# 3383 - Injury-death negative controls on the U.S. core-v2 context (NCHS multiple-cause microdata)

**Primary (M1, unintentional injury deaths at ages 0-5 in 2000-2004 on state x birth-cohort Lexis cells): PCV r 2.59 (1.15-5.76), P(r>1) 0.989, UNINFORMATIVE; RotaShield r 0.82 (0.17-3.13), P(r>1) 0.366, CONSISTENT. Autism, core-v2 U.S.-alone: PCV 1.17 (0.97-1.39); RotaShield 0.87 (0.36-1.83).**

The core-v2 U.S. likelihood (processed/run_us_joint_likelihood.py; exposure paths, kernels, priors, proposal and seeds imported read-only and unchanged) is fitted to child deaths from unintentional injuries - an outcome vaccination cannot cause - on state x birth-cohort cells of the PCV7 ramp and RotaShield cohorts. Expected r = 1. Generated by `07_analysis_code/3383_injury_negative_controls_us_deaths.py`; every number is also in `3383_injury_negative_controls_us_deaths.csv` / `3383_injury_negative_controls_us_deaths.json`; figure `08_figures/3383_injury_negative_controls_us_deaths.png`; fit records `project_results/negative_controls_3330/3383/`.

## Interpretation rule (documented)

Rule file `qa/negative_controls_20260925/3383_rule.json` (SHA-256 `3fdaeaccc34b75fccba41343ac61f37bb1300baef983c5b24cf2ca2b1c835013`, equal to the hash recorded when it was written in `qa/negative_controls_20260925/3383_rule.sha256`), last modified 2026-09-29 22:21:27; the NCHS files were parsed afterwards (parse record written 2026-09-29 22:30:04) and the first fit started 2026-09-29 22:30:53 (file times, local clock): rule before first fit = True. 27 records (likelihood receipts, result records, parse audit) checked at report time: 27 carry the rule hash, 0 another, 0 none (the report refuses to run otherwise).

Classes, in order (the round-1 U.S. classes of 3331_rule.json, same order and thresholds; the only wording difference is the width reference: 3331 compares with autism 'on the same cells', 3383 with the core-v2 autism posterior, since no autism outcome sits on the death cells):

1. UNINFORMATIVE: ln(high / low) of the injury 95% interval > 2.5 x ln(high / low) of the matching autism estimate's 95% interval (width_reference).
2. CONSISTENT: otherwise, if the 95% interval includes 1 (low <= 1 <= high).
3. BIAS SIGNAL: otherwise (the 95% interval excludes 1); direction 'r > 1 (harm direction)' if low > 1, 'r < 1 (protective direction)' if high < 1.

Width reference: No autism outcome exists on the death cells, so the matching autism estimate is the core-v2 U.S.-alone autism posterior of the same product (the estimate this control tests): log widths PCV 0.3625, RotaShield 1.6231, RV2006 0.4547.

Primary estimates: M1 PCV and M1 RotaShield. Everything else is SECONDARY (classified by the same rule, never part of the reading). RV2006 is NOT IDENTIFIED on these cells: the core sets rotavirus-2006 exposure to 0 for cohorts <= 2004 and every death cell is of cohorts 1997-2004, so its posterior is the N(0, log 2) prior up to importance-sampling noise (eta4 shares the proposal scale component with the other effects).

## Data obtained (this task)

| file | bytes | SHA-256 | URL | retrieved (UTC) |
|---|---|---|---|---|
| `raw/us/nchs_mcd/Mort1999us.zip` | 88,077,536 | `ffcc497167a064506e0316d4d1a8e38170704241760e9a0e481b364dcfecd0d9` | https://ftp.cdc.gov/pub/Health_Statistics/NCHS/Datasets/DVS/mortality/Mort1999us.zip | 2026-09-30T03:14:50Z |
| `raw/us/nchs_mcd/Mort2000us.zip` | 82,255,243 | `7be7704d170e84ed45a5a5098124b1d4d14058009ce3a8d6d7b767b45480ad7d` | https://ftp.cdc.gov/pub/Health_Statistics/NCHS/Datasets/DVS/mortality/Mort2000us.zip | 2026-09-30T03:14:50Z |
| `raw/us/nchs_mcd/Mort2001us.zip` | 83,108,556 | `de827b17e7699c5193c18029c00df4bfeb8e19a937da018219f3b304acba5946` | https://ftp.cdc.gov/pub/Health_Statistics/NCHS/Datasets/DVS/mortality/Mort2001us.zip | 2026-09-30T03:14:51Z |
| `raw/us/nchs_mcd/Mort2002us.zip` | 88,593,104 | `3ea85ea4ec2bc17f98892acf4357a25ffa8a0f5019c8a1380deaf227ec82e530` | https://ftp.cdc.gov/pub/Health_Statistics/NCHS/Datasets/DVS/mortality/Mort2002us.zip | 2026-09-30T03:14:51Z |
| `raw/us/nchs_mcd/mort2003us.zip` | 89,348,782 | `f7a86e71223e4ad71ca0fffff79b90423b10ce571545b45605931950f410d7fe` | https://ftp.cdc.gov/pub/Health_Statistics/NCHS/Datasets/DVS/mortality/mort2003us.zip | 2026-09-30T03:14:51Z |
| `raw/us/nchs_mcd/Mort2004us.zip` | 88,237,773 | `a242631d24ffedca9a10018cfd17b524892a051b5ee9066517bb984fcdb9ce24` | https://ftp.cdc.gov/pub/Health_Statistics/NCHS/Datasets/DVS/mortality/Mort2004us.zip | 2026-09-30T03:14:51Z |
| `raw/us/nchs_mcd/docs/Mort99doc.pdf` | 1,662,832 | `8ccc55d5d5a107b415b23a7584bfa8d5bd1c845f7c665e7df0e67ced64bd676e` | https://ftp.cdc.gov/pub/Health_Statistics/NCHS/Dataset_Documentation/DVS/mortality/Mort99doc.pdf | 2026-09-30T03:11:22Z |
| `raw/us/nchs_mcd/docs/Record_Layout_2004.pdf` | 15,650,765 | `4e1ce80a0fa3e9df3e2d392534a86eed01c3d59804c5434d9cf19330bc0f1293` | https://ftp.cdc.gov/pub/Health_Statistics/NCHS/Dataset_Documentation/DVS/mortality/Record_Layout_2004.pdf | 2026-09-30T03:11:22Z |
| `raw/us/nchs_mcd/docs/Record_Layout_2005.pdf` | 17,378,591 | `c04429524875ce5044c16ee19f7a48b48b453debd2bf1499a2781c51d79a2b9f` | https://ftp.cdc.gov/pub/Health_Statistics/NCHS/Dataset_Documentation/DVS/mortality/Record_Layout_2005.pdf | 2026-09-30T03:09:02Z |

Each file has a `.meta.json` sidecar (URL, retrieval time, SHA-256, size, terms). `raw/us/nchs_mcd` lies inside the project folder, which is itself an NTFS junction to `E:\MigratedFromC\...`, so the bytes are on E: (the new-e-backed-folder script refused a junction onto its own physical location).

Not obtained:

- `Mort2005us.zip` (82,992,987 bytes): not downloaded. Its own record layout (`Record_Layout_2005.pdf`) states that the policy excluding geographic identifiers takes effect with the 2005 data year and marks positions 21-60 (where 2004 carries state and county) as blank, so the public 2005 file has no state.
- CDC WONDER Underlying Cause of Death (state x single age x year, 1999-2020): not accessed. Its query pages require clicking "I Agree" to the data-use restrictions; the discovery record and the request list that as a user step, and a relayed authorization cannot stand in for the user's own click in this session. curl is blocked (HTTP 403) and the WONDER API cannot group by state. It is the only public state-level route to the RV2006 cohorts (2006+); without it RV2006 has no death-cell information.

Parse checks (every record of every file read; record length and data-year field constant within each file):

| data year | layout | records | foreign residents | U.S.-territory residents | age not stated | decedents aged 0-8 kept |
|---|---|---|---|---|---|---|
| 1999 | Mort99doc.pdf | 2,394,871 | 3,472 | 0 | 360 | 36,113 |
| 2000 | Mort99doc.pdf | 2,407,193 | 3,842 | 0 | 363 | 35,785 |
| 2001 | Mort99doc.pdf | 2,419,960 | 3,535 | 0 | 431 | 35,280 |
| 2002 | Mort99doc.pdf | 2,446,796 | 3,409 | 0 | 357 | 35,433 |
| 2003 | Record_Layout_2004.pdf | 2,452,154 | 3,642 | 224 | 359 | 35,454 |
| 2004 | Record_Layout_2004.pdf | 2,401,400 | 3,542 | 243 | 366 | 35,135 |

1999 against Mort99doc.pdf (U.S. data set): record count 2,394,871, by residence 2,391,399, foreign residents 3,472 - equal to the documented values: True. Death year 1999 is tabulated but not fitted (no held 1999 state x single-age denominator).

## How the death cells differ from the autism cells

| | core-v2 autism cells | 3383 death cells (M1) |
|---|---|---|
| measure | STOCK: children with an autism IEP on the count date (IDEA child count; CASEMIS December count) | FLOW: deaths during calendar year t (underlying cause V01-X59), state of residence |
| ages | 8 (IDEA, cohorts 1997-2003); 3, 4, 5 single ages 2005-2011 and the 3-5 band 2012-2015; California 7-9 2011-2018 | 0-5 (completed years), death years 2000-2004 |
| cohort assignment | birth cohort = count year - age (Dec-Nov window), one cohort per single-age cell; band cells population-weighted | Lexis square: half the person-time of (a, t) is cohort t-a-1 and half t-a (weights 0.5 / 0.5); every cohort in 1997-2004 |
| cells | 1,469 (A8 353, EC3 303, EC4 312, EC5 324, EC35 156, CA7-9 21) | 1,107 of 1,195 candidates (zeros dropped: 88); D0 244, D1 231, D2 228, D3 183, D4 131, D5 90 |
| denominator | Census civilian July-1 population at the cell age and count year | the same table (population_table(), v2009) at the death age and year |
| suppression | EDFacts small-cell suppression (x) | none (microdata) |
| exposure | A8: the core A8 rows (PCV, RotaShield, interaction); others: NewCellExposure (PCV, RV2006) | the same rows; squares containing a 1997-1999 cohort average the A8 rows of their two cohorts; squares of cohorts >= 2000 use NewCellExposure; RV2006 = 0 on every cell |
| exposure-to-outcome timing | exposure in infancy, autism counted years later | exposure in infancy, deaths at 0-5 overlap the exposure window for the youngest ages (deaths at age 0 can precede a dose) |

## Reproduction

The 3383 likelihood function (rule exposure routing) applied to the 1,469 core-v2 autism cells returns project_results/portfolio_v2/core_v2/us_alone_diagnostics_p18.json ['primary']['us']: every posterior field identical = True, both seeds' ESS identical = True (float equality) -> OK. A unit test checks that the component-mixture route equals NewCellExposure (and the A8 rows) on cells where both apply, to 1e-12.

## Results: primary estimates (the rule's verdicts)

| block | design | product | cells | deaths in cells | injury r (95% CrI) | P(r>1) | ESS (power) | autism, core-v2 | width / autism | class | flags | future+3 PCV placebo | 1-D fit, actual PCV path |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| M1 | Unintentional injury (V01-X59), ages 0-5 | PCV | 1,107 | 12,715 | 2.59 (1.15-5.76) | 0.989 | 474 (p18) | 1.17 (0.97-1.39) | 4.44 | UNINFORMATIVE | CrI excludes 1 (r > 1) | FAIL 0.18 (0.10-0.39) | 2.05 (1.12-3.67) |
| M1 | Unintentional injury (V01-X59), ages 0-5 | RotaShield | 1,107 | 12,715 | 0.82 (0.17-3.13) | 0.366 | 474 (p18) | 0.87 (0.36-1.83) | 1.81 | CONSISTENT | prior-dominated | n/a (PCV-path placebo only) | - |

width / autism = ln(high/low) of the injury CrI divided by that of the core-v2 autism CrI of the same product (UNINFORMATIVE above 2.5). future+3 = the 3326 PCV timing placebo on the same cells (reference kernel setting shown; PASS = 95% CrI includes 1 under all four settings). 1-D fit, actual PCV path = the same placebo module's grid fit of the actual point path on the same cells (one product, RV2006 and RotaShield point paths as nuisance columns, reference setting; not importance-sampled), which 3326 reports beside the placebo.

- M1 PCV: 2.59 (1.15-5.76), P(r>1) 0.989 -> **UNINFORMATIVE**. No conclusion.
- M1 RotaShield: 0.82 (0.17-3.13), P(r>1) 0.366 -> **CONSISTENT**. No non-specific bias detected at this precision; this does not validate the autism estimate.

## Reading

- PCV: the documented class of M1 is UNINFORMATIVE (CrI log width 4.4x autism's). Its CrI nevertheless excludes 1 (2.59 (1.15-5.76), P(r>1) 0.989), which the rule reports as a flag and does not let change the class. The future+3 PCV placebo on the same cells is FAIL (0.18 (0.10-0.39) at the reference setting): the core PCV path shifted three cohorts later also "finds an effect" on these deaths, in the opposite direction: on these cells the machinery returns associations that are not specific to the vaccine's timing (the placebo never changes the class). Secondary PCV fits: M2 2.47 (1.10-5.76) (future+3 FAIL); M3 2.69 (1.23-5.76) (future+3 FAIL); M4 37.45 (37.44-37.45) (future+3 PASS); M5 1.49 (0.66-3.60) (future+3 FAIL); M6 1.08 (0.50-2.78) (future+3 PASS). M4 is degenerate: pooled ESS 1.0 at p20 (one importance draw carries the posterior in both seeds), so its interval is not a posterior summary and its rule class (BIAS SIGNAL) carries no information; the 1-D fit of the actual path on the same cells is 0.59 (0.24-1.51).
- RotaShield: M1 0.82 (0.17-3.13), P(r>1) 0.366 -> CONSISTENT (prior-dominated: the interval is nearly as wide as the N(0, log 2) prior); autism core-v2 0.87 (0.36-1.83).
- RV2006: not testable on these cells (1.17 (0.33-4.49) is the prior up to sampling noise; class by the rule UNINFORMATIVE).
- Cohort-aligned DDD (autism relative to injury deaths on the core's own A8 units): PCV 1.03 (0.69-1.52), RotaShield 1.31 (0.48-2.84) (descriptive).

## Secondary estimates (classified by the same rule; not part of the reading)

| block | design | product | cells | deaths in cells | injury r (95% CrI) | P(r>1) | ESS (power) | autism, core-v2 | width / autism | class | flags | future+3 PCV placebo | 1-D fit, actual PCV path |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| M1 | Unintentional injury (V01-X59), ages 0-5 | Rotavirus 2006 | 1,107 | 12,715 | 1.17 (0.33-4.49) | 0.606 | 474 (p18) | 0.95 (0.76-1.19) | 5.72 | UNINFORMATIVE | SECONDARY; NOT IDENTIFIED (exposure identically 0 on these cells; the posterior is the N(0, log 2) prior up to importance-sampling noise); prior-dominated | n/a (PCV-path placebo only) | - |
| M2 | Unintentional injury, ages 0-5, complete series | PCV | 988 | 12,469 | 2.47 (1.10-5.76) | 0.986 | 776 (p18) | 1.17 (0.97-1.39) | 4.58 | UNINFORMATIVE | SECONDARY; CrI excludes 1 (r > 1) | FAIL 0.18 (0.10-0.39) | 2.03 (1.09-3.70) |
| M2 | Unintentional injury, ages 0-5, complete series | RotaShield | 988 | 12,469 | 0.84 (0.17-3.25) | 0.390 | 776 (p18) | 0.87 (0.36-1.83) | 1.84 | CONSISTENT | SECONDARY; prior-dominated | n/a (PCV-path placebo only) | - |
| M2 | Unintentional injury, ages 0-5, complete series | Rotavirus 2006 | 988 | 12,469 | 1.17 (0.33-4.47) | 0.601 | 776 (p18) | 0.95 (0.76-1.19) | 5.76 | UNINFORMATIVE | SECONDARY; NOT IDENTIFIED (exposure identically 0 on these cells; the posterior is the N(0, log 2) prior up to importance-sampling noise); prior-dominated | n/a (PCV-path placebo only) | - |
| M3 | Unintentional injury, ages 1-5 | PCV | 863 | 7,918 | 2.69 (1.23-5.76) | 0.992 | 782 (p18) | 1.17 (0.97-1.39) | 4.26 | UNINFORMATIVE | SECONDARY; CrI excludes 1 (r > 1) | FAIL 0.16 (0.08-0.38) | 2.16 (1.14-3.98) |
| M3 | Unintentional injury, ages 1-5 | RotaShield | 863 | 7,918 | 0.77 (0.17-2.83) | 0.340 | 782 (p18) | 0.87 (0.36-1.83) | 1.75 | CONSISTENT | SECONDARY; prior-dominated | n/a (PCV-path placebo only) | - |
| M3 | Unintentional injury, ages 1-5 | Rotavirus 2006 | 863 | 7,918 | 1.19 (0.33-4.47) | 0.617 | 782 (p18) | 0.95 (0.76-1.19) | 5.72 | UNINFORMATIVE | SECONDARY; NOT IDENTIFIED (exposure identically 0 on these cells; the posterior is the N(0, log 2) prior up to importance-sampling noise); prior-dominated | n/a (PCV-path placebo only) | - |
| M4 | Head injury (record axis S00-S09), ages 0-5 | PCV | 982 | 4,959 | 37.45 (37.44-37.45) | 1.000 | 1 (p20) | 1.17 (0.97-1.39) | 0.00 | BIAS SIGNAL (r > 1 (harm direction)) | SECONDARY; LOW ESS (pooled 1.0 at p20; the posterior is essentially one importance draw, so its interval is not a posterior summary) | PASS 1.07 (0.21-4.21) | 0.59 (0.24-1.51) |
| M4 | Head injury (record axis S00-S09), ages 0-5 | RotaShield | 982 | 4,959 | 0.98 (0.98-0.98) | 0.000 | 1 (p20) | 0.87 (0.36-1.83) | 0.00 | BIAS SIGNAL (r < 1 (protective direction)) | SECONDARY; LOW ESS (pooled 1.0 at p20; the posterior is essentially one importance draw, so its interval is not a posterior summary) | n/a (PCV-path placebo only) | - |
| M4 | Head injury (record axis S00-S09), ages 0-5 | Rotavirus 2006 | 982 | 4,959 | 5.30 (5.30-5.30) | 1.000 | 1 (p20) | 0.95 (0.76-1.19) | 0.00 | BIAS SIGNAL (r > 1 (harm direction)) | SECONDARY; NOT IDENTIFIED (exposure identically 0 on these cells; the posterior is the N(0, log 2) prior up to importance-sampling noise); LOW ESS (pooled 1.0 at p20; the posterior is essentially one importance draw, so its interval is not a posterior summary) | n/a (PCV-path placebo only) | - |
| M5 | All external causes (V01-Y89), ages 0-5 | PCV | 1,127 | 17,093 | 1.49 (0.66-3.60) | 0.828 | 2159 (p18) | 1.17 (0.97-1.39) | 4.68 | UNINFORMATIVE | SECONDARY | FAIL 0.36 (0.17-1.18) | 1.34 (0.67-2.66) |
| M5 | All external causes (V01-Y89), ages 0-5 | RotaShield | 1,127 | 17,093 | 0.55 (0.17-1.69) | 0.151 | 2159 (p18) | 0.87 (0.36-1.83) | 1.43 | CONSISTENT | SECONDARY; prior-dominated | n/a (PCV-path placebo only) | - |
| M5 | All external causes (V01-Y89), ages 0-5 | Rotavirus 2006 | 1,127 | 17,093 | 1.18 (0.31-4.53) | 0.601 | 2159 (p18) | 0.95 (0.76-1.19) | 5.87 | UNINFORMATIVE | SECONDARY; NOT IDENTIFIED (exposure identically 0 on these cells; the posterior is the N(0, log 2) prior up to importance-sampling noise); prior-dominated | n/a (PCV-path placebo only) | - |
| M6 | Unintentional injury, age bands 0 / 1-2 / 3-4 | PCV | 627 | 11,840 | 1.08 (0.50-2.78) | 0.572 | 169647 (p18) | 1.17 (0.97-1.39) | 4.71 | UNINFORMATIVE | SECONDARY | PASS 0.69 (0.22-2.87) | 1.07 (0.52-2.59) |
| M6 | Unintentional injury, age bands 0 / 1-2 / 3-4 | RotaShield | 627 | 11,840 | 0.85 (0.24-2.79) | 0.400 | 169647 (p18) | 0.87 (0.36-1.83) | 1.52 | CONSISTENT | SECONDARY; prior-dominated | n/a (PCV-path placebo only) | - |
| M6 | Unintentional injury, age bands 0 / 1-2 / 3-4 | Rotavirus 2006 | 627 | 11,840 | 1.08 (0.29-4.35) | 0.546 | 169647 (p18) | 0.95 (0.76-1.19) | 5.92 | UNINFORMATIVE | SECONDARY; NOT IDENTIFIED (exposure identically 0 on these cells; the posterior is the N(0, log 2) prior up to importance-sampling noise); prior-dominated | n/a (PCV-path placebo only) | - |

## Cohort-aligned DDD (M7, descriptive)

A shared-cell DDD is not meaningful (no death cell shares a state x age x year x measure with an autism cell). M7 refits the core's own 353 A8 state x cohort units with y = log(autism age-8 rate) - log(unintentional injury death rate of the same state x cohort at ages 0-8 in 2000-2004, Lexis half-weights) through the 3326 log-ratio device; 351 units (dropped for zero deaths: [['District of Columbia', 2001], ['Vermont', 2003]]). r below 1 = autism leans less than injury deaths on the same units. Descriptive; never changes a class.

| product | DDD r (95% CrI) | P(r>1) | ESS (power) | future+3 PCV placebo |
|---|---|---|---|---|
| PCV | 1.03 (0.69-1.52) | 0.551 | 25382 (p18) | PASS 1.20 (0.63-2.89) |
| RotaShield | 1.31 (0.48-2.84) | 0.728 | 25382 (p18) | n/a (PCV-path placebo only) |
| Rotavirus 2006 | 1.02 (0.28-4.22) | 0.512 | 25382 (p18) | n/a (PCV-path placebo only) (NOT IDENTIFIED) |

## Cells and cell loss

- M1 (Unintentional injury (V01-X59), ages 0-5): 1,107 cells used of 1,195 candidates; zero-count cells dropped 88; deaths in the used cells 12,715; 51 states. By class (used / candidates / zero): D0 244/253/9, D1 231/250/19, D2 228/249/21, D3 183/198/15, D4 131/147/16, D5 90/98/8. Squares dropped before the count rule because a 1997-1999 cohort has no core RotaShield path (DC 1999, MT 1998, VT 1999): 29.
- M2 (Unintentional injury, ages 0-5, complete series): 988 cells used of 1,195 candidates; zero-count cells dropped 88; series dropped (complete-series rule) 56; deaths in the used cells 12,469; 50 states. By class (used / candidates / zero): D0 240/253/9, D1 185/250/19, D2 193/249/21, D3 168/198/15, D4 115/147/16, D5 87/98/8. Squares dropped before the count rule because a 1997-1999 cohort has no core RotaShield path (DC 1999, MT 1998, VT 1999): 29.
- M3 (Unintentional injury, ages 1-5): 863 cells used of 942 candidates; zero-count cells dropped 79; deaths in the used cells 7,918; 51 states. By class (used / candidates / zero): D1 231/250/19, D2 228/249/21, D3 183/198/15, D4 131/147/16, D5 90/98/8. Squares dropped before the count rule because a 1997-1999 cohort has no core RotaShield path (DC 1999, MT 1998, VT 1999): 29.
- M4 (Head injury (record axis S00-S09), ages 0-5): 982 cells used of 1,195 candidates; zero-count cells dropped 213; deaths in the used cells 4,959; 51 states. By class (used / candidates / zero): D0 234/253/19, D1 212/250/38, D2 212/249/37, D3 152/198/46, D4 109/147/38, D5 63/98/35. Squares dropped before the count rule because a 1997-1999 cohort has no core RotaShield path (DC 1999, MT 1998, VT 1999): 29.
- M5 (All external causes (V01-Y89), ages 0-5): 1,127 cells used of 1,195 candidates; zero-count cells dropped 68; deaths in the used cells 17,093; 51 states. By class (used / candidates / zero): D0 247/253/6, D1 237/250/13, D2 233/249/16, D3 185/198/13, D4 135/147/12, D5 90/98/8. Squares dropped before the count rule because a 1997-1999 cohort has no core RotaShield path (DC 1999, MT 1998, VT 1999): 29.
- M6 (Unintentional injury, age bands 0 / 1-2 / 3-4): 627 cells used of 645 candidates; zero-count cells dropped 18; deaths in the used cells 11,840; 51 states. By class (used / candidates / zero): DB0 244/253/9, DB12 242/247/5, DB34 141/145/4. Squares dropped before the count rule because a 1997-1999 cohort has no core RotaShield path (DC 1999, MT 1998, VT 1999): 18.

## National counts (descriptive; residents of the 50 states + DC, underlying cause V01-X59)

| death year | age 0 | age 1 | age 2 | age 3 | age 4 | age 5 | ages 6-8 |
|---|---|---|---|---|---|---|---|
| 1999 (not fitted) | 845 | 640 | 492 | 410 | 354 | 293 | 886 |
| 2000 | 881 | 563 | 527 | 379 | 349 | 301 | 815 |
| 2001 | 976 | 573 | 441 | 359 | 329 | 275 | 752 |
| 2002 | 946 | 521 | 435 | 361 | 308 | 254 | 691 |
| 2003 | 945 | 526 | 488 | 389 | 306 | 263 | 618 |
| 2004 | 1,052 | 535 | 456 | 356 | 280 | 253 | 645 |

## Caveats

- Flow versus stock: an injury death in year t at age a is a single event, while an autism cell counts every child identified by the count date. A state x cohort bias that works through identification or reporting (special-education practice, diagnosis) would appear in autism but not in deaths; a bias that works through population composition, poverty or health-care access would appear in both. A CONSISTENT class here tests only the second kind.
- The death cells cover the PCV7 ramp and RotaShield cohorts at ages 0-5 in 2000-2004; they share no age, year or count date with the autism cells. The kappa plug-in and residual scale are re-estimated on the death cells (class-alone residual variance, rho 0, state-trend spec: death classes D0 0.074, D1 0.088, D2 0.120, D3 0.107, D4 0.130, D5 0.011; autism A8 0.0041, EC3-EC5 0.018, 0.012, 0.007), so the intervals are wider than autism's.
- Zero-count cells cannot enter a log-rate likelihood and are dropped (observed rule; M1 88 of 1,195 candidates, in 18 states). The complete-series (M2) and age-band (M6) sensitivities bound that selection.
- Accidental suffocation and strangulation in bed or elsewhere (W75-W84) is 65% of the infant unintentional-injury deaths in the fitted years; its certification can move between unintentional, undetermined and SIDS codes. M3 (ages 1-5) and M5 (all external causes) address that.
- Importance-sampling tails are coarse: the PCV upper bound is M1 5.76 / M2 5.76 / M3 5.76 and the RotaShield lower bound M1 0.17 / M2 0.17 / M3 0.17 / M5 0.17; equal values mean the same few heavy draws set those quantiles, so tail bounds are coarse even at ESS in the hundreds; medians and P(r>1) are less affected.
- The 1997-cohort RotaShield exposure uses the core 1997 coin; squares of DC 1999, MT 1998 and VT 1999 are dropped because the core has no A8 record (hence no RotaShield path) there.
- RV2006 cannot be tested with public microdata: the 2005+ public files carry no state, and the only public state-level route (CDC WONDER, 1-9 suppressed) was not accessed.