Statistical Inference to Evaluate and Compare Correlated Multi-State Classification Systems
Abstract
The current emphasis on including correlation when comparing diagnostic test performance is quite important, however, there are cases in which correlation effects may be negligible with respect to inference. This proposed work examines the impact of including correlation between classification systems with continuous features by comparing the optimal performance of two diagnostic tests with multiple outcomes as well as providing inference for a sequence of tests. We define the optimal point using Bayes Cost. Through simulation, we quantify the impact of correlation on standard errors comparing two tests and evaluate the resulting errors with respect to CI coverage and width under varying diagnostic test settings. When formulas are required for better inference to include correlation, we provide updated techniques that properly extend the Delta and Generalized method. Additionally, to date, no methods have been applied to quantify the performance of a sequence of tests. Therefore, the inference methods derived in this work are extended to sequenced tests where feature correlation is unavoidable and must be accounted for when developing inference on tests.
Document Details
- Document Type
- Technical Report
- Publication Date
- Aug 20, 2018
- Accession Number
- AD1063255
Entities
People
- Beau A. Nunnally
Organizations
- Air Force Institute of Technology