Neighborhood Vulnerability Index: Results Under Review

USDA's 2023 household food-security report publishes national rates by the race and ethnicity of the household reference person. Those household-level national estimates do not establish a food-insecurity rate for Hispanic neighborhoods in Santa Clara County.

The article reports the following patterns, all awaiting a matched public output:

• 29 census tracts with high Asian populations and low vehicle ownership in the top two index quintiles
• A tract with a reported index score of 0.437 and SNAP participation of 5.3%
• A 21.2% SNAP participation rate for one constructed subgroup

The problem with single-variable analysis is that it flattens complexity. "Hispanic = vulnerable" ignores that a high-income Hispanic professional in Palo Alto faces different barriers than a Hispanic single mother in East San Jose.

The article describes an Extended Vulnerability Index combining poverty, SNAP participation, grocery distance, vehicle access, and population density. The prose does not publish the variable-level weights, directions, transformations, or complete formula, and the validation and sensitivity results are not publicly reproduced.

The method can surface intersections for further study. It cannot, on the current record, establish neighborhood food insecurity or identify a causal intervention target.

This post documents an article-described neighborhood index and identifies the specifications required before its construction can be reproduced.

Why Build a Composite Index?

The Problem with Single Variables

Food insecurity is multidimensional:

• Economic: Can families afford food?
• Geographic: Can they access stores?
• Mobility: Can they get there?
• Social: Do they face discrimination or language barriers?

Single-variable analysis (poverty rate alone, distance to grocery alone) captures one dimension while ignoring others. A tract could have low poverty but terrible transit (mobility barrier), high poverty but excellent store access (economic barrier only), or moderate everything but compounding disadvantages.

When to Use a Composite Index

We can use a composite index when:

1. The outcome is multidimensional (food security involves economics, geography, mobility, social factors)
2. The goal is to rank or prioritize (which 20 neighborhoods need intervention first?)
3. We need to summarize co-occurring conditions in one additive score
4. Policy requires tractable targeting (resources can't be deployed everywhere at once)

We typically want to avoid a composite index when:

1. Causal identification is needed. A separate identification design must match the causal question.
2. A single dimension dominates (if 95% of variation is poverty, just use poverty)
3. Transparency matters more than precision (black-box indices obscure what's driving scores)
4. Components are highly correlated (double-counting the same thing)

The Santa Clara County Context

We needed an index because:

• A companion article reports zero food deserts by the USDA definition. That premise and any conclusion about geography are under reconciliation.
• The companion article reports 4.7x variation in SNAP rates, a comparison currently under reconciliation
• Policy question: Which specific neighborhoods need intervention?
• Single variables couldn't answer this: poverty alone missed mobility barriers, and distance alone missed economic barriers

Building the Extended Vulnerability Index

Theoretical Foundation

Our index draws on the WHO Social Determinants of Health framework (Solar & Irwin, 2010), which identifies structural determinants (socioeconomic position, education, income), intermediary determinants (material circumstances, behaviors), and health system factors (access, quality).

The article names five inputs for the food-security application:

Construction Details That Must Be Recovered

The article names poverty, SNAP participation, grocery distance, vehicle access, and population density as inputs. It does not publish the variable-level weights, whether each input raises or lowers the score, the exact normalization bounds, missing-data rules, or a complete formula.

The article reports that alternative weights produce similar rankings. That is the right sensitivity question, but neither the alternative rules nor the ranking files are public.

How the Index Would Need to Be Validated

Reported Association, Not Independent Criterion Validity

The article reports a Spearman correlation of r = 0.83 (p < 0.001) with SNAP participation. That value is not publicly reproduced. More importantly, SNAP participation is one of the index components, so correlating the finished index with SNAP participation partly compares the index with itself. A criterion-validity test needs an outcome that was not used to construct the index and that measures the intended food-insecurity concept.

Weight Sensitivity

The article reports correlations above 0.91 across five alternative weighting schemes and top-quintile agreement above 96%. A public sensitivity table should show each weighting rule, tract score, rank, and agreement calculation before the index is called robust.

Article-Reported Patterns to Reproduce

• A tract reported at 1.29 miles from a grocery store, 40% no-vehicle households, index score 0.437, and SNAP participation 5.3%.
• A reported 29 high-Asian, low-vehicle tracts in the top two index quintiles.
• A reported 21.2% SNAP participation rate for one constructed intersectional subgroup.
• Reported Q5-to-Q1 comparisons of 7.1× for SNAP participation, 4.5× for Hispanic population share, and 6.7× for non-diploma share.

These are hypotheses and descriptive checks to reproduce. They do not establish food insecurity, explain why households participate in SNAP, or identify a preferred intervention.

When to Use This Index (And When Not To)

Where the EVI Adds Value

Program targeting: A reproduced and externally validated index could help form outreach candidates, but it should not allocate resources from the unreconciled rankings on this page.

Resource allocation: An index can structure discussion, but observed needs, community input, uncertainty, and validated outputs must precede funding decisions.

Monitoring over time: Only comparable data vintages and a frozen construction rule can support change measurement.

Cross-jurisdiction comparison: A common index could support comparison after the inputs, coverage, and weights are harmonized across counties.

Where the EVI Falls Short

Not Causal: The index shows where vulnerability concentrates, not why. High EVI could result from labor market conditions, housing costs, discrimination, or historical disinvestment. Policy needs to address causes, not just correlations.

Aggregation Hides Variation: The index describes the average tract, missing within-tract heterogeneity. Some households in a moderate-EVI tract have cars and high incomes; others are transit-dependent with no savings.

Components Are Correlated: The article reports r = 0.71 between poverty and SNAP participation, but no public output derives it. If confirmed, including both inputs could double-count economic disadvantage. The article describes that overlap as intentional; the current record does not establish its effect on the score.

Static Snapshot: 2023 data doesn't capture seasonal variation, economic shocks, or program changes.

Final Thoughts

An index can be a heuristic for deciding where to investigate first. The present EVI should not yet be used for targeting because its inputs and outputs remain under reconciliation.

For similar indices, three checks matter: choose a criterion that actually measures the intended construct, publish sensitivity to weights, and report subgroup sample sizes. A matched public table is the prerequisite for claiming that those checks passed.

Methodology Note

Data Sources: Census ACS 5-Year Estimates (2019-2023), Google Maps Places API records with the published 6,613 population under reconciliation, Census TIGER/Line shapefiles

Sample: 408 census tracts, Santa Clara County, California

Index Construction: The article names min-max normalization, a weighted linear combination, and quintile classification. The variable-level weights, directions, normalization bounds, missing-data rules, and complete formula are not public.

Article-reported checks: r = 0.83 with SNAP participation and ranking correlations above 0.91 across five weighting schemes. Neither output is public, and the SNAP association is not independent because SNAP participation is included in the index.

Public materials: Article only. No complete public package currently derives these results.

References

Cohen, J. (1988). Statistical power analysis for the behavioral sciences (2nd ed.). Lawrence Erlbaum Associates.

Paruolo, P., Saisana, M., & Saltelli, A. (2013). Ratings and rankings: Voodoo or science? Journal of the Royal Statistical Society: Series A, 176(3), 609-634.

Rabbitt, M.P., Reed-Jones, M., Hales, L.J., & Burke, M.P. (2024). Household food security in the United States in 2023 (Report No. ERR-337). U.S. Department of Agriculture, Economic Research Service.

Schober, P., Boer, C., & Schwarte, L.A. (2018). Correlation coefficients: Appropriate use and interpretation. Anesthesia & Analgesia, 126(5), 1763-1768.

Solar, O., & Irwin, A. (2010). A conceptual framework for action on the social determinants of health (Social Determinants of Health Discussion Paper 2). World Health Organization.

Frequently Asked Questions

How well does the vulnerability index predict SNAP participation?

The article reports r = 0.83 with SNAP participation, but no public output derives it. SNAP participation is included in the index itself and is not a direct food-insecurity measure, so its association with the completed index is not an independent validation test.

Are the index rankings sensitive to how components are weighted?

The article reports correlations above 0.91 and top-quintile agreement above 96% across five weighting schemes. Those values are under reconciliation and should not yet be described as robust.

What Q5-to-Q1 comparisons does the article report?

The article reports a 7.1x Q5-to-Q1 comparison for SNAP participation, 4.5x for Hispanic population share, and 6.7x for non-diploma share. The classifications and comparison table are not public.

What components are included in the vulnerability index?

The article names five inputs: poverty, SNAP participation, grocery distance, vehicle access, and population density. The public prose does not specify the variable-level weights, directions, transformations, or complete formula needed to reproduce the index.