Population-level measures of perceived food access reveal barriers beyond geographic proximity

Introduction

• Population-level food-access measures still rely mainly on proximity, density, and travel time despite food access also reflecting affordability, acceptability, and responsiveness to residents’ needs.Food access emerges from the interaction between residents and their local food environment.Source: Introduction · p. 1
• Surveys, interviews, focus groups, and store audits capture perceived food environments but require extensive manual collection, limiting spatial coverage, frequency, and scalability.Source: Introduction · pp. 1–2
• Online grocery reviews contain unsolicited descriptions of affordability, quality, customer service, and store operations, but have mainly been used to study sentiment, satisfaction, and business performance.Source: Introduction · p. 2
• The paper develops a machine-learning framework to use online grocery reviews for population-level measurement of perceived food access.Source: Introduction · p. 2

Online grocery reviews enable population-level measurement of perceived food access

• The framework analyzed 25,125 reviews from 49 Raleigh-area stores, using unsupervised topic modeling and zero-shot classification to assign topics to access dimensions.Automated assignments agreed with manual labels for 85.4% of topics; topics could map to multiple dimensions.Source: Online grocery reviews enable… · p. 2
• The five dimensions—availability, accessibility, affordability, accommodation, and acceptability—organize review content into complementary aspects of residents’ interactions with their food environment.The framework is based on the multidimensional food-access framework described in the paper.Source: Online grocery reviews enable… · p. 2, Table 1 · p. 3
• 66.8% of reviews were assigned to at least one dimension, with acceptability most prevalent at 50.4% and affordability next at 12.1%.Acceptability commonly concerned food quality, cleanliness, and customer service; operational topics such as hours and product availability appeared less often.Source: Online grocery reviews enable… · p. 2
• Only two of 49 stores scored at or above the regional average across all five dimensions, and all pairwise dimension correlations were below |r| = 0.40.The strongest associations were accessibility with accommodation (r = 0.39) and accessibility with acceptability (r = 0.35).Source: Figure 1 · p. 3, Online grocery reviews enable… · p. 4
• Aldi’s affordability score was +1.88 standard deviations, while the same location scored +0.22 for accessibility; Whole Foods scored -2.03 for affordability.Store scores varied across dimensions, so strength in one dimension did not imply strength in others.Source: Online grocery reviews enable… · pp. 2, 4, Figure 1 · p. 3

Geographic proximity overestimates perceived food access

• A resident’s opportunity set is the subset of geographically available stores satisfying perceived needs, preferences, and constraints.The example begins with three stores within approximately one mile; a modest threshold excluding the bottom 15% removes one store.Source: Geographic proximity overestimates… · pp. 4–5
• Under the strictest illustrated acceptability threshold, which admits only stores in the top 7% of scores, the nearest acceptable grocery store was 7.5 miles away.Source: Geographic proximity overestimates… · p. 5
• Across the study region and all five dimensions, stricter perceived-access thresholds progressively reduced available stores, with multiple requirements compounding the reduction.Because the dimensions are only weakly correlated, the gap between geographic proximity and perceived access grows as requirements become more numerous or restrictive.Source: Geographic proximity overestimates… · p. 5
• The opportunity-set analysis characterizes stores meeting perceived requirements, not residents’ eventual shopping choices.Residents may travel farther for a better fit or shop nearby while accepting compromises.Source: Geographic proximity overestimates… · p. 5

Food access is jointly determined by store characteristics and resident needs

• The paired-store comparison minimized differences in brand, pricing structure, product offerings, and corporate policies to focus on differing resident perceptions.Source: Food access is jointly determined by… · p. 5
• After both Harris Teeter stores ended 24-hour operation under the same chain-wide policy, about 80% of accommodation-related reviews at one store mentioned the shortened hours, while none at the other did.The two stores had similar overall ratings and comparable scores on most dimensions, differing primarily in accommodation.Source: Food access is jointly determined by… · p. 5
• The same policy change was perceived differently across locations, consistent with food access reflecting the interaction between residents and their food environment.Differences in local management or staffing could plausibly contribute to contrasting checkout-efficiency experiences, but the passage presents this as a possibility.Source: Food access is jointly determined by… · p. 5

Perceived food access is structured by neighborhood socioeconomic characteristics

• Store-level dimension scores were aggregated to census tracts using a distance-decay spatial interaction model, then related to neighborhood characteristics with weighted least-squares regression.Source: Perceived food access is structured by… · p. 5
• Larger Black populations were associated with lower perceived availability, accessibility, accommodation, and acceptability, but higher affordability; only availability and affordability associations were statistically significant.Source: Perceived food access is structured by… · p. 5
• Nearby Harris Teeter stores under the same chain-wide policy differed in accommodation perceptions: approximately 80% of reviews at the lower-scoring store discussed reduced hours, versus none at the higher-scoring store.Source: Figure 3 · p. 6
• Higher-income census tracts had substantially higher perceived acceptability, the dimension with the strongest overall socioeconomic patterning (R2 = 0.415).Source: Perceived food access is structured by… · p. 6
• Perceived and geographic food access were not always aligned: availability reinforced geographic disadvantage in census tracts with larger Black populations, whereas affordability was higher in those communities.Incorporating perceived access does not uniformly increase or decrease measured disadvantage.Source: Perceived food access is structured by… · p. 6

Discussion

• Online reviews provide scalable measures of perceived access that complement geographic measures and capture distinct dimensions of residents’ experiences.The framework uses pretrained language models and can be applied elsewhere where sufficient review data are available.Source: Discussion · pp. 7–8
• Nearby stores in the same chain that changed operating hours identically received different resident responses, showing that store characteristics alone do not establish whether needs are met.Accommodation and acceptability require information about both the food environment and residents’ experiences.Source: Discussion · p. 7
• Review topics included customer service, social interactions, and cleanliness as components of acceptability, potentially capturing locally salient experiences that predetermined instruments miss.The authors suggest vector embeddings could help identify latent access dimensions absent from existing frameworks.Source: Discussion · p. 7
• In Blacker neighborhoods, poorer geographic access coincided with lower perceived availability but higher perceived affordability; higher-income neighborhoods had greater store distance but higher perceived acceptability.Which dimensions are considered can therefore change how food-access disadvantage is characterized across communities.Source: Discussion · p. 7
• The measures characterize constraints on food environments, not realized shopping behavior; distinguishing travel from compromises in nearby stores requires behavioral data such as mobility or household travel.Residents may travel farther for preferred retailers or shop nearby while compromising on access dimensions.Source: Discussion · p. 7
• Google Maps reviewers are self-selected, so the resulting scores measure expressed perceptions rather than population-representative resident experience.Multi-store reviewers rated stores somewhat higher overall, although dimension-specific differences were generally small.Source: Discussion · p. 7
• Store scores are standardized within each study region and indicate relative performance, not absolute perceived access; comparisons should stay within coherent retail markets.The authors suggest unstandardized scores for cross-region comparisons or mobility data to define store-specific markets.Source: Discussion · pp. 7–8
• Neighborhood associations may reflect both residents’ experiences and retailer location decisions, which this cross-sectional analysis cannot disentangle.For example, higher perceived affordability near tracts with larger Black populations could reflect perceptions as well as the types of retailers operating there.Source: Discussion · p. 8

Study area and review corpus

• The corpus covered 25,125 reviews from 49 nationally recognized grocery stores in Raleigh, including stores in the city and a one-mile buffer.The buffer was selected to include stores plausibly serving residents near the boundary under the USDA’s one-mile urban threshold.Source: Study area and review corpus · pp. 8–9

Text preprocessing and topic modeling

• BERTopic identified 59 topics, of which 41 were retained for dimension assignment after excluding 18 incoherent or non-substantive topics.The selected model had coherence 0.528 and diversity 0.709; eight candidate models were compared using coherence, diversity, and manual interpretability assessment.Source: Text preprocessing and topic modeling · p. 9

Mapping topics to dimensions of food access

• Zero-shot matching compared embeddings of grocery-review-style dimension definitions with topic centroids to assign topics to availability, accessibility, affordability, accommodation, and acceptability.The five definitions and reviews used the same embedding model, and cosine similarity measured topic-to-dimension similarity.Source: Mapping topics to dimensions of food… · p. 9, Table 3 · p. 10
• Automated and manual primary topic labels agreed for 85.4% of topics, with macro- and prevalence-weighted one-versus-rest ROC–AUC values of 0.915 and 0.861, respectively.One researcher manually labeled retained topics using keywords and representative reviews.Source: Mapping topics to dimensions of food… · p. 9

Store-level dimension scores; Bayesian partial pooling

• Store–dimension scores combined topic assignments with review signals and used separate Bayesian hierarchical models to partially pool estimates before standardizing them within each dimension.A score of zero represents the regional mean, while one represents one standard deviation above it.Source: Store-level dimension scores · p. 9, Bayesian partial pooling · p. 10

Census-tract exposure scores; Statistical analysis

• Census-tract exposure scores were normalized distance-weighted means of store scores, and population-weighted tract models examined income, the proportion identifying as Black, and their interaction.The analysis also measured pairwise Pearson correlations among store-level dimension scores.Source: Census-tract exposure scores · p. 10, Statistical analysis · p. 10

• Table 4 organizes each chain’s review count and dimension-assigned reviews; reviews may be assigned to multiple dimensions.Source: Table 4 · p. 14
• The association between reviewing one versus multiple stores and dimension assignment was small (Cramer’s V = 0.05).Source: Table 5 · p. 14
• Review-level scores assign −1 to one-star reviews, 0 to two- to four-star reviews, and +1 to five-star reviews; all ratings remain in the denominator.Table 6 reports mean Google stars and review-level food-access scores by reviewer scope and dimension.Source: Table 6 · p. 14