Inferences About Missing Attributes: Contingencies Affecting the Use of Alternative Information Sources

ABSTRACT - We examined the use of same- and other-brand information sources in inferences about missing attributes. Subjects often used a same-brand process-probabilistic inference--when the inferred attribute was expected to be correlated with a known attribute, but they did not use an other-brand process--average-value inference--when the inferred attribute was expected to be uncorrelated with any known attribute. Subjects' inference strategies were insensitive to attribute correlations in task product sets. Feature-absent inferences were made when (a) an undescribed attribute was described for other alternatives evaluated concurrently and (b) the inferred attribute was one that could logically be absent

Citation:

Carolyn J. Simmons and Nancy H. Leonard (1990) ,"Inferences About Missing Attributes: Contingencies Affecting the Use of Alternative Information Sources", in NA - Advances in Consumer Research Volume 17, eds. Marvin E. Goldberg, Gerald Gorn, and Richard W. Pollay, Provo, UT : Association for Consumer Research, Pages: 266-274.

Advances in Consumer Research Volume 17, 1990      Pages 266-274

INFERENCES ABOUT MISSING ATTRIBUTES: CONTINGENCIES AFFECTING THE USE OF ALTERNATIVE INFORMATION SOURCES

Carolyn J. Simmons, University of Illinois at Urbana-Champaign

Nancy H. Leonard, University of Illinois at Urbana-Champaign

ABSTRACT -

We examined the use of same- and other-brand information sources in inferences about missing attributes. Subjects often used a same-brand process-probabilistic inference--when the inferred attribute was expected to be correlated with a known attribute, but they did not use an other-brand process--average-value inference--when the inferred attribute was expected to be uncorrelated with any known attribute. Subjects' inference strategies were insensitive to attribute correlations in task product sets. Feature-absent inferences were made when (a) an undescribed attribute was described for other alternatives evaluated concurrently and (b) the inferred attribute was one that could logically be absent

Recently there has been considerable interest in models of consumers' evaluative processes which assume that consumers infer a value for salient attributes for which information is not available (Ford and Smith 1987; Gardial and Biehal 1987; Jaccard and Wood 1988; Johnson 1989; Johnson and Levin 1985; Meyer 1981; Zwick 1988). While there is some question about how broadly these models may be applied--there is evidence that such inferences are often not made unless explicitly prompted (Huber and McCann 1982; Lim, Olshavsky, and Kim 1988; Simmons 1988; Zwick 1988)--it is nonetheless clear that such inferences are sometimes made in the ordinary course of evaluating a product. In addition, there are naturally occurring situations which may be used to explicitly prompt inferences about a product--most obviously, in advertising. Therefore, it is important to understand not only the contingencies which encourage or dampen inferences about missing attributes, but also the processes by which such inferences are made.

ALTERNATIVE INFORMATION SOURCES

The literature has focused on two basic ideas about how inferred values are determined. The first is that inferences are based on some knowledge about the object of judgment. The second is that inferences are based on some knowledge or expectation about objects which are comparable to the object of judgment.

Knowledge about the ObJect of Judgment

Inferences based on knowledge about the object of judgment--specifically, a brand--have been labelled same-brand processes (Ford and Smith 1987), attribute-based processes (Zwick 1988), and intra-alternative processes (Johnson 1989). The term same-brand will be used in this paper. It is generally assumed that same-brand inferences follow a rule of consistency: the missing attribute is inferred to be in some way consistent with known attributes. Two types of consistency have been posited: evaluative consistency and probabilistic consistency. Evaluative consistency refers to a process in which the inferred attribute is assumed to be similar in evaluative implications to known attributes. In contrast, probabilistic consistency refers to an inference process which is based, not on the evaluative implications of known attributes, but rather on their expected-ecological correlation with the inferred attribute.

Research which has examined same-brand inference processes has generally concentrated on probabilistic inferences (Ford and Smith 1987; Huber and McCann 1982; Johnson and Levin 1985; Simmons 1988; Yamagishi and Hill 1981), although at least two studies have examined evaluatively consistent inferences (Jaccard and Wood 1988; Zwick 1988). There has been little systematic examination of the relative importance of these two types of consistency, however. Fishbein and Ajzen's (1975) review of research on trait inference suggests that probabilistic consistency is preferred as long as there is a strong correlation between a known attribute and the inferred attribute; otherwise, evaluative consistency is used. Dick, Chakravarti, and Biehal (1988) have presented results in a product evaluation study which are consistent with this conclusion.

Knowledge or Expectations about Comparable Objects

Inferences based on knowledge or expectations about comparable products have been labelled other-brand processes (Ford and Smith 1987), category-based processes (Zwick 1988), and inter-alternative processes (Johnson 1989). The term other-brand processes will be used in this paper. Other-brand processes involve the inference that an undescribed attribute has a value similar to that for comparable products. Most often, it has been assumed that such inferences are based on the expected average value for the product class or choice set.

There is evidence that alternatives are often evaluated in ways consistent with an average-value or discounted average-value inference for undescribed attributes (Ford and Smith 1987; Jaccard and Wood 1988; Meyer 1981; Slovic and MacPhillamy 1974; Yamagishi and Hill 1983; Yates, Jagacinski, and Faber 1978; Zwick 1988). It should be noted, however, that one can explain these findings without assuming that inferences are made. Using process-tracing data, Simmons (1988) found evidence that subjects treated the absence of information as a neutral or negative cue, but they did not actually make an inference about the missing attribute. This process results in evaluations that are consistent with what would be expected if neutral or negative inferences were made, and could lead researchers to overestimate the frequency of such inferences. The distinction between treating the absence of information as a neutral or negative cue and actually making an inference about a missing attribute is not trivial. Inferences create attribute "knowledge" which may subsequently affect responses to the product in a number of ways. For example, inferring that a product is expensive may discourage a person on a tight budget from considering the product. Treating the absence of price information as a negative cue--which lowers one's evaluation, but does not create a specific belief about price--may not have the same effect.

CONTINGENCIES AFFECTING USE OF ALTERNATIVE INFORMATION SOURCES

The Best-Predictor Hypothesis

A common theme in inference research is that inferences about missing attributes are made by a very rational process. It has often been assumed that a same-brand process--probabilistic inference--is used when there is an expected correlation between a known attribute and the inferred attribute, whereas an other-brand process--average-value or discounted average-value inference--is used when the inferred attribute is expected to be uncorrelated with any known attribute (Huber and McCann 1982; Johnson 1989; Johnson and Levin 1985; Meyer 1981; Simmons 1986; Yamagishi and Hill 1981). [The models are generally more complex than this in that inferences are predicted by the linear regression between a known attribute and the inferred attribute. The intercept term can capture any systematic bias in inferences, as well as any effect of expected average values on probalistic inferences.] This is essentially an assumption that consumers use the best predictor available. If there is no correlation between a known attribute and the missing attribute, the expected average value is the best predictor. If there is a correlation, then prediction can be improved by using the correlation (Simmons, 1986).

Although this prediction seems straightforward, it has not been directly tested. A strong test requires both that expected correlations be varied from high positive and negative to zero, and that process-tracing data be collected. When expected correlations are varied, probabilistic inferences can be diagnosed from patterns in overall evaluations (see Johnson and Levin, 1985, for these predictions). In such outcome data, however, average-value inferences are not distinguishable from a response in which no inference is made, but the absence of information influences the evaluation. Protocol data may thus be particularly useful in detecting average-value inferences. Protocols are also potentially useful in testing the competing hypothesis that when the inferred attribute is not expected to be correlated with a known attribute, evaluatively consistent--rather than average-value-inferences are made. Depending upon the correct model of information integration, evaluatively consistent inferences may or may not be detectable in outcome data alone.

Assessment of the Accuracy of Expectations

If the choice of an inference strategy depends upon the nature of expected correlations between the inferred attribute and known attributes, then one might expect inference makers to be sensitive to evidence about the validity of these expectations. We have evidence that consumers are able to detect attribute correlations (Bettman, John, and Scott 1986) and that decision processes may be affected by such correlations (Huber and Klein 1988). Therefore, one might expect that the probability of making a probabilistic inference decreases when the relevant attributes in the observed product set are uncorrelated, while the probability of making an average-value inference decreases when attributes in the observed product set are highly correlated. Although Simmons (1986, 1988) found no evidence that the extent to which observed correlations were consistent with expectations influenced the frequency of inference making, it remains possible that for those who do make inferences, the process is altered.

Spontaneous versus Prompted Inferences

There is some evidence that spontaneous inferences tend to be based on same-brand information, whereas inferences which have been explicitly prompted are more variable, drawing on both same- and other-brand information sources (Ford and Smith 1987; Gardial and Biehal 1987; Zwick 1988). At least one study, however, has presented conflicting evidence that spontaneous inferences are based- almost exclusively on other brand information (Jaccard and Wood 1988). This mixed evidence is not surprising because these studies have generally varied in their sensitivity to various inference processes.

There are, however, a number of reasons that one might expect prompted inferences to be more likely to draw on other-brand sources than are spontaneous inferences. First, such an effect could be caused by a selection bias. Huber and McCann (1982) and Hansen and Zinkham (1984) have argued that unprompted inferences are rarely made unless there is a strong expected correlation between a known attribute and the inferred attribute--it is the expected correlation which cues the inference. This implies that unprompted inferences are made primarily by those people who are likely to make a probabilistic inference. Prompting then encourages those who do not hold strong correlational expectations to make inferences, and these people are perhaps more likely to draw on other-brand information sources.

If this explanation is correct, then any external cue which encourages inferences should reduce the relative frequency of probabilistic inferences. For example, if an undescribed attribute becomes salient because competitive products emphasize that attribute, this may encourage inferences even though there is no correlational expectation. These inferences may be less likely to be probabilistic than inferences which are made in the absence of any such contextual cues.

Another reason that one might expect prompted inferences to be more likely to draw on other-brand information sources is motivational. Prompting may increase the motivation to make a valid inference by making the inference an integral part of the task. According to this explanation, the process by which inferences are made is altered. Feldman and Lynch (1988) have argued that people try to achieve their processing goals using the most easily accessible inputs. Only when these inputs are insufficiently diagnostic are other more diagnostic inputs retrieved from memory or sought externally. Since the explicit goal of a judgment task is to evaluate information about the object of judgment, same-brand information sources for inferences tend to be highly accessible. Inferences are therefore likely to be based on same-brand information. When the same-brand information is insufficiently diagnostic, as when prompting increases concern about the validity of the inference, the decision maker may supplement same-brand information with the less accessible other-brand information.

If this explanation is correct, then any manipulation which increases concern for the validity of the inference should have a similar effect. For example, being asked to report one's inference may increase one's concern for validity. On the other hand, an external cue which merely focuses attention on the missing attribute may not have this effect.

These two explanations for greater variability in the informational bases of prompted inferences-selection and motivation--are not mutually exclusive. Both processes may operate concurrently.

OVERVIEW OF STUDY

Process-tracing data were collected in a series of studies which examined spontaneous inferences made during product evaluation. Although these studies were not specifically designed to examine the contingencies governing the use of alternative information sources, they provide an opportunity to test a number of hypotheses suggested by the preceding discussion.

First, the "best-predictor" hypothesis is tested:

H1: When there is an expected correlation between a known attribute and the inferred attribute, inferences tend to be probabilistic, whereas when there is no expected correlation, inferences tend to be based on expected average values for the product class or choice set.

Consistent with this kind of rational inference strategy, it is predicted that inference makers are sensitive to evidence about the validity of their beliefs about attribute correlations

H2: If there is an expected correlation between a known attribute and the inferred attribute, the frequency of probabilistic inferences decreases when the attributes in observed product sets are uncorrelated. If there is no expected correlation between a known attribute and the inferred attribute, the frequency of average-value inferences decreases when the attributes in observed product - sets are highly correlated.

Finally, it was argued that prompted inferences are more likely to draw on other-brand information sources than are unprompted inferences. Two causal mechanisms were suggested: a selection bias and a motivational effect. It was argued that a selection bias effect would occur any time an external cue encourages inferences, whereas the motivational effect would occur only when concern for the validity of the inference is increased. In the current study, the selection bias effect is tested. Therefore, it is predicted that:

H3: The overall frequency of inference making increases when an external cue encourages inferences.

H4: The overall frequency of probabilistic inferences is not influenced by the presence of an external cue.

H5: The relative frequency of making a probabilistic inference is lower when an external cue encourages inferences.

Hypothesis 5 follows from Hypotheses 3 and 4. However, the relative frequency of making a probabilistic inference could be lowered for reasons other than a selection bias. Hypotheses 3 and 4 test the hypothesis that a selection bias is the causal mechanism.

METHOD

The data are from three studies (Experiments I, II, and IV) which are described in detail in Simmons (1986). A brief description of the method follows.

Subjects

Subjects were undergraduate psychology or business majors who participated in partial fulfillment of a course requirement. A total of 600 subjects participated in these studies, but with the exception of the tests of hypotheses 3 and 4, the following analyses are based only on the subjects who spontaneously made inferences (as indicated by protocols) during the course of the evaluation task.

TABLE 1

EXPECTED CORRELATIONS BETWEEN ATTRIBUTES

Pre test

Two product classes were examined: refrigerators and steam carpet cleaners. These product classes were chosen because it was possible to describe these products in terms of attributes which pretest subjects believed to be either highly correlated or uncorrelated with the undescribed attributes of interest. These attribute pairs and their expected correlations are illustrated in Table 1.

Design

The design can be thought of as a 2 x 3 between-subjects factorial, with two levels of expectations about the relationship between the known attribute and the inferred attribute (high and zero) and three levels of evidence regarding the validity of these expectations (confirming, disconfirming, and no evidence). [With the exception of the no-expected-correlation/ disconfirming-evidence condition, at least two of the original experiments (I, II, and IV) contributed subjects to each cell of this design.] Embedded in this design is a third factor, the presence or absence of an external cue. As will be described below, an external cue was provided in the confirming- and disconfirming-evidence conditions, but not in the no-evidence condition.

Expectations about correlations. One attribute descriptions of refrigerators or carpet cleaners (e.g., refrigerators described only in terms of capacity) were evaluated by some subjects in a context in which most other alternatives were described in terms of two attributes (e.g., capacity and warranty) and by other subjects in a context in which all other alternatives were also described only in terms of one attribute (capacity in this example). Interest focuses on the process by which the one-attribute descriptions were evaluated. It was expected that in the contexts in which most alternatives were described in terms of two attributes, inferences would tend to focus on the second attribute for which information was sometimes available, and this was found to be true. In a between-subjects manipulation, the undescribed attribute made salient by the context, and therefore the inferred attribute, was either one expected to be correlated with the known attribute or one that was expected to be uncorrelated with the known attribute. For example, the no-expected-correlation condition includes four groups of subjects: those who received information about capacity and made an inference about warranty, those who knew warranty and inferred capacity, those who knew area cleaned and inferred versatility, and those who knew versatility and inferred area cleaned. The expected-correlation condition includes all subjects who received one of the remaining twelve combinations selected in the pretest (see Table 1).

The remaining subjects--who never had information about a second attribute--were classified post hoc according to whether the attribute they inferred was expected (based on the pretest) to be correlated with the known attribute.

Evidence regarding validity of expectations. In the second between-subjects manipulation, subjects who did usually have information about a second attribute saw product sets in which correlations were either consistent or inconsistent with expectations. These were the confirming- and disconfirming-evidence conditions, respectively. The no-evidence condition consisted of those subjects who never had any information about a second attribute.

Presence or absence of an external cue. Subjects in the confirming- and disconfirming evidence conditions were cued by the context to think about specific undescribed attributes because information was provided about these attributes for other products being evaluated concurrently. These subjects constituted the external-cue condition. The subjects in the no-evidence condition never had any information about the inferred attribute, and thus were not cued by the context to make the inference. They constituted the no-external-cue condition.

Other design factors. In a within-subjects manipulation, product descriptions varied in attractiveness across four levels. Each subject made evaluations for a set of alternatives in each product class. As it turned out, only 10% of the subjects made inferences in both product classes. The two observations from these subjects are treated as independent observations in the following analyses.

Procedure

The task was presented as one of examining a catalog in anticipation of a purchase. A general description of the features common to all models (of refrigerators or carpet cleaners) was followed by specific information about individual models. Subjects were asked to evaluate each alternative. Following completion of this task, they were asked to retrospectively describe how they went about forming their evaluations.

Dependent Variable

Inferences revealed in the protocols were coded according to the type of inference process used. Only inferences about the attributes examined in the pretest are included in the following analyses. This is because the expected correlation between the known attribute and the inferred attribute is a variable in the design, and expected correlations were known only for the attributes from the pretest. Ninety-eight percent of these inferences were found to fall into one of the following categories: (a) probabilistic, (b) average-value, (c) feature-absent, and (d) indeterminate. An inference was coded as probabilistic if the subject stated a probabilistic inference rule (e.g., "the lower the price, the fewer cubic feet"), or if multiple inferences made by the subject varied as predicted by the expected correlation (e.g., for a low energy cost: "It's probably a small refrigerator;" for a high energy cost: "It's probably quite large"). [In the case of a positive expected correlation, it is possible that some subjects coded as making a probabilistic inference actually relied on evaluative consistency, but no subjects made comments which indicated such a process.] An inference was coded as average-value if subjects indicated assuming an average value (e.g., "I assumed [the area cleaned] was average amount"), or if the subject inferred the same value of the missing attribute for all alternatives evaluated (no examples of this category were observed). [This definition of an average-value inference was adopted to prevent underestimating the number of average-value inferences which would occur if subjects did not use terms like "average", "typical", etc.] An inference was coded as feature-absent if subjects assumed that the product did not have the feature (e.g., "Why get one without a warranty if you can get one with one?"). If it was not possible to determine the type of inference made. the inference was coded as indeterminate.

RESULTS

Inferences were coded independently by two coders, and discrepancies were resolved by discussion. There were initially disagreements on 13% of the inferences, but 60% of these differences involved the strictness with which the rule governing the use of the indeterminate category was applied, and were resolved by coding the inferences as indeterminate type.

Eighteen percent of all inferences made were coded as indeterminate type and were not included in the following analyses. This category occurred more frequently in the expected-correlation condition than in the no-expected-correlation condition (proportions =.18 and .03, respectively, p < .01). [All reported significance levels are for Fisher's exact test.] Based on the intercoder discrepancies, it appears that this occurred because at least 40% of the indeterminates were apparently probabilistic inferences, which were more likely to be made in the expected-correlation condition, as will be shown below. If this is correct, then it appears unlikely, given the results of the following analyses, that inability to classify these inferences introduced any bias.

Test of Hypothesis I

Hypothesis 1 predicts that when there is an expected correlation between the known attribute and the inferred attribute, inferences tend to be probabilistic, whereas when there is no expected correlation, inferences tend to be based on average values. The proportions of subjects making each kind of inference are presented in Table 2.

As predicted, the probability of making a probabilistic inference was higher when there was an expected correlation between the known attribute and the inferred attribute (p < .01). In fact, only one out of 39 subjects made a probabilistic inference in the no-expected-correlation condition, although more than three quarters of the subjects in the expected correlation condition did. This relationship held both when there was confirming evidence (p < .01) and when there was disconfirming evidence (p < .02). When there was no evidence about the validity of the expectation, there was no difference between cells, but this result is based on a sample of 1 in the no-expected-correlation cell.

Contrary to Hypothesis 1, the probability of making an average-value inference was not higher when there was no expected correlation between the known attribute and the inferred attribute. In fact, only one subject made an average-value inference and that subject was in the expected-correlation condition. For this reason, no further tests of effects on average-value inferences were conducted. [A feature-absent inference may be construed as an average-value inference if it is typical for members of a product class to not have a certain feature. Limitations of the design -- specifically, differential representation of attributes which can logically be absent in the expected-correlation and no-expected correlation conditions -- made it impossible for us to examine the effects of expectations about correlations on such inferences.]

TABLE 2

PORPORTIONS OF SUBJECTS MAKING VRAIOUS TYPES OF INFERENCES

There was also no evidence that subjects in the no-expected-correlation condition made evaluatively consistent inferences. No subject made comments which suggested such an inference process, nor were probabilistic inferences--which is how an evaluatively consistent inference could be misclassified- -common.

Test of Hypothesis 2

Hypothesis 2 predicts that inferences consistent with expectations about attribute correlations are more likely to be made when the observed product set confirms expectations than when it disconfirms expectations. In the expected correlation condition, the proportions of subjects making probabilistic inferences were in the predicted direction, but the difference was nonsignificant (p > .38). In the no-expected-correlation condition, the hypothesis could not be tested. No average-value inferences were made at all, so they could not vary as a function of the nature of evidence. However, feature-absent inferences may also be construed as being consistent with the expectation that the inferred attribute is uncorrelated with the known attribute. Whether evidence was confirming or disconfirming had no impact on the frequency of feature-absent inferences either; all inferences in these cells were feature-absent.

Tests of Hypotheses 3, 4 and 5

Hypothesis 3 predicts that the overall frequency of inference making increases when there is an external cue. For this test and the test of hypothesis 4, all subjects were included in the analysis, not just those who made inferences. Hypothesis 3 was supported, although the effect was small. The proportion of subjects making inferences when inferences were not cued by the context-was .15 and when inferences were cued by the context was .21 (p < .02). [See Simmons (1988) for detailed analyses of the effect of an external cue on inferences about individual attributes.]

Hypothesis 4 predicts that the overall frequency of making probabilistic inferences is not affected by the presence of an external cue. This hypothesis was not supported. The overall frequency of probabilistic inferences decreased when inferences were cued by the context. These proportions were .13 and .07 in < .02).

Hypothesis 5 predicts that the relative frequency of probabilistic inferences is smaller when there is an external cue. This hypothesis was supported. Of the subjects who made inferences, the proportion making a probabilistic inference was 1.00 when the inference was not cued by the context and .40 when the inference was cued by the context @ <.01).

Taken together, these results suggest that the external cue encouraged subjects who would not have otherwise done so to make an inference, and that it either suppressed inferences by some who would have made a probabilistic inference or altered the process by which inferences were made. The second explanation perhaps seems most plausible. Almost all non-probabilistic inferences made were feature-absent inferences. These inferences were only made in the external-cue conditions, and only for three of the ten inferred attributes examined: warranty, shelves, and versatility. Based on subjects' comments, a more accurate name for the versatility dimension is "restrictions on use," as this is how they interpreted versatility information (e.g., "can only be used on synthetic carpets and rugs"). In contrast to the other inferred attributes examined (energy cost, price, capacity, and weight), it is logically possible that a warranty, shelves, and restrictions on use could be absent. Therefore, it appears that the effect of the external cue was both to increase the frequency of inferences and to encourage a particular inference process for those attributes which could in fact be absent

DISCUSSION

When the results of this study are considered in the context of other research which has examined similar issues, we gain a number of insights into how alternative information sources may be used in inferences about missing attributes.

The "Best-Predictor" Hypothesis

As predicted, probabilistic inferences predominated when the inferred attribute was expected to be correlated with a known attribute. Contrary to prediction, average-value inferences were not made when these attributes were expected to be uncorrelated .

Many researchers have found evidence that incompletely described alternatives are evaluated in ways consistent with average-value or discounted average-value inferences about missing attributes (Jaccard and Wood 1988; Meyer 1981; Slovic and MacPhillamy 1974; Yamagishi and Hill 1983; Yates, Jagacinski and Faber 1978; Zwick 1988). In these studies, however, it is not always possible to distinguish between a strategy in which neutral or negative inferences are made about the-missing attribute and one in which the absence of information is simply treated as a neutral or negative cue. This is an important distinction because inferences can create product "knowledge" which may affect subsequent behavior.

Some of the strongest support for average-value inferences comes from a study in which evidence of average-value inferences in evaluations was found only when inferences were prompted (Zwick 1988)--presumably, the difference in evaluations between the unprompted and the prompted conditions was due to inferences--and from a study in which process-tracing data showed a high rate of average-value inferences made spontaneously during a choice task (Gardial and Biehal 1987). In contrast, the current study examined spontaneous inferences in a judgment task.

It may be that average-value inferences occur more frequently in these situations--when inferences are prompted and in choice tasks--than in judgment tasks in which inferences are not prompted. It was argued above that prompting may increase average-value inferences in two ways. First, prompting may create a selection bias by increasing the likelihood that those who do not hold correlational expectations Will make inferences. Such people are perhaps more likely to make average-value inferences. Second, prompting may increase motivation to make a valid inference, and thus increase the likelihood that the decision maker will supplement same-brand information with the less accessible other-brand information---such as an expected average value--as input to the inference.

Similarly, in choice tasks, the motivation to make a valid inference, or to make an inference at all, may be higher than in judgment tasks. In evaluation, processing tends to proceed by alternative. In choice tasks, processing tends to be dimensional; alternatives are compared on each dimension (Payne 1982). The tendency to make dimensional comparisons may increase the desire to make valid inferences about attributes for which information is available for some of the alternatives, but not for others.

This is not to claim that average-value inferences are' never made spontaneously in the course of evaluating a product, but rather that this and other inference processes may be influenced by variables such as prompting and processing goals. Process-tracing techniques may be especially useful for identifying other variables that encourage or dampen particular inference processes.

Some evidence suggests that evaluatively consistent inferences, rather than average-value inferences, are made when there is no expected correlation between the inferred attribute and a known attribute. In particular, Dick, Chakravarti, and Biehal (1988) found that evaluatively consistent inferences were made when information about correlated attributes was not available, but only among subjects who were prompted to make an inference. We found no evidence of such a process, but their task situation differed from ours in a number of ways. While the evidence is too limited to allow any conclusions, there is clearly a need, as with average-value inferences, to examine the task and contextual factors which lead one to the use of an evaluative consistency inference strategy.

Probabilistic Inference and Contingent Decision Behavior

Probabilistic inferences depend upon the expectation that the inferred attribute is correlated with a known attribute. One might expect, then, that a decision maker's inference strategy could be altered by varying the correlations in task product sets. We found little evidence that this occurred.

This result is consistent with a view of contingent decision behavior recently discussed by Bettman (1988). He argued that there are two distinctive ways in which decision strategies can be adaptive to environmental conditions. The first is a top-down mode in which the environment is assessed in a deliberate manner in order to select a strategy tailored to the environment. In a top-down mode, for example, an inference maker would assess attribute correlations before deciding whether to implement a probabilistic inference strategy. In a bottom-up mode, the process of adapting strategy to the environment is not so foresightful. Instead of choosing a strategy based on a thorough assessment of the environment, the decision maker begins with a strategy which seems reasonable, and then alters this strategy in response to task feedback. For example, an inference maker might begin by making probabilistic inferences because of a belief that two attributes are correlated, but alter this strategy as disconfirming evidence accumulates.

Bettman argues that much adaptation to the environment occurs in the bottom-up mode. In this mode, adaptation occurs over time. Therefore responsiveness to contextual factors such as attribute correlations may not be apparent in the short run. Our subjects exhibited a nonsignificant tendency to make fewer probabilistic inferences in the disconfirming-evidence condition. With more experience in the task environment, they might have shown stronger evidence of an adaptive strategy. Others have also reported results consistent with a bottom-up adaptation to correlations in task product sets (Huber and Klein 1988; Klein and Yadav 1989).

Effect of an External Cue

The external-cue manipulation was intended to test the hypothesis that a selection bias occurs when inferences are cued externally, as when one is prompted to make an inference. It was expected that the cue would decrease the relative frequency of probabilistic inferences by increasing the likelihood that people who did not hold correlational expectations would make inferences.

In the external-cue condition, some alternatives were described in terms of two attributes. It was expected that when the one-attribute descriptions were evaluated, this manipulation would encourage inferences about the second attribute, and this was found to be true. This particular cue, however, not only increased the overall frequency of inference making, but it also initiated specific kinds of inferences. When it was logically possible for a feature to be absent, this is what was often inferred. As it seems unlikely that prompting would have a similar effect, these results should not be taken as evidence regarding the effects of prompting on inferences.

These results have both practical and methodological importance, however. Since consumers often encounter just such differences in the availability of attribute information, feature-absent inferences may be relatively common. In addition, it should be recognized that the manipulations we employ to encourage inferences may alter the process under study. This criticism is relevant to the current study. However, the results in our no-external-cue condition are consistent with our conclusion that probabilistic inferences predominate when there is an expected correlation, whereas spontaneous average-value inferences may be rare in these kinds of judgment tasks.

Limitations of Study

Protocols are imperfect indicators of cognitive processes. In particular, retrospective protocols may be subject to errors in encoding and retrieval. Our subjects were always shown the product descriptions during the protocol procedure to aid recall, but it is possible that they inaccurately reported how they formed their evaluations, or were perhaps more likely to articulate probabilistic inferences than average-value inferences.

Although multiple exemplars of inferred attributes expected to be correlated or uncorrelated with known attributes were examined, some of the differences in the inferences made in these two conditions could be due to the particular exemplars examined. This is more of a limitation in the no-expected-correlation condition, in which there were four exemplars, than in the expected-correlation condition, in which there were twelve.

REFERENCES

Bettman, James R. (1988), "Processes of Adaptivity in Decision Making," in Advances in Consumer Research, Vol. 15, ed. Michael J. Houston, Provo, UT: Association for Consumer Research, 1-3.

Bettman, James R., Deborah Roedder John, and Carol A. Scott (1986), "Covariation Assessment by Consumers," Journal of Consumer Research, 13 (December), 316-326.

Dick, Alan, Dipankar Chakravarti and Gabriel Biehal (1988), "Memory Based Inferences During Consumer Choice," Unpublished working paper, State University of New York, Buffalo.

Feldman, Jack M. and John G. Lynch, Jr. (1988), "Self-Generated Validity and Other Effects of Measurement on Belief, Attitude, Intention, and Behavior," Journal of Applied Psychology, 73 (August), 421-435.

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

Ford, Gary T. and Ruth Ann Smith (1987), "Inferential Beliefs in Consumer Evaluations: An Assessment of Alternative Processing Strategies," Journal of Consumer Research, 14 (December), 363-371.

Gardial, Sarah and Gabriel Biehal (1987), "Measuring Consumers' Inferential Processing in Choice," in Advances in Consumer Research, Vol. 14, eds. Melanie Wallendorf and Paul F. Anderson, Provo, UT: Association for Consumer Research, 101-105.

Hansen, Chris J. and George M. Zinkham (1984), "When Do Consumers Infer Product Attribute Values?" in Advances in Consumer Research, Vol. 11, ed. Thomas C. Kinnear, Provo, UT: Association for Consumer Research, 187-192.

Huber, Joel and Noreen Klein (1988), "Choosing Which to Choose: A Model of Cutoff Selection in a Two-Stage Decision Process," Unpublished working paper, Duke University, Durham, NC.

Huber, Joel and John McCann (1982), "The Impact of Inferential Beliefs on Product Evaluations," Journal of Marketing Research, 19 (August), 324-333.

Jaccard, James and Gregory Wood (1988), 'The Effects of Incomplete Information on the Formation of Attitudes Toward Behavioral Alternatives," Journal of Personality and Social Psychology, 54 (April), 580-591.

Johnson, Richard D. (1989), "Making Decisions With Incomplete Information: The First Complete Test of the Inference Model," Advances in Consumer Research, 16, Thomas K. Srull, ed., Provo: UT: Association for Consumer Research.

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, 12 (September), 169-177.

Klein, Noreen M. and Manjit S. Yadav (1989), "Context Effects on Effort and Accuracy in Choice: An Enquiry into Adaptive Decision Making," Journal of Consumer Research, 15 (March) 411-421.

Lim, Jeen-Su, Richard W. Olshavsky, and John Kim (1988), "The Impact of Inferences on Product Evaluations: Replication and Extension," Journal of Marketing Research, 25 (August), 308316.

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

Payne, John W. (1982), "Contingent Decision Behavior," Psychological Bulletin, 92 (September), 382-402.

Simmons, Carolyn J. (1986), "Effects of Missing Information on Product Evaluations," Unpublished doctoral dissertation, University of Florida, Gainesville.

Simmons, Carolyn J. (1988), "Effects of Missing Information on Product Evaluations: The Influence of Attribute Information Provided by Competitors," Unpublished working paper, University of Illinois, Urbana-Champaign.

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

Yamagishi, Toshio and Charles T. Hill (1981), "Adding Versus Averaging Models Revisited: A Test of a Path-Analytic Integration Model," Journal of Personality and Social Psychology, 41 (July), 13-25.

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, 44 (May), 942-951.

Yates, J. Frank, Carolyn M. Jagacinski, and Mark D. Faber (1978), "Evaluation of Partially Described Multi-Attribute Options," Organizational Behavior and Human Performance, 21 (April), 240-251.

Zwick, Rami (1988), "Same- and Other-Brands Information Sources in the Formation of Inferential Beliefs about Partially Described Multiattribute Products," Working Paper 173, Working Series in Marketing Research, The Pennsylvania State University, University Park.

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