Assessing Parameter Importance in Decision Models. Application to Health Economic Evaluations

Abstract

Background: Uncertainty in parameters is present in many risk assessment and decision making problems and leads to uncertainty in model predictions. Therefore an analysis of the degree of uncertainty around the model inputs is often needed. Importance analysis involves use of quantitative methods aiming at identifying the contribution of uncertain input model parameters to output uncertainty. Expected value of partial perfect information (EVPPI) measure is a current gold- standard technique for measuring parameters importance in health economics models. The current standard approach of estimating EVPPI through performing double Monte Carlo simulation (MCS) can be associated with a long run time. Objective: To investigate different importance analysis techniques with an aim to find alternative technique with shorter run time that will identify parameters with greatest contribution to uncertainty in model output. Methods: A health economics model was updated and served as a tool to implement various importance analysis techniques. Twelve alternative techniques were applied: rank correlation analysis, contribution to variance analysis, mutual information analysis, dominance analysis, regression analysis, analysis of elasticity, ANCOVA, maximum separation distances analysis, sequential bifurcation, double MCS EVPPI,EVPPI-quadrature and EVPPI- single method. Results: Among all these techniques, the dominance measure resulted with the closest correlated calibrated scores when compared with EVPPI calibrated scores. Performing a dominance analysis as a screening method to identify subgroup of parameters as candidates for being most important parameters and subsequently only performing EVPPI analysis on the selected parameters will reduce the overall run time.