Adjustment for bias in discriminant analysis

Abstract

This paper demonstrates the use of an alternative technique when adjusting for bias in discriminant analysis (compatible with the use of R2 and adjusted R2 in regression analysis) for situations where a small sample size for groups exists. To demonstrate the technique, a simple random sample of 119 smokers was selected for the study. A discriminant analysis model was used to determine its predicting power in classifying the smokers into quitters, reducers, and nonchangers. Forty percent of the variability was attributed to group differences, which when adjusted, dropped to 34%. The model did significantly better than a random model in correctly classifying quitters and nonchangers, but was not effective in correctly classifying reducers.