Showing papers by "Ward Edwards published in 1980"

Equal Weights, Flat Maxima, and Trivial Decisions.

Abstract: : Most predictions are intended as a basis for decision making. The point of this paper is that prediction and decision require different methods. Equal weights, while often useful for prediction, are less useful for decision making. The action options available in any decision problem fall into three classes: sure winners, sure losers, and contenders. Sure winners and sure losers are defined by dominance, accepting sure winners and rejecting sure losers is trivial. Good decision rules should discriminate well among contenders. In the familiar pick-1 decision problem, options on the Pareto frontier (i.e. undominated options) almost always show negative correlations among attributes. Such negative correlations make equal weights inappropriate. This paper extends that result to the case in which a decision maker must pick k options out of n. In this case, the set of sure winners is usually not empty. It develops general procedures for identifying the set of contenders, given the options, k, and n. This set is a generalized Pareto frontier, of which the traditional kind is a special case. Simulations show that attribute intercorrelations among contenders are substantially depressed and typically negative, even if the intercorrelations in the whole set are positive. Such negative correlations among contenders strongly question the usefulness of equal weights for decision making.

Evaluating Credit Applications: A Validation of Multiattribute Utility Techniques Against a Real World Criterion,

Abstract: : Twenty-two credit officers from a major California lending institution served as subjects in a criterion validation of multiattribute utility elicitation techniques. The techniques tested were the Holistic Orthogonal Parameter Estimation (HOPE) technique (Barron and Person, 1978), Simple Multiattribute Rating Technique (SMART: Edwards, 1977), point distribution, and three rank weighting techniques as discussed in Stillwell and Edwards, 1979. Equal weighting of importance dimensions was also investigated. The criterion against which the judgments were compared was the lending institutions own credit scoring model. This model is based on statistical analysis of over 8,000 cases from the bank records and is a best fit prediction model. Results demonstrate that subjective judgments of importance weighting show a high degree of agreement in application selection and in total utility realized from that selection. Decomposition techniques did somewhat better than holistic techniques.

A Comparison of Importance Weights for Multiattribute Utility Analysis Derived from Holistic, Indifference, Direct Subjective and Rank Order Judgments.

Abstract: : Research done in the 1960's and early 1970's suggested that although statistical weights and subjective weights show some correspondence in regression-like situations, subjective weights tend to be too flat by comparison; statistical weights usually show that some attributes are quite important, while others are hardly important at all. More recent discussions of this literature, however, have pointed out a number of methodological problems with much of the early research, and have reached a more optimistic conclusion with respect to subjective weights. Several experiments support the more recent interpretation. The present study compared weight estimation procedures for additive, riskless, four-attribute value functions with linear single-attribute values. Self-explicated (subjective) weights were assessed from direct subjective and rank order estimates of attribute importance; observer-derived weights were determined both from indifference judgments (axiomatic approach) and from holistic evaluations (statistical approach) of alternatives. Assessed weights were compared to a true weight vector used to generate feedback during pre-assessment learning trials (constructed with zero inter-attribute correlations). Although self-explicated weights tended to be flatter than observer-derived weights, resulting composites correlated equally well with true composites. Only slight differences were found in ordinal correspondence between true and assessed weights.