An Empirical Investigation of the Evaluative Aspect of Certainty/Uncertainty

ABSTRACT - Several studies in marketing and psychology have investigated models of human decision making and attitude formation under uncertainty. Implicit assumption in most of these models has been that uncertainty leads the individual to calculate expected satisfactions and the uncertainty in itself is assumed to have no affective meaning. This study shows that individuals do attach negative evaluation to uncertainty and consequently the models should incorporate this additional element.

Citation:

Olli T. Ahtola (1980) ,"An Empirical Investigation of the Evaluative Aspect of Certainty/Uncertainty", in NA - Advances in Consumer Research Volume 07, eds. Jerry C. Olson, Ann Abor, MI : Association for Consumer Research, Pages: 345-349.

Advances in Consumer Research Volume 7, 1980     Pages 345-349

AN EMPIRICAL INVESTIGATION OF THE EVALUATIVE ASPECT OF CERTAINTY/UNCERTAINTY

Olli T. Ahtola, University of Florida

ABSTRACT -

Several studies in marketing and psychology have investigated models of human decision making and attitude formation under uncertainty. Implicit assumption in most of these models has been that uncertainty leads the individual to calculate expected satisfactions and the uncertainty in itself is assumed to have no affective meaning. This study shows that individuals do attach negative evaluation to uncertainty and consequently the models should incorporate this additional element.

INTRODUCTION

One of the most popular models to predict and explain human motivation, choice, preferences, attitude, and risk taking is the so called expectancy-value model. This model proposes that the above psychological variables in some way depend upon the strength of expectancy that the act (or attitude object) will be followed by (is associated with) a given consequence or goal (attribute) and the value (utility, evaluation, affective aspect, valence) of that consequence to the individual. These models have been proposed based on such varied theoretical foundations as learning theory (Fishbein, 1967), cognitive consistency theory (Rosenberg, 1956; Zajonc, 1954), field theory (e.g., Lewin, 1938), decision theory (e.g., Edwards, 1954), and expectancy theory (e.g., Tolman, 1951). Several of these different versions of expectancy-value model have been applied to the study of consumer attitudes and preferences (for a review see Wilkie and Pessemier, 1973).

Recently Ahtola (1975) has revised and extended the ex-pectancy-value model proposed by Fishbein to predict and understand consumer preferences and attitudes. This so called Vector model asserts that a consumer forms his/ her attitudes in the manner which can be mathematically expressed by calculating expected values (expected affec-rive evaluations) along product attribute dimensions using the person's subjective expectations (probabilities) of the product having various levels of the attribute and his/her subjective evaluations of these various levels of the attribute, and then summing over the expected evaluations of the salient attributes.

The attribute levels are mutually exclusive and collectively exhaustive of the whole dimension. In other words, the model asserts that an individual's attitude toward any alternative in a choice situation is the function of (1) the strength of his/her beliefs about the alternative and (2) the evaluative aspect of those beliefs. Algebraically, this may be expressed as follows:

EQUATION    (1)

where:

"_k = an individual's attitude toward alternative k,

B_ijk = his/her strength of belief ij about k (on the dimension i), that is, the probability that k is associated with some other concept (= category) ij,

a_ij = the evaluative aspect of ij, that is, the individual's evaluation of concept (= category ij),

g(i) = number of associated concepts (categories) on dimension i,

n = number of salient dimensions.

Central to this Vector model is the notion that an attitude object has a valence (affective value) which is a function (summative form is postulated) of the valences (affective values) of its attribute levels (such as consequences and characteristics) weighted by the subjective probability of their existence in the attitude object. This model was theoretically justified based on the learning theory but can be easily visualized as a type of filtering process. That is, if the subject is sure that the attitude object possesses a given level of the attribute, the filter is totally open for that attribute level and all of the valence of that attribute level is available to form the valence of the attitude object. In this case, the filter is totally closed for the other levels of that attribute, and none of their valence contributes to the valence of the attitude object. Various levels of uncertainty are reflected in the extent to which the filter is open or closed directly proportional to the subjective probability of the possession of the attribute level.

It can be seen that the filter itself is proposed to have no valence properties (or value creating properties), i.e., an individual does not like or dislike the filter whether it is open or closed or in between, he/ she only likes or dislikes (positive or negative valence exists for) the various attribute levels.

There seems to be evidence in social psychology (Osgood, Suci, and Tannenbaum, 1957; Edwards, 1953, 1954, 1961; Kogan and Wallach, 1967) and in marketing (Bauer, 1960) which is inconsistent with the postulate of certainty/ uncertainty having no direct affective properties. More specifically, there seems to be evidence that certainty is preferred over uncertainty. It is, however, very difficult to attribute the findings in these studies conclusively to the preference for certainty because these studies have used objective measures of value (e.g., dollar amounts) and/or objective measures of probabilities (true odds of winning bets) or they ignored some salient attributes or selected the attributes rather arbitrarily. Because of these weaknesses it can always be argued that they would not have found preference for certainty if they had used subjective (person's own) measures of value and subjective (person's own) probabilities, and considered all the salient attributes (e.g., including them in the model or controlling for those salient attributes not manipulated in the experiment). However, if this argument is valid, it is a strange coincidence to find rather consistently that situations or alternatives with higher certainty are preferred over situations or alternatives with lower certainty of salient outcomes when expectancy-value models would predict equal preference. It would seem unlikely that subjective probabilities and values always tend to be biased (i.e., different from their objective counterparts) to the direction of producing higher expected values for high certainty situations. Furthermore, there is direct evidence (Osgood, Suci, and Tannenbaum, 1957) that certainty-uncertainty adjective loads significantly on the evaluative (affect) dimension of the Semantic Differential.

THE ROLE OF UNCERTAINTY

What factors would cause this negative (at least relatively) connotation of uncertainty? First of all, what are the consequences of uncertainty? One common consequence is the need for additional information in most cases. Search for additional information is psychologically (and otherwise) costly. Another consequence is the knowledge that after the decision one may find out that he/she did not choose the best alternative. This anticipation of possible post-choice regret is naturally negative (Braden and Walster, 1964). Furthermore, if certainty has usually been associated with positive outcomes and uncertainty with negative outcomes we would expect that through conditioning the certainty would acquire positive evaluative response and uncertainty a negative evaluative response. There seems to be little evidence that the above would be generally true, except perhaps in the notion of selective perception, learning, and retention, where it is often argued that positive consequences (attributes) are perceived more readily, learned more easily, and retained better in the memory than negative consequences, i.e., they are, and will be, more strongly associated with situations and alternatives than negative consequences. Thus, irrespective of actual conditioning history these selective processes bias toward positive outcomes being more readily and strongly associated with past situations and experiences with objects. Also, perhaps in our relatively secure and affluent society good things really take place with higher certainty than bad things (irrespective of the saying that the only things certain are death and taxes, which are presumably negative things to most).

Based on this diverse conceptual and empirical evidence, it is hypothesized that the valence filtering process in itself elicits an internal evaluative response which is a function of the uncertainty associated with the evaluation process. The higher the certainty the more positive is the evaluation.

Information theory suggests the following ratio scaled formula for the determinants of uncertainty (e.g., Berlyne, 1957):

EQUATION    (2)

where:

U_ik = uncertainty with respect to alternative k on dimension i.

B_ijk and g(i) are the same as in Formula (1).

This formula has the properties which uncertainty can be postulated to possess, i.e.:

1.  U_ik > 0

That is, uncertainty is a unipolar concept which cannot take negative values, i.e., it is not meaningful to talk about uncertainty being less than zero.

2.  if g(i) includes only one nonzero category, U_ik = 0

That is, if only one category on dimension i is possible for alternative k, i.e., a person is certain that alternative k has category j, then uncertainty with respect to alternative k on dimension i is zero.

3.  U_ik reaches maximum when B_i1k = B_i2k = ... = B_ig(i)k

That is, uncertainty reaches a maximum if all possible (nonzero) categories on dimension i for alternative k have equal probability.

4.  if B_i1k = B_i2k = ... = B_ig(i)k, and a probability B_ig(i)+1k>0 is added to the set, U_ik increases.

That is, if the situation in 3 exists and then an additional category becomes possible (nonzero), uncertainty will increase.

To summarize, Postulate 1 limits the range of values uncertainty can take, Postulate 2 sets the natural zero for uncertainty, and Postulate 3 and 4 state that the more of the categories on a dimension are possible for a given alternative and the closer the probability values for these categories are to each other the higher the uncertainty.

One might ask why not use the more familiar variance or standard deviation as the measure for uncertainty. The major reason is that some dimensions by their very nature have categories which can be scaled only nominally or ordinally. With the exception of dichotomous categories, variance for these dimension is undefined. Also, even in the case where categories on a dimension can be ratio or interval scaled the variance (around the mean) does not seem logically good measure of uncertainty. For example, there seems to be no reason to expect that uncertainty is less if the possible (nonzero) categories are next to each other or if they are farther apart. Of course this issue could be empirically investigated. Perhaps a more important issue is that the proposed formula for uncertainty is insensitive to the "width" of a category. It might be argued that if the categories are very broad and general as compared to narrow and specific there should be more uncertainty. The proposed model, however, is perhaps often likely to predict more uncertainty in case of narrow categories, because it is likely that if the categories are narrow, more of them will be used and have nonzero probability. This issue also might be worth an empirical investigation.

At this point we can only say that, if it is found that increased uncertainty decreases attitudes, the above uncertainty formula is a candidate for the correction factor in the attitude formula. Thus, the attitude formula might become, for example, as follows:

EQUATION    (3)

However, the first step before trying to compare different uncertainty correction factors is to investigate whether any correction to the attitude model based on the expected affective evaluation formula is needed. This is the major purpose of this study.

PROCEDURE

It was decided that attitudes toward two hypothetical new citrus flavored soft drinks would be measured.

In the first part of the study, 64 undergraduate students were asked to indicate their evaluations (a_ij's) of various levels of four (4) attributes which were found in an earlier study (Ahtola, 1973) to be the salient attributes associated with the selection of a soft drink for refreshment in a similar undergraduate population. This was done using an instrument developed earlier by Ahtola (1973) which is based on measures along the evaluative dimension of the Semantic Differential (Osgood, Suci and Tannenbaum, 1957). Seven (7) subjects had to be eliminated because of incomplete data or failure to follow instructions.

One week later the subjects' evaluations of the two hypothetical (no drinks were actually tasted) brands were measured. One treatment level (brand) included no uncertainty while the second treatment level (the other brand) had uncertainty associated with it. This uncertainty manipulation was done in such a way that the expected effective evaluation according to the Vector model (i.e., attitude prediction based on the Vector model) stays the same under both levels of the treatment. The uncertainty manipulation took place only with respect to one attribute, sweetness, while the other three attributes were held constant between the two hypothetical brands.

In order to make the experiment natural to the subjects, they were told that these drinks were presently being tested for commercialization and that a certain number of students like them had tasted one can of both of these new soft drinks and had described the characteristics of each of the two brands on prespecified scales. They were also told that both of the brands used real fruit juice and that is why the level of sweetness may vary from can to can.

In order to make the situation natural it was decided that only unimodal probability vectors along the sweetness dimension were created in the uncertainty treatment. It seems most natural that if the sweetness varies among the cans of the same brand, the variation would be close and around some most common level instead of varying unsystematically along the whole sweetness dimension. It was also thought wise to keep the most common (modal) level to be the same for both brands to increase experimental control. In order for unimodal probability vectors to create the same expected satisfaction (according to the Vector model) as the certainty condition where the certainty is with respect to the modal category, it was necessary to use only monotonic sections of the affect vector along the sweetness dimension. It was further decided that three categories (sweetness levels) were utilized in the uncertainty condition, and that the number of students who were told to have tasted these brands were made to come as close to 100 as possible (it is not natural to say that e.g. "37 1/3 students said that brand K is fairly sweet," etc.)

It was also decided that the only attribute levels associated with the uncertainty brand were "fairly sweet," "slightly sweet," and "not Sweet, not bitter" in order to achieve as much homogeneity within the uncertainty treatment as possible. Consequently, "slightly sweet" was the only category associated with the certainty brand. Thirty-four of the respondents exhibited monotonic affect vectors along these three attribute levels and were selected for the experiment. Furthermore, for all these 34 subjects "fairly sweet" was the category with the highest evaluation.

To demonstrate the above methodology, let us assume that we have an individual whose affect vector is as follows:

DIAGRAM

For this individual the uncertainty brand's sweetness was described (in terms of number of students who described each level) as follows:

DIAGRAM

The purpose of the treatment manipulation was to give each subject cognitively identical treatment in each treatment condition, while keeping the expected satisfactions the same across the two treatments for each subject. Of course, it is realized that the overt uncertainty treatments varied slightly from subject to subject. However, this kind of adjustment of overt treatments to subjects is not totally unknown in the experiments. A good example is medical experiments where the drug is administered as a certain quantity per body weight unit. Of course, it is necessary to know each subject's body weight before the actual drug quantity can be determined for a given treatment level for that subject.

The order of certainty and uncertainty brands was randomized. The brands were labeled as "Brand L" and "Brand K" and it was randomly determined when each brand was in the certainty or uncertainty condition.

After the subjects were shown how their peers had described these two hypothetical brands they were asked to indicate their personal attitudes towards these two brands. This was accomplished by asking their feelings about both brands using a typical eight adjective Semantic Differential scale which included four adjectives found to consistently and strongly load on the evaluative dimension of the Semantic Differential. These scales were "good-bad," "pleasant-unpleasant," "tasty-distasteful," and satisfying-dissatisfying. No uncertainty measures about subjects' overall attitudes were taken. The average value of these four scales was used as the criterion measure for each subject and brand. In summary, the procedure included the following steps:

1.  Subjects' evaluations (a_ij's) of various levels of four soft drink attributes were taken.

2.  Belief vector (B_ijk) for sweetness was manipulated separately for each subject so that SB_ijka_ij stayed the same for both the certainty and uncertainty treatment conditions. (Within subjects design.) No uncertainty was provided for the other three attributes.

3.  Attitudes (A_k's toward the two soft drinks were measured.

RESULTS

Because it was hypothesized that the certainty treatment will result in a higher attitude than the uncertainty treatment and because each subject was exposed to both treatments, the difference score between the two attitudes was calculated for each subject, (certainty minus uncertainty attitude) and the one-tailed t-test was conducted. The null hypothesis was that the mean of this distribution of differences is zero. The computed t-statistic was t = 2.105 which is significant beyond .025 level (d.f. = 33). Consequently the null hypothesis was rejected and it was concluded that the certainty treatment resulted in higher attitudes than the uncertainty treatment. This result is consistent with the research hypothesis.

Even though the uncertainty in the uncertainty treatment was tried to keep as constant as possible, given the constraints of the experiment, it is quite obvious that some variation was inevitable (as demonstrated in the numerical examples in the method section). This being the case, it was of interest to test the hypothesis of whether the magnitude of uncertainty is correlated with the magnitude of differences in attitudes between certainty and uncertainty conditions. The uncertainty score for each subject was calculated using the uncertainty formula given earlier in this paper (Formula 2). The correlation between the uncertainty and the signed difference in attitudes between certainty and uncertainty conditions was calculated. Correlation coefficient was r = .3239, which is significant beyond .05 level (d.f. = 32, one-tailed). This significant correlation is especially interesting when it is kept in mind that the uncertainty was kept as constant as possible, thus reducing the likelihood of finding significant correlation even if a strong linear relationship in fact exists between these two variables.

In conclusion, this experiment seems to indicate that uncertainty is negatively valued (as compared to certainty) and furthermore that the higher the uncertainty the stronger the negative evaluation.

DISCUSSION

When interpreting these results, it should be kept in mind that some very restrictive conditions were put for subjects to be included in this study to achieve reasonable experimental control. Manipulations were conducted only along the monotonic section of the affect curve. Also, the certainty category, "slightly sweet," which also was the modal category in the uncertainty condition, was evaluated positive by all the subjects. Perhaps the results would have been different if this modal category had been an attribute level which has a negative evaluation.

At this stage of the investigation, these threats to the external validity were considered secondary to the elimination of the threats to the internal validity. Even though it is reasonable to argue that the experimental manipulation caused the observed differences in the criterion measure, it can still be argued that the theoretical explanation for the results has not been proven. This, of course, is always the case. A good theory is the one which conceivably could be proven false, but, in spite of vigorous testing, has not yet been proven false.

One alternative explanation for the findings in this study is that there really is no general preference for certainty, but some people like risks and some others do not, and in our sample we happened to have a large number of risk avoiding individuals. This reasoning, however, is consistent at least with the notion advanced in this paper, i.e., that the evaluative aspect of certainty/uncertainty is relevant to the attitude formation.

If the like or dislike of uncertainty is an individual difference variable, perhaps the uncertainty formula (Formula 2) needs to be weighted by a variable which can take both positive and negative values. How to independently measure this variable is an interesting problem, which solution is not attempted here.

Another alternative explanation for the results might be that it is not really the uncertainty which caused the lower evaluation of the brand but that the variability in the attribute implied poor quality control which was the real negatively evaluated concept. In this study, however, information was provided about the other salient characteristics, i.e., flavor, carbonation, and calorie content, and the two brands did not vary in these aspects. This should imply that the sweetness variation is not a matter of quality control but was caused by natural variation in sweetness among the oranges used in the drinks. In retrospect, it could have been stated more explicitly that the variability had nothing to do with the quality control. On the other hand, if the product variability in itself is considered to constitute a poor quality control, even if intentional and natural, it must be concluded that that would be totally consistent with the theory which states that certainty is preferred over uncertainty, exactly as hypothesized in this paper. If product variability alone is sufficient to cause people to consider the quality control to be poor, then this is just the manifestation of the phenomenon that people dislike uncertainty.

Still another alternative explanation, perhaps farfetched but not impossible, is that the uncertainty manipulation caused systematic changes in the affect vectors, such that the original model would have predicted the observed results. This, however, is very difficult to test, because if we observe affect vector changes after the attitude is formed these changes may be caused by the attitude and not the other way around. If it were found that belief vector changes cause affect vector changes (or vice versa) and then they together determine the attitude, the problems of predicting attitude formation and change become exceedingly complex, because we need the theory (and quantitative model) which explains these complex causal relationships.

In this study, uncertainty was operationalized as product variability with respect to one of its attributes. There is, however, another kind of uncertainty. We may know, i.e., be certain, that a brand does not vary with respect to some attribute but we do not know or remember what the right category of that attribute is. It should be made explicit that this kind of uncertainty was not investigated in this study. It is extremely difficult to manipulate this kind of uncertainty in a controlled fashion. Furthermore, there does not seem to exist any theoretical reason why this type of uncertainty would have different effects on attitudes from the variability based uncertainty.

It would have been useful to have some kind of independent direct measure of uncertainty as a manipulation check to eliminate empirically some of these alternative explanations.

REFERENCES

Ahtola, O. T. (1973), "An Investigation of Cognitive Structure within Expectancy-Value Response Models." Unpublished Doctoral Dissertation, University of Illinois, Urbana.

Ahtola, O. T. (1975), "The Vector Model of Preferences: An Alternative to the Fishbein Model," Journal of Marketing Research, 12, 52-59.

Bauer, R. A. (1960), "Consumer Behavior as Risk Taking," Proceedings, American Marketing Association, 389-398.

Berlyne, D. E. (1957), "Uncertainty and Conflict: A Point of Contact Between Information-Theory and Behavior-Theory Concepts," The Psychological Review, 64, 329-339.

Braden, M. and Walster, E. (1964), "The Effect of Anticipated Dissonance on Pre-decision Behavior," in L. Festinger, Conflict, Decision, and Dissonance, Stanford: Stanford University Press.

Edwards, W. (1953), "Probability-Preferences in Gambling,'' American Journal of Psychology, 66, 349-364.

Edwards, W. (1954), "The Theory of Decision Making," Psychological Bulletin, 51, 380-417.

Fishbein, M. (1967), "A Behavior Theory Approach to the Relations Between Beliefs About an Object and the Attitude Toward the Object," in M. Fishbein (Ed,), Readings in Attitude Theory and Measurement, New York: Wiley.

Kogan, N. and Wallach, M. A. (1972), "Risk Taking," in J. Cohen (Ed.), Behavioral Science Foundations of Consumer Behavior, New York: The Free Press.

Lewin, K. (1938), The Conceptual Representation and the Measurement of Psychological Forces, Durham: Duke University Press.

Osgood, C. E., Suci, G. J., and Tannenbaum, P. K. (1957), The Measurement of Meaning, Urbana: University of Illinois Press.

Rosenberg, M. J. (1956), "Cognitive Structure and Attitudinal Affect," Journal of Abnormal and Social Psychology, 53, 367-372.

Tolman, E. C. (1951), "A Psychological Model," in T. Parsons and E. A. Shils (Eds.), Toward a General Theory of Action, Cambridge: Harvard University Press.

Wilkie, W. L. and Pessemier, E. A. (1973), "Issues in Marketing's Use of Multi-Attribute Attitude Models," Journal of Marketing Research, 10, 428-441.

Zajonc, R. B. (1955), Cognitive Structure and Cognitive Tuning," Unpublished Doctoral Dissertation, University of Michigan, Ann Arbor.

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