# Analysis plan: Rural vs. Metropolitan Health Disparities Across Texas Counties

Vellora Research · plan version 1 · written 2026-10-23, before any group difference was computed for this study.
Independent exploratory analysis — not peer reviewed.

## Question

How do selected health outcomes and risk factors differ between metropolitan and nonmetropolitan Texas counties?

Design: ecological, cross-sectional comparison of the distributions of county values. The unit is the county, not the person. No causal claim is made about metropolitan or nonmetropolitan residence.

Prior Vellora work: the cardiovascular report (2026-10-23) already compared metropolitan and nonmetropolitan county medians for eight cardiovascular measures with Mann–Whitney tests. This study reuses some of the same verified data. It adds an observed cancer outcome, a prespecified primary family of observed (not modelled) outcomes, effect sizes with confidence intervals, a missing-data analysis and sensitivity analyses. Where the measures overlap, its group medians and Mann–Whitney tests must reproduce the earlier report's.

## Groups

USDA Economic Research Service Rural-Urban Continuum Codes 2023 (RUCC), as stored in `public/data/research/cardio-disparities-inputs.v1.json`:

- Metropolitan: RUCC 1–3 (expected 86 counties).
- Nonmetropolitan: RUCC 4–9 (expected 168 counties).

The stored codes were re-checked on 2026-10-23 against a fresh download of the USDA file (https://www.ers.usda.gov/media/5768/2023-rural-urban-continuum-codes.csv?v=97456, SHA-256 ec455ee2a8bc5fc8e070575ea5bee7dce46fc6037f8c3449cbf56e8b45331fa7): 254 Texas rows, no mismatches, 86 metropolitan and 168 nonmetropolitan. Counties are matched on five-digit FIPS codes. "Nonmetropolitan" is not equated with "rural": RUCC classifies whole counties, and both groups contain varied places.

## Indicators

Primary family (observed rates, not produced by a small-area model):

| ID | Measure | Source and period | Unit | Adjustment | CI |
|---|---|---|---|---|---|
| HDMORT | Heart disease death rate, adults 35+ | CDC/DHDSP, deaths 2022–2024, spatially smoothed | per 100,000 adults 35+ | Age-standardized, 2000 U.S. standard population | Not published |
| LUNG | Lung and bronchus cancer incidence, all ages | State Cancer Profiles (Texas Cancer Registry), diagnoses 2019–2023 | per 100,000 people | Age-adjusted, 2000 U.S. standard population | 95% CI published |

Secondary family (CDC PLACES 2025 release, BRFSS 2023, adults 18+, age-adjusted prevalence AgeAdjPrv, model-based; CDC's PLACES documentation does not name its standard population):

| ID | Measure |
|---|---|
| CHD | Coronary heart disease |
| BPHIGH | High blood pressure |
| DIABETES | Diagnosed diabetes |
| OBESITY | Obesity |
| CSMOKING | Current cigarette smoking |
| LPA | No leisure-time physical activity |

Not selected: stroke prevalence (CHD represents modelled cardiovascular conditions and the two are highly correlated in PLACES), other PLACES measures (outside the cardiometabolic and smoking focus), and other cancer sites (sex-specific or less tied to the selected risk factors). Crude and age-adjusted values are never combined.

## Inclusion and missing data

Per indicator, a county is included if its value is published. Suppressed values are excluded and listed with the source's reason; nothing is imputed or replaced with zero. Denominators are reported separately for each indicator and group. No statewide rate is computed from county values.

## Primary analysis

For each indicator, direction nonmetropolitan minus metropolitan:

1. Group n (published and suppressed), median, interquartile range, mean, SD, minimum and maximum.
2. Difference in group medians (descriptive).
3. Hodges–Lehmann estimate of the location shift (median of all nonmetropolitan − metropolitan pairwise differences) with a distribution-free 95% confidence interval from the order statistics of those differences (normal approximation for the critical rank, no tie correction).
4. Probability of superiority P = U / (n_nonmetro × n_metro), the probability that a randomly chosen nonmetropolitan county has the higher value (ties count ½), and the rank-biserial correlation r = 2P − 1.
5. Two-sided Mann–Whitney U test, normal approximation with tie and continuity corrections (the implementation used in the cardiovascular report, checked against SciPy).

Multiple testing: Benjamini–Hochberg false discovery rate at 0.05, applied separately within the primary family (2 tests) and within the secondary family (6 tests). Raw and adjusted p-values are both reported. A non-significant result is not evidence of equivalence.

## Exploratory and sensitivity analyses (unadjusted p-values, not used for conclusions)

- E1 Crude prevalence: repeat the primary analysis on PLACES crude prevalence (CrdPrv) for the six secondary measures.
- E2 Gradient: Spearman correlation between RUCC code (1–9) and each indicator across all counties with data.
- E3 Adjacency: within nonmetropolitan counties, compare those adjacent to a metro area (RUCC 4, 6, 8) with nonadjacent ones (5, 7, 9): medians, Hodges–Lehmann shift with 95% CI, Mann–Whitney p.
- E4 Missingness: for each indicator with any suppression, suppressed counts by group, two-sided Fisher's exact test, and the median 2020 population of suppressed counties.
- Population weighting: not performed. The county values are age-adjusted rates or model-based prevalence estimates; weighting them by 2020 population would not produce a valid pooled rate without age-specific counts and denominators, which are not in the verified data.

## Interpretation limits

Ecological fallacy, confounding (age structure beyond adjustment, income, education, insurance, occupation, environment, access to care), county-level classification and within-county heterogeneity, different reporting periods, model-based PLACES estimates whose covariates include county demographics, suppression in small counties and resulting selection, unweighted counties of very different size, different age-adjustment populations, statistical uncertainty, multiple comparisons, and no individual-level data.