A Comparative Analysis of Attitudinal Predictions of Brand Preference

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 10

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/271778072

A Comparative Analysis of Attitudinal Predictions of Brand Preference

Article in Journal of Marketing Research · August 1973


DOI: 10.2307/3149693

CITATIONS READS

129 339

2 authors, including:

Willam L Wilkie
University of Notre Dame
73 PUBLICATIONS 5,408 CITATIONS

SEE PROFILE

All content following this page was uploaded by Willam L Wilkie on 09 March 2015.

The user has requested enhancement of the downloaded file.


FRANK M. BASS and WILLIAM L. WILKIE*

A comparison of cross-sectional methods of analysis of multi-attribute attitude


models indicates striking differences in predictive power. Importance weights do
not detract from prediction, and correlations of attitude with preference com-
pare favorably with attitude-affect correlations found in social psychology.

A Comparative Analysis of Attitudinal Predictions


of Brand Preference

It is now generally accepted that attitudinal measures the probabilities of obtaining these satisfactions through
provide useful predictions of brand preference and such behavior. While the original proposals of Rosen-
choice. There is much uncertainty, however, about the berg [24] and Fishbein [13] in social psychology have
most appropriate form and components of attitude mod- provided impetus for initial marketing work, consider-
els and methods of testing and comparing models. For able extensions and alterations of these theories have
example, the need to include "importance weights" in been undertaken to better reflect marketing's consumer
multi-attribute models has been questioned. Methodo- context. The most significant changes involve:
logically, some studies have used cross-sectional infor- 1. Specification of attributes (satisfactions) which
mation in making predictions of preference, while others are idiosyncratic to product class perceptions,
have utilized information from individual consumers to 2. Use of "importance" ratings for each attribute
predict preference order for brands for individual con- by each individual to allow for differential
sumers. In this article we shall compare models and stresses in satisfactions sought rather than eval-
methods of analysis. We shall show that a few simple uations of whether or not the attribute is de-
changes in cross-sectional analysis methods provided a sirable,
160% increase in r^ over that obtained from a common 3. Simultaneous consideration of competing brands
method, and revealed contributions of importance as alternative attitude objects (behaviors),
weights which were not apparent in the earlier form. We 4. Measurements of "brand beliefs" on each attri-
also provide a comparison of cross-sectional and indi- bute in terms of degree of satisfaction expected
vidual preference predictions, rather than probability of attainment,
5. A resultant capability of "within-individual"
analysis as opposed to the former necessity of
MULTI-ATTRIBUTE ATTITUDE MODELS IN
cross-sectional methodologies.
MARKETING
Initial applications of multi-attribute attitude models
Much of the recent marketing research on attitudinal in marketing were basically concerned with investigat-
variables has concentrated on multi-attribute models ing the potentials of this approach in explaining brand
that transform consumer cognitions of brand offerings on preference and choice. All articles reported strong rela-
several dimensions to unidimensional measures of brand tionships and urged further development of the model
affect. These models are typically based on "expectancy" [2, 4, 14, 15, 19, 25, 27, 33]. Subsequent research, while
theories which assert that an individual's predispositions continuing to report positive results, has stressed modifi-
to given behaviors are governed by the set of satisfac- cations in model form, measurement, and analysis [ 5 , 8 ,
tions presumed to flow from such behavior together with 9, 10, 12, 17, 23, 29]. Recent reviews of these studies
show that researchers differ in their conceptualizations,
* Frank M, Bass is Professor of Industrial Administration and semantics, measurements, models, and analysis; signifi-
William L, Wilkie is Assistant Professor of Industrial Manage- cant issues remain for future research [21, 23, 31].
ment, Krannert Graduate School of Industrial Administration, The most debated of these issues concerns the extent
Purdue University, This research was supported by a grant to which "importance weights" add to the power of the
from the A,A,A,A, Educational Foundation,
model as contrasted with a simpler version using only
262

Journal of Marketing Research,


Vol, X (August 1973), 262-9
COMPARATIVE ANALYSIS OF ATTITUDINAL PREDICTIONS 263

brand beliefs. This article reviews conflicting results of in purpose between attitude theories developed in social
past research, investigates possible causes of the differ- psychology and brand preference studies suggest dif-
ences, and shows the significance of analytical method- ferent methods of analysis. In social psychology the pur-
ology in affecting results and conclusions of attitudinal pose, traditionally, has been to study the attitudes of
studies. different people for a single object. The analysis, of
necessity, was cross-sectional in character. Many market-
THE PROBLEM ing studies have also correlated the preference ranking
for a brand with the attitude toward that brand where the
Model Definitions correlations were computed across people. It is well-
One basic model utilized in marketing studies is de- known that interpersonal utility cannot be given rigorous
fined as: meaning; hence cross-sectional comparisons of attitude
with preference ranking for brands or with independent
(1) iijk measures of affect must be viewed with skepticism.
Cross-sectional studies which fail to adjust for within-
where: subject variance are especially vulnerable to criticism,
/ = attribute or product characteristic, since the interpersonal comparison problem, as we shall
/ = brand, show, is magnified in these cases. In marketing studies,
k = consumer or respondent, preferences and attitudes are obtained for each of sev-
such that: eral brands for each individual. It is possible in these
Aji: = consumer K's attitude score for brand /, studies to confine the analysis to individuals by using
Ijk = the importance weight given attribute / by the attitudes of each to predict the brand preference
consumer K, order of these same individuals. This method of analysis
avoids the problem of interpersonal utility comparison.
Biji- = consumer K'% belief as to the extent to
Both methods, properly applied, provide reasonably
which attribute / is offered by brand /.
strong predictions.
A simpler alternative (beliefs-only) model is defined as:
n
Inclusion of Importance Weights
(2) Ajk = y ] Bi k
The essence of the importance weight controversy
concerns whether or not model (2) outperforms model
Methods of Analysis
(1), The issue was originally raised in a different form by
Methods of analysis of attitudinal data have been re- Rosenberg, who noted that significant chi-square rela-
stricted to a comparison of predicted preference or
choice with stated preference or choice. Thus models are
Table 1
"tested" on a correlational basis. Unfortunately, it has^
not been possible to test attitude models on a scientific BRANDS AND PRODUCT ATTRIBUTES
basis, since a falsifiable attitude model has not yet been Brassieres Lipstick Mouthwash
developed. There has been a tendency to assume that
Comfort Color Color
better predictions imply better models. Although we Fit Container Effectiveness
question this assumption, we shall show, for whatever it Life Creaminess Kills germs
is worth, that predictions of brand preference with cross- Price Prestige factor Price
sectional data compare favorably with attitude-affect Style Taste/flavor Taste/flavor
correlations found in social psychology, Cohen, Fishbein, Lovable Avon Cepacol
and Ahtola [11], in commenting on Sheth-Talarzyk Maidenform Coty Colgate 100
[29] cross-sectional predictions of preference, state that Penneys Hazel Bishop Lavoris
the low r^'s found by Sheth and Talarzyk in comparison Playtex Max Factor Listerine
to correlation of .7 found with the Fishbein model [1] Sears Revlon Micrin
applied in a nonmarketing context suggests that the Orange juice (frozen) Toilet tissue Toothpaste
measures used in the Sheth-Talarzyk study are inferior
Nutritional value Color Decay prevention
to those employed in social psychology. We demon- Packaging Package size Freshens mouth
strate here that the data employed in the Sheth-Talarzyk Price Price Price
study, when properly adjusted, do in fact produce cor- Taste/flavor Strength Taste/flavor
relations which compare favorably with .7. Texture Texture Whitens teeth
Although we believe that individual analysis is a more A &P Aurora Colgate
appropriate basis for evaluating attitude models than Birds Eye Charmin Crest
cross-sectional methods, we focus attention in this Minute Maid Delsey Gleem
article on providing comparisons among a variety of Snow Crop Northern Macleans
Sunkist Scott Pepsodent
cross-sectional procedures and models. The differences
264 JOURNAL OF MARKETING RESEARCH, AUGUST 1973

tionships with an independent measure of affect were dryers), Lehmann by individual's Spearman rank cor-
found with model (1), model (2), and with a 2 Inc model relations (television shows), and Bass, Pessemier, and
structure. He concluded that importance represents Lehmann in terms of both Spearman rank correlations
"a distinguishable and important dimension of attitude- and percent correct predictions (soft drinks) [5, 15, 19].
related cognitive structures" [24]. Sheth's review of Ginter and Winter each found similar evidence favoring
Rosenberg's article noted, however, that for comparable model (1) in individual-level regressions for household
sets of attributes model (2) (beliefs-only) obtained higher products [14, 33]. In none of these studies, however, are
chi-square values than model (1) (beliefs times im- the differences in favor of model (1) as large as those re-
portance), which suggests that importance weights need ported earlier in support of model (2). It is clear that
not be included in summative attitude models [27]. This importance weights do not add strikingly to the ex-
conclusion has stimulated a number of empirical studies planatory power of the simpler beliefs-only structure
investigating the predictive power of model (1) versus of model (2).' The remaining controversy concerns
model (2). whether or not they detract and possible reasons for this
Three papers conclude that the inclusion of im- finding.
portance weights decreases predictive power. Sheth and A number of explanations have been advanced as to
Talarzyk report multiple regressions of brand preference the lack or low level of added explanatory power with
against alternative attitude model scores for five brands importance weights. Sheth and Talarzyk note that pos-
in each of six frequently purchased product categories sible causes of their suppression results could be:
[29]. In all thirty comparisons, model (2) yields con- 1. Importance weights being specific to product
siderably higher r^ than model (1); average r^ for beliefs- class rather than brands,
only equals .15 versus r^ of .04 for behefs times im- 2. Use of ordinal data for their dependent prefer-
portance. The authors conclude "there is a clear im- ence variable,
plication that (importance) not only does not contribute 3. Weighting and summing procedures which
toward the determination of the consumer's affect toward might remove variations in the independent
a brand, but also that it suppresses the determinant variables,
power of the (beliefs)" [29]. Lutz and Howard report a 4. Respondents implicitly incorporating impor-
similar result in their canonical correlation analysis of tance into their belief ratings,
instant breakfast preferences; "Beliefs-only measures 5. Use of cross-sectional regression analysis when
correlated more strongly (.92) with preference than did heterogeneity in the sample might be expected
the set of B-I measures (.81), for the first canonical [29].
variate, and similarly for the second variate (.47 to .41)" Beckwith and Lehmann show, with respect to point 4
[20]. Moinpour and MacLachlan, in a study of anal- above, that respondents do tend to spread beliefs ratings
gesics, likewise found that model (2) performed signifi- more on attributes rated as important [6]. Scott and
cantly better in seven of nine brand regressions, as well Bennett show, with respect to point 5 above, that prior
as providing somewhat superior results in Spearman clustering of respondents into two more homogeneous
rank and multidimensional scaling analysis [21]. segments leads to significantly different coefficients and
Seven papers conclude that there is little difference improved r^ for separate segment regressions as opposed
between models. Hansen and Bolland found model (2) to a single regression over the total sample [26]. Hughes
to outperform model (1) in individual predictions of first posits that the importance and beliefs ratings were highly
choice in beer halls, but model (1) superior for car correlated because of the types of scales employed,
washes [16], Cohen and Houston report very similar and suggests that each component be transformed to
r^ for each of six toothpaste brand regressions, while mean zero and included in a stepwise regression to ex-
multiple discriminant analysis on the same data pre- amine the explanatory contributions of beliefs, im-
sented by Cohen and Ahtola show very little differences portance weights, and their interaction term [17]. Cohen
in function significance, variation explained, and in per- and Ahtola point out that importance is often measured
cent correct classifications [8, 9]. Minor differences are in terms of the respondent designing an ideal brand, and
also found in the cross-sectional regression analyses of show evidence that this can lead to high mean impor-
Scott and Bennett [26] (industrial product), Hughes and tance ratings with little variance available for explaining
Guerrero [18] (automobiles), Churchill [7] (fountain variations in the dependent variable [8]. Wilkie and
pens), in the Spearman rank correlations presented by McCann discuss the negative impacts of correlated be-
Wilkie and Weinreich [32] (supermarkets), and in indi- liefs ratings across attributes; perfectly correlated be-
vidual-level regressions run by Beckwith and Lehmann liefs would allow no additional explanation through im-
(television shows) [6]. portance weights [30]. This point is also made by
Five studies conclude that inclusion of importance ' This statement applies only within the context of the meth-
weights increases prediction. Hansen bases this con- ods being compared here. Analysis of segments probably will
clusion on mean score differences between chosen and indicate the significance of importance weights for different
rejected alternatives (for restaurants, travel, and hair segments [2].
COMPARATIVE ANALYSIS OF AHITUDINAL PREDICTIONS
265
Beckwith and Lehmann who further suggest that model tional sample of 2000 female heads of households from
misspecification can lead to weak results for importance the Consumer Mail Panel of Market Facts, Inc. pro-
weights [6].
vided 1271 usable responses distributed in approxi-
The net impact of most of these arguments, however, mately the same proportions as the original balanced
is to indicate why results for model (1) and model (2) sample. The five leading brands in each of six product
should be similar. In considering why one form might be classes provided attitude objects for analysis; importance
found better than the other, attention turns to the studies weights and brand beliefs for each of five product
described earlier. All three studies which indicated sup- attributes were obtained (Table 1), as well as rankings
pression utilized cross-sectional methodologies; four of of brand preference for each product class. Only pur-
the five studies indicating improvements with importance
chasers of the product class whose favorite brand was
weights utilized individual-level analysis. It is possible,
included in thefivebrands were used in the analysis.
therefore, that analytical procedures could have led to
suppression conclusions. This possibility is examined in Beliefs as to the extent to which each brand offered
this study. satisfaction on each product attribute were obtained on
a 6-point scale ranging from "very satisfactory" to "very
unsatisfactory." Importance weights in terms of the re-
Normalization in Cross-Sectional Designs
spondent's designing an ideal brand were obtained on
A typical brand-level, cross-sectional analysis might a 6-point scale ranging from "important" to "unimpor-
involve univariate regressions, by brand, of ranked pref- tant,", and rescaled (e.g., " 1 " = "6", "2" = "5") for
erence ratings against summed attitude scores from analysis. Preferences were obtained in a forced ranking
model (1) or model (2).- Preference ranks are normal- from 1 through 5 in each product class.
ized by individual, not brand, such that the distribution
of the dependent variable can differ for each brand re- METHODOLOGY AND RESULTS
gression. One commonly overlooked assumption of Investigation of the hypothesized effects of normaliza-
these regressions is that scale level differences in belief tion in cross-sectional analyses required that four uni-
and importance ratings represent "true" differences in variate regressions of preference rank (Pj^) against atti-
consumer evaluations. This assumption may have serious tude scores be performed for each brand in each product
analytical consequences when violated via differential class. The four regressions reflected summated model
consumer response sets. combinations of beliefs-only versus beliefs times impor-
Normalization means a rescaling of numbers in such tance and raw versus normalized belief and importance
a manner as to adjust for within-subject variance in re- ratings:
sponses. Thus, if Bij^, is the belief of the A:th consumer
of the /th attribute for the /th brand, the normalized (1) Pj, =
belief of the kih consumer in BNau = Bijt/Y^jBuk.
Similarly, the normalized importance weight is INik = (beliefs-only, raw data)
A/ lJ
When predictor variables—brand beliefs and impor- (2)
tance weights—are not normalized by individual, the
effect is a regression model attempting to explain rela- (belief times importance, raw data)
tive brand preferences with attitude scores which, across
persons, do not reflect relative brand effects. It is there- (3) Pi,=
fore hypothesized that cross-sectional regressions using
raw data yield lower r^ of both models than identical
analyses run with normalized data. It is further hypoth- (beliefs-only, normalized).
esized that use of normalized data will show no suppres-
sion effect of importance weights. This article presents (4) Pi, =
a comparative analysis of the normalization hypothesis
in terms of changes in predictive and diagnostic results (beliefs-importance, both normalized).
of cross-sectional models with and without importance
weights. Following the selection of a "best" cross-sec- Results of these 120 regressions are summarized in
tional model, a reconciliation of cross-sectional results Table 2 in terms of average r- obtained for the five
brands in each product category from each form of the
with individual-level results is discussed.
attitude model. Results are internally consistent and
support expectations.
THE DATA Looking down column 1 (beliefs-only), it appears that
The data bank employed by Bass and Talarzyk and
Sheth and Talarzyk is utilized in order to afford com- - Disaggregated models in which each attribute represents a
separate predictor in a multiple regression are also representa-
parability with prior attitudinal studies [4, 29]. A na- tive of cross-sectional analysis.
266 JOURNAL OF MARKETING RESEARCH, AUGUST 1973

the ability to explain brand preference differs by product; Table 3


strongest predictions occur for brassieres and weakest AVERAGE NUMBER OF AHRIBUTES SIGNIFICANT (OF 5)
predictions for orange juice. The average r^ for all 30
brands is .15 for the beliefs-only model, and is consistent Disaggregated attribute attitude model
with results of prior research with this data base [29]. Product
f>B SBI bBN b BNIN
Comparison of column 2 with column 1, however, shows
the effects of including importance weights in the model Brassieres 3,0 3,0 4,8 5,0
and these results are significantly different from those Lipstick 3,6 3,0 4,0 5,0
reported by Sheth and Talarzyk. No systematic suppres- Mouthwash 3,6 3,6 3,8 5,0
sion effect is noted, but no significant increase in ex- Orange juice 2,2 2,0 2,4 4,6
Toilet tissue 2,8 3,2 3,6 5,0
planatory power is found either. Product differences re- Toothpaste 4,0 3,8 4,0 5,0
main. Average 3,2 3,1 3,8" 4,9''
The normalization hypothesis is tested by comparing
Significantly different than (1) and {2),p S ,05,
columns 1 versus 2 and 3 versus 4. Dramatic increases
Significantly different than (1), (2), and (3), p S ,001,
in r- are found with normalized ratings; explanatory
power jumps more than 150% in both models. Compari-
son of columns 3 and 4 again show no suppression effect Beta coefficients then represent the relative contribution
of importance weights. The increase from .38 to .39 is of each attribute within the attitude structure. Arguments
not statistically significant, but the weights do tend to be in favor of disaggregated analyses have been advanced
positive rather than negative. Slightly higher r- with by Sheth, Cohen and Houston, and Lutz and Howard
normalized importances were found in 25 of the 30 [9, 19, 28].
brand regressions and in 24 of 25 brands other than Disaggregated multiple regressions were performed
brassieres. to investigate the effects of normalization and impor-
The normalized beliefs times importance model yields tance weights on diagnostic power. The results summa-
an average r- = .39, corresponding to a multiple cor- rized in Table 3 indicate that use of nonnormalized data
relation coefficient between attitude score and brand (columns 1 and 2) seriously distorts diagnostic conclu-
preference of .63. The 30 brands ranged from .48 to sions by revealing far fewer dimensions of attitude struc-
.73, with product averages ranging from .54 to .67. ture than might actually be operative. Little differences
Given that these correlations are likely to be biased in number or type of significant attributes are found in
downward because of the constrained ordinal dependent either nonnormalized model form; it appears that
variable, results compare quite favorably to similar cor- product classes and brands vary substantially in dimen-
relations of .7 presented by Cohen, Fishbein, and Ahtola sionality, but that relatively few attributes (an average
as representative of attitude studies in social psychology of 3.2 in the list of 5) are operative. A shift to the
normalized beliefs-only model (column 3) shows a signif-
[Ill- icant increase in the average number of significant at-
Explanation of variance, however, is only one of two
major potentials of multi-attribute attitude models. Di- tributes (3.8 of 5), but continues to indicate variance by
agnosis of attitude structure is also an appealing pos- brand and product class. Inclusion of normalized im-
sibility that typically involves the use of disaggregated portance weights (column 4), however, provides a further
cross-sectional analyses in which each attribute retains significant increase in number of attributes (to 4.9 of 5)
its separate identity as a predictor variable. Estimated across all product claisses; 29 of the 30 brand regressions
showed all 5 attributes contributing to explanation of
preference. It is also interesting to note that r^ increased
Table 2
only from .39 to .40 in this disaggregated form as op-
VARIATION EXPLAINED IN BRAND PREFERENCE posed to prior summation, which tends to support orig-
(R2's AVERAGED OVER BRAND REGRESSIONS FOR EACH inal proposals for prior summation of the data.
PRODUCT CATEGORY) Shifting attention from importance weights to the
issue of cross-sectional versus individual-level analysis,
Attitude Modet a comparison of predictions made by each approach was
Product YB YBNI YBNitNi developed. Individual-level analysis and predictions with
V) (2) (3) (4) these data have previously been reported by Bass and
Talarzyk [2, 4]. The summed beliefs-importance scores
Brassieres ,19 ,19 ,45 ,45 for each individual on each brand are ranked in ascend-
Lipstick ,14 ,14 ,36 ,38
ing order by product category. A "confusion matrix"
Mouthwash ,18 ,19 ,42 ,43
Orange juice ,11 ,13 ,28 ,29 of the conditional probability of actual preference rank
Toilet tissue .14 ,13 ,38 ,41 given predicted preference is developed by summing over
Toothpaste ,14 ,14 ,37 ,38 all respondents. Table A presents the confusion matrix
Average r^ ,15 ,15 ,38" ,39' for five toothpaste brands. Note that this analysis forces
Significantly different than (1) and (2), p S ,001, an equal number of pre,dictions for each rank (1104).
COMPARATIVE ANALYSIS OF AHITUDINAL PREDICTIONS 267

The attitude model is especially effective in predicting Table 5


first and last choice brands, but the diagonal is dominant
CONDITIONAL PROBABILITY OF ACTUAL RANK GIVEN
for every rank; this analysis yields 53% correct pre-
PREDICTED RANK: TOOTHPASTE
dictions.
A related process was used to create a confusion (CROSS-SECTIONAL ANALYSIS)
matrix from the cross-sectional brand regressions for
Actuat rank
the summated, normalized beliefs-importance model. A
preference score was calculated for each respondent for 1 2 3 4 5 E n
each brand using the estimated regression coefficients:
1 ,82 ,14 ,03 ,01 ,00 1.0 423
Predicted 2 ,43 ,32 ,19 .06 ,01 1,0 1175
= di + hi 3 ,10 ,25 ,29 ,27 ,08 1,0 2342
rank
4 ,02 ,06 ,15 ,30 ,48 1,0 1216
Predictions of integer preference rank were calculated 5 ,00 ,01 ,01 ,09 ,89 1,0 364
from jyt: Percent correct predictions = 38%

Pile = 1 if - 0 0 < y ^ 1.5


1, 2, 4, and 5; note the preponderance of first ranks re-
= 2 if 1.5 < > ^ 2.5 ceiving second place predictions, second and fourth
= 3 if 2.5 < y ^ 3.5 ranks receiving third place predictions, and fifth ranked
brands predicted to be fourth. While normalization of
= 4 if 3,5 < > g 4.5 beliefs and importances has improved prediction com-
pared with predictions based on unadjusted data, it is
= 5 if 4,5 < y. clear that this cross-sectional regression model has not
Table 5 presents the resulting toothpaste confusion completely overcome the difficulties of interpersonal
matrix from this procedure. Comparison of these re- comparisons.
sults with those in Table 4 would seem to indicate that
cross-sectional regressions do an even better job than CONCLUSIONS
individual level analysis in predicting first and last
choice brands, while not doing quite as well for mid- It has been shown that methods of analysis can have
rated brands. The percentage of correct predictions major impacts upon the conclusions drawn on other
(38%), however, is substantially less than that of in- issues in attitude models. The issue of whether or not
dividual analysis (53%). Note that only 423 first place importance weights add to the power of a multi-attribute
and 364 last place predictions occur, while 2342 pre- model was given special attention here because of the
dictions of third rank result from the regression equa- controversy surrounding this question. A review of the
tion. Since there are an equal number of actual rankings controversy revealed a number of persuasive arguments
(1104), the effect of the cross-sectional model is to apart from analysis. It should be stressed that this study
significantly underpredict extreme ranks. This effect is has not dealt with suggestions involving specifications of
shown in a nondominant diagonal for Table 5, and is attributes or measurements or belief and importance;
clarified in Table 6, which shows, as diagonal elements, potential biases from these sources were effectively held
the probabilities of correct predictions given actual constant by using a data base which has previously been
ranks. The tendency of regression predictions to con- studied with respect to the importance question. A sec-
centrate on rank 3 results in poor predictions for ranks ondary analysis of these data, moreover, indicates that
arguments concerning the "halo effect" in brand beliefs
Table 4 and a lack of variance in obtained importance ratings
may be effective in explaining why importance weights
CONDITIONAL PROBABILITY OF ACTUAL RANK GIVEN
do not seem to add significantly to model predictions.
PREDICTED RANK: TOOTHPASTE A pervasive tendency of respondents in this data base
(INDIVIDUAL-LEVEL ANALYSIS) giving their most preferred brand higher belief scores
than other brands on all attributes has been shown by
Actuat rank
Bass and Talarzyk [3]. A breakdown of importance
1 2 3 4 5 E n ratings by brand most preferred shows little differences
between users of one brand and those of another brand.
1 ,75 .16 ,05 ,03 ,01 1,0 1104 The combination of these data characteristics would be
Predicted 2 ,14 ,46 ,22 ,13 ,06 1,0 1104 expected to allow importance weights little opportunity
rank 3 ,07 ,23 ,39 ,21 ,10 1,0 1104 to add significantly to the predictive power of beliefs.
4 ,03 ,12 ,25 ,41 ,19 1,0 1104
5 ,01 ,05 ,09 ,22 ,63 1,0 1104 Future research on the importance weight issue should
Percent correct predictions = 53% therefore examine attribute specification and measure-
ment alternatives.
268 JOURNAL OF MARKETING RESEARCH, AUGUST 1973

Table 6 art in empirical research in application of attitude theory


to consumer choice behavior is that models are evaluated
CONDITIONAL PROBABILITY OF PREDICTED RANK
and compared on a correlational basis. Attitudes are
GIVEN ACTUAL RANK: TOOTHPASTE
both a cause and a result of behavior. No formal theory
(CROSS-SECTIONAL ANALYSIS) of simultaneous causality has yet been developed. Thus
Predicted rank it is not possible to test attitude theories. Correlational
studies, such as this one, are of interest, but they provide
t 2 3 4 5 E n no information about causality.

1 ,31 ,46 ,21 ,02 ,00 1,0 1104 REFERENCES


2 ,06 ,34 ,54 ,07 ,00 1,0 1104
Actual 3 ,01 ,20 ,62 ,17 ,00 1,0 1104 1, Anderson, Lynn R. and Martin Fishbein, "Prediction of
rank 4 ,00 ,06 ,58 ,32 ,03 1,0 1104 Attitude from the Number, Strength, and Evaluative Aspect
5 ,00 ,01 ,17 ,52 ,29 1,0 1104 of Beliefs about the Attitude Object: A Comparison of
Summation and Contiguity Theories," in Martin Fishbein,
ed.. Readings in Attitude Theory and Measurement. New
York: Wiley, 1967, 437-43,
The following conclusions can be drawn from the 2, Bass, Frank M, and W, Wayne Talarzyk. "A Study of At-
study in this paper: titude Theory and Brand Preference," Institute Paper No.
1. The distinction between cross-sectional and in- 252, Institute for Research in the Behavioral, Economic, and
dividual-level analysis should be carefully con- Management Sciences, Krannert Graduate School of In-
sidered; cross-sectional analysis offers practical dustrial Administration, Purdue University, 1969,
3, . "Using Attitude to Predict Individual Brand Pref-
benefits in summarizing results on brands and erence," Occasional Papers in Advertising, 4 (May 1971),
attributes but requires very strong assumptions 63-72,
for validity of analysis. 4, . "An Attitude Model for the Study of Brand Prefer-
2. One typical assumption in cross-sectional anal- ence," Journal of Marketing Research, 9 (February 1972),
ysis is that scalar response differences are not the 93-6.
5, Bass, Frank M,, Edgar A, Pessemier, and Donald R, Leh-
result of response sets. Normalization of re- mann, "An Experimental Study of Relationships Between
sponses was tested and shown to lead to sig- Attitudes, Brand Preference and Choice," Behavioral
nificant increases in r^ (from .15 to .39) and Science, 17 (November 1972), 532-41,
number of significant attributes (from 3.2 to 6, Beckwith, Neil E. and Donald R, Lehmann, "The Impor-
4,9) produced by the attitude model. tance of Differential Weights in Multiple Attribute Models
of Consumer Attitude," Journal of Marketing Research, 10
3. Importance weights, when analyzed in this (May 1973), 141-5,
manner, do offer benefits in diagnosis of atti- 7, Churchill, Gilbert A,, Jr, "Linear Attitude Models: A Study
tude structure; prior theories using importances of Predictive Ability," Journal of Marketing Research, 9
in summated forms are supported. (November 1972), 423-6,
4. Predictions of brand preference ranking for 8, Cohen, Joel B. and Olli T, Ahtola, "An Expectancy X Value
Analysis of the Relationship Between Consumer Attitudes
each respondent can be used to judge the pre- and Behavior," Proceedings. Second Annual Conference,
dictive validity of cross-sectional analysis as Association for Consumer Research, 1971, 344-64,
compared to the theoretically preferred method 9, Cohen, Joel B. and Michael J, Houston, "The Structure of
of within-individual analysis. However, within- Consumer Attitudes: The Use of Attribute Possession and
Importance Scores," Faculty Working Paper No, 2, Col-
individual analysis is always the superior method lege of Commerce and Business Administration, University
of analysis when relative attitudes are used to of Illinois, 1971,
predict preference or choice, since this method 10, , "Cognitive Consequences of Brand Loyalty," Jour-
does not imply respondent homogeneity, a very nal of Marketing Research, 9 (February 1972), 97-9,
strong assumption necessarily implied by cross- 11, Cohen, Joel B,, Martin Fishbein, and Olli Ahtola, "The
Nature and Use of Expectancy-Value Models in Consumer
sectional analysis. Attitude Research," Journal of Marketing Research, 9
This article assesses the impacts of alternative ana- (November 1972), 456-60.
lytical procedures on the results and conclusions drawn 12, Day, George S, "Evaluating Models of Attitude Structure,"
in studies of the multi-attribute attitude model. It has Journal of Marketing Research, 9 (August 1972), 279-86,
been shown that cross-sectional analysis is particularly 13, Fishbein, Martin, "A Behavior Theory Approach to the
Relations Between Beliefs about an Object and the Attitude
susceptible to response assumptions; importance weights toward the Object," in Martin Fishbein, ed,. Readings in
not only do not suppress predictions, but also improve Attitude Theory and Measurement. New York: Wiley, 1967,
diagnosis of attitude structure when the data are normal- 389-99,
ized. Even though the cross-sectional model does not 14, Ginter, James L, "Attitude Change and Choice Behavior in
obtain the predictive levels of within-individual anal- New Product Introduction," unpublished doctoral disserta-
tion, Purdue University, 1972,
ysis, the results in this marketing context are comparable 15, Hansen, Flemming, "Consumer Choice Behavior: An Ex-
in level to those found in social psychological studies perimental Approach," Journal of Marketing Research, 6
of attitude, A major weakness of the present state of the (November 1969), 436-43,
COMPARATIVE ANALYSIS OF ATTITUDINAL PREDICTIONS 269
16, and Thomas Bolland. "Relationship Between Cog- bein," Journal of the Market Research Society, 12 (July
nitive Models of Choice and Non-metric Multi-dimensional 1970), 145-66,
Scaling," Proceedings. Second Annual Conference, Associa- 26, Scott, Jerome E, and Peter D, Bennett, "Cognitive Models
tion for Consumer Research, 1971, 376-88. '• of Attitude Structure: Value Importance is Important," Pro-
17, Hughes, G. David, "Distinguishing Salience and Valence," ceedings. Fall Conference, American Marketing Association,
paper presented to Attitude Research ,Workshop, University 1971,346-50.
of Illinois, 1970. 27, Sheth, Jagdish N. "Attitude as a Function of Evaluative
18, and Jose L, Guerrero. "Testing Cognitive Models Beliefs," paper presented to the AMA Conference Work-
Through Computer-Controlled Experiments," Journal of shop, Columbus, Ohio, 1969,
Marketing Research, 8 (August 1971), 291-7, 28, . "An Investigation of Relationships Among Eval-
19, Lehmann, Donald R, "Television Show Preference: Applica- uative Beliefs, Affect, Behavioral Intention and Behavior,"
tion of a Choice Model," Journal of Marketing Research, working paper. College of Commerce and Business Admin-
8 (February 1971), 47-55, istration, University of Illinois, 1970,
20, Lutz, Richard J, and John A, Howard. "Toward a Com- 29, and Wayne W, Talarzyk, "Perceived Instrumentality
prehensive View of the Attitude Behavior Relationship: The and Value Importance as Determinants of Attitudes," Jour-
Use of Multiple-set Canonical Analysis," Proceedings. Social nal of Marketing Research, 9 (February 1972), 6-9.
Statistics Section, American Statistical Association, 1971, 30, Wilkie, William L. and John M, McCann. "The Halo
215-25, Effect and Related Issues in Multi-Attribute Models—An
21, Moinpour, Reza and Douglas L, MacLachlan. "The Rela- Experiment," Institute Paper No. 377, Institute for Research
tions Among Attribute and Importance Components of in the Behavioral, Economic, and Management Sciences,
Rosenberg-Fishbein Type Attitude Models: An Empirical Krannert Graduate School of Industrial Administration,
Investigation," Proceedings. Second Annual Conference, Purdue University, 1972,
Association for Consumer Research, 1971, 365-75, 31, Wilkie, William L, and Edgar A, Pessemier, "Issues in
22, Moinpour, Reza and James Wiley. "An Empirical Investiga- Marketing's Use of Multi-Attribute Attitude Models," In-
tion of Expectancy-like Models in Marketing," working stitute Paper No, 365, Institute for Research in the Be-
paper. University of Washington, 1972. havioral, Economic, and Management Sciences, Krannert
23, Pessemier, Edgar A, "Measuring Stimulus Attributes to Graduate School of Industrial Administration, Purdue Uni-
Predict Individual Preference and Choice," Institute Paper versity, 1972,
No, 346, Institute for Research in the Behavioral, Economic, 32, Wilkie, William L, and Rolf P, Weinreich. "EfFects of the
and Management Sciences, Krannert Graduate School of Number of Attributes Included in an Attitude Model: More
Industrial Administration, Purdue University, 1972, is not Better," paper presented to Third Annual Conference,
24, Rosenberg, Milton J. "Cognitive Structure and Attitudinal Association for Consumer Research, 1972,
Affect," Journal of Abnormal and Social Psychology, 53 33, Winter, Frederick W, "A Laboratory Experimental Study
(November 1956), 368-72, of the Dynamics of Attitude and Choice Behavior," un-
25, Sampson, Peter and Paul Harris, "A User's Guide to Fish- published doctoral dissertation, Purdue University, 1972,
View publication stats

You might also like