BEYOND SINGLE-MODEL SURROGACY: A MULTI-MODEL EVALUATION OF PROGRESSION-FREE SURVIVAL AS A TRIAL-LEVEL SURROGATE FOR OVERALL SURVIVAL IN RELAPSED/REFRACTORY DIFFUSE LARGE B-CELL LYMPHOMA
Author(s)
Sneha Rai, MSc Biostats1, Mohd Kashif Siddiqui, MBA, MPH, PharmD2, Jatin Gupta, MBA, MPharm1.
1EBM Health Consultants, Delhi, India, 2EBM Health, London, United Kingdom.
1EBM Health Consultants, Delhi, India, 2EBM Health, London, United Kingdom.
OBJECTIVES: Progression-free survival (PFS) showed moderate-to-strong individual-level association with overall survival (OS) in relapsed/refractory diffuse large B-cell lymphoma (R/R DLBCL). However, individual-level association alone does not establish whether treatment effects (TE) on PFS reliably predict TE on OS. This study aimed at evaluating trial-level surrogacy between TE on PFS and OS in R/R DLBCL, using a multi-model surrogate endpoint framework.
METHODS: Trial-level log hazard ratios and standard errors for PFS and OS were extracted from 18 R/R DLBCL trials. Six models were fitted: weighted linear regression (WLR), fixed-effect meta-regression, Bayesian random-effects meta-regression, Daniels and Hughes Bayesian bivariate meta-analysis, Bayesian bivariate random-effects meta-analysis using product-normal formulation (BRMA PNF), and BRMA PNF with t-distribution. Trial-level association, surrogate threshold effect (STE), and leave-one-out cross-validation were assessed. Predictive performance was evaluated using coverage of observed OS effects within 95% prediction intervals and absolute prediction error on the log hazard ratio scale.
RESULTS: PFS and OS treatment effects showed moderate-to-strong trial-level association, with WLR weighted Pearson correlation of 0.788 (95%CI: 0.381-0.939). STE estimates varied across models, ranging from 0.612 to 0.878. Fixed-effect meta-regression produced the most conservative STE (0.878), whereas BRMA PNF with t-distribution produced the least conservative estimate (0.612). WLR and BRMA PNF produced similar STEs (0.673 and 0.662). Cross-validation showed that WLR had highest median predictive coverage (78.1%; range: 33.1%-98.7%). Among Bayesian models, BRMA PNF had the highest median coverage (69.9%; range: 13.9%-99.9%). Median absolute prediction errors were broadly comparable across models (0.080-0.119), indicating similar point prediction accuracy but differing uncertainty quantification.
CONCLUSIONS: PFS showed supportive, but not definitive, trial-level surrogacy for OS in R/R DLBCL. The association was moderate-to-strong, but STE estimates and predictive uncertainty were model-dependent. These findings support PFS as a candidate surrogate endpoint, while indicating that certainty of OS prediction remains sensitive to modelling assumptions and between-trial heterogeneity.
METHODS: Trial-level log hazard ratios and standard errors for PFS and OS were extracted from 18 R/R DLBCL trials. Six models were fitted: weighted linear regression (WLR), fixed-effect meta-regression, Bayesian random-effects meta-regression, Daniels and Hughes Bayesian bivariate meta-analysis, Bayesian bivariate random-effects meta-analysis using product-normal formulation (BRMA PNF), and BRMA PNF with t-distribution. Trial-level association, surrogate threshold effect (STE), and leave-one-out cross-validation were assessed. Predictive performance was evaluated using coverage of observed OS effects within 95% prediction intervals and absolute prediction error on the log hazard ratio scale.
RESULTS: PFS and OS treatment effects showed moderate-to-strong trial-level association, with WLR weighted Pearson correlation of 0.788 (95%CI: 0.381-0.939). STE estimates varied across models, ranging from 0.612 to 0.878. Fixed-effect meta-regression produced the most conservative STE (0.878), whereas BRMA PNF with t-distribution produced the least conservative estimate (0.612). WLR and BRMA PNF produced similar STEs (0.673 and 0.662). Cross-validation showed that WLR had highest median predictive coverage (78.1%; range: 33.1%-98.7%). Among Bayesian models, BRMA PNF had the highest median coverage (69.9%; range: 13.9%-99.9%). Median absolute prediction errors were broadly comparable across models (0.080-0.119), indicating similar point prediction accuracy but differing uncertainty quantification.
CONCLUSIONS: PFS showed supportive, but not definitive, trial-level surrogacy for OS in R/R DLBCL. The association was moderate-to-strong, but STE estimates and predictive uncertainty were model-dependent. These findings support PFS as a candidate surrogate endpoint, while indicating that certainty of OS prediction remains sensitive to modelling assumptions and between-trial heterogeneity.
Conference/Value in Health Info
2026-11, ISPOR Europe 2026, Vienna, Austria
Value in Health, Volume 29, Issue 12S
Code
SA96
Topic
Health Technology Assessment, Methodological & Statistical Research, Study Approaches
Topic Subcategory
Literature Review & Synthesis, Meta-Analysis & Indirect Comparisons
Disease
Oncology