PRM50 - A Frequentist Framework for Automated Generation of Node-Splitting Models for Assessment of Inconsistency in Network Meta-Analysis

Abstract

In conducting network meta-analysis to compare treatments across an array of clinical trials, potential inconsistencies between estimates of relative treatment effects are of importance. A node-splitting approach to detect inconsistency, while potentially labor intensive, is attractive because of its simplicity in interpretation, contrasting estimates from both direct and indirect evidence. An automated method for estimating node-splitting models was introduced by van Valkenhoef et al. (2014) using a Bayesian framework. Here, we introduce a frequentist framework for automated generation of node-splitting models to assess inconsistency in network meta-analysis. We assess our frequentist approach across two data sets and compare the results to the Bayesian approach. We developed an algorithm for automated generation of node-splitting models to assess inconsistency using a frequentist framework, implemented in SAS. First, we checked the decision rule to determine whether to split a specific comparison, based on properties of the evidence structure that are easily verified. Second, we used Parkinson’s disease data (Franchini et al., 2012) for normal data and Smoking data (Hasselblad, V., 1998) for binary data to assess the validity of our frequentist method vs. the Bayesian method of van Valeknhoef et al. For Parkinson’s disease data (4 treatments: A, B, C, D), the decision rule from our method detected the AC, AD, BD and CD comparisons as potential inconsistencies. We find that the method of van Valkenhoef et al. also selects the same treatment comparisons. Direct evidence and indirect evidence between our method and the Bayesian method produces similar results. For smoking data, both methods selected the same treatment comparisons and also showed similar results between direct evidence and indirect evidence. Based on the results from the Parkinson’s disease and smoking data sets, our proposed automated generation method using a frequentist framework can be used for detecting potential inconsistencies in network meta-analysis.