The goodness-of-fit of a single hypothetical discriminant function in the problem of discrimination between several groups has been considered by Bartlett (1951 a). Virtually, this is the same as testing the adequacy of a single canonical variable to bring out completely the relationship between two vector variables. A single hypothetical canonical variable will be adequate in this 'canonical' or 'external' analysis, if (i) there is only one non-zero canonical correlation, and (ii) the direction of the hypothetical canonical variable coincides with that of the true canonical variable corresponding to the non-zero canonical correlation. Bartlett (1951 a) andWilliams (1952, 1955) have considered these two aspects of' collinearity' and 'direction' in detail, and have derived certain exact tests for them. This situation in 'external' analysis has a parallel in 'internal' or 'principal components' analysis. If the population variance-covariance matrix of a number of variables has all its latent roots equal, except one, then the 'principal component' or latent vector corresponding to this 'anomalous' root can be called the single 'non-isotropic' component because if this component is removed, the variation in any other orthogonal direction is isotropic. In geometrical language, all the principal axes of the familiar ellipsoid (given by the exponent in the density function of a multivariate normal distribution), except one, are equal; in other words, the ellipsoid does not degenerate into a sphere because of this exceptional principal axis. Such a situation can arise in factor-analysis if there is only one 'common' factor (for details see ? 2) besides the 'specific' factor in the factor-structure. If, therefore, we are testing the adequacy of a single hypothetical non-isotropic principal component, it is necessary-as in the case of 'external' analysis-to consider the following two aspects: (i) departure from the hypothesis due to there being more than one non-isotropic principal components, and (ii) departure due to deviation in direction of the true principal component from the hypothetical one. In ? 2 of this paper, these two aspects are considered and an over-all x2 test is obtained, and the underlying assumptions are stated clearly. In ? 3, the over-all X2 iS partitioned and its 'direction' part Ad is separated, using a geometrical argument similar to the one used by Bartlett (1951 a) in deriving the 'direction factor' in external analysis. In ? 4. A is expressed in terms of rectangular co-ordinates. In ? 5, %d is expressed in terms of the sample latent roots and principal components in a manner similar to that of Williams (1955). This is also alternatively expressed in terms of 'residual' roots of the sample variance-covariance matrix, following Williams's (1952) idea of 'residual roots'. In ? 6, the problem of the distribution of these residual roots is considered; and finally in ? 7, the theory is illustrated by a numerical example and a method of deriving approximate confidence intervals for the elements of the single non-isotropic component is described and illustrated.
Read full abstract