OBJECTIVE: To illustrate the use of multivariable optimal discriminant analysis (MultiODA). DESIGN: Data from four previously published studies were reanalyzed using MultiODA. The original analysis was Fisher's linear discriminant analysis (FLDA) for two studies and logistic regression analysis (LRA) for two studies. MEASUREMENTS AND MAIN RESULTS: In Study 1, FLDA achieved an overall percentage accuracy in classification (PAC) for the training sample of 69.9%, compared with 73.5% for MultiODA. In Study 2, the LRA model required three attributes to achieve a 76.1% overall PAC for the training sample and a 79.4% overall PAC for the hold-out sample. Using only two attributes, the MultiODA model achieved similar values. In Study 3, the FLDA model achieved an overall PAC of 82.5%, compared with 87.5% for the MultiODA model. In Study 4, MultiODA identified a two-attribute model that achieved a 93.3% overall training PAC, when an LRA model could not be developed. CONCLUSIONS: MultiODA identified: a superior training model (Study 1); a more parsimonious model that achieved superior overall training and identical hold-out PAC (Study 2); a model that achieved a higher hold-out PAC (Study 3); and a two-attribute model that achieved a relatively high PAC when a multivariable LRA model could not be obtained (Study 4). These findings suggest that MultiODA has the potential to improve the accuracy of predictions made in general internal medicine research.
OBJECTIVE: To illustrate the use of multivariable optimal discriminant analysis (MultiODA). DESIGN: Data from four previously published studies were reanalyzed using MultiODA. The original analysis was Fisher's linear discriminant analysis (FLDA) for two studies and logistic regression analysis (LRA) for two studies. MEASUREMENTS AND MAIN RESULTS: In Study 1, FLDA achieved an overall percentage accuracy in classification (PAC) for the training sample of 69.9%, compared with 73.5% for MultiODA. In Study 2, the LRA model required three attributes to achieve a 76.1% overall PAC for the training sample and a 79.4% overall PAC for the hold-out sample. Using only two attributes, the MultiODA model achieved similar values. In Study 3, the FLDA model achieved an overall PAC of 82.5%, compared with 87.5% for the MultiODA model. In Study 4, MultiODA identified a two-attribute model that achieved a 93.3% overall training PAC, when an LRA model could not be developed. CONCLUSIONS: MultiODA identified: a superior training model (Study 1); a more parsimonious model that achieved superior overall training and identical hold-out PAC (Study 2); a model that achieved a higher hold-out PAC (Study 3); and a two-attribute model that achieved a relatively high PAC when a multivariable LRA model could not be obtained (Study 4). These findings suggest that MultiODA has the potential to improve the accuracy of predictions made in general internal medicine research.