An Idiothetic Analysis of Attitude-Behavior Models

James Jaccard, SUNY - Albany
Gregory Wood, SUNY - Albany
ABSTRACT - Limitations of nomothetic treatments of psychological data are discussed relative to consumer decision making. A framework is described for circumventing the limitations.
[ to cite ]:
James Jaccard and Gregory Wood (1986) ,"An Idiothetic Analysis of Attitude-Behavior Models", in NA - Advances in Consumer Research Volume 13, eds. Richard J. Lutz, Provo, UT : Association for Consumer Research, Pages: 600-605.

Advances in Consumer Research Volume 13, 1986      Pages 600-605

AN IDIOTHETIC ANALYSIS OF ATTITUDE-BEHAVIOR MODELS

James Jaccard, SUNY - Albany

Gregory Wood, SUNY - Albany

ABSTRACT -

Limitations of nomothetic treatments of psychological data are discussed relative to consumer decision making. A framework is described for circumventing the limitations.

INTRODUCTION

The need to develop methods for studying individuals has been evident for nearly fifty years. The vast majority of research in the social sciences has focused on aggregate level analyses, in which inferences are drawn about groups of individuals, considered as a whole. As Allport (1937) noted, the understanding of behavior at the aggregate level does not necessarily yield understanding of behavior at the individual (idiographic) level. When nomothetic (aggregate-based) principles are applied to individuals, there frequently exists considerable error. Because of such limitations, Allport suggested the creation of methodologies that would allow the social scientist to study the behavior of individuals.

Allport's call for the development of idiographic methods was greeted with skepticism. At the time, logical positivism and behaviorism were the dominant paradigms within the social sciences, and both approaches viewed the primary aim of science as the development of universal laws. Because of this, many critics felt that idiographic explanations of behavior were outside the realm of science. This skepticism was fueled by Allport s linking of idiographic approaches to disciplines such as history and literature, with the suggestion that methods for collecting data in these domains (e.g., analysis of letters, subjective biographical reports) be used in developing an idiographic science of human behavior. Allport's approach was criticized because of (a) the inability of the researcher to develop generalizable results, (b) difficulties with using single subject experimental designs, and (c) a lack of adequate methods for conducting idiographic research (for a review of these criticisms, see Runyan 1983). Despite these criticisms, Allport s motivation for suggesting new methodologies remained: Explanations of behavior based on aggregate or nomothetic methods of analysis fell short of explaining why a given individual performed behavior X in response to stimulus Y in context Z.

Since 1937, the terms "idiographic" and "nomothetic" have taken on a wide variety of meanings in the scientific literature. This proved so confusing that in 1962, Allport clarified his goals for the study of individual behavior in personality and labeled methodologies developed toward this end as "morphogenic." However, this new methodological description did not receive wide acceptance, and the terms "idiographic" and "nomothetic" are still used in a multitude of ways in the current literature (see Jaccard and Wood 1985 for an elaboration). The term "idiographic" invites controversy because of the multiple meanings attached to it. In the present paper, we will follow the suggestion of Lamiell (1981) and label the methodologies discussed in this chapter as idiothetic, a blending of the terms "idiographic" and "nomothetic." In the spirit of Allport, we recognize that idiothetic methods should help us to understand and to predict the behavior of individuals. The researcher should be capable of demonstrating the reliability and validity of findings derived from these methods. Further, these methods should not rely on group norms or other nomothetic information in their analysis of the individual. We to not deny (or minimize) the important contribution of nomothetic procedures to the social sciences and the analysis of consumer behavior. Ideally, information collected and analyzed at the individual level should be capable of later analysis at the nomothetic (aggregate) level. In the sections that follow, we will discuss some of the limitations of nomothetic procedures in constructing theories about the individual consumer. We will then briefly outline an idiothetic approach to consumer decision making which we have found to be useful in numerous applied settings.

Limitations of Nomothetic Treatment of Data

Consider the following experimental design in a study of consumer attitudes: A researcher is interested in whether social class is related to attitudes towards a given product and whether this attitude is differentially influenced by perceptions of the cost of the product in upper class as opposed to lower class individuals The researcher begins by developing a set of questions that will measure attitudes cowards the product in two groups, one consisting of upper class individuals and the other consisting of lower class individuals. He/she does this by consulting Osgood's cables of adjective descriptors that load on an evaluative factor via the semantic differential technique (e.g., good-bad, beautiful-ugly, clean-dirty, tasteful-distasteful). Each individual rates the product on the scales and an overall attitude score is derived by summing the responses across items. In addition, the product is rated on an inexpensive-expensive dimension. The mean attitude scores are contrasted in the two groups via a t test, and a statistically significant difference is observed, with lower class individuals revealing a higher (more positive) mean score. A correlation between the attitude score and the inexpensive-expensive rating is calculated for each group, separately. For lower class individuals, there is a moderate negative correlation between the two variables, whereas for upper class individuals, there is a moderate positive correlation between the variables. The difference in correlations is tested by means of the appropriate statistical test, and is found to be statistically significant. The investigator concludes that social class is related to attitudes towards the product, such that lower class individuals evaluate the product more positively than upper class individuals. In addition, the researcher concludes that the data support the proposition that perceptions of expense affect the attitudes differentially in the two groups. Upper class individuals attitudes are favorably influenced by cost, whereas lower class individuals attitudes are unfavorably influenced by cost.

Comparison of Mean Scores. A number of assumptions are required to compare mean scores to assess the relationship between social class and attitudes. First, one must assume that the rating scales are valid for both groups of individuals. To the extent that validity coefficients are low in magnitude for either group, any conclusions are called into question. Assuming valid measures (in a traditional psychometric sense), a second issue arises in the comparison of the mean scores. When an individual is asked to make a judgment about an object on a dimension, he/she must first make the judgment cognitively, and then translate that judgment onto the rating scale provided by the investigator. Two individuals might make the same cognitive judgment, but differ in their observed ratings, if they translate that judgment differently onto the rating scale. At issue is whether group differences reflect differences in the true underlying attitude or whether they reflect differences in the way individuals in the groups use and experience the rating scale. Research in psychophysics (e.g., Wegener, 1982; Upshaw 1962) suggests that the problem of response translation may be more serious than applied researchers realize.

How might an individual translate a judgment onto a numerical racing scale? One strategy suggested by psychophysicists is that the individual will position the center of the scale to correspond to the average subjective value of the stimuli he/she expects to judge. In this case, individuals who expect to evaluate stimuli with generally high values will make lower ratings than individuals who expect to rate stimuli with generally low values on the dimension in question. A second strategy, suggested by Parducci (1965), is that the individual will position the scale so that the two most extreme categories correspond to the most extreme values he/she would assign to stimuli of the type being studied. Having tone this, the individual equates the intermediate categories with a range of judgments between these extremes. Parducci discusses several situations in which this strategy would produce different ratings for two individuals, even though their cognitive Judgment is the same. In fact, Parducci has used such an interpretation as an alternative explanation for several "established" psychological phenomena. Although other plausible strategies could be elaborated, the point is that observed mean differences may reflect systematic differences in the response translation process. The assumption that all individuals use the rating scales in the same fashion, with similar experiences in mint, is required (or one must assume that the differences cancel or are irrelevant to the question at hand).

In the above example, suppose the product being rated was a certain type of appliance. One might reasonably assume that upper class individuals have had experiences with, on the average, better appliances than lower class individuals. If individuals use the first strategy identified above for response translation, then one would predict a lower mean attitude score for upper class individuals as opposed to lower class individuals. This was, in fact, observed.

Analysis of Correlations. The same issues discussed above apply with equal vigor to the analysis of correlations. First, one must assume that the scales are equally valid indicators within groups. Otherwise, differences in observed correlations may be an artifact of differential validity. Second, one must assume that all individuals within a group (but not necessarily between groups) use the response scale in the same fashion and with the same experiences as a referent. For example, consider the negative correlation between attitudes and perceived expense for the lower class individuals. Suppose individuals with more positive attitudes have tended to encounter more expensive appliances in their experiences. If all individuals within this group adopt the first response translation strategy mentioned above, then individuals with more positive attitudes would tent to rate the appliance as being less expensive than individuals with more negative attitudes (even though true differences in expense judgments might be minimal). This would yield the observed negative correlation, where possibly none exists.

These examples illustrate a frequently neglected requirement of nomothetic analyses in consumer research; namely, the assumption of a common response language across individuals. Violations of the assumption can affect either mean scores or correlations. The issue is particularly germane to consumer research that uses rating scales.

Inferences About Individuals. If the measurement assumptions noted above are met, then the theorist may still be restricted to aggregate level conclusions. To illustrate the problem in the case of correlation coefficients, consider the above study, but assume that observed correlations within social class groups are not artifacts of response language. Suppose, also, that the investigator did not measure social class. For all individuals considered together, the correlation between perceived expense and attitude would probably be near zero (assuming common mean centroids), since the investigator has mixed two groups, one exhibiting a positive correlation and the other a negative correlation. The implication would be that a change in perceived expense would not result in a change in attitude. On an aggregate level, this would be true. Shifting perceived expense downward would cause the lower class individuals to revise their attitude upward and the upper class individuals to revise their attitude downward. The net change in attitude would be minimal. But at the individual level, considerable change would occur, contrary to what one might infer from the correlation coefficient. In this instance, the correlation coefficient would only have been descriptive of individuals if all individuals "weighted" expense in the same way (sign and magnitude) in forming an attitude toward the product. This phenomenon has obvious implications for popular regression-based approaches of intention and attitude formation in the consumer literature (e.2.. models such as Fishbein, Sheth, etc.).

A second example of potentially misleading aggregate analysis for correlations is illustrated using market segments versus total market in Figure 1. In this example, the focus is on market segments versus the total market. There are three market segments (S1, S2, and S3), each of which exhibit a strong negative correlation between variables X and Y. If a correlation is computed ignoring segments, a strong positive correlation between X and Y results. One might infer from the across-segment analysis that positive changes in X will yield positive changes in Y, when exactly the opposite would occur, as revealed by the within-segment analyses.

FIGURE 1

EXAMPLE OF MISLEADING AGGREGATION IN CORRELATION

Inferences about individuals may also be problematic in the comparison of group means. This can be illustrated from an example in the family area. Researchers have asked newlyweds the number of children they intend to have in their completed family and then contrasted the expected mean number with the mean number of children the couples actually have. The correspondence is remarkably close, indicating that at the aggregate level, little change has occurred. At the individual level, however, expectations prove to be a poor match with actual behavior. Some couples have fewer children than expected, while others have more children than expected. The "overs" cancel the "unders" and the aggregate level means are quite similar. Simple examination of the means, in this case, does not permit accurate statements about the individual level. This process probably characterizes several marketing phenomena, such as the strong correspondence observed between actual prices of products and the average (calculated across individuals) perceived prices of those products (Louviere and Meyer 1981).

A second example of misleading aggregation for mean scores is illustrated in Table 1. There are four distinct segments of a market, the members of which were asked to rate each of three brands (A, B, and C) on a 10 point,"strongly dislike" (1) to "strongly like" (10) scale. Assume that each segment is homogeneous in its ratings and that individuals purchase the brand that they like the most. In this case, brand B would have a zero market share, yet if ratings are averaged across segments, brand B is the most liked. Mean score analysis would lead to the conclusion that preferences are unrelated to purchase behavior, when, in fact, a marketing strategy that changed preferences might have considerable impact on purchase behavior.

TABLE 1

EXAMPLE OF MISLEADING AGGREGATION WITH MEAN'S

A final example of potentially misleading aggregate analyses focuses on proportions. Consider two segments of a market in the analysis of brand switching, as illustrated in Table 2. For each segment considered separately, purchases at time 2 are independent of those at time 1. If the data for the segments are pooled, however, the brand switching matrix suggests a first-order process.

TABLE 2

EXAMPLE OF MISLEADING AGGREGATION WITH PROPORTIONS

In sum, inferences about individuals from group level data can encounter problems with respect to reliability, validity, response language used in interpreting scales, and within group variability. The above comments should not be taken to imply that means, proportions, and correlations in traditional consumer studies, when calculated on the aggregate level, will always be misleading about individual cases. However, they can be misleading and must be interpreted carefully when making inferences about individual behavior.

An Idiothetic Approach To Behavioral Decision Making

Jaccard and Wood (1985) describe three facets of an idiothetic approach to consumer decision making that circumvents many of the above problems. The first facet focuses on describing how consumers perceive different options and identifies idiographic perspectives on product space analysis. The second facet considers brand preferences, integrating aspects of choice theory with traditional attitude theory. The third facet concerns trade-off analyses and describes idiothetic methods for identifying how consumers trade-off the positive and negative features of a product in forming product preferences. Space limitations do not permit a discussion of each facet. We will focus our attention on the second facet.

Preference Structures. An individual s preference structure refers to his or her attitude toward each of the decision options. The preference structure is a key concept in our theoretical framework, because a person s choice is conceptualized as being, in large part, a function of the preference structure. An attitude is conceptualized in very restricted terms in our framework. It refers to the extent to which an individual feels favorable or unfavorable toward enacting a given behavioral option. A preference structure is measured by asking the individual to rate each decision option on a 10 to +10 unfavorable-favorable scale. Consistent with subjective-expected-utility (SEU) theory, it can be argued that an individual will choose to perform that option toward which the most positive attitude is held (exceptions to this principle will be discussed shortly). Thus, the attitude measures are analogous to global measures of SEU, without requiring the measurement of specific probabilities and utilities (see Jaccard, 1981). The essence of preference structure analysis is quite simple and certainly is not new to consumer psychology. Unlike traditional decision theory, however, the concept of attitude is a central construct in the framework. And, in contrast to traditional attitude theory, the focus is on within individual analysis of attitudes across competing brands, as opposed to across individual analysis. On a conceptual level, preference structure analysis represents a theory of the relationship between attitudes and behavior. It is quite distinct from traditional, nomothetic-based consumer models that attempt to explain discrepancies between attitudes and behaviors. We will elaborate these differences and then consider the applied implications of preference structure analysis.

Theoretical Implications. Several theorists have attempted to explain attitude-behavior discrepancies by suggesting that variables other than attitudes may influence behavioral decisions (or behavioral intentions, as they are called in most theories). These theories are typically stated in the form of a multiple regression equation such that

BI = w1 X1 + w2 X2 + ... + wn Xn    (1)

where BI = the intention to perform the behavior; X1 through Xn = the relevant predictor variables, of which the attitude toward performing the behavior is X1; and wl through wn = empirically determined regression weights which are said to reflect the importance of the respective variable in determining BI. According to these models, discrepancies between traditional measures of attitude and behavior occur because (1) the traditional attitude measured is usually not specific to the behavioral criterion (i.e., it is an attitude toward an object and not an attitude toward a behavior), (2) there may be other factors (e.g., normative beliefs) determining the behavior such that attitude is irrelevant, and (3) attitude may only determine intentions to perform a behavior and thus, should be predictive of the action only to the extent chat these intentions are highly related to the behavioral criterion. The present approach is consistent with point 1 (since the preference structure consists of a set of attitudes towards behaviors) and point 3. However, several differences can be highlighted with regard to point 2 and equation 1.

First, the preference structure approach differs from the regression approach in terms of an emphasis on behavioral predictability from a within-subject versus between-subject perspective. This difference may best be illustrated by considering only the first component of equation 1 (i.e., we will assume chat weights 2 through n are zero). The regression model would involve measuring different people s attitudes coward performing a behavior (e.g., using brand A) and correlating these measures with a behavioral criterion (e.g., use of brand A). It is assumed that the more favorable the attitude, the more likely it is the behavior will be performed. Table 3 presents a hypothetical example of three individuals and their attitudes towards brand A. According to the regression model, individual 1 should be most likely to use A, followed by individuals 2 and 3, respectively. In contrast, the preference structure approach suggests that behavioral prediction requires knowledge of the. distribution of attitudes across decision options. If a person possesses a positive attitude toward A , and an even more positive attitude toward B , it is unlikely he will choose A even though he has a positive attitude toward it. In Table 3, the preference structure approach would predict that individual 3 will use brand A, whereas individuals 1 and 2 will use brand B (given that the most positive attitude dictates the choice). In the regression model, it is assumed that the more positive the attitude relative to other people s attitudes, the more likely it is the individual will perform the behavior. By contrast, in the preference structure approach, it is assumed that the more positive the attitude relative to the person s attitude toward other decision options, the more likely it is the person will perform the behavior.

TABLE 3

PREFERENCE STRUCTURE ANALYSIS FOR THREE INDIVIDUALS

A second difference in the two approaches concerns the specification of factors other than attitudes that are relevant to behavioral prediction. A number of social scientists (e.g., Wicker, 1969; Triandis, 1977) have argued that a person s attitude toward an act is only one of a number of variables that influence behavior or behavioral intent. Several models have been proposed which investigate such additional factors as attitude toward the situation (Rokeach and Kliejunas, 1972), morals (Triandis, 1977), and normative beliefs (Fishbein, 1972). In the preference structure approach, the "additional variables" are attitudes towards performing the other decision options. Most regression based models do not focus on different decision options. Thus, the approaches differ in where they direct the theorist to look for other variables that can account for attitude-behavior discrepancies.

A third difference between the two approaches concerns the way in which the "other variables" are used within the theoretical network. In the regression model, the predictor variables are given weights representing the importance of the variable in determining intention. Generally, these weights are estimated via multiple regression procedures. The weighted predictor variables are summed to yield an index of predicted behavior. In the preference structure approach, no such weighting parameters are employed nor is there a summative relation among the weighted predictors. Rather, the attitudes toward performing each of the decision options are compared with one another and the option toward which the most positive attitude is held represents the predicted behavior.

One might be tempted to extend regression models of behavioral intention to include decision options, thus resulting in a more "complete" model of decision behavior. Such an approach was suggested by Ajzen and Fishbein (1969, 1980). In the context of Fishbein s model, this involves measuring Aact (attitude towards the behavior) and SN (subjective norm) toward each option and regressing an intention measure for each option onto the appropriate Aact and SN. The combination of Aact and SN would then be dictated by the results of the various regression equations.

There are several problems with this approach relative to the present one. First, the preference structure approach is idiothetic and allows prediction of an individual's behavior knowing only that individual s attitudes toward the decision options. The regression approach, in contrast, is not idiothetic. It requires that a group of individuals be studied so that least squares estimates of the weights (w1 and w2) can be obtained for any given option. This approach requires that all individuals within the group be homogeneous in terms of how they weight Aact and SN for a given option and also requires some rather stringent measurement assumptions (e.g., all individuals interpret the rating scales in the same way). Thus, deriving accurate estimates of how to weight the multiple predictors is problematic from the standpoint of psychological explanation (see Gordon 1968 for additional issues in this respect). These assumptions are not requisite in the preference structure approach, because no weighting parameters are involved. Second, if behavioral predictions are based on Aact and SN, the regression approach requires that the standard errors of estimate be low and homogeneous across all options. If the regression analysis yielded a poor goodness-of-fit for just one option, the entire analysis could be jeopardized. Finally, there is no empirical evidence to indicate that normative factors, when used in the context of the preference structure approach, will yield increased behavioral prediction over and above Aact. Initial research efforts therefore might best employ the more parsimonious perspective of focusing on just the attitude construct. This practice would be consistent with the literature on decision making and makes the assumption that the influence of other variables on behavioral decisions is mediated by attitudinal variables. Empirical support for the preference structure approach relative to regression models is presented in Jaccard (1981) and Jaccard and Becker (1985).

Preference structure analysis can also be contrasted with other formal choice models proposed in the consumer literature. Space limitations do not permit detailed comparisons. However, some general observations can be made. First, the importance of measuring attitudes (or some related concept) towards alternative brands has been widely recognized in consumer research (as opposed to psychological theories of behavioral intention which have tended to focus on the analysis of a single option). However, many applications (e.g., Wilkie and Pessimer

1973) use cross-sectional, regression-based strategies for relating the attitude indices to brand choice. Such strategies are subject to the limitations discussed above and contrast with the preference structure approach which uses a purely idiothetic method for generating predicted choice.

Second, the majority of applications in the consumer area have focused on attitudes towards "objects" as opposed to attitudes towards "behaviors" (Ajzen and Fishbein 1977). Thus, the attitude toward the brand per se (e.g. a Rolls Royce) is measured as opposed to the attitude toward actually purchasing (i.e. choosing) the brand. Research in psychology has demonstrated the superior predictive power of attitudes towards behaviors (e.g., Jaccard, Ring, and Pomazal 1977; Ajzen and Fishbein 1977). The preference structure approach uses attitudes towards behaviors.

Third, the preference structure approach conceptualizes attitude as the extent to which the individual feels favorable or unfavorable toward choosing the option in question. Measures of attitude follow directly from the attitude scaling literature, in which the attitude is measured on an unfavorable-favorable rating scale (Jaccard, Weber and Lundmark 1975). Many applications in the consumer area estimate attitude by use of compositional methods, in which perceptions about individual attributes of an option are combined (frequently using a variant of expected-utility theory) to yield an index of attitude. These approaches are generally inferior to the present measurement strategy, because they (a) make strong measurement assumptions (e.g., the presence of ratio level properties), (b) assume that all relevant attributes and no irrelevant attributes have been included in the analysis, (c) vary in the conceptualization of what the relevant attribute-level measures should be (e.g., measures might include indices of certainty, probability, evaluation, importance, salience, relevance) and/or (d) impose a combinatorial rule on the individual attribute measures which may not be valid. The resulting measures may not reflect the relevant attitude, as conceptualized in the preference structure approach.

Fourth, the preference structure approach states that an individual will choose to perform that option toward which the most positive attitude is held (however, see exceptions discussed later). Several models of choice functions have been suggested in the consumer literature (e.g., Louviere and Woodworth 1983; Currim, 1981; Batsell and Lodish 1981; Reibstein 1978), the most popular being a form of the Luce multinomial logistic function. The model states that the probability of choosing an option A from a set of options can be expressed as the (natural log) evaluation of option A divided by the sum of the (natural log) evaluations of all options considered. This function has primarily been evaluated for aggregate level phenomena, as opposed to individual choice functions. Batsell and Lodish (1981) evaluated the model at the individual level for repeat purchase decisions over time. Psychological research on the model at the individual level has found it to be unsatisfactory (e.g., Becker, DeGroot, and Marschak 1963; Tversky 1972). In its present form, the choice function of the preference structure approach is crude, but has proven to be effective. Future research is needed to better quantify and conceptualize preference structure choice functions.

Applied Implications. Preference structure analysis has numerous applied implications. Consider the case where a consumer psychologist is attempting to influence an individual to perform one of four behavioral options (purchase product A from the class A, B, C, D). The attitude toward each of the options can be measured. Depending on the distribution of the attitudes, different influence strategies would be dictated. First, it may be found that the attitudes toward options C and D are quite low and, hence, these options can be ignored. Three strategies are then possible: (1) make the attitude toward A more positive, (2) make the attitude toward B more negative, or (3) some combination of the above. The attitude scores will, in part, dictate the strategy used. For example, if the attitude toward option B is highly positive, the psychologist would be ill-advised to make the attitude toward alternative A more positive. There would probably be a "ceiling effect" and the only effective strategy would be to lower the attitude toward alternative B in conjunction with raising the attitude toward A. In contrast, if the attitude toward B was moderately positive, with the attitude toward A being slightly less positive, then any of the three strategies could be effective.

If one wanted to strengthen the decision to perform A (i.e., make the decision resistant to change), then this would involve maximizing the discrepancy between option A and option B. This could be accomplished by using any of the three influence strategies outlined above.

The preference structure approach also has implications for the selection of target populations. If the psychologist is trying to increase the number of individuals who perform A, then those individuals most likely to change their behavior will be those whose attitudes towards options A and B are roughly equal (i.e., those who are only slightly more positive toward B than A). In contrast, individuals who have much more positive attitudes toward B than A should be relatively difficult to influence because the discrepancy between attitudes is large. Initial change efforts therefore might be focused on the former group. The strategies used to induce change could differ for one group as opposed to the other. (Parenthetically, the size of the discrepancy between the two most positive options should be related to the stability of behavior over time: The larger the discrepancy, the more likely it is the behavior will be stable.)

The above comments are theoretical approximations, at best, and are made in the spirit of suggesting potential applied implications of preference structure analysis. More precise statements will be possible as research on the choice function is forthcoming.

Error Theory for Preference Structure Analysis. The comparison of attitudes across options for a single individual should address the problem of measurement error. Consumer research which has used individual choice functions to predict choice behavior has ignored the problem of unreliability of measures. In our research, we typically obtain ratings of attitudes on three separate occasions, and use the mean attitude score as an index of the "true" score. Differences between mean scores for any two options are then evaluated relative to an index of the average unreliability across options. This is accomplished using a one way analysis of variance model, in which the different options define the levels of the analysis, and replicates are the ratings in each session. Tukey-based (Kirk 1968, p. 88) critical differences are defined for pairwise comparisons of options to isolate attitude differences that can not be attributed to measurement error. This strategy is similar to the error theory of functional measurement (Anderson 1981, 1982). The major assumption is that within-cell variation reflects only random measurement error. The approach becomes problematic if the number of options is small, because the statistical tests may lack power. This can be offset by increasing the number of repeated assessments.

Non-Optimizing Decision Rules. Thus far, our statement of preference structure analysis assumes an optimizing choice process: The individual chooses that option towards which the most positive attitude is held. Although this will generally be the case, there are instances where behavior will not correspond to the most positive option. One such case can occur when the behavior is not volitional and is controlled by other people or events. In this instance, individuals may be unable to enact the option that they are most positive toward.

Our research on choice of banks has revealed a second moderating variable on the use of an optimizing rule. When individuals first move to a community, they choose the bank towards which they feel most positive. Over time, they might acquire information about a competing bank which leads them to be more positive toward that bank relative to their own bank. However, their behavior does not change because of the costs (both economic, social, and psychological) of switching from one bank to another. This suggests that the costs of switching options relative to the benefits to be gained will be an important mediator of attitude-behavior consistency from an optimizing perspective.

Nomothetic Analyses. Nomothetic (mean) level analyses, requiring minimal assumptions about a common response language, can be conducted on preference structures, if the mean comparisons are between two or more decision options (i.e. between "repeated measures" means). The necessary assumption is one of an approximately common scale unit across individuals (the origin being irrelevant). This assumption will probably be met if the measurement practices described in Jaccard and Wood (1985) are effected. Subgroup comparisons can also be made without recourse to common origin assumptions, if the focus is on group differences in the differences between means (e.g., the difference in mean attitude scores between options A and B for upper class individuals as compared with the difference between mean attitude scores for options A and B for lower class individuals).

CONCLUSIONS

We have outlined some of the difficulties with using aggregate level or across individual analyses in testing theories of consumer behavior. We have also briefly described one facet of an idiothetic approach to consumer decision making. Space limitations forced us to deal with the most simplistic aspect of our framework. However, even this facet has implications for much of the current consumer research using regression based attitude models to explain consumer behavior.

REFERENCES

Ajzen, I., & Fishbein, M. (1969), The prediction of behavioral intentions in a choice situation. Journal of Experimental Social Psychology, 5, 400-416.

Ajzen, I., & Fishbein, M. (1977). Attitude-behavior relations: A theoretical analysis and review of empirical research. Psychological Bulletin, 84, 888-918.

Ajzen, I., & Fishbein, M. (1980). Understanding social behavior. Englewood Cliffs: Prentice Hall.

Allport, G. (1937). Personality: A psychological interpretation. New York: Holt.

Allport, G. (1962). Pattern and growth in personality. New York: Holt, Rinehart and Winston.

Anderson, N. (1981). Foundations of information integration theory New York: Academic Press.

Anderson, N. (1982). Methods of information integration theory. New York: Academic Press.

Batsell, R., & Lodish, L. (1981). A model and measurement methodology for predicting individual consumer choice. Journal of Marketing Research, 18, 1-12.

Becker, G., DeGroot, H., & Marschak, J (1963). An experimental study of some stochastic models for wagers. Behavioral Science, 3, 199-202.

Currim, I. (1981). Using segmentation approaches for better prediction and understanding from consumer mode choice models. Journal of Marketing Research, 18, 301-309.

Fishbein, M. (1972) Towards an understanding of family planning behaviors. Journal of Applied Social Psychology, 2. 216-227.

Gordon, R. (1968). Issues in multiple regression. American Journal of Sociology, 73, 592-616.

Jaccard, J. (1981). Attitudes and behavior: Implications of attitudes towards behavioral alternatives. Journal of Experimental Social Psychology, 25, 41-78.

Jaccard, J., & Becker, M. Attitudes and behavior: An Information integration perspective. Journal of Experimental Social Psychology. In press.

Jaccard, J., Ring, G., & Pomazal, R. (1977). Attitudes and behavior: An analysis of specificity of attitudinal predictors. Human Relations, 30, 817-824.

Jaccard, J., Weber, J., & Lundmark, J. (1975). A multitrait-multimethod analysis of four attitude assessment procedures. Journal of Experimental Social Psychology, 11, 149-154.

Jaccard, J. and Wood, G. An idiothetic analysis of consumer decision making. In D. Brinberg and R. Lutz (eds.) Perspectives on methodology in consumer research. New York: Springer-Verlag, In press

Kirk, R. (1968). Experimental design procedures for the behavioral sciences. Belmont, Calif.: Wadsworth.

Lamiell, J. (1981) Toward and idiothetic psychology of personality. American Psychologist, 38, 276-289.

Louviere, J., & Meyer, R. (1981). A composite attitude-behavior model of traveler decision making. Transportation Research, 15b, 411-420.

Louviere, J., & Woodworth, G. (1983). Design and analysis of simulated consumer choice or allocation experiments: An approach based on aggregate data. Journal of Marketing Research, 20, 350-367.

Parducci, A. (1965). Category judgment: A range frequency model. Psychological Review, 66, 27-44.

Reibstein, D. (1978). The prediction of individual probabilities of brand choice. Journal of Consumer Research, 5, 163-168.

Rokeach,M., & Kliejunas P. (1972). Behavior as a function of attitude-toward-object and attitude-toward-situation. Journal of Personality and Social Psychology, 22, 194-201.

Runyan, W. (1983). Idiographic goals and methods in the study of lives. Journal of Personality, 51, 413-437.

Triandis, H. (1977). Interpersonal behavior. Monterey, Calif.: Brooks/Cole Publishing Co.

Tversky, A. (1972). Elimination by aspects: A theory of choice. Psychological Review, 79, 281-299.

Upshaw, H. (1962). Own attitude as an anchor in equal-appearing intervals. Journal of Abnormal and Social Psychology, 64, 85-96.

Wegener, B. (1982). Social attitudes and psychophysical measurement. Hillsdale, New Jersey: Lawrence Erlbaum Assoc.

Wicker, A. (1969). Attitudes vs. actions: The relationship of verbal and overt behavioral responses to attitude objects. Journal of Social Issues, 25, 41-78.

Wilkie, W., & Pessemier, E. (1973). Issues in marketing use of multi-attribute attitude models. Journal of Marketing Research, 10, 428-441.

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