Analysis Data Reviewer's Guide (ADRG) — GLPI103-301
| Field |
Value |
| Document ID |
DEF-301-ADRG |
| Version |
1.0 |
| Status |
Final (portfolio) |
| Study No. |
GLPI103-301 |
| CDISC STUDYID |
GLPI103 |
| Standards |
ADaMIG (v1.x) · Define-XML 2.1 |
| Confidentiality |
Confidential — portfolio use |
Companion to define.xml and SAP-301. Describes the analysis (ADaM) datasets and key derivations supporting the primary and secondary analyses.
Change History
| Version |
Date |
Author |
Summary |
| 1.0 |
2026-06-29 |
Biostatistics |
Initial version |
1. Introduction
This ADRG supports review of the analysis datasets used for the efficacy and safety analyses of Study GLPI103-301 (SAP-301). Derivations below correspond to src/cdisc/adam_builder.py.
2. Analysis Datasets
| Dataset |
Class |
Records |
Structure |
Key use |
| ADSL |
ADSL |
900 |
One record per subject |
Population flags, treatment, strata, covariates |
| ADLB |
BDS |
27,795 |
One per parameter/visit/subject |
HbA1c CFB analyses (ANCOVA, MMRM, estimands) |
3. ADSL — Key Variables & Derivations
| Variable |
Derivation |
TRT01P / TRT01A |
Planned / actual arm (no switching) → TRT01PN/TRT01AN (1=IV, 2=Oral, 3=Sema) |
AGEGR1 |
<65 vs >=65 from AGE |
FASFL / SAFFL |
"Y" — all randomized / all dosed |
PPSFL |
"Y"/"N" — derived from major protocol deviations (DEVFL); PPS = 847 (≠ FAS 900) |
COMPLFL / DCSREAS / DISCWK |
Completion flag, discontinuation reason and study week (disposition) |
RESCUEFL / RESCUEWK / DTHFL |
Rescue-medication intercurrent-event flag and week; death flag |
STRAT_HBA1C |
Randomization stratum, carried from the randomization record (IRT) by USUBJID, based on the screening HbA1c (<8.5 vs ≥8.5). It is highly concordant with the Day-1 baseline-derived category (~91%); minor discordance near the 8.5% threshold is expected because stratification (screening) and the analysis baseline (randomization visit) are separate measurements. |
STRAT_BMI |
Stratum (<30 vs ≥30) from baseline BMI |
STRAT_REGION / REGION |
Region group (Capital/Yeongnam/Honam/Chungcheong/Gangwon-Jeju) from province |
EDUCATION / MARITAL / EMPLOYMENT |
Baseline covariates (English CT); SDTM source = SC domain |
Derivation integrity note: an earlier build regenerated STRAT_* randomly; this was corrected so strata are carried from the randomization source and are consistent with the analyzed baseline values. This is the variable the primary ANCOVA adjusts for.
4. ADLB — Key Variables & Derivations
| Variable |
Derivation |
PARAM/PARAMCD |
Parameter (e.g., HBA1C) |
AVAL |
Analysis value (from LBSTRESN) |
BASE |
Baseline value (ABLFL="Y", VISITNUM=1) per subject/parameter |
CHG / PCHG |
Change / % change from baseline (AVAL − BASE) |
ABLFL |
Baseline record flag (VISITNUM=1) |
ANL01FL |
Analysis flag (non-baseline visits) |
5. Analyses Supported
- Primary ANCOVA (Week 52,
ANL01FL="Y", VISITNUM=8, FAS): CHG ~ TRT01A + BASE + STRAT_HBA1C + STRAT_BMI (SAP-301 §5; outputs/stats/ancova_results.csv).
- MMRM + estimands (
mmrm package, unstructured covariance, Kenward-Roger DF): treatment-policy vs hypothetical estimands (SAP-301 §6; outputs/stats/estimand_results.csv).
- Missing-data sensitivity (
rbmi): MAR / jump-to-reference / copy-reference MI + tipping-point (outputs/stats/missing_data_results.csv).
- Multiplicity fixed-sequence gatekeeping; responders (logistic + CMH); subgroups + interaction (
outputs/stats/{multiplicity,responder,subgroup}_results.csv).
- Time-to-event (KM + Cox) and recurrent-event hypoglycaemia (
outputs/stats/{survival,recurrent}_results.csv).
- Safety from SDTM AE (subject-level incidence by SOC/PT; nausea proportion test) —
outputs/stats/ae_prop_test.csv.
- Results are reconciled to CSR-301 by
tools/reconcile_csr.py (passes: 23/23 key numbers, incl. secondary endpoints).
6. Conformance & Tooling
define.xml (Define-XML 2.1) covers ADSL and ADLB. An automated subset validator (tools/validate_cdisc.py, report DEF-301_validation_report.md) checks ADaM required variables, analysis-set flags (Y/N), ADSL one-record-per-subject, ADLB↔ADSL referential integrity, and CHG = AVAL − BASE — 0 Errors, 0 Warnings. Official Pinnacle 21 Community validation is a documented manual step (registration-gated GUI tool; see SDRG §6).
7. Known Issues / Limitations (simulated dataset)
- PPS (847) ≠ FAS (900) — major protocol deviations define PPS exclusions.
- The AE domain includes serious events (
AESER="Y", 17 events) and 2 deaths.
- Region is a randomization stratum and Table-1 covariate but is not in the primary ANCOVA model (assessed descriptively/subgroup).