PMS31 - Quantifying the Importance of Model Parameters and Structural Assumptions on the Value of Treatments for Rheumatoid Arthritis Using Metamodeling

Abstract

Our aim was to quantify the relative importance of model parameters and model structures on the cost-effectiveness estimates of biologic disease modifying anti-rheumatic drugs (bDMARDs) vs conventional DMARDs (cDMARDs) for treatment of patients with moderate to severe rheumatoid arthritis with inadequate response to cDMARDs. We used the Innovation and Value Initiative’s open-source rheumatoid arthritis individual patient simulation to simulate outcomes. 1,000 sets of parameter values were sampled in a probabilistic sensitivity analysis (PSA) and 32 model structures were simulated. For each parameter set and model structure, 1,000 patients were simulated to calculate the mean incremental net monetary benefit (iNMB). The simulated mean iNMB was regressed on values of the model parameters and characteristics of the model structures using linear regression models. The absolute values of the coefficients of standardized variables were used to rank the model inputs by their relative importance. Structural assumptions related to treatment switching, long-term progression of the Health Assessment Questionnaire (HAQ) score, utility, and the effect of treatment on HAQ had large impacts on the iNMB. For example, models that use a latent class growth model to simulate HAQ progression for patients using cDMARDs were predicted to increase the iNMB by around $45,000 relative to models that assumed a linear rate of progression. Standardized input parameters with the largest coefficients in absolute value included the extent to which the HAQ score rebounded after treatment failure and the impact of changes in the HAQ score on mortality. Cost-effectiveness estimates for rheumatoid arthritis vary due to both parameter and structural uncertainty. New studies are needed to improve the quality of parameter estimates; consensus-driven approaches such as Delphi panels can help determine appropriate model structures. Research should prioritize the most sensitive model parameters and debates should focus on the structural assumptions most likely to influence results.