A Proxy Method for Classifying Grocery Stores at Scale

What an article-reported analysis suggests, and what remains to be reproduced

The Article-Reported 92.1% Share

The analysis described in this article reports that 92.1% of establishments tagged as "grocery stores" were assigned to Tiers 4-5 across seven California counties. That estimate comes from a proxy classification and has not been reproduced from a public artifact.

This matters because many food-environment measures treat grocery locations as equivalent. A related analysis on this site reports that 0.04% of its study population met a traditional food-desert definition, but that underlying population is under separate reconciliation and should not be treated as a verified benchmark here.

But that metric hides a deeper problem. Living near a store that sells mostly chips, sodas, and lottery tickets is not the same as living near a Safeway with a full produce section. The "grocery store" label obscures more than it reveals.

The methodological question is how to distinguish store types at scale when visiting the article-reported population of roughly 25,000 locations is not feasible.

The Audit Scalability Problem

The Nutrition Environment Measures Survey for Stores (NEMS-S), developed and evaluated by Glanz and colleagues in the Atlanta metropolitan area, is an in-person method for assessing availability, price, and quality. The original 2007 evaluation reported ten food categories or indicator food items. The current Penn NEMS-S instrument, which has been updated since that publication, lists 11 store-environment measures. The design problem here is scale.

Glanz, K., Sallis, J. F., Saelens, B. E., & Frank, L. D. (2007). Nutrition Environment Measures Survey in stores (NEMS-S): Development and evaluation. American Journal of Preventive Medicine, 32(4), 282–289. https://doi.org/10.1016/j.amepre.2006.12.019

Each NEMS-S audit requires a trained researcher to visit a store, check shelves, compare prices, and rate produce freshness. Using the article's 30-to-45-minute assumption, the upper-bound calculation for 25,000 locations exceeds 18,000 hours of fieldwork.

Alternative tools face similar constraints. The Consumer Food Environment Healthiness Score (CFEHS) and Healthy Food Supply (HFS) Score both require in-person assessment. These methods trade scalability for precision: a reasonable choice for targeted intervention studies, but impractical for population-level food environment mapping.

This creates a practical measurement question: can proxy data support broader geographic coverage without being mistaken for a direct store audit?

So we tried building one.

Our Method: Proxy-Based Tier Classification

The article describes a five-tier proxy classification using fields available through the Google Places API:

The classification relies on three proxy signals:

Chain affiliation. The described rule assumes that stores in a known chain have sufficiently similar inventory to share a tier. A lookup table of more than 200 chain names provides the starting assignment, but the assumption requires store-level validation.

Google Places type. The API distinguishes "supermarket" from "grocery_store" from "convenience_store." The method treats those categories as imperfect classification signals.

Review count. The described rule uses review count as a proxy for store scale. That relationship is an assumption in the classifier, not a validated result in the public record.

The article reports processing 38,157 raw listings, excluding 12,725 entries categorized as non-grocery establishments, and retaining 25,432 classified stores. Those counts have not been matched to a public input file or saved output.

The article-reported result assigns 7.9% of stores to Tiers 1-3 and 92.1% to Tiers 4-5. These are classification shares, not direct observations of inventory, price, affordability, or store quality.

The Article's Validation Exercise

A classification system is only useful if it reflects reality. The article describes a small comparison against product listings visible on Instacart storefronts.

What We Tested

The article reports collecting visible produce items, dairy items, and fresh-section indicators for 12 stores across four tiers. The displayed table names 11 stores, so the twelfth observation cannot be identified from the published material. The storefront extract and store-level coding are not public.

Article-Reported Results

What the Small Sample Suggests

The article reports a clear Tier 5 distinction in this sample. CVS and Walgreens were coded with zero fresh produce, zero fresh meat, and zero dairy beyond frozen desserts. Without the store-level extract, this should be read as a reported sample result rather than an independently checked validation.

The article reports substantial produce listings for the two Tier 3 markets sampled. The reported average was 7.5 visible items, comparable to the Tier 2 sample. A two-store comparison cannot establish performance for the tier as a whole.

The article reports separation in its storefront comparison. Tiers 1-3 averaged 8.8 visible produce items and Tier 5 averaged zero. The table names 11 stores even though the narrative reports 12, so these values remain unresolved. A larger, public validation set is needed to estimate classification performance.

What Needs Refinement

The Tier 1 vs. Tier 2 distinction was weak in the article-reported sample. Produce counts were 10.0 versus 8.8. That small comparison suggests testing whether the categories should be collapsed or whether dairy depth belongs in a future validation design.

Tier 4 remains unvalidated. The article says many small independent groceries were not represented on Instacart. The current evidence cannot establish whether a specific independent store belongs in Tier 4, and the absence of this category from the small validation sample limits interpretation of the reported 92.1% estimate.

Article-Reported County Comparisons

The article reports applying the classification to seven California counties and calculating a modified Retail Food Environment Index (mRFEI). The county values have not been reproduced from a public store-level table.

County-Level Results

Why Our Numbers Differ from CDC's

The discrepancy stems from what each index measures:

CDC's mRFEI includes fast food restaurants in the denominator. A county with many fast food outlets but decent supermarket coverage will show a low mRFEI. The metric captures healthy vs. unhealthy food retail broadly.

The article's proxy mRFEI excludes restaurants and focuses on establishments classified as grocery-type locations. It asks a narrower question: among places classified as selling groceries, what proportion were assigned to Tiers 1-3?

The measures are designed to answer different questions. The article proposes using the proxy index to study grocery composition specifically, but its usefulness for SNAP policy, supermarket-attraction initiatives, or food-security interventions has not been established by the current evidence.

Limitations

This is a scalable approach, not a perfect one. The limitations are real:

Proxy assumptions may fail. The method assumes chain affiliation predicts quality. A poorly stocked chain store can still receive Tier 1, while a well-stocked independent store can still receive Tier 4. Without broader validation, individual stores may be misclassified.

Tier 4 heterogeneity. Small groceries range from corner markets with fresh produce to stores selling mostly packaged goods. The described system groups them together.

No price or affordability data. A store can have produce but remain unaffordable to low-income residents. The described method addresses availability signals, not accessibility.

Static snapshot. Stores open and close. The article-reported data reflect one point in time.

No SNAP/WIC verification. The described method does not capture whether stores accept food-assistance benefits, a critical dimension for food security.

Implications

Given the current evidence status, the proxy-based approach is best treated as a candidate workflow to test:

For researchers: The described method provides a design for population-level food-environment mapping without a full field audit. Reproducibility cannot be assessed until the code, classification tables, inputs, and saved outputs are public.

For policymakers: The article-reported county values can motivate questions about retail composition, but they should not identify intervention priorities until the classification is validated and the calculations are reproduced.

For future work: Validating proxy classifications against a sample of NEMS-S audits would quantify error rates and enable calibration. If proxy-based scores correlate strongly with audit-based scores, the method could complement rather than replace traditional assessment.

The reported 92.1% estimate is hypothesis-generating, not a settled finding. It shows why store labels may need closer scrutiny, while the missing Tier 4 validation and nonpublic analysis package prevent a stronger conclusion about food access or policy.

This article reports an analysis of 25,432 classified stores across seven California counties and a storefront comparison from December 2025. Its narrative reports 12 stores, while the displayed table names 11; the missing observation is unresolved. The code, source data, classification table, store-level validation extract, and matched outputs are not publicly available, so the results have not been independently reproduced.