Empirical Evidence of Halo Effects in Store Image Research By Estimating True Locations

Nell E. Beckwith, University of Pennsylvania (Ph.D. Candidate), University of Pennsylvania
U. Victor Kubilius,
ABSTRACT - Individuals' judgments and responses often reflect many factors, such as their overall attitude toward the object being judged, its popularity, their familiarity with it, etc. These biases are similar to the well-known halo effect--an individual's tendency to bias his responses about an object on any specific attribute by his general, overall (global) impression. This article describes a procedure for estimating locations of the judged objects corrected for these types of halo-like effects.
[ to cite ]:
Nell E. Beckwith and U. Victor Kubilius (1978) ,"Empirical Evidence of Halo Effects in Store Image Research By Estimating True Locations", in NA - Advances in Consumer Research Volume 05, eds. Kent Hunt, Ann Abor, MI : Association for Consumer Research, Pages: 485-493.

Advances in Consumer Research Volume 5, 1978      Pages 485-493

EMPIRICAL EVIDENCE OF HALO EFFECTS IN STORE IMAGE RESEARCH BY ESTIMATING TRUE LOCATIONS

Nell E. Beckwith, University of Pennsylvania

U. Victor Kubilius (Ph.D. Candidate), University of Pennsylvania

[The authors are indebted to Charles D. Goldfine and Karen E. Jukobowski (1977), who originally collected the data and made it available to us for this further analysis, and to Randy Batsell, who critically reviewed an earlier draft.]

ABSTRACT -

Individuals' judgments and responses often reflect many factors, such as their overall attitude toward the object being judged, its popularity, their familiarity with it, etc. These biases are similar to the well-known halo effect--an individual's tendency to bias his responses about an object on any specific attribute by his general, overall (global) impression. This article describes a procedure for estimating locations of the judged objects corrected for these types of halo-like effects.

INTRODUCTION

Marketing researchers and psychologists frequently attempt to obtain evaluations or locations of particular alternatives (individuals, objects, concepts, or brands) on particular attributes (traits, factors, variables, or characteristics), as in image studies, attitude models, and product attribute positioning models (Fishbein and Ajzen, 1975; Huber and James, 1977; Rosenberg, et al, 1960). Convenient estimators of these locations are the averages of many individuals' responses on the attribute. However, individuals may commingle their knowledge about the objects on many different attributes, and thereby provide response ratings on each attribute which are biased. Such biases may be due to their evaluations on other attributes, their overall evaluation of the object, its popularity, their usage or familiarity with it, their beliefs about other people's assessment of the object (peer attitude), cognitive dissonance (Festinger, 1957), or consistency maintenance (Rosenberg, et al, 1960). The resulting averages may thereby be biased.

The halo effect is well recognized by psychologists and has been defined as the "tendency in rating to be influenced by general impression or attitude when trying to judge separate traits" (English, 1934). Additional halo-like systematic response biases may be associated with other sources.

We demonstrate a method for obtaining estimates of the locations of objects on particular attributes. We refer to these as estimates of the true or actual location of the objects on the attribute, even though the attribute may be subjective.

Basically, our objective is to use individuals' responses, which indicate their rating of an object on an attribute, to construct an estimate of the unmeasured (true or actual) location of the object on the attribute. We assume that individuals' responses are a function of the (true) location as well as a function of other variables. First, we estimate each individual's response function and then use an iterative technique to estimate the (true) locations of the objects across all individuals.

The store image data which was conveniently available to us has ordinal (ranking) rather than interval (rating) scales on some variables. Since the method we use assumes intervally scaled data, the empirical example is included only for illustration of the methodology. We draw no substantive inferences from it.

BACKGROUND

The halo effect is well known to psychologists. Thorndike (1920) noted that an individual tends to rate another person on an attribute by his perception of the person on some other attributes. Similar halo-like effects occur in many other circumstances. Bass and Talarzyk (1972) reported that respondents indicated higher average belief scores on attributes for their preferred brands than for less favored brands. Beckwith and Lehmann (1975) demonstrated that respondents also halo their responses when asked to rate television shows as commonly done in the product positioning attribute models. Bass and Wilkie (1973) suggested that the halo effect on indicated brand beliefs may partially explain why importance weights do not seem to add significantly to the predictive performance of such attitude models. In contrast, Moore and James (1977) found halo effects to be relatively unimportant in students' ratings of automobiles. A summary of halo effect research and implications is available elsewhere (Beckwith, Kassarjian, and Lehmann, 1977).

Store image studies provide a convenient example of attribute (i.e., trait) measurement problems and are of practical importance since objective measures on many of the attributes are difficult or impossible to obtain. [Examples of store image studies include those reported by Hawkins, Albaum, and Best (1975-76), Jain and Etgar (1976-77), Kunkel and Berry (1968), Lessig (1972), Linquist (1974-75), May (1971), Sharma and Doyle (1977), and Staples and Lockander (1975-76). Readers are also referred to the Special Store Image Issue of the Journal of Retailing (Winter 1974-75).]

Two recent investigations provided considerable motivation for the present study. Huber and James (1976) pointed out that cross-sectional averages of halo-biased ratings on an attribute are also biased because the alternative objects (such as stores or brands) will generally be favored by different proportions of the respondents. Peterson (1976) found such a haloing effect in six studies of retail store images where average ratings on many attributes were correlated with the market shares of the retailers in their geographic marketing areas. We attempt to obtain less biased estimates of the objects' locations on attributes by considering both relative usage and overall attitudes in the sample of respondents.

MODEL

In this section we develop an example of a class of models useful for estimating the true (or actual) locations of objects on specific attributes. The example selected is both rather simple and readily extendable to more elaborate functional forms if warranted.

As a simple illustration we examine a model having only three components. [The more general linear model is presented at (4). Here we display the special case used as an example.] First, each object j has an actual location Tjk on each attribute k. Tjk may be thought of as the true location of the object on the attribute. Respondents may have been exposed to the object, and thereby exposed to cues of Tjk (see Castellan, 1973). The analyst is attempting to estimate Tjk.

Second, individual i's evaluation Eijk of object j on each attribute k reflects influences of many other variables besides Tjk. For purposes of this example, assume a simple linear transformation:

Eijk =  bi0k  + bilk"ij + bi2kUij  + bi3kTij  (1)

The bilk"ij component corresponds to the usual halo effect. The bi2kUij component is included as an example of other halo-like biasing effects. [In our empirical work we originally attempted to include three separate effects: Store of Last Purchase, Store of Most Frequent Purchase, and Most Convenient Store. However, for most subjects these variables were too collinear. We settled on the single usage measure, Store of Most Frequent Purchase. Similar results were obtained with all three measures of usage except for the Value attribute.]

Third, we assume that each individual's response Bijk is a linear function of his evaluation Eijk plus a disturbance eijk:

Bijk  =  ai0kailk Eijk + eijk   (2)

Each individual uses his own linear scaling transformation with coefficients aiqk and ailk. Disturbances eijk are independently distributed with mean zero and idiosyncratic variance s2ik. Basically, this is a linear accommodation of the different response scales which individuals may employ to determine their own responses Bijk based on their own evaluations Eijk.

Substituting (1) into (2) yields the reduced form equation: [Where bi0k = ai0k + ailkbi0k, bilk = ailkbi1k, etc.]

Bijk = bi0k  + bilk"ij   +  bi2kUij  + bi3kTij + eijk .   (3)

The variables Bijk, Aij, and Uij are respondent i's indicated responses (data). However, Tjk, not empirically observable, is to be estimated. A naive estimator of Tjk is Bjk, the average across all i respondents (Beckwith and Lehmann, 1975). However, Huber and James (1976) have pointed out that such an estimator is biased because it neglects the differing fractions of respondents favoring each brand. Thus, we elect to estimate Tjk cross-sectionally.

ESTIMATION METHOD

Consider a simple response function for individual i on one attribute k. Dropping the k subscripts for the moment:

EQUATION   (4)

where i's response Bij about object j is determined as a function of L independent intervally scaled variables Xijl, l = 0,...,L (including the intercept), and also as a function of the location Tj of object j. The independent variables Xijl could include the overall evaluation of the object j (like-dislike), preference, popularity of market share, peer attitudes or i's belief about them, familiarity, etc. We seek to obtain estimates Tj for each j = 1,...,J. These initial guesses could be chosen arbitrarily. We could use them to estimate all the bs in (4) for an individual i by an ordinary least squares (OLS) regression across the J objects, or perhaps by some other method. If we then examine i's residuals Rij - Rijwe would find them to be positive for some objects and negative for others. Assuming bit> 0, we would then tend to guess that the corresponding original guess about Tj was too small (i.e., too negative) for those objects j with BiT> 0 and that the corresponding original guess about Tj was too large (i.e., too positive) for those objects j with eij < 0. [Of course, we could easily obtain all eij = 0 j = 1,...,J for any one particular individual i by rigging the Tjs. However, we cannot usually find a set of Tjs which will drive all eij = 0 for many individuals.]

Instead of using eijs from just one individual to improve the guesses about the Tjs, we suggest that the eijs be obtained for many individuals and that a Tj should be increased if Sieij> 0 and decreased if the sum of residuals for the object is negative, assuming the bits are generally positive. The scale for Tj and bit is arbitrary. To resolve this ambiguity, we simply constrain the Tj scale to be standardized with zero mean, unit variance, and sign such that bit > 0. [In retrospect we feel it might be better to constrain the estimated locations Tj to a scale of zero-one, rather than standardizing to zero mean and unit variance. Negative locations seem to make communication of findings more difficult.]

Thus, this procedure starts from an original set of arbitrary values for the Tjs. The bs of the response function are estimated for each individual by OLS. The Tjs are then revised. We used the updating rule:

EQUATION   (5)

The procedure is then repeated with the new Tjs replacing the previous values until convergence, when Sieij = 0.- Finally the signs of the Tjs are selected such that bit = 0, and the final Tjs are scaled to have zero mean and unit variance.

In our example, the procedure converged to the same final locations regardless of the initial Tjs arbitrarily selected. Since the procedure estimates (4) for each individual i separately, rather than conjoining them into some kind of a big pooled regression, each iteration of the estimation requires only a modest sized regression. Thus, the procedure can he performed on very modest computers, even for very large samples.

DATA

The data used in this study were collected by Goldfine and Jukobowski (1977) as part of an advanced study project at The Wharton School, University of Pennsylvania. They investigated the homogeneity of store images and the degree of halo effect upon these store images within a convenience sample of professional women.

Although the methodology is designed to address substantive issues, our purpose here is only to display the methodological procedure. We use this convenient data base as a means of illustrating the application of the procedure. The reader is cautioned to consider this report as only illustrative of the types of results possible. The names of the stores (objects) are included only to facilitate understanding of the procedure.

The sample was obtained mainly on a convenience basis. It consisted of 43 women who resided in the greater Philadelphia area, and were employed in professional/ administrative capacities or were studying toward that goal. Some of the respondents were acquaintances of the researchers. Responses from 9 of the women were not used due to missing data or errors. Also, individuals who indicated equal beliefs toward all 10 stores on an attribute were eliminated from analysis of that attribute since regression estimates of their weights were not possible. The 10 stores included diverse types familiar to many of the respondents.

Belief Responses Bijk

The respondents were questioned about the 10 stores on 8 attributes. [See Table 2 for the list of stores and attributes. The standardized average beliefs Bj across all individuals are displayed as B in the tables. Raw, rather than standardized, belief, attitude, and usage variables were used in the estimation procedure.] The attributes were selected on an a-priori basis. Thus, they are only illustrative of the types of attributes individuals might consider in making store evaluations. Each respondent i indicated her belief Bijk about each store j on each attribute k using a five-point ("Outstanding" = 4, "Unacceptable" = 0) scale which we assumed to be an interval measure. The questionnaire items used to measure two of these attributes, Convenience and Fashionability of Merchandise, were worded such that the individual's differing ideal points (or home/work site) preclude assumption of the existence of a true location of a store on the attribute for all individuals. [See the Appendix for questionnaire wording.] In addition, one of the usage measures, Most Convenient Store, also solicited convenience information. Consequently, estimates for these two attributes are not reported here.

Attitude Aij

Each respondent indicated her overall evaluative ranking of each store ("Most Preferred" = 9, "Least Preferred" = 0, no ties). This measure is, of course, only ordinally scaled. However, we used it as if it were an intervally scaled variable for purposes of illustrating the methodology.

Usage Uij

Three different measures of each individual's store usage were available: (1) Store of Last Purchase, (2) Store of Most Frequent Purchase, and (3) Most Convenient Store. These variables were coded as dummies (an indicated store = 1, otherwise = 0). Since these three variables measure somewhat different constructs, we would advocate using all three within the analytic procedure. However, in this small set of data, the variables were too collinear to permit their simultaneous inclusion. For purposes of comparison, we included them each one-by-one as a single measure of usage in separate analyses. These usage variables do not meet the interval scale assumption. [For an individual having usage variable value of 1 for only one store, that store is essentially ignored in estimating the individuals' coefficients b0, b1, b3, and determines b2. Some individuals indicated multiple stores.]

While these particular data from a convenience sample are not exactly consistent with the model's assumptions, they are adequate to simply illustrate the methodological procedure.

ILLUSTRATIVE APPLICATION

The Quality of Merchandise attribute is used in an example of the procedure. We selected the second operationalization (Store of Most Frequent Purchase) as the measure of usage. Other attributes and the importance of the halo effect are then examined in subsequent sections.

Quality of Merchandise Attribute

In lieu of the yet unavailable locations Tj we selected the cross-sectional averages Bj for the initial arbitrary starting point. We then used OLS to estimate bi C ks in (3) for each particular individual (one at a time). The 10 stores were used as 10 observations. The OLS estimated coefficients (and standard errors) for one particular individual were:

bijk  =  -.48 - .03"ij  + .28Uij   + 1.48Bjk  + eijk,

            (.18)     (.07)     (.36)         (.19)

R2 = .98 (unadjusted).

This procedure is replicated for each individual in the sample. The resulting residuals do not sum to zero across individuals for each store. For example, here the sum of residuals Siei1k = .10 for the first store, Bonwit Teller. Using an updating rule, such as (5), the estimates of Tjs are revised away fromBjk, and the OLS regressions are then repeated for all individuals. Eventually, the procedure converges to final estimates of the true locations Tj. For this same individual the final estimated coefficients (and standard errors) were:

bijk  =  1.64 + .15"ij  - .28Uij   + .92Tjk  + eijk,

            (.23)     (.05)     (.34)         (.13)

R2 = .98 (unadjusted).

After convergence the sum across individuals of the final residuals is zero for each store, e.g., Sieijk = 0. j = 1,...,J.

Although the analysis yields estimates like these for each individual, such results are too bulky to report in detail. [Final R2s exceeded initial R2s for about half of the respondents on ill attributes.] Figure 1 displays the distribution of Bs, standard errors, and t-statistics (H0 : bilk = 0) obtained for all 34 individuals. With only 10 objects (stores) the standard errors tend to be rather large. Even so, of the 34 t-statistics for B1, for example, 12 exceeded the critical value 2.447 for 6 degrees of freedom at the 95 percent confidence level. Since it is unlikely that this many estimated coefficients would have |t| $ 2.447 by chance alone, we conclude that b1 cannot be equal to zero for all individuals, although it could be zero for some individuals. Thus, we conclude that the halo effect cannot be ignored here.

FIGURE 1

DISTRIBUTION OF ESTIMATED COEFFICIENTS, STANDARD ERRORS, AND T-STATISTICS FOR EXPLAINING INDIVIDUALS' RESPONSES ON THE QUALITY OF MERCHANDISE ATTRIBUTE

Figure 2 displays the distribution of unadjusted R2 obtained for these individuals, which ranged from .54 to .98. Averages of these statistics are summarized in the bottom portion of Table 1, column 2. Columns 1 and 3 display similar average results for analyses using the other two measures of store usage: Store of Last Purchase (#1) and Most Convenient Store (#3), in place of Store of Most Frequent Purchase (#2).

FIGURE 2

CUMULATIVE DISTRIBUTION OF R2 STATISTICS FOR EXPLAINING INDIVIDUALS' RESPONSES ON THE QUALITY OF MERCHANDISE ATTRIBUTE

The average response Bjk and estimated true locations Tjk of the 10 stores on the Quality of Merchandise attribute are displayed at the top of Table 1. Note that the estimated locations Tjk were very similar to the average responses Bjk for several of the stores. However, for Nan Duskin, Gimbels, and Wanamaker, the estimated locations Tj were noticeably different from the average responses Bj. Nan Duskin is one of the less frequented stores; none of the respondents mentioned it as being either the Store of Last Purchase or Store of Most Frequent Purchase. Thus, it is not surprising that the estimated Tj is higher than Bj for this store. Wanamaker was mentioned in this sample as one of the more frequently shopped stores (although less than Loehmann's). Thus, it is not surprising that the estimated Tj is less than Bj for this store. Both of these differences are consistent with the notion that average responses Bj may be favorably biased estimates of Tj for stores (or brands) having larger market shares.

TABLE 1

COMPARISON OF RESULTS FOR QUALITY OF MERCHANDISE ATTRIBUTEA

Value Attribute

The same analysis was performed on the Value attribute generated by the questionnaire item: "Price should be considered in terms of value for money spent." The average responses Bj on this Value attribute varied widely among stores. However, the estimated locations Tj were roughly similar for all stores except Loehmann's (see Figure 3).

Loehmann's is rather distinct from the other stores. For example, it sells many items with labels excised. Loehmann's had the lowest estimated locations T of any store on several other attributes: Quality of Store Personnel, Customer Services, and Store Atmosphere (see Table 2). Readers familiar with Loehmann's may find these estimates plausible, as do the authors.

For many respondents the second measure of usage, Store of Most Frequent Purchase, was highly correlated with Tj. This collinearity caused large standard errors for B2 and B3 for many individuals. Consequently, for this attribute we used the first measure of usage, Store of Last Purchase.

Other Attributes

Corresponding summary results for the other attributes are displayed in Table 2. For two of the attributes, Quality of Store Personnel and Customer Services, the estimated locations T differ quite considerably from the average responses B for many of the stores. The procedure converged in between 19 and 58 iterations for all the attributes except Value, which needed 162 iterations. The average was 39 iterations before termination (when none of the Tjs changed more than 10-7).

FIGURE 3

AVERAGE RESPONSE Bj AND ESTIMATED LOCATIONS Tj FOR THE 10 STORES ON THE VALUE ATTRIBUTE

TABLE 2

SUMMARY OF RESULTS FOR ALL ATTRIBUTES

Importance of Halo Effect

In the previous section we concluded that the halo effect cannot be ignored here. The relative importance of the halo effect is indicated by comparing avgi(siABi1), avgi(siUB12), and avgi(siTB13) = BT for each of the attributes (see Table 3).

TABLE 3

RELATIVE IMPORTANCE OF HALO EFFECT

The halo (overall attitude) and (true) location components are appreciable and roughly equal for all attributes. The usage component is negligible for all the attributes, evidently because of the low variances s2iU of the 0-1 measure of usage. Comparison between attributes is not possible here because the dependent variable, belief response, was not standardized [In retrospect, we should have standardized the data to allow comparisons across attributes.].

CAVEATS

In retrospect, we are now concerned with several issues which were not adequately resolved by this study:

1. Simultaneity--We assumed that none of the right hand variables are determined as a function of the left hand variable. Our estimation is biased if this assumption is false. This type of analysis should be extended to allow for at least the possibility of simultaneous co-determination of overall attitude and the beliefs, perhaps along the lines of Beckwith and Lehmann (1975). Bemmaor and Huber (1977) have argued that this specification error may be relatively tolerable, but we have not verified it for this example.

2. Intervally scaled data--We used available rank order data as if it were intervally scaled in order to simply demonstrate the methodology. Instead, intervally scaled data should be used. In addition, the convenience sample of respondents precludes making any substantive inferences from this particular data. However, these data are useful for demonstrating the methodology and illustrating the types of results which might be expected.

3. Individual coefficient nonnegativity--On a-priori grounds, individuals' coefficients Bilk and BiTk could be assumed to be nonnegative. By using OLS to estimate each individual's coefficients, we allowed these estimates to be negative, and many were. A linear programming-based individual estimation procedure might be used instead to obtain estimates constrained to being nonnegative, along the lines of Pekelman and Sen (1974) and Shocker and Srinivasan (1974). This might also reduce the collinearity problem.

4. Statistical properties--We have not demonstrated that the suggested procedure yields estimators Tjk which are necessarily no more biased than Bjk as estimators of Tjk. Presently we only point out that the procedure appears to yield plausible estimates in our example case, and that it seems reason- able to us that the bias will usually be reduced since the procedure explicitly includes effects which would otherwise be "omitted variables," and therefore generally biasing. Also, the OLS single equation standard errors and R2s reported for individuals are suspect since T] is estimated along with the 8 coefficients.

5. Objective attribute recovery--We have not yet checked the ability of this procedure to retrieve accurate location estimates for attributes with objective physical measures.

6. Computational efficiency--Other computational schema may be found which estimate the true locations more economically. We have not yet attempted to reduce the computational burden of the procedure.

7. Underlying processes--The model, (1) and (2), used for this illustrative example seems to be an oversimplification of the individual's cognitive processes. We hope that other investigators will find this type of analysis useful in developing more complete understanding of such processes.

SUMMARY

A methodology for estimating the location of objects on attributes or traits using individuals' biased response ratings was demonstrated. A procedure like this should be useful for assessing images of objects, such as stores, in image studies, or brand locations in product attribute analyses. The iterative procedure does require a substantial, although not prohibitive, amount of computation. The procedure was demonstrated on an example set of data for purposes of illustrating the methodology and the types of results which might be obtained. However, because the conveniently available data are not completely adequate, we caution against drawing substantive inferences from the example results. They are displayed solely to facilitate description of the methodology. We hope that this work will' encourage other investigators to pursue more refined estimators of the unobservable locations or attribute levels of objects.

APPENDIX

Portions of the questionnaire are reproduced so that readers may judge for themselves the nature of the attributes measured.

Overall Attitude (Aij)

We are interested in your overall opinions of the following stores in a very general way. Please rank the stores numerically from 1 to 10 with 1 = most favored, 2 = second most favored, etc., with 10 meaning least favored. [The coding order was subsequently reversed, 9 to 0, for the analysis reported here.]

Store Usage (Uij)

At which of these stores did you most recently purchase an article of clothing for yourself? (If none, please specify) [Denoted Usage Variable #1. ]

At which of these stores do you most frequently purchase clothing for yourself? (If none, please specify) [Denoted Usage Variable #2.]

Which of these stores do you consider most convenient (i.e., easiest to get to) to either your place of business or your home? (Include branch stores when applicable. If more, please specify) [Denoted Usage Variable #3.]

Belief Responses (Bijk)

We are interested in determining your opinions on a number of characteristics for several stores. Please indicate your opinion of each store on each attribute by checking the appropriate space. A short description of these characteristics precedes each section.

A. Quality of Merchandise--This refers only to the materials used in the manufacture of the garment and the garment's construction without any reference to price, fashionability, or any other factors.

B. Price--This should be considered in terms of value for money spent. [Price is referred to as the value attribute in the report.]

C. Quality of Store Personnel--Under this heading you should consider such factors as helpfulness, friendliness, promptness, courtesy, and any other interpersonal dynamic that is important to you.

D. Store Atmosphere--Store atmosphere includes the store's cleanliness, lighting, interior design, background music (if any), and floor plan.

E. Appeal of Advertising--In answering this question, you might consider whether or not the store's ads attract your eye in a newspaper or magazine as well as whether the ads are likely to result in your visiting the store in question.

F. Convenience of Location--(Consider branch stores when appropriate) In this category consider such factors as traffic, parking, proximity to public transportation, how frequently you find yourself in the area where the store is located, and so forth.

G. Customer Services--Customer services include credit availability and terms, return or exchange policies, alterations, rest room facilities, in-store restaurants, or any other particular service that you consider important.

H. Fashionability of Merchandise--In determining the fashionability of a store's merchandise, ask yourself whether the store carries merchandise that you like, not whether the store's merchandise could be featured in a fashion magazine.

Five response categories were provided for each store on each of the above beliefs: "Outstanding," "Highly Acceptable,'' "Acceptable," "Somewhat Acceptable," and "Unacceptable." These were subsequently coded 4 to 0 respectively.

REFERENCES

A. H. Barton and R. W. Parsons. "Measuring Belief System Structure," Public Opinion Quarterly, 41 (Summer 1977), 159-180.

Frank M. Bass and W. Wayne Talarzyk, "An Attitude Model for the Study of Brand Preference. " Journal of Marketing Research, 9 (February 1972), 93-96.

Frank M. Bass and William L. Wilkie, "A Comparative Analysis of Attitudinal Predictions of Brand Preference. " Journal of Marketing Research, 10 (August 1973), 262-269.

Neil E. Beckwith and Donald R. Lehmann, "The Importance of Halo Effects in Multi-Attribute Attitude Models," Journal of Marketing Research, 12 (August 1976), 265-275.

Neil E. Beckwith, Harold H. Kassarjian, and Donald R. Lehmann, "Halo Effects in Marketing Research: Review and Prognosis," in Advances in Consumer Research, 5 (1977 Proceedings).

Albert C. Bemmaor and Joel C. Huber, "Econometric Estimation of Halo Effects: Single vs. Simultaneous Equation Models," in Advances in Consumer Research, 5 (1977 Proceedings).

N. John Castellan, Jr., "Comments on the 'Lens Model' Equation and the Analysis of Multiple-Cue Judgment Tasks," Psychometrika, 38 (March 1973), 87-100.

Horace B. English, A Student's Dictionary of Psychological Terms, fourth edition (New York: Harper Brothers, Inc., 1934).

Martin Fishbein and Icek Ajzen, Belief, Attitude, Intention, and Behavior: An Introduction to Theory and Research (Reading, Mass.: Addison-Wesley, 1975).

Charles D. Goldfine and Karen E. Jakubowski, "Store Image, Shopping Behavior, and the Halo Effect," unpublished paper, 1977.

Leon Festinger, A Theory of Cognitive Dissonance (Evanston, Iii.: Row-Peterson, 1957).

Del I. Hawkins, Gerald Albaum, and Roger Best, "Reliability of Retail Store Images as Measured by the Stapel Scale," Journal of Retailing, 52 (Winter 1975-1976), 31-38, 92-93.

Joel C. Huber and William L. James, "A Theory of Halo," working paper, Krannert School of Industrial Administration, Purdue University, March 1976.

Joel C. Huber, "A Concept of Halo," in Advances in Consumer Behavior, 5 (1977 Proceedings).

Arun K. Jain and Michael Etgar, "Measuring Store Image Through Multidimensional Scaling of Free Response Data," Journal of Retailing, 52 (Winter 1976-1977), 61-70, 95-96.

Journal of Retailing, Special Store Image Issue, 50 (Winter 1974-1975).

John H. Kunkel and Leonard L. Berry, "A Behavioral Conception of Retail Image," Journal of Retailing, 32 (October 1968), 21-27.

P. F. Lazarsfeld and N. W. Henry, Latent Structure Analysis (Boston: Houghton-Mifflin, 1968).

V. Parker Lessig, "Relating Multivariate Measures of Store Loyalty and Store Image, "in Proceedings, American Marketing Association, 1972.

Jay D. Lindquist, "Meaning of Image," Journal of Retailing, 50 (Winter 1974-1975), 29-38, 116.

Eleanor G. May, "Image Evaluation of a Department Store," Marketing Science Institute, Cambridge, 1971.

William L. Moore and William L. James, "A Study of the Amount of Halo in the Perception of Automobiles," in Advances in Consumer Research, 5 (1977 Proceedings).

Dov Pekelman and Subrata K. Sen, "Mathematical Programming Models for the Determination of Attribute Weights," Management Science, 20 (April 1974), 1217-1229.

Gene H. Peterson, "Design of Image Research," unpublished working paper, Deluxe Check Printers, Inc., St. Paul, January 1976.

M. J. Rosenberg, C. I. Hovland, W. J. McGuire, R. P. Abelson, and J. W. Brehn, Attitude Organization and Change, An Analysis of Consistency Among Attitude Components (New Haven: Yale University, 1960).

L. G. Schiffman, J. F. Dash, and W. R. Dillon, "The Contribution of Store-Image Characteristics to Store-Type Choice," Journal of Retailing, 53 (Summer 1977), 3-14, 46.

Alok Sharma and Peter Doyle, "A Model for Strategic Positioning in Retailing," in Barnett A. Greenberg and Danny N. Bellenger, eds., Contemporary Marketing Thought (Chicago: American Marketing Association, 1977), 10-14.

Allan P. Schocker and V. Srinivasan, "A Consumer-Based Methodology for the Identification of New Product Ideas," Management Science, 20 (February 1974), 921-937.

William A. Staples and William B. Locander, "Behaviorally Anchored Scales: A New Tool for Retail Management Evaluation and Control," Journal of Retailing, 52 (Winter 1975-1976), 39-48, 94-96.

Edward L. Thorndike, "A Consistent Error in Psychological Ratings," Journal of Applied Psychology, 4 (January 1920), 25-29.

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