A Quick and Simple Procedure For Benefit Segmentation Based on Conjoint Experiments

Citation:

Jagdish Agrawal and Wagner A. Kamakura (1987) ,"A Quick and Simple Procedure For Benefit Segmentation Based on Conjoint Experiments", in NA - Advances in Consumer Research Volume 14, eds. Melanie Wallendorf and Paul Anderson, Provo, UT : Association for Consumer Research, Pages: 567.

Advances in Consumer Research Volume 14, 1987      Page 567

A QUICK AND SIMPLE PROCEDURE FOR BENEFIT SEGMENTATION BASED ON CONJOINT EXPERIMENTS

Jagdish Agrawal, State University of New York at Buffalo

Wagner A. Kamakura, State University of New York at Buffalo

The traditional approach for benefit-segmentation based on conjoint analysis has been to estimate part-worths for each subject, and then to form clusters based on these subject-level estimates (Green, Wilt and Jain 1972, Moore 1980). Since fractionated designs are often used in the collection of preference data (Green and Srinivasan 1978), the estimates of part worths at the subject-level are usually based on relatively scarce data which can make the estimates quite sensitive to errors in the data.

More recently, Hagerty (1985) proposed a Q-Factor-Analytic procedure which provides direct segment-level estimates of part-worths, starting from the observed preference ratings of stimulus profiles. This procedure groups the subjects based on the stated preferences, eliminating one step in the segmentation process. However, Stewart (1981) demonstrates that in O-Factor clustering a factor does not represent a cluster and that, in fact, a solution with a single factor may contain more than one cluster, which would make the identification of segments rather complex. The proposed simple procedure also uses stated preferences for clustering subjects into segments. However, unlike Hagerty's (1985) Q-Factor analytic procedure, a complete linkage algorithm is used to cluster subjects according to the similarity of their stated preference rankings, leading to non-overlapping hierarchical segments.

The Proposed Procedure

We assume that when presented wit h the stimulus profiles, each subject assigns utilities to each, according to the attributes it possesses and the value El/he assigns to each attribute level. This process, however, involves some random error, so that the utilities assigned at each occasion are samplings from distributions of unobservable "true" utilities. In reality, since the subjects provide preference rankings for the stimulus profiles, one can only observe the ranking of a sample of these unobservable "true" utilities. Our objective is to cluster subjects into homogeneous segments according to the similarity in their stated preferences.

Empirical Comparisons

In order to segment the sample according to the preference-based procedure, the correlation between the preference rankings of every pair of subjects was computed, generating a distance matrix. A complete linkage clustering procedure was then a plied to the distance matrix, providing the benefit segments. For the procedure based on part-worth estimates, subject-level part-worths were first estimated and transformed into standardized ranges. These standardized ranges indicate the relative importance of each attribute. A minimum-variance clustering procedure (Howard and Harris 1966) was then used to form the benefit segments. Multinomial Logit was used for the estimation of subject-level and segment-level part-worths.

In attaining the classification objective, we first used a simulation experiment in which preference data were generated from known "true" utility functions plus a random disturbance. In order to emulate different error conditions, the standard deviation of the random disturbances was manipulated in three different simulations. In the first condition (no error), the average rank correlation between the observed (generated) preference rankings and the "true" interval-scaled utilities was equal to 0.993 The average rank correlation for the other two error conditions (*=1 and *=2 ) were 0.770 and 0.616, respectively. The two segmentation procedures were then applied to these data, to measure the extent to which subjects with identical "true" utility functions were classified into the sane segments. In the absence of error (*=0), both the procedures correctly reproduced the actual membership (100%) in each of the three segments. When *=1, the preference based procedure again correctly classified all the 90 hypothetical subjects to their respective a priori established segments. But the performance of the part-worth procedure decreased as indicated by the measure of fit (Cramer's V = 0.770) and percentage of classification (67.82) At a higher error condition (*=2), even though the preference based procedure could not repeat its high performance, its overall performance was still better (Cramer's V = .643) than that of the part-worth procedure (Cramer's V = .332).

Data from an actual conjoint experiment were also used for another empirical comparison. In the experiment, a sample of 100 subjects evaluated 27 profiles forming a fractional factorial design of 5 attributes at three levels each. Four segments were identified by both segmentation procedures. The "average" part-worths obtained by the preference-based procedure for each segment fit best to the preference rankings of each segment member (Kendall's Tau .62 vs .56) This superiority is significant (at the 02 level) Similar conclusions would be drawn by comparing the cumulative log-likelihoods (-5205 for the preference-based clustering, versus - 5392 for the part worths-based clustering).

These results, are not presented to support a claim of general superiority for the proposed method over the traditional method due to the lack of a formal proof. However, the findings tend to reject the hypothesis that the classic benefit segmentation procedure based on subject level part-worths estimates is always better than the simpler procedure based directly on preferences.

REFERENCES

Gensch, D H. (1985), "Empirically Testing a Disaggregate Choice Model for Segments," Journal of Marketing Research, 22 (November), 462-467.

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

Green, P.E., Y. Wind and A.K. Jain (1972), "Benefit Bundle Analysis," Journal of Advertising Research, 12, 31-36.

Hagerty, M.B. (1985), "Improving the Predictive Power of Conjoint Analysis The Use of Factor Analysis and Cluster Analysis," Journal of Marketing Research, 22, 168-184.

Moore, V.L. (1980), "Levels of Aggregation in Conjoint Analysis: An Empirical Comparisons" Journal of Marketing Research, 17 (November), 516-523.

Stewart, D (1981), "The Application and Misapplication of Factor Analysis in Marketing Research," Journal of Marketing Research, 18 (February), 51-62.

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