A student I am working with ran an inadvertant experiement and we are having some trouble trying to account for the results -- primarily because I am not a math-head and have trouble wrapping my brain around some of the more esoteric aspects of MDS. He stacked 4 25 x 25 symmetrical (correlation) matrxes and ran the INDSCAL options in ALSCAL -- but defined the input as shape=asymmetric condition = row model = INDSCAL rather than going with the default input (shape = symmetric, no specified condition) model = INDSCAL. After correcting his error with another run, he compared the results and found that the "inappropriate" input was a "better" solution (higher RSQ by 20 points, lower stress) than the default which more accurately described the input. I ran a similar test on my own data (3 correlation matrixes, 31 x 31) and found slightly better RSQ for the asymm, row-cond run (1.5 points) and slightly better stress.

If we assume that he input the matrixes correctly, then can someone more knowledgeable than I am suggest a reason for this? Why would defining a set of square symmetric matrixes as asymmetric and row-conditional give a more satisfactory solution?

All enlightenment gratefully received.

Kate McCain College of Information Science & Technology Drexel University

mccainkw@duvm.ocs.drexel.edu