Decision SupportImplausible alternatives in eliciting multi-attribute value functions
Introduction
Many approaches to multi-attribute decision making attempt to infer a decision maker’s preferences from the direct evaluations of a set of sample alternatives presented to him or her. Several approaches of this type have been developed independently in different literature streams. Some methods use statistical regression techniques (Schoemaker & Waid, 1982) to estimate weights of additive multi-attribute utility functions from cardinal scores assigned to the sample alternatives. Most approaches, however, are based on pairwise comparisons or a ranking of alternatives. Such approaches are widely considered to be even less demanding in terms of cognitive effort than providing scores. A very early reference to such a method is Srinivasan and Shocker (1973). One of the best known techniques using this concept is the UTA method of Jacquet-Lagreze and Siskos (1982), which is the basis of several similar methods collectively known as preference disaggregation methods (Jacquet-Lagreze & Siskos, 2001). The Linmap method (Horsky & Rao, 1984) is also based on holistic comparisons of alternatives and uses a linear programming model to estimate attribute weights. Similarly, case-based methods, which infer preferences from a set of examples, have become popular in multi-attribute sorting (Chen et al., 2007, Doumpos and Zopounidis, 2004). Recently developed methods to calculate a representative value function (Greco, Kadzinski, & Slowinski, 2011) for decision models under incomplete information (Greco, Mousseau, & Slowinski, 2008) also follow a similar approach. In marketing, the method of Conjoint Analysis was developed to model consumer preferences (Green & Srinivasan, 1978). Conjoint Analysis is based on an additive utility model, and some approaches also use holistic comparisons between alternatives to infer preference parameters. Over the last decades, Conjoint Analysis has evolved into one of the most widely used techniques to investigate consumer preferences (Eggers and Sattler, 2011, Green et al., 2001).
All these techniques rely on the responses of decision makers on a set of alternatives (stimuli). Often, the fact that respondents holistically evaluate entire alternatives, rather than having to specify detailed information on single parameters, is described as a major advantage of such methods. Given the central role of sample alternatives in these methods, it is surprising that the question which alternatives to present to the decision maker is rarely addressed in literature. Usually, the selection of those alternatives is only described in rather general terms, e.g. as alternatives used in previous decisions, or even more general as “fictitious alternatives … which can easily be judged by the decision maker” (Jacquet-Lagreze & Siskos, 2001, p. 235/236).
While the decision making literature thus rarely has investigated the question which alternatives to present in decompositional preference elicitation methods, similar questions have received more attention in the literature on Conjoint Analysis. Given the strong statistical focus of marketing methods, the optimal design of stimuli sets is an important topic in that literature (e.g., DeSarbo et al., 1985, Green and Helsen, 1989).
Based on statistical considerations concerning the efficiency of parameter estimation, orthogonal designs are often recommended. Since orthogonal designs require the evaluation of a large number of alternatives, methods to reduce the required number of comparisons have been developed. These methods minimize the total number of questions to be asked (Holloway & White, 2003) or formulate comparisons which allow to “cut off” particularly large subsets of the remaining parameter space (Toubia, Hauser, & Simester, 2004). However, these methods usually assume that attributes are not correlated and determine the questions to be asked from a technical perspective, while the perspective of the decision maker is rarely addressed.
In real-world decision problems, alternatives often exhibit a correlation of attributes, which is caused by the trade-offs inherent to existing alternatives. For example, when selecting a car, higher speed or a stronger engine usually are accompanied by a higher fuel consumption. Any other attribute combination would be implausible. Another example for an implausible alternative would be a low-priced laptop computer with the fastest processor on the market. However, orthogonal designs (or fractional designs created by an optimization approach) usually ignore such real world correlations, and thus might require the subject to evaluate alternatives which are unlikely to exist in reality (Moore & Holbrook, 1990).
This problem was also recognized in the literature on Conjoint Analysis (e.g., Green, Helsen, & Shandler, 1988). Methods were developed to create optimal designs which avoid such alternatives (Hair, Black, Babin, & Anderson, 2010), and some empirical studies were conducted to investigate whether or not the presence of implausible alternatives has a negative impact on the preference models estimated. While these contributions offer only limited insights into the role of implausible alternatives in Conjoint measurement, they nevertheless highlight the importance of the topic.
Conjoint Analysis and multi-attribute decision models differ in the goals of preference elicitation. In Conjoint Analysis, preferences are elicited in order to predict consumer behavior in future purchases, and thus the predictive abilities of the model over an extended time span are a major concern. In contrast, multi-attribute decision models elicit preferences in order to support decision makers to make complex decisions in a way that is consistent with their preferences. Therefore, consistency is a major concern when preferences are elicited for this purpose.
Due to the focus on the models’ predictive ability, most existing studies on the effects of implausible alternatives in Conjoint Analysis use a two-stage design, in which choices predicted by the elicited model are compared to later choices (of similar complexity) made by consumers. Since the model is used for prediction of consumer behavior, respondents are not aware of the elicited model when they make their choice in the second stage.
In contrast, preference elicitation for multi-attribute decision models is mainly concerned with obtaining a consistent representation of the decision maker’s preference at the time the elicitation is performed. This representation is then used by decision makers themselves to solve more complex problems. Inconsistencies in responses, although they can be corrected by statistical means, are seen as an indicator that the decision maker is not sure about his or her preferences, or is confused by the questions being asked. This could impede the reliability of the elicitation. Therefore, consistency of the responses is a major concern, while the ability of the model to predict choices a considerable time later is not relevant for this application. Therefore, we focus on immediate results rather than on choices in a later stage.
Analyzing the consistency of responses requires to develop a method to determine compatibility of subjects’ choices with plausible assumptions about preference parameters. We introduce a new approach, and apply it in an experimental setting to study the impact of implausible alternatives on the elicitation process. More specifically, in our research we want to identify negative effects of implausible stimuli directly on the elicitation process and the estimated parameters.
The remainder of the paper is structured as follows: Section 2 provides a literature review, based mainly on literature in Conjoint Analysis, on possible effects of implausible alternatives in preference elicitation. In Section 3, we use these results to formulate research questions for our empirical study, which is described in Section 4. Section 5 describes the model we use to test compatibility of parameters with plausible assumptions, and Section 6 presents the empirical results. Section 7 concludes the paper by summarizing its main results and providing an outlook on future research.
Section snippets
Literature review
Since its development in the 1970s, Conjoint Analysis has become one of the most widely applicable techniques for identifying consumers’ preferences (Green & Srinivasan, 1990). A considerable amount of literature has been devoted to algorithms and applications of this approach (e.g., Johnson, 1974, Srinivasan and Shocker, 1973). However, the perspective of respondents is less prominent in existing literature.
The most widely used method for eliciting preference parameters is the “full-profile”
Research questions
As our literature review has indicated, there still exists only limited evidence on whether considering plausible designs leads to a better elicitation of preferences. More than 20 years ago, Moore and Holbrook (1990) have noted that a large number of applications of Conjoint Analysis have ignored reliability and validity issues. Green and Srinivasan (1990) also ask for more research on the topic of robustness of orthogonal designs in correlated attribute environments. This call remained widely
Method
To answer our research questions, we performed an empirical study in a typical application area of Conjoint Analysis. Following recent research streams in marketing (e.g., Darian, Wiman, & Tucci, 2005), we used a store choice problem as an application example. An essential step is the identification of the relevant attributes that are responsible for forming the consumers’ preferences and therefore have to be included in the stimulus set.
In recent years, a comprehensive body of literature was
Model
Research question RQ1 refers to dominance between the alternatives presented to subjects. Dominance can only be established if the ranking of outcomes within each attribute is known. While a plausible ranking can be assumed for the attributes distance, price, and quality, this is not the case for store size. However, both stimuli sets contain dominating alternatives for different preferences with regard to store size. If a subject prefers small over large stores, then alternative number 4 in
Dominance
To test consistency with the dominance condition (RQ1), we examined whether the preferences indicated by subjects violated the dominance relations indicated in Section 5, depending on the subjects’ stated preferences for store sizes. We then applied a χ2 test to verify the significance of results. The outcomes of this analysis are presented in Table 4.
In contrast to our assumptions, the implausible group shows a lower level of dominance violation (9.0%) compared to the plausible group (17.9%).
Conclusions and further research
In our experiment, we tested the effect of presenting implausible alternatives to subjects on three types of consistency of responses and the elicited utility functions: Consistency with the principle of dominance, consistency with (stated or reasonable) intra-attribute preferences, and consistency with stated inter-attribute preferences. By considering three different types of consistency, we were able to draw a more differentiated picture of the impact of implausible alternatives on utility
References (42)
- et al.
Comparative analysis of UTA multicriteria methods
European Journal of Operational Research
(2001) - et al.
Retail patronage intentions: The relative importance of perceived prices and salesperson service attributes
Journal of Retailing and Consumer Services
(2005) - et al.
Developing sorting models using preference disaggregation analysis: An experimental investigation
European Journal of Operational Research
(2004) - et al.
Selection of a representative value function in robust multiple criteria sorting
Computers and Operations Research
(2011) - et al.
Ordinal regression revisited: Multiple criteria ranking using a set of additive value functions
European Journal of Operational Research
(2008) - et al.
Planning merchandising decisions to account for regional and product assortment differences
Journal of Retailing
(1999) - et al.
Question selection for multi-attribute decision-aiding
European Journal of Operational Research
(2003) - et al.
Assessing a set of additive utility functions for multicriteria decision-making, the UTA method
European Journal of Operational Research
(1982) - et al.
Preference disaggregation: 20 Years of MCDA experience
European Journal of Operational Research
(2001) - et al.
Determinants of retail patronage: A meta-analytical perspective
Journal of Retailing
(2006)
A hierarchical Bayes model of choice between supermarket formats
Journal of Retailing and Consumer Services
Learning about preferences in electronic negotiations – A volume based measurement method
European Journal of Operational Research
Behavioral influences on weight judgments in multiattribute decision making
European Journal of Operational Research
Incorporating prior knowledge into the analysis of conjoint studies
Journal of Marketing Research
Shopping behavior and consumer preference for store price format: Why “large basket” shoppers prefer EDLP
Marketing Science
Multiple-criteria sorting using case-based distance models with an application in water resources management
IEEE Transactions on Systems, Man, and Cybernetics Part A: Systems and Humans
International expansion and strategies of discount grocery retailers: The winning models
International Journal of Retail and Distribution Management
On the creation of feasible conjoint analysis experimental designs
Preference measurement with conjoint analysis. Overview of state-of-the-art approaches and recent developments
New Methods
On the design of choice experiments involving multifactor alternatives
Journal of Consumer Research
Cross-validation assessment of alternatives to individual-level conjoint analysis: A case study
Journal of Marketing Research
Cited by (12)
An active preference learning approach to aid the selection of validators in blockchain environments
2023, Omega (United Kingdom)Best-worst Tradeoff method
2022, Information SciencesCitation Excerpt :See, for instance, Fig. 5. This suits the results obtained by Vetschera et al. [23] according to which alternatives with extreme values may improve the consistency of results. However, this should not be interpreted as binding.
Ranking range model in multiple attribute decision making: A comparison of selected methods
2021, Computers and Industrial EngineeringCitation Excerpt :Multi-attribute utility theory (MAUT) (Baucells & Sarin, 2003; Deutsch & Malmborg, 1986; Wallenius et al., 2008). The MAUT provides an axiomatic foundation to rank alternatives involving multiple attributes via constructing the appropriate preference representation functions (Tsetlin & Winkler, 2009; Vetschera, Weitzl, & Wolfsteiner, 2014; Weber, 1985). For example, Keeney (1976) introduced additive and multiplicative utility function forms as the general utility forms satisfying mutual utility independence, which have been widely used in MADM.
Stakeholder interviews with two MAVT preference elicitation philosophies in a Swiss water infrastructure decision: Aggregation using SWING-weighting and disaggregation using UTA<sup>GMS</sup>
2018, European Journal of Operational ResearchCitation Excerpt :On the other hand, a larger magnitude of differences increases the use of indifference statements when only indifference and preference answers are permitted. Moreover, first studies indicate that the plausibility of hypothetical alternatives can have an effect on the consistency of the decision-makers responses; but there seem to be differences between intra- and inter-attribute preferences (van Valkenhoef & Tervonen, 2016; Vetschera, Weitzl, & Wolfsteiner, 2014). Thus, there is certainly opportunity for further research in this area.
Modeling project preferences in multiattribute portfolio decision analysis
2017, European Journal of Operational ResearchCitation Excerpt :On the other hand, eliciting preferences using real projects may result in higher DM involvement that could in turn lead to less biases. In standard single alternative choice problems, the use of hypothetical reference alternatives has been shown to lead to more consistent preference statements (Vetschera, Weitzl, & Wolfsteiner, 2014), but it is unclear whether this result is generalizable to the PDA context. Second, our results suggest that potential optimality may provide a more intuitive and readily extendable solution concept than dominance for multiattribute portfolio decision analysis under incomplete information.
How well does the OWA operator represent real preferences?
2017, European Journal of Operational Research