Modeling Covariate-Adjusted Survival for Economic Evaluations in Oncology.

Abstract

In economic evaluations in oncology, adjusted survival should be generated if imbalances in prognostic/predictive factors across treatment arms are present. To date, no formal guidance has been developed regarding how such adjustments should be made. We compared various covariate-adjusted survival modeling approaches, as applied to the ENDEAVOR trial in multiple myeloma that assessed carfilzomib plus dexamethasone (Cd) versus bortezomib plus dexamethasone (Vd). Overall survival (OS) data and baseline characteristics were used for a subgroup (bortezomib-naïve/one prior therapy). Four adjusted survival modeling approaches were compared: propensity score weighting followed by fitting a Weibull model to the two arms of the balanced data (weighted data approach); fitting a multiple Weibull regression model including prognostic/predictive covariates to the two arms to predict survival using the mean value of each covariate and using the average of patient-specific survival predictions; and applying an adjusted hazard ratio (HR) derived from a Cox proportional hazard model to the baseline risk estimated for Vd. The mean OS estimated by the weighted data approach was 6.85years (95% confidence interval [CI] 4.62-10.70) for Cd, 4.68years (95% CI 3.46-6.74) for Vd, and 2.17years (95% CI 0.18-5.06) for the difference. Although other approaches estimated similar differences, using the mean value of covariates appeared to yield skewed survival estimates (mean OS was 7.65years for Cd and 5.40years for Vd), using the average of individual predictions had limited external validity (implausible long-term OS predictions with > 10% of the Vd population alive after 30years), and using the adjusted HR approach overestimated uncertainty (difference in mean OS was 2.03, 95% CI - 0.17 to 6.19). Adjusted survival modeling based on weighted or matched data approaches provides a flexible and robust method to correct for covariate imbalances in economic evaluations. The conclusions of our study may be generalizable to other settings. ClinicalTrials.gov identifier NCT01568866 (ENDEAVOR trial).