To Register      SMDM Homepage

Sunday, 17 October 2004

This presentation is part of: Poster Session - Public Health; Methodological Advances

WEB-BASED BAYESIAN COMMUNICATION: THE BAYESIAN Z-TEST

Harold P. Lehmann, MD, PhD1, Alexander Barshay, MS1, L. Allan Grimm, BS1, Karen A. Robinson, MSc2, and Cynthia Sheffield, MLS, MBA3. (1) Johns Hopkins, Health Sciences Informatics, Baltimore, MD, (2) Johns Hopkins, Medicine, Baltimore, MD, (3) Johns Hopkins, Welch Medical Library, Baltimore, MD

Purpose: Bayesian Communication (BC) provides an explicit and quantitative way to combine a reader’s preconceived notions with data from a study to help in making decisions, and thus implements the decision-analytic paradigm in the setting of interpreting and adapting research results. To date, BC has been available only for statisticians or for end-users, interpreted through statisticians. The current research addresses whether BC can be provided over the Web to non-statistical clinicians using a knowledge-based paradigm.

Methods: The domain was clinical trials whose outcomes in two groups were expressed as proportions (i.e., the Bayesian z-test for proportions). A User Panel of clinicians who read research reports (as well as generate them) was employed to elicit front-end specifications and to provide formative evaluation. Usability tests with other end users were further employed during development.

Results: A three-tier architecture was implemented; see http://www.hopkinsmedicine.org/bayes. The Web-based front end guides users through the assessment process: First, the user’s minimally clinically important difference for the absolute difference in percentage is elicited. Then, the user chooses the appropriate statistical model, depending on which parameters they have the most prior knowledge: absolute difference alone; absolute difference plus baseline control rate; absolute difference plus experimental rate; relative difference plus control rate; and relative difference experimental rate. For each parameter, they are then asked to specify the mean and level of certainty. The application then displays the posterior distributions, after the system’s back end utilizes the BUGS Bayesian-statistical updater to perform the calculations. Finally, the system further displays one-way sensitivity analyses for each prior mean and certainty, again, using BUGS, but using novel models for this purpose.

Conclusion: This is the first attempt at delivering non-trivial BC to non-statisticians and provides a model for carrying the decision-analytic agenda forward into the domain of clinical trials.


See more of Poster Session - Public Health; Methodological Advances
See more of The 26th Annual Meeting of the Society for Medical Decision Making (October 17-20, 2004)