Close

PCV64 FACTOR SCREENING FOR PARSIMONY OF VARIABLES IN PHARMACOECONOMIC MODELS

Abstract

OBJECTIVES: This research applies factor screening techniques to a decision tree model of the first 30 days following an acute coronary event. The objective is to determine which input variables most affect the total cost and to evaluate how well a metamodel composed of fewer variables can replace the decision tree model. METHODS: Factor screening techniques are often used in defense simulations to examine simultaneously many input variables to identify those with a significant effect on a selected response (output). The original model was built using TreeAge Pro Health care and parameterized to replicate 30-day results from the Clopidogrel in Unstable angina to prevent Recurrent ischemic Events (CURE) trial. The model had 44 input variables (28 relatively independent factors) comparing 30-day results for clopidogrel plus aspirin to a hypothetical alternative. Ranges of uncertainty were assigned to the input variables. A 2 k Fractional Factorial of Resolution IV (FF) and a Folded-over Sequential Bifurcation (SB-X) method were used to identify important factors. A metamodel was generated using a Central Composite Design and dropping the effects with F-ratio less than 2.0. RESULTS: Both the FF and SB-X methods identified the same seven factors (five costs and two probabilities). In addition to the main effects, the metamodel had only two second-order terms, and no quadratic effects. When a Nearly Orthogonal LatinHypercube Design with 256 observations was used to sample the entire experimental region (including all 28 factors), the reduced metamodel could account for more than 99% of the variation observed over the entire experimental region. (Error: 0.702%) CONCLUSION: Results from the 44-variable decision-tree model were matched by a nine-term metamodel. When factor screening can reduce the number inputs that must be specified, the data needed to predict responses can be greatly reduced. Further work investigates factor screening and developing corresponding metamodels for stochastic models.

MIDAS Network Members

Citation: