# 3361 Schedule-load and same-product meta-regressions

Generated by `07_analysis_code/3361_meta_regressions.py` (deterministic, no random draws; 131 s of computation). Documented rule: `qa/layers_20260929/3361_rule.json`, sha256 `144a66ac38743a48317da334acd37d66af90271d1ad690369d176c708efb5768`, written before any statistic below was computed; priors, grids, context sets and decision rules are read from it. The 36 input files listed in the rule were verified unchanged at run time (hierarchy_final_p18.json sha256 `e103ab7ab14be768603cc2c741373160b39698794d8c3081c873879612f480ca`). Files: `3361_meta_regressions.csv` (results, long format), `3361_meta_regressions_inputs.csv` (one row per context or series: loads, ages, own-data r, bias), `3361_meta_regressions.json`, `08_figures/3361_meta_regressions.png`.

## Headline

- **(a) Load dependence ('too many too soon'), primary A1**: 10 pooled hierarchy contexts that have a 3307 load, one likelihood per series.
  - Per **+5 antigens** due by the product's first dose: ratio **0.999 (0.507-1.965)**, P(slope > 0) = 0.499; prior 0.507-1.972; information ratio (posterior SD / prior SD) 0.996. Predicted r at the lowest load (5 antigens) 1.001 (0.551-1.757), at the highest (6.29) 1.001 (0.556-1.736). Verdict: NOT SUPPORTED; NOT INFORMATIVE (information ratio >= 0.8).
  - Per **+5 injections**: ratio **0.982 (0.516-1.865)**, P(slope > 0) = 0.477; prior 0.507-1.972; information ratio 0.945. Predicted r at 1 injection 1.007 (0.539-1.817), at 5 injections 0.991 (0.517-1.851). Verdict: NOT SUPPORTED; NOT INFORMATIVE (information ratio >= 0.8).
- **(a)(ii) Jump rows**: 2 gate-passing 3309 jump rows share no overlap group with the hierarchy (US_PCV7_2000__INDIVIDUAL_SURVEYS, US_PCV7_2000__INDIVIDUAL_SURVEYS__NHIS). Both are U.S. PCV7 2000 rows at one load value, so on their own (A6) the slope is not identified (information ratio 1.000: posterior = prior). Added to A1 (A7): +5 antigens 1.000 (0.508-1.967), +5 injections 0.996 (0.530-1.872).
- **(b) Same-product consistency** (statistical version: the contexts' own data, no bias allowance): modern rotavirus (5 contexts) tau 0.09 (0.00-0.50), Q = 1.93 (df 4, p = 0.749); with U.S. rotavirus 2006 tau 0.06 (0.00-0.33), Q = 1.94 (df 5, p = 0.857); PCV (4) tau 0.16 (0.01-0.67), Q = 2.64 (df 3, p = 0.450); MMR (2) tau 0.30 (0.01-1.03), Q = 1.03 (df 1, p = 0.310). Outliers (leave-one-out homogeneity conflict p < 0.05): none (smallest p 0.108). Verdict for every product and version: contexts AGREE. The tau posteriors are wide: 2 to 6 contexts per product cannot pin down heterogeneity, so agreement means no detectable conflict, not proven homogeneity.
- **U.S. PCV** (own data 1.17 (0.97-1.39)) against the other PCV contexts (Sweden, Germany, Hong Kong), whose pooled value under homogeneity is 0.95 (0.79-1.14): conflict p = 0.108 (random effects p = 0.442; with the assumed bias SDs p = 0.828). By the documented rule U.S. PCV is **CONSISTENT** with the other PCV contexts.
- **Within vs between products** (two-level model, 15 pooled contexts, statistical): within-product SD tau_w 0.069 (0.000-0.240), between-product SD tau_b 0.102 (0.000-0.408), P(tau_w > tau_b) = 0.379; Q_within = 5.60 (df 8, p = 0.692), Q_between = 5.59 (df 6, p = 0.471). With the bias SDs: tau_w 0.125 (0.000-0.467), tau_b 0.153 (0.001-0.591), P(tau_w > tau_b) = 0.437. Neither spread is distinguishable from zero, and neither dominates.
- **(c) Age at exposure** (critical window; C1 = the A1 contexts with their first-dose ages): ratio per +10 months **0.906 (0.321-2.523)**, P(slope < 0) = 0.574; prior 0.322-3.102; information ratio 0.910. Verdict: NOT SUPPORTED; NOT INFORMATIVE (information ratio >= 0.8). Only Japan MMR (12 months) and Sweden PCV (3 months) differ from the 2-month products, so the age slope rests on Japan MMR (a weak likelihood) and is confounded with that product.

In plain terms: the 10 contexts with a verified schedule load span 5 to 6.3 antigens and 1 to 5 injections at the product's first dose, and once the hierarchy's assumed bias allowances (0.35 or 0.70 on the log scale) are added each context carries little weight. The data leave the slope priors almost where they started: the analysis neither supports nor rules out load dependence of the size the prior allows (one prior SD is a doubling or halving over 10 antigens or injections). The same holds for age at first dose. Within products the contexts agree with one another, including U.S. PCV (no outlier anywhere).

## Data

Loads are counted from the 3307 schedule strings: in the regime the product joined (`schedule_before` of its introduction row), the distinct antigens and the injections (route `inj`; oral and percutaneous doses are not injections) scheduled at or before the product's first scheduled age. The parser reproduces the published 3307 0-24 month loads of all 583 event rows (4664 values, 0 mismatches). Own-data r: the published own-data posterior (hierarchy_final_p18.json 'unpooled': the context's own likelihood alone with an N(0, log 2) prior); the curves the models use reproduce it (see Checks). Jump rows: the 3309 calibrated interval.

| Context | Product | Bias SD | Own-data r | 3307 load event | Antigens | Injections | 0-24 m antigens / injections | First dose (months) |
|---|---|---|---|---|---|---|---|---|
| U.S. PCV | PCV | 0.70 | 1.17 (0.97-1.39) | US_PCV7_2000 | 6 | 5 | 10 / 16 | 2 |
| U.S. RotaShield | RotaShield | 0.70 | 0.87 (0.36-1.83) | US_ROTASHIELD_1998 | 6 | 5 | 10 / 15 | 2 |
| Japan MMR | MMR | 0.70 | 0.85 (0.28-2.32) | JPN_MMR_OPTION_1989 | 6 | 4 | 6 / 5 | 12 |
| Sweden PCV | PCV | 0.70 | 1.01 (0.79-1.56) | SWE_PCV_2009 | 5 | 1 | 8 / 4 | 3 |
| Germany PCV | PCV | 0.70 | 1.03 (0.27-3.96) | DEU_PCV_2006 | 6 | 1 | 11 / 8 | 2 |
| Hong Kong PCV | PCV | 0.70 | 0.89 (0.66-1.16) | - | no 3307 load | - | - | - |
| England Rotavirus | modern_rotavirus | 0.35 | 0.94 (0.86-1.02) | ENG_ROTA_2013 | 6 | 2 | 10 / 9 | 2 |
| Scotland Rotavirus | modern_rotavirus | 0.35 | 1.55 (0.38-5.31) | SCO_ROTA_2013 | 6 | 2 | 10 / 9 | 2 |
| Northern Ireland Rotavirus | modern_rotavirus | 0.35 | 1.01 (0.88-1.16) | NIR_ROTA_2013 | 6 | 2 | 10 / 9 | 2 |
| Canada Rotavirus (Excl. Quebec) | modern_rotavirus | 0.35 | 0.99 (0.32-1.89) | CAN_PE_ROTA_2010 and 6 more provinces | 6.29 | 2.14 | 11.29 / 11.43 | 2 |
| Quebec Rotavirus | modern_rotavirus | 0.35 | 1.07 (0.74-1.48) | CAN_QC_ROTA_2011 | 6 | 2 | 11 / 11 | 2 |
| Czechia MMR | MMR | 0.70 | 1.04 (0.32-3.66) | - | no 3307 load | - | - | - |
| Brazil Maternal DTaP | MATERNAL_DTAP | 0.70 | 0.77 (0.51-1.14) | - | no 3307 load | - | - | - |
| Colombia Varicella | VARICELLA | 0.35 | 0.76 (0.51-1.07) | - | no 3307 load | - | - | - |
| Tanzania Schedule | SCHEDULE | 0.35 | 1.12 (0.54-1.69) | - | no 3307 load | - | - | - |
| U.S. Rotavirus 2006 (untied; sensitivity) | modern_rotavirus | 0.35 | 0.95 (0.76-1.19) | US_ROTATEQ_2006 | 7 | 6 | 13 / 25 | 2 |
| U.S. PCV7 (NSCH Survey Jump) (jump row, (ii)) | PCV | 0.70 | 1.93 (0.43-8.58) | US_PCV7_2000 | 6 | 5 | 10 / 16 | 2 |
| U.S. PCV7 (NHIS Survey Jump) (jump row, (ii)) | PCV | 0.70 | 0.88 (0.28-2.77) | US_PCV7_2000 | 6 | 5 | 10 / 16 | 2 |

Contexts without a 3307 load (left out of (a) and (c), not imputed): Hong Kong PCV: Hong Kong is not one of the 17 jurisdictions of 3307_schedule_load_events.csv; Czechia MMR: Czechia is not one of the 17 jurisdictions of 3307_schedule_load_events.csv; Brazil Maternal DTaP: Brazil is not in 3307; maternal dTpa is given in pregnancy, and no verified count of the load in place at that time exists in the project (not imputed, not set to 0); Colombia Varicella: Colombia is not one of the 17 jurisdictions of 3307_schedule_load_events.csv; Tanzania Schedule: Tanzania is not one of the 17 jurisdictions of 3307_schedule_load_events.csv.

Flags (reported, not excluded): the Japan MMR row is option / uptake-dependent with secondary verification; Scotland and Northern Ireland assume UK-wide dates; Canada excl. Quebec averages the 7 provinces of its block (6 from the knowledge-based national-typical track, British Columbia from its verified track: 6-8 antigens, 2-3 injections); Quebec uses the national-typical track; U.S. injection counts assume separate products (3307: an upper bound).

## (a) Load dependence, all variants

Model: theta_c = alpha + pi_product + beta (x_c - xbar) + u_c, with product and context random intercepts on the hierarchy's heterogeneity grid (product SD 0.1 / 0.35 / 0.7 x context SD 0.1 / 0.35, equal mass). Each context's own likelihood curve is evaluated at theta_c + e_c + B_group (e_c ~ N(0, bias SD^2); B = the shared UK / Canada / care-bridge terms, SD 0.20 / 0.20 / 0.25) and integrated exactly on the 801-point grid. alpha ~ N(0, log 2^2 + 0.15^2); beta ~ N(0, (log 2 / 10)^2) per antigen and per injection.

| Variant | Metric | n | Ratio per +5 (95% CrI) | Prior 95% | P(slope > 0) | Information ratio | Predicted r at lowest load | Predicted r at highest load |
|---|---|---|---|---|---|---|---|---|
| A1 | antigens | 10 | 0.999 (0.507-1.965) | 0.507-1.972 | 0.499 | 0.996 | 5: 1.001 (0.551-1.757) | 6.29: 1.001 (0.556-1.736) |
| A1 | injections | 10 | 0.982 (0.516-1.865) | 0.507-1.972 | 0.477 | 0.945 | 1: 1.007 (0.539-1.817) | 5: 0.991 (0.517-1.851) |
| A2 | antigens | 11 | 0.991 (0.507-1.935) | 0.507-1.972 | 0.489 | 0.985 | 5: 0.991 (0.556-1.709) | 7: 0.988 (0.556-1.693) |
| A2 | injections | 11 | 0.965 (0.544-1.712) | 0.507-1.972 | 0.452 | 0.843 | 1: 1.005 (0.542-1.807) | 6: 0.969 (0.515-1.782) |
| A3 | antigens | 10 | 1.032 (0.541-1.969) | 0.507-1.972 | 0.538 | 0.950 | 6: 0.974 (0.458-2.036) | 11.29: 1.010 (0.545-1.807) |
| A3 | injections | 10 | 1.018 (0.620-1.669) | 0.507-1.972 | 0.528 | 0.728 | 4: 0.976 (0.422-2.212) | 16: 1.019 (0.449-2.264) |
| A4 | antigens | 10 | 1.049 (0.548-1.998) | 0.507-1.972 | 0.558 | 0.951 | 5: 1.010 (0.688-1.375) | 6.29: 1.022 (0.704-1.373) |
| A4 | injections | 10 | 1.101 (0.712-1.648) | 0.507-1.972 | 0.676 | 0.614 | 1: 0.983 (0.657-1.381) | 5: 1.061 (0.702-1.478) |
| A5 | antigens | 10 | 0.995 (0.261-3.797) | 0.257-3.891 | 0.497 | 0.985 | 5: 1.000 (0.533-1.826) | 6.29: 1.000 (0.553-1.746) |
| A5 | injections | 10 | 0.945 (0.308-2.895) | 0.257-3.891 | 0.461 | 0.823 | 1: 1.019 (0.506-1.988) | 5: 0.972 (0.448-2.072) |
| A6 | antigens | 2 | 1.000 (0.507-1.973) | 0.507-1.972 | 0.500 | 1.000 | 6: 1.094 (0.377-3.174) | 6: 1.094 (0.377-3.174) |
| A6 | injections | 2 | 1.000 (0.507-1.973) | 0.507-1.972 | 0.500 | 1.000 | 5: 1.094 (0.377-3.174) | 5: 1.094 (0.377-3.174) |
| A7 | antigens | 12 | 1.000 (0.508-1.967) | 0.507-1.972 | 0.500 | 0.996 | 5: 1.018 (0.573-1.748) | 6.29: 1.018 (0.579-1.726) |
| A7 | injections | 12 | 0.996 (0.530-1.872) | 0.507-1.972 | 0.495 | 0.928 | 1: 1.019 (0.551-1.822) | 5: 1.016 (0.548-1.829) |
| A7b | antigens | 11 | 1.001 (0.508-1.969) | 0.507-1.972 | 0.501 | 0.996 | 5: 1.028 (0.572-1.788) | 6.29: 1.029 (0.578-1.768) |
| A7b | injections | 11 | 1.006 (0.532-1.900) | 0.507-1.972 | 0.507 | 0.937 | 1: 1.026 (0.552-1.848) | 5: 1.030 (0.545-1.892) |

- A1: primary: (i) the pooled hierarchy contexts that have a 3307 load (one likelihood per series), primary load.
- A2: A1 + U.S. rotavirus 2006 (untied in the hierarchy) as a modern-rotavirus context.
- A3: A1 with the 0-24 month before-regime load (3307 antigens_before / injections_before).
- A4: A1 with every bias term removed (context and shared group SDs 0; statistical uncertainty only).
- A5: A1 with the slope prior SD doubled (log 2 / 5 per unit).
- A6: (ii) the gate-passing jump rows without hierarchy overlap, alone.
- A7: (ii) A1 + those jump rows as PCV contexts (U.S. country bias group).
- A7b: A7 without jump rows whose own placebo check fails.

The variants in which the data move the prior most are A4 (no bias terms: +5 injections 1.101 (0.712-1.648), information ratio 0.614) and A3 (0-24 month before-regime load, 4-16 injections: 1.018 (0.620-1.669), information ratio 0.728); both CrIs include 1. Savage-Dickey Bayes factors for a zero slope (descriptive, not documented) lie between 1.00 and 1.48 (the CSV has each).

## (b) Same-product consistency

Per product: theta_c = mu_p + u_c, u_c ~ N(0, tau^2), tau ~ HalfNormal(0.5), mu_p ~ N(0, log 2^2 + 0.15^2), exact own curves. Q: Cochran Q (GLS) on the mean and SD of each own curve normalised on the grid. Conflict p: two-sided leave-one-out cross-validatory conflict p-value of the context's own curve against the predictive from the other contexts of the product, under homogeneity (tau = 0) and under random effects (tau from the others). With-bias version: the assumed bias SDs and the shared group terms are added (the shared term integrated jointly).

| Product | Version | k | tau (95% CrI) | P(tau > 0.1) | P(tau > 0.35) | tau, HalfNormal(1) prior | Product mean r | Q (df, p) | I^2 | Verdict |
|---|---|---|---|---|---|---|---|---|---|---|
| modern_rotavirus | statistical | 5 | 0.088 (0.002-0.498) | 0.430 | 0.064 | 0.095 (0.002-0.669) | 0.987 (0.822-1.279) | 1.93 (4, 0.749) | 0.00 | contexts AGREE |
| modern_rotavirus | with_bias | 5 | 0.187 (0.008-0.746) | 0.700 | 0.226 | 0.232 (0.010-1.088) | 1.042 (0.615-1.768) | 0.93 (4, 0.921) | 0.00 | contexts AGREE |
| modern_rotavirus_plus_US_RV2006 | statistical | 6 | 0.062 (0.000-0.332) | 0.285 | 0.020 | 0.064 (0.000-0.381) | 0.978 (0.862-1.156) | 1.94 (5, 0.857) | 0.00 | contexts AGREE |
| modern_rotavirus_plus_US_RV2006 | with_bias | 6 | 0.159 (0.006-0.644) | 0.652 | 0.163 | 0.184 (0.008-0.858) | 1.016 (0.654-1.587) | 0.94 (5, 0.967) | 0.00 | contexts AGREE |
| PCV | statistical | 4 | 0.159 (0.008-0.672) | 0.666 | 0.160 | 0.185 (0.010-0.994) | 1.035 (0.736-1.431) | 2.64 (3, 0.450) | 0.00 | contexts AGREE |
| PCV | with_bias | 4 | 0.253 (0.012-0.897) | 0.778 | 0.348 | 0.359 (0.016-1.436) | 1.024 (0.475-2.206) | 0.08 (3, 0.994) | 0.00 | contexts AGREE |
| MMR | statistical | 2 | 0.304 (0.014-1.035) | 0.815 | 0.431 | 0.518 (0.024-1.847) | 0.897 (0.319-2.508) | 1.03 (1, 0.310) | 0.03 | contexts AGREE |
| MMR | with_bias | 2 | 0.313 (0.014-1.054) | 0.820 | 0.444 | 0.546 (0.025-1.885) | 0.928 (0.300-2.873) | 0.90 (1, 0.342) | 0.00 | contexts AGREE |

Leave-one-out conflict checks (statistical version; the with-bias homogeneity p in the last p column):

| Product set | Context | Own r (flat prior) | Pooled r of the others (homogeneity) | Conflict p, homogeneity | Conflict p, random effects | Conflict p, with bias | Flag |
|---|---|---|---|---|---|---|---|
| modern_rotavirus | England Rotavirus | 0.94 (0.86-1.02) | 1.03 (0.91-1.16) | 0.238 | 0.548 | 0.743 | consistent |
| modern_rotavirus | Northern Ireland Rotavirus | 1.01 (0.88-1.16) | 0.95 (0.88-1.03) | 0.457 | 0.899 | 0.939 | consistent |
| modern_rotavirus | Scotland Rotavirus | 4.45 (0.01-27.32) | 0.97 (0.90-1.04) | 0.367 | 0.374 | 0.387 | consistent |
| modern_rotavirus | Canada Rotavirus (Excl. Quebec) | 0.49 (0.00-1.86) | 0.97 (0.90-1.03) | 0.573 | 0.534 | 0.483 | consistent |
| modern_rotavirus | Quebec Rotavirus | 1.07 (0.73-1.50) | 0.96 (0.90-1.03) | 0.569 | 0.757 | 0.915 | consistent |
| modern_rotavirus_plus_US_RV2006 | England Rotavirus | 0.94 (0.86-1.02) | 1.01 (0.90-1.13) | 0.303 | 0.553 | 0.788 | consistent |
| modern_rotavirus_plus_US_RV2006 | Northern Ireland Rotavirus | 1.01 (0.88-1.16) | 0.95 (0.88-1.03) | 0.449 | 0.767 | 0.979 | consistent |
| modern_rotavirus_plus_US_RV2006 | Scotland Rotavirus | 4.45 (0.01-27.32) | 0.97 (0.90-1.03) | 0.367 | 0.371 | 0.385 | consistent |
| modern_rotavirus_plus_US_RV2006 | Canada Rotavirus (Excl. Quebec) | 0.49 (0.00-1.86) | 0.97 (0.90-1.03) | 0.575 | 0.552 | 0.498 | consistent |
| modern_rotavirus_plus_US_RV2006 | Quebec Rotavirus | 1.07 (0.73-1.50) | 0.96 (0.90-1.03) | 0.565 | 0.676 | 0.881 | consistent |
| modern_rotavirus_plus_US_RV2006 | U.S. Rotavirus 2006 | 0.95 (0.75-1.20) | 0.97 (0.90-1.04) | 0.872 | 0.811 | 0.846 | consistent |
| PCV | U.S. PCV | 1.17 (0.97-1.40) | 0.95 (0.79-1.14) | 0.108 | 0.442 | 0.828 | consistent |
| PCV | Sweden PCV | 1.01 (0.78-1.62) | 1.07 (0.92-1.26) | 0.774 | 0.967 | 0.990 | consistent |
| PCV | Germany PCV | 1.30 (0.00-276.39) | 1.05 (0.92-1.21) | 0.960 | 0.958 | 0.959 | consistent |
| PCV | Hong Kong PCV | 0.89 (0.65-1.17) | 1.11 (0.96-1.32) | 0.149 | 0.450 | 0.809 | consistent |
| MMR | Japan MMR | 0.34 (0.00-2.65) | 1.04 (0.32-3.76) | 0.413 | 0.429 | 0.448 | consistent |
| MMR | Czechia MMR | 2.23 (0.14-161.70) | 0.84 (0.27-2.34) | 0.568 | 0.582 | 0.600 | consistent |

Own r (flat prior) is the context's own curve normalised on the grid [-6, 6] (the conflict test's own-data side). For the flat or one-sided curves (Germany, Scotland, Canada excl. Quebec, Japan, Czechia) it depends on the grid limits and carries little weight; the own-data r with the N(0, log 2) prior is in the Data table.

| Two-level model | Version | Contexts | tau_w within products | tau_b between products | P(tau_w > tau_b) | Q_within (df, p) | Q_between (df, p) |
|---|---|---|---|---|---|---|---|
| pooled_15 | statistical | 15 | 0.069 (0.000-0.240) | 0.102 (0.000-0.408) | 0.379 | 5.60 (8, 0.692) | 5.59 (6, 0.471) |
| pooled_15 | with_bias | 15 | 0.125 (0.000-0.467) | 0.153 (0.001-0.591) | 0.437 | 1.91 (8, 0.984) | 0.68 (6, 0.995) |
| plus_US_RV2006 | statistical | 16 | 0.058 (0.000-0.205) | 0.101 (0.000-0.404) | 0.335 | 5.61 (9, 0.778) | 5.61 (6, 0.468) |
| plus_US_RV2006 | with_bias | 16 | 0.115 (0.000-0.425) | 0.149 (0.001-0.581) | 0.419 | 1.92 (9, 0.993) | 0.67 (6, 0.995) |

## (c) Age at exposure

Availability: 02_policy_events.csv carries a context-specific first-dose age only for Japan MMR (12 months, JPN_MMR_1989_1993: Honda et al. 2005 primary study); England Rotavirus (2 months, UK_ROTA_2013: https://www.gov.uk/government/collections/rotavirus-immunisation-programme-for-infants); Scotland Rotavirus (2 months, UK_ROTA_2013: https://www.gov.uk/government/collections/rotavirus-immunisation-programme-for-infants); Northern Ireland Rotavirus (2 months, UK_ROTA_2013: https://www.gov.uk/government/collections/rotavirus-immunisation-programme-for-infants). For U.S. PCV, U.S. RotaShield, U.S. Rotavirus 2006 the field is empty; for Sweden PCV (6, vaccine_class), Germany PCV (6, vaccine_class), Canada Rotavirus (Excl. Quebec) (6, vaccine_class), Quebec Rotavirus (6, vaccine_class) it holds vaccine-class placeholders, which are not scheduled ages; Hong Kong PCV, Czechia MMR, Brazil Maternal DTaP, Colombia Varicella, Tanzania Schedule have no 02 row. 3300_jump_inventory.csv has no scheduled-age column (45 columns, none an age). The scheduled first-dose age is therefore taken from the 3307 introduction rows (the event files that give the loads); it agrees with 02 wherever 02 is context-specific (4 of 4). Contexts without a 3307 row have no event-file age and are left out. Prior: N(0, (log 2 / 12)^2) per month.

| Variant | n | Ratio per +10 months (95% CrI) | Prior 95% | P(slope < 0) | Information ratio | Predicted r at 2 months | Predicted r at 12 months |
|---|---|---|---|---|---|---|---|
| C1 | 10 | 0.906 (0.321-2.523) | 0.322-3.102 | 0.574 | 0.910 | 1.006 (0.556-1.759) | 0.910 (0.297-2.708) |
| C2 | 11 | 0.909 (0.322-2.527) | 0.322-3.102 | 0.572 | 0.909 | 0.994 (0.563-1.698) | 0.901 (0.296-2.664) |
| C3 | 10 | 0.847 (0.326-2.108) | 0.322-3.102 | 0.639 | 0.821 | 1.024 (0.703-1.392) | 0.863 (0.327-2.158) |

- C1: primary: contexts of A1 with their 3307 first-dose age.
- C2: C1 + U.S. rotavirus 2006.
- C3: C1 without bias terms.

## Checks

- Own curves: the 12 non-core curves (global_hierarchy.structure) reproduce the hierarchy's published own-data posteriors to 0.0e+00 on the log scale; the core curves reproduce the us_alone diagnostics within U.S. PCV 0.0030, U.S. RotaShield 0.0123, Japan MMR 0.0015, U.S. Rotavirus 2006 0.0028 (tolerance 0.02).
- Joint U.S. likelihood (Gaussian approximation): likelihood correlations US_PCV|US_RotaShield -0.185, US_PCV|US_RV2006 +0.042, US_RotaShield|US_RV2006 -0.026. It enters as the product of its per-context marginal likelihoods; after the 0.70 bias SDs the U.S. PCV / RotaShield correlation is -0.014.
- 3307 parser: 0 mismatches over 4664 published values.
- Jump rows: 57 rows; hierarchy context validation jump (decided in hierarchy): 10; reconciliation gate FAIL: 16; interval not calibrated: 6; selected: 2; shares an overlap group with a hierarchy context: 3; no finite r / interval: 20. No selected row shares an overlap group with a hierarchy context.
- The shared bias groups and bias SDs equal those of global_hierarchy.structure() and hierarchy_final_p18.json (asserted).

## Caveats

- The context posteriors are likelihood-scale exposure-mixture parameters whose spread mixes heterogeneity and bias (3320). The regressions use each series once (own curves, never the pooled posteriors) and keep the hierarchy's bias model, so the slopes are conditional on that model.
- 10 contexts, 4 products and loads clustered at 5-6.3 antigens and 1-5 injections: a null slope here is absence of evidence, not evidence of absence. The information ratios say how little the data move the priors.
- Load is a property of the schedule, not of the individual child; uptake, catch-up and co-timed changes sit inside the context estimates.
- The age analysis rests on 1 context(s) at 12 months or later (Japan MMR, a weak likelihood) and is confounded with the MMR product.
- The (ii) jump rows are 2 survey rows for one U.S. introduction; own placebo check failing: U.S. PCV7 (NHIS Survey Jump) (A7b drops them).