Modeling Subjective Health Outcomes

Abstract

of Medical Care, Krabbe 1 provides us with a much needed discussion of the potential usefulness of Thurstone’s approach for quantifying subjective health outcomes using ordinal information, such as rankings. As pointed out eloquently by Krabbe, Thurstone’s articles, 2,3 have had a tremendous influence on the development of methods for collecting and analyzing evaluative judgments in the form of preferences, attitudes, or more specifically, health outcomes. In fact, it is our view that Krabbe’s discussion is rather modest. We feel more exuberant and present a list of top-10 reasons for using Thurstonian models. Although presented separately, these reasons should be viewed in conjunction because in this way they demonstrate the power and generality of this approach for studying and understanding evaluative judgments. For many years, it was not possible to apply Thurstone’s model in its full generality because of computational limitations. In technical terms, given rankings of p health outcomes p-1 dimensional integration over a normal distribution is needed if the model is to be estimated by maximum likelihood. Before the computer revolution, this integration was unfeasible and when Thurstone proposed his model more than 80 years ago only univariate integration could be done. Having been trained as an engineer, Thurstone 2 proposed an effective 2-stage estimation procedure that has been the skeleton upon which most classic research on Thurstonian models have been built. 4 Krabbe discusses this method very nicely. Given the proportions by which one health outcome is preferred over another (the P matrix in Krabbe’s article), one obtains the corresponding normal deviates (Z values) by univariate integration. Then, the parameters of the model are obtained from the Z values by minimum distance methods. If only univariate integration is to be performed, only highly restrictive versions of Thurstone’s general model can be estimated (such as the so-called Case V model considered by Krabbe). Fortunately, these computational limitations are now overcome and Thurstone’s model can be estimated in its full generality. As a result, the applicability of this approach has widened tremendously as our list of top-10 reasons illustrates. To put it bluntly, we are comparing a 1930s model car to the latest model when we look at Thurstone’s classic Case V representation with its unrestricted counterpart. Still, the basic advantages of Thurstone’s models should not be dismissed as illustrated by Krabbe, and our list of reasons highlights a number of them before we discuss the current capabilities of a Thurstonian