OBJECTIVES:

Randomized controlled trials (RCTs) and real-world evidence (RWE) are often synthesized separately in health technology assessment (HTA). One reason is that RCTs and RWE show great heterogeneity in methodology and risk of bias which make merging the two data sources technically difficult. To address this problem, Bayesian Network Meta-regression (BNMR) models have been applied for evidence synthesis in the HTA setting. However, BNMR models vary significantly in algorithms, including naïve pooling (NP), inclusion in the form of prior information (PI), and three-level hierarchical modelling (HM), and information on their performance are still lacking. Hence, we aimed to estimate and compare the performance of BNMR models in a case study of Myelodysplastic Syndromes (MDS).

METHODS:

A Meta-analysis was conducted in MDS patients receiving conditioning regimens of allogeneic hematopoietic stem cell transplantation. The patients were treated with either reduced intensity conditioning (RIC) or myeloablative conditioning (MAC). The primary outcome was overall survival. We collected all data from published systematic reviews, and applied four BNMR models (i.e. one NP, one PI, and two HM models developed by Dias2010 or Verde2021) to synthesize evidence using the R package “Crossnma”. We compared point estimates of the BNMR models in forest plots with 95% confidence intervals (CIs).

RESULTS:

Estimates gained from the NP and HM (Dias2010) model were similar (i.e. -0.01 [-0.14, 0.14] and -0.01 [-0.15, 0.15]), but they differed significantly from the estimate gained from the PI and HM (Verde2021) model (i.e. -0.06 [-0.2, 0.06] and 0.13 [-0.12, 0.47]). Also, all these estimates gained from both RCTs and RWE differed significantly from the estimate obtained from only RCTs (0.1 [-0.2, 0.43]).

CONCLUSIONS:

Estimates obtained from the BNMR models are sensitive to model algorithms. Further research is needed to confirm our findings by validating these algorithms in other case studies.