A Monte Carlo Evaluation of Maximum Likelihood Multidimensional Scaling Methods

We compare three alternative Maximum Likelihood Multidimensional Scaling methods for pairwise dissimilarity ratings, namely MULTISCALE, MAXSCAL, and gurations very well. The recovery of the true dimensionality depends on the test criterion (likelihood ratio test, AIC, or CAIC), as well as on the MLMDS method. The three MLMDS methods test the dissimilarity data equally well. The methods are relatively robust against violations of their distributional assumptions. MULTISCALE outperforms separate Monte Carlo study, it is shown that the MLMDS methods frequently converge to local optima, especially if a random start is used. Rational starts, however, turn out to provide a satisfactory solution for the local optima problem. Implications for researchers intending to apply MLMDS are provided