[feedback][clinical-trial-ipd-sim] AEREL Assignment Not Realistic

[lengning]

Gap in Skill Specification

The skill documents hazard-driven AE generation extensively (per-visit hazard, frailty loading, severity assignment) but does not specify how AEREL (causality assessment) should be generated. The scm_spec_template.md §6.3 describes:

• λ(t) = exp(log_haz_base + β_arm + f_cluster + β_recurrence) for incidence
• p_severe for grade conditioning
• P(reported | grade) for reporting probability

But there is no guidance on P(RELATED | arm, AE_type, grade) — the investigator's causality assessment.

Consequence

Implementers default to deterministic assignment:

AEREL = if (arm == "PEMBRO") "RELATED" else "NOT RELATED"

This produces:

1. TRAE ≡ Any AE in the active arm (100% relatedness)
2. Placebo immune AEs invisible in TRAE-filtered outputs (0% relatedness)
3. Both violate the blinding principle: in a double-blind trial, investigators assess causality without knowing assignment

Proposed Addition to Skill

Add to scm_spec_template.md §6.3 a required specification for AEREL:

§6.3.5 Causality Assessment (AEREL)

AEREL is an INVESTIGATOR ASSESSMENT, not a ground truth. In blinded trials:
- AEREL is conditionally independent of true arm given AE type and temporal pattern
- Model as: P(RELATED | AE_type, temporal_onset) with type-specific priors:
    - Immune-mediated AEs (thyroid, colitis, hepatitis): P(RELATED) ~ 0.80 (both arms)
    - General AEs with known drug signal (fatigue, rash): P(RELATED) ~ 0.70 (both arms)
    - General AEs with high background (arthralgia, headache): P(RELATED) ~ 0.50
- Slight arm modulation allowed ONLY in open-label trials:
    P(RELATED | active_arm) = base + 0.10
    P(RELATED | control_arm) = base - 0.05

This ensures TRAE ⊂ Any AE (proper subset) and creates realistic separation between
categories in Table S3/S5 outputs.