Sponsoring Section/Society: ASA-SBSS
Session Slot: 2:00- 3:50 Wednesday
Estimated Audience Size: 100
AudioVisual Request: none
Session Title: Bayesian Methods in Health-Care Policy
Theme Session: Yes
Applied Session: No
Session Organizer: Stangl, Dalene Duke University
Address: Dalene Stangl Box 90251, ISDS Duke University Durham, NC 27708-0251
Phone: 919-684-4263
Fax: 919-684-8594
Email: dalene@stat.duke.edu
Session Timing: 110 minutes total (Sorry about format):
Opening Remarks by Chair - 5 minutes First Speaker - 30 minutes Second Speaker - 30 minutes Third Speaker - 30 minutes Discussant - 10 minutes Floor Discusion - 5 minutes
Session Chair: Sargent, Dan Mayo Clinic
Address: Dan Sargent Cancer Center Statistics Mayo Clinic 200 1st St. SW Rochester, MN 55905
Phone: 507-284-5380
Fax: 507-284-1902
Email: Sargent.Daniel@mayo.edu
1. Stochastic Optimization Methods For Cost-Effective Quality Assessment in Health Care
Draper, David, University of Bath, UK
Address: David Draper Statistics Group School of Mathematical Sciences University of BAth Claverton Down Bath BA2 7AY, England
Phone: +44-1225-826222
Fax: +44-1225-826492
Email: d.draper@math.bath.ac.uk
Abstract: Government health services worldwide have recently been looking more closely at the quality of care they provide, focusing in particular on what goes on inside hospitals. One method for indirectly assessing the quality of hospital care is a kind of input-output (I/O) approach, in which hospital outcomes are compared after adjusting for differences in inputs. With death as the monitored outcome, the input-output approach takes the form of a contrast between observed and expected mortality rates, given how sick patients are when they arrive at the hospital. A broad implementation of I/O hospital quality assessment would involve data-gathering on many thousands of patients per year, making the cost-effective measurement of admission sickness crucial to the success of health policy initiatives of this kind.Clinical expert judgment typically identifies about 100 variables relevant to the construction of a sickness scale for each disease. The standard way to construct such a scale uses logistic regression with death as the outcome, sifting through available sickness indicators with standard variable-selection methods to find a parsimonious and clinically reasonable subset. From a cost-effectiveness point of view this approach is deficient in that it takes no account of differences in cost of data collection among the available predictors. When both data-collection cost and accuracy of prediction of mortality are considered, a large optimization problem arises in which costly variables that do not predict well enough should be omitted from the scale.
In this paper I take a Bayesian decision-theoretic approach (based on maximization of expected utility) to solving this optimization problem. I compare the usefulness of several methods of stochastic optimization based on MCMC, including simulated annealing and simulated tempering, and competitors including genetic algorithms. I illustrate the methods with data on 2700 elderly pneumonia patients from a large nationally-representative study of Medicare patients hospitalized in the US between 1980 and 1986.
2. Evaluating profiling methods used in the health care industry to identify extreme performance
Christiansen, Cindy L., Harvard Medical School and Harvard Pilgrim Health Care
Address: Cindy L. Christiansen, Ph.D. Assistant Professor of Ambulatory Care & Prevention Harvard Medical School & Harvard Pilgrim Health Care 126 Brookline Ave., Suite 200 Boston, MA 02215
Phone: 617-421-6013
Fax: 617-421-2716
Email: cindy@hphc.org
Abstract: Health care profiles are performed and reported routinely by government, the media, and by private organizations. Often, profile reports are used to identify extreme performance, e.g., which providers have extremely high or extremely low utilization, patient satisfaction, or patient outcome rates. Health care statisticians recently have suggested and used hierarchical models to improve the accuracy of these profile reports and to enhance the value of information for decision makers and health care consumers. We use simulation studies to compare hierarchical modeling methods for detecting outliers with commonly-used exact and approximate p-value methods. These evaluations are made for several settings which mimic the characteristics of actual data.
3. Conditional Categorial Response Models with Application to Treatment of Acute Myocardial Infarction
Gelfand, Alan E., Univ. of Connecticut
Address: Alan E. Gelfand Department of Statistics, U-120P University of Connecticut Storrs, Connecticut 06269-3120
Phone: 860-486-3416
Fax:
Email: GELFAND@UCONNVM.UCONN.EDU
Abstract: For a sample of 2361 patients admitted with suspected acute myocardial infarction (AMI) to a set of 37 hospitals, recorded patient response variables inc lude eligibility for treatment with aspirin, eligibility for treatment with thrombolytics, treatment with aspirin received, treatment with thrombolytics received, and short- term patient survival. Each of these five variables has two levels resulting in a 25 contingency table. Covariate information includes age, sex, race and comorbidity status. Because the responses arrive in sequence we model this data in three st ages: eligibility, then treatment received given eligibility and finally short-term survival given eligibility and treatment received, all given the covariates. Issues of i nterest include extent to which treatment received matches eligibility, whether pobability of survival is affected by treatment status and how the chance of mortality is affected by whether or not treatment received matches eligibility. Influence of covariate information on these quantities is examined. These quantities are studi ed at the hospital level adjusted for case mix and also in aggregate, marginalizing over hospitals.
Discussant: Stangl, Dalene Duke University
Address: Dalene Stangl Box 90251, ISDS Duke University Durham, NC 27708-0251
Phone: 919-684-4263
Fax: 919-684-8594
Email: dalene@stat.duke.edu
List of speakers who are nonmembers: None