Explaining Behavior With Weighted, Unweighted and Standardized Attitude Scores

Sherren Waung, Lever Brothers Company [Group Research Manager, Lever Brothers Company, and Doctoral Candidate at the University of Illinois.]
[ to cite ]:
Sherren Waung (1975) ,"Explaining Behavior With Weighted, Unweighted and Standardized Attitude Scores", in NA - Advances in Consumer Research Volume 02, eds. Mary Jane Schlinger, Ann Abor, MI : Association for Consumer Research, Pages: 345-356.

Advances in Consumer Research Volume 2, 1975      Pages 345-356

EXPLAINING BEHAVIOR WITH WEIGHTED, UNWEIGHTED AND STANDARDIZED ATTITUDE SCORES

Sherren Waung, Lever Brothers Company [Group Research Manager, Lever Brothers Company, and Doctoral Candidate at the University of Illinois.]

[The author wishes to gratefully acknowledge financial support from the Lever Brothers Company.]

This investigation contrasted two alternative multi-attribute attitude models in terms of their relative ability to explain and diagnose attitude and actual behavior. The specific dependent variables were represented by physician attitudes toward prescribing specific drugs and product loyalty over eleven prescription trials. Of the two alternative models, one, an unweighted beliefs formulation, appeared to generate greater explanatory power than its weighted beliefs counterpart. This occurred in the case of explaining both attitude and product loyaltY. This analysis also showed that little contribution was added by controlling for individual differences in response style via the creation of standardized scores.

INTRODUCTION

It is apparent that the extensive development, application and refinement of multi-attribute attitude models has enabled them to assume a posture of acceptability and credibility in consumer research. Yet as it has been pointed out, major issues pertaining to conceptual development, measurement and testing remain far from resolution (Wilkie and Pessemier, 1973). This investigation attempts to shed some insight into three specific issues, namely:

(1) Which model structures offer greater potential explanatory and diagnostic potential when product affect is the dependent variable?

(2) Do these same structures perform satisfactorily when behavior over time represents the dependent variable?

(3)What contribution to explanatory and diagnostic capability is made evident by the control of individual response tendencies?

BACKGROUND

The focal point of this paper is to describe the prescribing patterns of medical doctors via multi-attribute attitude models. Selection of this phenomenon for examination was generated by a number of disparate, but equally cogent concerns. One was an observation that many, indeed most marketing studies dealing with multi-attribute attitude models stopped short of relating these formulations to purchase behavior. Indeed, of the forty_two empirical studies summarized in the Wilkie and Pessemier (1973) article, only eight attempted to predict or explain actual choice decisions and even fewer did so in the context of a real world environment.

A second reason for investigating physician prescribing behavior was because it occurs in a data rich and relatively noise free environment that appears most appropriate for testing attitude behavior models. The value and extent of these characteristics are more clearly delineated if one considers how the physician product selection process might be analyzed as compared to that of the housewife as typically investigated in consumer behavior or marketing research. One basic assumption concerning consumer brand choice is that the consumer is rational (Howard and Sheth, 1969). To the extent that an individual is prone to impulse purchases or becomes satiated with a preferred brand or product, errors in predicting behavior will occur. In contrast to the housewife, decision-making by the physician is assumed to be the result of rigorous training and adherence to professional standards, both of which emphasize the formation of carefully weighed judgements. The physician being less prone to impulsive selection behavior appears to be a more suitable subject upon which to test the validity of behavioral models than the shopping housewife. The physician, when treating a common disease, often is the sole decision-maker, not likely to be directly influenced in drug selection by a professional colleague. This characteristic eliminates another potential source of error which is often difficult to control for, i.e. situations where a consumer purchase may be influenced by numerous interested individuals such as family members. The existence of a hard criterion variable, specifically recorded prescriptions, reduces sources of error that are likely to arise with purchase measures based on recall. In this regard, Parfitt (1967) has noted that attempts to recall behavior beyond the recent past produces exaggeration or oversimplification which markedly biases purchase data. Situational variables are often capable of being measured or controlled in analyzing each prescription act. Data pertaining to disease, patient age and sex, presence of drug allergies, history of previous illness and laboratory tests are available through patient records or concurrent data gathering. Numerous researchers in social psychology (Fishbein, 1967; Rokeach, 1968; Wicker, 1969) and marketing (Day, 1970; Sheth, 1970) have noted the potential of situational variables for explaining behavior not accounted for by direct attitude and multi-attribute models. Finally, since major new drug introductions are rather infrequent and most diseases (except for influenzas and respiratory infections) are not seasonal, drug markets are relatively stable. This characteristic suggests that individual prescribing patterns are likely to be more consistent over time than housewife purchases of non-durables. In addition, the high incidence of commonly treated diseases often requires physicians to repeat similar decisions within a short period of time. This short "re-prescription" cycle suggests the likelihood of an attitude change is rather small since fewer variables are likely to intervene between decisions and attitudes may have become strong and therefore resistent to change because of extensive experience with the drugs in question.

A final, but no less important basis for this research stemmed from the relative absence of key information regarding physician prescribing habits and their determinants. Attention has been drawn by the medical professional itself to the existing problems of drug selection and therapy. Dowling, a former member of the American Medical Association Council on Drugs, (1970, 1971),voiced criticism of his own medical profession for not attaining high standards in the use of therapeutic drugs. He especially deplored the reliance and dependence of his fellow physicians on commercial sources for new drug information. Two medical educators, Stolley and Lasagna (1969), emphasized the need to understand physician prescribing patterns and the rationale underlying drug selection. They attributed this need to the widespread use of new, powerful drugs, and the increasing recognition of adverse side effects. Among the questions they posed were: [P.D. Stolley and L. Lasagna, Prescribing patterns of physicians. Journal of Chronic Diseases, 22, 1969, p.2.]

(1) What are the factors leading to new drug adoption or persistent use of an older drug?

(2) Why are some drugs prescribed with the hope of producing a pharmacological effect they do not possess?

(3) Why are potent antibiotics still prescribed by a significant number of physicians for trivial infections?

and (4) Why do certain drugs retain their popularity while most expert opinion decries their use?

THE THEORY AND HYPOTHESES

Much controversy has concerned the alternative structures of the multi-attribute attitude models. Both the original Rosenberg (1956) and Fishbein (1963) paradigms from social psychology have spawned new reformulations, yet a common trend still remains among the initial models and their subsequent variations: to explain consumer predispositions toward objects (brands/products) via an appropriate weighting of decision choice-related beliefs. The formulations below represent the two main structures at issue, and have undergone some translation in the context of physician drug-related attitudes and behavior.

Model I

Ak = f (b1 [P.I.lkx V.I.1] +...+ bn[P.I.nkx V.I.n])

where:

Ak = attitude toward prescribing the kth drug

P.I.jk = the extent that the kth drug meets the jth selection criterion.

V.I.j = the value importance associated with meeting the jth selection criterion, i.e., the desirability of meeting that criterion.

Model II

Ak = f (b1 [P.I.lk] +...+ bn[P.I.nk])

A number of observations require mention. Both models incorporate the perceived instrumentality component of Rosenberg (1956), however, the variable shown is probably most closely related to Sheth's (1969) definition of evaluative beliefs. In this case, the term expresses the extent to which drug selection criteria (secondary goals) are met by the drug in question. Model I possesses the value importance term of Rosenberg (1956), however restated in terms of desirability rather than personal satisfaction. As the dependent variable, attitude relates to the act of prescribing a drug and therefore is similar to Fishbein's (1967) construct of attitude towards an act (Aact).

The advisability of including the value importance component is the issue at hand. In assessing the empirical support for and against inclusion of this term, Bass and Wilkie (1973) note that the bulk of evidence suggests importance weights are not likely to improve the explanatory capability of multi-attribute models, or decrease it as first proposed by Sheth and Talarzyk (1972). A number of recent studies, however have indicated significant increases in predictive and diagnostic power attributable to the use of normalized scores (Bass and Wilkie, 1973; Ginter, 1973; Wilkie, McCann, and Reibstein, 1973).

Given this background, the research objective is to re-examine this issue by comparing the explanatory and diagnostic power of Models I and II. The two relevant null hypotheses are:

H1 No differences exist among the proposed models in explaining physician attitudes toward prescribing the kth drug.

H2 No differences exist among the proposed models in explaining physician product loyalty towards the kth drug.

METHOD

Subjects

Fifty-five physicians from four university health centers and a local community volunteered to participate in this study.

Procedure

Self-administered questions related to the behavioral phenomenon of interest, the treatment of a specific common disease, were posed in the context of a typical prescribing situation. Specifically, the physician was asked to imagine treating a first episode case which was to be treated prior to the positive identification of the organism and its susceptibility to different drugs. Responses to the following were elicited in the framework of this prescribing situation.

(1) Affect (Direct). Respondents indicated the strength of their preference for prescribing three different drugs on 7-point scales. This measure was intended to reflect predisposition toward the act of using a drug in a well-defined situation, and not toward the drug itself in a general context.

(2) Value Importance. Respondents indicated how desirable they consider each of seven choice criteria in selecting a drug under the specific situational conditions defined.

(3) Perceived Instrumentality. Physicians were asked to indicate the degree to which each drug meets each of the drug choice criteria on a 7-point scale. Five attributes, evolved from exploratory research and pre-tests, were used.

Patient medical records were screened until Information pertaining to eleven prescribing occasions or trials were gathered. From these data two measures of product loyalty were formed, one simply being the proportion of total prescriptions devoted to each drug. The other measure was derived by calculating loyalty scores based on a factor analytic procedure proposed by Sheth (1970). This latter measure reflected, for each physician, the pattern and frequency followed in prescribing -each drug.

RESULTS

Intrasubject Reliability

Twelve perceived instrumentality judgements were repeated after completion of the main questionnaire. The mean reliability was .91, the range .38 to l.O. A final sample size of forty-five was determined as a result of these specific respondent assessments.

Validation Assessment of Affect Measure

A multitrait - multimethod approach (Campbell and Fiske, 1959) was employed to assess the validity of affect. The mean r of affect with product loyalty was .55 (p < .01) thus adding assurance that the measure was appropriate.

Testing of Hypothesis I

Prior to testing Hypothesis 1, two transformations were performed on the data in order to account for possible individual differences in response style. These within subject standardizations of raw responses were suggested by a number of considerations. Osgood, Suci and Tannenbaum (1957) in their research on the semantic differential noted numerous instances where response style differences among individuals were apparent. While discussing the benefits and pitfalls of within subject standardization, they noted that better educated subjects tend to use the intermediate positions relatively more frequently than the polar or neutral positions (Stagner and Osgood, 1946). In addition, intelligence scores appear to be related to position usage in responding to the semantic differential, subjects with lower scores being more polarized in their judgements (Kerrick, 1954). Heise (1969) in a recent review of methodological issues in semantic differential research noted the presence of important differences in response styles, particularly a tendency for some subjects to use end points more often than the discriminatory intermediate positions. Peabody (1962) in turn concluded that this propensity is a stable trait of individuals over time and over different sets of concepts. Finally, Bass and Wilkie (1973) supported hypotheses that cross-sectional regression models utilizing normalized scores would yield higher proportions of explained variance and more significant coefficients than those employing raw responses. The authors attributed their findings to the fact that: (1) preference ranks are normalized by individual, so that the distribution of the dependent variable will differ for the separate brand regressions; (2) scale differences in belief and importance ratings may not represent true scores. These two conditions result in regressions which attempt to explain brand preference (ranked) via belief and importance scores. which reflect individual response style differences.

In view of the above mentioned considerations, within subject standardizations were conducted on perceived instrumentality ratings;(l) across all three products and all five attributes so as to control for differences in response style which might be reflected across all drug criteria ratings; (2) across all three products, but within each drug criteria in order to control for response style variations in perceived instrumentality ratings of products against a given drug criteria. The option of standardizing value importance (desirability) scores within each individual was given serious consideration; however, the decision to proceed with this was made contingent upon the results stemming from the use of standardized perceived instrumentality scores. This seemed consistent with the Bass and Wilkie (1973) results which demonstrated that normalization of belief, rather than importance scores, yielded substantially greater explanatory power over raw responses.

The testing of Hypothesis 1 therefore entailed formulating the two different attitude models in terms of raw scores and two different sets of standardized scores. Then a total of six regressions of affect against different attitude formulations were performed for each drug product. Results of the 18 runs regressing affect onto the proposed models are summarized on the next page.

TABLE 1

PERCENT VARIANCE EXPLAINED IN AFFECT  (ADJUSTED R2)

The disaggregated versions provided the greatest explanatory power in physician affect toward the three drug products. The mean percent of variance explained was 37% and 18%, for Models II and I respectively. Each equation of Model II was significant with respect to explaining affect toward each drug. Specifically, 9 out of 9 Model II regression equations were significant. In contrast, Model I yielded 5 significant equations and had no single formulation which yielded a significant amount of explained variance in attitude towards each of the three products. A different assessment of relative predictive power involved F-tests for significant differences in R2. On no occasion did the weighted model generate a statistically higher R2 than its unweighted counterpart. Instead,in 4 out of the 9 contrasts, the unweighted formulation was superior. No significant differences in the mean number of coefficients were found among Models I and II. Thus each was generally equivalent in terms of this operational definition of diagnostic power. To summarize, differences in ability to explain affect among the proposed models are apparent, therefore the null hypothesis is rejected. These differences were manifest in terms of percent variance explained in affect and the number-of products towards which physician affect could be explained. No differences, however, were found with respect to the number of significant coefficients provided by the models.

Testing of Hypothesis 2

Eighteen regressions of physician product loyalty (represented by respondent factor scores) onto the hypothesized attitude models were performed.

TABLE 2

PERCENT VARIANCE EXPLAINED IN PRODUCT LOYALTY  (ADJUSTED R2)

In contrast to previous findings, the product loyalty regressions were characterized by generally low explanatory power. For example, the mean R was .42 and .39 for the two equations which provided the greatest explanation in affect. In contrast, the highest R2 attained among the product loyalty regressions was .23. Seven equations of Model II proved significant, while four formulations of Model I attained significance.In terms of significantly different R2s, once again the unweighted model held a marginal edge over the weighted version.

TABLE 3

NUMBER OF SIGNIFICANTLY HIGHER R2s AMONG LOYALTY REGRESSIONS

The differences in the number of significant coefficients generated were insignificant across models, once again demonstrating equal diagnostic capability via this narrow definition.

Because differences among the models were apparent with respect to their capability to explain loyalty the null hypothesis is rejected. As has been previously found, the disaggregated unweighted beliefs model provided the greater explanation in loyalty.

Comparison of Raw and Standard Score Regressions

A series of F-Tests were performed to determine if the R2s produced by the raw score equations were significantly nigher than those generated by the two standardized score formulations.

TABLE 4

NUMBER OF SIGNIFICANTLY HIGHER R2s

In no instance did any standard score regression produce a higher R2 than its raw score counterPart. The obverse held true in four comParisons.

Stability of Results

The limited sample size of the study necessitated a comprehensive investigation of the stability of the regression results. In particular, a procedure suggested by Mosteller and Tukey (1968) was employed for the purpose of providing refined parameter estimates which incorporated direct assessments of the variability in the data. These authors contend that major sources of variation must be assessed in deriving results, citing group or cluster variation as being a major determinant of stability, particularly in the case of small samples. The jackknife is a means which allows one to determine directly the variability in a data set by estimating the same statistic computed on several overlapping groups of data. Thus, this method was employed to asses the effect of physician sample clusters upon the aggregate regression results. This was conducted by successively eliminating each physician cluster from the total sample over a series of regression analyses. As a result of this procedure, R2s, regression coefficients and standard errors of coefficients were re-estimated for the affect and loyalty regressions. As it turned out, all re-estimated parameters were highly consistent with the original ones. In no instance were the conclusions derived from testing either of the hypotheses altered. In sum, the small sample based estimates were not altered by within group variation thereby indicating stability of the results.

SUMMARY AND IMPLICATIONS

The findings contribute further support to the contention that multi-attribute attitude models can explain behavior, in this particular instance prescribing patterns over time. While the amount of variance explained in loyalty would likely shrink in a cross validation of these results, nonetheless they appear satisfactory. This appears so given the R of .10 typically found in social psychology studies which attempted to explain or predict behavior from attitude models (Wicker, 1969).

The control of individual differences through standardization, with a few exceptions, did not markedly enhance the explanatory power of the models. This does not mean or imply that other transformations might not have improved the results. Clearly, normalization has improved results as demonstrated in a number of recent studies. A relevant point to be raised regarding transformations, however, is that as consumer researchers experiment with such data manipulation, they should continue to report all particular circumstances where model performance is or is not improved. The necessity for this stems from the recognition that sub-populations differ in their response tendencies and that whole-scale alteration of data sets may badly distort information from these key entities.

Any claims of superiority for the unweighted over the weighted model must of course be tempered by recognizing the sample size employed in this study. Nonetheless, the pattern of greater explanatory power was consistent in explaining both affect and loyalty. This type of stability ought to serve as an evaluative criteria for assessing multi-attribute models in addition to predictive and diagnostic considerations.

REFERENCES

Bass, F.M. & Wilkie, W.L. A comparative analysis of attitudinal predictors of brand preference. Journal of Marketing Research, 1973, 10, 262-269.

Bopp, J. A quantitative semantic analysis of word association in schizophrenia. Unpublished doctoral dissertation, University of Illinois, 1955.

Campbell, D.T. & Fiske, D.W. Convergent and discriminant validation by the multitrait-multimethod matrix. Psychological Bulletin, 1959, 56, 81-105.

Day, G.S. Buyer attitudes and brand choice behavior. New York: Free Press, 1970.

Dowling, H.F. Medicines for man. New York: Alfred A. Knopf, Inc., 1970.

Dowling, H.F. The prescribed environment. Saturday Review of Literature, 1971, 58-60.

Fishbein, M. An investigation of the relationships between beliefs about an object and the attitude toward that object. Human Relations, 1963, 16, 223-239.

Ginter, J.L. The effects of normalization on the multi-attribute attitude model. Paper presented at the Fourth Annual Meeting of the Association For Consumer Research, Boston, Mass., November, 1973.

Heise, D.R. Some methodological issues in semantic differential research. Psychological Bulletin, 1969, 72, 6, 406-422.

Howard, J.A. & Sheth, J.N. The theory of buyer behavior. New York: Wiley, 1969.

Mosteller, F. & Tukey, J.W. Data analysis, including statistics. In G. Lindzey & E. Aronson (Eds.), Handbook of social psychology. (2nd ed.) Vol. 2. Reading, Mass.: Addison - Wesley, 1968.

Kerrick, J.S. The effects of intelligence and manifest anxiety on attitude change through communications. Unpublished doctoral dissertation, University of Illinois, 1954.

Osgood, C.E., Suci, G.J. & Tannenbaum, P.H. The measurement of meaning. Urbana: University of Illinois Press, 1957.

Parfitt, J.H. A comparison of purchase recall with diary panel records. Journal of Advertising Research, 1967, 7, 16-31.

Peabody, D. Two components in bi-polar scales: Direction and extremeness. Psychological Review, 1962, 69, 65-73.

Rokeach, M. Beliefs attitudes and values: A theory of organization and change. San Francisco: Jossey-Bass, 1965.

Rosenberg, M.H. Cognitive structure and attitudinal affect. Journal of Abnormal and Social Psychology, 1956, 53, 367-372.

Sheth, J.N. A factor analytic model of brand loyalty. Journal of Marketing Research, 1968, 5, 395-404.

Sheth, J.N. Attitude as a function of evaluative beliefs. Paper presented at the American Marketing Association Conference Workshop, Columbus, Ohio, 1969.

Sheth, J.X. & Talarzyk, W.W. Relative Contribution of perceived instrumentality and value importance components in determining attitudes. University of Illinois, College of Commerce and Business Administration. Working paper number 15, 1971.

Stagner, R., and Osgood, C.E. Impact of war on a nationalistic frame of reference: Changes in general approval and qualitative patterning of certain stereotypes. Journal of Social Psychology, 1946, 24, 187-215.

Stolley, P.D. & Lasagna, L. Prescribing patterns of physicians. Journal of Chronic Diseases, 1969, 22, 395-405.

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

Wilkie, W.L., McCann, J.M. & Reibstein, D.J. Halo effects in brand belief measurement: Implications for attitude model development. Paper presented at the Fourth Annual Meeting of the Association For Consumer Research, Boston, Mass., November, 1973.

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

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