Deriving Attribute Utilities From Consideration Sets: an Alternative to Self-Explicated Utilities

David B. Klenosky, DePaul University
W. Steven Perkins, Pennsylvania State University
ABSTRACT - Questions have arisen concerning the use of self-explicated utilities as an element in conjoint analysis. This research proposes an alternative method by deriving utilities from consideration sets. Fifty subjects screened conjoint profiles into groups of alternatives they would accept for consideration, put on hold, or reject. They then completed traditional conjoint and self-explicated tasks. Attribute utilities derived from their consideration sets differed markedly from their self-explicated utilities, apparently reflecting a different aspect of the consumer choice process. In subsequent hybrid conjoint analyses, including the consideration set utilities in place of the self-explicated utilities resulted in a higher R2 and alleviated the bias in the parameter estimates representing the attribute-level effects.
[ to cite ]:
David B. Klenosky and W. Steven Perkins (1992) ,"Deriving Attribute Utilities From Consideration Sets: an Alternative to Self-Explicated Utilities", in NA - Advances in Consumer Research Volume 19, eds. John F. Sherry, Jr. and Brian Sternthal, Provo, UT : Association for Consumer Research, Pages: 657-663.

Advances in Consumer Research Volume 19, 1992      Pages 657-663


David B. Klenosky, DePaul University

W. Steven Perkins, Pennsylvania State University


Questions have arisen concerning the use of self-explicated utilities as an element in conjoint analysis. This research proposes an alternative method by deriving utilities from consideration sets. Fifty subjects screened conjoint profiles into groups of alternatives they would accept for consideration, put on hold, or reject. They then completed traditional conjoint and self-explicated tasks. Attribute utilities derived from their consideration sets differed markedly from their self-explicated utilities, apparently reflecting a different aspect of the consumer choice process. In subsequent hybrid conjoint analyses, including the consideration set utilities in place of the self-explicated utilities resulted in a higher R2 and alleviated the bias in the parameter estimates representing the attribute-level effects.


Researchers attempting to study and predict consumer preferences frequently rely on attribute importance weights and attribute-level desirabilities that are directly elicited from respondents. The collection of such "self-explicated" data, as it is commonly called, is recognized as a useful and relatively efficient approach for modeling consumer choice behavior (Leigh, MacKay and Summers 1984; Srinivasan 1988; Green and Helsen 1989). Interest in self-explicated data has become particularly important recently with the development of "hybrid" conjoint models that combine the data-collection efficiencies of the self-explicated approach with the generality and power of regression-like conjoint preference models (see Green 1984 for a review of hybrid conjoint models). One popular conjoint software package even includes a self-explicated task before respondents evaluate profiles (Johnson 1987). Recent research, however, has noted possible drawbacks to the use of self-explicated data. First, the information-processing assumptions underlying the self-explicated approach may be violated by consumers' actual choice behavior. Second, estimates of attribute parameters may become biased when self-explicated data are included in a regression equation. This paper presents an alternative method for generating attribute utilities that addresses both of these issues.

Choice Behavior Inconsistent with Underlying Assumptions

Fundamental to the self-explicated approach is the assumption that respondents can accurately state which levels within an attribute are acceptable versus unacceptable. Furthermore, it is assumed that respondents use a conjunctive evaluation procedure in which alternatives possessing unacceptable attribute levels will be rejected from further consideration, while those possessing acceptable levels may be considered depending on the levels of other attributes.

Previous research, however, indicates that people do not always adhere to the conjunctive implications of the self-explicated procedure. In particular, Klein (1986) found that respondents often fail to reject alternatives with attribute levels which they themselves had previously designated as unacceptable. More specifically, she found that 15 per cent of the alternatives chosen as best by the respondents in her study possessed an attribute level that was supposedly "completely unacceptable." In a later study, Green, Krieger and Bansal (1988) demonstrated that the tendency to consider alternatives with "completely unacceptable" levels can be somewhat reduced by strengthening the instructions used to rate the attribute-level desirabilities. Interestingly, however, these researchers found that there is still a 47% chance, at best, that respondents will fail to adhere to the conjunctive implications of the completely unacceptable level. In sum, researchers rely on the self-explicated approach, despite the fact that its assumptions are often violated by actual choice behavior.

Bias in Parameter Estimation

Several studies have compared the predictive accuracy of self-explicated models, traditional conjoint models and hybrid conjoint models (Akaah and Korgaonkar 1983; Cattin, Hermet and Pioche 1982; Green, Goldberg and Wiley 1982; Moore and Semenik 1988). Although the results of these studies have been somewhat inconsistent, the traditional conjoint and hybrid conjoint models typically outperform the self-explicated model on a number of cross validation measures (i.e., adjusted R-squared, percent of first-choice predictions and average holdout sample correlations). Interestingly, however, the hybrid conjoint model's improvement over the traditional conjoint model is often not very substantial. According to Green, Goldberg and Montemayor (1981), the variation captured by the self-explicated component of the hybrid model may fail to increase the amount of variance accounted for by the conjoint component. In other words, the information captured by the self-explicated component of the hybrid model may be partially redundant with the information captured by the "conjoint" component.

This problem can be seen by analyzing the effect of adding the self-explicated component to the other terms in the hybrid conjoint model, i.e., the terms reflecting the attribute-level dummy-variables. Adding the self-explicated term to the hybrid model may have an adverse, biasing effect on the parameter estimates of the dummy variables. In particular, dummy-variable parameter estimates that were previously significant may become insignificant when the self-explicated term is added to the hybrid model. Interestingly, this problem has yet to be addressed in published studies of the hybrid conjoint model.

Given the above problems with the self-explicated approach in general and in its use in hybrid conjoint analysis in particular, it is important to consider alternative approaches for estimating attribute-level desirabilities.

An Alternative to the Self-Explicated Approach

One alternative to the collection of self-stated attribute utility estimates is to derive those estimates based on actual choice behavior. In this paper we describe such an approach, called the "screening" approach, in which respondents are shown full-profile choice alternatives (like those used in traditional conjoint studies) and are asked to decide which of three groups each alternative belongs in: (1) those the respondent would definitely consider selecting, (2) those he might consider selecting and (3) those he would definitely reject from further consideration. As shown below, these "screening decision" responses can then be used to derive utility estimates representing the desirability of each level of each attribute for each respondent.

Study Objectives

In general, one would expect that these screening-derived attribute-level utilities would be somewhat similar to the attribute-level desirabilities obtained using the self-explicated approach. One purpose of this research is to determine whether this is in fact the case. Thus, our first objective is to assess the convergent validity of the self-explicated approach with an alternative approach based on actual screening-decision behavior.

As noted above, the information reflected in the self-explicated term is often redundant with the other terms of the hybrid conjoint model, resulting in biased dummy-variable parameter estimates. Using an alternative term based on the screening-derived attribute utilities may be one way to get around this problem. Thus, the second objective of this study is to assess the extent to which a screening-derived utility term (used in place of the self-explicated term in an alternative hybrid model) reduces the amount of bias in the parameter estimates typically found in hybrid conjoint models.


Typically in hybrid conjoint, self-explicated utilities are gathered by having the respondent rate the desirability of each level of each attribute. These utilities are denoted as uij for the utility of level i on attribute j. Though scales of varying lengths have been employed, a simple three point scale--best, acceptable, unacceptable--is not uncommon (e.g., Green 1984, Figure 3). The respondent notes which single level of the attribute is best, then marks the remaining levels as either acceptable or unacceptable. Then the analyst assigns a score to each point on the scale, such as best=3, acceptable=2, and unacceptable=1. (Respondents may also be asked to state their self-explicated importance weight for attribute j, but to keep the focus on the attribute level, all attributes in this study were unit weighted.) For the estimation phase of the project, the self-explicated utilities are summed by profile according to the corresponding attribute levels in the profile. This sum represents a holistic, compositional estimate of the utility for that profile. (In some studies, the utilities are not summed; instead weights are estimated for each attribute, though calculating the summed utility appears often, as in Moore and Semenik 1988).

In contrast, this project investigated a different approach to generating the uij's, though the ultimate use of the summed utility in estimation was the same. Specifically, the respondent's task differed from the rating approach seen in most hybrid studies. To derive the utilities, respondents screened each profile by assigning it to one of three categories--accept for consideration, hold (undecided), or reject from consideration. Then working from those observed screening decisions, utilities were developed by calculating the probability of accepting, holding, or rejecting, for each level on an attribute. For example, a respondent might screen four profiles with attribute j at level i, rejecting three of the profiles and putting one on hold. The derived probability of rejecting equals 75%, with a 25% chance of holding, and a 0% chance of accepting. A utility value for that level could be calculated by weighing the score for each category by the probability of that category occurring. Thus for an attribute j:



pik = the probability that a profile at level i (of attribute j) will be screened into category k

sk = the numeric desirability score for category k

Note that the typical rating task employed in gathering self-explicated data can also be expressed in these terms; however, the scale value marked by the respondent receives a probability of 100% while all other scale values for that attribute level equal 0% probability.

A brief numeric example should make the steps in deriving the utilities more clear. Assume that a respondent screens twelve profiles, four profiles at each of the three levels on attribute j. The results might be as in Table 1. These same steps would be followed for each attribute. The screening derived uij could then be treated in the same manner as self-explicated utilities in the estimation phase.

Though mathematically parallel to the self-explicated utilities, the screening derived utilities capture a different aspect of the judgment process. In this project, the subjects' task was specifically to screen the profiles for consideration, not to choose one. In a later task they rank ordered the same profiles and completed the self-explicated utility ratings. Thus the screening derived utilities represent the weights subjects placed on attribute levels during consideration set formation, not during final evaluation and choice. Several researchers have hypothesized that at this earlier consideration set stage subjects would exhibit more non-compensatory behavior, compared to more compensatory behavior at a later evaluation stage (Roberts and Lattin 1991; Wright and Barbour 1977). That is, at an earlier stage in the process, subjects would be more likely to use a conjunctive rule and reject a choice alternative (profile) if any of its attribute levels were below an acceptable level. Whereas, in a later stage those alternatives that have passed screening would be evaluated in a more compensatory manner.




The data for this research was collected as part of a larger study designed to investigate the consumer consideration set formation process (Klenosky 1990). The study involved a simulated decision-making task in which subjects were asked to select a restaurant, from a pool of available restaurants, for a hypothetical restaurant choice situation. (Though not relevant for the present study, subjects were randomly assigned to one of two situations: (1) choose a restaurant to go to when you're feeling too lazy to cook and (2) choose a restaurant to go to for dinner with your boy/girlfriend and his/her parents).


A paid convenience sample of 50 university students were recruited for the study. Subjects were selected based on a screening questionnaire to ensure they had some familiarity and interest in the choice domain involved.

Stimuli Used as Choice Alternatives

The stimuli used in the study consist of descriptions of hypothetical restaurants constructed to simulate condensed restaurant reviews. Based on a literature review and a series of pilot studies, six attributes were selected to describe the restaurant alternatives (see Table 2 for a summary of the attributes and attribute levels used in the study). As a starting point, the three levels of the four most important attributes (Atmosphere, Food Quality, Service Quality and Price) were fully crossed to create a base set of 34, or 81 profiles. Profiles with unrealistic attribute level combinations were then discarded reducing the available set to 63 profiles. Finally, the levels of two additional attributes (Food Variety and Waiting Time) were randomly assigned to each profile to create 63 restaurant alternatives described on a total of 6 attributes. The orthogonality of the attributes in these profiles was a concern. However, a subsequent analysis indicated that the degree of correlation among the attributes used in the profiles was low, indicating the design was still relatively orthogonal.


Each subject participated individually in a session whose average time was 45 minutes. Subjects were first shown a brief description of the choice situation and then began to examine the profiles. The profiles were not in any particular order and were drawn at random by the experimenter. Subjects were allowed to examine as many or as few profiles as they wanted or needed to in order to make their choice. Subjects were instructed to indicate what they wanted to do with each profile by placing those that they wanted to consider for the situation in a pile to their left and those that they wanted to reject from consideration in a pile to their far right. They were also told they could use a middle pile if they liked as a "hold" pile for alternatives they were unsure of. Subjects indicated when they were ready to stop examining profiles and make their choice. They then ranked each of the profiles examined from their most preferred to their least preferred. (Note that the 50 subjects participating in the study examined an average of 6.4 profiles, with a range of 1 to 22. Of the 306 profiles screened, 21% were accepted, 15% put on hold, and 64% rejected).

Once the process of examining profiles ended, subjects completed the self-explicated task by indicating the desirability of each attribute level using the simple three-point scale described earlier. That is, they first indicated, for each attribute, which level was the best or most preferred, and then rated the remaining levels as either acceptable or unacceptable.




As noted above, the first purpose of this research was to assess the convergence between self-explicated utilities and screening derived utilities. Table 2 presents these results by attribute level; the means across attributes appear at the bottom of the table. There are noticeable differences between the two approaches. In particular, on average the unacceptable level was chosen 27% of the time but 64% of the profiles were rejected. On fourteen of the seventeen attribute levels, respondents rejected the profiles at a higher rate than they rated the attributes as unacceptable. The differences between the two approaches are quite large on some attributes, such as "casual atmosphere" where none of the respondents rated that as unacceptable, but 60% of the time profiles at that level were in fact rejected. Of course, such differences are understandable since our methodology for deriving utilities based on screening behavior enforces a strict conjunctive rule that if a profile is rejected, all the attribute levels for that profile are considered rejected.

The deviations between the self-explicated and screening derived utilities can also be examined by correlating the two sets of utilities produced by each individual. The mean correlation coefficient across the 50 subjects equalled .40 (after a Fisher Z transformation) with a range from +.90 to -.73.For 46% of the respondents the two sets of utilities were significantly correlated; or conversely, for 54% the two sets were not related using a 1 tail test at alpha = .05. There is one caveat: the observations within an individual's set of utilities are not completely independent. Thus, while these correlation results indicate some lack correspondence between the two utility measures, they are not definitive.

Table 3 summarizes the extent to which subjects violated the conjunctive implications of the self-explicated procedure by crosstabulating the outcome of the screening decisions by the presence of unacceptable attribute levels in the profiles. If any attribute level in a profile had been rated as unacceptable by the respondent, then the profile was classified as "unacceptable" (see Green, Krieger, and Bansal 1988, Table 4). We might expect that all unacceptable profiles should have been rejected, and in fact of the 245 profiles classified as "unacceptable," over 75% were rejected. On the other hand, however, respondents failed to reject 25% of these supposedly "unacceptable" profiles. Note also that of the 64 profiles that were accepted by the respondents, 40% contained at least one unacceptable attribute level, indicating once again a moderate level of inconsistency.





The median rank orders for the profiles which fell into each screening by self-explicated cell are also presented in table 3. (The ranks were rescaled by respondent such that the most highly ranked profile equalled 1 and the least highly equalled 0.) The rankings monotonically decline from accept to reject decisions and from acceptable to unacceptable attribute levels, as expected. Respondents' rankings were congruent with their screening and self-explicated results. Thus, the screening and self-explicated utilities should be useful in predicting preferences, even though the respondents were not always consistent in following the conjunctive rule.

The second purpose of the research was to assess the effect of including a summed screening derived utility term in a hybrid conjoint regression equation in place of the typical summed self-explicated term. In previous studies the inclusion of the self-explicated term, along with the traditional dummy variables representing the levels on the profile attributes, has improved the predictive ability of the equation (e.g., Green 1984). However, this increase may come at the cost of introducing some bias into the estimation of the regression parameters because the self-explicated term can be redundant with the attribute levels. Thus, the coefficients for the dummy variables will be altered if the self-explicated utility term is included.

Using the rescaled respondents' rankings of the choice alternatives as the dependent variable, three equations were estimated with the CONJOINT macro in SAS which performs monotonic regression (SAS 1985). All equations were estimated at a total group level, n=308. As a baseline for comparison, the first equation included only the dummy variables for the profile attribute levels. The estimated parameters appear in Table 4, along with the R2 of .35. Note that seven of the eleven attribute level parameters are significant.

A second equation included the self-explicated utility term, summed by profile, along with the dummy variables, following the same formulation shown in Moore and Semenik (1988, their equation 3). The R2 increased to .46 as might be expected, but in this case four of the attribute level parameters which had been significant in the dummy variable equation are now insignificant. Only four parameters are significant when the self-explicated term is included. This is important from a managerial perspective because the meaning of the parameters changes between the two equations. Considering the self-explicated term alone, without dummies, results in an R2 of .39.

Finally, a third equation substituted the summed screening derived utility in place of the self-explicated term in the second equation. In this case the R2 improved to .50, but more importantly the attribute parameters remained as they were in the first dummy variable equation. The same seven which had been significant previously, remain significant in this equation indicating that the screening-derived term is bringing in new information, uncorrelated with the profile dummy variables. An equation with the screening derived term alone, without dummies, produces an R2 of .36.


This research has merged aspects of consideration set formation with the modelling of consumer preferences. Previous research has noted the consideration set formation as an essential step before product evaluation (Roberts and Lattin 1991), but many models ignore that step and focus only on the final step of the process: choice. Obviously if a product is not included in the consumer's consideration set, it cannot be chosen. To model choice more realistically, the earlier processes should also be included.

The typical self-explicated utility task asks consumers to declare whether any levels of an attribute are unacceptable. Thus, if a product had that level, it would not be considered for choice. The method described in this paper seeks to obtain the same information from the consumer; however, in this method, attribute utilities are derived by screening profiles for consideration. Our approach entails a less direct task from the consumer's perspective--they do not overtly rate attribute levels. On the other hand, this approach may provide a more direct, behaviorally-based measure of attribute utilities.

Summarizing the research results, we find noticeable inconsistencies between the screening decisions and the self-explicated utilities. For example, over 40% of the profiles respondents accepted for consideration contained at least one "unacceptable" attribute level, while 25% of the profiles with "unacceptable" levels were not rejected from consideration. We also find that, compared to a summed self-explicated utility term, a summed screening-derived utility term is less redundant with the dummy variables in a hybrid conjoint equation.

These results suggests that the utilities generated by the screening task differ from those generated by traditional self-explicated methods. These differences may arise from both our mathematical formulation as well as the respondent's task. First, our formulation calculates the utility of a profile by weighing the desirability scores for an attribute level by the probabilities derived from the screening decisions. This probabilistic formulation of utility may be a more realistic representation of the respondent's behavior, but further research is needed on this point. To improve the comparability of utilities across subjects, future research could require all respondents to screen the same set of, say 18 factorially designed, profiles. An additional possibility would be to substitute a profile screening task for the self-explicated task now built into the Adaptive Conjoint Analysis (ACA) software program (Johnson 1987). Green, Krieger, and Agarwal (1991) have recently questioned several aspects of ACA, and a screening task may help resolve some of those issues.

Second, in addition to differences due to our mathematical formulation, differences between the screening derived and self-explicated utilities may also result from the task the respondent is asked to perform. The screening task captures the consideration set formation stage of the choice process. Thus, the weights placed on attribute levels during consideration set formation may well differ from those associated with the final choice stage. Future research could explore the factors which increase or decrease the differences between the weights used in the two stages, including factors related to product class, respondents, purchase situation, and task instructions.


Akaah, Ismael P. and Pradeep K. Korgaonkar (1983), "An Empirical Comparison of the Predictive Validity of Compositional, Decompositional, and Hybrid Multiattribute Preference Models," Journal of Marketing Research, 20 (May), 187-97.

Cattin, Philippe, Gerard Hermet, and Alain Pioche (1982), "Alternative Hybrid Models of Conjoint Analysis: Some Empirical Results," in Analytical Approaches to Product and Market Planning: The Second Conference, Cambridge, MA: Marketing Science Institute (October), 142-52.

Green, Paul E. (1984), "Hybrid Models for Conjoint Analysis: An Expository Review," Journal of Marketing Research, 21 (May), 155-9.

Green, Paul E., Steven M. Goldberg, and Mila Montemayor (1981), "A Hybrid Utility Estimation Model for Conjoint Analysis," Journal of Marketing, 45 (Winter), 33-41.

Green, Paul E., Steven M. Goldberg and James B. Wiley (1982), "A Cross Validation Test of Hybrid Conjoint Models," Advances in Consumer Research, Vol. 10, R. Bagozzi and A. Tybout, eds., Ann Arbor, MI: Association for Consumer Research, 147-50.

Green, Paul E. and Kristiaan Helsen (1989), "Cross-Validation Assessment of Alternatives to Individual-Level Conjoint Analysis," Journal of Marketing Research, 26 (August), 346-350.

Green, Paul E., Abba M. Krieger, and Pradeep Bansal (1988), "Completely Unacceptable Levels in Conjoint Analysis: A Cautionary Note," Journal of Marketing Research, 25 (August), 293-300.

Green, Paul E., Abba M. Krieger and Manoj K. Agarwal (1991), "Adaptive Conjoint Analysis: Some Caveats and Suggestions," Journal of Marketing Research, 28 (May), 215-222.

Johnson, Richard M. (1987), "Adaptive Conjoint Analysis," Sawtooth Software Conference on Perceptual Mapping, Conjoint Analysis, and Computer Interviewing, Ketchum, ID: Sawtooth Software, 253-265.

Klein, Noreen M. (1986), "Assessing Unacceptable Attribute Levels in Conjoint Analysis," in Advances in Consumer Research, Vol. 14, M. Wallendorf and P. Anderson, eds., Provo, UT: Association for Consumer Research, 154-8.

Klenosky, David B. (1990), The Consumer Choice Set Formation Process: The Role of Screening and Stopping Decisions, Unpublished Ph.D. Dissertation, Pennsylvania State University.

Leigh, Thomas W., David B. MacKay, and John O. Summers (1984), "Reliability and Validity of Conjoint Analysis and Self-Explicated Weights: A Comparison," Journal of Marketing Research, 21 (November), 456-62.

Moore, William L. and Richard J. Semenik (1988), "Measuring Preferences with Hybrid Conjoint Analysis: The Impact of a Different Number of Attributes in the Master Design," Journal of Business Research, 16, 261-74.

Roberts, John H. and James M. Lattin (1991) "Development and Testing of a Model of Consideration Set Composition," Journal of Marketing Research, forthcoming.

SAS Institute Inc. (1985) The MATRIX Procedure, Technical Report P-135, Cary, NC: SAS Institute Inc.

Srinivasan, V. (1988), "A Conjunctive-Compensatory Approach to the Self-Explication of Multiattributed Preferences," Decision Sciences, 19 (Spring), 295-305.

Wright, Peter and Frederick Barbour (1977), "Phased Decision Strategies: Sequels to an Initial Screening," Unpublished Research Paper (No. 353), Stanford University, Graduate School of Business.