Title: On Attribute Weighting in Value Trees

Author: Mari Pöyhönen

Date: April 1998

Status: Systems Analysis Laboratory Research Reports A73, March 1998

 


The thesis focuses on the biases appearing when decision makers are asked to give numerical preference statements on attribute weights in multiattribute value tree analysis (MAVT). The thesis focuses on two problems. First, different attribute weight elicitation methods yield different weights although they are based on the same theoretical principles. Second, the attribute weights change when the structure of a value tree is varied. We run two experiments to study the differences between five different weight elicitation methods (Analytic Hierarchy Process, direct weighting, Simple Multi Attribute Rating Technique, SWING weighting, and tradeoff weighting) and to study how verbal expressions are used in preference elicitation. The weighting methods do yield different weights. These differences originate from the way decision makers restrict their responses depending on the numbers that the methods explicitly or implicitly propose. With the biases related to the structural variation of value trees we first point out that earlier experiments are insufficient because they have drawn conclusions of individual behavior based on averages over large group of subjects. In an experiment we then show that the weights change when the structure of a value tree changes because the subjects again give numbers to describe their preferences so that they clearly favor some response scales. We propose, based on the results from our experiments and earlier observations, that two same origins for many different problems in attribute weighting are that the decision makers give numbers that reflect ordinal information on preferences only and that the weights are normalized to sum up to one. The decision makersí interpretation of the numbers that they use differs from the assumptions of value theory. One way to avoid these problems is to develop more interactive weighting procedures. Different weighting methods should be combined and decision makers should interactively evaluate the results during the weighting. The role of the interactiveness was tested in two experiments. The other focused on the so called preference programming techniques that were tested in a group decision making situation. The other experiment showed that the increased awareness of the methodology can decrease the biases.