05335v3.qxd

Using Linear Regressions to Approximate Results of Decision Analysis: An Application
to a Cost Comparison Across Three First-Line Drug Strategies in Type 2 Diabetes
Botteman MF, Gao X, Stephens JM
HERQuLES, Abt Associates Clinical Trials, Bethesda, MD, USA
„ A literature-based decision tree model was Clinical and Economic Inputs
Introduction
developed to project the number of patients Conclusions
„ The long-term, aggressive treatment of diabetes therapies and the associated costs over a „ Results from the baseline analysis and the „ Both methods demonstrate similar results that is cost effective but often requires the use of three-year timeframe (Figure 1).
SOMCS are presented in Table 1.
a sulfonyluea strategy with glipizide GITS is the Primary failures after first-line therapy leading multiple therapies to achieve American Diabetes least expensive first-line therapy for newly „ A hypothetical cohort of 1,000 patients was „ Regression models fitted from SOMCS results diagnosed type 2 diabetes patients, followed by Discontinuation due to adverse events leading assigned to begin one of the three first-line are presented in Table 2.
„ Few studies have addressed the clinical and strategies. For each first-line treatment cohort, „ Comparison of Results of SOMCS Model and economic impact of initiating patients on specific z Proportion of patients controlled with addition the model generated the cumulative cost per of a second or third agent if the patient was treated patient at the end of three years.
z An additional MCS was run and the inputs „ The analytical perspective was that of a payer, z Secondary failures over time leading to were used in both the SOMCS and LR models.
incorporating only direct medical costs. The cycle The LR can be used as a quick tool for use in useful tool to address these impacts. However, z Both methods led to identical conclusions length in the model was three months to reflect „ Medical resource use(1) and associated costs these techniques are complex and often not the frequency of physician office visits.(1) expensive in >97% of cases. The accordance decision tree, and providing decision support understood by decision makers. Simplified Long-term diabetic complications were excluded z Office visits, laboratory tests, patient between projected costs across methods was methods that allow end users to easily use and due to the short-term nature of the model.
statistically significant (Kappa>0.80, p<0.001) customize these decision models are very much in each head-to-head comparison, confirming the feasibility of using the LR to approximate References
the results of the decision analysis.
Objectives
z Treatment of clinical failures and adverse 1. Standards of medical care for patients with Figure 1. Model Pathways and Outcomes
Table 1. Cost/Net Cost of Three-Year Treatment Per Patient*
diabetes mellitus. Diabetes Care 2001; 24:1-24.
„ To project and compare via a previously „ Costs assigned for resources were 2001-2002 Drug Comparison
Base Case Model
F irst-L in e T h erapy
2. Ramsdell et al. Economic model of first-line In itial M on oth erap y
glyburide/m etform in, rosiglitazone, or repaglinide three-year direct medical costs associated with average wholesale prices. All input data are glycemic control in newly diagnosed type 2 three common first-line oral antidiabetic classes diabetes. Pharmacoeconomics 2003; in press.
Second-Order Monte-Carlo Simulation and
Sulfonylurea: glipizide gastrointestinal therapeutic system (Glucotrol XL - GXL) Regression Analysis
„ Triangular distributions were used to describe the z Biguanide: generic metformin (MET), and The “Net Cost” of Drug 1 vs. Drug 2 is defined as Drug 1 cost minus Drug 2 cost. variability of the model input in the SOMCS.
Table 2. Regression Results for Three-Year Treatment Cost Per Patient
„ To summarize the model in an LR form, the costs (dependent variables) estimated via 1,000 MCS Regression 1
Regression 2
Regression 3
DV = GXL Cost
DV = METIR Cost
DV = ROS Cost
„ To evaluate the feasibility and usefulness of runs were summarized through ordinary least- Variables
Variables
Variables
summarizing the decision analytic model’s square regressions, using the most sensitive and/or relevant variables from the decision model as predictors (identified via Tornado diagrams).
„ Specifically, each MCS run was treated as a ran- dom observation. Three regression models were generated to predict three-year treatment cost of Overview
„ The target treatment population consisted of -1328.60* GXL monotherapy drug cost as switch „ The results generated via each method were patients newly diagnosed with type 2 diabetes compared. Multiple tests were conducted to determine the predictive abilities of the regres- 1. Unless specified, all p < 0.001. *p > 0.05.
sion against the results of the traditional model.
2. Regression Model 1: R2 =0.53, F = 111.20, p <0.001; Regression Model 2: R2 =0.49, F = 95.29, p < 0.001; Regression Model 3: R2 = 0.49, F = 96.20, p < 0.001.
Poster prepared at www.SciFor.com

Source: http://www.abtassociates.com/presentations/ISPOR2003-linear.pdf