Holistic Conjoint

Caroline M. Henderson, Dartmouth College
David J. Reibstein, University of Pennsylvania
ABSTRACT - Considerable research has focused on the role of product attributes in consumer preferences. Many approaches present the subjects with a set of attributes on which to rate available choices. This study contends that such a presentation heightens the salience of the attributes. An alternative procedure is presented where the subject is given a set of choices, but the underlying attributes under investigation are not made explicit. The nonstated attributes are orthogonally arrayed which allows for simple estimation of their partworths.
[ to cite ]:
Caroline M. Henderson and David J. Reibstein (1986) ,"Holistic Conjoint", in NA - Advances in Consumer Research Volume 13, eds. Richard J. Lutz, Provo, UT : Association for Consumer Research, Pages: 282-285.

Advances in Consumer Research Volume 13, 1986      Pages 282-285


Caroline M. Henderson, Dartmouth College

David J. Reibstein, University of Pennsylvania


Considerable research has focused on the role of product attributes in consumer preferences. Many approaches present the subjects with a set of attributes on which to rate available choices. This study contends that such a presentation heightens the salience of the attributes. An alternative procedure is presented where the subject is given a set of choices, but the underlying attributes under investigation are not made explicit. The nonstated attributes are orthogonally arrayed which allows for simple estimation of their partworths.


A critical issue in marketing is the determination of the importance of characteristics to consumers in their evaluation of choice alternatives. From such information, product and advertising decisions can be directed towards improving consumer's judgments of the products being offered by a firm. The task facing researchers in this field has been the measurement of the consumer importances associated with the product attributes. In this paper we present an alternative method for measuring product attribute importances without distorting the salience of the attributes through the data collection instrument.


Multi-Attribute Attitude Models

Throughout the late sixties and early seventies the marketing literature was filled with studies which focused on the evaluation of choices along a set of product attributes, and the assessment of their importance in forming an overall attitude or preference toward the choices. The origin of such work is often associated with Fishbein (1967) and Rosenberg (1956). Such efforts, founded in psychology, quickly evolved in the marketing literature, and were well reviewed by Wilkie and Pessemier (1973).

The underlying theory in this literature was that a consumer had an attitude toward an object which was based upon the object's intrinsic characteristics. The issue was then one of identifying the characteristics, evaluating the choices on these characteristics, weighting the characteristics, and then aggregating them (usually though some linear additive model) into an overall attitude or preference towards the object. The measurement task usually entailed the tedious respondent evaluation of each of the objects across each of the characteristics.

Aside from possible consumer fatigue, several possible problems arose (Wilkie and Pessemier, 1973). First was the identification of the correct set of underlying characteristics which led to overall choice. It was generally believed that this was best accomplished though management judgment and a pretest of consumers where the attributes were elicited. This presumed that the managers or the consumers would be aware of which characteristics were critical in leading to choice. Any unconscious attributes that were part of the evaluation would be overlooked by the method.

A second problem was termed "halo effects" (Wilkie, McCann and Reibstein, 1973; Beckwith and Lehmann, 1975). "Halo effects" refer to the respondents mapping their preferences back onto the attribute evaluations, thereby rating each of the favored choices highly on each of the attributes, and the disfavored choices poorly on each of the attributes, beyond their true levels. This implied that it was the affect which drove the beliefs and not the reverse. Such response made the importance weights associated with the attributes difficult to assess.

The initial process for evaluating the attribute importances was through a direct evaluation of the attributes (Lehmann, 1969). This, too, was problematic because a) consumers may be incapable of evaluating the true importance of a characteristic in their overall evaluation, b) consumers may be unwilling to confess the true weights because they may feel their process was irrational, i.e., the use of color in their assessment, c) they may not remember what characteristics were instrumental in their initial evaluation, d) there may be socially desirable answers which could make their answer "look better", i.e., fuel efficiency in their evaluation of automobiles, or e) the attribute may be salient, although not determinant, because of minimal variance across the choice alternatives

This difficulty led Bass and Wilkie (1977) to recommend deriving the importance weights via regression based on the preferences (the dependent measure) and the attribute beliefs (the independent measure).

Conjoint Measurement

An alternative approach, with the same underlying theory, evolved and gained in popularity throughout the mid to late seventies and into the eighties. This approach was conjoint measurement. Although not specifically designed for these purposes, conjoint measurement answered many of the problems in multi-attribute attitude models.

Again, the approach was to describe choices along a set of attributes and ask the respondent to rate, rank or choose between two choices based on these characteristics. (See Green and Srinavasan (1978) for a thorough review of conjoint measurement.) However, in conjoint, no consumer beliefs were measured, only the overall choice evaluation or preference. With a carefully balanced design of choice attributes, it was possible to assess the impact of each attribute on the ultimate evaluation using analysis of variance (either MONANOVA or ordinary least squares). The belief still remained that the overall evaluation was based on the underlying choice characteristics.

A major possible problem with this approach is the cueing of respondents created by the data collection approach. The respondents are explicitly told to evaluate the choices based on these characteristics. Generally, the products are nonidentified, or, more likely, hypothetical. Hence, the respondent can only evaluate the choices based on these characteristics. The respondent is usually told to ignore all other characteristics or to treat the choices as if they are equal on all nonstated attributes.

All that can be measured from a conjoint study is the relative importances of the attributes included in the exercise. It is unrealistic to expect to have included all of the relevant attributes. A method for measuring whether the relevant attributes have been included is to look at the individual model's fit or, if OLS is used, R . A low degree of fit would indicate either that the respondent is inconsistent or that the model is misspecified, i.e., some relevant attributes have been excluded. The inconsistency may even be the result of the truly determinant attributes being excluded.

Alternatively, in many conjoint studies the model's fit may be quite high; unfortunately, this may be the result of heightening the salience of the attributes on which the choices are described. This heightened salience may distort the natural Process of evaluation.

Holistic Evaluation

An alternative procedure, which we propose, is to present the choices to the subject, where the attributes are orthogonally arrayed, but the attributes are not directly explicated to the subject. The subject makes evaluative judgments about the choices from which the attribute importances or partworths can be described.

Holistic judgments are not new. One of the distinguishing characteristics about multidimensional scaling (MDS) techniques is that they are based on holistic or gestalt evaluations, from which the underlying dimensions can be derived.

Unlike HDS where the relevant dimensions are not known by the researcher until after the data are collected, in our approach the attributes will be known by the researcher, but not the subject. This is very similar to the economic technique of hedonic regression, or to psychophysical research. There are also several examples in the marketing literature.

Huber (1975) gave subjects iced tea to drink, compare, and evaluate. Presented in a full factorial design, the various iced teas differed in amounts of tea and sugar. Based on subject judgments, he was able to determine the ideal levels of sugar and tea and their importance in forming preference. Holbrook (1981) analyzed both perceptions and affect, in an integrative model of evaluative judgments, for presentation of a piece of classical music. Each presentation differed in tempo, rhythm, dynamics, and phrasing. Buchanan and Morrison (1984) have analyzed holistic judgments where subjects make judgments in blind taste tests, either in pairs or triplets, and select either which choice is preferred or different. From this they can test on individual's ability to discriminate, although they do not relate the judgments to the underlying attributes.


In many cases it is difficult, sometimes even impossible, to ask subjects to evaluate choices without describing the choices along some set of attributes. This is particularly true when one is dealing with hypothetical products which could neither be shown or even described by brand name. In such cases, the only alternative may be to use a conjoint design. However, in other cases, it is possible to present the choices either physically, visually, or verbally without making it explicit which attributes are under investigation.

We selected food products which could be presented to consumers without making the basic ingredients explicit. In particular, we were interested in the role of basic nutrients in food choice and preference formation. In this case, the respondents would have a low level of awareness of different nutrients, might not know how the nutrients influence their food preference, or, if aware, would know what is socially designated as appropriate behavior. There has been a considerable amount of publicity on "eating right. n All of these conditions would make it extremely difficult for a respondent to assess the importance of various food nutrients or to rank/rate their choice of foods described on a set of nutrients.

Hence, the subjects were given pairs of meals from which they would have to select their preferred meal. The meals were described in terms of the foods. not their underlying nutritional content.

Each respondent had to rate sixteen pairs of meals (32 meals in total). The pairs were constructed, with the help of a nutritionist, to represent an orthogonal array of six basic nutrients - protein, carbohydrates, fiber, fat, sodium, and cholesterol.

Respondents were handed a deck of sixteen cards, each containing one pair of meals. They were first asked to go through the deck, one card at a time, and to indicate which of the two meals they would personally prefer to eat on an individual taste basis. They then were asked to repeat the task with respect to which meal they would be most likely to eat at home, considering cost, time, etc. They repeated the task a third time, but this time they were to choose the meal that they thought was the most healthful. The three measures were intended to indicate preference, behavioral intentions, and nutritional perception or knowledge.

The study was conducted with a national probability sample of 700 consumers - participating in an on-going research project funded by the United States Department of Agriculture in a grant to the Marketing Science Institute (Schmalensee, et al., 1982). Subjects were selected as the "primary food purchaser and preparer in the home" and were 93% female. Each subject was interviewed in the home and also completed a lengthy leave-behind questionnaire.


Individual regression analyses were performed for each of the three dependent measures. Unlike a conjoint study or the Huber (1975) study, it was not necessary to dummy-variable code the individual factors, since they were on a continuous scale. Each observation was choice "A" meals, coded either a zero (if rejected) or a one (if selected), and the independent variables were the actual differences (grams, milligrams) between the meal pairs on each of the six nutrients. The "B" choices were excluded since they would simply provide redundant (mirrored images) information.

The choices for "prefer" and "prepare" were most similar, with a Pearson correlation coefficient of .70, while "nutritious" and "prefer" were least correlated with a correlation coefficient of .47.

As an example of the results, one individual's regressions are shown in Table 1, in this case, we may assume she "prefers" meals with higher fat content, with none of the other partworths being significant at the P<.10 level. However, when asked which meal she is actually likely to prepare at home for her family, a different selection pattern begins to emerge. It appears this person is more likely to "prepare" meals higher in protein (P<.01) and sodium (p<.01) and lower in carbohydrates (P<.01), fiber (p<.05) and cholesterol (p<.10). Her assessment of what she considers most nutritious is quite different. It should be noted that for this individual the R is significant for both the "prepare" and "nutritious" models.

Given that the respondents are not cued as to the underlying attributes, it would be expected that for many individuals the models would not be significant. This is different than what one generally finds for traditional conjoint models, where the subjects are specifically directed to make choices based on the dimensions of the models. The average R for the three models across the population was .41 for preference, .42 for "prepare" or behavioral intention, and .46 for "nutritious, n with the variance of the R 's equal to .17 for each model.

The average coefficients (partworths) for each of the three models are shown in Table 2. The mean values and the standard deviations across the population are shown in the first two columns. The next column is the percentage of individuals that had the coefficient significant (P<.10).



As can be seen, the percentage is relatively small for any one attribute. In fact, only 15.2% of the sample had two or more significant factors. This implies either that a) respondents base their evaluations on a particular nutrient, b) that the stimuli set (sixteen cards) was not large enough to detect underlying nutritional preferences without undue distortion by specific food likes and dislikes, or c) that there are a number of other factors which are not included in the model; that is, the respondents make their choices relying on a number of factors other than nutrients. This is totally realistic, and this approach does not force the evaluations to be based on the factors which are included.

It should be noted from Table 2 that the two attributes which are most commonly positive for the preference and behavioral intention models are protein and cholesterol, while fat, sodium and carbohydrate are negative. There apparently is some nutritional knowledge, as in the third model, the fiber dimension grows in importance. It would appear that consumers know fiber is good for them, they just don't like it nor intend to eat it. The results for cholesterol are also illuminating, as it appears to be a consistently positive predictor of all three dependent variables. This positive effect would be unlikely if consumers are specifically cued to cholesterol content. It would appear that consumers are either unable to detect high cholesterol foots or have natural tastes for such foods. This finding would be obscured in a traditional conjoint design.

In many cases it is essential that we identify which characteristics are important to consumers in their choice process. The question facing researchers is one of how to measure these importances. In some cases, it is difficult for respondents to indicate directly what is important to them. It may even be difficult for them to indicate not just the degree of importance, but even which attributes are used in their choice process. Describing choices along a set of dimensions may overly cue the respondents to give added weight to the selected dimensions.



We have presented a methodology which should ameliorate this problem. Rather than listing the attributes for product evaluation, the holistic choices themselves are presented. From these choices, the respondents make their selections. This provides the information from which the partworths can be derived.

In many cases the proposed approach would not be logical. When the product category has numerous potential/hypothetical products it would be impossible to present the subjects with the actual choices. Alternatively, a product category may contain far too few alternatives to allow the model to be specified. In many cases, however, the researcher is interested in the role of specific attributes which play a significant role, but may not be directly recognizable to the respondent. The example we chose was a nutritionist's concern about the role of basic nutrients - general characteristics unknown to the subject.

The potential with food products should be clear; but the opportunities are also apparent in other product categories. in the context of verbal v. pictorial presentation, Holbrook and Moore (1981) note that holistic evaluations may be appropriate if evaluative judgments depend on aesthetics, taste, symbolic meaning, or sensory experience. For example, fashion designers may be concerned about the influence of specific designs, fabrics colors, and patterns. Rather than describing alternative apparel which vary on these dimensions, thereby heightening their salience, specific pictorial or actual choices could be presented.

Future Direction

The next step in this research is to collect direct importance measures in a conjoint design which would allow weights to be derived. These weights could then be compared with weights derived from a holistic conjoint. This would allow one to determine if the weights were different, but would not serve as a basis for determining which set of weights were true.

What clearly remains, for all procedures, is the need to determine the validity of the various alternative measures. Future research needs to collect behavioral data with real stimuli that can be compared with responses predicted by sets of conjoint weights.


Bass, Frank M., and William L. Wilkie (1973), "A Comparative Analysis of Attitudinal Predictions of Brand Preferences," Journal of Marketing Research, Vol. 10, August, pp. 262-69.

Beckwith, N.E. and D.R. Lehmann (1975), "The Importance of Halo Effects in Multi-Attribute Attitude Models, n Journal of Marketing Research, Vol. XII, August, 265-75.

Buchanan, B.S. and D.G. Morrison (1984), "Measuring Simple Preferences: An Approach to Blind, Forced Choice Product Testing," Marketing Science, Vol. 4 #2, Spring, 93-109.

Fishbein, M., ed. (1967), Reading in Attitude Theory and Measurement, Reading, MA: Addison-Wesley.

Green, P.E. and V. Srinivasan (1978), "Conjoint Analysis in Consumer Research: Issues and Outlook," Journal of Consumer Research, Vol. 5, September, 103-23.

Holbrook, M.B. (1981), "Integrating Compositional and Decompositional Analysis to Represent the Intervening Role of Perceptions in Evaluative Judgments," Journal of Marketing Research, XVIII, February, 13-28.

Holbrook, M.B. and W.L. Moore (1981), "Feature Interactions in Consumer Judgments of Verbal Versus Pictorial Presentations," Journal of Consumer Research, Vol. 8, 103-13.

Huber, J. (1975), "Predicting Preferences on Experimental Bundles of Attributes: A Comparison of Models," Journal of Marketing Research, Vol. X, August, 290-99.

Lehmann, Donald R. (1969), "Choice Among Similar Alternatives: An Application of a Model of Individual Preference to the Selection of Television Shows by Viewers," Unpublished Doctoral Dissertation, Purdue University.

Rosenberg, M.J. (1956), "Cognitive Structure and Attitudinal Affect", Journal of Abnormal and Social Psychology, 53, 367-72.

Schmalensee, D.H., C.M. Henderson, A.G. Clayton, E. Haas, R. Leonard, J. Quelch, L. Smith, and J. Williams-Jones (1982), Determinants of Food Consumption in American Households, Cambridge, MA: Marketing Science Institute.

Wilkie, W.L. and E.A. Pessemier (1973), "Issues in Marketing's Use of Multi-Attribute Attitude Models," Journal of Marketing Research, Vol. X, November, 428-41.

Wilkie, William L., John M. McCann, and David J. Reibstein (1973), "Halo Effects in Brand Belief Measurement: Implications for Attitude Model Development," Proceedings, Fourth Annual Conference, Association for Consumer Research, November.