Statistical methods for investigating structural independence in multi-attribute health utility instruments

Full Name: Nicholas Mitsakakis

Academic Affiliation: Institute of Health Policy, Management and Evaluation, University of Toronto

Position: Assistant Professor

Abstract: In health economics cost effectiveness analyses rely on accurate estimation of health utility, a single global measure of health related quality of life (HRQoL). Health utilities are often measured with the use of questionnaire-type instruments, containing a number of items describing specific domains of HRQoL. The construction of these instruments relies on multi-attribute utility theory that makes a number of assumptions. One of them is the so called “structural independence” among the attributes, indicating that being at a particular level on one attribute does not preclude being at any level on another. Here, I am considering these concepts from a statistical point of view, using contingency tables to describe how the responses on the different items are distributed. Structural independence is then equivalent in having zero cell-probabilities in the table. In statistics, there is developed methodology dealing with cases where we know a-priori that the probability is zero (“structural zeros”), or with cases where the data are giving us zero counts due to sampling variation while the true probability is strictly positive (“random zeros”). Here, we face with a different problem, since we do not know a-priori if the true cell-probability is zero or not. I will describe some of my efforts to deal with this problem, using as an example data from a prostate cancer specific instrument.