The Identification of Consumer Judgmental Combination Rules: Statistical Prediction Vs. Structured Protocol

ABSTRACT - Two approaches to the identification of consumer judgmental information processing rules, the statistical prediction method, and the modified structured protocol method were used in a laboratory setting. The degree of consistency between the judgmental rules identified by these two approaches was examined in a number of analyses. A significant degree of consistency between the two methods was found, although the two methods were not perfectly consistent for a substantial portion of the sample. Possible sources of biases in the two methods were examined to evaluate their validity, and suggestions were made for future research.

Citation:

C. Whan Park and Charles M. Schaninger (1976) ,"The Identification of Consumer Judgmental Combination Rules: Statistical Prediction Vs. Structured Protocol", in NA - Advances in Consumer Research Volume 03, eds. Beverlee B. Anderson, Cincinnati, OH : Association for Consumer Research, Pages: 184-190.

Advances in Consumer Research Volume 3, 1976      Pages 184-190

THE IDENTIFICATION OF CONSUMER JUDGMENTAL COMBINATION RULES: STATISTICAL PREDICTION VS. STRUCTURED PROTOCOL

C. Whan Park, The University of Kansas

Charles M. Schaninger, The University of Kansas

ABSTRACT -

Two approaches to the identification of consumer judgmental information processing rules, the statistical prediction method, and the modified structured protocol method were used in a laboratory setting. The degree of consistency between the judgmental rules identified by these two approaches was examined in a number of analyses. A significant degree of consistency between the two methods was found, although the two methods were not perfectly consistent for a substantial portion of the sample. Possible sources of biases in the two methods were examined to evaluate their validity, and suggestions were made for future research.

INTRODUCTION

An increasing number of studies on consumer information processing have been generated since the late 1960's (e.g., Alexis et al., 1968; Bettman, 1970; Haines, 1969, 1970; King, 1969; Russ, 1971; Wright, 1973). These studies can be broadly classified into two types: those concerned with the structural aspects of consumer information processing (Alexis et al., 1968; Bettman, 1970; Haines, 1969; King, 1969), and those concerned with the functional aspects (Russ, 1971; Wright, 1973). The structural approaches are concerned with the structural and sequential aspects of information processing or the paths by which a consumer as a problem solver or decision maker goes about performing the evaluation task, while the functional approaches deal with the specific attribute combination rules used by the consumer for evaluative judgments.

The identification and specification of judgmental rules offers potential not only for increasing the degree of understanding of consumer judgmental processes, but also for the development of practical promotional strategies, providing an appropriate framework to the advertising manager for the adaptation and adjustment of specific advertising plans.

The present study focuses on the functional approach toward consumer information processing. In the next section, several approaches used to identify consumer judgmental models (attribute combination rules) are examined. Two of these approaches are then operationalized to identify the consumer judgmental rules used by subjects to evaluate hypothetical product profiles in an experimental setting.

MAJOR APPROACHES USED TO IDENTIFY THE CONSUMER'S JUDGMENTAL RULES

A number of approaches have been proposed (Wright, 1973) and partially used in several studies to identify the consumer's judgmental rules. This paper focuses on three major approaches: the unstructured protocol method, the structured protocol method, and the statistical prediction method.

The Unstructured Protocol Method

Unstructured protocol data are usually obtained by having individuals write down or verbalize their thought processes during decision making. Obtaining unstructured protocol data is a very tedious and time-consuming procedure. Furthermore, once obtained, the analysis is quite informal, i.e., no formal procedure has been developed for inferring and identifying various response models or rules.

For the most part, unstructured protocol data have been analyzed through content analysis. Berelson (1952) has proposed a set of requirements for an effective content coding system: (a) the categories should be mutually exclusive, (b) the categories should be collectively exhaustive, and (c) the categories should record the information as accurately as possible.

Two inherent problems exist in using unstructured protocol data to infer and identify consumer response models (rules). The first is essentially a form of respondent error, stemming from the relative inability of the respondent to recognize and accurately identify his own sequential thought processes. The second is essentially a form of interviewer error, i.e., coding error--that of assigning the respondent's description accurately to appropriate content categories. From the respondent's viewpoint, although the underlying psychological processes which occurred up to the final stages of evaluation may have been quite different, the actual judgmental rules used in the final stages of evaluation (for different underlying processes) are quite similar. This similarity at the final evaluation stage, of different judgmental processes leading up to that point, results in difficulty on the respondent's part in accurately specifying the specific type of judgmental rule(s) actually employed. Thus, the subjects' responses, from the coder's viewpoint, are ambiguous, confounding the two sources of error and possibly increasing the total magnitude of error for this method.

Despite these faults, when no a priori defined models or rules for describing the individual's decision processes have been formulated, the unstructured protocol method is appropriate.

The Structured Protocol Method

The structured protocol method is a modified version of the unstructured protocol method. It is designed to avoid these problems which arise with the use of the unstructured protocol method. The structured protocol method is constructed as follows:

1) Establish the content categories. Each category represents a type of judgmental rule.

2) Each content category contains a detailed explanation of a judgmental rule corresponding to that particular category.

3) Let the respondent select the content category which represents his thought processes most accurately.

Thus, the respondent's judgmental rule is identified on the basis of his selection of one of several statements (categories) describing judgmental rules, rather than on the researcher's classification.

The structured protocol method is not without inherent problems either. Due to the researcher's specification of "rational" decision making categories, the structured protocol method may measure what the respondent ought to do rather than what he did, i.e., he may choose a more "rational" response because he likes to think himself (or wants to appear) as rational, or because he cannot remember or identify the earlier stages of his underlying cognitive processes.

One way to avoid or reduce the impact of these biases is to combine both the unstructured and structured protocol methods. A pilot study conducted by the author found that response bias problems were reduced by combining both methods: utilizing the unstructured protocol method first, followed by the structured protocol method. This combined form of the unstructured and structured protocol methods was used in this research study, and is referred to hereafter as the Modified Structured Protocol Method. The unstructured protocol method is useful in making the respondent more aware of his own judgmental thinking processes, thus enhancing his ability to accurately select the appropriate statements (categories) representing his thought processes in the structured protocol method which follows. For the modified structured protocol method, the identification of the respondent's judgmental rule is based not on the unstructured protocol data, but on the structured protocol data which follows.

The Statistical Prediction Method

The statistical prediction method identifies the consumer's judgmental rules by selecting that judgmental rule which yields the highest correlation between the individual's actual and predicted judgment.

Two major interpretational problems may exist with the statistical prediction method: (1) mathematical functional equations developed to represent different underlying cognitive judgmental processes may be algebraic equivalents, and (2) different types of mathematical functions may predict the person's actual judgment with equal effectiveness. These two subproblems are referred to as the "paramorphic" problem by Hoffman (1960). Given the existence of this "paramorphic" problem, it is necessary to institute careful controls on both the informational input (stimulus) and the judgmental evaluation (response) dimensions, when using the statistical prediction method to identify the consumer's underlying judgmental rules.

Four functional equation, corresponding to judgmental models (rules) which are examined by the statistical prediction method in this study, are specified below.

EQUATION

where

E = individual's evaluation of an alternative based on its profile of information related to product attributes;

X_i = attribute possession score of the ith element in that profile of information;

C_i = attribute possession score of the ith salient element (evaluative criterion) in that profile of information;

W_i = subjective importance of the ith element;

min C_i = a minimum of the possession scores on evaluative criteria (salient attributes);

max C_i = a maximum of the possession scores on evaluative criteria.

Of the approaches used to identify the underlying judgmental rules or models, the modified structured protocol method and the statistical prediction method offer the most promise. Yet, both present possible problems of interpretation and measurement biases. This study utilizes both of these methods, and examines the degree of consistency between them.

Several approaches are used to examine the degree of congruency between the judgmental rules identified by the two methods. The degree of binary consistency between the judgmental rule identified by both methods of identification is examined first, to test the significance of the number of subjects identified as using the same judgmental rule by both the statistical prediction and modified structured protocol methods.

The second approach to examine the degree of consistency focuses on the correlations between predicted and actual judgments. Although the predicted--actual correlation for the judgmental model identified by the structured protocol method cannot be higher than that for the statistical prediction method, it should be higher than the correlations for the other judgmental models not identified by either method. Therefore, the difference between the highest correlation (statistical prediction) and that for the judgmental rule identified by the modified structured protocol method is compared to the differences between the highest correlation and the correlations for the other judgmental rules not identified by either method. Thus, the relative predictive accuracies of the two methods are examined.

The third approach used to examine the degree of consistency between the judgmental rules identified by the two methods of identification is an examination of the consistency between the predicted judgments based upon the judgmental rules identified by the structured protocol and statistical prediction methods, using correlational analyses.

The relationships between the judgmental rule identified and the variables of perceived product complexity and prior familiarity are examined for both methods of identification, and the relationships found are compared to each other and to theoretical formulations, to determine the degree to which the relationships found for both methods of identification are consistent with theory and with each other.

METHODOLOGY

Questionnaire data were collected from two hundred and eighty-four junior and senior undergraduate students enrolled in marketing management courses at the University of Illinois and the University of Iowa, in a controlled classroom setting. Each respondent was presented with a questionnaire providing a critically controlled profile of the ratings of eight product attributes for each of eight hypothetical brands in the product class randomly assigned to him. Based on a previous pilot study, seven consumer products and their attributes were chosen for this study (hamburger, automobiles, toothpaste, stereo cassette decks, exterior trim paint, suntan preparations, and automobile tires). These products were selected to allow a sufficient variation on the dimensions of familiarity and complexity, to be described in the following paragraphs. The respondent was asked to rate the importance of each of the eight attributes for the product class assigned to him, on a seven point scale, ranging from "the most important to me" (7) to "not important to me at all" (1).

Any product criterion (attribute) with an importance score of greater than four was operationally defined as an evaluative criterion. All products with five or more choice criteria were classified as high in complexity; those with less than five as low in complexity. A previous pilot study indicated that this cutoff value was highly associated with respondents' binary classifications of products into high and low complexity groups.

The degree of the individual's prior familiarity with a product was measured by the respondent's agreement with one of three statements assessing the degree to which he had established a set of evaluative criteria for product evaluation. The three statements were the operationalized versions of the three levels of learning phases (extensive, limited and routinized learning) discussed by Howard and Sheth (1969).

Each subject was asked to evaluate along a seven point scale (poor to excellent) for each of eight hypothetical brands in the product class assigned to him. Immediately following each evaluation the subject was asked to specify (in writing) his judgmental processes. Then five structured protocols (the conjunctive, disjunctive, weighted, unweighted linear-compensatory model and none of the above four models) describing each judgmental rule were presented to the respondent. He was asked to identify that particular process which he utilized in evaluating the eight brands.

The critically controlled product attribute profiles were designed to control the informational input dimension on the basis of the following two criteria:

1) The rating on each product attribute was designed to reflect no consistent bias (favorable or unfavorable) across all product attributes.

2) Not all product attributes which were likely to be important were related favorably (or unfavorably).

These controls were used to minimize the likelihood of different judgmental rules producing the same judgment score, and to make sure that the conjunctive and disjunctive models were as likely to be used as the weighted or unweighted linear-compensatory models.

Judgmental evaluations were obtained for eight brands to control the judgment (response) dimension, in order to achieve a balance between response reliability and statistical reliability. Although it may be argued that a larger number of brands would result in a greater degree of statistical reliability for the correlational analysis used to identify the individual's judgmental model or rule, it was felt that the critical control on the stimulus dimension would reduce this problem. Furthermore, the response reliability resulting from a larger number of brands (judgments) was expected to be substantially lower, due to intentional refusal to put up with the time-consuming procedure, confused or ambiguous responses resulting from the intensive mental stress required to evaluate an unreasonably large number of brands.

RESULTS

As stated previously, two different approaches were used to identify the respondent's judgmental rules: the statistical prediction method, and the modified structured protocol method. The criterion used to identify the respondent's judgmental rules for the statistical prediction method was the highest of the four correlations between the actual and predicted judgmental evaluation (of the four models presented previously) on each of the eight brands evaluated. Subjects whose highest correlations were less than .3 were dropped from the analysis presented in this paper because it was felt that none of the four judgmental models adequately represented the judgmental processes of these subjects. These (highest) correlations ranged from .3 to 1.0 and were densely clustered between .7 and 1.0.

TABLE 1

CROSS-TABULATION OF THE JUDGMENTAL RULES IDENTIFIED BY THE STRUCTURED PROTOCOL AND STATISTICAL PREDITION METHODS

Table 1 presents the cross classification of the judgmental rules identified by the statistical prediction and structured protocol methods. The diagonal frequencies in this table represent the number of respondents demonstrating perfect binary consistency. These observed proportions may be tested against the proportions expected by chance. Two approaches were considered to determine the chance proportions:

1) Given the judgmental rule identified by the structured protocol method, the proportion of .25 may be assigned as that expected by chance for the proportion falling in each diagonal cell.

2) The expected proportions may be calculated for each diagonal cell by calculating joint probabilities based upon the product of the proportion observed for a given judgmental rule for the statistical prediction method times that observed for the structured protocol method.

The former approach is reasonable only if the expected overall proportions (column and row subtotal proportions) for the four decision rules are equal. For both methods of identification, the observed proportions were significantly different from .25 at an alpha of .01 for all four decision rules. Therefore, it was decided to use the latter approach to determine the chance proportions, treating the observed proportions as estimates of the chance proportions for each decision rule's identification. These expected proportions for each diagonal cell and for the total diagonal are presented at the bottom of Table 1. None of the z-values was significant for a one-sided alpha of .05, although the z-values for the overall diagonal and for the unweighted linear compensatory model were significant at .10. Thus, only a small degree of consistency between the judgmental rules identified by the two methods was demonstrated for this binary analysis.

The above analysis does not provide insight about the nature of the differences between the judgmental rules identified by the two methods, and is a very stringent examination of the degree of consistency. Even if the rules identified by the two methods were different, the correlations between respondents actual and predicted judgments may be quite similar for the decision rules identified by the two methods, and the predicted judgments for the two methods may also be quite consistent.

If the mean difference score (d₁) between the correlations based on statistical prediction (the highest correlation) and structured protocol is less than that (d₂) between the highest correlation and the next highest correlation not identified by either method, and less than that (d₃) between the highest correlation and the lowest of the correlations not identified by either method, further evidence of the relative consistency between the two methods is obtained.

It is necessary to apply adjusted Fisher's z-transformations to normalize the skewness of the four correlations in order to apply the t-test of differences. The formula used for transformation was z = [0.5 Zn((l+ r)/(1 - r))] - [r/(2(N - 1))].

To examine the above hypotheses (d₁ < d₂, d₁ < d₃), t-tests of differences were run for the differences between d₁ and d₂ and between d₁ and d₃. Both of these differences were significant at an alpha of .001 in the expected directions (for (d₂ - d₁), t_diff = 3.81; for (d₃ - d₁), t_diff = 16.81). Thus, the correlations between predicted and actual judgments for the modified structured protocol method were significantly higher than those for the decision rules not identified by either method of identification, demonstrating consistency with regard to predictive accuracy between the two identification methods.

TABLE 2

DISTRIBUTION OF INDIVIDUAL CORRELATIONS BETWEEN EIGHT PREDICTED JUDGMENTS BASED ON THE DECISION RULE IDENTIFIED BY THE STATISTICAL PREDICTION METHOD AND THAT IDENTIFIED BY THE STRUCTURED PROTOCOL METHOD

TABLE 3

CROSSTABULATION BETWEEN THE LEVEL OF PRIOR FAMILIARITY AND THE DECISION RULE IDENTIFIED BY THE STRUCTURED PROTOCOL AND STATISTICAL PREDICTION METHODS

Table 2 presents the distribution of the correlations between the judgments predicted by the decision rule identified by the statistical prediction method and those predicted by that identified by the modified structured protocol method. An examination of these correlations demonstrates that 73.6% of the sample had correlations greater than .621, the critical level of the correlation at the .05 level. Similarly, 55.4% of the reduced sample (excluding the 102 subjects with perfect correlations, the same judgmental model identified by both methods), had correlations greater than the .621 critical level. Thus, there is a relatively high degree of consistency at the individual level between the predicted judgments of both methods of identification.

TABLE

CROSS-TABULATION BETWEEN THE LEVEL OF PERCEIVED PRODUCT COMPLEXITY AND THE DECISION RULE IDENTIFIED BY THE STRUCTURED PROTOCOL AND STATISTICAL PREDICTION

Several authors have suggested theoretical formulations concerning the relationships between an individual's information processing modes and his familiarity with the object of concern and the degree of complexity of cognitive structure toward an object (Rosenberg, 1968; Brunet et al., 1962; Ostrom and Brock, 1969). The relationships between judgmental rules with prior familiarity and with perceived product complexity are examined and compared for both methods of identification. Tables 3 and 4 present cross tabulations of the relationships between the judgmental rules identified with prior familiarity and product complexity, respectively, for both the structured protocol and statistical prediction methods. The top value in each row represents the actual frequency for the structured protocol method, while the value directly underneath it represents the frequency for the statistical prediction method; the expected frequencies are presented in parentheses.

Three Chi square values are presented with each table: the value for the judgmental rule identified by the structured protocol method; the value for the statistical prediction method; and a third value representing the significance test of the equality of two Chi square distributions, as suggested by Mood et al. (1974). This modified Chi square test is a form of the test for the equality of two multinomial distributions, and can be expressed as follows:

EQUATION

with (jk) - 1 df

Nijk = the observed frequency in cell jk for distribution i;

i = the number of distributions compared

j = the number of columns (levels of familiarity)

k = the number of rows (types of judgmental rules)

If the relationships between the judgmental rule used by the individual with prior familiarity and product complexity are equivalent (consistent) for both methods of identification, further evidence is provided for the relationships (between judgmental models with familiarity and complexity), and for the consistency between the two identification methods, although aggregate analysis does not necessarily represent the underlying relationships at an individual level analysis.

Theoretical speculations (Rosenberg, 1968; Park, 1974) suggested that the unweighted linear-compensatory model was more likely to be used when an individual is unfamiliar with a product (class), and that the weighted linear-compensatory model was more likely to be used when an individual is quite familiar with a product (class). Although neither Chi square value was significant, the results of both methods were in the direction of the theoretical formulations. The modified Chi square value testing the equality of the two distribution was not significantly different. A closer examination of the two results reveals that the statistical prediction method exhibited stronger results for those classified as low in familiarity (they were more likely to use the unweighted and less likely to use the weighted linear-compensatory model), while the structured protocol method exhibited stronger results for those high in familiarity (they were more likely to use the weighted and less likely to use the unweighted linear-compensatory model).

Bruner et al. (1962) introduced two types of information processing, careful-but-slow strategy and risky-but-fast strategy, when referring to the conjunctive and the disjunctive models. It is strongly believed that the risky-but-fast strategy is not an appropriate strategy for the person whose cognitive structure toward a product is relatively complex (when a large number of product attributes are viewed as important). An examination of the relationships for both methods of identifying decision models reveals that significant relationships were found for both methods. But, only the results for the statistical prediction method were in accord with the theoretical speculations. The Chi square value for the equality of the two distributions was not significant, although the interpretation of the findings of the two Chi square distributions would be different. The data for the structured protocol method did not indicate a relationship between complexity and use of the disjunctive model, while the data for the statistical prediction model did. The results of the structured protocol model indicated that those high in complexity were more likely to use the unweighted and less likely to use the weighted linear-compensatory model than those low in complexity. A comparable tendency was found with regard to use of the unweighted linear-compensatory model for the statistical prediction method, but not for the weighted model.

It should be noted that subjects were classified as high in complexity if five or more (out of eight) attributes were rated as high (above 4 on a 7 point scale) in importance. One feasible explanation for the differences in use of the unweighted and weighted linear-compensatory models between those grouped as high versus low in complexity may be that those high in complexity dealt with a larger number of "evaluative'' criteria which they perceived as being more equal in importance (all high in importance) than those classified as low in complexity. The greater number of attributes considered as important, and the lesser degree to which the importance ratings differed across attributes, may have led to the greater reported usage of the unweighted and lower reported usage of the weighted linear-compensatory models by the subjects for the structured protocol method. The lack of difference for the weighted model for the statistical prediction method may have been due to the statistical artifact problem.

DISCUSSION

Weak evidence of consistency between the results of the statistical prediction and structured protocol methods was found when their binary consistency was examined. However, when the correlations between predicted and actual judgments were examined, it was found that the correlations for the decision rules identified by the modified structured protocol method were higher than those for the other decision rules not identified by either method. It was also found that high positive correlations existed between the predicted judgments for the judgmental rules identified by the two methods. Both of these findings indicate a significant degree of consistency between the decision rules identified by the modified structured protocol and statistical prediction methods. No significant differences were found between the relationships of judgmental rules with the theoretical constructs of prior familiarity and perceived complexity for the two methods of identification, although examination of the two sets of relationships could lead to different conclusions about the nature of the relationships. The results for the statistical prediction method were more consistent with theoretical speculations.

More detailed information concerning the underlying judgmental processes is needed to determine which of the two methods of identification is more valid. Both approaches have sources of possible bias. For the modified structured protocol method, respondents may be unable to differentiate one model from others while tracing back their thought processes; or their preferences for "rational" or thought oriented decision-making rules may influence their responses; or the set of judgmental rules used may be less than comprehensive. The statistical prediction method may be subject to biases due to: the paramorphic and statistical artifact problems (Hoffman, 1960); incomplete sets of evaluative criteria presented by the researcher; or halo effects of researcher-defined "salient" criteria. Incomplete specification of the full set of judgmental rules, or combinations of rules, as suggested by a number of authors (Russ, 1971; Wright, 1975; Dawes, 1964; etc.) may lead to biases in the results of both the statistical prediction and modified structured protocol methods.

A more process oriented approach may overcome some of these biases. Modifications of the discrimination net approach (Bettman, 1970; Clarkson, 1961; Swinth, 1975) may prove fruitful if modified to allow the examination of the specific impact of each stimulus on evaluations or choices and, if modified, to allow the incorporation of configural processes or compensatory models. The determination of the salience of various stimuli would be more accurate when examined in a decision net process oriented approach than with questionnaire measures such as importance scales. The response bias problems due to preference for "rational" judgmental rules and to difficulty in differentiating one model from others, inherent in the structured protocol method may also be eliminated with such a process oriented approach. Such an approach would also provide information about the sequential or multi-stage aspects of decision processes which utilize combined characteristics of two or more models. Although such process-based information is costlier and more time consuming to collect, and harder to make generalizations with, such information would permit the evaluation of the validity of other identification approaches such as the statistical prediction and structured protocol methods.

Wright's (1975) approach of training subjects to use different attribute combination rules may also be quite fruitful if combined with actual choice situations where respondents are asked which combination rule, or sequential combination of rules they actually used. This approach allows the researcher to control for the number of alternatives, and in actual choice situations, allows the individual to specify which evaluative criteria (and their number) he actually used. Information obtained in such a setting would alleviate some of the biases due to measurement problems, encountered with the statistical prediction method in this study.

Furthermore, after such training, it would be possible to utilize questionnaire measures to determine the number of alternatives, number of salient attributes, and sequential combination rules reported as typically used, by the individual for a number of product classes, and to investigate their interrelationships.

REFERENCES

M. Alexis, G. Haines, and L. Simon, "Consumer Information Processing: The Case of Women's Clothing," Proceedings (American Marketing Association Fall Conference, 1968), 197-205.

B. Berelson, Content Analysis in Communication Research (Glenco, Illinois: The Free Press, 1962).

J. R. Bettman, "Information Processing Models of Consumer Behavior," Journal of Marketing Research, 7 (August, 1970), 370-376.

J. S. Brunet, J. J. Goodnow, and G. A. Austin, Study of Thinking (Science Editions, 1962).

G. P. Clarkson, "Decision-Making in Small Groups: A Simulation Study," Behavioral Science, 13 (July, 1968), 288-305.

G. H. Coombs, Theory of Data (New York: John Wiley and Sons, 1964).

R. M. Dawes, Toward a General Framework for Evaluation (Ann Arbor: University of Michigan, Department of Psychology, 1964)

R. M. Dawes, "Social Selection Based on Multidimensional Criteria," Journal of Abnormal and Social Psychology, 68 (January, 1964), .104-109.

L. R. Goldberg, "Simple Model or Simple Process? Some Research on Clinical Judgments," American Psychologist, 23 (July, 1968), 483-496.

G. H. Haines, Jr., "Process Models of Consumer Decision Making," presented at the Third Annual Conference on Buyer Behavior, Columbia University, May 22 and 23, 1969.

P. J. Hoffman, "The Paramorphic Representation of Clinical Judgment," Psychological Bulletin, 57 (March, 1960), 116-131.

J. A. Howard and J. A. Sheth, The Theory of Buyer Behavior (New York: John Wiley and Sons, 1969).

R. H. King, "A Study of the Problem of Building a Model to Simulate the Cognitive Processes of a Shopper in a Supermarket," in G. H. Haines, Consumer Behavior: Learning Models of Purchasing (New York: Free Press, 1969), 22-67.

A. Mood, F. A. Graybill, and D. C. Boes, Introduction to the Theory of Statistics, Third Edition (New York: McGraw-Hill Publishing Company, 1974).

T. Ostrom and T. Brock, "Cognitive Bonding to Central Values and Resistance to a Communication Advocating Changes in Policy Orientation," Journal of Experimental Research in Personality, 1 (July, 1969), 30-41.

C. Whan Park, "An Exploration of the Consumer's Judgmental Rules," unpublished Doctoral Dissertation, University of Illinois, 1974.

M. J. Rosenberg, "Discussion: Information Processing and the Evaluation of New and Old Subjects," in R. P. Abelson et al., Theories of Cognitive Consistency: A Sourcebook (Chicago: Rand McNally, 1968).

F. A. Russ, "Evaluation Process Models and the Prediction of Preference," Proceedings (Second Annual Conference, Association for Consumer Research, 1971), 256-261.

R. L. Swinth, "A Decision Process Model for Predicting Job Preferences," Faculty Working Paper, School of Business, University of Kansas, 1975.

P. L. Wright, "The Simplifying Consumer: Perspectives on Information Processing Strategies," presented at the American Marketing Association, Doctoral Consortium, East Lansing, Michigan, August, 1973.

P. L. Wright, "Analyzing Consumer Judgment Strategies: Paradigms, Pressures and Priorities," Faculty Working Paper 094, College of Commerce and Business Administration, University of Illinois at Urbana-Champaign, 1973.

P. L. Wright, "Consumer Choice Strategies: Simplifying vs. Optimizing," Journal of Marketing Research, 12 (February, 1975), 60-77.

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