AN OUTCOMES REGRESSION APPROACH FOR INDIRECT COMPARISONS OF SURVIVAL OUTCOMES WHEN STANDARD NETWORK META-ANALYSIS IS NOT FEASIBLE

OBJECTIVES: In many cases the relevant evidence base of competing interventions cannot be reflected with one connected evidence network of randomized controlled trials (RCTs) to perform standard network meta-analysis. We present an outcomes regression method to perform indirect comparisons regarding time-to-event outcomes when the evidence bases consist of disconnected RCTs and/or single-arm trials. METHODS: This method requires access to individual patient data for at least one index interventions from which a set of bootstrap samples are obtained with replacement. For each bootstrap sample multiple competing multivariable survival models are estimated that describe the log-hazard over time as a function of prognostic factors and effect-modifiers. Their predictive performance is assessed based on the "out-of-bag" samples. Next, for each trial for which only summary data is available and not connected to other trials, a large number of hypothetical individuals are simulated based on the reported marginal distributions of the covariates of interest and their assumed correlation. For each of these populations the average log-hazards over time with the index intervention is predicted based on the model with best predictive performance for each of the bootstraps. Their summary distribution by trial effectively represents the outcome with an index intervention-based control group for each of these trials, which in turn allows all trials of relevance to be connected and to proceed with between-trial comparisons using standard network meta-analysis models. RESULTS: The method is illustrated with an indirect comparison of interventions for advanced melanoma. CONCLUSIONS: The proposed outcomes regression method uses cross-validated models, ensures that the prediction of outcomes and indirect comparisons are performed on the same (transformed) scale, and uncertainty associated with estimation of the outcome model parameters is propagated through the indirect comparison. It can be considered a useful addition to previously proposed methods for indirect comparisons in the presence of disconnected networks.