Report · Phase 2

Esophageal cancer: what the areas with excess have in common

Hypothesis generation with public data on agricultural production, water, rurality, GDP and geography, to guide a study with patients. Versão em português.

Author: Clériston Almeida Capistrano (ORCID 0009-0008-5280-1869) · independent researcher with no formal training in health sciences · cledon.org

Abstract

Phases 1 and 1b showed a persistent excess of esophageal cancer mortality in inland areas of Minas Gerais, southern Brazil and Bahia. Phase 2 describes which municipal characteristics, in public data, go along with this excess, with an analysis plan recorded in git before the data (commit 0d88889). This version incorporates an AI-assisted technical review (not independent peer review): missing data are no longer treated as zero, and the analyses were redone.

• Comparing each municipality with the expectation from its state's rates, esophageal cancer mortality is higher where there is more spring or well water, yerba mate extraction, rural population, firewood and tobacco production, and lower where municipal GDP per capita is higher. The directions hold with state fixed effects, with the old missing-data rule and with 2000 and 2010 measures.
• For lung and other cancers, the rural and water variables are associated in the opposite direction, and the difference from esophagus is statistically clear for 8 of the 10 variables. This indicates that the pattern is not the same for all cancers; it does not rule out recording or diagnostic biases that affect cancers differently.
• By region (not pre-specified): in the South, the excess follows extracted yerba mate and tobacco; in Minas Gerais, where there is almost no mate or tobacco, it follows rurality, lower GDP per capita and lower altitude, with an inverse association with coffee area.
• The comparison of the 24 clusters with their neighbours is weaker: after correction for the 10 comparisons, no variable has q < 0.05.

1. Methods

Outcome. Esophageal cancer deaths 2015–2024 by municipality, with expected deaths from the state's own rates by sex and age group, used as offset. This compares each municipality with its state's pattern, but is not equivalent to including state effects in the regression; hence a sensitivity analysis with state fixed effects.

Exposures and units. Tobacco, cultivated yerba mate and coffee: harvested area of the crop as a percentage of the municipality's territorial area (PAM/IBGE). Extracted yerba mate: tonnes per 1,000 inhabitants; firewood: m³ per inhabitant, adding extraction and forestry (PEVS/IBGE). Production averaged over 2005–2010. Well water and spring water: % of households, 2022 Census. Rural population: %, 2022 Census. Municipal GDP per capita in 2010 (not household income). Centroid altitude (Copernicus DEM). Production variables enter as log(1 + x) and all are standardized: the rate ratio (RR) is per standard deviation of the transformed variable, not per original unit.

Missing data. In IBGE tables, “-” is an absolute zero and “...” means data not available. In the first version, “...” and absent municipalities became zero, confusing missing information with no production. Now: “-” = 0; “...”, “X” and absence from the response = missing; the average uses only years with data; with no year, the municipality is left out of that variable's analysis. Plant extraction has “...” in the same municipalities and years for yerba mate and firewood, concentrated in SP, RJ, PR, SC and MG; the forestry table omits many municipalities, so firewood (extraction + forestry, as in the plan) has data for few municipalities and there is a sensitivity analysis with extraction only.

Table 1. Audit of symbols in the production tables (6 years × 5,570 municipalities).

Model. Poisson regression with offset, sandwich standard errors clustered by IBGE immediate region (for spatial dependence and overdispersion), Benjamini–Hochberg across the 10 variables.

2. Associations by variable

Mortality RR per standard deviation of each variable, with 95% CI: esophagus (blue) and the comparison outcomes lung (orange) and other cancers except esophagus (green).

Table 2. Esophagus: main estimate and sensitivity analyses.

3. Comparison outcomes

Lung and all cancers except esophagus. They are not negative controls in the formal sense: lung shares exposures (tobacco), and recording or diagnostic biases may affect cancers differently. They show whether the pattern is specific to the esophagus. The difference was formally tested in a joint model, in the same municipalities.

An RR ratio above 1 indicates a stronger (or more positive) association for the esophagus than for the comparison outcome. There was no statistical evidence of a difference for coffee and altitude.

4. Joint models

With all 10 variables, only municipalities with data on all of them enter (firewood greatly limits the sample); hence also a model without firewood and extracted yerba mate. Several variables are correlated (cultivated and extracted yerba mate, ρ = 0.81; rurality and GDP, ρ = −0.47).

5. Clusters versus neighbours

Mean within each of the 24 clusters minus the mean of neighbours in the same state (up to 100 km, outside clusters), in standard deviations. The number of clusters varies with data availability (21 for firewood and 23 for extracted yerba mate; 24 for the others). Sign test excluding ties and Benjamini–Hochberg correction across the 10 variables. Clusters share neighbours, so the comparisons are not independent.

Clusters tend to have lower GDP per capita than their neighbours (19 of 24), but after correction no variable has q < 0.05 (GDP: q = 0.07). This analysis adds little evidence to the previous ones.

6. South and Minas Gerais (not pre-specified)

Separate models by region; variables with fewer than 20 municipalities with values above zero in the region were omitted (—), and firewood was left out of the joint models.

South: extracted yerba mate and tobacco remain associated in the joint model, together with well water and lower GDP; coffee area is inversely associated. Minas Gerais: rurality, lower GDP and lower altitude; coffee area is also inversely associated; well water, strong in the single-variable analysis, weakens in the joint model. None of the measured production variables has a statistically clear positive association with the excess in Minas Gerais.

7. Timing of the measures

Water and rurality come from the 2022 Census, after deaths start (2015), and do not document conditions before disease onset. As a historical comparison (not pre-specified): rural population in 2000 and 2010 and households with well or spring water in 2010 (the 2010 Census does not separate wells from springs). Rurality is very stable (correlation with 2022: 0.93 for 2000 and 0.95 for 2010); well water in 2022 and well or spring water in 2010 have correlation 0.73.

More wells do not demonstrate water contamination. Nitrate and other contaminants are hypotheses to investigate, not findings of this analysis.

8. The 16 municipalities

Value in the municipality and, in parentheses, the state median; “n/a” = no data. 14 of the 16 have a share of households using well water above the state median in 2022 (14 for well or spring water in 2010).

9. Hypotheses for a study with patients

Supported, at the municipal level, by these data:

• South: yerba mate (extractive production as a regional indicator) and tobacco (cultivated area).
• Minas Gerais and South: rurality and lower municipal GDP per capita.
• Well water: in the South, the association remains in the joint model; in Minas Gerais, it appears in the single-variable analysis but loses precision in the joint model (the confidence interval includes 1). Spring water: associated in the single-variable analysis and in the national joint model without firewood, but not in the South and Minas Gerais joint models.

Motivated by the literature, not tested here (no suitable municipal data): temperature of chimarrão (mate) and other drinks (IARC classified very hot beverages, above 65 °C, as probably carcinogenic, group 2A); hand-rolled or rope tobacco; cachaça and home-distilled spirits; indoor wood stoves; nitrate and other water contaminants; diets poor in fruit and vegetables. Phase 2b (https://cledon.org/fase2b/en) looks further into Minas Gerais.

10. Limitations

• Ecological study: associations between municipalities do not show what happens to people (ecological fallacy).
• Local production is a rough proxy for consumption; firewood produced does not measure wood-stove use.
• Data not missing at random: plant extraction is missing mainly in some states, which may bias those estimates.
• Residual confounding by poverty and rurality is likely; GDP per capita is not household income.
• Regional and sensitivity analyses were done after seeing the national results.
• Analyses by an independent researcher with no formal training in health sciences, with AI assistance (Anthropic's Claude) for code and text, under human review. No machine learning model was trained or evaluated.

11. Reproduction

Code at https://cledon.org/codigo. Plan: PLANO-fase2.md (commit 0d88889). After Phases 1 and 1b: exposicoes.py, hipoteses.py, gerar_fase2.py. CSV results on the map page and in the Zenodo deposit (fase2_*.csv).

In memory of Francisco Cledio Capistrano (F.C.C.).