CANONICAL CORRELATION AND DISCRIMINATION WITH MISSING OBSERVATIONS

A procedure is proposed as a solution to the problem of estimating canonical correlations using samples with missing observations. Simulation studies are made to compare this procedure to the common procedure of ignoring the incomplete observations in the estimation of the correlations. An application is made to the problem of constructing discriminant functions based on training data with incomplete observations. Also proposed are a measure of the possible reduction in variance of the estimates achieved by using the incomplete observations and a hypothesis test on the amount of information contained in the incomplete data.