Technical report · Phase 1

Cancer mortality by municipality in Brazil, 2020–2024: an exploratory study

A document for anyone who wants to check or reproduce the analysis: sources and versions, definitions, exclusions, methods, parameters, sensitivity tests and limitations. Versão em português.

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

Abstract

Objective: to identify Brazilian municipalities with cancer mortality above what their population's age and sex would predict, 2020–2024, as hypotheses for investigation. Methods: ecological study using 7,913,003 death certificates from SIM/DATASUS (1,202,950 with cancer as the underlying cause) and the 2022 Census population, across 5,570 municipalities; indirect standardization by age and sex, exact confidence intervals, Poisson tests with Benjamini-Hochberg correction within each of 13 analysis groups and empirical Bayes smoothing; exploratory geographic criteria and sensitivity analyses. Results: 145 statistical signals in 93 municipalities; 45 remain after the geographic criteria. Esophageal cancer accounts for 16 of these signals (5 in the South, 9 in Minas Gerais and 2 in Bahia), 14 of which persist in every sensitivity scenario; in Minas Gerais and Bahia, the smoothed SMR relative to Brazil ranges from 2.39 to 3.63. In 2015–2019, a period not used to select the signals, and with an analysis plan recorded locally before the data were examined, 29 of the 45 municipality–group pairs after the geographic criteria and 12 of the 16 for esophageal cancer again showed excess mortality under the defined criterion. These results support the temporal persistence of the observed patterns, but do not distinguish real excesses from persistent biases, nor do they establish causation. Limitations: ecological design, mortality rather than incidence, possible errors in underlying cause and address, unvalidated geographic criteria, effects of the COVID-19 pandemic and lack of independent review. The signals are hypotheses, not conclusions.

1. Question and design

Question: in which municipalities was cancer mortality, from 2020 to 2024, above what their population's age and sex would predict, with strong statistical evidence? Design: ecological, descriptive and exploratory study, with the municipality as the unit of analysis. Outcome: death with cancer as the underlying cause (CAUSABAS field of the death certificate). Each death is assigned to the municipality of residence (CODMUNRES field), not the place of death.

2. Data sources and versions

3. Definitions

Group names follow the codes used: "Colon, rectum and anus" (C18–C21) includes colon, rectosigmoid junction, rectum, anus and anal canal; "Liver and intrahepatic bile ducts" corresponds to C22; "Lip, oral cavity and pharynx" to C00–C14. "All cancers" contains the other groups, which is why we speak of 13 analysis groups rather than 13 cancer types.

Auxiliary groups, used only for quality and context: all causes (A00–Z99) and ill-defined causes (R00–R99). Matching uses the first three characters of the underlying cause. Age groups: 0-4, 5-9, 10-14, 15-19, 20-24, 25-29, 30-34, 35-39, 40-44, 45-49, 50-54, 55-59, 60-64, 65-69, 70-74, 75-79, 80+ years. Age: IDADE field unit 0–3 (under 1 year) = 0; unit 4 = years; unit 5 = 100 years or older.

4. Processing and exclusions

7,913,003 all-cause death certificates were kept, of which 1,202,950 had cancer (C00–C97) as the underlying cause and 403,341 (5.1%) an ill-defined cause. 18,361 certificates (0.23% of the total) were excluded for lacking information required for standardization:

Exclusions are applied in this order, to all causes. Missing values are not imputed. All 5,570 municipalities in the 2022 Census are included, including those with zero deaths in a group.

5. Statistical methods

Expected deaths (indirect standardization). For each group and reference (Brazil or the municipality's state), the reference rate is r_s,a = deaths_s,a / person-years_s,a, by sex s and age group a. Expected deaths for municipality i are E_i = Σ_s,a r_s,a × P_i,s,a × 5, where P is the 2022 Census population. By construction, the sum of expected deaths equals the sum of observed deaths for each reference.

SMR (standardized mortality ratio) = O_i / E_i, with an exact 95% confidence interval (Garwood method, via the chi-square distribution). Test: one-sided Poisson, p = P(X ≥ O_i | E_i). Multiple-comparison correction: Benjamini-Hochberg (FDR) applied within each analysis group, across the 5,570 municipalities. Correction across all 13 analysis groups together was tested in the sensitivity analysis (scenario C).

Empirical Bayes smoothing (Marshall, 1991): Poisson-Gamma model, with a Gamma prior estimated by the method of moments, one per group and reference. Smoothed SMR = (O + α) / (E + β), and the posterior probability that SMR > 1 comes from the Gamma(O + α, E + β) distribution. This is why the smoothed SMR is usually lower than the direct ratio O / E in small municipalities: in Capelinha (esophagus), for example, 37 / 9.75 = 3.79 unsmoothed and 2.44 smoothed, with a 95% CI of 2.67 to 5.23 for the unsmoothed SMR.

Signal criteria (reference: the municipality's state): q < 0.01; P(SMR > 1) > 0.99; smoothed SMR ≥ 1.3; at least 10 observed deaths. The state reference reduces, but does not eliminate, regional differences in diagnosis and recording.

Exploratory geographic criteria. Each signal receives a context: state capital (27 capitals); capital surroundings (approximate center of the municipality within 40 km of a capital's center); treatment hub (at least 200 cancer deaths of non-residents occurring in the municipality and at least 25% of its residents' cancer deaths). The rest are shown as "after geographic criteria". The thresholds were chosen in an exploratory way, without external validation. These criteria do not demonstrate that biases were removed and may exclude real excesses. The CDC guidelines stress that geographic choices can create or hide patterns and require expert interpretation. The thresholds were therefore varied in the sensitivity analysis.

6. Results

145 signals in 93 municipalities: 47 in state capitals, 9 in capital surroundings, 44 in treatment hubs and 45 after the exploratory geographic criteria. Counts by analysis group are in the table in section 3.

7. Sensitivity analyses

Each scenario changes one methodological choice and keeps the others. "In common" counts municipality–group pairs that remain after the geographic criteria both in the scenario and in the main analysis. In scenario D, the 2022 age- and sex-specific population is rescaled by each year's estimated total. The 2020 and 2021 estimates were made before the Census and are higher than it, so this scenario tests sensitivity to population, not a "correct" population.

8. Esophageal cancer: stability of the signals

14 of the 16 municipalities with a signal after the geographic criteria remain in every sensitivity scenario. Period SMRs are unsmoothed: deaths observed in the period divided by expected deaths relative to Brazil, split in proportion to the number of years (3/5 and 2/5). This is a stability check, not independent validation: the municipalities were selected using all five years together, so the excess tends to appear in both periods by construction. Validation in an independent period is in section 9.

5 of these municipalities are in the South (PR, SC and RS). In this region, the literature already associates esophageal cancer (especially squamous cell carcinoma) with drinking very hot beverages, such as maté (chimarrão) at very high temperatures, as well as alcohol and tobacco; IARC classifies drinking very hot beverages, above 65 °C, as probably carcinogenic (group 2A). This study did not measure any of these exposures, and consistency with the literature does not confirm causation. For the municipalities in Minas Gerais and Bahia, the data analyzed do not point to an explanation.

Table 8a. Deaths, SMR and confidence interval (2020–2024).

Table 8b. Stability by period and across sensitivity scenarios.

9. Validation in an independent period (2015–2019)

The signals were selected using 2020–2024 deaths. To test whether they could be due to chance, the same municipality–group pairs were tested in 2015–2019, a period not used for selection. Nothing was reselected: there are no new thresholds or geographic criteria. Analysis plan recorded locally: the criterion was defined and recorded in version control (git, commit 1112c34, 2026-10-03 22:35 (UTC−3)) before the 2015–2019 data were processed; the validar.py file used in the analysis is identical to the one recorded in that commit. A signal replicates if, in 2015–2019, the unsmoothed SMR relative to its own state is greater than 1 with a one-sided Poisson p < 0.05. The number of replications was compared with the 5% expected if there were no excess (one-sided binomial test, as specified). The record is kept in the author's repository: it documents the order of steps but does not independently prove when the data were examined, as a dated record on a public platform would. We therefore do not call it a formal pre-registration.

Data. The DATASUS FTP server was unavailable on 2026-10-03 and 2026-10-04. Deaths for 2015–2019 were obtained from the national OpenDataSUS CSV files, which have the same fields. Before use, both sources were compared for 2020: counts by municipality, sex, age group and analysis group were identical in 100% of the 292,219 cells (commit 7a233dd). The 2015–2019 population was estimated by linear interpolation between the 2010 and 2022 Censuses at the central year (2017), by municipality, sex and age group; the 5 municipalities created after 2010 use the 2022 population.

Inclusions and exclusions in 2015–2019. 6,529,667 all-cause death certificates were kept, of which 1,089,385 had cancer as the underlying cause. 23,465 were excluded, for the same reasons and in the same order as in the main analysis: municipality of residence unknown or invalid (15,169), sex unknown (2,157) and age unknown (6,139). The SHA-256 of each of the 755 input files (of this analysis, the validation and Phases 1b to 2b, with the uso column giving the stage) is in hashes_entradas.csv.

Table 9a. Replication of the signals in 2015–2019. "Pairs" are municipality–group combinations; the number of distinct municipalities is in parentheses.

On expected values and p. The 5% is a nominal reference: because the Poisson test is discrete, the effective probability that a pair "replicates" without real excess is lower (3.3% to 4.6% on average across the sets), computed exactly from each pair's expected deaths (complementary, not pre-specified analysis, in validar_complementar.py). The p-value in the table assumes independence between pairs, which does not fully hold: some municipalities have more than one pair, groups overlap ("All cancers" contains the others) and state references are shared. These p-values, however small, are therefore not the main support for the interpretation, and a p-value does not give the probability that the data were produced by chance alone. The esophagus set is the least affected: 16 distinct municipalities and a single group.

Table 9b. Esophageal cancer: pooled SMR in 2015–2019 relative to each municipality's state.

Table 9c. The 16 municipalities with an esophageal cancer signal, in 2015–2019.

Interpretation. In 2015–2019, a period not used to select the signals, 29 of the 45 municipality–group pairs after the geographic criteria and 12 of the 16 for esophageal cancer again showed excess mortality under the defined criterion; all 16 esophageal pairs kept an SMR above 1. These results support the temporal persistence of the observed patterns, but do not distinguish real excesses from persistent biases, nor do they establish causation. This is a temporal replication within the same recording system. Among pairs in state capitals, capital surroundings and treatment hubs, potentially subject to differences in diagnosis, recording and address, 91 of 100 also replicated; replication cannot tell recording differences from real excesses in those places. The geographic criteria identify contexts but do not prove the presence or absence of bias. 2015–2019 values are lower than 2020–2024 values (median ratio of 0.61 for the esophagus), which is consistent with regression to the mean but may also reflect changes over time and in the denominators, since the 2015–2019 population is interpolated between Censuses. The most informative next step is to compare these results with record quality, cancer incidence and local information on diagnosis and treatment, with the participation of epidemiologists.

Additional checks (Phase 1b). A separate report, with its own analysis plan, cross-checks the 16 esophageal cancer municipalities against the INCA Hospital Cancer Registry, searches for spatial clusters and describes annual trends from 2015 to 2024: https://cledon.org/fase1b/en. Phase 1c goes back to 1996–2014, with data that played no part in the selection: https://cledon.org/fase1c/en. Phase 2 describes municipal characteristics associated with the excess, to generate hypotheses: https://cledon.org/fase2/en; Phase 2b looks further into Minas Gerais: https://cledon.org/fase2b/en.

10. Where artificial intelligence was used

AI (Anthropic's Claude) was used as a programming and writing assistant: it wrote the code from specifications and helped explore the data, review and translate the texts, always under human review. The analysis is classical epidemiology and statistics (indirect standardization, Poisson tests, FDR and empirical Bayes smoothing). No machine learning model was trained, tuned or evaluated in this phase. For the future clinical pilot with predictive models, reporting will follow the TRIPOD+AI guideline.

11. Limitations

• Ecological study: associations at the municipal level do not apply to individuals (ecological fallacy).
• Mortality is not incidence: it also reflects diagnosis, staging and access to treatment.
• The quality of underlying-cause certification varies across municipalities; ill-defined causes underestimate cancer.
• The residential address may be wrong, for example when patients report the address of the town where they are treated.
• Fixed 2022 population in the main analysis; municipalities with large population changes may have imprecise expected deaths.
• Many comparisons: even with FDR, some signals may be false positives, and small municipalities have little statistical power.
• The signals were not compared with incidence data (population-based cancer registries) or with exposures.
• COVID-19 pandemic: in 2020–2021, cancer diagnoses and treatments were postponed and there was excess mortality from other causes, which may affect cancer mortality unevenly across municipalities.
• The 2015–2019 validation is a temporal replication within the same recording system: it supports the persistence of the patterns but does not distinguish real excesses from persistent biases; and the analysis plan was recorded in the author's repository, not on an independent public platform.

12. Statements

• Ethics: processing used SIM microdata made publicly available without access restrictions by DATASUS (death certificates without names, with fields such as municipality of residence, sex, age and underlying cause), as well as aggregated IBGE data. The microdata were read only on the author's computer to produce counts by municipality, group, sex and age group; no individual data were published, linked to other sources or used to identify anyone. Published results are aggregated by municipality and group over five years, a level of detail equal to or coarser than DATASUS's own public tabulations. In the author's assessment, the study falls under the cases exempt from registration and review by the Brazilian CEP/Conep research ethics system in National Health Council Resolution No. 510/2016, art. 1, sole paragraph, item II (publicly accessible information, under Law No. 12,527/2011) and item V (aggregated results with no possibility of individual identification). The study was not submitted to an ethics committee; this assessment is not a formal waiver issued by a committee.
• Funding: none.
• Competing interests: none.
• Data and code availability: source data are public (DATASUS and IBGE). Aggregated results, code and this report are at https://cledon.org/relatorio/en and in the Zenodo deposit doi.org/10.5281/zenodo.24195057. Licenses: CC BY 4.0 for the report and results; MIT for the code.
• Authorship: conception, analysis and writing by Clériston Almeida Capistrano, an independent researcher with no formal training in health sciences, with AI assistance as described in section 10.

13. How to reproduce

The complete code (https://cledon.org/codigo; comments in Portuguese) runs on Python 3.14.4 (pandas 3.0.6, numpy 2.5.3, scipy 1.18.1, dbfread 2.0.7, datasus-dbc 0.1.3). Order: baixar.py → preparar.py → analisar.py → sensibilidade.py → gerar_mapa.py → gerar_relatorio.py. Validation: baixar.py validacao (or validacao-csv if the FTP is unavailable) → preparar.py validacao → validar.py; the source equivalence test is in equivalencia_fontes.py; the complementary analysis in validar_complementar.py; the hashes in registrar_hashes.py. Validation results: validacao_sinais.csv, validacao_resumo.csv, validacao_complementar.csv and hashes_entradas.csv. Full results are in municipios_rmp.csv, sinais.csv, sensibilidade.csv and parametros_suavizacao.csv (at https://cledon.org/mapa); the Zenodo deposit (doi.org/10.5281/zenodo.24195057) contains the same files, the code of this version and this report. DATASUS files may be republished with corrections, which is why the dates in section 2 matter.

14. References

• Marshall RJ. Mapping disease and mortality rates using empirical Bayes estimators. Applied Statistics, 1991;40(2):283–294.
• Benjamini Y, Hochberg Y. Controlling the false discovery rate. Journal of the Royal Statistical Society B, 1995;57(1):289–300.
• CDC/ATSDR. Guidelines for Examining Unusual Patterns of Cancer and Environmental Concerns, 2022.
• IARC. Drinking Coffee, Mate, and Very Hot Beverages. IARC Monographs, vol. 116, 2018.
• Collins GS et al. TRIPOD+AI statement. BMJ, 2024;385:e078378.