Tislelizumab first-line (1L) gastric/gastroesophageal junction cancer (G/GEJ) treatment efficacy on patient-reported outcome (PRO)-based symptom endpoints adjusting for informative missing data bias: Results from RATIONALE 305.

2605 Background: PRO-based symptom endpoints are rarely associated with treatment efficacy in oncology trials. PRO analyses emphasize early treatment cycles to avoid bias arising from study dropout (SDO; eg, due to death or other intercurrent events). SDO is thought to induce data missing not at random (MNAR), thereby biasing estimates and inference. However, separation between arms on other efficacy endpoints (eg, overall survival [OS]) typically occurs at cycles subsequent to the analysis period cutoff for PRO-based endpoints. This research examined treatment efficacy in PRO-based symptom endpoints in treatment cycles occurring later than typically analyzed. To protect against MNAR bias associated with later cycles, joint models (JMs) were employed. Methods: RATIONALE 305 was a randomized, double-blind, placebo-controlled, phase 3 trial comparing efficacy and safety of tislelizumab + chemotherapy (tislelizumab) with placebo + chemotherapy (chemo) as 1L treatment for patients with locally advanced, unresectable, or metastatic G/GEJ adenocarcinoma. EORTC-STO22 symptom domain scores were analyzed. Change from baseline (CFBL) in each domain was analyzed for up to 21 cycles between baseline and Cycle 38 (cycle with no less than 10% of baseline surviving). Treatment efficacy for STO22 endpoints was analyzed using JMs. The JM linked a linear mixed model (LMM) for STO22-domain CFBL to a Cox proportional hazards model for OS, adjusting LMM estimates for the potential MNAR bias arising from OS-based SDO. The LMM and Cox model were adjusted for study randomization factors of region (East Asia vs Rest of World), PD-L1 status (positive vs. negative), and peritoneal metastases (present vs. absent). All analyses were conducted using JMBayes2 package in R version 4.3.2. Results: STO22 pain results are reported; all domains will be presented. For tislelizumab (n=465) vs. chemo (n=467), subjects remaining at Cycle 38 were 46 (10%) and 28 (6%), respectively. Tislelizumab was associated with a constant lower level of pain compared with chemo (-2.08; P=0.0109). Tislelizumab was associated with a 21% reduction in risk of death ( P=0.0269). Conclusions: In this analysis, compared with chemo, tislelizumab + chemo was associated with lower pain and longer survival. Joint models for PRO-based symptom endpoints may illuminate patient-centric therapeutic benefits. [Table: see text]