Drug Therapy Decision Rules Among Physicians: Decision Rule Segmentation

Phyllis A. Rosenberg, ARBOR, Inc.
Sandra K. Webster, Westminister College
ABSTRACT - Much past research on consumer cognitive processing has focused on their use of compensatory versus lexicographic decision models. Two studies of physicians' decisions among simulated drug profiles revealed a number of distinct decision rule segments rather than supporting one decision model. Implications of decision rule segmentation for future cognitive processing research and market research applications are discussed.
[ to cite ]:
Phyllis A. Rosenberg and Sandra K. Webster (1984) ,"Drug Therapy Decision Rules Among Physicians: Decision Rule Segmentation", in NA - Advances in Consumer Research Volume 11, eds. Thomas C. Kinnear, Provo, UT : Association for Consumer Research, Pages: 218-223.

Advances in Consumer Research Volume 11, 1984      Pages 218-223

DRUG THERAPY DECISION RULES AMONG PHYSICIANS: DECISION RULE SEGMENTATION

Phyllis A. Rosenberg, ARBOR, Inc.

Sandra K. Webster, Westminister College

ABSTRACT -

Much past research on consumer cognitive processing has focused on their use of compensatory versus lexicographic decision models. Two studies of physicians' decisions among simulated drug profiles revealed a number of distinct decision rule segments rather than supporting one decision model. Implications of decision rule segmentation for future cognitive processing research and market research applications are discussed.

INTRODUCTION

In recent years research on consumer decisions among multi-attribute brands has shifted from primarily statistical prediction of decisions to a greater emphasis on the cognitive processes underlying decisions (Park and Schaninger 1981; Pras and Summers 1981; Frey and Kinnear 1979). The typical research approach has been to define two or more hypothesized decision models which are then statistically fitted to consumers product decisions (e.g., Park 1975; Bettman, Capon and Lutz 1975b; Pras and Summers 1975; Nakanishi and Bettman 1974). The model which shows the greatest degree of fit (i.e., ability to predict choice behavior) is then concluded to be most representative of the cognitive process underlying the brand decisions. Two general classes of models taken from cognitive psychology have received a significant amount of attention in the consumer decision research literature compensatory and lexicographic models.

Compensatory Models

Compensatory decision models posit that decision makers attempt to process each piece of available product information (product attributes) in making choices between brands so that they can choose the brand which is most satisfactory to them on the greatest number of (or most important) attributes (Rappaport and Wallsten 1972; Coombs 1964). The models are compensatory in that positive attributes are assumed to compensate for negative ones. In other words, the decision maker is willing to trade off one (or more) attribute(s) in order to obtain the most desirable combination of those available. A number of studies focus on the specific cognitive algebra used by consumers to combine attribute information in forming judgements (additive, averaging, weighted averaging, etc., Park and Schaninger 1981; Bettman, Capon and Lutz 1975a and 1975b). However, none have yielded definitive results on the specific cognitive algebras used by consumers to combine product attribute information in making brand decisions.

Lexicographic Models

A class of alternative decision models, the lexicographic models, assumes that the decision maker bases brand decisions on the product attribute which is perceived to be most important (Coombs, 1964; Dawes 1964; Tversky 1979). If brands are equally desirable on the basis of the most important attribute, the decision maker then takes into consideration the second most important attribute, and so forth until only one brand remains. A variety of lexicographic models has been tested in the consumer judgment context (e.g., Wahlers 1979; Park and Schaninger 1931; Park 1978). However, as with compensatory models, no single lexicographic model has shown strong support across studies.

Comparison of the Models

The compensatory models suggest an active, rational cognitive process which first involves evaluation of each product attribute and then a complex combination of these evaluations to yield the most satisfactory brand decision. The lexicographic models, in-contrast, suggest a decision maker who processes only one type of information at a time and who is motivated by a desire to simplify the cognitive task.

Past research has shown that compensatory decision rules are more likely to be employed in high involvement (Howard 1977) decisions in which the decision maker is familiar with the product class and its attributes (Taschian, Taschian, and Slama 1981).

Furthermore, more educated respondents (Taschian, Taschian, and Slama 1981), and those with a higher tolerance for ambiguity (Malhotra, Pinson and Jain 1977) have been shown to be more likely to provide valid judgement data and to use compensators decision rules.

Decision Strategy Segmentation

Two types of previous consumer research have identified respondent segments based upon their choice behavior -Product Benefit Segmentation and tests of competing decision models. Neither of these approaches has been employed explicitly to identify distinct decision strategy segments. The research reported here represents an attempt to extend the Benefit segmentation approach to identification of decision strategy segments which may utilize compensatory and/or lexicographic decision rules.

The use of Benefit segmentation as a market research tool preceded the development of sophisticated multivariate methods of estimating attribute utilities from choice data. Since 1961 Benefit segmentation has been used to define homogenous market segments made up of individuals who seek the same pattern of benefits in a particular product class (Haley 1968). Haley (1968) has postulated that other types of market segments (e.o., Purchase interest, Demographic and psychographic segments) exist because individuals in these segments are similar in terms of the product benefits they seek out. The early studies based Benefit segmentation on respondent supplied direct estimates of the importance attached to specific product attributes (Johnson 1971; Haley 1968). A few snore recent studies have reported benefit segments based upon segmentation of attribute importance weights derived from simulated product rankings (Wind, Grashof and Goldhar 1978; Moriarity and Venkatesan 1978). The results of these studies have been used to suggest product positioning geared for specific "benefit segments."

The second research approach which was lead to the identification of respondent segments based upon their estimated attribute utilities has been the test of competing decision models. This approach has employed predictive choice criteria derived from competing decision models (e.o., compensatory vs. lexicographic models, Park 1978; adding vs. averaging compensatory models; Bettman, Capon and Lutz 1975a). Respondents are classified into segments based on the model which best predicts their choice behavior. Typically in these studies which do show identifiable decision strategy segments, the research objective is to identifY the most predictive decision model(s). Hagerty (1983) has recently proposed the situation model. Its basic assumption is that the situation and time will change individual attribute utiLities and therefore must be incorporated into the decision model. The extension Suggested here is that the situation may not only influence the attribute utilities but also the decision strategies employed. Furthermore, different decision strategy models are not viewed as competing, but rather, as being utilized differentially in different respondent segments.

Hypotheses

Based on the findings of past research on compensatory and lexicographic decision models, it has hypothesized that physicians would be more likely to use compensatory decision rules, particuLarly when evaluating; new drug profiles in terms of their preference for using them in therapy. This hypothesis was based on the following considerations: a) Physicians have a high level of education and have demonstrated their abilities to undertake cognitively complex tasks, D) Physicians are extremely familiar with the product categories in which they prescribe treatment and with the product attributes used to describe them, and c) Physician's drug decisions are relatively high in involvement since they are prescribing therapy for potentially life threatening medical conditions .

A second hypothesis tested in this research was that physicians would show different patterns of attribute importances in making drug decisions. It was hypothesized in the current studies that among physicians, distinct drug attribute benefit segments also exist, and that these segments may differ in the extent to which they employ compensatory decision rules.

Overview of Method

Drug decision data frown two market studies which assessed physicians' responses to two different types of drugs and which involved different drug profile designs and analysis strategies will be presented in this paper. In both studies doctors were presented with simulated drug profile cards containing specific attribute information about the major dimensions of the drugs. They were asked to rank order the drug profiles in order of their preference for therapy within a given treatment context. In each study, doctors' choice data sere first analyzed using conjoint analysis to develop individual part-worth utilities scores (importance estimates) on each attribute level included in the drug profiles. A subsequent series of analyses were then performed to determine attribute segments and potential differences in decision rules. Different analytic strategies were used in the two studies because of the differences in the structures of the decision tasks (conjoint design) and number of respondents between the two studies.

STUDY 1

Method

One hundred physicians (5() specializing in General Practice, GP's, and 50 specializing in Internal Medicine, IM's) were personally interviewed by trained medical interviewers. Interviews were conducted in 10 different, regionally distinct, U.S. cities during October of 1982. Doctors were asked to rank profiles simulating 12 different antihypertensive medications in order of their perceived appropriateness for treating advanced hypertension patients. Each profile contained information on the product attributes: a) therapy regimen (3 levels), b) dose (2 levels), and c) side effects (4 levels). Table 1 shows the product attributes and their levels. The 12 profiles were constructed using a nested hierarchical design which allowed for analysis of main effects of the therapy and dose dimensions as well as the effects of the side effects dimension nested within the therapy dimension.

Aggregate Results

Part-worth utiLities for each drug, dimension for each doctor were obtained through case by case multiple regression utilizing the hierarchical nested design Of the decision base. The analysis resulted in part-worth utility estimated (regression coefficients) and individual analyses of variance of the main effects of therapy and dose, and the effects of side effects nested within therapy (interaction). Salience scores for each dimension were derived from the part-worth utilities of the attributes within that dimension. The salience score represents an estimate of the importance of the dimension, overall, in the drug profile preferences.

Table 1 shows results of the conjoint analyses averaged over the total sample. Interpretation of these results without further analysis would lead to the following conclusions: a) Each dimension received a significant average salience score [Adjusted to account the different number of levels in each drug attribute.]) indicating that physicians were likely to consider all the drug information available to them in ranking the antihypertensive drug profiles (i.e., use compensatory decision rules); b) Therapy and side effects were about equally weighted in doctors' drug profile judgments; c) The dose attribute was also considered, but was less important than the other two characteristics.

Attribute Utility Segmentation Results

Repeated measures 2 x 3 x 4 (nested design) analyses of variance were performed on each respondent's rankings of the 12 drug profiles. The results sere used to classify the doctors into decision model groups based on the number and type of effects which were statistically significant (p.'10)) in predicting the drug profile ranking [A statistical significance criteria of p.'10 was used in order to increase the power of the statistical test.]). Respondents with only one significant effect were classified into the Single Dimension models segment; those with two significant effects into the Two Dimension Model segment; and those with all three effects significant into the Three Dimension Models segment. Then the respondents within each segment were categorized With respect to the specific effects which were related to their drug profile rankings.

Example #1 shows a respondent who was classified into the Two Dimension Models segment. This respondent's rankings were significantly related to Therapy (p'.01) and to Side Effects (p'.001), but not to Dose. Example #2 summarizes the results for a respondent, who was classified into the Single Dimension Models segment. For this respondent only the Dose dimension was significantly (pS.05) related to his/her drug profile rankings.

The results of the segmentation are summarized in Table 2. One third of the respondents showed only a single dimension which related to their drug, profile ranking. Of these doctors, the majority (21% out of 34%) showed significant main effects on the Therapy dimension. Twenty-eight percent (28%) of the doctors showed significant relationships between drug profile rankings and two drug dimensions. The two dimensions shown most frequently sere a combination of Therapy and Side Effects (18%). Only 212 of the doctors showed drug profile rankings which were significantly relate.1 to all three drug dimensions. (This segment would appear to be the best example of a compensatory decision process which takes into account all the information available to them.)

TABLE 1

PART-WORTH UTILITIES AND SALIENCE SCORES FOR THE DRUG DECISION PARAMETERS

Seventeen percent of the sample showed no significant relations between drug profiles and any of the dimensions. This segment has been labeled "Random." They may have been indifferent to the judgment task, or to the drug attributes used to describe the profiles and therefore ranked the profiles in a random process. Alternatively, they may have based their judgements on criteria other than the dimensions provided in the profiles or used decision strategies which were so comPlex as to elude identification in this analYsis.

STUDY 2

Method

Personal interviews were conducted with 230 physicians (101 psychiatrists, 77 GP's and 52 IM's) in 12 geographically distinct U.S. cities. These doctors were asked to rank eight profiles simulating antidepressant drugs in order of their preferred first line agent for treating a typical depressed patient. Each profile contained information on seven antidepressant attributes (2 levels each).

After ranking the first eight profiles, physicians were asked to rank a second set of eight profiles. These profiles differed from the previous set in that they contained information on cost but did not contain information on cardiovascular side effects [The two stage conjoint tasks allowed for an estimation of the importance of specific anti-depressant attributes which may have been masked by the inclusion of a more salient dimension. (In this study cost was hypothesized to be that salient dimension.)]). The Profiles were constructed using an orthogonal main effects fractional factorial design (Adelman 1962).

EXAMPLE #1

EXAMPLE #2

TABLE 2

A SUMMARY OF FOUR ATTRIBUTE DECISION SEGMENTS BASED ON AN ANTIHYPERTENSIVE EXAMPLE

Results

Each set of ranking data were subjected to conjoint analysis after an initial aggregate regression analysis confirmed the appropriateness of the models to be tested [Only six parameters could be estimated in the eight profile ranking tasks. Aggregate analyses showed almost identical parameter estimates in both tasks for the Dose and Sedation attributes. Therefore, Dose was included in the conjoint analysis of the first ranking and Sedation was included in the conjoint analysis of the second ranking task.]). Part-worth utilities were generated for the 12 anti-depressant attributes for each respondent (six from each ranKing task). Drug decision segments were identified through a cluster analysis [The cluster analysis technique was chosen rather than the analysis of variance technique because it better fit the eight dimension model and because the larger number or cases (230 compared to 100) allowed for cluster identification with greater statistical confidence.]) of the part-worth utilities for each attribute. A six-cluster solution showed the best discrimination between clusters and the highest homogeneity within clusters. Confirmatory discriminant function analysis based on the six cluster solution snowed that it correctly categorized 87 percent of the doctors. One resulting cluster contained only four doctors and therefore was not included in the analysis. Each of the five remaining clusters was labeled for the attribute(s) which most distinguished it from the total sample and from the other four clusters.

First Profile Judgment Task

Table 3 shows the mean part-worth utilities for each attribute estimated from the two antidepressant profile ranking tasks. The aggregate results (total sample) showed four salient attributes used in ranking the anti-depressant profiles in each of the two ranking tasks. A typical interpretation of this data would be that doctors used a modified compensatory decision model and based their judgements on a weighted combination of the information presented on Cardiovascular side effects (weighted most heavily), Dosing (moderate weight), Cholinergic side effects (moderate weight) and with lesser but some weight on the Onset attribute. A further implication is that the Excitation and Success attributes were not considered in the antidepressant decision process. An examination of the five decision clusters, however, revealed that no single cluster showed a pattern of attribute part-worth utilities similar to that of the total sample. Also, the attributes which showed very low part-worth utilities in the total sample analyses averaged high part-worth utilities for two of the five clusters.

A simple analysis of the number of attributes which contributed to the antidepressant ranking decision was accomplished by counting the number of attributes with mean salience values of .10 or more in each of the five clusters. Of the five clusters, four showed three attributes with mean saliences greater than .10. Interpretation of the aggregate analysis of the first profile ranking task indicated more attributes considered in the decision process (four compared to three) and a different pattern of attribute utilities than were observed in any of the five decision clusters.

Second Profile Judgment Task

The results of the second profile judgment task paralleled those of the first judgment task. In the aggregate anaLysis four attributes showed high part--worth utilities. In contrast, the decision clusters showed a totally different pattern of results.

First and Second Profile Task Compared

In both judgment tasks, the aggregate results would support a modified compensatory model, in which the four most important dimensions were differentially weighted. However, an interesting difference between the segmentation results of the first and second ranking tasks was the number of attributes shown to be important in the decision process. In the second judgment task four of the six attributes showed as important for all but one of the clusters. Thus, changing the attribute array on the antidepressant profile by including cost informal ion rather than information relative to cardiovascular side effects, had several effects. A most noticeable effect was that it increased the importance placed on the cholinergic side effects information. In most clusters, it also seemed to influence doctors to make decisions based on more of the information provided in the profiles. One might hypothesize then, that when given information on a particularly salient variable cardiovascular side effects - doctors tended toward a lexicographic decision process, but moved to a compensatory process when this salient variable was removed from the decision process.

DISCUSSION

In terms of theoretical implications, the results presented argue for a shift in paradigm toward the identification of segments of the population who tend to use specific decision rules or strategies in specific circumstances rather than a test of one, or more models against another. In this research well educated, cognitively sophisticated decision makers (physicians) varied in the degree to which they employed all information available (compensatory decision rules) to make decisions on a relatively high involvement product class.

The results suggest a need to recognize that individual decision makers may be flexible in their decision strategies and may change the way in which they use attribute information depending upon the specific information Available, For examPle, doctors in Study II appeared to be more willing to make trade-offs on a greater number of attributes when the most important attribute (cardiovascular side effects) was not presented.

TABLE 3

PART-WORTH UTILITIES

In terms of practical marketing, applications, the findings suggest that a complete understanding of consumer decision rules requires segmentation of the market based on decision strategies. Market research which uses the aggregate approach to cognitive processing risks development of marketing strategies targeted to the "Grand Mean Consumer," who may not in fact exist Instead market researchers should design studies which can, a) identify decision rule segments and b) identify the purchase situations in which different decision rules might be applied.

In order to identify decision rule segments market research will continue to depend upon accurate, a priori, definitions of the relevant product attributes. We as marketers may not be in a position to know what attributes are considered important to other decision rule segments (research probably tend to similar compensatory rules). In the sane vein, decision rule segmentation research should be based on samples widely representative of a product's market. For example, cognitive processing research based on MBA students, who are trained in a style of cognitive processing, would only be appropriate for products marketed exclusively to MBA students. Finally, decision rule segmentation should rely on sample sizes large enough to statistically identify segments (based in part on the number of product parameters being estimated) and to reveal final sample segments large enough to be described with adequate statistical precision.

REFERENCES

Adelman, S. (1962), "Orthogonal Main-Effect Plans for Asymmetrical Factorial Experiments," Technometrics, 4, 21-46.

Bettman, J. R., Capon, N. and Lutz, R. .T. (1975a), "Multi-attribute Measurement Models and Multi-attribute Attitude Theory: A Test of Construct Validity," Journal of Consumer Research, 1, 1-15.

Bettman, J. R., Capon, N. and Lutz, J. (1975b) "Cognitive Algebra in Multiattribute Attitude Models," Journal of Marketing Research, 12, 151-164.

Carrol, J. D., Green, P. E., and DeSarbo, W. S. (1979) "Hierarchical Model Testing in Conjoint Analyses." Advances in Consumer Research, 6 688-691.

Coombs, C. H. (1964), A Theory of Data, New York: John Wiley & Sons, Inc.

Dawes, R. ;4. (1963), "Social Selection Based on Multidimensional Criteria," Journal of Abnormal and Social Psychology, 68, 104-109.

Frey, C. l. and Kinnear, T. C. (1979), "Information Integration Theory; An alternative Attitude Model for Consumer Behavior," Advances in Consumer Research, 6, 350-353.

Hagerty, M. R. (1983), "Individual Choice Models in Marketing," Research in Marketing, 6, 99-130. Greenwich, Connecticut: JAI Press, Inc.

Haley, R. I. (1968), "Benefit Se,mentation: A Decision Oriented Research," Journal of Marketing, 32, 30-35.

Howard, J. A. (1977), Consumer Behavior: Application of Theory, New York: McGraw-Hill.

Johnson, R. M. (1971) "Market Segmentation: A Strategic Management Tool." Journal of Marketing Research, 8, 13-18

Malhotra, N. &., Pinson, C., and Jain, A. K. (1977), "Accommodative Cognitive Style Differences in Consumer Reduction of Alternatives," Advances in Consumer Research, 4, 541-546.

Moriarty, M. and Venkatesan, M. (1978), "Concept Evaluation and Market Segmentation ," Journal of Marketing, 42, 82-80.

Nakanishi, M. and Bettman, J. R. (1971), "Attitude Models Revisited: An Individual Level Analysis," Journal of Consumer Research, 1, 16-21.

Park, C. W. (197-8), "A Conflict Resolution Choice Model," Journal of Consumer Research, 5, 124-137.

Park, C. W. and Schaninger, C. M. (1981), "The Identification of Consumer Judgmental Combination Rules: Statistical Prediction vs. Structured Protocol," Advances in Consumer Research, 8, 184-190.

Pras, B. and Summers, J. (1975), "A Comparison of Linear and Nonlinear Evaluation Process Models," Journal of Marketing Research, 12, 276-281.

Pras, B. and Summers, J. O. (1981), "Operationalizing Risk in Multiattribute Decision Models," Advances in Consumer Research, 8, 92-97.

Rappaport, A. and Wallsten, T. S. (1972), "Individual Decision Behavior," Annual Review of Psychology, 93, 131-179.

Taschian, A., Taschian, R. O. and Slama, M. E. (1991), "the Impact of Individual Differences on the Validity of Conjoint Analysis," Advances in Consumer Research, 8, 363-366

Tversky, A. (1972), "Elimination by Aspects: A Theory of Choice," PsYchological Review, 79, 281-299.

Wahlers, R. G. (1979), "Number of Choice Alternatives and Number of Product Characteristics as Determinants of the Consumer's Choice of an Evaluation Process Strategy," Advances in Consumer Research, 6, 544-549.

Wind, Y., Grashof, J. F. and Goldhar, J. D. (1976), "Market based Guidelines for Design of Industrial Products," Journal of Marketing, 42, 27-37.

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