The Measurement and Moderating Role of Confidence in Attributions

Donald R. Lichtenstein, Louisiana State University
Scot Burton, Louisiana State University
ABSTRACT - Although attribution theory explicitly recognizes the role of both the extremity and confidence with which an attribution is made, consumer researchers have largely failed to explicitly consider attribution confidence in their studies. It has been suggested that attribution confidence may moderate the relationship between attribution extremity and subsequent decision-making processes. If this suggestion has validity, then the relationships reported between attributions and related constructs -may be attenuated. The present study employs three alternative measures of attribution confidence in investigating the moderating role of attribution confidence on attribution extremity-attribution outcome relationships. Directional support for two of these three measures as moderating variables was found.
[ to cite ]:
Donald R. Lichtenstein and Scot Burton (1988) ,"The Measurement and Moderating Role of Confidence in Attributions", in NA - Advances in Consumer Research Volume 15, eds. Micheal J. Houston, Provo, UT : Association for Consumer Research, Pages: 468-475.

Advances in Consumer Research Volume 15, 1988      Pages 468-475

THE MEASUREMENT AND MODERATING ROLE OF CONFIDENCE IN ATTRIBUTIONS

Donald R. Lichtenstein, Louisiana State University

Scot Burton, Louisiana State University

ABSTRACT -

Although attribution theory explicitly recognizes the role of both the extremity and confidence with which an attribution is made, consumer researchers have largely failed to explicitly consider attribution confidence in their studies. It has been suggested that attribution confidence may moderate the relationship between attribution extremity and subsequent decision-making processes. If this suggestion has validity, then the relationships reported between attributions and related constructs -may be attenuated. The present study employs three alternative measures of attribution confidence in investigating the moderating role of attribution confidence on attribution extremity-attribution outcome relationships. Directional support for two of these three measures as moderating variables was found.

INTRODUCTION

Attribution theory "has to do with the processes by which man 'knows' his world and, more importantly, knows that he knows, that is, has a sense that his beliefs and judgements are veridical" (Kelley 1973, p.107). Consistent with Kelley's perspective, there are two dimensions of an attribution: the extremity of the attribution and the confidence with which it is made. This two-dimensional perspective is also consistent with correspondent inference theory (Jones and Davis 1965). For example, in making correspondent inference (i.e., trait) attributions, the observer's confidence that the actor is extreme on a trait represents the strength of the correspondent inference attribution.

Despite the theoretical importance of attribution confidence, consumer attribution researchers have generally failed to incorporate measures of attribution confidence into their attribution operationalizations. Instead, attribution confidence is most often omitted from attribution studies (e.g., Settle and Golden 1974; Smith and Hunt 1978; Sparkman and Locander 1980), or it is measured but treated solely as a dependent variable. In general, the effect of confidence on the relationship between attribution extremity and other variables is ignored (e.g., Calder and Burnkrant 1977). At this point, the implications of not assessing attribution confidence, or treating it solely as a dependent variable, are unknown.

ATTRIBUTION EXTREMITY - CONFIDENCE MEASUREMENT

Extremity in attribution has usually been measured by the degree to which a subject responds toward one of the anchors on bipolar adjective scales (e.g., sincere-insincere (cf. Calder and Burnkrant 1977)), or similarly, toward one of the anchors on internal-external attribution scales (cf. Mizerski 1975). Extremity also has been measured as the degree to which subjects respond toward the probable anchor of a probable-improbable attribution scale (cf. Lichtenstein and Bearden 1986), or the degree to which subjects assign a high percentage probability on an ipsative causal allocation scale (cf. Mizerski 1978). When consumer researchers have measured attribution confidence, the measures have primarily consisted of subjects' self,reported confidence on a "confident-not confident" scale (cf. Calder and Burnkrant 1977; Mizerski 1975).

One school of thought is that the confidence with which an attribution is made is captured by the extremity of response to an attribution scale; that is, as we are more extreme in our attributions, we are also more confident. Therefore, confidence measures are not needed because they are redundant with measures of attribution extremity. However, a second school of thought is that the confidence with which an attribution is made is not totally captured by the extremity of response to an attribution scale. Mizerski, Golden, and Kernan (1979) have hypothesized that higher attributional confidence may result in a higher probability of actions resulting from attributions. If this hypothesis has merit, failure to explicitly measure confidence may result in an attenuation of relationships of attributions with both antecedent and outcome variables. In other words, attribution confidence may moderate the impact of information on attribution extremity. Also, consistent with the hypothesis of Mizerski et al. (1979), it may moderate the impact of attributions on outcomes such as subsequent perceptions, attitudes, and behavior [Although attribution theory only specifically addresses the process of information input to attribution output, the impact of attributions on subsequent decision making processes is a natural extension (Bem 1972; Kelley 1973; Mizerski et al. 1979).].

The purpose of the present study is to address the following two questions: (1) do measures of attribution extremity fully reflect attribution confidence, and (2) if not, does attribution confidence moderate the relationship between attribution extremity and subsequent cognitions. A necessary condition for evidence of attribution confidence acting as a moderator variable is that attribution extremity measures not totally encompass attribution confidence. Therefore, positive evidence pertaining to the second question above will also serve as negative evidence for the first question. A study is described in which three different measures of attribution confidence are used and the moderating role of confidence on the relationship between attribution extremity and attribution outcomes is investigated for its consistency with theoretically-based hypotheses. (Because the substantive meaning of the hypotheses are not at issue here, but rather, are provided as only a theoretical basis for testing for moderating relationships, the hypotheses are provided after the measures are discus$ed.) Results are discussed and implications for attribution research are provided.

Attribution Extremity Measurement

One of the most common methods of measuring attribution extremity is by the use of independent rating scales. Independent rating scales are multi-item scales in which a response to one scale-item does not preclude any response to another scale-item. Attribution extremity has been measured via independent rating scales in the following three ways: (1) the degree to which subjects responded toward either anchor on bipolar trait attribution scales (cf. Calder and Burnkrant 1977), (2) through the use of internal-external attribution scales (Mizerski 1975), and (3) the degree to which subjects respond toward the probable anchor on probable-improbable scales (cf. Lichtenstein and Bearden 1986). There is evidence which suggest that independent rating scales are more reliable and valid than open-ended and ipsative methods of measuring attribution extremity (Elig and Frieze 1979). Consequently, independent rating scales are used in the present study to measure attribution extremity.

Self-Report Measure of Confidence

Calder and Burnkrant (1977) asked respondents to make twenty-seven trait extremity attributions on semantic differential scales. Following the twenty-seven scales, a single confidence scale was employed to assess average confidence for all twenty-seven attributions. This confidence measure may not accurately reflect the confidence of those who make different attributions with varying degrees of confidence. However, because of its previous use, it is employed here.

Cognitive Conflict Measures of Confidence

There appears to be support in the literature for two cognitive conflict-based measures of confidence: a measure of causal complexity (Mizerski 1978), and a measure of response latency (Aaker et al. 1980; Tyebjee 1979). Cognitive conflict occurs when an individual experiences competing cognitions. As the number and/or equality (in terms of strength) of competing cognitions increases, cognitive conflict also increases. As conflict increases, individuals should feel less strongly about any particular cognition. Thus, confidence in any particular cognition decreases, and therefore, measures of cognitive conflict can be taken as indicators of confidence (Tyebjee 1979).

Causal complexity refers to a perceiver's attribution allocation among various possible causes. The assumption inherent in using causal complexity as a measure of confidence is that those individuals who: a) recognize more possible causes of an event, and b) weigh potential causes equally with respect to causal probability, are more causally complex, and thus, are less confident in their attribution allocation (Mizerski 1978). Holding condition "a" constant enables a causal complexity score to be calculated as a respondent's dispersion of causal probability assignments across a given number of possible attributions [The number of competing alternatives has also been held constant in other studies of cognitive conflict. For example, in the Tyebjee (1979) and Aaker et al. (1980) studies, subjects' response latency scores were measured by the time it took subjects to choose between two and five brand alternatives, respectively.]. For example, an individual responding that five possible attributions had an equal 20% chance of being the cause of an effect would be of high causal complexity, and thus should have low confidence in any particular attribution. Conversely, an individual who responded that potential cause A had a 96% chance of being the cause, while potential causes B, C, D, and E each had a 1% chance, would be of low causal complexity and should have high confidence in cause A (Mizerski 1978).

Using a set number of probable-improbable scales across respondents to measure attributions, individuals of high causal complexity can be classified as those who respond to many attribution scales with equal probability. These individuals cannot discriminate which attributions are more likely vis-a-vis the other attributions, and therefore, should have less attribution confidence. However, a person of low causal complexity would be one who responded to some attributions as very probable, and to others as very improbable. These individuals can discriminate probable from improbable causes, and therefore, should show more confidence in their attributions.

The second cognitive conflict measure, response latency, is typically defined as the time it takes a subject to make a response (Tyebjee 1979). The assumption inherent in using response latency as a conflict-based measure of confidence is that those individuals with low conflict are more confident of their responses, and thus, respond quicker. Support for this assumption has been found by Tyebjee (1979) and Aaker et al. (1980).

These two conflict-based measures have the desirable property of being "maximally different" from the method used to measure attribution extremity. In addition, all three methods of assessing attribution confidence, the single self-report measure, the response latency measure, and the measure of causal complexity, appear to have the desirable property of being "maximally different" from each other.

METHOD AND HYPOTHESES

Study Procedure

Two hundred and seventy-eight undergraduate business majors at a large state university were randomly assigned to cells in a 2x2x3 laboratory experiment. Because of no-shows, cell sizes ranged from 21 to 28. Manipulations included the consistency with which an advertising merchant made a cash discount claim (high-low), the distinctiveness of the claim vis-a-vis competitors (high-low), and the level of discount (high-medium-low). Measured variables included consumer attributions for the merchant-advertised discount claim, the confidence with which the attributions were made, and also the perception of the value of the advertised deal and attitude towards the deal. Because the purpose of the present study is to investigate relationships between attribution extremity, attribution confidence, and attribution outcomes, the manipulations will not be discussed further. Rather, the analyses are performed within each of the experimental cells in order to obtain results that are unaffected by the manipulations.

Twenty-four experimental sessions were conducted; two each for each of the twelve treatment conditions. Similar to procedures employed by Aaker et al. (1980), upon entering a computer laboratory, subjects were seated at individual desks, each with an IBM personal computer. Subjects were told that they were going to be participating in a study designed to get their reactions to an advertisement that Dawson's Furniture Store (fictitious), a store in another state, was considering for use in the upcoming months. They also were told that the study was being conducted in their city rather than the city where the furniture store operated because the furniture store wanted to take every precaution to maintain its anonymity among potential customers. Subjects were informed that although the store name in the advertisement was changed to add further insurance of anonymity, everything else in the advertisement was factual (i.e., exactly as the advertisement would be used).

TABLE 1

ATTRIBUTION MEASURES AND FACTOR LOADINGS

After exposure to the test advertisement and the manipulations, subjects were told that their feelings about why the store was offering this discount were of interest. Instructions and sample questions were provided both verbally and on the computer before the actual study questions were presented. After all subjects were comfortable with the computer questionnaire procedure, they were allowed to begin.

Study Variables

Attribution Extremity: Subjects responded to thirty-six probable-improbable attribution extremity scales in response to the question, "Why do you think Dawson's furniture store is offering this price deal?" The thirty-six attribution extremity items (shown in Table 1) were determined using the two-step procedure suggested by Elig and Frieze (1979) and Weiner (1985). First, a pretest sample of 66 undergraduate business majors provided reasons why a furniture store might offer a discount on a desk. Thirty-six distinct responses were obtained. These responses were then converted to probable-improbable scales for this study.

The attribution extremity scales were presented one at a time on the computer monitor. Subjects responded to a low probability attribution by depressing a low number key and a high probability attribution by depressing a higher number key. Keys 1 through 9 were used; thus the scale had nine-places. As soon as subjects responded to a particular attribution, both their attribution extremity response (i.e., keys 1 through 9) and their response time (in seconds) were recorded on a data diskette, and the next attribution extremity scale appeared on the screen.

Extremity responses to the thirty-six items were subjected to a principal components analysis to reduce the number of variables to a more manageable size for testing for the moderating role of confidence. Based upon Kelley's (1973) contention that person, stimulus, and circumstance attributions exhaust the causal space (i.e., any cause of an effect must be due to at least one of these three factors), and the empirical support for this taxonomy found by Lichtenstein and Bearden (1986), three factors were extracted [A common exploratory factor analysis using a varimax rotation was also conducted with no substantive differences in structure.]. Items not showing evidence of simple structure were deleted and the remaining items were reanalyzed. The resulting structure consisted of the twenty-one items shown in Table 1. Consistent with Kelley's (1973) contention and the findings of Lichtenstein and Bearden (1986), it does appear that the factors can be interpreted as person, product (stimulus), and circumstance factors, respectively. As such, person, product, and circumstance attribution extremity variables were operationalized by creating summed scales of the respective attribution extremity items (cf. Lichtenstein and Bearden 1986; Sujan 1986). The correlation between the summated product attribution extremity scale and the summated person and circumstance attribution extremity scales was -.15 and .15, respectively. The correlation between the latter two scales was .03. Descriptive statistics and reliability estimates for the attribution extremity scales, as well as other study variables, are provided in Table 2. The reliability for the three attribution extremity scales of .70, .74, and .64, respectively, are consistent with those obtained for attribution constructs in other studies (Sujan 1986).

Attribution Outcomes: The attribution outcome variables were perceptions of the value of the deal offered in the advertisement and attitude towards the deal. The perceived value of the deal was operationalized using four nine-point semantic differential scales reflecting the dimensions of perceived worth, perceived savings, price acceptability, and value for the money. Attitude towards the deal was measured using five nine-point semantic differential scales (e.g., favorable-unfavorable, good-bad). The correlation between the two variables was .77. Coefficient alpha for these two variables were .80 and .92, respectively.

TABLE 2

DESCRIPTIVE STATISTICS AND RELIABILITY ESTIMATES FOR STUDY VARIABLES

Attribution Confidence: Three methods of measuring attribution confidence were employed. Similar to the measure employed by Calder and Burnkrant (1977), a single, self-report confidence in attribution scale was placed after the thirty-six attribution extremity scales. Specifically, the scale was "In general, how confident are you in the answers you just provided?" anchored by "very confident-not at all confident."

The second method of measuring confidence was a single measure of causal complexity. This measure was not a direct one, but rather, was derived by calculating the dispersion of responses to the attribution extremity scales. Specifically, causal complexity was calculated as an individual's standard deviation across all thirty-six attribution extremity scales [Although only 21 of the items loaded on the three factors, items that did not load should still affect causal complexity in the hypothesized manner. Thus, all 36 were used to calculate the causal complexity based measure of confidence.].

The third confidence measure method was a response latency measure which assessed the time interval in seconds from the time that the attribution extremity measure appeared on the computer monitor until the time at which the subject depressed a key to respond to the extremity measure. This time interval was recorded directly on the data diskette; thus the latency measure was unobtrusive. A separate latency score for each of the three summated attribution extremity scales was-calculated by summing individual latency scores for those items that loaded on each of the three attribution extremity factors. Therefore, there was a person latency score comprised of response times for nine items, a product latency score comprised of response times for seven items, and a circumstance latency score comprised of response times for five items. The reciprocal of each latency scale was then calculated and used as the response latency measure to make the direction of latency scores comparable with other confidence measures (i.e., higher scores reflecting higher levels of confidence). Coefficient alpha for these three scales was .64, .57,and .54, respectively.

HYPOTHESES

The primary purpose of this study concerns the role of attribution confidence as a moderator of the relationship between attribution extremity and attribution outcomes. However, a necessary condition for a moderator variable is that there be a relationship to moderate; in this case, this concerns the significance of the relationship between attribution extremity and outcomes. The first set of hypotheses pertain to this necessary relationship between attribution extremity and outcomes.

In the present study, subjects were asked to make attributions for the reason that the furniture dealership was offering a price discount on the advertised desk. It appears plausible to suggest that any product attribution should imply something negative about the desk in order to justify the discount. An examination of the items that loaded on the product factor in Table 1 reveals that these items do suggest something negative about the desk. Therefore, product attribution extremity should be negatively related to the two attribution outcome variables, perceptions of the value of the deal and attitude towards the deal.

Person attributions, on the other hand, imply nothing about the quality of the desk itself, but rather, reflect the motives of the dealership in day-to-day business operations. The items that loaded on the person attribution extremity factor in Table 1 appear consistent with this explanation; the items appear to describe various motives of the dealer. Because the dealership is offering a discount, and the discount is not attributed to an inferior product, person attribution extremity should be positively related to attribution outcomes.

Circumstance attributions have been viewed as a residual category used by observers when person or stimulus attributions are not appropriate. As such, some researchers have called circumstance attributions nothing more than "a confession of ignorance" (Jones and Davis 1965). A review of the items that loaded on the circumstance factor in Table 1 reveal that some appear to imply something negative about the desk (i.e., items 3 and 4), while others do not (i.e., items 2, 6 and 10). Given the conjecture of Jones and Davis (1965) and Jones and McGillis (1976), and the nature of the items comprising the circumstance extremity factor, it seems plausible to suggest that circumstance attribution extremity should not be related to attribution outcomes, and as such, not moderated by attribution confidence.

Based on the above explanations, the following three hypotheses are offered for the impact of attribution extremity on attribution outcomes:

H1: Product attribution extremity will have a negative effect on attribution outcomes.

H2: Person attribution extremity will have a positive effect on attribution outcomes.

H3: Circumstance attribution extremity will not affect attribution outcomes.

The effect of attribution extremity on the attribution outcomes should be stronger when confidence is high. For example, if a perceiver believed that a price discount was due to an inferior product, for a given level of extremity, the product attribution should have a stronger negative impact on attribution outcomes under higher levels of confidence. Consequently, hypotheses concerning the moderating role of attribution confidence are offered below.

H4: The impact of product attribution extremity on attribution outcomes is moderated by attribution confidence.

H5: The impact of person attribution extremity on attribution outcomes is moderated by attribution confidence.

H6: The impact of circumstance attribution extremity on attribution outcomes is not moderated by attribution confidence.

MODERATED REGRESSION ANALYSIS

Moderated regression analysis was used to test whether attribution confidence was a significant moderator of the relationship between attribution extremity and attribution outcomes (Sharma, Durand, and Gur-Arie 1981). Moderated regression analysis is based on a comparison between the following regression equations:

y = a + b1x   (1)

y = a + b1x + b2z   (2)

y = a + b1x + b2z + b3xz    (3)

where y is the dependent variable, a is a constant, x is the hypothesized predictor variable, and z is the suspected moderator. If partial F-tests reveal a significant difference between equations 2 and 3, z is a moderator variable.

In terms of the present study, if attribution confidence moderates the attribution extremity-attribution outcome relationship, higher levels of attribution confidence should result in a stronger attribution extremity-attribution outcome relationship. For product attributions (H4), this means that the b 3 in equation 3 should be negative. For person attributions (H5), b 3 should be positive. For circumstance attributions (H6), it should not be significant.

In order to avoid manipulation-caused effects on the study variables, twelve different moderated regression analyses were performed (one within each of the twelve experimental conditions) using each of the outcome measures as dependent variables. The results of the moderated regression analyses are reported in Tables 3 and 4. Each row of the table represents a summary of the twelve different moderated regression analyses, one for each experimental cell [Pooling data across experimental cells for an overall moderated regression analysis would have aided in the interpretation of results. However, results of the within cell moderated regression analyses revealed differences in the beta coefficients across experimental cells; thus, a pooled analysis across cells would have been inappropriate. For this reason, the data are presented in a within-cell, rather than across-cell, framework.]. The attribution extremity column in Table 3 corresponds to the direction and significance of t-values associated with b1 in equation 1, the confidence column corresponds to the direction and significance of t-values associated with b2 in equation 2, and the interaction column corresponds to the direction and significance of t-values associated with b3 in equation 3.

For example, equation A in Table 3 is read as follows: person attribution extremity had a positive effect on the dependent variable, perceptions of the value of the deal, in 11 of the 12 conditions, and a negative effect on perceptions in only 1 of the 12 conditions. In 3 of the 11 cases in -which there was a positive effect, the effect was significant at the .05 level. The average adjusted R2 across all 12 conditions was .12. When self-report confidence was added to the equation, in 5 of the 12 conditions it had a positive effect on perceptions, in 7 of the 12 conditions it had a negative effect, and none of these effects were significant. The average adjusted R2 across all 12 conditions was .18. When the interaction between person attribution extremity and self-report confidence was added to the equation, in 8 of the 12 conditions, the effect was positive (with p<.05 for two of the conditions), and in 4 of the 12 conditions, the effect was negative. The average adjusted R2 across all 12 conditions when all three terms are in the equation is .27. Because pattern of results are very similar across both dependent variables, for the sake of brevity, the results reported in Tables 3 and 4 are discussed jointly.

A theoretically consistent pattern of results is evident for the attribution extremity-attribution outcome relationship (equation 1) across both outcome variables. That is, product attribution extremity had a negative impact on outcomes (H1 supported), person attribution extremity had a positive impact on outcomes (H2 supported), and circumstance attribution extremity did not affect attribution outcomes in a directionally consistent manner (H3 supported). Further, in those experimental conditions in which attribution extremity had a significant effect, with the exception of the one significant effect of circumstance extremity on attitude, the significance was always in the hypothesized direction for both dependent variables. In sum, the results provide support for the first three hypotheses.

To test if the relationships between attribution extremity and attribution outcomes are moderated by any of the three attributional confidence measures, the individual confidence measures (as indicated by equation 2) were added to the regression equations, followed by a confidence by attribution interaction (as indicated by equation 3). If higher levels of confidence result in a stronger attribution-attribution outcome relationship, a positive person by confidence and a negative product by confidence interaction should result.

TABLE 3

MODERATED REGRESSION ANALYSIS SUMMARY OF POSITIVE (SIGNIFICANT POSITIVE)/NEGATIVE (SIGNIFICANT NEGATIVE) T-VALUES WITHIN TWELVE CONDITIONS FOR THE IMPACT OF ATTRIBUTION EXTREMITY ON PERCEPTIONS AS MODERATOED BY CONFIDENCE

TABLE 4

MODERATED REGRESSION ANALYSIS SUMMARY OF POSITIVE (SIGNIFICANT POSITIVE)/NEGATIVE (SIGNIFICANT NEGATIVE) T-VALUES WITHIN TWELVE CONDITIONS FOR THE IMPACT OF ATTRIBUTION EXTREMITY ON ATTITUDES AS MODERATED BY CONFIDENCE

There appears to be a theoretically consistent pattern of directional support for the moderating impact of the self-report and causal complexity confidence measures. Results are consistent for person attribution extremity on both outcome measures (perceptions and attitude), and product attribution extremity on both outcome measures, with a single exception (equation J in Table 4). For this single situation in which directional support was not found, the three significant interaction effects were all in the hypothesized direction. In fact, for all situations in which product and person attribution extremity was moderated by either self-report or causal complexity confidence (i.e., equations A, B, D, and E in Table 3, and equations 1, K, M, and N in Table 4), the significant effects were all in the hypothesized direction.

The response latency measure failed to show any evidence of a consistent moderating relationship across any of the analyses. In fact, in the four analyses in -which response latency was used as the measure of confidence and a significant interaction was predicted (equations G and H in Table 3, and P and Q in Table 4), the interaction term was in the opposite direction of that which was hypothesized (cf. H4 and H5) in 23 of the 48 analyses (48%).

To further investigate the relationships between the confidence measures, the correlations between each of the three latency scales and the self report and causal complexity measures were examined. The correlation between the latter two measures was .23 to<.01), thus providing evidence of convergent validity. However, the range of correlations of the three latency measures with the two other confidence measures was from -.05 to .08, with an average (absolute value) correlation of .05. These results fail to provide support for the validity of response latency as a measure of attribution confidence. In sum, these results support for H4-H6 for the self-report and causal complexity measures of confidence, but do not support the hypothesized relationships using response latency as a measure of confidence.

DISCUSSION

The evidence from the within-cell analyses appear to lend some support for the contention that attribution confidence moderates the relationship between attribution extremity and attribution outcomes. As such, support also is provided for: a) the contention that attribution extremity does not totally capture attribution confidence, and b) the validity of causal complexity and general self-report measures of confidence.

From a strict statistical significance standpoint, the results of the within cell analyses are not strong. However, it has been argued that statistical significance is in many cases a poor criterion to gauge the qualitative significance of results (Sawyer and Peter 1983). They suggest that more emphasis should be placed on replicability of findings because if a finding can be replicated sufficiently, statistical significance tests are unimportant.

The within-cell analyses conducted in the present study, while not providing the evidence of external validity that a series of twelve successful independent replications would, does provide moderate evidence of replicability across the twelve information conditions. For the 96 moderated regression analyses (2 x attribution extremity measures x 2 confidence measures x 2 attribution outcome variables x 12 conditions) in which causal complexity and self-report were used as confidence measures and a direction was hypothesized for the interaction variable (i.e., H4 and H5), the interaction was in the hypothesized direction in 69 cases, and in the opposite direction in 27 cases. Further, when the moderating effect was statistically significant, the effect was in the hypothesized direction 14 times, and in the opposite direction 0 times. For circumstance attributions, it was hypothesized that confidence would not moderate the attribution extremity-attribution outcome relationship. Consistent with this hypothesis, no consistent pattern of results emerged; of the 48 relevant analyses (2 confidence measures x 2 outcome variables x 12 conditions), 21 of the interactions were positive and 27 were negative.

The results of the self-report and causal complexity confidence measurement methods across the three types of attributions can be taken as evidence of nomological validity. That is, attribution confidence acted as a moderator when there was theoretical justification for it to do so, and did not act as a moderator when there was theoretical justification for it not to do so.

Despite past findings that suggested that response latency was an appropriate measure of response confidence (cf. Aaker et al. 1980; Tyebjee 1979), no evidence was found supporting the validity of response latency as a measure of attribution confidence in the present study. Perhaps other factors such as student familiarity with the computer, student keyboard proficiency, the novelty of the study procedure, or some other facet specific to the present study procedure contaminated the response latency measure. Although the response latency measure employed in the present study was similar to measures employed by Aaker et al. (1980) and Tyebjee (1979), it is possible that the construct "attribution confidence" differs in some respect from types of response confidence used in these studies, and therefore requires two different measurement procedures.

In any event, given that a) the concept of attribution confidence is specifically incorporated into attribution theory (cf. Jones and Davis 1965; Kelley 1973), b) people make attributions with varying degrees of confidence, c) attribution extremity does not appear to totally capture attribution confidence, and d) there is some support that attribution confidence moderates the relationship between attribution extremity and attribution outcomes (in the manner suggested by Mizerski et al. (1979)), it seems appropriate for researchers to assess attribution confidence in future attribution studies which investigate the impact of attribution extremity on attribution outcomes. Fortunately, the results of the present study suggest that attribution confidence can be measured with considerable ease with two "maximally dissimilar" measures. The self-report confidence measure can be assessed by adding one scale after a series of attribution extremity scales. The causal complexity measure, being a derivative measure of the probable-improbable rating scales, can be employed with DO additional measures on a questionnaire, and also without respondent knowledge that confidence is being assessed. Therefore, it seems that confidence measures offer some explanatory power at very little cost to the researcher, indicating that confidence should be assessed in studies concerning attribution relationships.

REFERENCES

Aaker, D. A., R. P. Bagozzi, J. M. Carman, and J. M. Maclachan (1980), "On Using Response Latency To Measure Preference," Journal of Marketing Research, 17 (May), 237-244.

Bem, D. J. (1972), "Self-Perception Theory," in Advances in Experimental Social Psychology, ed, L. Berkowitz, New York: Academic Press, Inc.

Calder, B. J. and R. E. Burnkrant (1977), "Interpersonal Influence on Consumer Behavior An Attribution Theory Approach," Journal of Consumer Research, 4 (June), 29-38.

Elig, T. W. and I. H. Frieze (1979), "Measuring Causal Attributions for Success and Failure," Journal of Personality and Social Psychology, 37, 621-634.

Jones, E. E. and K. E. Davis (1965), "From Acts to Dispositions: the Attribution Process in Person Perception," in Advances in Experimental Social Psychology, Vol 2, ed. L. Berkowitz, New York: Academic Press, Inc.

Kelley, H. H. (1973), "The Process of Causal Attribution," American Psychologist, 28, 107-128.

Lichtenstein, D. R. and W. O. Bearden (1986), "Measurement and Structure of Kelley's Covariance Theory," Journal of Consumer Research, 13 (September), 290-296.

Mizerski, R. W. (1975), "A Test of the Relationship Between Trait and Causal Attribution," in Advances in Consumer Behavior, Vol 2, M. J. Schlinger, ed., Chicago: Association for Consumer Research, 471-480.

Mizerski, R. W. (1978), "Causal Complexity: A Measure of Consumer Causal Attribution," Journal of Marketing Research, 15, 220-228.

Mizerski, R. W., L. L. Golden, and J. B. Kernan (1979), "The Attribution Process in Consumer Decision Making," Journal of Consumer Research, 6, 123-140.

Sawyer, A. G. and J. P. Peter (1983), "The significance of Statistical Tests in Marketing Research," Journal of Marketing Research, 20, 122-133.

Sharma, S., R. M. Durand, and O. Gur-Arie (1981), "Identification and Analysis of Moderator Variables," Journal of Marketing Research, 18, 291-300.

Sparkman, Richard M. and William B. Locander (1980), "Attribution Theory and Advertising Effectiveness," Journal of Consumer Research, 7, 219-224.

Sujan, H. (1986), "Smarter Versus Harder An Exploratory Attributional Analysis of Salespeople's Motivation," Journal of Marketing Research, 23, 41-49.

Tyebjee, T. T. (1979), "Response Time, Conflict, and Involvement in Brand Choice," Journal of Consumer Research, 6, 295-304.

Weiner, B. (1985), "An Attributional Theory of Achievement Motivation and Emotion," Psychological Review, 92. 548-573.

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