GENERATING LINEAR AND QUADRATIC COX MIXTURE-MODELS

While Scheffe canonical models have much to offer in the analysis of m ixture data, their strong point does not lie in the ease with which ef fects of mixture components can be inferred from parameter estimates. The motivation for Cox mixture models, which are overparameterized ver sions of Scheffe models, is that inferences are easily made from Cox p arameters about gradients and curvature along Cox-effect directions. G iven a Scheffe model, there are an infinite number of equivalent Cox m odels, but once a reference blend, or base point, is specified, there is a unique Cox model. A FORTRAN program is described that will calcul ate linear and quadratic Cox polynomials for 2 to 10 mixture component s. Required inputs are the design matrix, responses, and the compositi on of base points of interest.