OBJECTIVES: In trials, surrogate endpoints are often used instead of final clinical outcomes to more quickly obtain treatment effects, which are used to support regulatory and reimbursement decisions. Before use in decision-making, surrogate endpoints require validation to ensure treatment effects observed on surrogate endpoints can accurately predict treatment effects on final outcomes. Traditionally, validation of surrogate endpoints has only used RCT data. However, RCT data are often too limited for validation. The objective of this study is to investigate approaches for including comparative real world evidence (cRWE) and single-arm RWE (sRWE) alongside RCT data in surrogate endpoint evaluation, accounting for potential differences in treatment effects between sources of data, to improve surrogate endpoint validation and prediction of treatment effects on final outcomes.

METHODS: Bayesian bivariate random-effects meta-analysis (BRMA) is used to evaluate progression-free survival (PFS) as a surrogate endpoint for overall survival (OS) in colorectal cancer. The model is extended to include bias terms to account for differences in treatment effects between RCTs and RWE. Cross-validation is used to assess whether including RWE improves prediction of treatment effects on OS.

RESULTS: Using RCT data alone (7 studies) estimated between-studies correlation of 0.75 (95% CrI: -0.25, 0.99). Adding cRWE and sRWE (7 studies) more precisely estimated correlation at 0.74 (95% CrI: -0.056, 0.98). Inclusion of RWE reduced median absolute discrepancy of predictions of treatment effects on OS from 0.16 to 0.12. Bias adjustment resulted in weaker correlation of 0.71 (95% CrI: -0.25, 0.98) and less accurate predictions of treatment effects on OS with median absolute discrepancy of 0.18.

CONCLUSIONS: Adding RWE to RCTs in BRMA without bias adjustment improves precision of surrogacy parameters and prediction of treatment effects on the final outcome. However, using bias adjustment lowered the correlation, providing weaker evidence for surrogacy. This could be because bias adjustment reduces between-studies heterogeneity.