Probabilistic multidimensional unfolding: An anisotropic model for preference ratio judgments (Q1896622)

Multidimensional unfolding models represent stimuli and individuals as points in a multidimensional psychological space. Anisotropic unfolding models represent some or all of the points in terms of probabilistic distributions with variances that may differ from dimension to dimension. In contrast, isotropic unfolding models assume that the variance of any stimulus or ideal point is the same on each dimension. This paper extends earlier probabilistic generalizations of multidimensional unfolding models in two key ways: by admitting anisotropic spaces and by allowing for dependent as well as independent sampling. Section 1 of this paper presents the anisotropic unfolding model, describes how preference ratio judgments are modeled as a ratio of quadratic forms in normal variables, and shows how dependent and independent sampling procedures are accommodated. In Section 2, a procedure for obtaining maximum likelihood estimates of the model's parameters is described. Exact and approximate procedures for computing the probability density function (PDF) of distance ratios are presented. The responsiveness of the PDF to changes in the variance structure is also illustrated. To illustrate the importance of anisotropic distributions, Section 3 gives an example that demonstrates the sensitivity of different measures to changes in the variance structure of the stimuli. This is followed in Section 4 by a discussion of sampling properties. If an isotropic model is mistakenly used when the anisotropic model would be appropriate, systematic biases will occur. Even worse results are obtained when deterministic nonmetric methods are employed. A Monte Carlo analysis is used to illustrate these effects in Section 5. Section 6 uses the hypothesis testing abilities of probabilistic models to evaluate the structure of a set of residential choice data. These empirical results support the value of using anisotropic unfolding models with dependent sampling. Finally, a brief discussion section suggests some directions for future research efforts.