# 3382 - Brazil Census design: injury negative controls, round 2 (SIM external-cause deaths)

**Documented reading for the pooled context (Brazil Maternal DTaP, BRA_MDTPA_census_autism): DESIGN BIAS DETECTED (r < 1).** Unintentional injury deaths (V01-X59, SIM, death years 2021-2023, age at death) on the same 108 state x age-band cells, same exposure path, kernels and prior: r 0.36 (0.15-0.79), P(r > 1) = 0.005, class BIAS SIGNAL (r < 1); autism on the same cells 0.77 (0.51-1.14).

In plain terms: fitted to deaths from unintentional injuries, which no vaccine can cause, the Brazil Census design returns a maternal dTpa 'effect' of 0.36 (0.15-0.79) (autism on the same cells: 0.77 (0.51-1.14)). By the documented rule the autism estimate on these cells carries non-specific bias in the protective direction and is not interpretable as a vaccine effect by itself; the rule does not say how much of the autism lean that bias explains (autism relative to injury deaths, DDD: 1.92 (0.97-3.57)). Round 1 found SUS injury admissions on the same cells at 0.51 (0.29-0.83) (BIAS SIGNAL, r < 1); deaths come from a different register and depend less on SUS hospital access, and the two tests are reported side by side without being combined.

Generated by `07_analysis_code/3382_injury_negative_controls_brazil_sim.py` (every number below is computed by that script and also written to `3382_injury_negative_controls_brazil_sim.csv` / `3382_injury_negative_controls_brazil_sim.json`). Figures: `08_figures/3382_injury_negative_controls_brazil_sim.png` (pooled context) and `08_figures/3382_injury_negative_controls_brazil_sim_products.png` (autism vs unintentional injury deaths under each product's design); hollow light markers = no weight under the rule (UNINFORMATIVE). Tests: `07_analysis_code/test_3382_injury_negative_controls_brazil_sim.py`. Fits cached under `project_results/negative_controls_3330/3382/`. Raw responses: `raw/brazil/sim_injury/`. Status: falsification audit of the Brazil Census design; ecological; no causal pool weight of its own; nothing here changes the hierarchy, `blocks.csv` or `gate_decisions.json`.

## Interpretation rule (documented)

Rule file `qa/negative_controls_20260925/3382_rule.json` (sha256 `cadfe07783ea11be7da2a42657b4018e34fd9b537e0cf327212793a0a2d30664`, recorded beside it in `qa/negative_controls_20260925/3382_rule.sha256` when written). 'written' field: 2026-09-29 22:09:22 Central Daylight Time (2026-09-30T03:09:22Z), before any SIM death count was downloaded or seen by this task (no TabNet SIM data POST had been sent; raw/brazil/sim_injury/ held no data response), and before any likelihood, posterior, placebo, DDD or gate of any outcome, and any refit of an autism block, was computed by this task (3382). The fit folder project_results/negative_controls_3330/3382/ did not exist or was empty. File mtime 2026-09-29T22:09:22-05:00; first SIM data POST 2026-09-30T03:14:01Z (rule before it: True); first fit 2026-09-29T22:23:42-05:00 (fit logs; rule before it: True). The hash was recorded at fit time in all 21 fit receipts, 9 gate records and 30 fit-log lines and is re-checked here; the report refuses to run on any mismatch.

- Round-1 class rules (3332) unchanged: expected r = 1; UNINFORMATIVE (a zero count in any of the 108 cells, or 95% CrI log width > 2.5x the autism CrI's on the same cells); CONSISTENT (95% CrI includes 1); BIAS SIGNAL (95% CrI excludes 1; direction reported).
- Design reading (primary window): DESIGN BIAS DETECTED (direction d) if all unintentional injuries is BIAS SIGNAL in d, or at least two informative components are BIAS SIGNAL in d; NO BIAS DETECTED if the composite is CONSISTENT and no informative component is BIAS SIGNAL; UNTESTED if the composite is UNINFORMATIVE and no component is informative; MIXED otherwise. Components: road traffic, falls, drowning, fire, accidental poisoning. Variants (Suffocation (W75-W84), All External Causes excl. Care (V01-Y98 excl. Y40-Y84)) are reported, never counted.
- Specified deviations from round 1 (in the rule): age at death is the only age convention (TabNet SIM has no birth-date dimension, so the round-1 primary census-cohort age is not available; the round-1 band_admit sensitivity is the like-for-like convention); the sensitivity window is the 2020-2024 pool (the round-1 2022-only window was not among the approved requests); no utilization reference (the table holds external-cause deaths only).
- Beside the reading, never changing it: future+3 placebo, DDD, the 2020-2024 sensitivity, the four report-only products, the harm-injection gate, grid checks, and the round-1 admission results.

## Data acquired (approved request BRA_NC_SIM_EXTCAUSE_CENSUS)

Eight TabNet POSTs to `http://tabnet.datasus.gov.br/cgi/tabcgi.exe?sim/cnv/ext10uf.def` (Linha Grupo CID10, Coluna Unidade da Federação, Incremento Óbitos p/Residênc, one Faixa Etária selection per band, every other filter 'Todas as categorias', `zeradas=exibirlz`), strictly sequential with at least 6 s between any two DATASUS requests; for each, the result page and the CSV it links were stored with a `.meta.json` sidecar (URL, POST payload, UTC time, SHA-256, size, HTTP headers, validation record).

| window | band | CSV (bytes) | sha256 | CSV URL | server response time (HTTP Date) | checks a-g |
|---|---|---|---|---|---|---|
| 2021-2023 | 0-4 | `raw/brazil/sim_injury/sim_ext10uf__grupo_x_uf__deaths_2021_2023__age_0_4.csv` (5,045) | `7f31afbd4009003b5593848803b48d9d362b7111412c9fb81258c46ba5ca4521` | http://tabnet.datasus.gov.br/csv/sim_cnv_ext10uf001501104_188_169_166.csv | Wed, 30 Sep 2026 03:15:07 GMT | all pass |
| 2021-2023 | 5-9 | `raw/brazil/sim_injury/sim_ext10uf__grupo_x_uf__deaths_2021_2023__age_5_9.csv` (4,876) | `9251d75cf7791ff9bde9bb4390fb9b9eafc46765e9ac95d12e76155eddc0be77` | http://tabnet.datasus.gov.br/csv/sim_cnv_ext10uf001514104_188_169_166.csv | Wed, 30 Sep 2026 03:15:21 GMT | all pass |
| 2021-2023 | 10-14 | `raw/brazil/sim_injury/sim_ext10uf__grupo_x_uf__deaths_2021_2023__age_10_14.csv` (4,961) | `d34f7d8da58ad002a60f2445bb1ac17ba8d6487424b8d8cc9e7476ce7c3945d6` | http://tabnet.datasus.gov.br/csv/sim_cnv_ext10uf001527104_188_169_166.csv | Wed, 30 Sep 2026 03:15:34 GMT | all pass |
| 2021-2023 | 15-19 | `raw/brazil/sim_injury/sim_ext10uf__grupo_x_uf__deaths_2021_2023__age_15_19.csv` (5,242) | `ab62b92374e1d23ef72cb23642b3334ef33f09cbfc87bddb56be2ec87bf7e1b2` | http://tabnet.datasus.gov.br/csv/sim_cnv_ext10uf001541104_188_169_166.csv | Wed, 30 Sep 2026 03:15:47 GMT | all pass |
| 2020-2024 | 0-4 | `raw/brazil/sim_injury/sim_ext10uf__grupo_x_uf__deaths_2020_2024__age_0_4.csv` (5,146) | `a6ee8ef416139a6fbfe26040ba2015d14a7259826856b136c35c3ef7de87ed6d` | http://tabnet.datasus.gov.br/csv/sim_cnv_ext10uf001554104_188_169_166.csv | Wed, 30 Sep 2026 03:16:00 GMT | all pass |
| 2020-2024 | 5-9 | `raw/brazil/sim_injury/sim_ext10uf__grupo_x_uf__deaths_2020_2024__age_5_9.csv` (4,976) | `06bf8c47dcc7732ab05b92864fc62f519b991d8fb8d387bf04a131cfee991536` | http://tabnet.datasus.gov.br/csv/sim_cnv_ext10uf001607104_188_169_166.csv | Wed, 30 Sep 2026 03:16:13 GMT | all pass |
| 2020-2024 | 10-14 | `raw/brazil/sim_injury/sim_ext10uf__grupo_x_uf__deaths_2020_2024__age_10_14.csv` (5,065) | `ac514917a2e93fb837f9b60e0110a1cd352287de615d21267feb2dd33ae1f78f` | http://tabnet.datasus.gov.br/csv/sim_cnv_ext10uf001620104_188_169_166.csv | Wed, 30 Sep 2026 03:16:26 GMT | all pass |
| 2020-2024 | 15-19 | `raw/brazil/sim_injury/sim_ext10uf__grupo_x_uf__deaths_2020_2024__age_15_19.csv` (5,373) | `78382af058466937e988d43d8c8182bb2aab63aaf0a045d83fbc0ead0787730d` | http://tabnet.datasus.gov.br/csv/sim_cnv_ext10uf001633104_188_169_166.csv | Wed, 30 Sep 2026 03:16:39 GMT | all pass |

Sidecar `retrieved_utc` is the client clock when the request was issued (before the pacing wait); the server Date header of every response is also stored. Consecutive DATASUS responses (metadata included) are 5-7 s apart by the server clock.

Result pages (HTML, same data): `sim_ext10uf__grupo_x_uf__deaths_2021_2023__age_0_4.result.html` 14,126 B sha256 `56be69ecac58`; `sim_ext10uf__grupo_x_uf__deaths_2021_2023__age_5_9.result.html` 13,937 B sha256 `e34cf8018b52`; `sim_ext10uf__grupo_x_uf__deaths_2021_2023__age_10_14.result.html` 14,023 B sha256 `764626f2ac34`; `sim_ext10uf__grupo_x_uf__deaths_2021_2023__age_15_19.result.html` 14,328 B sha256 `23a4861a0ac0`; `sim_ext10uf__grupo_x_uf__deaths_2020_2024__age_0_4.result.html` 14,240 B sha256 `3908b61e95cb`; `sim_ext10uf__grupo_x_uf__deaths_2020_2024__age_5_9.result.html` 14,044 B sha256 `7fc7be93d067`; `sim_ext10uf__grupo_x_uf__deaths_2020_2024__age_10_14.result.html` 14,134 B sha256 `913528aa2673`; `sim_ext10uf__grupo_x_uf__deaths_2020_2024__age_15_19.result.html` 14,478 B sha256 `b727429258d6`.

Metadata (GET): `br_uf.cnv` 1,802 B (http://tabnet.datasus.gov.br/cgi/territorio/br_uf.cnv); `br_ufalfa.cnv` 1,802 B (http://tabnet.datasus.gov.br/cgi/territorio/br_ufalfa.cnv); `br_ufsigla.cnv` 1,802 B (http://tabnet.datasus.gov.br/cgi/territorio/br_ufsigla.cnv); `CID10-grupos-CapXX.cnv` 2,739 B (http://tabnet.datasus.gov.br/cgi/sim/cnv/CID10-grupos-CapXX.cnv); `CID10-grupos-sel-CapXX.cnv` 2,458 B (http://tabnet.datasus.gov.br/cgi/sim/cnv/CID10-grupos-sel-CapXX.cnv); `ext10uf.def` 4,731 B (http://tabnet.datasus.gov.br/cgi/sim/cnv/ext10uf.def); `ext10uf_form.html` 51,213 B (http://tabnet.datasus.gov.br/cgi/deftohtm.exe?sim/cnv/ext10uf.def); `fxidad2.cnv` 926 B (http://tabnet.datasus.gov.br/cgi/sim/cnv/fxidad2.cnv); `fxidadp.cnv` 1,460 B (http://tabnet.datasus.gov.br/cgi/sim/cnv/fxidadp.cnv); `fxidops.cnv` 784 B (http://tabnet.datasus.gov.br/cgi/sim/cnv/fxidops.cnv). The live form's option lists equal the discovery snapshot; the Grupo CID10 row code list (35 leaf groups + 4 subtotal rows) and the UF column code list were used to map every row and column.

Rejected responses (kept under `raw/brazil/sim_injury/rejected/`): `sim_ext10uf__grupo_x_uf__deaths_2021_2023__age_0_4__attempt1.csv` (failed f_html_table_equals_csv; byte-identical to the accepted re-query: True). The one rejection was a parser defect in this script (the result page lists the TOTAL row first and uses "." thousands separators); after the parser was fixed the request was re-sent and accepted.

Acceptance checks per response (rule `acquisition.acceptance`): (a) title states Óbitos p/Residênc by Grupo CID10 and Unidade da Federação; (b) the Faixa Etária header line equals the requested bands; (c) the Período header line equals the requested death years, and no other filter line; (d) all 39 code-list rows (35 leaf groups mapped one-to-one onto groups 1-35 by ICD-10 block) and a Total row, every column a UF sigla or Total; (e) UF columns sum to the Total column, the 35 leaves to the Total row, and each subtotal row to its children; (f) the result page's HTML table, parsed independently, equals the CSV cell by cell; (g) no response byte-identical to another and no two value tables equal; (h) every 2021-2023 cell <= the same 2020-2024 cell: True. All eight re-validated from disk at the count stage.

## Provenance of the counts, fits and this report

- Counts (`project_results/negative_controls_3330/3382/counts/sim_ext_counts.csv`, sha256 `062ddfa7c41522e7...`): built 2026-09-29T22:23:16-05:00 by script version e51ed790; equal to the value in every fit receipt: True. Deaths with residence Ignorado/exterior in the requested bands: 2021_2023 0_4: 0, 2021_2023 5_9: 0, 2021_2023 10_14: 0, 2021_2023 15_19: 0, 2020_2024 0_4: 0, 2020_2024 5_9: 0, 2020_2024 10_14: 0, 2020_2024 15_19: 0.
- Fits and gates: all 21 fit receipts and 9 gate records record script version e51ed790; byte-exact copy `qa/negative_controls_20260925/snapshots/07_analysis_code/3382_injury_negative_controls_brazil_sim.e51ed790b7b4.py`.
- This report: script version 3d4946c1; differs from the fit-time version only in report-stage definitions (`REPORT_STAGE`, `acquisition_summary`, `plain_terms`, `write_md`).
- Round-1 results read from `3332_injury_negative_controls_brazil.json` (sha256 `d3109833df7e`), read-only.

## Design

- Cells and model: the production block built by `processed/global_adapters_americas.brazil_census` (imported read-only) fitted with `processed/international_likelihoods.fit` (unchanged): 6 kernels x the block's exposure scenarios (maternal dTpa: 7), equal weights, 801-point grid of ln r, 'reference' setting, N(0, ln 2) prior. Only y changes: y = ln(deaths / Census 2022 population of the cell); an outcome with zero deaths in a cell is fitted on its non-zero cells for information only (and classed UNINFORMATIVE).
- Outcome counts: SIM deaths by residence, underlying cause (chapter XX Grupo CID10 rows summed per outcome), residence UF, age at death, death years 2021-2023 pooled (primary) and 2020-2024 (sensitivity).
- Future+3 placebo: the block's own future-shifted exposure scenarios (birth times +3 years; `global_adapters_americas.FUTURE`) on the same cells; PASS iff the 95% CrI includes 1 under all four kernel settings.
- DDD: y = ln(autism / control deaths) on the control's cells, same likelihood.
- ESS: not applicable (deterministic grid quadrature). Kish effective number of grid points and a 1601-point refit are reported instead (largest relative change of median / CrI bounds over the 22 primary blocks: 2.3e-04).

## Autism reproduction

| product | published (own data) | refit here | max abs curve diff | future+3 published | future+3 refit | max abs curve diff | validated |
|---|---|---|---|---|---|---|---|
| Brazil Maternal DTaP (pooled context) | 0.77 (0.51-1.14) | 0.77 (0.51-1.14) | 0.0e+00 | 0.84 (0.64-1.14) | 0.84 (0.64-1.14) | 0.0e+00 | True |
| Brazil rotavirus (report only) | 0.95 (0.62-1.57) | 0.95 (0.62-1.57) | 0.0e+00 | 0.76 (0.54-1.12) | 0.76 (0.54-1.12) | 0.0e+00 | True |
| Brazil PCV10 + MenC 2010 (report only) | 1.19 (0.76-2.05) | 1.19 (0.76-2.05) | 0.0e+00 | 0.90 (0.63-1.41) | 0.90 (0.63-1.41) | 0.0e+00 | True |
| Brazil MMRV / varicella (report only) | 0.82 (0.64-1.83) | 0.82 (0.64-1.83) | 0.0e+00 | 0.87 (0.75-1.11) | 0.87 (0.75-1.11) | 0.0e+00 | True |
| Brazil hepatitis A (report only) | 2.17 (1.36-3.98) | 2.17 (1.36-3.98) | 0.0e+00 | 1.85 (1.23-3.09) | 1.85 (1.23-3.09) | 0.0e+00 | True |

## Deaths on the 108 cells and cell losses

| endpoint | window | zero cells (cells lost) | total deaths | min (cell) | median | national 0-4 / 5-9 / 10-14 / 15-19 |
|---|---|---|---|---|---|---|
| All Unintentional Injuries (V01-X59) | 2021-2023 | 0 | 18,670 | 8 (AP 5_9) | 99.5 | 5,875 / 1,774 / 2,239 / 8,782 |
| Road Traffic (V01-V89) | 2021-2023 | 1 | 8,635 | 0 (AP 5_9) | 35.0 | 858 / 710 / 1,032 / 6,035 |
| Falls (W00-W19) | 2021-2023 | 13 | 609 | 0 (RO 0_4) | 3.0 | 256 / 75 / 99 / 179 |
| Drowning (W65-W74) | 2021-2023 | 0 | 3,906 | 2 (SE 5_9) | 25.0 | 1,494 / 453 / 579 / 1,380 |
| Fire (X00-X09) | 2021-2023 | 29 | 288 | 0 (RO 5_9) | 2.0 | 108 / 65 / 40 / 75 |
| Accidental Poisoning (X40-X49) | 2021-2023 | 55 | 193 | 0 (RO 0_4) | 0.0 | 46 / 7 / 19 / 121 |
| Suffocation (W75-W84) | 2021-2023 | 7 | 3,043 | 0 (AC 5_9) | 6.0 | 2,612 / 151 / 117 / 163 |
| All External Causes excl. Care (V01-Y98 excl. Y40-Y84) | 2021-2023 | 0 | 42,556 | 12 (AP 5_9) | 158.0 | 7,036 / 2,129 / 3,993 / 29,398 |
| All Unintentional Injuries (V01-X59) | 2020-2024 | 0 | 31,486 | 21 (AC 5_9) | 171.0 | 9,570 / 3,013 / 3,753 / 15,150 |
| Road Traffic (V01-V89) | 2020-2024 | 0 | 14,823 | 1 (AP 5_9) | 57.5 | 1,472 / 1,214 / 1,772 / 10,365 |
| Falls (W00-W19) | 2020-2024 | 5 | 1,038 | 0 (RO 0_4) | 5.0 | 432 / 137 / 159 / 310 |
| Drowning (W65-W74) | 2020-2024 | 0 | 6,618 | 3 (DF 10_14) | 38.5 | 2,438 / 796 / 981 / 2,403 |
| Fire (X00-X09) | 2020-2024 | 15 | 455 | 0 (AC 0_4) | 2.0 | 191 / 98 / 59 / 107 |
| Accidental Poisoning (X40-X49) | 2020-2024 | 39 | 395 | 0 (RO 0_4) | 1.0 | 77 / 18 / 36 / 264 |
| Suffocation (W75-W84) | 2020-2024 | 4 | 4,871 | 0 (AC 5_9) | 9.0 | 4,162 / 242 / 200 / 267 |
| All External Causes excl. Care (V01-Y98 excl. Y40-Y84) | 2020-2024 | 0 | 71,964 | 26 (AC 5_9) | 263.5 | 11,515 / 3,618 / 6,641 / 50,190 |

Cell losses (primary window): Road Traffic (V01-V89): 1 of 108 cells with zero deaths (AP 5_9); Falls (W00-W19): 13 of 108 cells with zero deaths (RO 0_4; RO 5_9; AC 5_9; AC 10_14; AC 15_19; AP 5_9; RN 5_9; AL 10_14; SE 5_9; PR 10_14; DF 5_9; DF 10_14; DF 15_19); Fire (X00-X09): 29 of 108 cells with zero deaths (RO 5_9; AC 0_4; AC 10_14; AC 15_19; RR 5_9; AP 0_4; AP 5_9; AP 15_19; TO 0_4; TO 5_9; TO 10_14; TO 15_19; PI 15_19; CE 10_14; RN 0_4; RN 10_14; RN 15_19; PB 10_14; PB 15_19; SE 10_14; SE 15_19; ES 10_14; SC 10_14; MS 0_4; MS 5_9; MT 15_19; GO 10_14; DF 0_4; DF 10_14); Accidental Poisoning (X40-X49): 55 of 108 cells with zero deaths (RO 0_4; RO 5_9; RO 10_14; RO 15_19; AC 0_4; AC 5_9; AC 10_14; AC 15_19; AM 5_9; AM 10_14; RR 15_19; PA 0_4; PA 5_9; AP 0_4; AP 15_19; TO 5_9; TO 15_19; MA 10_14; PI 5_9; PI 10_14; CE 5_9; CE 10_14; RN 5_9; RN 10_14; RN 15_19; PB 0_4; PB 5_9; PE 5_9; AL 0_4; AL 5_9; AL 10_14; AL 15_19; SE 5_9; SE 10_14; SE 15_19; MG 5_9; ES 5_9; SP 5_9; SP 10_14; PR 10_14; SC 5_9; SC 10_14; RS 0_4; RS 5_9; RS 10_14; MS 0_4; MS 5_9; MS 10_14; MS 15_19; MT 5_9; MT 15_19; GO 5_9; GO 10_14; DF 0_4; DF 5_9); Suffocation (W75-W84): 7 of 108 cells with zero deaths (AC 5_9; RR 10_14; TO 10_14; PI 10_14; RN 5_9; RN 10_14; RN 15_19). Such an outcome is fitted on its non-zero cells for information only and is UNINFORMATIVE by rule. Not counted in any cell: deaths with residence Ignorado/exterior (0 in every requested band and window) and deaths with age 'Idade ignorada' (outside every requested band; their number was not queried). Structural losses against round 1: no census-cohort age convention and no 2022-only window (see rule).

## Results: pooled context (maternal dTpa), primary window (death years 2021-2023, age at death)

ESS is not applicable (exact grid quadrature); the column gives the Kish effective number of grid points (warning below 20). Asimov: posterior medians with nothing injected (R=1) and an injected doubling (R=2) in the noise-free null projection. Plasmode (lead decision D4): the doubling injected into the OBSERVED outcome; PASS iff its R=2 median > 1 (a level test, see 3332). Autism on the same cells is the published estimate wherever cells = 108; otherwise autism refitted on the control's cells.

| outcome | role | cells | deaths | r (95% CrI) | P(r>1) | autism same cells | CrI width / autism | class | future+3 r (verdict) | DDD autism / control | Kish ESS | Asimov R=1 / R=2 (verdict) | plasmode R=2 (D4) |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Autism (Census 2022) | reference | 108 | - | 0.77 (0.51-1.14) | 0.096 | - | 1.00 | - | 0.84 (0.64-1.14) (PASS) | - | 48 | 1.03 / 2.09 (PASS) | 1.70 (PASS) |
| All Unintentional Injuries (V01-X59) | composite | 108 | 18,670 | 0.36 (0.15-0.79) | 0.005 | 0.77 (0.51-1.14) | 2.08 | BIAS SIGNAL (r < 1) | 0.31 (0.15-0.57) (FAIL) | 1.92 (0.97-3.57) | 99 | 0.69 / 1.38 (PASS; null biased) | 0.63 (FAIL) |
| Road Traffic (V01-V89) | component | 107 | 8,635 | 0.81 (0.34-1.66) | 0.291 | 0.76 (0.51-1.14) | 1.99 | UNINFORMATIVE (zero deaths in 1 cell) | 0.77 (0.27-1.67) (PASS) | 1.05 (0.37-2.53) | 93 | 0.92 / 1.86 (PASS) | 1.28 (PASS) |
| Falls (W00-W19) | component | 95 | 609 | 1.26 (0.43-3.27) | 0.674 | 0.80 (0.52-1.20) | 2.45 | UNINFORMATIVE (zero deaths in 13 cells) | 2.54 (0.87-7.00) (PASS) | 0.63 (0.18-2.27) | 120 | 0.81 / 1.41 (PASS) | 1.51 (PASS) |
| Drowning (W65-W74) | component | 108 | 3,906 | 0.75 (0.27-1.90) | 0.279 | 0.77 (0.51-1.14) | 2.45 | CONSISTENT | 0.68 (0.31-1.38) (PASS) | 1.08 (0.41-2.53) | 117 | 0.68 / 1.29 (PASS; null biased) | 1.19 (PASS) |
| Fire (X00-X09) | component | 79 | 288 | 0.69 (0.20-2.38) | 0.279 | 0.80 (0.50-1.28) | 2.64 | UNINFORMATIVE (zero deaths in 29 cells) | 1.18 (0.37-3.40) (PASS) | 1.29 (0.39-3.82) | 150 | 0.75 / 1.45 (PASS; null biased) | 0.84 (FAIL) |
| Accidental Poisoning (X40-X49) | component | 53 | 193 | 0.77 (0.23-2.58) | 0.336 | 0.94 (0.58-1.52) | 2.50 | UNINFORMATIVE (zero deaths in 55 cells) | 0.92 (0.27-2.94) (PASS) | 1.37 (0.34-5.18) | 145 | 0.82 / 1.27 (PASS) | 0.85 (FAIL) |
| Suffocation (W75-W84) | variant | 101 | 3,043 | 0.66 (0.21-1.88) | 0.219 | 0.80 (0.52-1.21) | 2.63 | UNINFORMATIVE (zero deaths in 7 cells) | 0.38 (0.13-1.03) (PASS) | 1.53 (0.44-4.74) | 131 | 1.24 / 2.47 (PASS) | 0.80 (FAIL) |
| All External Causes excl. Care (V01-Y98 excl. Y40-Y84) | variant | 108 | 42,556 | 0.26 (0.10-0.58) | 0.000 | 0.77 (0.51-1.14) | 2.17 | BIAS SIGNAL (r < 1) | 0.17 (0.07-0.35) (FAIL) | 3.19 (1.57-6.33) | 104 | 0.82 / 1.69 (PASS) | 0.36 (FAIL) |

Class stability (descriptive, never changes the class; linear-only, spline-only, rho .85 kernel settings and the wide N(0, 1.2) and harm-centred N(ln 1.25, ln 2) priors): the class changes for all unintentional injuries (prior wide: UNINFORMATIVE); drowning (prior wide: UNINFORMATIVE); all external causes excl. care (prior wide: UNINFORMATIVE). The composite class changes (above).

## Reading

- Documented reading: **DESIGN BIAS DETECTED (r < 1)**. All unintentional injuries: 0.36 (0.15-0.79), P(r > 1) 0.005, class BIAS SIGNAL (r < 1), CrI log width 2.08x autism's (autism 0.77 (0.51-1.14) on the same cells). Informative components: 1 of 5: drowning 0.75 (0.27-1.90) CONSISTENT.
- Not counted (UNINFORMATIVE): road traffic 0.81 (0.34-1.66) (zero deaths in 1 cell); falls 1.26 (0.43-3.27) (zero deaths in 13 cells); fire 0.69 (0.20-2.38) (zero deaths in 29 cells); accidental poisoning 0.77 (0.23-2.58) (zero deaths in 55 cells). By the rule these carry no weight in either direction.
- Variants (never counted): suffocation (W75-W84) 0.66 (0.21-1.88) UNINFORMATIVE (zero deaths in 7 cells); all external causes excl. care 0.26 (0.10-0.58) BIAS SIGNAL (r < 1).
- DDD, autism relative to unintentional injury deaths on the same 108 cells: 1.92 (0.97-3.57), P(r > 1) 0.969. It restates the autism-versus-deaths comparison cell by cell; the rule does not let it change the reading.
- Future+3 placebo: autism 0.84 (0.64-1.14) (PASS); unintentional deaths 0.31 (0.15-0.57) (FAIL); fails for all unintentional injuries.
- Sensitivity (reported, not read): death years 2020-2024: reading DESIGN BIAS DETECTED (r < 1), unintentional 0.36 (0.15-0.77); classes differ from the primary for road traffic.
- Round 1 on the same cells (admissions, 3332): all injuries 0.51 (0.29-0.83) (BIAS SIGNAL, r < 1) under census-cohort age and 0.49 (0.28-0.81) under age at admission. Deaths here: 0.36 (0.15-0.79) (BIAS SIGNAL, r < 1). Each is read under its own documented rule; this task does not revise the round-1 reading.
- Report-only products on the same cells (separate readings): Brazil rotavirus: autism 0.95 (0.62-1.57) vs unintentional deaths 0.37 (0.19-0.86) (BIAS SIGNAL, r < 1; reading DESIGN BIAS DETECTED (r < 1)); Brazil PCV10 + MenC 2010: autism 1.19 (0.76-2.05) vs unintentional deaths 0.26 (0.15-0.54) (BIAS SIGNAL, r < 1; reading DESIGN BIAS DETECTED (r < 1)); Brazil MMRV / varicella: autism 0.82 (0.64-1.83) vs unintentional deaths 1.22 (0.61-1.96) (CONSISTENT; reading NO BIAS DETECTED); Brazil hepatitis A: autism 2.17 (1.36-3.98) vs unintentional deaths 0.75 (0.38-1.74) (CONSISTENT; reading MIXED).
- Detectability (production harm-injection gate on the maternal dTpa path; never part of a class). Asimov: autism PASS (R=2 median 2.09); 8 of 8 controls PASS (R=2 median: all unintentional injuries 1.38; road traffic 1.86; falls 1.41; drowning 1.29; fire 1.45; accidental poisoning 1.27; suffocation (W75-W84) 2.47; all external causes excl. care 1.69); null biased (R=1 median outside 0.8-1.25): all unintentional injuries (R=1 0.69); drowning (R=1 0.68); fire (R=1 0.75). D4 plasmode: autism PASS (1.70); 3 of 8 controls PASS, FAIL for all unintentional injuries (R=2 0.63, observed R=1 0.36); fire (R=2 0.84, observed R=1 0.69); accidental poisoning (R=2 0.85, observed R=1 0.77); suffocation (W75-W84) (R=2 0.80, observed R=1 0.66); all external causes excl. care (R=2 0.36, observed R=1 0.26).

## Deaths (this task) beside admissions (round 1, 3332) on the same cells: pooled context

Definitions differ (deaths: underlying cause; admissions: principal diagnosis or any-field external cause), so a row pairs each death outcome with its closest admission outcome. Reported only; the two are not combined and neither changes the other's reading.

| death outcome (3382) | r deaths, 2021-2023, age at death | class | admission outcome (3332) | r admissions, band_admit (age at admission) | class | r admissions, round-1 primary (census cohort) | class |
|---|---|---|---|---|---|---|---|
| All Unintentional Injuries (V01-X59) | 0.36 (0.15-0.79) | BIAS SIGNAL (r < 1) | All Injuries (S00-T98 excl. T78, T80-T88) | 0.49 (0.28-0.81) | BIAS SIGNAL (r < 1) | 0.51 (0.29-0.83) | BIAS SIGNAL (r < 1) |
| Road Traffic (V01-V89) | 0.81 (0.34-1.66) | UNINFORMATIVE (zero deaths in 1 cell) | Road Traffic (V01-V89) | 1.13 (0.56-2.10) | CONSISTENT | 1.13 (0.56-2.08) | CONSISTENT |
| Falls (W00-W19) | 1.26 (0.43-3.27) | UNINFORMATIVE (zero deaths in 13 cells) | Falls (W00-W19) | 0.36 (0.14-0.84) | BIAS SIGNAL (r < 1) | 0.38 (0.15-0.86) | BIAS SIGNAL (r < 1) |
| Drowning (W65-W74) | 0.75 (0.27-1.90) | CONSISTENT | Drowning (T75.1 / W65-W74) | 0.83 (0.25-2.60) | UNINFORMATIVE | 0.94 (0.27-3.07) | UNINFORMATIVE |
| Fire (X00-X09) | 0.69 (0.20-2.38) | UNINFORMATIVE (zero deaths in 29 cells) | Burns (T20-T32) | 1.72 (0.75-3.53) | CONSISTENT | 1.47 (0.65-2.99) | CONSISTENT |
| Accidental Poisoning (X40-X49) | 0.77 (0.23-2.58) | UNINFORMATIVE (zero deaths in 55 cells) | Accidental Poisoning (X40-X49) | 0.74 (0.23-2.32) | UNINFORMATIVE | 0.71 (0.22-2.05) | UNINFORMATIVE |
| Suffocation (W75-W84) | 0.66 (0.21-1.88) | UNINFORMATIVE (zero deaths in 7 cells) | Foreign Body (T15-T19) | 0.94 (0.30-2.74) | UNINFORMATIVE | 0.86 (0.28-2.52) | UNINFORMATIVE |
| All External Causes excl. Care (V01-Y98 excl. Y40-Y84) | 0.26 (0.10-0.58) | BIAS SIGNAL (r < 1) | All Injuries (S00-T98 excl. T78, T80-T88) | 0.49 (0.28-0.81) | BIAS SIGNAL (r < 1) | 0.51 (0.29-0.83) | BIAS SIGNAL (r < 1) |

Round-1 readings (3332, primary): DESIGN BIAS DETECTED (r < 1) (pooled context); this task: DESIGN BIAS DETECTED (r < 1).

## Report-only products on the same cells (primary window)

| product | autism (own data) | autism future+3 | unintentional deaths r | class | future+3 | DDD autism / unintentional | components BIAS SIGNAL | informative components | reading | round-1 reading (admissions) |
|---|---|---|---|---|---|---|---|---|---|---|
| Brazil Maternal DTaP (pooled context) | 0.77 (0.51-1.14) | 0.84 (0.64-1.14) (PASS) | 0.36 (0.15-0.79) | BIAS SIGNAL (r < 1) | 0.31 (0.15-0.57) (FAIL) | 1.92 (0.97-3.57) | none | 1 of 5 | DESIGN BIAS DETECTED (r < 1) | DESIGN BIAS DETECTED (r < 1) |
| Brazil rotavirus (report only) | 0.95 (0.62-1.57) | 0.76 (0.54-1.12) (PASS) | 0.37 (0.19-0.86) | BIAS SIGNAL (r < 1) | 0.26 (0.13-0.54) (FAIL) | 1.91 (0.86-4.75) | none | 0 of 5 | DESIGN BIAS DETECTED (r < 1) | NO BIAS DETECTED |
| Brazil PCV10 + MenC 2010 (report only) | 1.19 (0.76-2.05) | 0.90 (0.63-1.41) (PASS) | 0.26 (0.15-0.54) | BIAS SIGNAL (r < 1) | 0.28 (0.18-0.46) (FAIL) | 3.04 (1.29-8.00) | Drowning (W65-W74) (r < 1) | 1 of 5 | DESIGN BIAS DETECTED (r < 1) | DESIGN BIAS DETECTED (r < 1) |
| Brazil MMRV / varicella (report only) | 0.82 (0.64-1.83) | 0.87 (0.75-1.11) (PASS) | 1.22 (0.61-1.96) | CONSISTENT | 0.71 (0.37-1.16) (PASS) | 0.66 (0.37-2.26) | none | 1 of 5 | NO BIAS DETECTED | MIXED |
| Brazil hepatitis A (report only) | 2.17 (1.36-3.98) | 1.85 (1.23-3.09) (FAIL) | 0.75 (0.38-1.74) | CONSISTENT | 0.47 (0.26-0.89) (FAIL) | 2.24 (1.00-6.01) | Drowning (W65-W74) (r < 1) | 1 of 5 | MIXED | DESIGN BIAS DETECTED (r > 1) |

Each product's reading uses that product's own autism refit (same 108 cells) as the width reference. The products share the cells and the counts, and each block carries the other products as nuisance columns, so the five readings are not independent tests.

## Sensitivity (pooled context): death years 2020-2024

| outcome | primary (2021-2023) | pool5 (2020-2024) |
|---|---|---|
| All Unintentional Injuries (V01-X59) | 0.36 (0.15-0.79) BIAS SIGNAL (r < 1) | 0.36 (0.15-0.77) BIAS SIGNAL (r < 1) |
| Road Traffic (V01-V89) | 0.81 (0.34-1.66) UNINFORMATIVE (zero deaths in 1 cell) | 0.63 (0.29-1.18) CONSISTENT |
| Falls (W00-W19) | 1.26 (0.43-3.27) UNINFORMATIVE (zero deaths in 13 cells) | 0.97 (0.33-2.53) UNINFORMATIVE (zero deaths in 5 cells) |
| Drowning (W65-W74) | 0.75 (0.27-1.90) CONSISTENT | 0.88 (0.33-2.06) CONSISTENT |
| Fire (X00-X09) | 0.69 (0.20-2.38) UNINFORMATIVE (zero deaths in 29 cells) | 0.69 (0.20-2.42) UNINFORMATIVE (zero deaths in 15 cells) |
| Accidental Poisoning (X40-X49) | 0.77 (0.23-2.58) UNINFORMATIVE (zero deaths in 55 cells) | 0.82 (0.23-2.94) UNINFORMATIVE (zero deaths in 39 cells) |
| Suffocation (W75-W84) | 0.66 (0.21-1.88) UNINFORMATIVE (zero deaths in 7 cells) | 0.66 (0.22-1.75) UNINFORMATIVE (zero deaths in 4 cells) |
| All External Causes excl. Care (V01-Y98 excl. Y40-Y84) | 0.26 (0.10-0.58) BIAS SIGNAL (r < 1) | 0.25 (0.10-0.56) BIAS SIGNAL (r < 1) |

Readings: primary: DESIGN BIAS DETECTED (r < 1); pool5: DESIGN BIAS DETECTED (r < 1) (the rule reads the primary only).

## Future+3 placebo detail (pooled context, primary; reference prior, four kernel settings)

| outcome | reference | linear only | spline only | rho .85 | verdict |
|---|---|---|---|---|---|
| All Unintentional Injuries (V01-X59) | 0.31 (0.15-0.57) | 0.31 (0.15-0.57) | 0.31 (0.15-0.57) | 0.31 (0.15-0.57) | FAIL |
| Road Traffic (V01-V89) | 0.77 (0.27-1.67) | 0.77 (0.27-1.67) | 0.77 (0.27-1.67) | 0.77 (0.27-1.67) | PASS |
| Falls (W00-W19) | 2.54 (0.87-7.00) | 2.54 (0.87-7.00) | 2.54 (0.87-7.00) | 2.54 (0.87-7.00) | PASS |
| Drowning (W65-W74) | 0.68 (0.31-1.38) | 0.68 (0.31-1.38) | 0.68 (0.31-1.38) | 0.68 (0.31-1.38) | PASS |
| Fire (X00-X09) | 1.18 (0.37-3.40) | 1.18 (0.37-3.40) | 1.18 (0.37-3.40) | 1.18 (0.37-3.40) | PASS |
| Accidental Poisoning (X40-X49) | 0.92 (0.27-2.94) | 0.92 (0.27-2.94) | 0.92 (0.27-2.94) | 0.92 (0.27-2.94) | PASS |

## Candidates not fitted here

- `BRA_NC_SIM_HEADINJ_MULTCAUSE_CENSUS`: LOW priority in the request; 674,338,991 bytes (DOBR2020-DOBR2024.dbc); the discovery record states that the multiple-cause field names and completeness are unverified on any held file and that child head-injury deaths per state x band are likely sparse. Skipped under the task's 'only if cheap and clearly specified' instruction; not downloaded.
- `BRA_NC_SINAN_ACCPOISON_CENSUS / BRA_NC_SINAN_DOGBITE_CENSUS / BRA_NC_SINAN_VENOMOUS_CENSUS`: Not assigned to this task.

## Caveats

- Age at death, not census-cohort age: deaths in 2021-2023 at a given band come from birth cohorts up to about 1.5 years either side of the Census band cohorts (2.5 years in the 2020-2024 pool), and the 0-4 band includes deaths in 2023 of children born after the Census reference date. The exposure path is the Census cells' own.
- Flow versus stock: the controls are deaths over 2021-2023 per Census-2022 resident; autism is a Census stock. State and band effects absorb level differences; identification is the state x band interaction with the exposure path, as in the autism block.
- Denominators are the Census 2022 population of each band, not person-time at risk in the window.
- Death registration completeness and the share of ill-defined causes differ by state; chapter XX group coding (e.g. unintentional vs undetermined intent) also varies by state. Such state x band differences are what the controls are meant to expose; they are not removed.
- Underlying cause only: a death is counted once, under its chapter XX underlying cause; nature-of-injury codes (S/T, e.g. head injury) are not available from TabNet (BRA_NC_SIM_HEADINJ_MULTCAUSE_CENSUS was not acquired).
- Deaths are far fewer than admissions: falls, fire and accidental poisoning have zero-death cells and are UNINFORMATIVE by rule; the reading rests on the composite and the informative components.
- 2021 deaths fall in the COVID-19 period.
- The Census cells are a single cross-section: the six kernels coincide, so the four kernel settings return the same interval.
- Individual-scale r is the cohort-mixture parameter of the working likelihood; for an injury outcome it measures how the design converts state x band structure into an apparent effect, not a causal effect.