Discriminant analysis, a method of analyzing grouped multivariate data, is often used in ecological investigations. It has both a predictive and an explanatory function, the former aiming at classification of individuals of unknown group membership. The goal of the latter function is to exhibit group separation by means of linear transforms, and the corresponding method is called canonical analysis. This discussion focuses on the application of canonical analysis in ecology. In order to clarify its meaning, a parametric approach is taken instead of the usual data-based formulation. For certain assumptions the data-based canonical variates are shown to result from maximum likelihood estimation, thus insuring consistency and asymptotic efficiency. The distorting effects of covariance heterogeneity are examined, as are certain difficulties which arise in interpreting the canonical functions. A 'distortion metric' is defined, by means of which distortions resulting from the canonical transformation can be assessed. Several sampling problems which arise in ecological applications are considered. It is concluded that the method may prove valuable for data exploration, but is of limited value as an inferential procedure.