# 3326 - Is the U.S. core-v2 PCV lean specific to autism? Negative-control IDEA outcomes

**Documented verdict: MIXED – only hearing impairment is precise enough to count, and it leans like autism.** Under MIXED the documented rule draws neither conclusion: the data do not establish that the autism PCV r is shared drift, nor that it is specific to autism. Everything else below (the UNINFORMATIVE controls, the DDDs, the complete-series sensitivity, the placebo) is reported beside the verdict and, by the rule, does not change it.

Generated by `07_analysis_code/3326_us_core_negative_control_outcomes.py` (every number below is computed by that script and also written to `3326_us_core_negative_control_outcomes.csv` / `3326_us_core_negative_control_outcomes.json`). Figure: `08_figures/3326_us_core_negative_control_outcomes.png` (hollow light markers = no weight under the rule: UNINFORMATIVE or unstable across cell rules; italic flags at right). Status: falsification audit of the core-v2 U.S. PCV context; ecological; no causal pool weight of its own.

## Interpretation rule (documented)

Rule file `qa/stage2b_20260925/3326_interpretation_rule_preregistered.json` last modified 2026-09-25 12:50:29 (file mtime, local clock), before the first fit, which started at 2026-09-25 13:02:06 (`project_results/us_core_negative_controls_3326/observed/autism/likelihood_p18_s23091`, receipt time minus its run time; first likelihood file written 2026-09-25 13:02:50). Its 'written' field: 2026-09-25, before any non-autism outcome was fitted on the core-v2 cells (only the autism counts had been rebuilt from the raw files and matched to the 1,469 core-v2 cells; no likelihood, posterior or placebo of any other category had been computed). SHA-256 of the rule text `3da1e9ea747c24c0...` recorded at report time. None of the 96 fit records (likelihood receipts and result.json) and none of the 3 fit logs carry it: they predate hash recording, so for these fits the hash adds nothing to the timing and the precedence evidence is the file times alone. The script now writes the hash into every likelihood receipt and result.json and prints it into the fit log when fitting, and the report refuses to run if any recorded hash differs from the rule file.

> Quantity: the U.S.-alone posterior of the PCV individual-scale r (eta0 of the core-v2 joint U.S. likelihood; primary configuration, 'population' kernels, N(0, log 2) priors, seeds 23091 and 23092 pooled, p18) fitted on each outcome's observed cells. Each negative control (traumatic brain injury, orthopedic impairment, hearing impairment, visual impairment) is classified as: UNINFORMATIVE if its 95% CrI is more than twice as wide as autism's on the log scale, or if its pooled ESS is below 100 at p18 and also at p20; otherwise SIMILAR if median r > 1, P(r > 1) >= 0.80 and its 95% CrI overlaps autism's 95% CrI; NULL-CENTRED if its 95% CrI includes 1 and its median lies in [0.95, 1.05]; BELOW if its median is < 0.95 (drift of the opposite sign); OTHER otherwise. Verdict: NON-SPECIFIC (the autism 1.17 is shared state-cohort drift) if at least two informative negative controls are SIMILAR and SIMILAR outnumbers NULL-CENTRED; AUTISM-SPECIFIC (reported as an unresolved harm-direction signal) if at least two informative negative controls are NULL-CENTRED, none is SIMILAR, and autism's P(r > 1) >= 0.80; MIXED otherwise (reported as such; neither conclusion is drawn). The DDD log-ratio fits (autism relative to each negative control), the complete-series sensitivity and the future+3 PCV placebo are reported beside the verdict and do not change it. The claim-relevant categories (intellectual disability, speech or language impairment, specific learning disability, emotional disturbance, other health impairment, all disabilities) are not negative controls and never enter the verdict.

In words: if the negative controls give a PCV r of similar size to autism (overlapping and centred above 1), the autism 1.17 is non-specific state-cohort drift; if they centre on 1 while autism stays above, the autism lean is specific and must be reported as an unresolved harm-direction signal.

## Design

- Likelihood: `processed/run_us_joint_likelihood.py` imported and used unchanged (cells, cohort weights, Census denominators, time keys, splines, HepA / influenza / Hib nuisance columns, kernels with the series-variance plug-in re-estimated on each outcome, R_any PCV path, RotaShield and RV2006 paths, priors, proposal, seeds 23091 / 23092, p18). Posterior = `us_alone` (U.S. cells only, `population` kernels, N(0, log 2) priors), the diagnostic that gives the core its 1.17. Only the six `population` U.S. kernels (and the nine Japan models, so `us_alone` runs unchanged) are evaluated; the other kernel variants never enter that diagnostic.
- Outcome swap: the autism count of each core-v2 cell is replaced by the category count from the same files (EDFacts `bchildcount2005-2011` ages 3/4/5/8, `env2012-2015` Total Age 3-5; CDE DataQuest CASEMIS state reports ages 7-9, the HTML source of the CSV the core reads). Denominators are unchanged.
- DDD: log-ratio model, y = log(autism / control) on cells where both are observed, same likelihood; the fitted r multiplies autism risk relative to the control (shared state x cohort drift cancels). Because autism is observed in every core-v2 cell, the DDD cells are the control's cells and the DDD restates the control-versus-autism comparison; it is not an independent line of evidence.
- Cell rules: `observed` (primary: numeric and > 0; every other cell dropped and counted by kind) and `complete_series` (sensitivity: whole state x age-class series dropped if any core cell is not numeric and > 0, plus the core reporting-break rule on the category's own EDFacts single-age counts).
- Cell status coding, from the raw files' own footnotes: EDFacts `x` / `X` = "Data suppressed due to small cell size" (suppressed); EDFacts `-` = "Data not available" (not published; defined in every bchildcount2005-2011 and env2012-2015 file); EDFacts `*` = "Data flagged due to questionable data quality" (quality-flagged; env2013 / env2015); CASEMIS `*` = "Denotes values under 11" (suppressed). A category row absent from a file is also not published. Every kind is excluded under either rule; none is imputed.
- Future+3 PCV placebo: point R_any path of cohort c + 3 on the same cells, RV2006 and RotaShield as nuisance columns, six `population` kernels, 1-D grid, four kernel settings (run_us_joint_gates conventions). PASS = the 95% CrI includes 1 under all four settings.
- Low ESS: an outcome with pooled p18 ESS < 100 is re-fitted at p20 and read at p20.

## Autism reproduction

With the autism outcome the script returns the core-v2 U.S.-alone posterior EXACTLY (project_results/portfolio_v2/core_v2/us_alone_diagnostics_p18.json ['primary']['us']): PCV r 1.17 (0.97-1.39), P(r > 1) = 0.948; every posterior field and both seeds' ESS are identical (float equality). The autism counts rebuilt from the raw files equal the core-v2 counts in all 1,469 cells. Bit-for-bit reproduction needs the core's 4 BLAS threads (OpenBLAS rounding changes with the thread count, ~1e-9 in the log-likelihood); the script refuses to fit under any other count, and `07_analysis_code/test_3326_us_core_negative_control_outcomes.py` checks the written p18 result, a fresh p16 fit and (T3326_FRESH_P18=1) a fresh p18 fit against the stored core-v2 files.

## Cells used (primary rule `observed`)

| outcome | role | cells used | A8 | EC3 | EC4 | EC5 | EC35 | CA7-9 | suppressed (x / X; CASEMIS *) | not published (-) | quality-flagged (*) | zero | states |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Autism | reference | 1,469 | 353 | 303 | 312 | 324 | 156 | 21 | 0 | 0 | 0 | 0 | 51 |
| Traumatic brain injury | negative control | 472 | 110 | 43 | 62 | 84 | 152 | 21 | 867 | 0 | 0 | 130 | 50 |
| Orthopedic impairment | negative control | 1,261 | 303 | 248 | 254 | 279 | 156 | 21 | 191 | 0 | 0 | 17 | 51 |
| Hearing impairment | negative control | 1,381 | 341 | 273 | 283 | 307 | 156 | 21 | 83 | 0 | 0 | 5 | 51 |
| Visual impairment | negative control | 1,073 | 250 | 196 | 217 | 233 | 156 | 21 | 361 | 0 | 0 | 35 | 51 |
| Intellectual disability | claim relevant | 1,117 | 349 | 160 | 163 | 268 | 156 | 21 | 282 | 6 | 0 | 64 | 51 |
| Speech or language impairment | claim relevant | 1,459 | 353 | 299 | 310 | 320 | 156 | 21 | 10 | 0 | 0 | 0 | 51 |
| Specific learning disability | claim relevant | 845 | 353 | 54 | 67 | 199 | 151 | 21 | 397 | 7 | 0 | 220 | 51 |
| Emotional disturbance | claim relevant | 809 | 350 | 54 | 63 | 171 | 150 | 21 | 420 | 8 | 0 | 232 | 51 |
| Other health impairment | claim relevant | 1,364 | 346 | 264 | 273 | 305 | 155 | 21 | 76 | 16 | 0 | 13 | 51 |
| All disabilities | claim relevant | 1,448 | 353 | 303 | 312 | 324 | 156 | 0 | 21 | 0 | 0 | 0 | 51 |
| Autism / traumatic brain injury | ddd | 472 | 110 | 43 | 62 | 84 | 152 | 21 | - | - | - | - | 50 |
| Autism / orthopedic impairment | ddd | 1,261 | 303 | 248 | 254 | 279 | 156 | 21 | - | - | - | - | 51 |
| Autism / hearing impairment | ddd | 1,381 | 341 | 273 | 283 | 307 | 156 | 21 | - | - | - | - | 51 |
| Autism / visual impairment | ddd | 1,073 | 250 | 196 | 217 | 233 | 156 | 21 | - | - | - | - | 51 |

Core-v2 cells: 1,469 (A8 353, EC3 303, EC4 312, EC5 324, EC35 156, CA7-9 21). For a DDD the excluded-cell columns are those of the control; the DDD uses cells kept for both. CASEMIS HTML vs derived CSV: 522 overlapping values checked, 0 mismatches.
All disabilities: 0 of the 21 CA79 cells are usable (CASEMIS publishes no all-disability total by age and suppresses deaf-blindness below 11, so a CA79 total exists only where all 13 categories are numeric); the remaining CA79 cells are counted as suppressed.

## Results (primary rule `observed`)

| outcome | role | power | cells | PCV r (95% CrI) | P(r>1) | RotaShield r | P(r>1) | RV2006 r | P(r>1) | ESS pooled (s23091 / s23092) | future+3 PCV r (reference) | future+3 verdict |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Autism | reference | p18 | 1,469 | 1.17 (0.97-1.39) | 0.948 | 0.87 (0.36-1.83) | 0.349 | 0.95 (0.76-1.19) | 0.325 | 1279 (1038 / 322) | 0.86 (0.66-1.17) | PASS |
| Traumatic brain injury | negative control | p18 | 472 | 1.65 (0.77-4.29) | 0.892 | 0.69 (0.20-2.21) | 0.286 | 2.37 (1.09-4.58) | 0.985 | 640 (116 / 542) | 1.14 (0.45-3.78) | PASS |
| Orthopedic impairment | negative control | p18 | 1,261 | 1.43 (0.91-2.26) | 0.937 | 0.97 (0.31-2.35) | 0.472 | 1.01 (0.70-1.46) | 0.529 | 8937 (7694 / 1284) | 0.59 (0.39-1.05) | PASS |
| Hearing impairment | negative control | p18 | 1,381 | 1.41 (1.05-1.91) | 0.990 | 1.41 (0.48-3.16) | 0.771 | 1.41 (1.07-1.82) | 0.991 | 199 (150 / 64) | 0.97 (0.63-1.74) | PASS |
| Visual impairment | negative control | p18 | 1,073 | 1.45 (0.74-2.72) | 0.855 | 1.14 (0.35-2.90) | 0.599 | 1.16 (0.78-1.71) | 0.778 | 5153 (5310 / 585) | 0.95 (0.54-2.07) | PASS |
| Intellectual disability | claim relevant | p18 | 1,117 | 1.18 (0.94-1.49) | 0.925 | 0.72 (0.25-1.74) | 0.230 | 1.42 (0.96-2.05) | 0.960 | 3109 (1983 / 1206) | 0.55 (0.40-0.81) | FAIL |
| Speech or language impairment | claim relevant | p18 | 1,459 | 1.09 (0.95-1.24) | 0.898 | 1.05 (0.50-2.01) | 0.565 | 1.11 (0.94-1.31) | 0.885 | 271 (228 / 207) | 0.80 (0.65-1.00) | FAIL |
| Specific learning disability | claim relevant | p18 | 845 | 1.13 (0.95-1.33) | 0.918 | 0.79 (0.35-1.63) | 0.210 | 0.89 (0.58-1.36) | 0.279 | 5436 (4064 / 1989) | 0.70 (0.54-0.96) | FAIL |
| Emotional disturbance | claim relevant | p18 | 809 | 1.02 (0.79-1.30) | 0.561 | 1.13 (0.48-2.15) | 0.642 | 0.47 (0.20-0.86) | 0.005 | 777 (5169 / 562) | 0.66 (0.45-1.06) | PASS |
| Other health impairment | claim relevant | p18 | 1,364 | 1.40 (1.12-1.77) | 0.997 | 0.71 (0.28-1.51) | 0.216 | 1.40 (1.08-1.84) | 0.993 | 361 (236 / 127) | 0.48 (0.37-0.66) | FAIL |
| All disabilities | claim relevant | p20 | 1,448 | 1.09 (1.00-1.19) | 0.980 | 0.96 (0.49-1.92) | 0.415 | 1.14 (1.00-1.27) | 0.977 | 101 (90 / 16) | 0.94 (0.81-1.11) | PASS |
| Autism / traumatic brain injury | ddd | p20 | 472 | 0.33 (0.20-0.86) | 0.013 | 1.58 (0.45-3.75) | 0.792 | 0.68 (0.39-1.29) | 0.109 | 250 (90 / 192) | 1.02 (0.33-3.90) | PASS |
| Autism / orthopedic impairment | ddd | p18 | 1,261 | 0.81 (0.54-1.23) | 0.156 | 0.72 (0.23-1.92) | 0.257 | 1.00 (0.69-1.43) | 0.491 | 12768 (10777 / 2203) | 1.67 (0.83-4.35) | PASS |
| Autism / hearing impairment | ddd | p18 | 1,381 | 0.79 (0.56-1.12) | 0.094 | 0.64 (0.22-1.50) | 0.138 | 0.76 (0.60-1.03) | 0.042 | 459 (389 / 74) | 0.90 (0.60-1.54) | PASS |
| Autism / visual impairment | ddd | p18 | 1,073 | 0.90 (0.61-1.34) | 0.294 | 0.66 (0.21-1.86) | 0.212 | 0.86 (0.54-1.36) | 0.264 | 13241 (10791 / 2488) | 1.04 (0.51-2.98) | PASS |

Re-fitted at p20 because the pooled p18 ESS was < 100: All disabilities (observed): p18 ESS 47, p20 ESS 101 (s23091 90 / s23092 16); Autism / traumatic brain injury (observed): p18 ESS 40, p20 ESS 250 (s23091 90 / s23092 192); All disabilities (complete_series): p18 ESS 47, p20 ESS 101 (s23091 90 / s23092 16).

## Verdict under the documented rule

| negative control | PCV r (95% CrI) | P(r>1) | CrI log width / autism's | class | complete-series PCV r | DDD autism / control r (restates the comparison on the control's cells) | DDD P(r>1) |
|---|---|---|---|---|---|---|---|
| Traumatic brain injury | 1.65 (0.77-4.29) | 0.892 | 4.75 | UNINFORMATIVE | 0.83 (0.42-1.59) | 0.33 (0.20-0.86) | 0.013 |
| Orthopedic impairment | 1.43 (0.91-2.26) | 0.937 | 2.52 | UNINFORMATIVE | 1.40 (0.87-2.23) | 0.81 (0.54-1.23) | 0.156 |
| Hearing impairment | 1.41 (1.05-1.91) | 0.990 | 1.65 | SIMILAR | 1.41 (1.05-1.92) | 0.79 (0.56-1.12) | 0.094 |
| Visual impairment | 1.45 (0.74-2.72) | 0.855 | 3.61 | UNINFORMATIVE | 1.36 (0.70-2.69) | 0.90 (0.61-1.34) | 0.294 |

Autism: 1.17 (0.97-1.39) on observed cells, 1.17 (0.97-1.39) on complete series. **Primary verdict: MIXED.** Complete-series sensitivity: MIXED (Traumatic brain injury UNINFORMATIVE, Orthopedic impairment UNINFORMATIVE, Hearing impairment SIMILAR, Visual impairment UNINFORMATIVE).

## Reading

- Documented verdict: MIXED – only hearing impairment is precise enough to count, and it leans like autism. Hearing impairment is SIMILAR: 1.41 (1.05-1.91), P(r > 1) 0.99, against autism's 1.17 (0.97-1.39) (CrI log width 1.65x autism's). NON-SPECIFIC needs at least two informative SIMILAR controls (there is 1); AUTISM-SPECIFIC needs at least two informative NULL-CENTRED controls and none SIMILAR (there are 0 NULL-CENTRED and 1 SIMILAR). Under MIXED the rule draws neither conclusion. What the data cannot establish, by the documented bar, is that the 1.17 is drift, because only 1 control is precise enough to count.
- Not counted (UNINFORMATIVE: 95% CrI more than 2x as wide as autism's on the log scale): traumatic brain injury 1.65 (0.77-4.29), 4.8x; orthopedic impairment 1.43 (0.91-2.26), 2.5x; visual impairment 1.45 (0.74-2.72), 3.6x. By the rule these carry no weight in either direction. Nor is the controls' position relative to autism stable across cell rules: 4 of 4 medians lie above autism's 1.17 on observed cells, but 3 of 4 on complete series, where traumatic brain injury is 0.83 (0.42-1.59), below autism's 1.17.
- DDD (autism relative to each control, log-ratio model): traumatic brain injury 0.33 (0.20-0.86); orthopedic impairment 0.81 (0.54-1.23); hearing impairment 0.79 (0.56-1.12); visual impairment 0.90 (0.61-1.34). These are not a second line of evidence. Autism is observed in all 1,469 core-v2 cells under both cell rules, so each DDD is fitted on exactly the control's cells, and cell by cell y_DDD = y_autism - y_control: the DDD restates the control-versus-autism comparison on the control's cells. As a rough check, the ratio of the separate-fit medians (autism / control, the DDD value the separate fits imply at full coverage) is traumatic brain injury 0.70, orthopedic impairment 0.81, hearing impairment 0.83, visual impairment 0.81, against DDD fits of 0.33, 0.81, 0.79, 0.90. The largest gap is traumatic brain injury (0.33 vs 0.70), where the DDD refits autism on 472 of its 1,469 cells. The documented rule does not let the DDDs change the verdict.
- Unstable across cell rules: traumatic brain injury 1.65 (0.77-4.29) observed vs 0.83 (0.42-1.59) complete series (472 vs 222 cells); autism / traumatic brain injury 0.33 (0.20-0.86) observed vs 1.22 (0.47-3.20) complete series (472 vs 222 cells). A median that moves by more than 1.5x with the cell rule carries no weight here.
- Claim-relevant categories (not negative controls; outside the verdict): intellectual disability 1.18 (0.94-1.49), speech or language impairment 1.09 (0.95-1.24), specific learning disability 1.13 (0.95-1.33), emotional disturbance 1.02 (0.79-1.30), other health impairment 1.40 (1.12-1.77), all disabilities 1.09 (1.00-1.19). The future+3 PCV placebo (reference setting shown; FAIL = some setting excludes 1) fails for intellectual disability (0.555, 0.405-0.810), speech or language impairment (0.797, 0.655-0.998), specific learning disability (0.705, 0.542-0.957), other health impairment (0.483, 0.373-0.658): the path shifted three cohorts early is also associated with these outcomes, so their PCV leans are not interpretable as vaccine effects. Autism and all four negative controls pass it. All disabilities is read at p20 and stays borderline: pooled ESS 101, seed 23092 alone 16.
- Negative controls also carry nominal rotavirus-2006 leans (descriptive; the documented rule concerns the PCV r only): traumatic brain injury RV2006 r 2.37 (1.09-4.58), P(r > 1) 0.985, hearing impairment RV2006 r 1.41 (1.07-1.82), P(r > 1) 0.991 (autism 0.95 (0.76-1.19)).

## Sensitivity: complete series

| outcome | cells | PCV r (95% CrI) | P(r>1) | RV2006 r | RotaShield r | ESS pooled | future+3 verdict |
|---|---|---|---|---|---|---|---|
| Autism | 1,469 | 1.17 (0.97-1.39) | 0.948 | 0.95 (0.76-1.19) | 0.87 (0.36-1.83) | 1279 | PASS |
| Traumatic brain injury | 222 | 0.83 (0.42-1.59) | 0.282 | 0.65 (0.34-1.25) | 0.98 (0.28-3.36) | 53045 | PASS |
| Orthopedic impairment | 1,027 | 1.40 (0.87-2.23) | 0.917 | 0.93 (0.62-1.37) | 0.87 (0.28-2.25) | 15093 | PASS |
| Hearing impairment | 1,232 | 1.41 (1.05-1.92) | 0.990 | 1.40 (1.04-1.80) | 1.35 (0.45-3.16) | 292 | PASS |
| Visual impairment | 722 | 1.36 (0.70-2.69) | 0.813 | 1.34 (0.87-2.03) | 1.13 (0.31-3.18) | 20134 | PASS |
| Intellectual disability | 973 | 1.19 (0.95-1.50) | 0.933 | 1.30 (0.87-1.93) | 0.72 (0.25-1.72) | 3497 | FAIL |
| Speech or language impairment | 1,420 | 1.10 (0.96-1.25) | 0.919 | 1.12 (0.95-1.31) | 1.05 (0.52-1.97) | 216 | PASS |
| Specific learning disability | 604 | 1.10 (0.93-1.29) | 0.871 | 0.99 (0.64-1.57) | 0.78 (0.35-1.63) | 4946 | FAIL |
| Emotional disturbance | 652 | 1.03 (0.81-1.32) | 0.603 | 0.73 (0.40-1.28) | 1.17 (0.50-2.35) | 8858 | PASS |
| Other health impairment | 1,301 | 1.34 (1.05-1.69) | 0.989 | 1.23 (0.92-1.70) | 0.66 (0.27-1.51) | 126 | FAIL |
| All disabilities | 1,448 | 1.09 (1.00-1.19) | 0.980 | 1.14 (1.00-1.27) | 0.96 (0.49-1.92) | 101 | PASS |
| Autism / traumatic brain injury | 222 | 1.22 (0.47-3.20) | 0.664 | 1.25 (0.70-2.19) | 1.01 (0.25-4.24) | 96864 | PASS |
| Autism / orthopedic impairment | 1,027 | 0.72 (0.45-1.21) | 0.106 | 1.05 (0.72-1.52) | 0.90 (0.27-2.34) | 10808 | PASS |
| Autism / hearing impairment | 1,232 | 0.76 (0.54-1.10) | 0.073 | 0.77 (0.60-1.05) | 0.65 (0.22-1.50) | 457 | PASS |
| Autism / visual impairment | 722 | 0.93 (0.54-1.71) | 0.401 | 0.72 (0.42-1.18) | 0.71 (0.23-2.01) | 10500 | PASS |

## Future+3 PCV placebo detail (primary rule; N(0, log 2) posterior median and 95% CrI per kernel setting)

| outcome | cells | actual path, reference | future+3 reference | linear only | spline only | rho .85 | verdict |
|---|---|---|---|---|---|---|---|
| Autism | 1,469 | 1.13 (0.96-1.35) | 0.86 (0.66-1.17) | 0.86 (0.66-1.17) | 0.85 (0.64-1.20) | 0.85 (0.65-1.17) | PASS |
| Traumatic brain injury | 472 | 1.67 (0.77-3.88) | 1.14 (0.45-3.78) | 0.84 (0.38-2.58) | 1.19 (0.47-3.88) | 1.01 (0.41-3.37) | PASS |
| Orthopedic impairment | 1,261 | 1.38 (0.90-2.16) | 0.59 (0.39-1.05) | 0.59 (0.39-1.05) | 0.60 (0.38-1.15) | 0.75 (0.44-1.64) | PASS |
| Hearing impairment | 1,381 | 1.34 (1.02-1.77) | 0.97 (0.63-1.74) | 0.97 (0.63-1.74) | 0.90 (0.58-1.70) | 1.08 (0.66-2.17) | PASS |
| Visual impairment | 1,073 | 1.51 (0.87-2.67) | 0.95 (0.54-2.07) | 0.95 (0.54-2.07) | 0.85 (0.49-1.83) | 0.97 (0.52-2.32) | PASS |
| Intellectual disability | 1,117 | 1.17 (0.93-1.47) | 0.55 (0.40-0.81) | 0.55 (0.40-0.81) | 0.52 (0.38-0.78) | 0.62 (0.44-0.94) | FAIL |
| Speech or language impairment | 1,459 | 1.09 (0.95-1.24) | 0.80 (0.65-1.00) | 0.80 (0.65-1.00) | 0.82 (0.66-1.05) | 0.80 (0.66-1.00) | FAIL |
| Specific learning disability | 845 | 1.12 (0.95-1.33) | 0.70 (0.54-0.96) | 0.70 (0.54-0.96) | 0.77 (0.57-1.12) | 0.74 (0.57-0.99) | FAIL |
| Emotional disturbance | 809 | 1.02 (0.80-1.30) | 0.66 (0.45-1.06) | 0.66 (0.45-1.06) | 0.64 (0.43-1.07) | 0.67 (0.46-1.09) | PASS |
| Other health impairment | 1,364 | 1.38 (1.09-1.75) | 0.48 (0.37-0.66) | 0.48 (0.37-0.66) | 0.48 (0.37-0.68) | 0.57 (0.44-0.80) | FAIL |
| All disabilities | 1,448 | 1.10 (1.01-1.20) | 0.94 (0.81-1.11) | 0.94 (0.81-1.10) | 1.01 (0.86-1.21) | 0.96 (0.83-1.12) | PASS |
| Autism / traumatic brain injury | 472 | 0.43 (0.23-1.04) | 1.02 (0.33-3.90) | 1.20 (0.41-4.26) | 0.77 (0.29-3.06) | 0.97 (0.34-3.59) | PASS |
| Autism / orthopedic impairment | 1,261 | 0.84 (0.58-1.25) | 1.67 (0.83-4.35) | 1.67 (0.83-4.35) | 1.63 (0.78-4.53) | 1.26 (0.63-3.31) | PASS |
| Autism / hearing impairment | 1,381 | 0.86 (0.64-1.17) | 0.90 (0.60-1.54) | 0.90 (0.60-1.54) | 0.97 (0.61-1.87) | 0.86 (0.54-1.66) | PASS |
| Autism / visual impairment | 1,073 | 0.92 (0.63-1.38) | 1.04 (0.51-2.98) | 1.04 (0.51-2.98) | 1.32 (0.56-4.32) | 1.11 (0.48-3.47) | PASS |

The future+3 path has no cohort at the zero (pre-licensure) level on these cells (every c + 3 >= 2000), so it is fittable but never a feasible breakpoint in the global_placebo_family sense; the verdict is the D4 primary rule only. The 1-D fits use point paths and one product at a time, so their actual-path values differ from the importance-sampled joint posteriors above.

## Caveats

- Suppression is informative: EDFacts suppresses small cells with `x` / `X` (complementary suppression included), so rare categories (traumatic brain injury, visual impairment, and several categories at ages 3-5) are observed mainly in large states and in their higher-count years. The `observed` rule therefore selects cells partly on the outcome; the `complete_series` rule removes within-series selection at the cost of fewer states. Separately, `-` marks data not available (intellectual disability 6; specific learning disability 7; emotional disturbance 8; other health impairment 16 core-v2 cells, counted as not published, not as suppressed) and `*` marks data flagged for questionable quality in env2013 / env2015 (no core-v2 cell in these categories carries it); both are excluded under either rule, as suppressed cells are, so no estimate depends on the relabelling.
- Rare-category log rates carry large Poisson noise; the kernel's population scaling and the per-class variance plug-in absorb it by widening the intervals (they do not remove the suppression selection above), so the negative controls are less precise than autism by construction; 3 of 4 are too imprecise to count under the documented width limit.
- `All disabilities` contains autism itself (a mechanical share of any autism signal) and developmental delay; CA79 cells for it exist only where all 13 CASEMIS categories are numeric.
- Developmental delay (ages 3-9 where states use it) competes with the specific categories at young ages; state changes in its use are exactly the kind of state x cohort reporting drift the negative controls are meant to reveal.
- CASEMIS hearing impairment = hard of hearing + deaf (the federal category). The CASEMIS CSV the core reads lacks four categories (its parser dropped float-typed columns); the counts here come from the HTML reports it was derived from, checked equal on every overlapping value.