Evaluation of Five Discrimination Procedures for Binary Variables

Abstract Five procedures for discrimination with binary variables and small samples are discussed and evaluated. Two procedures are specific for binary variables and are based on first and second order approximations to multinomial probabilities. The third procedure, based on the full multinomial model, is completely general. The fourth and fifth procedures are the linear and quadratic discriminants. Evaluation is in terms of mean actual error and mean correlation between observed and true log likelihood ratios determined by Monte Carlo sampling. The concept of a "reversal" in log likelihood ratios is introduced to explain the results.