Presentation (CTI)

OBJECTIVES: A virtual inconsistency-adjusted analysis (VIAA) has previously been proposed as a sensitivity framework for star-shaped network meta-analysis (NMA) to assess robustness of treatment rankings to potential inconsistency. The present study aimed to develop a one-stage likelihood-augmented implementation of VIAA that targets the same estimand as the original numerical procedure, and to illustrate its use.
METHODS: Design: Methodological development and illustration. In the original VIAA, a user-controlled inconsistency parameter was introduced to shift the mean of virtual head-to-head effects between non-reference treatments. Virtual contrasts were generated from a prespecified distribution to the observed star-shaped network, analyzed using Bayesian random-effects NMA models across multiple imputations, and combined to assess robustness based on which the original treatment ranking was preserved over the acceptable range of inconsistency values. In this study, we developed a likelihood-augmented VIAA that incorporated the same working distribution into a single extended NMA model through pseudo-likelihood terms for the virtual contrasts. Under normal distribution assumptions, we derived closed-form posterior summaries and demonstrated that the posterior mean of the basic treatment effects coincided with the large-imputation limit of the multiple-imputation VIAA estimator. The method was illustrated using a published star-shaped NMA comparing dipeptidyl peptidase-4 inhibitors (DPP4i), glucagon-like peptide-1 receptor agonists (GLP-1RA), and sodium-glucose co-transporter-2 inhibitors (SGLT2i) versus placebo in patients with cardiovascular disease, focusing on cardiovascular mortality.
RESULTS: In the example on cardiovascular mortality, SGLT2i ranked highest, followed by GLP-1RA and DPP4i. Treatment rankings were fully robust (100%) to virtual inconsistency up to a value of 1.4. The likelihood-augmented VIAA achieved these results using a single model fit with good MCMC convergence, matching the conclusions of the multiple-imputation approach with substantially reduced computation.
CONCLUSIONS: The likelihood-augmented VIAA provides a theoretically grounded and computationally efficient one-stage implementation of virtual inconsistency-based sensitivity analysis for star-shaped NMAs, facilitating routine robustness assessment of treatment rankings.