A maximum likelihood method for an asymmetric MDS model

A maximum likelihood estimation method is proposed to fit an asymmetric multidimensional scaling model to a set of asymmetric data. This method is based on successive categories scaling, and enables us to analyze asymmetric proximity data measured, at least, at an ordinal scale level. It enables us to examine not only the appropriate scaling level of the data, but also the appropriate dimensionality of the model, using AIC. Prior to or in fitting the asymmetric MDS model, it is important to verify that the data are sufficiently asymmetric. Some variants of symmetry hypotheses are developed for this purpose. Since the emphasis in our paper is not on hypothesis testing, but on model diagnosis, we compare several candidate models including models with these hypotheses based on a similar model comparison idea using AIC. The method is applied to artificial data and a set of friendship data among nations in East Asia and the USA. Relations to other methods are also discussed.