Related Experiment Videos
Misclassification among methods used for multiple group discrimination--the effects of distributional properties
1Department of Preventive Medicine and Biometrics, University of Colorado Health Sciences Center, Denver 80262.
Statistics in Medicine
|May 1, 1991
Summary
This study compares multiple group classification methods for normal and non-normal data. Logistic discrimination performs best for non-normal distributions, matching linear discrimination for normal data.
Area of Science:
- Statistics
- Machine Learning
Background:
- Multiple group discriminant analysis is understudied for non-normal distributions.
- Classification performance varies with covariate distributions.
Purpose of the Study:
- To evaluate multiple discrimination methods for classifying more than two populations.
- To compare performance under normal and various non-normal covariate distributions.
Main Methods:
- Simulated continuous normal and non-normal covariate distributions.
- Evaluated methods: polychotomous logistic regression, multiple group linear discriminant analysis, kernel density estimation, rank transformations.
- Varied parameters: population distance, mean vector configuration, skewness, kurtosis, bimodality.
Main Results:
- Logistic discrimination performed near-optimally under Neyman-Pearson allocation across all distributions.
- Performance was comparable for logistic and linear discrimination with normal data.
Conclusions:
- Logistic discrimination is a preferred method for multiple group classification with non-normal data.
- It offers performance comparable to linear discrimination for normal data.