MS3 NEW METHODOLOGIES IN PARAMETRIC NETWORK META-ANALYSES (PNMA): MULTI-ARM CORRECTION FOR RANDOM EFFECTS AND GENERAL POPULATION MORTALITY (GPM)

Abstract

The PNMA was introduced by Ouwens et al in 2010, and according to the NICE DSU 14 it should be considered as an alternative to the hazard ratio network meta-analysis (NMA). No published PNMA considered GPM and/or multi-arm trial correction for random effects (RE). The objective was to present a PNMA that considers GPM in the likelihood function of the parametric distributions, and a RE PNMA with multi-arm correction. Four trials in newly-diagnosed multiple myeloma were considered for the analysis. One trial comparing melphalan-prednisone-thalidomide (MPT), lenalidomide-dexamethasone 18 cycles (Rd18) and Rd-continuous, and three trials comparing melphalan-prednisone (MP) and MPT. Three types of PNMAs were performed 1) fixed effects (FE), 2) RE and 3) FE + GPM. The following distributions were tested; Exponential, Gompertz, Loglogistic, Lognormal, and Weibull. PNMA fit was assessed by Watanabe–Akaike information criterion (WAIC). The best fitting distribution was the Weibull with a WAIC for FE 14905, FE + GPM 14874 and for RE 14909. For the three best fitting distributions the mean survival in months by arm in the FE PNMA for Weibull was MPT=27.3, MP=17.8, Rd-continuous=44.3 and Rd18=27.6. Corresponding results for Loglogistic are MPT=35.8, MP=23.2, Rd-continuous=63.1 and Rd18=33.4, and for Gompertz MPT=27.3, MP=17.1, Rd-continuous=87.0 and Rd18=33.4. Considering GPM resulted in the following mean survival Weibull (MPT=27.3, MP=17.7, Rd-continuous=42.9, and Rd18=27.6), Loglogistic (MPT=31.3, MP=21.0, Rd-continuous=42.3, and Rd18=30.0) and Gompertz (MPT=27.3, MP=16.6, Rd-continuous=49.5, and Rd18=27.8). This research presents the first multi-arm RE PNMA and first PNMA considering GPM. The RE model increased the uncertainty around the treatment effect of shape and scale parameters, including GPM seems to improve the fit for all distributions, and the GPM reduced the variation over predicted mean survival over the tested distributions (structural uncertainty).