Fishbein's Theory of Reasoned Action models behavior as based on beliefs and evaluations on a small set of salient attributes. Two methods of reducing large sets of potentially salient attributes into a smaller set of salient attributes are proposed. The methods are based on expectancy valuation analysis and logistic regression analysis. When applied to consumer beliefs and evaluations on 59 attributes over three milk types (whole, low-fat, and skim milk), both methods identify reduced sets of attributes. The reduced attribute sets are then used to model whether or not respondents drink a particular milk type. Results indicate that the reduced models are statistically significant in explaining choice of m"ilk type although there is some loss of information as compared to models with 59 attributes. Furthermore, the data indicate that statistically-imputed evaluation ratings differ from self-stated evaluation ratings.