Sponsoring Section/Society: ASA Health Policy Statistics
Session Slot: 10:30-12:20 Tuesday
Estimated Audience Size: 75-125
AudioVisual Request: None
Session Title: Statistical Tools for Studying Health Care Costs: What
Works and Where Do We Need Improvements?
Hillary Clinton, CEO's, drug companies, government officials and patients share concerns over health care costs. This session focuses on several aspects of the problem. Breast cancer screening has been shown to reduce mortality rates in some age groups of women. The first talk considers the cost-effectivenss of targeting certain risk groups for this screening. This talk is followed by one which discusses pitfalls one often encounters when using quality of life years and discounting of later life years in a cost-effective analysis. Examples will be discussed in which the study results (and conclusions) were highly dependent upon the particular weighting method used. The third talk investigates statistical modeling of health care costs. The discussion is applied to the problem of casemix adjustment when setting payments to health plans for their patients. The talks are discussed by Sue Rosenkranz from Harvard Pilgrim Health Care, a large health maintenance organization in the Northeast.
Theme Session: Yes
Applied Session: Yes
Session Organizer: Christiansen, Cindy L. Harvard Medical School and Harvard Pilgrim Health Care
Address: Harvard Medical School and Harvard Pilgrim Health Care, 126 Brookline Ave., Suite 200, Boston, MA 02215
Session Timing: 110 minutes total (Sorry about format):
3 speakers and 1 discussant Opening Remarks by Chair - 3 minutes First Speaker - 30 minutes Second Speaker - 30 minutes Third Speaker - 30 minutes Discussant - 10 minutes Floor Discusion - 7 minutes
Session Chair: Hilbe, Joe
1. Breast Cancer Risk and Cost-Effectiveness of Screening Mammography
McGuigan, KA, UCLA School of Public Health, Dept Health Services
Address: RAND, 1700 Main St., PO Box 2138 Santa Monica, CA 90407-2138
Phone: 310-393-0411 x7698
Abstract: The cost-effectiveness of a screening program can be facilitated by targeting high-risk populations, as the performance characteristics (positive predictive value) of a screener improve as a function of increased prevalence of disease in a population. Using data from the Breast Cancer Detection Demonstration Project (BCDDP), parametric (logistic regression) and nonparametric (recursive partitioning, artificial neural networks) models are compared to determine which methods better discriminate cancer cases from controls based on known risk factors. Performance characteristics include model sensitivity, specificity, positive predictive value, and area under the receiver operating characteristic curve (ROC). The best-performing model is used to distinguish low vs normal and high risk groups for alternative screening schedules. Discussion addresses whether current information about epidemiologic risk factors can be used to obtain improved yields with lower overall costs, subject to the constraint of minimizing missed cases (false negatives).
2. Dangers of Discounting and QALYs
Palmer, J. Lynn, M.D. Anderson Cancer Center
Address: Department of Biomathematics, 1515 Holcombe Blvd., Houston, TX 77030
Abstract: Many traditional methods of analysis of health care policy data use weighting methods which may make interpretation of their results difficult. Any analysis using quality of life years as an outcome variable should make explicit in its conclusions how weighting or discounting of early versus late life years was determined, and how its conclusions would have changed if slight adjustments in these weights would have been made. This presentation will include examples of how these difficulties can be (easily) encountered and how results can be clearly presented.
3. How Well Do Models Work? Predicting Health Care Costs
Ash, Arlene S., Boston University School of Medicine
Address: Boston University School of Medicine 720 Harrison Ave, #1108, Boston, MA 02118
Abstract: The goal is to pay health plans (such as HMOs) for the expected cost of care for the persons they enroll. Models based on age and sex can reliably distinguish population subgroups whose costs differ by as much as 10 to 1; however, R-squares for age-sex models rarely exceed 0.02. Health-based predictions, such as those produced by Diagnostic Cost Group (DCG) models, explain more of the variation, can detect subgroups whose costs differ by more than 100 to 1 and have larger R-square values, but still, generally less than 0.10. In the context of producing models to be used in setting payments to health plans for their enrollees, we address questions such as: how much better is one model than another? in what way does a model fail? how good is good enough?
Discussant: Rosenkranz, Susan L. Harvard Pilgrim Health Care
Address: Deputy Medical Director's Office, Harvard Pilgrim Health Care, BPW-2, 10 Brookline Place West, Brookline, MA 02146
List of speakers who are nonmembers: None