Report · Phase 1b

Esophageal cancer: hospital registry, spatial clusters and trends

Three additional checks on the Phase 1 esophageal cancer signals (technical report). Versão em português.

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

Exploratory study — not conclusions. Ecological, descriptive studies that have not yet been reviewed by independent specialists. They do not indicate causation, do not allow estimating the risk of any specific person and do not change any medical recommendation. Version 1.7 · November 13, 2026. DOI: 10.5281/zenodo.24195057. This is a translation of the Portuguese report; in case of discrepancy, the Portuguese version prevails.

Abstract

Phase 1 found 16 municipalities (outside state capitals, their surroundings and treatment hubs) with esophageal cancer mortality above the expectation for their own state in 2020–2024, and showed that the excess was already present in 2015–2019. Phase 1b runs three additional checks, with an analysis plan recorded locally in git (commit 84533e1) before any result. Anything not in the plan is labelled as not pre-specified.

Hospital cancer registry (RHC/INCA), a source different from death certificates. We compared the share of esophageal cancer among the cancer cases of residents treated in hospitals with that of their own state (proportional incidence ratio, PIR). All 16 had a PIR above 1; 14 met the criterion PIR > 1 and p < 0.05. The pooled PIR of 2.69 (95% CI 2.38 – 3.04) describes the share of esophageal cancer among cancers recorded in hospitals; it does not mean that the population risk is 2.69 times higher.
Spatial clusters. The scan found 24 areas with excess esophageal cancer mortality within their own state. 15 of the 16 municipalities lie inside them, and 15 areas contain none of the 16. 22 of the 24 areas also had an excess in 2015–2019.
Trends. Considering the 16 municipalities together, the esophageal cancer SMR remained above 1 in every year from 2015 to 2024; individually, this held for 6 of them. In the scan areas, the excess persists throughout the period, with a slight increase.
These results support that the excess is persistent and appears in two different sources. They do not distinguish a real excess from biases shared by both sources, nor do they establish causation.

1. Cross-check with the Hospital Cancer Registry (RHC/INCA)

Why. Mortality comes from death certificates. The RHC records cases treated in hospitals, with the diagnosis and usually the tumour histology. If the excess also appears there, it is less likely to be only a death-certificate completion problem.

How. National bases of the RHC Integrator (INCA), by year of first visit, 2015–2022 (2023 as sensitivity, since it is incomplete); analytic cases; patient's municipality of residence; sites C00–C97 except C44. Because the RHC does not cover all hospitals, we do not compare with the population. We use the proportional incidence ratio (PIR): observed ÷ expected esophageal cases, with the expectation given by the share of esophageal cancer among cancers in the same state, by sex and age group. "Corroborates" = PIR > 1 and p < 0.05 (one-sided Poisson).

A problem found and fixed. In the national base downloaded without São Paulo hospitals (3,970,130 rows over 9 years), every row appears an even number of times: in each year, all identical rows (all fields, including hospital and dates) occur 2, 4, 6 or more times, and none occurs an odd number of times. In 2019, for example, 257,655 rows occurred twice, 585 four times and 82 six or more times. We interpret this as the whole file being duplicated; rows occurring 4, 6 or more times would be records already repeated at the source. We therefore divide by two the number of times each row appears. An independent check supports this: in 2019, half of the rows in this base (259,093) is exactly the number of rows from hospitals outside São Paulo in the national base that includes SP, which is not duplicated. The correction does not change the ratios, but avoids wrongly narrow confidence intervals.

Table 1. The 16 municipalities in the RHC, 2015–2022.

Municipality	UF	SMR (SIM 2020–24)	Cancer cases in RHC	Esophagus obs.	Esophagus exp.	PIR	95% CI	p	Corroborates	RHC cases ÷ SIM deaths
Matina	BA	5.00	51	6	1.7	3.44	1.26 – 7.49	0.009	yes	0.40
Palmas de Monte Alto	BA	3.40	75	8	2.0	3.90	1.69 – 7.69	0.001	yes	0.50
Brasília de Minas	MG	2.65	229	30	10.4	2.90	1.95 – 4.13	< 0.001	yes	0.88
Braúnas	MG	5.44	56	6	2.2	2.73	1.00 – 5.94	0.025	yes	0.46
Capelinha	MG	3.79	297	30	11.8	2.54	1.72 – 3.63	< 0.001	yes	0.75
Caraí	MG	3.06	131	19	5.8	3.30	1.99 – 5.15	< 0.001	yes	0.66
Cipotânea	MG	5.78	100	20	4.7	4.22	2.58 – 6.52	< 0.001	yes	1.25
Ervália	MG	3.03	170	17	6.7	2.55	1.49 – 4.09	< 0.001	yes	0.77
Poté	MG	3.68	124	15	5.5	2.73	1.53 – 4.50	< 0.001	yes	0.71
Santa Maria do Suaçuí	MG	4.14	96	17	4.6	3.69	2.15 – 5.91	< 0.001	yes	0.57
Serro	MG	3.29	138	13	6.5	1.99	1.06 – 3.41	0.016	yes	0.42
Mallet	PR	4.32	147	14	4.8	2.94	1.61 – 4.94	< 0.001	yes	0.64
Prudentópolis	PR	2.48	512	38	16.3	2.33	1.65 – 3.19	< 0.001	yes	0.64
Dom Feliciano	RS	3.37	139	10	5.5	1.80	0.87 – 3.32	0.056	no	0.42
Riozinho	RS	6.18	29	3	1.1	2.80	0.58 – 8.19	0.094	no	0.27
Itaiópolis	SC	3.28	255	16	7.7	2.08	1.19 – 3.37	0.006	yes	0.59

Table 2. Pooled PIR by region, 2015–2022.

Region	Municipalities	Esophagus obs.	Esophagus exp.	Pooled PIR	95% CI	p
Minas Gerais	9	167	58.1	2.87	2.45 – 3.34	< 0.001
South (PR, SC, RS)	5	81	35.4	2.29	1.82 – 2.84	< 0.001
Bahia	2	14	3.8	3.69	2.02 – 6.19	< 0.001
All 16	16	262	97.3	2.69	2.38 – 3.04	< 0.001

Sensitivity. Including 2023: 14 of 16. With the national base that includes São Paulo hospitals (downloaded separately, without duplication; not pre-specified): 14 of 16, pooled PIR 2.71 (95% CI 2.39 – 3.06).

Overall agreement. Among municipalities with at least 5 expected esophageal cancer deaths, the mortality SMR and the RHC PIR move together: Spearman correlation 0.52 in the states with signals (746 municipalities) and 0.40 across Brazil, with the base that includes SP.

Table 3. Descriptive data on esophageal cases in the RHC, 2015–2022 (no threshold).

Set	Esophageal cases	Squamous (%)	Adenocarcinoma (%)	Alcohol recorded (%)	Alcohol: current or former (%)*	Tobacco recorded (%)	Tobacco: current or former (%)*
The 16 municipalities	262	84.7	8.8	88.5	71.5	91.6	82.1
State: MG	7,708	81.5	9.6	94.0	80.3	95.2	83.7
State: PR	2,624	76.2	13.3	79.9	67.0	83.8	82.5
State: SC	1,788	54.4	21.7	55.8	75.8	67.3	85.1
State: RS	5,007	66.1	14.5	48.9	65.0	56.8	81.0
State: BA	2,407	75.6	10.1	80.5	66.8	81.5	71.7

* Among cases with the field filled in. Alcohol and tobacco have many "no information" entries and do not measure amount or type of drink.

In the 16 municipalities, 84.7% of esophageal cases are squamous cell carcinoma, the type associated with very hot drinks, alcohol and tobacco, slightly above the states. The recorded shares of alcohol and tobacco use are above those of some states and below others, with different completeness and patient mix. These descriptive data cannot determine whether differences in alcohol and tobacco exposure explain the signals found.

Caveats. The RHC is not population-based, and INCA itself notes that hospital registries do not measure population incidence. In most of the 16 municipalities there are fewer RHC cases than SIM deaths in the same years (last column of Table 1). This ratio may suggest coverage differences but does not directly measure completeness, because first visits and deaths in the same period are not necessarily the same people. The PIR is valid only if coverage does not depend on cancer type; for example, if esophageal cancer patients are not referred to RHC hospitals more often than patients with other cancers. The RHC years partly overlap the period used to select the municipalities (2020–2024), but the source is different.

2. Spatial clusters (Kulldorff scan)

How. Across Brazil, we searched for circles of neighbouring municipalities, up to 100 km in radius and 1 million inhabitants, with more esophageal cancer deaths in 2020–2024 than expected for their own state (Poisson model). The p-value comes from 999 simulations. The method looks at regions, not isolated municipalities: it can find excesses spread over several small municipalities that would not be signals on their own. Non-overlapping secondary clusters were kept if p < 0.05. Each area was tested in 2015–2019 (replicates if SMR > 1 and p < 0.05).

municipalities inside the 24 esophageal cancer clusters (p < 0.05) · the 16 municipalities with a signal. Sources: SIM/DATASUS, 2022 Census and IBGE boundaries.

Table 4. The 24 esophageal cancer clusters, in order of statistical strength.

#	Centre	UF	Radius (km)	Municipalities	Signals inside	Deaths obs.	Exp.	SMR 2020–24	p	SMR 2015–19	Replicates	PIR in RHC*	Annual change
1	Angelândia	MG	98	41	4	405	150	2.71	0.001	2.35	yes	2.75 (2.50–3.03)	2.8%
2	São Mateus do Sul	PR	94	36	2	364	189	1.93	0.001	2.10	yes	1.67 (1.50–1.85)	-0.9%
3	Sebastião Laranjeiras	BA	94	22	2	209	105	1.99	0.001	1.44	yes	2.00 (1.65–2.40)	4.7%
4	São José do Belmonte	PE	87	40	0	260	148	1.76	0.001	1.66	yes	1.83 (1.60–2.09)	1.2%
5	Alvorada de Minas	MG	81	38	2	194	102	1.89	0.001	1.79	yes	1.70 (1.47–1.96)	-0.6%
6	Patis	MG	98	24	1	352	224	1.57	0.001	1.64	yes	1.62 (1.46–1.79)	0.4%
7	Goioxim	PR	99	40	1	286	175	1.63	0.001	1.57	yes	1.67 (1.48–1.89)	-0.7%
8	Itaocara	RJ	61	27	0	194	107	1.82	0.001	1.75	yes	1.65 (1.42–1.91)	-0.2%
9	Santa Cruz do Escalvado	MG	74	60	1	410	279	1.47	0.001	1.46	yes	1.68 (1.54–1.83)	0.5%
10	Itambé	BA	79	20	0	275	172	1.59	0.001	1.26	yes	1.41 (1.18–1.67)	4.5%
11	Nanuque	MG	93	30	0	309	202	1.53	0.001	1.54	yes	1.93 (1.74–2.14)	0.7%
12	Alpercata	MG	54	25	0	236	148	1.59	0.001	1.50	yes	1.76 (1.56–1.99)	0.4%
13	Ivorá	RS	88	42	0	357	249	1.43	0.001	1.08	no	1.30 (1.17–1.43)	3.4%
14	Rio das Flores	RJ	46	18	0	207	133	1.56	0.001	1.62	yes	0.72 (0.53–0.96)	-0.8%
15	São Domingos	GO	90	13	0	48	18	2.68	0.001	2.10	yes	few cases	6.8%
16	Padre Carvalho	MG	60	13	0	92	47	1.94	0.001	1.58	yes	1.62 (1.25–2.08)	7.4%
17	Poço Dantas	PB	84	72	0	290	206	1.41	0.001	1.42	yes	1.51 (1.32–1.73)	0.2%
18	Riozinho	RS	15	2	1	33	11	3.10	0.002	1.75	yes	1.95 (1.07–3.27)	13.0%
19	São José do Cedro	SC	43	22	0	96	53	1.80	0.006	1.18	no	1.40 (1.12–1.73)	7.4%
20	Casa Branca	SP	56	24	0	274	197	1.39	0.008	1.14	yes	1.11 (0.97–1.27)	3.5%
21	Alto Rio Doce	MG	16	4	1	34	12	2.79	0.008	2.28	yes	2.55 (1.86–3.42)	3.3%
22	Rosário do Sul	RS	98	8	0	208	146	1.43	0.021	1.37	yes	1.42 (1.22–1.65)	-0.7%
23	Arapeí	SP	57	20	0	243	177	1.37	0.041	1.58	yes	0.60 (0.36–0.94)	-1.6%
24	Eunápolis	BA	44	5	0	69	37	1.86	0.047	1.41	yes	2.24 (1.64–2.98)	3.3%

* Not pre-specified: the same RHC PIR, 2015–2022, applied to the area of each cluster, using the base that includes São Paulo hospitals. With 999 simulations, the smallest possible p-value is 0.001. Annual change: see section 3.

22 of the 24 clusters also had an excess in 2015–2019. In the RHC, 20 of the 24 have a PIR above 1 with p < 0.05. In the other four the sources disagree: Rio das Flores (RJ) and Arapeí (SP) have a share of esophageal cancer below expectation in the RHC (PIR 0.72 and 0.60); Casa Branca (SP) is slightly above, not significant (1.11); and São Domingos (GO) has few RHC cases. Of the 16 municipalities with a signal, only Dom Feliciano (RS) is outside the clusters.

Other cancer groups (exploratory). The same scan for the other 12 groups, without correction across them. These clusters include state capitals and large cities, where differences in diagnosis, recording and address weigh more, and were not examined in detail.

Group	Clusters (p < 0.05)	Include a capital	Single municipality
All cancers	69	12	28
Lung, trachea and bronchus	23	0	3
Colon, rectum and anus	21	7	10
Stomach	17	1	2
Prostate	14	0	1
Breast (women)	11	5	3
Cervix	11	1	2
Lip, oral cavity and pharynx	9	1	1
Brain, meninges and other CNS	6	0	2
Liver and intrahepatic bile ducts	5	0	1
Pancreas	5	2	3
Leukemias	0	0	0

3. Annual trends, 2015–2024

How. Deaths by year, read again from the same SIM files (the totals match the Phase 1 bases exactly). Expected deaths use the state's rates in the same year; population interpolated between the 2010 and 2022 Censuses up to 2021, and the 2022 population afterwards. The annual change comes from a Poisson regression with expected deaths as offset; standard errors were corrected for overdispersion (quasi-Poisson), a detail not in the plan.

Standardized mortality ratio (SMR) for esophageal cancer in the 16 municipalities, by region. Above 1 = more deaths than expected for the state, at the same age and sex. The shaded band is the period used to select the municipalities.

Considering the 16 municipalities together, the SMR remained above 1 in every year from 2015 to 2024 (from 1.58 to 3.58). Individually, 6 of the 16 were above 1 in all years; Dom Feliciano was above in 6 of the 10 years and Riozinho in 7, with few deaths per year. The apparent rise (6.5% per year; 95% CI 3.4 to 9.7) is largely expected by construction, since the municipalities were selected for having an excess in 2020–2024. In the scan areas, the excess persists throughout the period but is not stable: together, the SMR changed 1.2% per year (95% CI 0.6 to 1.9), a slight increase; 18 of the 24 clusters were above expectation in all 10 years; 4 show an upward trend and 0 a downward one. The same caveat applies: the clusters were also selected for their excess in 2020–2024, which tends to produce positive slopes by construction. Phase 1c extends the series back to 1996–2014, with data that played no part in the selection: https://cledon.org/fase1c/en.

4. What this suggests, with caution

The excess esophageal cancer mortality in these municipalities (1) is not only in death certificates: it also appears in cases treated in hospitals, with histology; (2) is part of larger regional areas, some without any municipality that stood out on its own; and (3) persists throughout 2015–2024, considering the municipalities together. This makes investigating these regions more justified, but does not say why. Biases shared by both sources, such as access to diagnosis or the address recorded, remain possible, and the data do not measure individual exposures. The most informative next step is to bring these findings to epidemiologists and state health departments, and to compare them with population-based cancer registries where they exist. Phase 2 (https://cledon.org/fase2/en) describes municipal characteristics associated with this excess.

5. Limitations

Ecological studies: they refer to municipalities and regions, not to people.
The RHC covers only hospitals that submit data; coverage varies across states (Goiás, for example, has few cases). The national base without SP was duplicated (fixed).
Scan with circular windows and limits set in the plan (100 km, 1 million inhabitants); other limits may produce different areas. Clusters for the other 12 groups were not examined.
Interpolated annual population; the trends depend on this approximation.
Analyses by an independent researcher with no formal training in health sciences, with AI assistance (Anthropic's Claude) to write code and text, under human review. No machine learning model was trained or evaluated.

6. Reproduction

Code at https://cledon.org/codigo. Plan: PLANO-fase1b.md (commit 84533e1). Steps: after the Phase 1 steps (including validar.py): baixar.py rhc rhc-sp; contar_anual.py; varredura.py; rhc.py and rhc.py com-sp (which read the scan clusters); tendencias.py; gerar_fase1b.py. CSV results on the map page (https://cledon.org/mapa) and in the Zenodo deposit. SHA-256 of the RHC bases in hashes_entradas.csv.

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