choice and Behavioral Changes during COVID-19 Using Machine Learning Methods

Abstract

This study leverages machine learning (ML) models to predict grocery shopping channel choices in Florida, comparing findings across two waves of stated preference (SP) survey data (collected February to April and November to December 2021). The SP survey design considered three alternatives (online, curbside, in store) and five cost and time attributes, (product price, delivery cost, travel time, delivery time, shopping time). The data were analyzed to examine changes in the importance of the explanatory variables including cost and time attributes, personal attitudes, and sociodemographic characteristics. In addition to ML models, a mixed logit model was applied to provide a reference to compare with the ML models. Among the six ML models evaluated, XGBoost was the best performing, and product price was the most influential variable predicting grocery shopping channel choices. The results also showed that recreational shoppers tended to prefer the online and curbside options over in-store shopping, during both waves. Furthermore, cost-consciousness was linked to a preference for the curbside pickup alternative during the second wave, and some attitudes changed between the two waves. The comparison revealed that both mixed logit and ML models consistently identified cost-related and demographic factors as key influences on shopping choices. ML models offered more nuanced insights into evolving factors like vehicle ownership, whereas mixed logit models provided better interpretability, especially in capturing the stable impact of socioeconomic factors. The findings of this study provide a better understanding of consumer behaviors, and their changing attitudes toward and perceptions of different shopping channels.