Measuring Consumers' Inferenential Processing in Choice

Sarah Gardial, University of Tennessee
Gabriel Biehal, University of Houston
ABSTRACT - A conceptual framework is used to define a verbal protocol coding scheme that measures two important aspects of consumer inference making: (1) the inferential process used and (2) the nature of the inferential outcome. Use of the coding scheme is illustrated in a study that did not explicitly require subjects to make inferences during a choice task. Differences in inference making for low versus high product knowledge subjects are highlighted.
[ to cite ]:
Sarah Gardial and Gabriel Biehal (1987) ,"Measuring Consumers' Inferenential Processing in Choice", in NA - Advances in Consumer Research Volume 14, eds. Melanie Wallendorf and Paul Anderson, Provo, UT : Association for Consumer Research, Pages: 101-105.

Advances in Consumer Research Volume 14, 1987      Pages 101-105

MEASURING CONSUMERS' INFERENENTIAL PROCESSING IN CHOICE

Sarah Gardial, University of Tennessee

Gabriel Biehal, University of Houston

ABSTRACT -

A conceptual framework is used to define a verbal protocol coding scheme that measures two important aspects of consumer inference making: (1) the inferential process used and (2) the nature of the inferential outcome. Use of the coding scheme is illustrated in a study that did not explicitly require subjects to make inferences during a choice task. Differences in inference making for low versus high product knowledge subjects are highlighted.

INTRODUCTION

How do consumers make choices from brands when relevant information is missing? One possibility is that missing information is ignored and choice is based on information already available, externally or in memory. Alternatively, the consumer may infer unknown properties and integrate the inferred values with known information. Since brand-related information sources, e.g., advertising or salespeople, rarely offer "complete" information, inferential processes in consumer choice may be widespread. Despite this possibility, our current understanding of inferential processes in consumer choice remains very limited.

Existing research procedures may have contributed to this limited understanding, i.e., researchers have often measured inferences by showing that they affect overall evaluations (e.g., Johnson and Levin 1985, Huber and McCann 1982). However, this gives little insight into the amount of inferencing done, the types of inferential outcomes generated and the underlying processes employed. Also, studies have often either required subjects to make inferences or so structured the situation that subjects may have been "tipped off." Prompting inferences shows that consumers can infer if asked but does not indicate their natural propensity to do so.

In contrast, this paper proposes to directly measure and describe consumers' inferencing using concurrent verbal protocols. Protocols, if properly used, can provide valid indications of information processing (Ericsson and Simon 1980), but the researcher needs a well-defined coding scheme to convert them into useful data. While some schemes have been used to investigate complex phenomena like consumers' use of memory and external information during choice (Bettman and Park 1980, Biehal and Chakravarti 1983), at present no scheme measures consumers' inferential processing. This paper offers such a scheme. It also illustrates its use in a study that examined inferencing differences for low versus high knowledge subjects using an experimental task that did not explicitly request inferences.

CONCEPTUAL BACKGROUND

A Consumer Decision Making Scenario

In a typical choice the consumer encounters information about a set of brands. The consumer's task is to relate this possibly incomplete or incomparable information to a specific usage context and make a choice. Information, which may be presented in various formats, may be found in ads, in-store, etc. Should the consumer feel the need to gather more information before deciding, s/he may consult external sources, e.g., ads, store displays, friends, etc., or s/he may retrieve prior knowledge from memory. If these sources yield sufficient information, a choice may occur. However, if there remain perceived information gaps on relevant evaluative criteria, the consumer may try to infer the brands' performance on the 'unknown dimensions.

The need to infer may be prompted by other circumstances as well. For example, the consumer may seek certain product-related benefits, e.g., convenience. If external information does not refer explicitly to these benefits, the consumer may have to infer from other feature information how much of the benefit a brand contains. Thus, inferential processes transform available information into abstract forms more congruent with the consumer's evaluation criteria. Also, at times product information may not be sufficiently explicit. For example, certain ads substitute evaluatively-laden phrases, e.g., "quality performance," for ore specific, functional information (Shimp and Preston 1981). The consumer may try to infer the concrete product characteristics indicated by such phrases.

Finally, consumers' processing may be affected by many factors, both external and internal (Bettman 1979). One factor of particular relevance here is the consumer's product category knowledge. Several studies have shown the importance of prior product knowledge in consumer decision making (e.g., Brucks 1985), and noted the possibility that there should be some fundamental differences in inferencing for low versus high knowledge consumers (Dover 1981, Taylor and Crocker 1981).

Inferences Defined

Inferences are beliefs that go beyond directly observable events (Fishbein and Ajzen 1975). However, this statement hides some complexities that need clarification if a protocol coding scheme is to be useful. For the present purpose inferences are defined as beliefs generated by the individual (as opposed to being obtained directly from external sources) and attached to or associated with a particular stimulus, e.g., a brand, a person or a situation

This definition highlights two issues. First, inferences are generated by the consumer, in contrast to other types of information which may be simply retrieved from memory (Camp, Lachman and Lachman 1980). For example, a consumer viewing a Toyota ad may retrieve the information that it is a Japanese car. This known information (which may have been previously inferred) is stored in the consumer's product category schema and retrieved, not inferred. However, in the absence of prior knowledge, the consumer may infer that the Toyota gets good gas mileage from information about its size.

Second, because of attentional shifts, interrupts, or inability to resolve the issue, a consumer may begin inferential processing without actually creating a new, inferred belief. Thus the definition states that an inferential outcome must be attached to a particular stimulus object. For example, the consumer may retrieve from memory "Host Toyotas get good mileage." However, until s/he has taken the next step, i.e., "Then this particular Toyota must also get good mileage," no inference has been made.

For coding purposes inferences may differ on two important dimensions: (1) the underlying inferential process used and (2) the nature of the inferential outcome.

Inferential Processes

The literature suggests at least three inferential processes. Evaluative Consistency processes yield inferred values consistent with prior overall evaluations, i.e., liked brands have positive attributes and disliked brands have negative attributes (Fishbein and Ajzen 1975). Similarly, good performance on one dimension may imply g,.od performance on another. For example, social impression formation research shows that subjects may rely on prior overall evaluations stored in memory to guide subSequent evaluations (Lingle et al. 1979).

Probabilistic Consistency inferences are based on perceived associations between stimulus properties. These associations may be: (1) based on prior experience and stored in memory, e.g., price-quality perceptions (Huber and McCann 1982, Johnson and Levin 1985), (2) derived from external information, e.g., perceived inter-attribute correlations found in a set of brands, or (3) based on logically derived relationships, e.g., "heavier cameras are likely to be more durable and hence more reliable." Thus, these processes represent subjective conditional probabilities that certain properties go with other properties. However, these distinctions, while conceptuallY important, are hard to make in practice and the proposed coding scheme does not include them.

Finally, consumers may infer using Distributional Knowledge, i.e., by determining typical or average values of a product feature which are then used to fill in informational gaps (Yamagishi and Hill 1983, Meyer 1981). Distributional knowledge may be derived from external sources or consumers may have stored in memory "default" values for particular product categories. These values may then be retrieved and used to fill in missing information. For example, if information is not available about a camera's mode of operation (automatic or manual), the consumer may retrieve from memory a prototypical camera's properties and use the retrieved information to infer that a particular exemplar is, say, automatic. Thus, this type of inference may require two stages. First, the consumer categorizes the target brand, e.g., "that is a German camera." Then the consumer retrieves information from the category schema and attaches the retrieved value(s) to the target brand (Rosch and Lloyd 1978), e.g., "then it oust have superior optics." However, in practice the full two-staged process is not always captured by the protocol.

The Nature of the Inferential Outcome

Inferences may include many different types of outcomes (Russo and Johnson 1979) which need to be included in the protocol coding scheme. Thus, at "low" levels, the consumer may infer a concrete, objectively verifiable product feature, e.g., "that camera has an automatic timer," or how the camera can be used, e.g., "on that cal era you have to set the aperture manually." "Higher" level inferences may reflect the integration of product relevant information and certain subjective judgments, e.g., "that camera sounds quite sophisticated," or reactions to a brand's performance on a dimension, e.g., "that sounds like a good lens." Finally, at the "highest" level are overall evaluations, e.g., "That would be a good camera to buy." In practice, however, it is hard to unequivocally determine the level of a given inference because the underlying attentional and information integration processes are unclear. Thus, while the proposed coding scheme captures different inferential outcomes, it does not attempt to scale them by inferential level.

AN INFERENTIAL PROCESSING CODING SCHEME

This section defines several classification codes and gives examples of each. Since the study presented later used 35-m cameras as stimuli, examples are taken from that product category. Code categories were designed to capture: (1) the type of inferential outcome generated about target brands and (2) the inferential Process that appeared to underlie the inferential outcome. Thus each product inference is categorized by I-codes (inferential outcomes) and P-codes (inferential processes).

Inferential Outcome (I) Codes

I1. Concrete Feature Inferences. This code identifies inferences about specific, objectively verifiable product features or characteristics. For example, "This camera probably has a timer."

I2. Abstract Feature Inferences. This code identifies product feature inferences that cannot be objectively verified because they refer to subjective, perceptual dimensions. For example: "This camera looks pretty convenient.

I3. Product-in-use Inferences. This code is for inferences that elaborate how the product may be used, e.g., "With this camera you'll probably have to make a lot of tricky adjustments.

I4. Product APProPriateness for a Usage Situation. This code is used for statements that assess the appropriateness of a target brand for a given situation, e.g., "This camera is more for vacation or family picture taking."

I5. Product Appropriateness for User Type. This code is for inferences that reflect an assessment of the product's appropriateness for a type of user, e.g., "This is a beginner's camera."

I6. Feature Evaluation. This code is used for statements that evaluate a product's performance on an evaluative dimension, e.g., "That sounds like a good lens on that camera.

I7. Brand Category Member. This code is used for statements that show classification of the brand relative to another, known brand stored in memory, e.g., "I believe this is like a Minolta."

I8. Product Class Category Member. This code is for statements that classify a specific brand as belonging to a particular product class, e.g., "This looks like one of those 'sure-shot' cameras."

I9. Overall Non-Affective Evaluation. This code is used for statements about the brand as a whole that do not contain an evaluative component. Such statements may reflect a counting of desirable brand features or integration of information on several evaluative criteria, e.g., "All these cameras are about the same," or "They all offer something a little different."

I10. Overall Affective Evaluation. At the "highest" inferential level, this code is used for statements that reflect an overall evaluation of the brand, e.g., "This is the best camera."

Inferential Process (P) Codes

P1. Distribution Knowledge Processes. Typical characteristics of the product category are generated and attached to a specific target brand, e.g., "Host 35"m cameras have a non al lens, so I will assume that this one does, also." If this process is based on prior knowledge, it is assumed to require two stages, i.e., categorization and retrieval. Sometimes the protocol contains the classification process, e.g., "That camera is like a Minolta, so I bet it has..." At other times the categorization is not explicitly stated. Two qualifier codes (see below) are used to distinguish distributional knowledge inferences which are explicitly derived from memory categorizations.

P2. Evaluative Consistency Codes. Two codes are used to reflect different types of evaluative consistency processing.

a. Evaluative Consistency at the Attribute Level. This code is used when affect from one or ore known dimensions is related to other, unknown dimensions, e.g., "It says here that the camera has superior optics, so I would think that it probably has a good metering system, too."

b. Overall Evaluative Consistency. In this process a previously determined overall evaluation about a target brand is used to make an inference about its unknown features in a manner evaluatively consistent with the prior evaluation, e.g., "This is a very nice camera, so it probably has a good lens."

P3. Probabilistic Consistency. The inferential process appears to be based on a perceived associative rule of the general form: "Other cameras with an (x) also tend to have a (y), so I assume that, since this camera has (x), it has (y) too."

Qualifier Codes

Some protocol coding schemes have a descriptive focus (Bettman and Park 1980), i.e., codes describe a phrase's processing content without further interpretation. Even with these schemes the coder needs to carefully consider the phrase's context. Proper use of the proposed coding scheme, which is concerned with both underlying processes and outcomes, requires even greater care. Nevertheless, the essential incompleteness of protocol data causes interpretation problems. If these problems are explicitly recognized and incorporated in the coding scheme, the researcher can develop a conservative view of subjects' inferential processing by screening out ambiguous data. To this end, eight qualifier codes were defined. Thus, each coded inference phrase contained an outcome (I) code, a process (P) code and one or more qualifier (Q) codes.

Q1. Inference Implied by Context - No Explicit Attachment. Sometimes an inference is not explicitly stated but only i-plied by the context. Because attachment to a brand is critical, this code is used to reflect the possibility, but not the certainty, that an inference was made. For example, the subject may state: "If John wants a camera that will grow with him, then I would choose camera A." From the context surrounding the first part of the statement, the inference implied, but not directly stated; is that camera A will grow with the user's needs. However, to be cautious in interpreting the data, this conclusion should be qualified.

Q2. Inference Attempted but Unresolved. Sometimes inferential processing occurs but the subject does not or cannot reach a clear conclusion. For example: "Well, the camera might have lenses other than the one in the picture. Host do. I'm just not sure." These statements have theoretical relevance and should be recorded because they may reflect differences in subjects' product category knowledge. This qualifier identifies unresolved inferential processing and thereby permits a search for its correlates.

Q3. Possible Reinterpretation of Stimulus Information. It may be difficult to clearly determine if a statement is an inference or a recoding of externally available information. For example, upon reading that a camera has an easy loading system, the subject may state: "This is a self-loading camera." In fact, the easy loading benefit may be delivered by several functional properties e.g., a film take-up device or an electric film advance, whose presence the consumer may infer. Alternatively, s/he may simply be recoding the externally available information into a more appropriate form. While it may be argued that any recoding of available information is an inference, this qualifier is used to mark the possible indeterminacy.

Q4. Repeated Inferences. This code avoids double counting previously made inferences. When repeated inferences are encountered, it is not possible to clearly determine if the inferential process was actually repeated or, perhaps more likely, that subsequent statements reflect retrieval from memory of the prior inference.

Q5. Multiple Inferences. Sometimes a statement contains several implied inferences, yet it is not possible to determine the actual number, e.g., "All of these cameras are light-weight." The researcher may take the statement at face value and conclude that, if four brands are available, four inferences were made. However, this may be inappropriate because the statement may refer to only a subset of the brands. The recommended approach is to count the phrase as indicating one inference and to use the qualifier to identify the possibility of multiple inferences.

Q6. Information Source versus Product. This qualifier is used when the coder cannot tell if the statement is a product inference or a comment about some aspect of the stimulus information. The problem is most likely to occur when pictures are shown of products, e.g., in ads. Thus the statement: "This camera is a s all one" may be an inference about the camera's actual size or it may be a comment about the comparative size of the illustrations contained in various advertisements. While the researcher could interrupt the subject and ask for clarification, this runs counter to the method's spirit, i.e., a "freeflow" of thought. It may also disrupt the subjects's train of thought. This could cause attentional shifts and perhaps prompt concern with explaining what s/he is doing, thereby encouraging a rationality bias.

Q7. Brand Categorization. This code is used when a distributional knowledge process (P1) is preceded by explicit categorization of the stimulus into a known brand category. For example, "This is a lot like a Minolta, therefore it probably has additional lenses."

Q8. Product Class Categorization. This qualifier code is used when a distributional knowledge process (P1) is preceded by explicit categorization of the stimulus brand into a product class group. For example, "This is one of those old, box-type cameras, so I bet it's pretty heavy."

These qualifiers capture most eventualities. A more general problem is differentiating inferences from retrieval of prior knowledge. For example, if a subject states: 'Most 35mm cameras are expensive," does this imply the inference that the target 35mm camera is also expensive? Although such an inference say have occurred, it is also possible that this product category information was simply retrieved and reviewed. This indeterminacy is conservatively resolved by labeling such statements as inferences only when followed by explicit protocol evidence that a new belief was generated with respect to a target stimulus. Otherwise, they are considered prior knowledge. Bettman and Park (1980) provide a useful set of knowledge retrieval codes.

A STUDY OF INFERENTIAL PROCESSES

Overview

A study examined consumers' inferential processing during choice. Subjects selected the best 35mm camera from three hypothetical brands. These were described on separate sheets of paper using an ad-like format, i.e., there was a headline and a photograph of a disguised, actual brand followed by a few paragraphs of copy. The ads, which gave information on three attributes, did not give complete information for all brands.

Subjects had to choose bearing in mind a specific context, namely, selecting the brand most appropriate for a third party. While choosing, subjects talked out loud and described what they were thinking. At no time were they asked to make inferences. Finally, subjects of varying experience with the product category were recruited. Prior knowledge was assessed in several debriefing questions and two levels (low/high) were defined. Data was gathered in individual sessions that lasted approximately 33 minutes.

Subjects

100 student volunteer subjects were recruited from the campus community. Their average age was 25 years, and 60 percent were male.

Protocol Coding

Tape recorded protocols were transcribed and coded independently by the authors. To reconcile the codings, the coders first agreed on the phrases that contained inferential content. Then they discussed the codes they had assigned to each phrase. All discrepancies were identified, logged and resolved.

The coders first agreed on 565 phrases which contained inferences. From these phrases, a total of 1882 I, P, and Q codes were assigned, yielding 220 (12%) disagreements of which 42.3% were P codes and 40.5% were I codes. In addition, several phrases were coded by only one researcher. After discussion, 44 of these phrases and the I, P and Q codes assigned by the original coder were accepted in full by the second coder. Thus, a total of 609 inferential phrases were coded with an agreement rate of 85.4%. Of the 609 phrases, 139 were removed because they had qualifiers (Q1, Q2, Q4 and Q6) that indicated some degree of uncertainty about the-. The 470 remaining inferences form the basis of the data presented here.

Data Analysis

To identify low (LE) and high experience (HE) groups, subjects' prior experience with 35mm cameras was assessed by a multiple choice quiz (Sujan 1985). The original test, which had 15 items, was reduced to eight items based on (1) pretests that indicated some items confused the subjects and (2) an item reliability analysis. The reduced instrument had a KR-20 test value of 0.8654, indicating a high degree of reliability (Peter 1979). Subjects' experience scores were computed by summing the number of correct responses on the eight item scale. Low (n=22) and high experience subjects (n=26) were then defined at the extreme 22-265 ends of the distribution (zero and six or more correct answers, respectively). To check this classification, two-way chi-squared tests were made with measures of prior 35mm camera use, prior shopping for 35mm cameras, photographic magazine subscription and taking photography courses. All were significant (p<0.001). Also, while experience differences say be confounded with other factors, e.g., intelligence, and need careful interpretation, analyses of the subjects' task perception measures showed no confounding with motivation to process.

Results

Table 1 reports both frequencies and proportions, or the relative use, of the various inference outcomes for LE and HE subjects. On average, HE subjects generated significantly ore inferences (5.85 phrases/protocol) than LE subjects (3.32 phrases, t=-2.71, p<0.01). These data suggest that inferencing say be an important constituent of consumer information processing regardless of experience level. There were also group differences in inferential outcomes. HE subjects made proportionately ore inferences about concrete features (I1, z=2.56, p<0.01), about how the product may be used (I3, z=1.41, p<0.05) and about brand category membership (I7, z=1.76, p<0.05).

These inferences reflect HE subjects' greater knowledge of the product class that allows the to "fill in" quite specific brand characteristics during evaluation. In contrast, LE subjects placed proportionately ore emphasis on evaluating the product's situation appropriateness (I4, z=1.96, p<O.05) and user appropriateness (I5, z=1.35, p<O.10). Thus, compared to HE subjects, LE subjects see to make more integrative and evaluative inferences, rather than specific feature level inferences.

Table 2 shows the two groups' relative use of each inferential process. A comparison of the proportions of processes used shows that both LE and RE subjects relied most on probabilistic consistency processes (79.5% and 57.9%, respectively), though the LE proportion was higher (z=3.49, p<0.001). HE subjects used significantly more distributional knowledge processes (24.3% versus 4.1%, z=4.83, p<0.0001). Interestingly, the LE group made little use of evaluative consistency processing. Thus they may have had sufficient general consumer knowledge to use probabilistic consistency processes in the experimental task. Fishbein and Ajzen (1975) propose that evaluative consistency processing will be used only if probabilistic processing is ineffectual. These data sees consistent with their conclusion. They are also consistent with theoretically based expectations that HE subjects should make greater use of distributional knowledge processes.

TABLE 1

TYPES OF INFERENTIAL OUTCOMES FOR LOW (LE) VERSUS HIGH EXPERIENCED (HE) SUBJECTS

TABLE 2

TYPES OF INFERENTIAL PROCESSES FOR LOW (LE) VERSUS HIGH EXPERIENCED (HE) SUBJECTS

In summary, these results indicate that (a) even in experimental situations where they are not asked to infer, subjects a}e inferences during decision making and (b) inferential outcomes and processes may differ as a function of experience level.

Conclusions

The proposed coding scheme helps develop our understanding of consumers' inferential processing. It lets the researcher capture the amount of inferencing and classify it with respect to both type of outcome and type of processing. Thus it moves beyond the aggregate, indirect measures used in previous research and facilitates examination of the underlying dynamics of consumer inferential processing

Results obtained from using the coding scheme should raise many further questions. Thus, the scheme serves as a guideline for future research. For example, it appears that probabilistic inferential processes are often used. However, several associational processes are included in this category, e.g., ecological correlations, logical propositions and true probabilistic estimates. One area for future research would be unraveling the factors that affect their relative use in inference making.

REFERENCES

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

Bettman, James R. and C. Whan Park (1980), "Implications of a Constructive View of Choice for Analysis of Protocol Data: A Coding Scheme for Elements of Choice Processes," in Advances in Consumer Research, Vol. 7, ed. Jerry Olson, Ann Arbor. MI: Association for Consumer Research, 148-153.

Biehal, Gabriel J. and Dipankar Chakravarti (1983), "Information Accessibility as a Moderator of Consumer Choice." Journal of Consumer Research, Vol. 10(1), 1-14.

Brucks, Merrie (1985), "The Effects of Product Class Knowledge on Information Search Behavior," Journal of Consumer Research, Volume 12 (June), 1-16.

Camp, Cameron J., Lachman, Janet L. and Lachman, Roy (1980), "Evidence for Direct-Access and Inferential Retrieval in Question-Answering," Journal of Verbal Learning and Verbal Behavior, 19 (October), 583-596.

Dover, Phillip A. (1981), "Inferential Belief Formation: An Overlooked Concept in Information Processing Research," Advances in Consumer Research, Vol. XI, Ann Arbor, MI: Association for Consumer Research, 187-189.

Ericsson, K. Anders and Herbert A. Simon (1980), "Verbal Reports as Data," Psychological Review, Vol. 87(3), 215-251.

Fishbein, Martin and Icek Ajzen (1975), Belief, Attitude, Intention and Behavior: An Introduction to Theory and Research, Reading, MA: Addison-Wesley.

Huber, Joel and John McCann (1182), "The Impact of Inferential Beliefs on Product Evaluations," Journal of Marketing Research, Volume 14, 324-333.

Johnson, Richard D. and Irwin P. Levin (1985), "More Than Meets the Eye: The Effect of Missing Information on Purchase Evaluations," Journal of Consumer Research, Volume 12, 169-177.

Lingle, John H., Thomas Ostrum, Nehemia Geva and Michael Leippe (1979), "Thematic Effects of Person Judgments on Impression Formation," Journal of Personality and Social Psychology, Volume 37(2), 674-677.

Meyer, Robert J. (1981), "A Model of Multiattribute Judgments Under Attribute Uncertainty and Informational Constraint," Journal of Marketing Research, Volume 18, 428-441.

Peter, J. Paul (1979), "Reliability: A Review of Psychometric Basics and Recent Marketing Practices," Journal of Marketing Research, Volume 16, 6-17.

Rosch, Eleanor and Barbara B. Lloyd (1978), Cognition and Categorization, Lawrence Erlbaum Associates, Hillsdale, NJ.

Russo, J. Edward and Eric Johnson (1979), "What do Consumers Know about Familiar Products?" in Advances in Consumer Research, Vol. 7, ed. Jerry Olson, Ann Arbor, MI: Association for Consumer Research, 417-423.

Shimp, Terence A. and Ivan L. Preston (1981), "Deceptive and Non-Deceptive Consequences of Evaluative Advertising," Journal of Marketing, Volume 45(1), 22-32.

Sujan, Mita (1985), "Consumer Knowledge: Effects on Evaluation Strategies Mediating Consumer Judgements," Journal of Consumer Research, Vol. 12, no. 1, 31-46.

Taylor, S. E. and Jennifer Crocker (1981), "Schematic Bases of Social Information Processing," in Social Cognition: The Ontario Symposium Vol. 1, E. T. Higgins, P. Herman, and M. Zanna, eds., Hillsdale, NJ: Lawrence Erlbaum Associates.

Yamagishi, Toshio and Charles T. Hill (1983), "Initial Impression Versus Missing Information as Explanations of the Set-size Effect," Journal of Personality and Social Psychology, Volume 44(5), 942-951.

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