Implications of a Constructive View of Choice For Analysis of Protocol Data: a Coding Scheme For Elements of Choice Processes

ABSTRACT - A Constructive view of choice is outlined, and its implications concerning a more detailed contingency notion for choice heuristics are considered. A scheme for coding protocol data is presented which follows from this view. An initial application of the scheme to protocol data is considered. Potential uses of the coding system for choice process research are suggested.

Citation:

James R. Bettman and C.W. Park (1980) ,"Implications of a Constructive View of Choice For Analysis of Protocol Data: a Coding Scheme For Elements of Choice Processes", in NA - Advances in Consumer Research Volume 07, eds. Jerry C. Olson, Ann Abor, MI : Association for Consumer Research, Pages: 148-153.

Advances in Consumer Research Volume 7, 1980     Pages 148-153

IMPLICATIONS OF A CONSTRUCTIVE VIEW OF CHOICE FOR ANALYSIS OF PROTOCOL DATA: A CODING SCHEME FOR ELEMENTS OF CHOICE PROCESSES

James R. Bettman, University of California, Los Angeles

C.W. Park, University of Pittsburgh

ABSTRACT -

A Constructive view of choice is outlined, and its implications concerning a more detailed contingency notion for choice heuristics are considered. A scheme for coding protocol data is presented which follows from this view. An initial application of the scheme to protocol data is considered. Potential uses of the coding system for choice process research are suggested.

INTRODUCTION

Consumer researchers have displayed a great deal of interest in characterizing the choice heuristics used by consumers. In attempting to study these heuristics, many different methods have been used to gather data on details of processing. One method used by many researchers is the collection and analysis of verbal protocol data (e.g., Alexis, Haines, and Simon 1968; Bettman 1970; Payne 1976; Russo and Rosen 1975; Lussier and Olshavsky 1979). Despite this activity, the most useful ways for analyzing these data have not been obvious. Prior approaches have ranged from building complex discrimination net models of choice (Alexis, Haines, and Simon 1968; Bettman 1970) to supporting and supplementing insights into heuristics drawn from studies of information acquisition which used other process-tracing methodologies (e.g., Russo and Rosen 1975; Payne 1976).

The purpose of the present paper is to propose another approach for analyzing protocol data which relies on a detailed coding of the content of the phrases in a protocol. This approach flows from a consideration of the nature of choice processes. Many choice researchers tend to focus on well-structured "rules" (e.g., lexicographic, conjunctive, linear compensatory). However, protocol data seem to show that such rules are not systematically used. Rather, there is evidence that consumers do not always have fully developed choice rules, but sometimes make up or construct a heuristic when they need to make a choice (Bettman and Zins 1977). Heuristics constructed in this fashion tend to be more adaptive to the situation and less well-structured than the standard rules mentioned above. Thus, attempts to study the components of such heuristics, rather than some overall structure, may yield useful insights. In the following we consider how choice processes are constructed, and the implications of this view for analyzing protocol data are noted. Then a scheme for coding protocol data which follows from this view is presented, and an initial application of this scheme is briefly outlined. Finally, potential uses of the coding system in future research on choice are considered.

A CONSTRUCTIVE VIEW OF CHOICE AND ITS IMPLICATIONS

Only an outline of a constructive view of choice is presented here, as more detailed treatments can be found elsewhere (Bettman 1979; Bettman and Zins 1977). The basic concept is that consumers may not have complete rules or heuristics stored in memory which they use to make a choice. Rather, consumers may have only fragments or elements of heuristics in memory, which are put together during the actual choice process to "construct" a heuristic. Thus, in some cases completed heuristics or rules do not already exist, but must be built dynamically at the time of choice from elements or subparts. Such elements may be beliefs about alternatives, evaluations, simple rules of thumb (e.g., "compare these brands on attribute A to see if they differ very much", "compare this attribute level to a standard"), rules for integrating data (e.g., "count how many attributes alternative X is best on) and so on. Note that some of these "elements" are also simple heuristics. Heuristics can be defined at several levels of detail, so the issue of concern is whether "detailed" elements or more aggregate rules are the focus.

To the extent that such elements are put together at the time of choice, the elements used for a particular choice and the sequence in which they are used will be a function of such factors as: what information is available (e.g., whether the same data are available for all brands); the format in which the information is presented (e.g., prices may not be compared if unit prices are not provided and different brands have different-sized packages); the salience of various pieces of information; intermediate processing results; and other task specific factors. Thus, the choice heuristics which result from the construction process may vary from one situation to the next, depending on how the underlying elements were put together.

Constructive processes will not always be used, of course. In cases where there is a good deal of prior experience with a choice, consumers may very well have completed heuristics available in memory. However, when choices are to be made where there is little prior knowledge, constructive processes are likely.

This notion of constructive process has important implications for the analysis of protocol data on choice processes. If consumers build up heuristics as they go along, and the elements used are sensitive to many task specific factors (salience, format, and so on), the resulting choice "heuristic" may consist of a sequence of elements with no necessarily coherent overall structure. The following sequence illustrates such a structure: (1) A consumer uses the element "compare attribute levels to a standard" for several brands (a component of a conjunctive rule) (2) The consumer then notes a good value for a particular brand on some attribute and eliminates all brands still being considered which are not "close" to that value (like an element of an elimination by aspects heuristic--see Tversky 1972), (3) Next the consumer compares two brands to see which is better on more attributes (a procedure which is part of a heuristic additive difference rule (Russo and Dosher 1975)), and so on. Thus, each element or short sequence of elements may be used to process only a few brands. Different sequences of elements are used for different brands. Lest this appear far-fetched, examples of this fragmented overall structure appeared in many of the individual protocols in the application described further below. One individual went through the following sequence: (1) Several brands which were not satisfactory on a particular attribute were eliminated, (2) Pairwise differences were compared between attribute values for two brands, (3) The attributes of each of several brands were compared to standards for those attributes, (4) All brands that were not satisfactory on a particular attribute were eliminated, and (5) Pairwise differences for two brands were finally compared. Such sequences might be quite common if the view of constructive processes presented above is valid. Thus, looking for an overall strategy applied to all brands may be fruitless (this also implies that applying algebraic rules to brands in an attempt to predict consumers' choices is unlikely to be productive).

The arguments above lead to an extension of the notion that the choice heuristic used is contingent upon properties of the choice task. Prior research guided by this contingency perspective has examined relatively stable properties of the task: time pressure (Wright 1974), format (Bettman and Zins 1979), incomplete information (Slovic and MacPhillamy 1974), type of response required (Slovic 1972; Grether and Plott forthcoming), and so on. However, the constructive view implies a more complex notion of contingency, where many properties of the "choice task" itself change as the consumer progresses. Thus the elements of choice heuristics used may not even be the same for all brands during a given choice. The task is not the same for all brands. The elements used to process a given brand may depend upon which brands have already been processed (e.g. whether a "good" alternative has appeared yet or not); upon the particular sequence of elements already used (e.g. if certain brands have been eliminated because of their values on a given attribute, that attribute will have a relatively restricted range when further operations involving that attribute are considered); upon which other brands happen to be near a given brand in a shelf display (e.g., because this will affect the magnitude of the differences on various attributes); and so on. The constructive view implies a more detailed contingency notion: the elements of choice heuristics used at any given time are contingent upon the properties of the choice task at that particular time.

This detailed contingency notion is similar to the production system concept of Newell and Simon (1972). A production system consists of a set of individual productions, each consisting of a condition and an action. When the condition is met, the corresponding action is executed. A production system operates by having control start at the top of the list of productions, with the condition portion of each production tested until a condition is met. Then the corresponding action is taken and control once again reverts to the top of the set of productions (Newell and Simon 1972, pp. 44-46). An example of a production for a consumer choice might be "If the number of brands to be compared is two, and the number of attributes for each is greater than three, then count the number of desirable features for each brand." Newell and Simon use protocol data to develop detailed production system models for individuals performing a particular problem solving task. That is, the protocol data are used to infer a concise set of conditions and contingent actions which will reproduce the sequence of events occurring in the protocol.

Despite the theoretical attractiveness of this approach, its implementation is problematic, given our current level of understanding of choice heuristic elements and the associated contingencies which determine when such elements might be used. If consumers use constructive processes in relatively complex task environments, the detailed analysis of contingencies implied in building production system models or in fully operationalizing the more complex contingency notion outlined above would be too ambitious an undertaking at this time (see Lussier and Olshavsky 1979, for a similar view). Therefore, a much more limited operationalization of the detailed contingency notion was considered: isolating those factors which influence the frequency with which various fragments or elements are used to build up choice heuristics.

In sum, looking for some overall strategy or heuristic which is applied to all brands seems inappropriate, given a constructive view of choice. Rather, the detailed elements of choice heuristics seem more likely to be the appropriate units of analysis. Research focussing on the factors influencing the frequency of use of these seems to be a fruitful first step. To carry out this research, a scheme for coding such elements is needed.

DEVELOPMENT OF A SCHEME FOR CODING ELEMENTS OF CHOICE PROCESSES

Three main sources of input were used to provide potential elements for the coding scheme: prior research on coding protocol data; theoretical analysis of the elements used in various choice heuristics; and a scan of a sample of actual protocol data. Each of these sources is discussed briefly below.

Prior Research

Svenson (1974) obtained protocols from subjects making choices among seven hypothetical houses. He coded these protocols with respect to usage of such elements as comparison of an attribute to a standard and comparison of two alternatives on a single attribute. Although Svenson performed some fascinating analyses on his data, his coding scheme was limited. Montgomery (1977) developed a coding scheme for protocols taken during choices between two gambles. His scheme of nine codes focused on comparison of attribute values and evaluation of differences in attribute values between the two alternatives. Johnson (1978), in an analysis of memory structure after a choice task, considered such basic elements of choice heuristics as overall ratings, statements of brand-attribute values, comparison of two brands on an attribute, and comparison of two attributes for a given brand.

Payne and Ragsdale (1978) developed a coding scheme for consumer protocol data. Their scheme contained twenty process codes, under five general headings: 1) information processing within product categories; 2) statements of goals, needs, or strategies; 3) product choice decisions; 4) statements of simple awareness; and 5) miscellaneous. Finally, Lussier and Olshavsky (1979) propose a scheme which concentrates on input, comparison, trade-off, ranking, range, and attribute-relationship operations. These previous efforts helped to generate a partial list of elements for inclusion in the coding scheme presented below. However, the proposed scheme is at a much greater level of detail than any of these earlier codes.

Analysis of Choice Heuristics

By analyzing choice heuristics proposed by previous researchers, insights can be gained into the elements or components comprising such heuristics. For example, those elements required for the linear compensatory heuristic might include statements of attribute weights, brand beliefs, and compensatory combination across attributes; one element underlying a conjunctive heuristic would be comparison of an attribute to a standard or desired level; elements contained in a heuristic additive difference rule might include computing pairwise differences between brands on individual attributes, counting the number of attributes where one brand is best, making trade-offs among pairwise differences on several attributes, and so on. Thus, by considering an array of such rules (e.g., those mentioned above plus disjunctive, elimination by aspects, and lexicographic), a relatively extensive list of potential elements of choice processes was developed.

Examination of Protocol Data

As a final method for generating elements of choice processes which might be included in the coding scheme, a sample of the protocol data described below was scanned, This process led to the identification of several additional elements.

Based upon the inputs from each of these three sources, a coding scheme containing five broad categories and seventy individual codes was developed. This scheme is outlined below. It should be noted, however, that details of choice heuristics were the focus of this particular set of codes. Researchers with different interests, of course, could develop other coding systems which emphasized different aspects of the protocols (e.g., the various sources of influence upon consumers). Finally, our objective was to develop a very detailed code initially. If required, the detailed codes could be aggregated for later analysis. Initial coding at a more aggregate level was rejected, because we wished to preserve the ability to retrieve detailed information if desired.

THE CODING SCHEME

Basic Structure

The coding scheme is presented in Table 1. The basic structure is provided by five broad categories: attribute comparison processes, within-brand processes, use of prior knowledge, statements of plans or needs, and general statements (this is similar to the division in Payne and Ragsdale 1978). Each of these broad categories is now briefly considered, using examples drawn from the protocol data on microwave oven choices described in more detail below. The coding scheme is applied by first segmenting the protocol into short phrases, and then coding each individual phrase. It is beyond the scope of this paper to consider explanations of the use of each individual code. For a more extensive description and set of examples for each code, see Bettman and Park (1979).

Attribute Comparison Processes

The basic organizational structure of this section is provided by whether one or more than one attribute is involved; and whether the process is based on two brands, more than two brands, or no specific brand. All of the codes within this section specifically refer to cases where the individual is processing within an attribute or set of attributes, across brands. Note that some of these elements might be sub-components of common choice heuristics (e.g., lexicographic (A12, A5, A6); elimination by aspects (A12, A9); or heuristic additive differences (A1, A16, A20).

As an example of the delineations between the individual codes, consider A10, A12, and A14. A10 is used when a phrase refers to the evaluation of a particular attribute level for a continuous attribute (e.g., number of cooking levels) if that evaluation is general, and not for a particular brand. If one level of a discrete attribute is evaluated (e.g., whether or not a browner is present), A14 should be used. Finally, for statements about the weight or importance of a continuous attribute, A12 is appropriate. Examples of the usage of these codes are "three cooking levels would probably be enough" (A10), "price is important to me" (A12), and "I definitely want one with a safety start" (A14).

Within-Brand Processes

This section is organized by the number of attributes considered. All of the codes refer to processes carried out within one brand, over one or more attributes. Some of these elements might also be components of standard heuristics (e.g., linear compensatory (B1, B2, B9, B11) or conjunctive (B4)). The codes in this section are among the most difficult to distinguish, particularly BI, B2, B3, B5, and B6. B1 is used to code single, non-evaluative statements that a brand possesses a particular feature. B2 refers to a single evaluation of an attribute level for a particular brand. If a phrase is a single statement of how a brand is rated on a particular attribute relative to the range for that attribute, use of B3 is appropriate. The distinguishing aspect of B5 and B6, relative to B1, B2, or B3, is that B5 and B6 are used when a phrase is part of a string of features being enumerated by the individual. B5 is used for phrases noting desirable features, and B6 for phrases noting undesirable features as part of a string. If there is a string of features enumerated, B5 or B6 should take precedence over B1, B2, or B3, even if these latter codes might seem appropriate for individual elements in the string of features. Examples of these codes are "there are four cooking levels here" (B1); "The price was good" (B2), "I see that the Sears has a pretty good size" (B3), "The Sharpe has a very low leakage/ it has a safety start/ it has a browner" (B5-the / denotes boundaries between phrases), and "but it doesn't have a temperature setting/it doesn't have a browner/it doesn't have a scale timer/it doesn't have a safety start" (B6).

Use of Prior Knowledge

Note that all of the codes in the two previous categories refer to aspects of or the results of current processing. However, some phrases clearly refer to results of prior processing or some other type of prior knowledge. Statements which emphasize prior knowledge are the focus of this section. The organization of the section is provided by aspects of the multi-attribute attitude model: overall evaluations, brand-attribute values, evaluations of attribute levels, and attribute weights are considered. For each, five sources of prior knowledge are used: prior choice processes, ownership, or usage; word of mouth; generalization from other products; advertising; and other. Finally, statements about level of knowledge and current brand ownership are included. As examples of some of these codes, consider "OK, I like the Wards that I have. I haven't really had any problems with it" (E1), "I have other Sears things that are good" (E3), "I have the fan and it's fine" (E11), "I don't know anything about Sharpe" (E22), "I do own a Litton" (E24).

Statement of Plans or Needs

This section consists of codes for various plan, strategy, or need statements. Examples include "so I would really want to check into it before I would consider it" (P3), "I'd really need to look at the machine to really know" (P4), "but if it came down to choosing between two, I would look for the most safety involved" (P5).

General

These codes refer to input, choice, and task statements. Examples are "I guess I'll pick the Amana" (G2), "Now, the safety start" (G4), "Wait, this shows Litton twice. Oh, those are two different kinds" (G8).

The above provides an introduction to the coding scheme. The scheme was also applied to a set of actual protocol data to investigate its potential usefulness. This initial application is briefly discussed below.

AN INITIAL APPLICATION OF THE CODING SCHEME

Data and Subjects

The protocol data were collected during a choice among brands of microwave ovens. Fifteen different microwave ovens were described on nine attributes in a table of information provided to each subject. Subjects were first asked to find all brands which would be acceptable to then (the first phase of a phased choice process (Wright and Barbour 1977)). Then they were requested to choose the one brand most preferred (the second phase). Subjects were encouraged to think aloud during these two phases, and their comments were tape recorded. The data from both phases was combined for the analysis reported below.

TABLE 1

A PROTOCOL CODING SCHEME FOR ELEMENTS OF CHOICE PROCESSES

Subjects were 68 individuals from a small midwestern town. They were recruited to serve in the study and were selected to have different levels of familiarity with microwave ovens (this aspect of the data is beyond the scope of the current paper).

Procedure

The data were transcribed from the tape-recorded protocols and broken into short phrases. Each phrase consisted of a single task-related statement (Newell and Simon 1972; Payne and Ragsdale 1978). The length of the phrases depended to some extent on the content of the phrase and the focus of the research. For example, the coding scheme presented above focuses on processing details. Thus, phrases related to such details tended to be fairly short. However, phrases recounting prior knowledge in some detail could be rather lengthy.

After pilot testing and revising the codes on six of the protocols, the authors served as independent judges in coding the remaining 62. In general, use of such schemes requires multiple judges, as the decisions are often quite difficult. In the present case, each judge sometimes selected more than one code as a potential descriptor of a phrase. If the two judges agreed on one of these codes, that phrase was counted as one on which they agreed.

Results

For the 68 protocols, 2328 phrases were coded. For the 62 protocols with two coders, 1977 phrases were coded. The two coders agreed on 1547 phrases, or 78.3% of the total. The percentage of agreement ranged from 48% to 100% across the 62 individuals examined.

The number and percentage of phrases falling into each of the categories is presented in Table 1. Attribute comparisons were the most common type of element, with statements of plans or needs the least frequent. It is obvious from inspection of the individual codes that many appear rather infrequently. However, related codes can be aggregated for further analysis if desired. Finally, since all of the examples given above represent single phrases, and context is obviously important in making coding decisions, an excerpt from a complete protocol is presented in Table 2.

TABLE 2

EXCERPTS FROM A CODED PROTOCOL

CONCLUSIONS

The results of the initial application of the scheme are encouraging, as inter-coder reliability seems good and the scheme seems to be relatively comprehensive. The elements in the coding scheme can be used to begin to examine the complex contingency notions for elements of choice heuristics described above. That is, one can formulate hypotheses about which factors will influence the frequency with which various elements (or aggregates of elements) are used.

For example, based on prior theoretical work on extensive and limited problem solving (Howard 1977), one might hypothesize that attribute evaluation and weight codes (Al0, Al2, Al4) would be used more frequently in early stages of the choice process by individuals with low levels of familiarity or experience with the product class. Hypotheses about information format could also be developed: if information on alternatives were presented simultaneously, more comparisons or tradeoffs across brands would be expected (e.g. A1, A2, A16); if it were presented sequentially, more comparisons of attributes to standards (B4) or tradeoffs within brands (B9) might be used. Finally, one can examine propositions about the type of processing characterizing phases of the choice process: comparisons to standards may be more extensive during the first phase, with compensatory elements used more frequently during the second phase.

The above examples suggest that examining the elements of choice processes is potentially a valuable approach. A constructive view of choice implies that such elements, rather than overall choice rules, may be the most appropriate units of analysis for examining consumer choice. The coding scheme presented above provides a tool for examining these elements.

REFERENCES

Alexis, Marcus, Haines, George H., and Simon, Leonard (1968) "Consumer Information Processing: The Case of Women's Clothing," in Proceedings of the Fall Conference, Marketing and The New Science of Planning (Chicago: American Marketing Association, 197-205).

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

Bettman, James R. (1979), An Information Processing Theory of Consumer Choice (Rending, Mass.: Addison-Wesley) .

Bettman, James R. and Park, C.W. (1979), "Description and Examples of a Protocol Coding Scheme for Elements of Choice Processes," Paper No. 76, Center for Marketing Studies, Graduate School of Management, University of California, Los Angeles.

Bettman, James R. and Zins, Michel A. (1977), "Constructive Processes in Consumer Choice," Journal of Consumer Research, 4, 75-85.

Bettman, James R. and Zins, Michel A. (1979), "Information Format and Choice Task Effects in Consumer Decision Making," Journal of Consumer Research, 6, in press.

Grether, David M. and Plott, Charles R. (forthcoming), "Economic Theory of Choice and the Preference Reversal Phenomenon," American Economic Review.

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

Johnson, Eric J. (1978), "What Is Remembered About Consumer Decisions?" unpublished manuscript, Department of Psychology, Carnegie-Mellon University.

Lussier, Denis A. and Olshavsky, Richard W. (1979), "Task Complexity and Contingent Processing in Brand Choice," Journal of Consumer Research, 6, in press.

Montgomery, Henry (1977), "A Study of Intransitive Preferences Using a Think Aloud Procedure," in H. Lungermann and G. De Zeeuw, eds., Decision Making and Change in Human Affairs (Amsterdam: Reidel)

Newell, Allen and Simon, Herbert A. (1972), Human Problem Solving (Englewood Cliffs, NJ: Prentice-Hall).

Payne, John W. (1976), "Task Complexity and Contingent Processing in Decision Making: An Information Search and Protocol Analysis," Organizational Behavior and Human Performance, 16, 366-387.

Payne, John W. and Ragsdale, E.K. Easton (1978), "Verbal Protocols and Direct Observation of Supermarket Shopping Behavior: Some Findings and a Discussion of Methods," in H. Keith Hunt, ed., Advances in Consumer Research, Volume 5 (Chicago: Association for Consumer Research, 571-577).

Russo, J. Edward and Dosher, Barbara A. (1975), "Dimensional Evaluation: A Heuristic for Binary Choice," Unpublished paper, Department of Psychology, University of California, San Diego.

Russo, J. Edward and Rosen, Larry D. (1975), "An Eye Fixation Analysis of Multi-alternative Choice," Memory and Cognition, 3,267-276.

Slovic, Paul (1972), "Information Processing, Situation Specificity, and the Generality of Risk-Taking Behavior," Journal of Personality and Social Psychology, 22, 128-134.

Slovic, Paul and MacPhillamy, Douglas (1974), "Dimensional Commensurability and Cue Utilization in Comparative Judgment," Organizational Behavior and Human Performance, 11, 172-194.

Svenson, Ola (1974), "Coded Think Aloud Protocols Obtained When Making a Choice to Purchase One of Seven Hypothetically Offered Houses: Some Examples," Unpublished paper, University of Stockholm.

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

Wright, Peter L. (1974), "The Harassed Decision Maker: Time Pressures, Distractions, and the Use of Evidence," Journal of Applied Psychology, 59, 555-561.

Wright, Peter L. and Barbour, Frederic (1977), "Phased Decision Strategies: Sequels to an Initial Screening," in Martin K. Starr and Milan Zaleny, eds., North-Holland TIMS Studies in the Management Sciences, Volume 6: Multiple Criteria Decision Making (Amsterdam: North Holland, 91-109).

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