|
Activity Number:
|
182
|
|
Type:
|
Topic Contributed
|
|
Date/Time:
|
Tuesday, August 10, 2004 : 8:30 AM to 10:20 AM
|
|
Sponsor:
|
Biopharmaceutical Section
|
| Abstract - #300933 |
|
Title:
|
Bayesian Analysis of Mislabeling of Training Observations in Discriminant Analysis: Applications to Diagnostic Testing in the Medical Device Industry
|
|
Author(s):
|
Laura A. Thompson*+
|
|
Companies:
|
U.S. Food and Drug Administration
|
|
Address:
|
Center for Devices and Radiological Health, Rockville, MD, 20850,
|
|
Keywords:
|
classification ; Bayesian ; medical devices ; mislabel
|
|
Abstract:
|
The standard case of discriminant analysis constructs a classification rule using training observations that are assumed to be correctly labeled as belonging to their respective classes by some labeling mechanism. However, in practice, the labeling mechanism can be fallible, with an unknown probability of mislabeling training observations. The probability of mislabeling can depend on the class label (class-dependent mislabeling) or even on the value of the observation vector to be labeled (observation-dependent mislabeling). We use a Bayesian framework to develop models for both of these types of mislabeling for the k-class multivariate Gaussian case, and suggest alternatives for relabeling the mislabeled training observations using the posterior distribution of the class labels for the training observations. We also derive predictive classification probabilities for new observations. Our analyses are investigated on several simulated datasets and on example scenarios related to diagnostic testing in the medical device industry.
|
