How well does the OWA operator represent real preferences?

We present the results of an empirical study, in which we analyze the ability of the Ordered Weighted Average (OWA) operator to model actual preferences in a multi-attribute ranking task. We compare the OWA to a competing model (Simple Additive Weighting SAW), and we also study whether its characteristic feature, the ability to represent different attitudes toward compensation among attributes, is reflected in the preferences provided by subjects. We find that in general, the OWA model fits slightly less well to empirical data. Subjects whose preferences are better explained by the SAW model are also more consistent in their choice behavior. Our results furthermore indicate that preferences of most subjects are not fully compensatory. Thus the ability of the OWA operator to represent different attitudes toward compensation can be a useful feature in modeling actual preferences. The structure of the weights estimated for the OWA operator also suggests that it might provide a good approximation to certain types of decision heuristics.