Baron, J., Hershey, J. C., and Kunreuther, H. (2000). Determinants of priority for risk reduction: the role of worry. Risk Analysis, 20, 413-428.

Determinants of priority for risk reduction: the role of worry

Jonathan Baron1
John C. Hershey
Howard Kunreuther
University of Pennsylvania

Abstract

One hundred twenty two members of the Society for Risk Analysis (experts) completed a mail questionnaire and 150 non-experts completed a similar questionnaire on the World Wide Web. The questionnaires asked about priorities for personal and government action for risk reduction, badness of the risk, number of people affected, worry, and probabilities for self and others. Individual differences in mean desire for action were largely explained in terms of worry. Worry, in turn, was largely affected by probability judgments, which were lower for experts than for nonexperts. Differences across risks in the desire for action, within each subject, were also determined largely by worry and probability. Belief in expert knowledge about the risk increased worry and the priority for risk reduction. A second study using 91 non-experts (42 interviewed and 49 on the Web) replicated the main findings for non-experts from the first study. Interviews also probed the determinants of worry, attitudes toward government vs. personal control, and protective behaviors.

Introduction

Inspired by Starr's seminal article, researchers have studied the determinants of perceived risk. The resulting "psychometric approach to risk perception" has taught us two things. First, average citizens differ from experts in the determinants of what they consider "risky." Second, citizens' judgments of risk are affected by factors other than the probability and disutility of bad events.2,3,4
Implicit in previous studies is the idea that the term "risk" itself refers to an important percept. This idea - originally found in the literature on risk vs. expected return in financial investments - finds clear expression in the work of Fischhoff et al. and more recently in the work of Holtgrave and Weber, who explicitly postulate that perceived risk is a mediating variable for decision making under uncertainty. Much of the work in this tradition has examined the factors influencing risk judgments. Usually risk judgments are predicted by two judgments: probability and disutility of the possible outcome. Sometimes one of these factors accounts for much more variance than the other, depending on the person and the kinds of items under discussion.8,9,10 Other studies find that risk judgments are affected by other factors - such as catastrophic potential, dread, novelty, naturalness, voluntariness, degree to which risk is unknown, and equity.8,11,12,13
The meaning of "risk," like that of other common-language terms, is dependent on the context. Furthermore, the desire to avoid a risk may depend on other factors aside from its judged riskiness. It is thus worthwhile to do additional studies, such as this one, that ask people about decisions, even if the choices are hypothetical ones. We asked subjects about their priority for risk reduction when they could completely eliminate each risk either from their own lives or from the lives of average citizens of their country. According to expected-utility theory, this priority should depend on probability and disutility. We asked subjects not to consider cost or the effectiveness of measures designed to reduce each risk.
In the same study, we also asked about several determinants of these priority judgments, specifically, probability, disutility (which we called "badness"), and the number of people affected if a bad event occurs. We also asked about two factors that are not considered in a simplistic application of expected-utility theory: how much is known about the risk, and how much the subject worries about it. Previous studies, beginning with Fischhoff et al., have found risk judgments to increase for "unknown risks."
Our primary interest is the role of worry. Previous researchers have proposed that worry is a determinant of action priority. People may worry about bad events for reasons other than the probability and disutility of those events. Drotz-Sjöberg and Sjöberg and Sjöberg found that worry about risks was correlated with judgments of riskiness, but they did not examine how worry affected priority.14,15 Myers et al. did find that worry was correlated with personal action to reduce risk and with desire for regulation of risk, but they did not ask whether this effect was subsumed by the effects of probability and disutility. We ask here whether worry contributes to the prediction of priority for risk reduction even when probability and disutility are taken into account.
We also revisit questions about group differences. Much of the literature on risk perception has been driven by the finding that experts and nonexperts differ in their perceptions. We examine this issue with respect to all the variables of interest: priority for action, probability, disutility, and worry. We also examine sex differences. Flynn et al. found that white women rated risks as higher than white men, on a verbal scale. We can examine the source of such differences, if we find them.
The next section describes an experiment comparing judgments of experts, defined as members of the Society for Risk Analysis living in the U.S., with non-experts. (Although members of the Society are not experts on each risk, we assume that they are better informed than the general public about risks in general.) The questionnaire for the two groups was worded identically but differed in visual appearance: it was sent by mail to the experts but was completed on the World Wide Web by non-experts. In a second experiment we compared the responses of non-experts with those who had one-on-one interviews and were asked why they responded in the way they did. In both experiments, worry played an important role in influencing people's decisions on which risks they felt they should protect against, and on which ones the government should reduce.

Experiment 1

The questionnaire in Experiment 1 presented a list of 32 different risks. Each risk was defined in terms of both an outcome and a cause, e.g., cancer from electromagnetic fields (EMFs). For each risk, the subject was asked to judge the priority of reducing it for himself or herself and for the country. The subject was also asked the lifetime probability of the bad outcome for the average American family, and for her own family, the badness (disutility) of that outcome, the number of people affected, the degree of expert knowledge about the risk in question, and how much he or she worries about the risk. We also asked about the importance of eliminating each risk completely, assuming that this could be accomplished and that cost was not an issue This was done first with regard to the subject and the subject's family, and then repeated for the American public as a whole.

Method

We mailed a questionnaire to 312 members of the Society for Risk Analysis and received 122 replies. We consider these respondents to be risk experts. The non-expert subjects (N=150) completed the questionnaire on the web. They found the study through postings in newsgroups and links from other web pages, such as that of the American Psychological Society. The two groups differed in a number of socio-economic dimensions. The experts were older (median ages 46 vs. 21, ranges 21-75 vs. 16-54), more likely to be male (80% vs. 43%), more likely to be married (87% vs. 16%) and more likely to have children (72% vs. 16%). (Age did not correlate with any of the main measures within each group.) The groups did not differ, however, on a scale of political liberalism, both being slightly more liberal than "middle of the road" (.28 and .32 on a scale of 3 to -3). All but three of the experts had advanced degrees; 53% had a Ph.D. Seventy-six percent of the non-experts were students.
After a brief introduction, subjects answered a question for each of the 32 risks shown in Table 1 and then proceeded to the next question for all these risks until the questionnaire was completed.
Table 0.1: Risks used in Experiment 1.


Bacterial infections
from contaminated food
from drinking water
Cancer
from food additives
from pesticides
from chemicals in water
from chemicals in air
from radon gas in homes (lung)
from smoking (lung)
from electromagnetic fields
from nuclear radiation
AIDS
from sex
from other causes
Other health problems
from unhealthy diet
from vaccines
from drugs or alcohol
from malnutrition
from air pollution (lung diseases)
from new diseases (such as ebola)
Injury or death
from hurricanes
from tornadoes
from earthquakes
from lightning
from floods
from violent crime
from war
from terrorism
from fire at home
from accidents at home
from accidents at work
from auto accidents
from airplane accidents
from an asteroid hitting Earth
All subjects answered the following questions:
1. How much do you worry about each of these risks, on the average, for you and your immediate family? Answer on a scale of 1 to 7, where 1 means "not at all" and 7 means "a great deal." A "great deal" means that you think about it often and that you are greatly bothered by the thought of the bad event in question.
2. Imagine there were a way to completely eliminate some of these risks to you and your family. Indicate the priority you would give to eliminating each risk to you and your family. [This is what we call "personal action." The responses were High, Medium, Low, or None, coded as 4, 3, 2, and 1, respectively.]
3. Imagine there were a way to completely eliminate some of these risks to the American public as a whole. Indicate the priority you would give to eliminating each risk to the American public as a whole. [This is what we call "government action."]
4. What is the lifetime probability for the average American family [4a], and what is the lifetime probability for you and your family [4b], for each of the outcomes listed? To answer this question, use the risk ladder on the last page, which may help you in thinking about low probabilities or small proportions. When something can happen more than once, we want the probability that it happens at least once .
[Figure 1 shows the risk ladder, which was logarthmically spaced. The web subjects saw only the text, not the vertical line.]
5. On a scale of 1 to 7, how accurate do you think that experts are in estimating the probability of this outcome for the average person? 7 means extremely accurate. 1 means extremely inaccurate, so that the true probability could be several times more or less than what any given expert thinks.
6. If this outcome happens, how many people, on the average, suffer the direct effect at the same time, from the same episode? Answer on the following scale:
1. This happens to one person at a time.
< 10. ... more than one person but usually less than 10.
< 100. ... more than 10 people but usually less than 100.
< 1000. ... more than 100 people but usually less than 1000.
More. ... more than 1000 people at once.
7. Of those who suffer this effect, how bad, on the average, is the impact for each person? Answer on a scale of 0 to 100, where 0 means not bad at all and 100 means as bad as a painful death. (You may go beyond 100 if you need to.)
Response        Probability        Example (chance of)
A 1 in 1 Certain to happen
B 1 in 3 A 80 year old dying by 85
C 1 in 10 A 65 year old dying by 70
D 1 in 30 A 51 year old dying by 56
E 1 in 100 A 35 year old dying by 40
F 1 in 300 A 20 year old dying by 23
G 1 in 1,000 A 20 year old dying in the next year
H 1 in 3,000 A 20 year old dying in the next 4 months
I 1 in 10,000 A 20 year old dying in the next 5 weeks
J 1 in 30,000 A 20 year old dying in the next 2 weeks
K 1 in 100,000 A 20 year old dying in the next 4 days
L less than 1 in 100,000 (specify the probability)


Figure 0.1: Risk ladder used for experts in Experiment 1 and interview subjects in Experiment 2.

Results

We analyze the data in two main ways, by subject, and by item within subjects. Figure 2 shows a schematic layout of the data. Each rectangular plane represents a subject. The four rows in the plane represent the 32 risks. The columns in the plan represent the various measures, the answers to our questions. We symbolize the dependent variable with y, and the predictors with a, b, c, and d. For example, the priority for personal action might be the dependent measure y, and we might predict this from worry, personal probability (expressed logarithmically), badness, and knowledge. Each variable has two subscripts. The first refers to the subject, and the second refers to the risk. At the bottom of each column are variables representing the mean response of a given subject, e.g., a1· for the first subject's answer to the first predictor.
In one type of analysis, we consider only the subject's marginal means, the values at the bottom of each rectangle. We regress these means - y·· - against the means of the predictors - a··, b··, .... But we use only these means. The individual items are relevant only because they contribute to each subject's average. This kind of analysis is relevant to questions about individual differences among subjects. Naturally, we can also use other subject characteristics as predictors in this kind of analysis, such as sex, age, and expertise.
In the other type of analysis, we carry out a regression within each subject. We predict the y's from the various predictors, and we find a regression weight for each predictor, for each subject. To test statistical significance, we treat these regression weights as our primary data and ask questions about their mean value across subjects.
protfig0.png
Figure 0.2: Schematic illustration of the structure of the data. The first subscript refers to subject, and the second refers to risk items.

Individual and group differences in determinants of mean responses

We computed each subject's mean response to each question, across the 32 risks. These means allow us to ask whether the general tendency of people to support action, to worry, and to assign high probabilities to bad events were related to sex, expertise, politics, age, marriage, and parenthood. Of primary interest are the effects of expertise and sex, but, as we noted earlier, expertise was correlated with other variables, including sex. So, we used multiple linear regression as our primary method of analysis.
Tables 2 shows the main results. Table 2A shows the means of the main measures for the two groups. Table 2B shows the regressions of government and personal action, probability, and worry on expertise, age, sex, marriage, parenthood, and political liberalism. The bottom half (B) shows regressions of government and personal action and worry on probability and the three most important demographic variables, expertise, liberalism, and sex (coded 1 for males, 0 for females). In this part of the table, we include regressions of worry on the other variables. These regressions are relevant to questions about the mediating role of worry.
We coded probabilities logarithmically, with two steps on the ladder corresponding to one unit. (Probabilities below 10-10 were set at 10-10.)
Table 0.2: Section A shows the mean responses of expert and non-expert subjects to questions about priority, probability (log scale), worry, and certainty (with ranges in parentheses). Sections B and C show the standardized regression weights for predictors of individual and group differences, Experiment 1. * indicates p < .05; **, p < .01. N » 265. Sex is coded as 1 for male, 0 for female. Probability is coded logarithmically, as described in the text.
A. Means of experts and non-experts
Question (range) Expert mean Non-expert mean
Pers. action (1-4) 2.45 2.76
Gov't. action (1-4) 2.82 3.12
Worry (1-7) 2.45 2.71
Pers. risk (log prob.) -5.53 -4.04
Avg. risk (log prob.) -4.80 -3.42
Certainty (1-7) 4.71 4.53


B. Standardized regression weights, demographic variables only.
Dependent variable:
Personal Gov't. Personal Average
Predictor: action action Prob. prob. Worry
Expert -.35** -.41** -.54** -.53** -.16
Liberal .15* .20** .06 .07 .05
Sex -.07 -.18** .05 -.01 -.03
Age .09 .05 .15 .06 .02
Parenthood -.07 -.02 .04 .03 -.03
Married .07 .19* -.16 -.06 .04
R2 .12** .18** .26** .27** .02


C. Standardized regression weights. main demographic variables, probability, and worry.
Dependent variable:
Pers. Pers. Gov't. Gov't.
Predictor: action action Worry action action Worry
Worry .42** .27**
Pers. Prob..15* .08 .17*
Avg. Prob. .28** .22** .20**
Expert -.20** -.18** -.05 -.12 -.11 -.04
Liberal .14* .13* .05 .17** .15** .05
Sex -.06 -.05 -.02 -.16** -.15** -.01
R2 .14** .30 .05* .29** .22** .05*


Experts were less supportive of personal action and government action than non-experts (Table 2A and 2B). The mean rating for personal action was lower for experts (mean, adjusted for sex, 2.48 on a scale where high priority is 4 and no priority is 1) than nonexperts (mean 2.75). Likewise, the mean rating for government action was lower in experts than nonexperts (2.85 vs.  3.11 adjusted for sex). Experts also gave much lower probabilities than nonexperts (means -5.53 vs. -4.04 for own risk, t=8.65, -4.80 vs. -3.24 for average log probability, t=9.04). One possible explanation of this difference is that experts are more willing to use the bottom of the scale (shown in Figure 1) to indicate low probabilities, rather than to a real difference in belief.
Females supported government action more than males (3.06 vs.  2.95 adjusted for expertise). Support of both personal and government action correlated positively with political liberalism. Both of these effects may be related to political attitudes more than to beliefs about probability. Sex and liberalism were unrelated to beliefs about probability.
Table 2C shows that worry is strongly related to personal and government action. Probability also affects action, but the effect of personal probability on personal action (and, to a lesser extent, on government action) is reduced when worry is included in the regression equation (second column vs. first column, and fourth vs. third). These results suggest a model in which worry mediates the effect of probability on personal or government action tendency. To test this model, we formed two new variables for each subject: probability, the mean of personal and average probability, and action, the mean of personal and government action. We regressed these two variables on sex, expertise, and liberalism and calculated the residuals. We then repeated this procedure but added mean worry as a predictor. Consistent with the hypothesis that worry mediates the correlation between action and probability, we found that the correlation between the residuals was smaller when worry was included as a predictor (.1490) than when it was not (.2091). As these are dependent correlations, we tested their difference with the method proposed by Steiger and found a z of 2.207 (p=.0137).
In sum, the correlation between probability and action is reduced significantly when the correlation of each of these variable with worry is taken into account. It appears that worry mediates the correlation between action tendency and probability. Most likely, the perception that risk is great induces worry, which, in turn, creates a positive attitude toward action. It is also possible that worry is induced directly by some other factor, such as personal experience or exposure to news reports, and that worry affects probability judgments, or that the other factors affect both worry and probability judgments.
Subjects thought, in general, that their own probabilities of bad events were lower than the average probabilities (mean -4.70 for own risks, -4.03 for average on a base-10 logarithmic scale, a ratio of more than 4; t=13.50, p=.000). This may, of course, be true, as the groups were unrepresentative of the U.S. population, but it could also result from an optimistic bias (Weinstein, 1980).

Individual differences concerning particular risks

In this section, we examine the determinants of the pattern of responses across the 32 risks. In particular, we are interested in whether experts and non-experts differ in which risks they rate high on each scale and which risks they rate low. Such differences have been reported before (e.g., Slovic, 1987), but our sample of risks includes some that are current sources of public controversy.
Experts and non-experts differed in the risks they worried about. The interaction between expertise and risk was significant (F31,7967=6.41, p=.0000). Figure 3 plots mean expert worry against mean non-expert worry. We examined group differences by looking at correlations between expertise and worry for each of the 32 risks. The four highest correlations (indicating that experts worry more) were with bacteria from food, lightning, fire at home, and accidents at home. The four lowest correlations (all negative) were AIDS from sex, AIDS from other causes, malnutrition, and cancer from chemicals in water. Some of these differences may reflect the fact that experts were older and more settled, hence less exposed to the risk of AIDS, for example. But the expertise-risk interaction was still significant when marital status and age were statistically controlled. Now, the four risks that experts worried relatively most about were radon gas, unhealthy diet, drugs or alcohol, and accidents at home. The four risks that experts worried relatively least about were cancer from pesticides, AIDS (both causes), and accidents at work. These low ratings could result either from greater knowledge or from less exposure to each risk.
protfig2.png
Figure 0.3: Mean worry ratings of experts as a function of the corresponding means for nonexperts. Each point is one of the 32 risks. Those starting with "Ca" are cancer risks. Some labels are omitted.
Figure 4 shows the mean (logged) average-family probability judgments of experts plotted against those of non-experts.
protfig3.png
Figure 0.4: Mean log probability judgments of experts (for average family) as a function of the corresponding means for nonexperts. Each point is one of the 32 risks. Those starting with "Ca" are cancer risks. Some labels are omitted.
Experts gave lower probabilities than non-experts, so, to ask whether these differences were statistically meaningful, we standardized each subject's (logged) probability estimates across the 32 risks (replacing missing data with the mean standardized score across all subjects). This expressed probabilities in comparison to other risks rated by the same subject. An analysis of variance showed a significant interaction between risk and expertise (F31,7812=10.66, p=.0000). That is, experts' standardized probabilities were higher for some risks than non-experts and lower for others. Age, sex, liberalism, and marital status did not interact significantly with risk when added to the regression. Thus, expertise seems to be the main determinant of the pattern of beliefs about probability of risk.
Experts had "significantly" higher probabilities then non-experts (p < .05 uncorrected, with correlations between log probability and expertise [coded 1 or 0] in parentheses) for bacterial infections from food (.29), infections from water (.15), unhealthy diet (.15) fires at home (.28), accidents at home (.24), accidents at work (.21), and auto accidents (.19). Experts had significantly lower probabilities for cancer from food additives (-.44), cancer from pesticides (-.40), cancer from chemicals in water (-.34), cancer from chemicals in air (-.20), cancer from electromagnetic fields (-.29), cancer from nuclear radiation (-.22), AIDS from sex (-.22), and AIDS from other causes (-.14). Note that these "probabilities" are relative to each subject's other judgments, so they reflect the relative ranking of risks relative to other risks. We found (but do not report here) a similar pattern of expert/non-expert differences in support for government action. It is possible that non-experts are more influenced by coverage in news media (Lichtenstein et al., 1978) and experts are more influenced by statistical data. News reports may concentrate disproportionately on low risks of cancer.

Individual differences in within-subject effects

We carried out correlational analyses across the 32 risks within each subject. Table 3 shows the means, across subjects, of these within-subject correlations between measures across the 32 risks. Of interest here is that worry was more highly correlated with personal action than with government action (t=15.66, p=.0000, across subjects), and more highly correlated with personal probability than with average probability (t=3.02, p=.0028). Personal action is more highly correlated with personal probability than with average probability (t=3.38, p=.0008), and government action is more highly correlated with average probability than with personal probability (t=6.89, p=.0000). These results are not unexpected: people worry about risks that affect them and want to take personal actions that reduce their own risks. The demand for action taken by the government is more likely to be triggered by the estimate of risk to the average American family.
Table 0.3: Mean within-subject correlations, across risks.
Pers. Gov't Pers. Avg.
Worry action action risk risk Cert. Scope
Pers. action .6443
Gov't. action .4555 .6187
Pers. risk .4906 .4916 .3707
Avg. risk .4474 .4462 .4608 .6688
Certainty .1363 .1326 .1767 .1587 .2404
Scope -.0633 -.0625 .0206 -.0779 -.0827 -.0462
Badness .0492 .0835 .1113 -.1311 -.1175 -.0007 .0581


We analyzed the determinants of worry, priority for government action, and priority for personal action, within each subject. To do this, we regressed each of these measures on predictor variables, such as judged probability of a bad event, within each subject, across the 32 risks. The coefficients tell us what accounts for differences in each subject's worry and priorities. What makes a subject more inclined to take action for some risks than for others? What makes her worry more about some risks than others? Note that this question is independent of the question we have already asked, which concerns the mean tendency to act or to worry, across many different risks.
We regressed personal action on personal probability, badness, certainty (expert knowledge about probability), and scope (number of people affected), with and without worry as an additional predictor. We also regressed government action on average probability, badness, certainty, and scope, again with and without worry.
Table 4 shows the mean standardized regression coefficients. (Standardized coefficients allow comparison of effect sizes.) Of interest is the fact that probabilities were strong predictors of all dependent measures, even stronger than badness ratings. Support for action was higher when the subject thought that experts knew more about the probability of the risk, a result opposite to that of Slovic. Finally, scope (number of people affected) predicted government action but not personal action; this result is, in a way, a manipulation check. If individuals are concerned mainly about themselves and their families, then scope should have little effect, but it should matter for government policy.
As in the analysis of means, worry was a strong predictor of support for action, especially personal action, and probability was a strong predictor of worry.
Table 0.4: Mean within-subject (standardized) regression weights. Parentheses indicate not significant at p < .05.
Dependent variable:
Personal Personal Gov't. Gov't.
action action action action Worry
Predictor
Worry .475 .252
Pers. Prob..248 .493 .490
Avg. Prob. .304 .416
Badness .098 .166 .131 .162 .120
Certainty .032 .058 .076 .085 .057
Scope (-.003) (-.028) .118 .101 (-.028)


We compared regression weights using standardized coefficients (shown in Table 4). To test the mediating role of worry, we compared the mean coefficients of personal probability for predicting personal action, with and without worry in the model. When worry was included, the coefficient for probability was much lower (.248 vs. .493, t239=14.09, p=.0000). Likewise, the coefficient for average probability for predicting government action was lower when worry was included (t237=9.15, p=.0000). It appears that probability beliefs affect support for action toward particular risks in part because they are related to worry about these risks. (Again, the causal direction of this relation may go in both directions. Worrying about a risk may lead people to exaggerate its probability, relative to other risks.)
The effect of worry was greater on personal action (.310) than on government action (.147, p=.0000 by t test across subjects). In a separate analysis in which we regressed worry on personal and average probability together (and the other three predictors), the coefficient for personal probability (.358) was higher than that for average probability (.143, p=.0012). Thus, the main role of worry is in mediating personal action, and the main determinant of worry is personal probability, although worry also plays some role in government action and is somewhat influenced by average probability.
Scope affected government action but not personal action, although badness affected both. The difference in the two scope coefficients was significant (t=4.26, p=.0000, in the regressions without worry). This suggests that subjects understood that scope is most relevant for government action rather than personal action.
We also examined the effect of sex, expertise, and liberalism on the regression weights. Again, because expertise and sex were confounded, we used multiple regression. Such analyses tell us about how people differ in what determines their support for action or their worry.
In personal action, men were more influenced by worry (.068 for the unstandardized coefficient of the worry coefficient for sex, p=.0354; this result does not imply that men worry more then women, so it is consistent with our finding that they do not do so). None of the remaining coefficients (in the regression with worry removed) showed any significant effects of sex, liberalism, or expertise. For government action, no predictors were significant.
The largest individual differences were found in the determinants of worry. Worry in women was more affected by personal probability than worry in men (.154, p=.0031). Worry in nonexperts was also somewhat more influenced by personal probability than worry in experts (.104, p=.044).
In sum, despite the large differences in age, expertise, and other variables between experts and subjects tested on the Web, the within-subject determinants of action and worry were much the same for both groups. Men's personal action was more affected by worry than women's, and women's worry was more affected by personal probability than men's, but these effects are post hoc. The main finding here is that, while groups differ in mean levels of action tendency and worry, they do not differ much in how they respond to different risks.

Experiment 2

Experiment 2 attempted to replicate the major results of Experiment 1 while serving two additional purposes. First, it compares two groups of non-expert subjects, one group completing the questionnaire in personal interviews and the other completing the same questionnaire on the Web. This permits a more direct comparison of these two methods.
Second, the interview subjects, after completing the questionnaire, were asked to explain the reasons for discrepancies in their answers.

Method

Forty-nine subjects completed a questionnaire on the web like that of Experiment 1. The only substantive difference was in the questions about action to reduce the risk, where there was no indication that the risk could be completely reduced. These questions were worded as follows:
If you had more money to spend, which of these risks would you spend it on? Circle the priority you would give to each risk: Hi=high, Med=medium, Lo=low, No=no money at all. [Note: We are interested in all expenditures in advance, whether they reduce the probability or magnitude.]
If the government had more money to spend, which of these risks would you like the government to spend it on? Circle the priority you would give to each risk: hi=high, med=medium, lo=low, no=no money at all.
Forty-two others completed the same questionnaire in a one-on-one interview. The interviewer answered the subjects' questions, checked the answers for completeness, and asked a series of follow-up questions. Interview subjects were limited to those living on their own (i.e., not in university residences or with their parents) because the follow-up questions concerned protective behaviors that would be relevant only for such people. We attempted to get a varied sample of people living near the University of Pennsylvania, in West Philadelphia, a community consisting of students, university staff, and a variety of ethnic groups. The web subjects were similar to those in Experiment 1. Web subjects were younger than those who had interviews (median age 19 vs. 30.5, ranges 18-47 vs. 18-81, p=.0000 by U test), had fewer children (means 0.3 and 0.8, p=.0011 by U test), more likely to be students (92% vs. 45%, p=.0000 by Fisher test), but no different in sex (29% male vs. 36%) or politics (means of 2.8 vs. 2.6 on a 5 point scale of liberalism).
Twenty-eight of the interview subjects were contacted for follow-up interviews. These interviews were directed at two issues. The first issue concerned the difference between worry and expected disutility (probability times badness). We wanted to know why people worry about risks that they perceive as not bad or not very probable, and why they fail to worry about risks that they perceive as likely and serious? We thus looked for risks that were ranked high in worry and low in expected disutility, and vice versa. Four risks for the worry vs. the expected-disutility question were selected as follows: the probability was multiplied by the badness rating for each risk and the resulting expected-disutility estimates were ranked; the worry ratings were also ranked; the ranks were subtracted, with the two highest and two lowest (most negative) differences selected for follow-up questioning (with arbitrary choices made in the case of ties).
The second issue concerned the difference between personal and government responsibility for risk reduction. The interviewer selected (arbitrarily) two of the risks with the greatest discrepancy favoring government action, and two with the greatest discrepancy favoring personal action.

Results

Group and demographic differences in determinants of mean responses

As in Experiment 1, we first computed subjects' mean response to each question across all 32 risks. We first determined whether web subjects differed from interview subjects, a question of methodological interest. In an analysis of variance of the means of all major variables (government concern, personal concern, worry, knowledge, badness, average probability, personal probability, and scope - the number of people affected), the effect of group (web vs. interview) was significant overall (F8,83=3.16, p=.0036), but the only variable that showed a significant effect was average probability, which was higher in the web group (mean logs -3.49 for interview, -3.13 for web; t=2.39, p=.0188 - the interaction between web vs. interview and average vs. personal probability was also significant: F1,90=6.12, p=.0153, so the difference is specific to average probability). The overall analysis was no longer significant when average probability was removed.
One explanation of the difference is the format of the two questionnaires. On the web, subjects answered questions about average probability first, then questions about personal probability. In the interview, the two questions were side by side, a format which may have encouraged subjects to give the same answer to both questions. Of course, this explains only the interaction, and not the fact that average probability was higher on the web. That may have to do with the format of the risk ladder, which was all text on the web but with a visual display in the interview. It is encouraging that other variables did not differ. In sum, the differences between web and interview are easily understood in terms of display factors.
We also examined other determinants of mean responses: age, sex, student status, politics, and children. A canonical correlation analysis suggested that a single dimension of individual differences accounted for differences in worry, probability, badness, scope, and certainty (p=.0174 for the entire analysis) and that worry was the main variable that captured these effects. Separate regression analyses confirmed that only worry showed significant effects. In particular, mean worry increased with number of children (t=2.84, p=.0057, for the regression coefficient) and decreased with being married (t=2.05, p=.0436). In sum, the ones who worry the least are those who are married but childless.
Separate analyses in which Black (vs. White - excluding other groups) was included also found that blacks worried more than whites (t=3.52, p=.0010). There was no difference between Asians and Whites, however. Men and women did not differ in this study.

Determinants of mean personal and government action
Separate regression analyses predicting individual differences in personal and government action from worry, badness, certainty, scope, and probability (personal probability for personal action, average probability for government action) were significant (R2=.27 and .36; p=.0000 for both). The only significant predictor was worry (t=4.37 and 5.45, p=.0000 for both), except that scope was significant one-tailed for government action (t=1.77, p=.0405). In all the regressions reported below we looked for an effect of group (interview vs. web) by comparing the standardized coefficients. In no analysis was there any significant effect. Thus, predictability of mean responses appears similar for web and interview subjects.

Determinants of mean worry
Worry itself was predicted by average probability (t=2.12, p=.0366) but not by personal probability in a regression on both. When it was regressed on average probability, badness, knowledge and scope, the overall regression was significant (R2=.10, p=.0482) and the only significant predictor was average probability (t=2.77, p=.0069).
In sum, it appears that individual differences in desire for action are determined largely by differences in worry, which, in turn, is determined largely by differences in probabilities, a finding similar to that of Experiment 1.
Determinants of mean probability judgments
As in Experiment 1, we used the logarithmic scale for probabilities, and we report the results that way. Subjects thought, as in Experiment 1, that their own risks were lower than average (mean -3.91 for own risks, -3.30 for average, a ratio of more than 4; t=9.67, p=.0000).

Individual differences in within-subject effects

Table 5 shows the mean within-subject correlations between measures across the 32 risks. Of interest here is that worry was more highly correlated with personal action than with government action (t=8.64, p=.0000, across subjects), and more highly correlated with personal probability than with average probability (t=1.89, p=.0309 one tailed). Worry thus seems more important for personal risks than for societal risks.
In contrast to Experiment 1, men and women did not differ in the correlation between personal action and worry.
Table 0.5: Mean within-subject correlations, across risks, Experiment 2.
Pers. Gov't Pers. Avg.
Worry action action risk risk Cert. Scope
Pers. action .5768
Gov't. action .3663 .4861
Pers. risk .4546 .4066 .2786
Avg. risk .4198 .4095 .3209 .6448
Certainty .1557 .1818 .1102 .1316 .2118
Scope -.0460 -.0157 .1430 -.0423 -.0165 .0646
Badness .1411 .1274 .1456 -.0408 -.0200 .0237 .1496


We analyzed the determinants of responses within each subject by regressing each dependent measure on the predictor measures within each subject and examining the regression coefficients. We regressed personal action on personal probability, badness, certainty, and scope, with and without worry as an additional predictor. We also regressed government action on average probability, badness, certainty, and scope, again with and without worry. Finally, we regressed worry on personal probability, badness, certainty, and scope.
Table 6 shows the mean standardized regression coefficients. Of interest is the fact that probabilities were strong predictors of all dependent measures, even stronger than badness ratings. Personal action tendency was higher when subjects thought that experts were more certain about the probability of the risk. Finally, scope predicted government action but not personal action (again, suggesting that subjects understood that scope is more relevant to government action).
As in the analysis of means, worry was a strong predictor of action tendency, especially personal action, and probability was a strong predictor of worry.
Table 0.6: Mean within-subject (standardized) regression weights. Parentheses indicate not significant at p < .05.
Dependent variable:
Personal Personal Gov't. Gov't.
action action action action Worry
Predictor
Worry .444 .256
Pers. Prob..134 .325 .428
Avg. Prob. .179 .300
Badness .051 .107 .068 .102 .147
Certainty .083 .139 (.008) (.030) .107
Scope (.004) (.004) .150 .125 (-.023)


Scope affected government action but not personal action, although badness affected both. The difference in the two scope coefficients was significant (t=2.82, p=.0059, in the regressions without worry). This suggests that subjects took the distinction seriously.

Follow-up interview

For expected-disutility vs. worry, the follow-up interview asked:
1. Where have you heard about this risk? Do you know people who have experienced it? Have you heard about it in the news? In movies or TV shows?
2. Do you remember hearing or reading any statistical information about this risk?
3. Why do you worry (not worry) about this risk?
Almost all subjects mentioned news or TV for every risk. We looked for a difference between worrisome and non-worrisome risks in terms of personal experience and knowledge of statistics. We might expect people to worry more when they have personal experience (or know someone who had such experience). The relation between recalling statistics and worry could go either way. People might worry more about risks that they have heard more about, or find their worries allayed by statistical information. In fact, both responses were more frequent for worrisome risks. Subjects mentioned statistics for 52% of the most worrisome risks and for only 32% of the least worrisome risks (z=2.52, p=.0116, Wilcoxon test). And they said they had knowledge of personal experience for 40% of the most worrisome risks and 31% of the least worrisome risks (z=2.06, p=.0397).
In response to the questions about reasons for worry (or not), Table 7 shows the percent of subjects giving each of several reasons for worrying or not worrying.
Table 0.7: Percent of subjects (N=28) giving each reason for worrying or not worrying about some risk.
Reason Worry Not worry
Personal control 4 32
Personal lack of control 21 25
Government control or lack 7 7
Lack of knowledge 11 0
Probability overall 29 18
Probability for self 18 64
Bad or not bad 36 11
"Social problem" 11 0
Personal experience 21 0


Personal control was mostly given as a justification for not worrying. The implication is that the person had in fact taken steps to control the risk. Only one person (4%) gave control as a reason for worry. Control could be a reason for worry, if worry functions as a way of motivating people to remember to take protective action. Others (21%) worried about risks because they could not control them, e.g., worrying about auto accidents because they are caused by other drivers. Such worry could also help people remember to take protective if the lack of control is only partial. For example, people who think they cannot control automobile accidents may still think that they can protect themselves by wearing seat belts. The third category of "government control or lack" involved giving lack of government control as a reason for worry (analogous to lack of personal control) or the presence of adequate government control as a reason for not worrying. For example, one subject worried about new diseases because "World Health people are behind and strains become immune and we could be heading back to the days of epidemics because the World Health can't keep up with all the different strains."
The association between lack of knowledge and worry refers more to personal lack of knowledge than to beliefs about what the experts know. For example, one subject worried about electromagnetic fields because "it is something people talk about but I don't really understand it." Thus, the prevalence of this response as a reason for worry is not inconsistent with the finding that expert knowledge was, if anything, positively related to worry.
People gave average probability as a reason for worry, but they often justified not worrying in terms of their own perceived low probability, e.g., not being in a high-risk area for natural disasters or not engaging in behaviors that could lead to AIDS. Badness of the outcome was also given as a reason for worry, as was the statement that something was a "big problem for society" (e.g., AIDS, or drugs).
Personal experience was the item most consistently associated with worry about risks that were relatively low in expected disutility. For example, one woman worried about air pollution because her husband had asthma, and a man worried about it because his father died of lung cancer.
Concerning risks with disagreement between what government should do and what individuals should do, the interview asked:
1. Why do you think that government should (should not) try to reduce this risk?
2. How much do you think that individuals can reduce the probability and magnitude of this risk?
3. How much do you think that government can reduce the probability and magnitude of this risk?
It was difficult to find risks in both categories for everyone, since our sample tended strongly to favor government responsibility. (Although they were predominantly "middle of the road" politically, most residents of West Philadelphia, both students and nonstudents, tend to vote Democrat and favor government action.) Still, we could classify risks roughly in terms of attitudes about who should control them. Of course, we did not classify all the risks in the study because we asked each subject only about four. Attitudes toward who should reduce the risk were most strongly justified in terms of who could control it most easily.
The largest category included risks that subjects felt should be dealt with by government because the government could control them and individuals could not: pesticides, cancer from food additives, contaminated food, vaccines, new diseases, AIDS from causes other than sex, chemicals in air and water, air pollution, nuclear radiation, airplane accidents, floods, war, and asteroids.
A few risks followed the opposite pattern, with respondents feeling that government should play a lesser role in reducing the risk because individuals had more control over the risks than government: fires and accidents in the home, unhealthy diet, lightning, and tornadoes. For these risks, subjects tended to say that individuals were "responsible" for control as well as saying that they could control the risk more easily.
For other risks, subjects said that both the government and the individual could control the risk, and, in these cases, they typically said that the government should control it: auto accidents, drugs/alcohol, AIDS from sex, and hurricanes. For example, individuals can drive safely, but government can enforce speed limits. In response to the question about why government should control the risk, many answers referred to the seriousness of the risk rather than the relative ease of control, e.g., "There is too much drunk driving and it destroys family life."
Some risks fell into a mixed category, with subjects responding differently from each other: crime, accidents at work, radon gas, earthquakes (where some respondents said that nobody could do anything). For example, two subjects said that government can educate people about radon, but another subject said that government cannot do much and that this is up to homeowners.
It is interesting that, for almost every risk, one could argue that both the individual and the government could do something. This is particularly true when we count as individual actions voting and giving money to causes - as some subjects explicitly did for malnutrition and war - and when we count education of individuals as a function of government - as subjects did for unhealthy diet and radon. Yet despite this possibility, 38% of the answers indicated that the individual and government had little or no control over the risk. In general, subjects seemed unaware of some of the possibilities for both the individual and the government to take action to reduce risks.
The last question asked what protective behaviors the subject took, and, in particular, why she did not do anything to prevent risks with the highest expected disutility and why she did take action to protect against those with the lowest. Selection of risks for this question were not, however, systematic, and subjects tended to answer about risks discussed in other parts of the interview.
Two subjects said they did not protect against any of the risks. Other subjects mentioned a variety of protective actions, e.g., avoiding fast food restaurants, avoiding pesticides, boiling water, using condoms, and getting a smoke detector. Subjects also mentioned not taking any protective actions against certain risks, because they thought that these actions were unnecessary, ineffective, or both. These include hurricanes, war, and bacterial infections and accidents at work. One subject did not want to protect against cancer from food additives because, "I "don't know what to do and don't want to give up foods I like, just in case." Another subject thought that protection against asteroids was unnecessary "because if they hit it is more likely they'll hit in a cornfield than in a city."

Conclusion

A major result of these two experiments is that concern for action, both personal and government, is strongly related to worry. Worry, in turn, is affected mainly by beliefs about probability. Although we did not compare our risk ladder to other methods, the strength of the probability effect suggests that this method is effective for eliciting probability judgments. It certainly did allow for a wide range of responses. (Eiser & Hoepfner discuss effects of response options on probability judgment.)
The relatively small effect of badness, on the other hand, may result from our choice of risks to include in the questionnaire. Most of our risks were serious ones. Of course, in a regression model, the weight of this dimension would be determined by even a few cases in which the risks were not so bad (e.g., bacterial infections), but range of variation in probability may still have been greater (in logarithmic terms) than the range of variation in badness of outcomes. The small effect may also result from our way of asking the questions. Others have found larger effects in similar studies.
Experts and non-experts did not differ much in what determined their worries or their desire for action, but they did differ in their beliefs about particular risks. Non-experts were much more concerned about what experts consider to be small risks of cancer from environmental sources. Experts were more concerned about the statistically more frequent but more mundane events such as auto accidents. These differences may result from attention in news reporting, with non-experts being more influenced by stories. This is, of course, what was found by Lichtenstein et al., but the particular risks subject to public concern may have changed in the last 20 years.
We are unsure of the direction of causality between worry and beliefs about probability. Possibly people inflate their probability estimates of risks they worry about, in order to make their worry seem rationally justified. If, however, probability beliefs affect worry, then it may be possible to change what people worry about by providing them with more accurate and comparative information about probabilities. If worry then affects protective behavior and political action, then people may engage in more effective protective behavior, and they may support the efforts of government to reduce more serious risks. The interview study supported the relation between worry and perceived probability. In addition, worry was sometimes related to control and sometimes to lack of control. If worry is a determining factor in people's attitudes toward certain risks, then it may be helpful to provide individuals with information about actions that are now being taken by industry and government as well as what they can do to reduce either the likelihood or consequences of these future events.
We conclude that, even though worry is an unpleasant emotion, it may be important in moving people to protect themselves against harm, both individually and collectively. More research is needed to better understand the role that worry plays in people's perception of risk and the actions that they take to reduce their future losses. In particular, although worry may motivate protective action, it may also be somewhat autonomous from beliefs. People may worry too much about risks they know to be minor and too little about risks they know to be serious. This discrepancy may weaken the effectiveness of attempts to inform people about the relative probability of risks. If so, it might be fruitful to try to address the worry itself.

References

1. Starr, C. (1969). Social benefit versus technological risk: What is our society willing to pay for safety? Science, 165, 1232-1238.
2. Slovic, P. (1987). Perception of risk. Science, 236, 280-285.
3. Slovic, P. (1998). Trust, emotion, sex, politics, and science: Surveying the risk-assessment battlefield. In M. H. Bazerman, D. M. Messck, A. E. Tenbrunsel, & K. A. Wade-Benzoni (Eds.) Environment, ethics and behavior: The psychology of environmental valuation and degradation. San Francisco: New Lexington Press.
4. Brun, W. (1994). Risk perception: Main issues, approaches, and findings. In G. Wright & P. Ayton (Eds.), Subjective probability, pp. 295-320. Chichester: Wiley.
5. Markowitz, H. M. (1959). Portfolio selection. New York: Wiley.
6. Fischhoff, B., Slovic, P., Lichtenstein, S., Read, S., & Combs, B. (1978). How safe is safe enough? A psychometric study of attitudes towards technological risks and benefits. Policy Sciences, 8, 127-152.
7. Holtgrave, D. R., & Weber, E. U. (1993). Dimensions of risk perception for financial and health risks. Risk Analysis, 13, 553-558.
8. Brun, W. (1992). Cognitive components of risk perception: natural versus manmade risks. Journal of Behavioral Decision Making, 5, 117-132.
9. Kuhn, K. M., & Budescu, D. V. (1996). The relative importance of probabilities, outcomes, and vagueness in hazard risk decisions. Organizational Behavior and Human Decision Processes, 68, 301-317.
10. Slovic, P. (1967). The relative influence of probabilities and payoffs upon perceived risk of a gamble. Psychonomic Science, 9, 223-224.
11. Gardner, G. T., & Gould, L. C. (1989). Public perceptions of the risks and benefits of technology. Risk Analysis, 9, 225-242.
12. Kraus, N. N., & Slovic, P. (1988). Taxonomic analysis of perceive risk: Modeling individual and group perceptions within homogeneous hazard domains. Risk Analysis, 8, 435-455.
13. Teigen, K. H., Brun, W., & Slovic, P. (1988). Societal risks as seen by a Norwegian public. Journal of Behavioral Decision Making, 1, 111-130.
14. Drottz-Sjöberg, B-M, Sjöberg, L. (1990). Risk perception and worries after the Chernobyl accident. Journal Environmental Psychology, 10, 135-149.
15. Sjöberg, L. (1998). Worry and risk perception. Risk Analysis, 18, 85-93.
16. Myers, J. R., Henderson-King, D. H., Henderson-King, E. I. (1997.) Facing technological risks: The importance of individual differences. Journal of Research in Personality, 31, 1-20.
17. Flynn, J., Slovic, P., & Mertz, C. K. (1994). Gender, race, and perception of environmental health risks. Risk Analysis, 14, 1101-1108.
18. Steiger, J. H. (1980). Tests for comparing elements of a correlation matrix. Psychological Bulletin, 87 , 245-251.
19. Birnbaum, M. H. (in press). Psychological experiments on the internet. New York: Academic Press.
20. Eiser, J. R., & Hoepfner, F. (1991). Accidents, disease, and the greenhouse effect: Effects of response categories on estimates of risk. Basic and Applied Social Psychology, 12, 195-210.
21. Weinstein, N. D. (2000). Perceived probability, perceived severity, and health-protective behavior. Health Psychology, 19, 1-10
22. Lichtenstein, S., Slovic, P., Fischhoff, B., Layman, M., & Combs, B. (1978). Judged frequency of lethal events. Journal of Experimental Psychology: Human Learning and Memory, 4, 551-578.
23. Baron, J. (1992). The effect of normative beliefs on anticipated emotions. Journal of Personality and Social Psychology, 63, 320-330.

Footnotes:

1This research was supported by N.S.F. grants SBR95-20288 and 524603 and by a grant from the Annenberg Foundation. Send correspondence to Jonathan Baron, Department of Psychology, University of Pennsylvania, 3815 Walnut St., Philadelphia, PA 19104-6196, or (e-mail) baron@psych.upenn.edu.

File translated from TEX by TTH, version 3.74.
On 24 Sep 2006, 14:20.