Perceived Risk: a Measurement Methodology and Preliminary Findings


James R. Bettman (1972) ,"Perceived Risk: a Measurement Methodology and Preliminary Findings", in SV - Proceedings of the Third Annual Conference of the Association for Consumer Research, eds. M. Venkatesan, Chicago, IL : Association for Consumer Research, Pages: 394-403.

Proceedings of the Third Annual Conference of the Association for Consumer Research, 1972      Pages 394-403


James R. Bettman, University of California, Los Angeles

Since Bauer (1960) introduced the concept of perceived risk, many research studies have explored the ramifications of this notion. However, research in this area has been hampered by the lack of a suitable measurement of perceived risk itself. The measures proposed to date have been arbitrary for the most part, and have often differed from one study to the next. This study proposes a measurement methodology for perceived risk and reports some initial research using this measure.

An initial measurement for perceived risk was proposed by Cunningham (1967), utilizing ordinal certainty and danger scales. Cunningham then arbitrarily combined these scales multiplicatively and developed risk categories. This final ordinal measure has been used by several other studies. Bettman (1969) proposed an ordinal measure of perceived risk based on paired comparisons. Both of the above measures suffer from having only ordinal properties; an interval-scaled measure would be more desirable, as it would allow a wider range of analytical tools to be employed in examining research results. Two different rating scale measures of risk that are assumed to be interval were proposed by Spence, Engel, and Blackwell (1970) and by Perry and Hamm (1969). Finally, some studies utilizing perceived risk notions have used very special purpose and arbitrary notions rather than attempting to develop a general measure of risk (For example, Cox and Rich, 1964 and Sheth and Venkatesan, 1968).

In view of this variety of measures and sometimes questionable methodology, a theoretical perspective for measurement is needed. First, a distinction should be made between two different types of risk, which have been confused in previous studies. Inherent risk is the latent risk a product class holds for a consumer, the innate degree of conflict the product class arouses in the consumer. Handled risk is the amount of conflict a product class engenders when the buyer chooses a brand from that product class in his usual buying situation. Thus handled risk includes the effects of information and risk reduction processes as they have acted on inherent risk. Cunningham's (1967) measure seems to deal with inherent risk. However, Cox and Rich (1964) and Spence, Engel, and Blackwell (1970) seem to be dealing with handled risk. The purpose of making the distinction between the two constructs is to allow for greater precision in measurement.


The measures developed for inherent and handled risk are based on the notion of distinguishing the concepts by means of the choice situation utilized for measurement. In both cases an extended paired comparison method was used.

In each case, a brief description of the concept of risk, outlining the ideas of uncertainty, consequences, economic, and social risks, was given. Then, for inherent risk the S was asked, for pairs of product types, to rate which product type of the pair would be more risky to shop for in an imaginary store where all packages were labeled only with a product type and size, no brand labels or prices. Ss were told that this store had a full selection of all the usual brands for each product type; they were just to assume they could not tell which brand was which. After choosing which product type was more risky, Ss then rated how much more risky that type was on a scale ranging from 0 (equally risky) to 9 (very much more risky). Handled risk was measured in essentially the same way, but by evoking a different choice situation. The S was asked to rate which product type in a pair was more risky to shop for in her own usual grocery store. Hence the effect of brand knowledge and information is included.



These measures were obtained from a convenience sample of 97 housewives, using nine product types: paper towels, dry spaghetti, furniture polish, toothpaste, beer, instant coffee, aspirin, margarine, and fabric softener. Scale values for inherent and handled risk for each of the nine product types were developed from the paired comparison data for each subject. The scaling method is quite simple. For the 36 product type pairs (9 8/2) rated as outlined above, the rating for each ordered pair (i,j) is denoted by dij. The sign convention used is d >0 if i is more risky than j, and dij<0 if j is more risky than i for the ordered pair (i,j). By convention dii = 0. Since only half of the off diagonal pairs are rated, set dji = -di.. Then scale values are simply obtained by setting the scale value Si for stimulus i to Si = EJ d ./9. Note that by using this procedure the Si sum to zero. The mean has mean removed. Hence we have relative measures for inherent and handled risk. The measures can range from -8 to +8 because of the rating scale used.

In addition to the proposed measures of inherent and handled risk, a modified version of Cunningham's (1967) certainty and danger questions was employed. Instead of using Cunningham's scale, a ten point rating scale was used for each component. For the certainty component, the scaling was then reversed from Cunningham's convention, so that in this study the scale ranged from 1 (very certain an untried brand will work as well as my present brand) to 10 (almost never certain it would work as well). Hence it is an uncertainty measure in this format. For the danger component, the convention utilized by Cunningham was used: 1 (no danger in trying a brand you never tried before) to 10 (a great deal of danger).


Several sets of analyses were performed utilizing these measures. The methodology used for each analysis will be discussed in the appropriate section.

Perceived Risk over the Population of Subjects

Properties of the distributions of inherent and handled risk for each product type were computed, and are presented in Table 1. Since there are only 97 Ss, it is premature to authoritatively discuss these distributions. However, bearing this caveat in mind, note that toothpaste and margarine had the highest ratings for relative inherent risk, and beer and instant coffee had the highest ratings for handled risk. Aspirin and beer seem to have the highest variation in the risk scale values; this result seems intuitively plausible, since housewives presumably do not know much about brands of beer and perceptions of differences among brands of aspirin vary widely. Finally, note that there is a tendency for the curves to be skewed to the left. Typical histograms for handled risk are shown in Figure 1, and for inherent risk in Figure 2.

To examine the properties of the risk scales one step further, a cluster analysis using a modified K-Means algorithm was performed on the inherent risk scale values (McRae, 1970). Euclidian distance was used in the analysis. This analysis was also performed for handled risk, but the results were similar and are not presented here. The cluster centroids for the six inherent risk clusters developed are given in Table 2. The clusters are quite interesting. For example, cluster 5 seems to associate risk with household chore products, whereas cluster 3 associates high risk with the socially oriented types of beer and instant coffee. More research needs to be done here to see if cluster membership can be predicted on the basis of S characteristics.





The Relationship between Inherent and Handled Risk

The second set of findings examines the relationship between the inherent and handled risk ratings. One hypothesis is that the inherent risk scales are more extreme than the handled risk scales, that they depart more from zero. This seems plausible on intuitive grounds; the less the information used in developing the rating, the more pronounced the tendencies to differentiate product types should be.

To test this hypothesis, the sums of the squared scale values for both handled and inherent risk were developed for each S. The hypothesis is then that the sum of squared values for inherent risk is greater than the sum for handled risk. By making several assumptions, it would be possible to show that these sums of squares approximate chi-squared variables, use the additive property of chi-square over Ss and form an F-ratio to test the equality of the sums of squares for inherent and handled risk. However, it was felt that the assumptions necessary were not justified. Hence, a more conservative alternative was taken. The sign test was used to test the hypothesis. If it rejects the null hypothesis that the sum of squares for inherent risk is equally likely to be either greater or less than the sum of squares for handled risk, we can be fairly confident in the result, since its power-efficiency is low. For the 97 Ss, 65 had higher sums of squares for inherent risk, 31 for handled risk, and there was one tie. The one-tailed probability of this result under the null hypothesis is .0005. Thus, we can reject the null hypothesis.

If the scale values for inherent risk are more extreme than those for handled risk, there are several ways this could occur. One possibility is that inherent risk values tend to expand uniformly away from zero compared to handled risk values. That is, negative and positive values of handled risk are associated with more negative and more positive values of inherent risk respectively. The ordering of product types on risk would stay the same, but the scales would expand outward from zero for inherent risk. One test of this hypothesis is to perform a simple regression analysis using inherent risk as the dependent variable and handled risk as the independent variable. If the regression weight were greater than one with an intercept of zero, this would support the hypothesis above. Two sets of nine product type regressions were run, one set for all 97 Ss and one set for just the 65 Ss for whom the sums of squared scale values were larger for inherent risk. In both cases, the fits of the regressions were good, with R2 ranging from .177 to .662 for the 97 S runs and from .149 to .623 for the 65 S runs. However, in only one case, aspirin for the 65 S cases, was the regression weight greater than one. Hence, it seems clear that uniform expansion is not responsible for the more extreme values for inherent risk. As can be seen looking at the averages in Table 1, there is some switching of product type order occurring.

Relationships with Cunningham's Components

As noted above, Cunningham's certainty and danger measures were transformed to ten point rating scales for uncertainty and danger. After this transformation, both scales should be directly proportional to risk. Cunningham used an arbitrary multiplicative model for risk. However, parameters can be estimated for both additive and multiplicative models relating the scales derived from Cunningham's work to the inherent risk measure proposed in this study in an attempt to derive less arbitrary weightings for the two components. Inherent risk is used since it seems to be most related to the measures Cunningham employed. Thus, if Y represents inherent risk, X1 uncertainty, and X2 danger, the two models examined are the simple linear model:



(1) Y = ao + a1 X1 + a2 X2

and the multiplicative model

(2) Y = bo X1b1 X2b2

In both models the theoretically correct sign for the X1 and X2 parameters is positive. The parameters in both models can be estimated simply by least squares regression if logarithms are taken in (2). Before doing this transformation, however, since inherent risk can be negative it must be transformed. Therefore, the dependent variable in (2) was inherent risk +9, yielding a minimum possible value of +1 (inherent risk can be -8) and hence a minimum logarithm of zero. There can be problems with such a transformation. Results do not stay invariant over such transformations, as pointed out by Goldberg (1971). Hence, the results of this model must be interpreted with caution.

The models were run for all nine product classes and for the data pooled over product classes using dummy variables for the product classes. The results of these runs (coefficients for the dummy variables in the pooled analysis are not shown) are given in Table 3.

Note that the linear models seem to fit slightly better than the multiplicative models. This may be because we have removed the mean and are dealing with only relative risk measures. The parameters have the right sign in all cases where they are significantly different from zero. Finally, note that danger seems to be the more important component in these analyses.

Two additional comparisons with Cunningham's findings can be made. Cunningham used the product classes headache remedies, fabric softener, and dry spaghetti in his study. He found 23.7%, 6.1%, and 1.2% high risk perceivers (by his definition of high risk) for his sample. By using Cunningham's definitions and his ten point certainty and danger measures from this study, 29.9%, 18.6%, and 12.4% of the Ss were classified as high risk perceivers. Thus, Ss in the current study seem to perceive a good deal more risk, by Cunningham's definitions. Finally, Cunningham claims that the certainty and danger components seem relatively independent (a maximum correlation of .22 for his study). However, the correlations between uncertainty and danger for this study range from a low of .294 to a high of .698 across product types, with a correlation of .519 for the data collapsed across product types. Although the correlations for this study will be higher because ten scale points were used as opposed to only four for Cunningham's study, we may conclude that uncertainty and danger are definitely not independent for this study. It is not clear how these differences in findings should be reconciled.


The above findings support to some degree the usefulness of the proposed measures. There are problems, to be sure. For product types with widely disparate risk values (e.g.,*automobiles and grocery products), it is not clear how to compare the results of different analyses, since we have only relative risk scores. However, the true test of the measure must be based on its usefulness in research or decision making. Studies of the components of risk are in progress using the measures of this study as dependent variables. Other research is needed. In particular, as Spence, Engel, and Blackwell (1970) point out, much more work is needed on the problems of developing criterion measures for risk.




Bauer, R. Consumer Behavior as Risk Taking. In Hancock, R. (ed.), Proceedings of the 43rd American Marketing Association. 1960. 389-398.

Bettman, J. Behavioral Simulation Models in Marketing Systems. Unpublished doctoral dissertation, Yale University, 1969.

Cox, D. & Rich, S. Perceived Risk and Consumer Decision Making - The Case of Telephone Shopping. Journal of Marketing Research. November 1964, 32-39.

Cunningham, S. The Major Dimensions of Perceived Risk. In Cox, D. (ed.), Risk Taking and Information Handling in Consumer Behavior. Boston: Harvard Business School, 1967, 82-108.

Goldberg, L. Five Models of Clinical Judgment: An Empirical Comparison Between Linear and Nonlinear Representations of the Human Inference Process, Organizational Behavior and Human Performance, 1971, 458-79.

McRae, D. MIKCA: A FORTRAN IV Iterative K-Means Cluster Analysis Program. Mimeo. September 1970.

Perry, M. & Hamm, B. Canonical Analysis of Relations between Socioeconomic Risk and Personal Influence in Purchase Decisions, Journal of Marketing Research. August 1969, 351-354.

Sheth, J. & Venkatesan, M. Risk-Reduction Processes in Repetitive Consumer Behavior, Journal of Marketing Research. August 1968, 307-310.

Spence, H., Engel, J., & Blackwell, R. Perceived Risk in Mail-Order and Retail Store Buying. Journal of Marketing Research. August 1970, 364-369.



James R. Bettman, University of California, Los Angeles


SV - Proceedings of the Third Annual Conference of the Association for Consumer Research | 1972

Share Proceeding

Featured papers

See More


Penny for Your Preferences: Leveraging Self-Expression to Increase Prosocial Giving

Jacqueline R. Rifkin, Duke University, USA
Katherine Crain, Duke University, USA
Jonah Berger, University of Pennsylvania, USA

Read More


How Awe Might Be Awesome: The Role of Awe in Consumers’ Food Consumption and Perceptions of Misshapen Produce

Begum Oz, University of Massachusetts, USA
Elizabeth Miller, University of Massachusetts, USA

Read More


Associative versus Relational Processing: The Role of Elaboration in Evaluative Conditioning

Xiaomeng Fan, Northwestern University, USA
Galen V. Bodenhausen, Northwestern University, USA

Read More

Engage with Us

Becoming an Association for Consumer Research member is simple. Membership in ACR is relatively inexpensive, but brings significant benefits to its members.