Cost-Effectiveness Uncertainty Analysis Methods: A Comparison of One-Way Sensitivity, Analysis of Covariance, and Expected Value of Partial Perfect Information

Objectives. To compare model input influence on incremental net monetary benefit (INMB) across 3 uncertainty methods: 1) 1-way sensitivity analysis; 2) probabilistic analysis of covariance (ANCOVA); and 3) expected value of partial perfect information (EVPPI). Methods. In a preliminary model, we used a published cost-effectiveness model and assumed 20,000 pound per quality-adjusted life-year (QALY) willingness-to-pay (Case 1: lower decision uncertainty) and 8000 pound/QALY willingness-to-pay (Case 2: higher decision uncertainty). We conducted 1-way sensitivity, ANCOVA (10,000 Monte Carlo draws), and EVPPI for each model input (1000 inner and 1000 outer draws). We ranked inputs based on influence of INMB and compared input ranks across methods within case using Spearman's rank correlation. We replicated this approach in 3 follow-up models: an additional linear model, a less linear model with uncorrelated inputs, and a less linear model with correlated inputs. Results. In the preliminary model, lower and higher decision uncertainty cases had the same top 3 influential parameters across uncertainty methods. The 2 most influential inputs contributed 78% and 49% of variation in outcome based on ANCOVA for lower decision uncertainty and higher decision uncertainty cases, respectively. In the follow-up models, input rank order correlations were higher across uncertainty methods in the linear model compared with both of the less linear models. Conclusions. Evidence across models suggests influential input rank agreement between 1-way and more advanced uncertainty analyses for relatively linear models with uncorrelated parameters but less agreement for less linear models. Although each method provides unique information, the additional resources needed to generate and communicate advanced analyses should be weighed, especially when outcome decision uncertainty is low. For less linear models or those with correlated inputs, performing and reporting deterministic and probabilistic uncertainty analyses appear prudent and conservative.