In both the statistical and clinical trial's literature, little work has been published concerning the use of correlated primary efficacy measures at baseline as a potential source for reducing the variability of statistical models and increasing the ability of F-tests to detect the presence of a treatment effect. Traditionally, only the baseline score of the response being analyzed is included in the model as a covariate. The correlation structure among these measures and its impact upon the contribution of the additional covariates in reducing the model's MSE is assessed.

Three models are compared to determine their ability to detect treatment effects in primary measures of efficacy. The first model is a traditional ANOCOVA model, which includes only the baseline measure for the response being analyzed as a covariate. The newly proposed multiple regression model, which includes all baseline scores of primary efficacy measures as covariates, and the multivariate analysis of covariance model are also reviewed. Three actual clinical trial experiments and 40 simulated clinical experiments, designed to represent a large class of correlation patterns among measures, were analyzed with each of the models.

Results from actual clinical trials and simulated experiments indicated the newly proposed multiple regression model established statistical evidence of efficacy more frequently than the ANOCOVA model under a wide range of correlation structures. The use of the additional baseline covariates led to significantly reducing the MSE of the model and increasing the ability to detect treatment differences.

The multivariate model failed to provide any benefits in detecting treatment differences over those observed with the multiple regression model. However, the multivariate approach was valuable in assessing treatment comparability at baseline. In this situation, the multivariate model maintained an overall 5% Type I error rate for assessing baseline comparability among all primary measures, while performing repeated univariate tests led to Type I error rates as high as 23%.