Uncovering food purchase behaviors and their shifts using Hidden Markov Modeling

Abstract

Abstract Background Investigating food purchase data is a novel and effective approach to large-scale dietary assessment, but many traditional analysis methods often overlook the temporal structure of food purchases. This study demonstrates a data-driven approach for identifying and describing changes in food purchasing behaviors over time, using Hidden Markov Modeling (HMM) to assess the underlying structure of individual-level time-series purchase data. Methods The loyalty cardholders of the LoCard study provided access to purchase data from 9/2016 to 12/2018 and responded to a background questionnaire providing cross-sectional information on sociodemographic characteristics and self-estimated share of purchases from the retailer (purchase loyalty) ( n = 29,245). Food group-level data of participants with ≥ 40% purchase loyalty to the retailer were aggregated monthly, subjected to a dimension reduction, and then an HMM was fitted on the monthly data. Food purchase behaviors (i.e. modeled states) were labeled, and their estimated prevalence and transition probabilities across behaviors were examined by gender, age, household structure, and household income. Results Nine food purchase behaviors were grouped as health-oriented: (1) ‘balanced’ (2) ‘home cook’, (3) ‘flexitarian’, and (4) ‘planetary’; discretionary-oriented: (5) ‘traditional’, (6) ‘convenience’, (7) ‘sugar-sweetened beverages and snacks’, (8) ‘beer, sweets, and treats’; or baby-oriented: (9) ‘baby formula and foods’. Participants’ food purchase behaviors remained relatively stable, as demonstrated by low month-to-month transition probabilities (< 5%). Compared to other age and household categories, younger adults and households with two adults and children were most likely to change food purchase behaviors. Transitions typically occurred within the health-oriented or discretionary-oriented groups, or away from the baby-oriented group. Women, older adults, and households without children more often exhibited overall health-oriented behaviors, while younger adults specifically leaned towards plant-based purchasing behaviors, such as ‘flexitarian’ and ‘planetary’. Conclusions HMMs offer an interpretable approach for analyzing longitudinal food purchase data collected at high frequency over extended periods, having an ability to capture temporal changes in purchasing behaviors over time. The identified food purchase behaviors differed across sociodemographic characteristics, revealing distinct and relatively stable purchasing behaviors. Trends in their prevalences were only gradual, and transitions took place mainly between similar purchase behaviors, which highlights that purchase behavior is not easily altered.