Vellora Research · Independent exploratory research — not peer reviewed
Vellora research: a summary for reviewers
Three exploratory, county-level studies of public health data for all 254 Texas counties, with their data, code, analysis plans and an independent recomputation of every reported statistic.
Agamvir Singh — Independent Researcher & Founder of Vellora
About Vellora
Vellora is an independent public health education and research site covering all 254 Texas counties. It publishes a health profile and a PDF report for every county, an interactive map and comparison tools, and plain-language articles, built from verified copies of public datasets from CDC, the National Cancer Institute and Texas Cancer Registry, the U.S. Census Bureau, HRSA and USDA. See Methodology for every source and County profiles for the county pages.
The three research reports below are secondary analyses of those public data. They have not been peer reviewed, are not affiliated with any university or public agency, and are not medical advice. They describe patterns between counties; none can show what causes those patterns or describe any individual’s health.
The three studies
Study 1 · Published 2026-10-23 · Version 1.1 (updated 2026-10-24)
Geographic Variation in Cardiovascular Health Across Texas Counties
How do county estimates of cardiovascular conditions, their risk factors and heart disease death rates vary across Texas, do they cluster geographically or differ between metropolitan and nonmetropolitan counties, and how strongly do the risk factors track the outcomes?
Verified findings
- All 8 measures were geographically clustered (global Moran’s I 0.35 to 0.57; permutation p = 0.0001 for each, the smallest attainable).
- Every measure was higher in nonmetropolitan counties; the median heart disease death rate was 415.1 against 361.2 per 100,000 (BH-adjusted p < 0.0001).
- Of five risk factors, current smoking had the strongest association with the heart disease death rate (Spearman ρ = 0.54, 95% CI 0.44 to 0.63); diagnosed diabetes showed no clear association (ρ = 0.09, adjusted p = 0.132).
- Data and years
- CDC PLACES 2025 release, age-adjusted prevalence (BRFSS 2023) for five risk factors and two conditions; CDC heart disease death rates, adults 35+, 2022–2024; USDA Rural-Urban Continuum Codes 2023.
- Statistical methods
- Distributions; global Moran’s I with 9,999 permutations; Mann–Whitney tests by metropolitan status; Spearman correlations with Fisher z intervals; Benjamini–Hochberg adjustment within each table; crude prevalence as a sensitivity analysis.
- Sample
- 253 of 254 counties (Loving County withheld by CDC in every source).
- Main limitations
- Ecological and cross-sectional; PLACES risk factors and conditions come from one model, so their high correlations are partly built in; spatially smoothed death rates; unweighted counties; no adjustment for confounders.
Study 2 · Published 2026-10-23 · Version 1.2 (updated 2026-10-24)
Smoking Prevalence and Lung Cancer Incidence Across Texas Counties
Is age-adjusted adult current cigarette smoking prevalence associated with age-adjusted lung and bronchus cancer incidence across Texas counties?
Verified findings
- Age-adjusted adult smoking prevalence and age-adjusted lung cancer incidence were positively associated across counties (Spearman ρ = 0.44, 95% CI 0.32 to 0.55, p < 0.0001, n = 203).
- Exploratory: the association was stronger among metropolitan (ρ = 0.68) than nonmetropolitan counties (ρ = 0.32); removing any one county left ρ between 0.43 and 0.46.
- All 51 excluded counties are small counties whose lung cancer rate is suppressed (fewer than 16 cases over 2019–2023).
- Data and years
- CDC PLACES 2025 release, current cigarette smoking among adults, age-adjusted (BRFSS 2023); State Cancer Profiles lung and bronchus cancer incidence, age-adjusted, 2019–2023 (Texas Cancer Registry); USDA RUCC 2023.
- Statistical methods
- One prespecified confirmatory test: Spearman’s ρ with a Fisher z interval (Bonett–Wright standard error). Exploratory: Pearson’s r, metropolitan and nonmetropolitan subgroups, leave-one-out influence, crude smoking prevalence and the profile of excluded counties.
- Sample
- 203 of 254 counties with both values; 51 excluded.
- Main limitations
- Ecological and cross-sectional; smoking (2023) and incidence (2019–2023) differ in period and lung cancer has long latency; model-based smoking estimates; suppression removes the smallest counties; no confounder adjustment; the two measures use different age adjustments.
Study 3 · Published 2026-10-24 · Version 1.1 (updated 2026-10-24)
Rural vs. Metropolitan Health Disparities Across Texas Counties
How do selected health outcomes and risk factors differ between metropolitan and nonmetropolitan Texas counties?
Verified findings
- Heart disease death rates were higher in nonmetropolitan counties: Hodges–Lehmann shift +45.4 per 100,000 (95% CI +25.7 to +65.0), BH-adjusted p < 0.0001.
- Lung cancer incidence showed no clear difference (shift +3.5, 95% CI −0.9 to +7.9 per 100,000, adjusted p = 0.134); its rate is suppressed in 43 of 168 nonmetropolitan and 8 of 86 metropolitan counties, so the result is inconclusive.
- All six CDC PLACES age-adjusted estimates were higher in nonmetropolitan counties (probability of superiority 0.61 to 0.73; adjusted p ≤ 0.0033).
- Data and years
- USDA RUCC 2023 (re-verified against the USDA file); CDC heart disease death rates, adults 35+, 2022–2024; State Cancer Profiles lung cancer incidence, 2019–2023; CDC PLACES 2025 age-adjusted estimates (BRFSS 2023).
- Statistical methods
- Prespecified primary family (two observed rates) and secondary family (six PLACES estimates); medians, Hodges–Lehmann shifts with distribution-free 95% CIs, probability of superiority, Mann–Whitney tests with Benjamini–Hochberg adjustment within each family; Fisher’s exact test for suppression; exploratory crude, gradient and adjacency analyses.
- Sample
- 86 metropolitan and 168 nonmetropolitan counties; per-indicator sample sizes reported separately (lung cancer: 78 and 125).
- Main limitations
- Ecological and cross-sectional; county classification masks within-county variation; confounding; model-based PLACES estimates; suppression concentrated in nonmetropolitan counties; unweighted counties; different reporting periods and age adjustments.
Where external review would help most
The analyses have been checked computationally, not by independent scientists. These are the questions on which expert review would add the most:
- Ecological inference
- Every study compares county populations. Reviewers could judge whether the reports keep county-level findings from being read as individual-level ones.
- Confounding
- No study adjusts for income, education, insurance, occupation, environment or access to care. Advice on which adjusted analyses would be defensible with public county data would help.
- Spatial dependence
- Neighbouring counties are not independent. Only the cardiovascular study measures clustering (Moran’s I); the correlation and group tests treat counties as independent, which can understate uncertainty.
- Uncertainty in county estimates
- PLACES estimates and cancer rates carry their own confidence intervals, which the tests do not propagate. Small-county values are especially uncertain.
- Suppression and selection
- Lung cancer rates are suppressed for 51 small counties, mostly nonmetropolitan. The effect of this non-random missingness on the smoking–lung cancer correlation and the metropolitan comparison deserves scrutiny.
- Model-based estimates
- CDC PLACES values come from a small-area model that uses county demographics; associations among PLACES measures, and differences by metropolitan status, may partly reflect the model.
- Statistical methods
- Choice of rank-based tests, the Hodges–Lehmann interval without tie correction, Benjamini–Hochberg families, and the unweighted county design.
Reproducing the results
The reproducibility package contains Vellora’s verified copies of the input data, the results files, the analysis code, the analysis plans, the download scripts and the independent Python verifiers, with a README covering provenance, suppression rules, licences and expected outputs. Two checks run from a clean folder:
- Verify: each study’s Python verifier (SciPy 1.18.1, NumPy 2.5.4) recomputes every reported statistic from the inputs and compares it with the results file.
- Regenerate:
npx tsx regenerate.tsreruns the original analysis code on the inputs and confirms each results file is reproduced exactly.
Limits: the agencies may revise their data, so rerunning the download scripts today may not reproduce the stored inputs; the cardiovascular study had no separately committed analysis plan; and automated recomputation shows that the numbers follow from the code and data, not that the methods are the right ones.
Versions and corrections
Each report keeps a stable URL and a public version history, and material corrections are recorded. See the corrections policy and full version history.
Feedback from reviewers
A reviewer contact method has not been published yet. When it is, this section will explain how to report a question, a methodological concern or a suspected error for a specific study.
AI assistance. Anthropic’s Claude (via Claude Code) was used for software development, implementation of the statistical analyses and of their verification scripts, data visualization, and drafting of the analysis plans and report text. Results were checked with reproducible computational methods: a separate Python implementation (SciPy and NumPy) recomputes every reported statistic. These automated checks are not independent human scientific review. The research author, Agamvir Singh, is responsible for the interpretation and conclusions.