Do the Parameters of Choice Models Depend on Differences in Stimulus Presentation: Visual Versus Verbal Presentation?

Jordan J. Louviere, University of Alberta
Herb Schroeder, North Central Forest Experiment Station, U.S. Forest Service
Cathy H. Louviere, Data Sciences (Associates) of North America
George G. Woodworth, University of Iowa
ABSTRACT - The purpose of this paper is to compare the parameters of conjoint choice models derived from choice tasks in which some of the attributes are described either verbally or visually. A task involving choice of state parks was chosen to test the hypothesis because parks have a substantial visual component such as scenicness that should influence choice. Results suggest little statistical difference in the parameters of models developed from either type of information. Because considerable effort and expense is necessary to insure comparability of verbal and visual attribute descriptions, the research suggests reliance on visual wherever possible.
[ to cite ]:
Jordan J. Louviere, Herb Schroeder, Cathy H. Louviere, and George G. Woodworth (1987) ,"Do the Parameters of Choice Models Depend on Differences in Stimulus Presentation: Visual Versus Verbal Presentation?", in NA - Advances in Consumer Research Volume 14, eds. Melanie Wallendorf and Paul Anderson, Provo, UT : Association for Consumer Research, Pages: 79-82.

Advances in Consumer Research Volume 14, 1987      Pages 79-82

DO THE PARAMETERS OF CHOICE MODELS DEPEND ON DIFFERENCES IN STIMULUS PRESENTATION: VISUAL VERSUS VERBAL PRESENTATION?

Jordan J. Louviere, University of Alberta

Herb Schroeder, North Central Forest Experiment Station, U.S. Forest Service

Cathy H. Louviere, Data Sciences (Associates) of North America

George G. Woodworth, University of Iowa

ABSTRACT -

The purpose of this paper is to compare the parameters of conjoint choice models derived from choice tasks in which some of the attributes are described either verbally or visually. A task involving choice of state parks was chosen to test the hypothesis because parks have a substantial visual component such as scenicness that should influence choice. Results suggest little statistical difference in the parameters of models developed from either type of information. Because considerable effort and expense is necessary to insure comparability of verbal and visual attribute descriptions, the research suggests reliance on visual wherever possible.

INTRODUCTION

An issue of considerable importance to conjoint analysis in general, and to consumer choice modeling in particular, is whether one obtains the same results using visual as opposed to verbal representations of choice alternatives. Our review of the available literature discloses little more than anecdotal evidence: for example, Green and Srinivasan (1978) report no evidence that visual and verbal stimulus displays yield the same results, although they do state that, "Alpert, Betak and Golden (1978) report that a combination of pictures and words produced roughly the same results as the purely verbal approach..." (p.111). The reader is left to decide what "roughly" means. The single major study we could find by Holbrook and Moore (1981) also compared the outcomes of conjoint tasks using verbal and visual descriptions of sweaters. However, our understanding of their experiment indicates to us that they made no attempt to insure that the words used corresponded to the drawings. Thus, their conclusions regarding processing differences may have been due to the lack of correspondence. In this paper we address this problem by going to considerable lengths to insure verbal and visual stimulus correspondence.

Of course, there has been considerable research into the effects of pictorial presentations of information in advertising and other contexts (See, e.g., Shanteau and Nagy 1979; Edell and Staelin 1983). This latter research has demonstrated that pictures interact with words to produce effects that are significantly larger than words alone. Thus, despite the vast amount of research available on consumer response to conjoint stimuli, we have been unable to find any direct comparisons of model or process differences that arise from providing information verbally versus pictorially.

A direct comparison of verbal and pictorial attribute representations is important because information about attribute levels is often verbally expressed, or expressed as a combination of verbal and visual information in commercial as well as academic applications. Furthermore, because conjoint tasks are often used to test academic hypotheses, the content of the attribute information may affect the generalizability of the findings. Thus, as with other context effects in conjoint tasks (see, e.g., Huber and Sheluga 1980; Meyer and Eagle 1982), generality is an issue. The purpose of this paper, therefore, is to examine whether there are statistically significant differences in the parameters (and their associated inferences) of conjoint choice models developed from statistically identical tasks that differ only in the mode of presentation of the attribute information used to convey stimulus levels in conjoint choice tasks.

We concentrate on a choice task in order to eliminate the possibility that scale effects could account for the results. We also use a task that should involve a highly visual component in order to create a "worst case" example: if we fail to find differences in this kind of task, there would be less concern about other less visually involved tasks. Hence, we concentrate on consumer choice among state parks because state parks are often selected by consumers on the basis of their scenicness and environmental distinctiveness. These latter elements should make it relatively difficult for verbal descriptions to provide comparable results to visual descriptions, however, our results will show that such was not the case.

THE PARK CHOICE EXPERIMENT

Based on previous research into the park choice behavior of consumers funded by the U.S. Forest Service (Curry, et al. 1983; Louviere and Woodworth 1984), a set of park choice decision factors was compiled. These factors were shown to significantly influence consumer choices among parks in previous empirical work using conjoint choice tasks. As well, the predictions of the choice models derived from such tasks were consistent with the actual park choices that the surveyed consumers were observed to make (Louviere and Woodworth 1984). Questions, however, were raised concerning whether consumers' mental images of the verbal descriptions were isomorphic to visual images.

In particular, U.S. Forest Service researchers strongly felt that visual descriptions of terrain, vegetation and bodies of water were important to add realism to park choice tasks, and therefore the results of the verbal tasks were suspect despite good predictive ability. A key research hypothesis, therefore, is that there are no differences in the parameters of choice models derived from separate park choice tasks in which visual and verbal attribute descriptions were separately manipulated.

A test of the research hypothesis focuses on the three decision attributes mentioned earlier as being of particular concern to the U.S. Forest Service researchers: terrain, vegetation density and water bodies. In order to test the research hypothesis, a conjoint choice task involving hypothetical state parks was developed. The choice task was structured as follows: (a) Each of the three attributes of research interest were assigned visual or verbal levels as described in Table 1, and (b). These attributes were combined with other park choice attributes (Table 1) to produce choice alternatives. (Refer to Table 1).

The visual levels of the three key attributes involved in the test of visual versus verbal hypothesis were selected by the research team from a very large library of color photographs possessed by a member of the U.S. Forest Service research team. Terrain and vegetation tensity photographs were selected so as to provide perceptually significant monotonic gradations as can be noted in the corresponding Table 1 verbal descriptions. Photographs of creeks, rivers, lakes and ponds were selected to be typical of such features in the Midwest U.S. (where the study was conducted), and so that consumers would closely agree about which was which.

Candidate lists of photographs were first presented to a convenience sample of 125 consumers recruited from a local shopping mall in a Midwestern city. The screening criterion was that the interviewee had visited a park ln the last 2-3 years. Interviewees were shown candidate park scene photographs of the three key attribute dimensions that had been enlarged and reproduced using a color xerox process. They were asked to provide verbal descriptions of each photograph as if they were describing the scene in the picture to someone who hat not seen it.

TABLE 1

ATTRIBUTES AND LEVELS USED IN THE PARK CHOICE TASKS

The verbal descriptions were content analyzed, and a frequency distribution of word and phrase usage was developed. Three photographs were selected to represent the levels of each attribute based on the following arbitrary, but (we believe) reasonable, criteria: (a) the photographs evoked verbal descriptions that connoted a significant monotonic difference between the levels of terrain and vegetation density, or a recognizable and correct differentiation in the case of the water body pictures; and (b) the interviewees appeared to agree reasonably well in their use of words and phrases to describe these pictures as indicated by the results of the content analyses.

As a further check, the final 10 photographs (three for terrain and vegetation density, and four for water bodies) were enlarged and reproduced by color xerox methods. These reproductions were shown to 25 undergraduate students at a Midwestern university, who were asked to match each photograph with a list of 10 candidate verbal descriptions. The verbal descriptions included the ones assembled by the research team from the content analysis of the verbal descriptions given by the previous interviewees, as well as 'ringers' consisting of combinations of similar words and phrases deliberately constructed to be inaccurate in some significant way with respect to the target photographs representing each attribute.

The 25 students correctly matched the photographs with the verbal descriptions developed by the research team over 85% of the time (low of 76%, high of 92%). We concluded, therefore, that the verbal descriptions were as close as possible to matching the visual scenes given the time and resource constraints involved. In fact, we suggest that it is unlikely that many research projects have gone to such lengths to insure a match between visual and verbal descriptions. The purpose of these efforts was to insure to the extent possible that we could attribute any differences in choice model parameters to visual vs. verbal differences not experimenter ineptitude or laziness. That is, we wished to try to rule out differences due to poor matches in verbal descriptions, thus giving verbal descriptions best possible chance to produce same results.

The experimental design was a near orthogonal main effects plan derived from a 4 x 34 x 28 factorial involving 40 treatments. The 40 treatments were placed into 20 sets of pairs using a balanced incomplete block procedure. Each treatment in a pair was placed on a separate 8 1/2 x 11 inch sheet of standard typewriter paper, and the sheets were assembled into a looseleaf notebook such that when the notebook was open the pairs faced each other on opposite pages. The designs for visual and verbal information were identical in all respects, except that descriptions of the three key attributes was displayed in color xerox reproductions in the visual condition, and in words in the verbal condition.

Ten different notebooks for each task were assembled for use in ten sampled state parks. The order of the pairs was identical in each notebook to control for differences due to order effects and interviewers. Thus, the order of appearance of the pairs in notebooks was random, but in every notebook the pairs appeared in the same random order. Subjects were 224 individuals intercepted in ten different state parks located in the center of a Midwestern state. A commercial research firm was contracted to obtain no less than 20 interviews in each of the parks.

Color-coded response sheets were prepared for subjects to use to respond to the choice task in each correspondingly colored notebook (white for visual, blue for verbal). In addition to the choice task, selected demographic and park visitation questions were asked after subjects completed the choice task in order to permit comparisons of the sample with those involved in previous park choice research and to allow segments to be developed and compared. The color-coded response sheets were randomly ordered and given to each interviewer. Interviewers were instructed to have each subject complete the task in the notebook that had the same color as the topmost response sheet. In this way, it was possible to control for differences in the way interviewers recruited subjects, differences in the types of subjects recruited and differences due to the parks in which the interviews took place.

Thus, regardless of interviewer or park, respondents were randomly assigned into either verbal or visual tasks by virtue of the random assignment of response sheets. 101 completed interviews were obtained in the verbal task condition and 123 in the visual task condition. Interviewers reported that approximately one percent of individuals approached to do the interview refused to participate in the task, which took an average of 20-30 minutes to complete.

Analytical interest centers entirely on the aggregate choice responses in the two task conditions. The two sets of data were combined to test the null hypothesis of no significant differences: a common set of main effects was fit to the choice data based on the experimental attributes, and the interaction of each main effect with a dummy variable representing the two different tasks was used to test the null hypothesis. Multinominal Logit Regression was used to analyze the choice data employing the generalized least squares regression procedure suggested by Louviere and Woodworth (1983). The first-pass vector of parameters was operated upon by iteratively reweighting and updating the parameter vector to obtain the maximum likelihood estimates (Woodworth and Louviere 1985). The results are given below.

RESULTS AND CONCLUSIONS

The critical statistical tests for the null hypothesis center on the interaction effects representing departures from the hypothesis that the estimates are the same in the two task conditions except for sampling fluctuations. The results, contained in Table 2, surprisingly indicate that only two of the seven interaction parameters that bear on the null hypothesis approach conventional significance levels. In fact, only the pond interaction parameter is significant at conventional levels; improved camping is also significant but it has no bearing on the null hypothesis.

Thus, we tentatively conclude that there is little difference in the choice model results due to the differences in verbal versus visual representation of attribute levels. Our conclusion is tentative because our results may be due to pooling subjects who reacted differently to the different levels of attributes in the two tasks, but whose differences averaged out in the aggregation. We could not test this eventuality because the subjects participated in one or the other of the two tasks, making it impossible to group them into comparable subsets on which separate tests could be conducted. If all subjects had participated in both tasks, comparable tests could have been performed, but the length of the interview precluded this.

Additionally, it should be noted that the effects in Table 2 associated with the visual vs verbal terms are small. Thus, retaining the null associated with the such small differences may simply be an artifact of such small statistical effects in the aggregate. We grouped individuals from both conditions (verbal and visual) in an attempt to see whether the aggregation effects were the likely source of the small effects. Recognizing that we cannot disassociate the effect of conditions, these analyses suggest considerable between group differences in responses to the three key decision variables of interest. Thus, we may be understating the differences quite considerably.

Because our study design could not rule out the aggregation hypothesis to account for the results, we normally would suggest that future work be directed toward testing whether aggregation could be ruled out. However, in this case, we believe that practicality argues against pursuing such a course. In particular, even if we assume that our results are correct (that is, there are only minor differences in verbal compared with visual representations of stimuli), the amount of effort required to insure that verbal descriptions correspond to visual is considerable.

Thus, we would conclude that both academic and practical researchers should seriously consider the benefits of visual stimulus presentations in terms of realism and accuracy. We recognize that the use of visual information in some instances may escalate research costs. However, the time and effort involved in developing appropriate verbal descriptions may considerably outweigh the additional costs of preparing and administering visual materials. Furthermore, one must have the visual materials first in order to develop the verbal materials; thus, we can see few instances in which researchers would want to use verbal information if there are questions about the ability of the verbal descriptions to convey the same information.

TABLE 2

STATISTICAL RESULTS PERTINENT TO THE NULL HYPOTHESIS

REFERENCES

Alpert, Mark I., Betak, John F., and Golden, Linda L. (1978), "Data Gathering Issues In Conjoint Measurement," Working Paper, Graduate School of Business, University of Texas. Austin.

Curry, Davit J., Louviere, Jordan J., Rausch, Dennis and Woodworth, George G. (1983), "Recreation Destination Choice," Final Report No. 35, Institute of Urban and Regional Research, The University of Iowa, Oakdale, Iowa.

Edell, Julie A., and Richard Staelin (1983), "The Information Processing of Pictures in Print Advertising," Journal of Consumer Research, 10 (1), 45-61.

Green, Paul E. and V. Srinivasan (1978), "Conjoint Analysis in Consumer Research: Issues and Outlook," Journal of Consumer Research, 5, 103-123.

Holbrook, Morris B. and William L. Moore (1981) ,"Feature Interactions In Consumer Judgements of Verbal Versus Pictorial Presentations," Journal of Consumer Research, 8 (1), 103-113.

Huber, Joel and Davit Sheluga (1980), "The Effect of Pair Similarity on Dollarmetric Profile Comparisons," Advances In Consumer Research, 7, 134-139.

Louviere, Jordan J. and Woodworth, George G. (1983), "Design and Analysis of Simulated Consumer Choice or Allocation Experiments: An Approach Based On Aggregate Data," Journal of Marketing Research, 20, 350-367.

Louviere, Jordan J. and Woodworth, George G. (1984), "Models of Park Choice Derived From Experimental and Observational Data: A Case Study In Johnson County, Iowa" Unpublished Final Report to the U.S. Forest Service, North Central Forest Experiment Station, Chicago, Illinois (available from the authors on request).

Meyer, Robert J. and Eagle, Thomas C. (1982), "Context Induced Parameter Instability In A Disaggregate-Stochastic Model of Store Choice," Journal Of Marketing Research, 3, 62-71.

Shanteau, James and Nagy, Geraldine F. (1979), "Probability of Acceptance in Dating Choice," Journal of Personality and Social Psychology, 37 (4), 522-533.

Woodworth, George G. and Louviere, Jordan J. (1985), "Fitting Large Multinomial Logit Models Using Iteratively Reweighted Least Squares," Paper presented to the Annual ORSA/TIMS Marketing Science Meetings in Nashville, TN (March).

----------------------------------------