Risk Acceptability in Segments With Distinct Value Orientations

Gordon G. Bechtel, University of Florida
Jaime Ribera, University of Florida
ABSTRACT - The acceptability of risk is measured for various products, services and activities. These measures are shown to have different multi-attribute structures in different population segments. The idiosyncratic patterns are revealed through the application of Zellner's "seemingly unrelated regressions" to dependent and independent variables measured in a log-odds metric. The feasibility of extending this approach to national probability samples is emphasized.
[ to cite ]:
Gordon G. Bechtel and Jaime Ribera (1983) ,"Risk Acceptability in Segments With Distinct Value Orientations", in NA - Advances in Consumer Research Volume 10, eds. Richard P. Bagozzi and Alice M. Tybout, Ann Abor, MI : Association for Consumer Research, Pages: 590-595.

Advances in Consumer Research Volume 10, 1983      Pages 590-595

RISK ACCEPTABILITY IN SEGMENTS WITH DISTINCT VALUE ORIENTATIONS

Gordon G. Bechtel, University of Florida

Jaime Ribera, University of Florida

[The data analyses in the present paper were carried out with the support of the Center for Econometrics and Decision Sciences, University of Florida.]

ABSTRACT -

The acceptability of risk is measured for various products, services and activities. These measures are shown to have different multi-attribute structures in different population segments. The idiosyncratic patterns are revealed through the application of Zellner's "seemingly unrelated regressions" to dependent and independent variables measured in a log-odds metric. The feasibility of extending this approach to national probability samples is emphasized.

INTRODUCTION

Growing international concern about societal risks has created a need for assessing the public perception and acceptance of risk. This need has been answered by various research groups in the United States and Europe (Slovic, Fischhoff, and Lichtenstein 1980, Green 1980, Vlek 1980), who have been motivated by the following considerations:

"For people, and the institutions that serve them, the question 'How safe is safe enough?' appears likely to be one of the major policy issues of the 1980's. The daily discovery of new hazards and their widespread publicity is causing more and more individuals to see themselves as victims of technology, rather than as its beneficiaries. Growing fears and opposition to technology have puzzled and frustrated industry promoters and policy makers, many of whom believe that pursuit of a 'zero risk' society threatens the nation's political and economic stability" (Slovic, Fischhoff, and Lichtenstein 19815.

This new area of research is at the crossroads of marketing, consumer behavior, and environmental psychology, since the majority of activities and substances of concern represent goods and services connected with the marketplace. While earlier risk research has concentrated upon identifying the components of perceived risk (Starr 1969, Bettman 1973, Ross 1975, Lowrance 1976, Rowe 1977), recent efforts are being directed toward explaining risk acceptability in terms of these components (Slovic, Fischhoff, and Lichtenstein 1980). The present paper pursues this latter effort one step further by exploring inter-group variations in the functional form linking risk acceptability to the components of perceived risk. That is, we will assess the differential predictive power of perceptual risk components in accounting for the acceptability of risk in different population segments. Segmentation is a particularly important strategy in this area of public inquiry, where many suspect the existence of idiosyncratic functional forms in any pluralistic society (Green and Brown 1980). It is hoped that our method for estimating and testing these forms will ultimately be useful in predicting the segmented purchase and participatory behavior of consumers vis-a-vis products, services, and activities. The present approach may also be useful to policy makers attempting to understand differential acceptability levels which different groups have for the risk in a given action.

THE DATA

The data for addressing the present problem have been made available by Paul Slovic, who has kindly given permission for the secondary analyses described below. [The authors would also like to thank Mark Layman at Decision Research for his competent workup and transmission of these data.] These data have been chosen because a) they contain information on risk perception and acceptability for the same subjects, and b) they are segmented a priori into subgroups of subjects with distinct value orientations (see below). These features permit us to observe differential functions relating risk acceptability to risk attributes for distinct population segments.

The data were collected from three different groups of people who rated each of-thirty products, services, and activities on the acceptability of its current level of risk (the dependent variable). These subjects also rated the thirty hazards on nine "semantic differential" scales defining various risk attributes (the independent variables).

Stimuli

Eight of the hazards examined here were investigated by Starr (1969) in his pioneering study of societal risk. The other twenty-two were selected to vary broadly in both the type and extent of their risks. These stimuli are listed in Table 1.

TABLE 1

THE THIRTY HAZARDS

Respondents

The first group of seventy-six subjects were members or the Eugene, Oregon League or Women Voters and their spouses (52 women and 24 men). The second sample consisted of sixty-nine college students at the University of Oregon (33 women and 36 men). The final group of forty-seven subjects were members of the Eugene Active Club, which consists of business and professional people involved in community service activities (3 women and 44 men).

These groups were convenience (rather than probability) samples from their respective organizations, and,therefore, the results below are not representative of the views of the organizations involved. However, due to the industrial-academic composition of this particular SMSA, we presume a "liberal" orientation on the part of these League-of-Women-Voter and student respondents. Similarity, we assume the "conservative" orientation of the business professionals who participated in the study. These presumptions have been partially validated by the voting behavior of two of the samples. On a referendum in Oregon, which occurred after the data collection, 95% of the Eugene League-of-Women-Voter sample indicated voting against nuclear power. In contrast, 85% of the Eugene Active Club sample, upon revisitation, reported voting in favor of nuclear power. (Slovic, Fischhoff, and Lichtenstein 1981).

Measures in Log-Odds Metric

Risk Acceptability. For each hazard the subject was asked "to consider the risk of dying (across all U.S. society as a whole) as a consequence of this activity or technology." The subject subsequently selected one of the following response options: a) "Could be riskier: It would be acceptable if it were ___ times riskier;" b) "It is presently acceptable;" and c) "Too risky: To be acceptable, it would have to be ___ times safer." In the present analysis the ratio estimates in (a) and (c) are discarded, ant, since category (a) attracted very few responses, the data are dichotomized into two categories of acceptability. That is, (a) and (b) are regarded as "acceptable" responses, while (c) is "unacceptable". This procedure was invoked to make the dependent variable commensurate with the independent variables in the subsequent analyses.

In each group the proportion of "acceptable" responses was recorded for each hazard and then transformed to its logit, which is the log odds of an "acceptable" response. Since the logit of .5 = 0, if more than half of the group regard a hazard as "acceptable", our measure takes a positive value. If less than half of the group regard this hazard as "acceptable", it takes a negative value (Green, Carmone, and Wachspress 1977). These logits are the dependent variables in the analyses described below.

Risk Attributes. The nine attributes in this analysis have been postulated by several writers as components of perceived risk (Starr 1969, Lowrance 1976, Rowe 1977, pp. 31-67). The subjects rat d each hazard on these nine "semantic differential" scales, which are given in Table 2. Although each was a 7-point scale in the original study, the present analysis utilizes just three response categories, i.e., the middle category was treated as "in between", with all other responses counted into the outside categories labeled in Table 2. Thus, a group's ratings for a given hazard on a given attribute provided three proportions (of subjects) falling into three response categories.

These proportions were processed by a logit scaling method given by Bechtel and Wiley (1982). In the case of the first attribute in Table 2, this procedure postulates a middle interval representing the in-between category on the "voluntary" continuum. This category is thus defined by a lower boundary to the left of in between and an upper boundary to the right of this middle category. Utilizing the category proportions described in the preceding paragraph, the scaling method gives generalized least squares (GLS) estimates of these boundaries on the "voluntary" attribute, along with a GLS estimate of each hazard's value on this attribute. Moreover, this latter scale value is interpretable as the log odds of the hazard being rated to the right of the midpoint of the in-between interval, i.e., to the right of the midpoint of the "voluntary" scale (Bechtel and Wiley 1982).

A separate application of this procedure was made to the proportions for each attribute in each segment. For example, in the League-of-Women-Voters group we were able to estimate the log odds of a "voluntary" response, that of an "effect delayed" response, etc. to each of the thirty hazards in Table 1. Since a hazard's nine log-odds values are comparable over attributes, and are also comparable to its log odds "acceptability" (defined above), the present analysis converts our dependent and independent variables to a common metric. This advantage, of course, can be maintained in any survey in which the dependent and independent variables are observed on two- or three-category scales.

METHOD

The aggregation of respondents within each segment, along with the application of our scaling procedure to the resulting proportions, produced three sets of group scales for subsequent analysis. Each set consisted of a segment-specific acceptability scale of thirty hazards (the dependent variable) and nine segment specific attribute scales for these hazards (the independent variables). In each segment the dependent variable was regressed upon the nine independent variables across the thirty hazards in Table ]. That is, each regression consisted of thirty "observations" or "cases" -- one for each hazard.

In parallel regressions of this nature it is likely that corresponding values of the dependent variable (and their errors) will be correlated. For example, the three acceptability values of nuclear power (for the League of Women Voters, the students, and the business professionals) are likely to covary over our three segmental regressions. When this is the case, a gain in efficiency of the estimated coefficients is obtained by applying Zellner's method of "Seemingly Unrelated Regressions". [The authors would like to thank Jim Wiley and Chezy Ofir for making them aware of this technique.] This method embraces our three regressions in one analysis, and its gain in efficiency, over three separate ordinary-least-squares runs, increases directly with the magnitudes of the inter-segment correlations (Zellner 1962, p. 354, Johnston 1972, p. 241). A further advantage of this method lies in its capability of testing hypotheses about regression coefficients across models. For example, in the present analysis we may test League-of-Women-Voter coefficients against those estimated for the business professionals. The power of these tests is enhanced considerably by the reduction in the coefficients' sampling errors achieved by the Zellner procedure.

The "seemingly unrelated regressions" below were obtained from the SYSREG procedure in the SAS computer package (Statistical Analysis System 1979). The appropriateness of this method was supported by the sizable correlations it revealed among the residual errors for the different segments. These correlations ranged between .48 and .86 in the present analyses.

RESULTS

Attribute Selection

First, an exploratory SYSREG regression, including all nine attributes in Table 2, was carried out. This run consistently selected the common and effect delayed attributes in each of the three groups, with the other seven characteristics falling short of a statistically significant relationship with acceptability. Hence, only common and delay were retained in the subsequent analyses. These two attributes are very highly loaded upon separate factors which Slovic, Fischhoff, and Lichtenstein (1981) discovered among the set of characteristics in Table 2. These authors labeled their factors as "dread risk" and "unknown risk," which load common and delay respectively. Their two - factor solution, which accounted for a very high proportion of the variance in each of the nine attributes, was extremely similar across the three segments studied here. It should also be noted that the construct of immediate vs. delayed effect has been elicited in Repertory Grid studies of hazards (Green and Brown 1980).

TABLE 2

RISK ATTRIBUTES IN THREE-CATEGORY FORM

The Two-Dimensional Attribute Structures

Figures 1, 2, and 3 present the hazard configurations in common-delay space for the League of Women Voters, the students, and the business professionals. The coordinates of each hazard are the attribute measures in the log-odds metric described above.

Correlations ranging between .92 and .97 were observed between the coordinates of the three pairs of groups on the two separate dimensions. Although these correlations indicate the similarity among groups, an inspection of Figures 1, 2, and 3 reveals that the three perceptual structures are not congruent. For the League of Women Voters (Figure 1) and the students (Figure 2) the common dimension is bounded by nuclear power and home appliances, whose risks appear to display low and high degrees of societal assimilation. These contrasting assimilations also characterize the business professionals (Figure 3) who, however, regard power mowers as the most innocuous item in Table 1.

FIGURE 1

HAZARD CONFIGURATION FOR THE EUGENE LEAGUE OF WOMEN VOTERS

FIGURE 2

HAZARD CONFIGURATION FOR THE UNIVERSITY OF OREGON STUDENTS

FIGURE 3

HAZARD CONFIGURATION FOR THE EUGENE BUSINESS PROFESSIONALS

Concerning the vertical dimension, all three segments regard the cluster < contraceptives, food coloring, food preservatives, pesticides, smoking, spray cans, X-rays > as a set of hazards producing delayed effects. On the other hand, while the effects of handguns, hunting, and motorcycles are viewed as immediate in each group, the order of these latter three hazards is reversed for the students and business professionals and virtually tied for the League of Women Voters.

Attribute-Acceptability Relationships

To this point our analysis has selected two components of perceived risk for further investigation as determinants of risk acceptability. However, unlike the situation in confirmatory research, these relatively new components are not embedded in any theoretical framework solid enough to permit apriori hypothesis generation. Hence, our explanatory probe is more safely conducted by continuing the model selection approach into the final phase of analysis. This approach permits the discovery of particular functional forms, e.g., linear, optimal, etc., linking the common and delay attributes to acceptability. Given forms, uncovered in this way, are then available as hypotheses for subsequent testing (cf. Tukey 1977).

This final stage of analysis was carried out by running several polynomial regressions by SYSREG. This procedure permits the observation of differential linear slopes and quadratic curvatives for the different segments. These runs pointed to the solution given in Table 3, which has an R2 of .56 for the system. The gain in efficiency for this solution is indicated by its 332 reduction, on average, in the sampling variances of the coefficients in Table 3, when compared with those for three separate regressions.

In Table 3 common and delay are linearly related to acceptability for the League of Women Voters and for the students. It will also be noted that these two monotone relations are in the opposite direction, i.e., the more common the risk the more acceptable, and the more delayed the effect the less acceptable. Although the pattern of relationships is similar for these two segments, separate hypothesis tests across groups indicate that their common and delay slopes differ significantly. Nevertheless, their resemblance is reflected by the similar orientations of the arrows in Figures 1 and 2. The arrow in Figure 1 connects the origin to the point (.60, -.19), which contains the League-of-Women-Voter slopes, while that in figure 2 passes through the corresponding point (.75, -.30) for the students. These arrows highlight the purely monotone character of acceptability in each of these segments. That is, each arrow defines a direction of increasing acceptability, and the (perpendicular) projections of the hazards on this arrow are estimates of the 30 acceptability measures for its segment (cf. Carroll 1972, p. 115, Bechtel 1976, pp. 3438).

TABLE 3

THE THREE "SEEMINGLY UNRELATED REGRESSIONS"

Turning to the business professionals, Table 3 reveals a different picture. Here common again enters linearly, but cross-motels tests show its slope to be significantly less than that in either of the other two segments. Moreover, delay now enters quadratically, suggesting the existence of an optimal point of acceptability on this second attribute. The form of the regression for this segment indicates that this optimal point is at the origin of the delay scale in Figure 3. This relation between delay and acceptability is the only nonmonotone relation in the present analysis. Thus, unlike the pure monotone models illustrated in Figures 1 and 2, the business professionals are described by a "mixed" monotone and nonmonotone structure (Green and Srinivasan 1978, Bechtel 1981). For this group, movement toward the right in Figure 3 increases acceptability. However, in contrast to the other segments, acceptability maximizes at the origin of the vertical axis for any fixed value on the horizontal axis, decreasing with upward or downward movements away from this optimal point.

DISCUSSION

The preceding results are worthy of further confirmatory investigation in this new area of research. First, concerning the attribute selection problem, they point to the common-dread dimension as a major determinant of risk acceptability. The definition of this attribute in Table 2 identifies it with societal assimilation of risk. A second dimension, delayed effect, also displays a significant secondary presence here. These two attributes, which load separate factors in the correlational studies of Slovic et al., should be included in future surveys addressing risk acceptability.

Second, the relational structures revealed for these three segments attest to the effectiveness of the "seemingly unrelated regressions" approach, especially when employed in conjunction with polynomial regression. As an exploratory tool, this methodology permits model selection on a group-by-groups basis, while simultaneously testing for inter-group differences. This capability is particularly useful in studying risk acceptability and perception in a value-segmented society. The particular structures revealed here, i.e., purely monotone in two segments and mixed (monotone and non-monotone) in the third, are results which should be replicated in future studies in this area.

Finally, it should be emphasized that the present systems regressions are fully applicable to segment-by-segment studies involving national probability samples. The extension to larger surveys, with their richer segmentations and representative results. has been called for by various workers in this field (Slovic, Fischhoff, and Lichtenstein 1981). In fact, the present log-odds metric for encoding perceptual and acceptability responses is probably most advantageous in large-scale surveys, where two- and three- category judgments are solicited by telephone and other methods of communication. This common metric for dependent and independent variables lends a direct interpretability to the coefficients of several "segments", or, alternatively, to those of several "years," when societal change is being monitored.

REFERENCES

Bechtel, Gordon G. (1976), Multidimensional Preference Scaling, The Hague: Mouton.

Bechtel, Gordon G. (1981), "Metric Information for Group Representations," in Multidimensional Data Representations: When and Why, ed. Ingwer Borg, Ann Arbor. HI: Mathesis Press, pp . 441-78.

Bechtel, Gordon G. and Wiley , James B. (1982), "Probabilistic Measurement of Attributes: A Logit Analysis by Generalized Least Squares," Discussion Paper 67, Center for Econometrics and Decision Sciences, University of Florida.

Bettman, James R. (1973), "Perceived Risk and its Components: A Model and Empirical Test," Journal of Marketing Research, 10, 184-90.

Carroll, J. Douglas (1972), "Individual Differences and Multidimensional Scaling," in Multidimensional Scaling: Theory and Applications in the Behavioral Sciences, Volume 1, eds. R.N. Shepard, A.K. Romney, and S. B. Nerlove New York: Seminar Press. OD. 105-55.

Green, C. H., (1980), "Revealed Preference Theory: Assumptions and Presumptions," in Society, Technology, and Risk Assessment, ed. J. Conrad, London: Academic Press, pp. 49-56.

Green, C. H., and Brown, R. A. (1980), "Through a Glass Darkly: Perceiving Perceived Risks to Health and Safety," Paper presented at Decision Research's Risk Perception Workshop, Eugene, Oregon, December, 1980

Green, Paul E., Carmone, Frank J., and Wachspress, David P. (1977), "On the Analysis of Qualitative Data in Marketing Research," Journal of Marketing Research, 14, 52-59.

Green Paul E., and Srinivasan, V. (1978), "Conjoint Analysis in Consumer Research: Issues and Outlook," Journal of Consumer Research, 5, 103-23.

Johnston, J. (1972), Econometric Methods, 2nd Edition, New York: McGraw-Hill.

Lowrance, W. W. (1976), Of AccePtable Risk, Los Altos, CA: W.W. Kaufmann.

Ross, Ivan (1975), "Perceived Risk and Consumer Behavior: A Critical Review," in Advances in Consumer Research, Vol. II, ed. Mary Jane Schlinger, Chicago: Association for Consumer Research, pp. 1-19.

Rowe, W. D. (1977), An Anatomy of Risk, New York: Wiley.

Slovic, Paul, Fischhoff, Baruch, and Lichtenstein, Sarah (1980), "Perceived Risk," in Societal Risk Assessment: How Safe is Safe Enough?, eds. R.C. Schwing and W. A. Albers. New York: Plenum Press.

Slovic, Paul, Fischhoff, Baruch, and Lichtenstein, Sarah (1981). "Characterizing Perceived Risk," in Technological Hazard Management, eds. R. W. Kates and C. Hohenemser, Cambridge, MA: Oelgeschlager, Gunn, and Hain.

Starr, C. (1969), "Social Benefit versus Technological Risk," Science, 165, 1232-38.

Statistical Analysis System (1979), SAS User's Guide, 1979 Edition, Cary, NC: SAS Institute, Inc.

Tukey, John W. (1977), Exploratory Data Analysis, Reading, MA: Addison Wesley.

Vlek, C., and Stallen, P.J. (1980), "Rational and Personal Aspects of Risk," Acta Psychologica, 45, 273-300.

Zellner, Arnold (1962), "An Efficient Method of Estimating Seemingly Unrelated Regressions and Tests for Aggregation Bias,". Journal of the American Statistical Association, 57, 348-68.

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