Presentation (CTI)

OBJECTIVES: Differences in study populations can introduce substantial bias in naïve unanchored comparisons of disconnected evidence. Existing population-adjusted methods are limited to pairwise settings and can only transport estimates to the comparator population. To address these challenges, we introduce the mlumr R package, implementing multilevel unanchored meta-regression (ML-UMR).
METHODS: ML-UMR extends the multilevel network meta-regression (ML-NMR) framework to unanchored settings. An outcome model is fitted to the individual patient data and integrated over the aggregate-data covariate distribution to recover population-average estimands. The integration uses quasi-Monte Carlo with Sobol sequences and a Gaussian copula. Two specifications are provided: a shared prognostic-factor assumption (SPFA) model, in which covariate effects on the outcome are common across treatments, and a relaxed model that permits treatment-specific covariate effects. Robustness to unmeasured confounding is assessed with quantitative bias analysis (E-values and tipping-point analyses). The package is implemented in R, with Bayesian estimation performed in Stan via rstan/cmdstanr. Binary, continuous, count, and time-to-event (parametric and flexible M-spline baseline hazards) outcomes are supported.
RESULTS: The workflow involves five steps: declare the individual patient data (set_ipd) and the published aggregate data (set_agd, or set_agd_surv), combine them (combine_data), construct the integration grid (add_integration), and fit the model (mlumr). marginal_effects estimates marginal treatment effects in any decision-relevant population, and conditional_effects returns covariate-conditional effects. The same fit transports effects to any target population and accommodates any number of treatments, including disconnected networks of single-arm studies. Built-in unadjusted (naive) and G-computation (stc) methods provide benchmark comparators. calculate_dic, calculate_loo, and calculate_waic support model selection. Dedicated functions quantify the sensitivity of unanchored estimates to unmeasured confounding.
CONCLUSIONS: mlumr is an R implementation of ML-UMR that transports treatment effects to any target population, supports all common outcome types, accommodates multiple treatments and studies, and provides bias analysis for unmeasured confounding, supporting transparent and robust indirect comparisons for health technology assessment.