Decisions 45-50: scenario tuning, NC truths, oracle own estimand, separate PVI phases, validation

Co-Authored-By: Claude Opus 5.5 <[email protected]>

docs/decisions.md

@@ -91,3 +91,14 @@ Each decision is listed with the date and the reason for it. Settled decisions a
 | 42 | **Deferral:** in W0–W3, W5 and WR, deferred PVI matches Sorber (3.6% late PVI among patients without early PVI). **W4 is calibrated to 15%**, which is trial-like crossover, and in W4 deferral is driven by severity. In WR, deferral is at the Sorber rate but driven by severity. PVI after CLTI was lowered to about 1% per month, so most late PVI is elective. | At the Medicare rate, the deferral threat would be almost invisible. Trials (SUPER 29% in year 1) show much more crossover. |
 | 43 | **Physician style coefficient 1.6** (W2, W3, WR). Among physicians with ≥11 patients this gives a median early-PVI rate of 0.8%, a maximum of 75%, and 4.7% of physicians ≥2 SD outliers who perform 38% of early PVIs. Hicks 2020 reported 0%, 58% and 5.6%. | Close enough to Hicks 2020. Not tuned further. |
 | 44 | **Within a month the order is:** severity update → outcome events → death → elective PVI decision. A MALE event and death in the same month count the MALE event first. | Matches discrete-time pooled logistic analysis. |
+
+## Build notes: truths, estimators, validation (2026-10-27)
+
+| # | Decision | Why |
+|---|---|---|
+| 45 | **Scenario effects, tuned in W0 against the disease composite:** harm (relief 0.3, kappa 0.039) gives RD(A−B1) **+5.1 pp**. Benefit-then-harm (relief 2.0, kappa 0.075) gives **−1.2 pp at 24 months and +3.1 pp at 60 months**. The same mechanism is used in every world. Where skill varies (W1, WR) the true effects are larger (+7 pp), because exp(−skill) is convex: on average a PVI is done worse than one at average skill. | D34 sets the headline on the disease composite. MALE carries an extra +7 to +15 pp that comes purely from the outcome definition. |
+| 46 | **Truths for the negative controls are taken from the truths table, not assumed to be 0.** In the non-null scenarios PVI changes death, so a negative-control outcome's 5-year risk moves slightly through the competing risk (e.g. NCO-1 −0.08 pp). In the null scenario the effect is exactly 0. | Honest scoring. |
+| 47 | **The own estimand of landmark and naive is estimated within each replicate by an oracle run:** the same estimator with exact severity, true frailty and smoking, and the latent skill and style. Estimation bias is the paired difference (estimate − oracle). | A large-n limit was too noisy: the landmark MCSE at n = 10⁵ was 2.8 pp. |
+| 48 | **The PVI model is fitted separately inside the window (month indicators) and in the deferred period (spline in month).** | A pooled model misspecified the window probabilities that the CCW and sequential weights depend on. |
+| 49 | **V2 is the V1 fit with a physician-cluster-robust SE.** Under V4, physicians without an estimable fixed effect in a phase get probability 0, and the V4 population is restricted to physicians with variation in early PVI. | Same point estimate by construction (D15). |
+| 50 | **Validation (W0, harm, 40 replicates of 50,000):** CCW A−B1 is +5.12 vs a truth of +5.11 on the disease composite (z = 0.0); sequential ITT is +5.83 vs +5.71 (z = 0.6); MALE and NCO-1 have |z| ≤ 0.7. The model SE is 0–20% below the empirical SE for CCW, so the bootstrap check in D28 still needs doing. | Plan §15 validation (a). |