OBJECTIVES: Analysis of treatment effects on surrogate endpoints may be useful where final endpoints are not recorded in trials (e.g. long-term cardiovascular endpoints) or where estimated treatment effects on final endpoints are uncertain (e.g. overall survival where the proportion of subjects experiencing an event is limited). The estimated relationship between final and surrogate endpoints may be used for the purposes of inference regarding the existence of treatment effects or estimation regarding the magnitude of treatment effects on final endpoints. In this simulation study we compare the performance of linear regression (Reg), Daniel and Hughes (DH) and Bivariate Random effect Meta-Analysis (BRMA) in estimating the relationship between a surrogate and decision-relevant (final) endpoints.

METHODS: Sets of simulated treatment effect data were estimated based on given values for the between and within study variance and between study correlation for final and surrogate endpoints. The regression co-efficients describing the relationship between treatment effects on surrogate and final endpoints were estimated using the Reg, DH and BRMA models. The Bias, variance and coverage were estimated for each estimator. The results were compared for scenarios where the within study variance for treatment effects regarding the final and surrogate endpoints were identical and where the within study variance was greater for the final endpoint.

RESULTS: The BRMA model was the best performing model where the within study variance was equal for treatments effect on the final and surrogate endpoints. However, the DH model was the best performing model, with lower Bias, compared with the BRMA model where the variance was greater for treatments effects on the final endpoint

CONCLUSIONS: Given that the variance will commonly be greater for the treatment effects on final endpoints, caution should be exercised with respect to the interpretation of the results from BRMA models.