FACTOR SCREENING TO IDENTIFY FACTORS WHICH INFLUENCE STRATEGY SELECTION
Reza Yaesoubi, BS1, Stephen D. Roberts, PhD1, Robert W. Klein, MS2, Naga P. Chalasani, MD3, and Thomas F. Imperiale, MD3. (1) North Carolina State University, Raleigh, NC, (2) Medical Decision Modeling, Inc., Indianapolis, IN, (3) Indiana University School of Medicine, Indianapolis, IN
Purpose: To evaluate the use of sequential bifurcation as a factor screening technique using a Markov model of screening for esophageal varices in patients with cirrhosis. Background: One- and two-way sensitivity analyses can be misleading when there are many correlated factors. Probabilistic sensitivity analysis provides information on variability but not the factors that drive it. Factor screening techniques, including sequential bifurcation, have been used to identify key drivers in military and manufacturing simulations but have not been applied to healthcare. Additionally, identifying factors that are unimportant can direct studies to more important factors. Methods: Using TreeAge, a Markov model with many uncertain inputs found large variation in the cost effectiveness of the choice between endoscopic screening of all cirrhotic patients and screening only those determined to be high risk using a clinical decision aid. However, using traditional one-way sensitivity analysis no single variable greatly influenced the incremental cost-effectiveness ratio. Therefore, ranges were specified for all uncertain input variables, QALYs and cost/QALY were selected as the response functions, and sequential bifurcation was performed for each strategy at different delta levels of response. Several approaches to identify factors in the competing strategies that can change such that the competing strategy may become preferred to the base strategy were tried. Metamodels (R1 and R2), predicting the response for each strategy (S1 and S2), were built from the lists of important variables. Factors that maximize change in metamodels of the absolute value of the difference between strategies (Rd=R1-R2) were found. Results: Using traditional 3-way sensitivity analysis combinations of reduced prevalence of large varices and better sensitivity or specificity of the clinical decision rule altered the preferred strategy. Although the Markov model had 39 variables, metamodels of 8 to 12 variables predict essentially the same responses (R2 > 0.98). The metamodels of differences also indicate that bleeding rate, progression rate from small to large varices, and the frequency of bleeds requiring TIPs are important variables. Conclusions: The number of factors considered important is very sensitive to the delta level chosen for the response function. Factor screening offers promise in prioritizing variables for which further study narrowing their uncertainty is needed to insure confidence in the choice of clinical strategies as well as targeting strategies to particular subpopulations.