This paper is concerned with a multidimensional scaling method for data measured with rating items which have ordered categorical choices. A method which gives maximum likelihood estimation of configuration of rated objects and rating items in a latent rating space was proposed. This method has a problem concerned with computational complexity. In order to improve the problem, a probabilistic assumption is introduced to the model of the method. The assumption is that distribution of dissimilarity's error between the points in the rating space is double exponential distribution which is type I extreme value distribution. Monte Carlo experiments have revealed that the method gives considerably good fit to theoretical configuration in the rating space.