A number of researchers in geography [5; 6; 7; 10; 11; 12; 13; 24; 28] have carried out studies which incorporate various disaggregate spatial choice models (integration theory and functional measurement; multiattribute attitudinal, markovian, MDS and other scaling models; portfolio theory). The findings are inconclusive in the absence of commonly accepted theories of choice behavior, and the models themselves vary substantially in their ability to predict actual patterns of spatial preference and spatial choice. Details of these approaches are, however, well-documented [4; 9; 14]. A noticeable gap in behavioral research has been the limited number of empirical comparisons of alternative modeling approaches. Certainly, there are exceptions [7; 25], but most geographic studies merely highlight conceptual differences [1; 18; 27]. This has hampered necessary advances in behavioral geography. Central to the development of a theory of choice is a procedure (i.e., a decision rule) that specifies how information is to be processed in order to arrive at a selection. The functional form of the decision rule has been a basic concern of geographers. However, a point of controversy which has not been satisfactorily resolved is: to what extent is compensation (or non-compensation)' a modeling assumption, and to what extent is it a substantive finding? The issue is complicated by the fact that empirical tests demonstrate that both compensatory and noncompensatory decision rules can exhibit a reasonable degree of convergent validity under experimental conditions [7; 13; 23]. This paper examines the predictive efficacy of three different modes of processing subjective information in spatial choice situations. The research is innovative in two distinct ways. First, the results are validated by using data collected from the same sample population to predict how subjects make preference and patronage judgments among spatial opportunities. Second, in terms of methodology, two compensatory models (additive conjoint measurement-a decomposition model; and multinomial logit-a composition model)2 are compared di-