Date of Award

1983

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Biometry

College

College of Graduate Studies

First Advisor

Alan B. Cantor

Second Advisor

M. Clinton Miller

Third Advisor

Loren Cobb

Fourth Advisor

Alan H. Johnson

Fifth Advisor

William E. Groves

Sixth Advisor

Thomas J. Stewart

Abstract

A new Chernoff-type face in color has been developed for purposes of representing and analyzing multidimensional data. This cartoon-like but fairly realistic face is defined by 20 parameters, including 4 color parameters. The programming was done in extended BASIC on the Hewlett-Packard 9845C color graphics computer. A method based on the mean pooled variances of parameter values within observed clusters was developed in order to establish an empirical rank order of importance among the face parameters. It was found experimentally that the smile, the outline of the face, and certain eye parameters were among the most important. Using a model consisting of a mixture of multivariate normal distributions, data were generated artificially from four known populations in order to compare different schemes for assigning data coordinates to face parameters. Five different schemes were experimentally evaluated with regard to their ability to recover known clusterings. The five methods were compared with one another, with random clusterings, and with the results of applying numerical algorithms to the artificial data. The assignment scheme best able experimentally to recover the known clustering was one where principal component scores were used to construct the faces rather than the original, raw data. Numerical algorithms which operated on the component scores were also generally superior to those operating on the original data. Using the new faces, a method was developed to cluster variables rather than the customary clustering of cases. This was compared with the clustering of variables through principal component analysis (varimax orthogonal rotations), and with numerical clustering algorithms which use the product moment correlation as a similarity measure. A data set consisting of psychological profiles of nine entering classes of physicians in a Family Medicine residency was utilized to illustrate some of the foregoing, and also to depict and analyze changes over time of entering class characteristics.

Rights

All rights reserved. Copyright is held by the author.

Download

Share

COinS