8.9 Risk perception and communication*
Oxford Textbook of Public Health
Risk perception and communication*
Baruch Fischhoff, Ann Bostrom, and Marilyn Jacobs Quadrel
Role of risk perceptions in public health
Role of perceptions about risk perceptions in public health
Estimating the size of risks
Response mode problems
Cumulative risk—a case in point
Mental models of risk decisions
Mental models of risk processes
Role of risk perceptions in public health
Many health risks are the result of deliberate decisions by individuals consciously trying to make the best choices for themselves and for those important to them. Some of these choices are private. They include decisions such as whether to wear bicycle helmets and seatbelts, whether to read and follow safety warnings, whether to buy and use condoms, and how to select and cook food. Other choices involve societal issues. They include such decisions as whether to protest the siting of hazardous waste incinerators and half-way houses, whether to vote for fluoridation and ‘green’ candidates, and whether to support sex education in schools.
In some cases, single choices can have a large effect on health risks (e.g. buying a car with airbags, taking a dangerous job, becoming pregnant). In other cases, the effects of individual choices are small, but can accumulate over multiple decisions (e.g. repeatedly ordering broccoli, wearing a seatbelt, using the escort service in parking garages). In still other cases, choices intended to affect health risks do nothing at all or the opposite of what is expected (e.g. responses to baseless cancer scares, subscription to quack treatments).
In order to make health decisions wisely, individuals need to understand the risks and the benefits associated with alternative courses of action. They also need to understand the limits to their own knowledge and to the advice proffered by various experts. This chapter reviews the research base for systematically describing people’s degree of understanding about health-risk issues. Some fundamental topics in designing and evaluating messages for improving that understanding are also considered. Following convention, these pursuits will be called risk perception and risk communication research, respectively.
Role of perceptions about risk perceptions in public health
The fundamental assumption of this chapter is that statements about other people’s understanding must be disciplined by systematic data. People can be hurt by inaccuracies in their risk perceptions. They can also be hurt by inaccuracies in what various other people believe about those perceptions. Particularly significant others include doctors, nurses, public health officials, legislators, regulators, and engineers—all of whom have some say in what risks are created, what is communicated about them, and what role laypeople have in determining their own fates.
If their understanding is overestimated, people may be thrust into situations that they are ill-prepared to handle (e.g. choosing among complex medical procedures). If their understanding is underestimated, people may be disenfranchised from decisions that they could and should make. The price of such misperceptions of risk perceptions may be exacted over the long run as well as in individual decisions. The outcomes of health-risk decisions partly determine people’s physical and financial resources. Furthermore, the processes by which health-risk decisions are made partly determine people’s degree of autonomy in managing their own affairs and in shaping their society.
In addition to citing relevant research results, this chapter will emphasize research methods. One conventional reason for doing so is improving access to material that is scattered over specialist literatures or is part of the implicit knowledge conveyed in professional training. A second conventional reason is to help readers to evaluate the substantive results reported here by giving a feeling for how they were produced.
A less conventional reason is to make the point that method matters. We are routinely struck by the strong statements made about other people’s competence to manage risks, solely on the basis of anecdotal observation and intuition. These statements appear directly in pronouncements about, for instance, why people mistrust various technologies or fail to ‘eat right’. Such claims appear more subtly in the myriad of health advisories, advertisements, and warnings directed at the public without any systematic evaluation. These practices assume that the communicator knows what people currently know, what they need to learn, what they want to hear, and how they will interpret a message.
Even casual testing with a focus group shows a willingness to have those (smug) assumptions challenged. The research methods presented here show the details that need attention and, conversely, the pitfalls to casual observation. This chapter also shows the limits to such research, in terms of how far current methods can go and how quickly they can get there. It has been our experience that, once the case has been made for conducting behavioural research, it is expected to produce results immediately. That is, of course, a prescription for failure and for undermining the perceived value of future behavioural research. Furthermore, the cumulative attack on public competence can lead to its disenfranchisement and the transfer of authority to technical experts, be they doctors or technology managers. Indeed, some attacks seem designed to effect such a change in the balance of political power. There are those who would just as soon have no research (or ambiguous research) regarding public perceptions of risk so that they can fill the void with their own punditry (Fischhoff et al. 1983; Fischhoff 1990; Leiss and Chociolko 1994; Okrent and Pidgeon 1998).
The next section, on quantitative assessment, treats the most obvious question about laypeople’s risk perceptions: Do they understand how large (and how small) various risks are? It begins with representative results regarding the quality of laypeople’s quantitative judgements, along with some relevant psychological theory. It continues with issues in survey design, focused on how design choices can affect respondents’ apparent competence. Some of these methodological issues reveal substantive aspects of lay risk perceptions.
The following section, on qualitative assessment, shifts the focus from summary estimates of risk to judgements about qualitative features of the events being considered. It begins with the barriers to communication created when experts and laypeople unwittingly use terms differently. For example, when experts tell (or ask) people about the risks of drinking and driving, what do people think is meant regarding the kinds and amounts of ‘drinking’ and of ‘driving’? We then ask how people believe that risks are created and controlled, as a basis for generating and evaluating action options.
The final section provides a general process for developing communications about health risks. That process begins by identifying the information to communicate, based on (a) descriptive study of what recipients know already, and (b) formal analysis of what they need to know in order to make informed decisions. The process continues by selecting an appropriate format for presenting that information. It concludes with an explicit empirical evaluation of the resulting communication, with the process being iterated if the results are wanting. The process is illustrated with examples taken from several case studies, looking at such diverse health risks as those posed by radon, Lyme disease, electromagnetic fields, carotid endarterectomy, and nuclear energy sources in space.
This chapter will not address several issues that belong in a fuller account. These include the roles of emotion, individual differences, culture, and social processes in decisions about risk. This set of restrictions suits the chapter’s focus on how individuals think about risks. It may also suit a public health perspective, where it is often necessary to ‘treat’ populations with information. The success of that effort both depends on and shapes recipients’ social and emotional resources for acting on its contents. Good communication can expand the envelope within which people feel that they can understand and act on the facts of risk. Access to missing topics might begin with Jasanoff (1986), Weinstein (1987), Heimer (1988), Krimsky and Plough (1988), National Research Council (1989), Otway and Wynne (1989), Douglas (1992), Krimsky and Golding (1992), Royal Society (1992), Yates (1992), and Leiss and Chociolko (1994).
Estimating the size of risks
A common presenting symptom in experts’ complaints about lay decision-making is that ‘they (the public) do not realize how small (or large) the risk is’. If that were the case, then the mission of risk communication would be conceptually simple (although still technically challenging)—transmit credible estimates of the magnitude of risks. The need for such communication can be seen in research showing that lay estimates of risk are, indeed, subject to biases (Kahneman et al. 1982; Slovic 1987; Weinstein 1987; Stallen and Thomas 1988). Rather less evidence directly implicates these biases in inappropriate risk decisions, or substantiates the idealized notion of people waiting for crisp risk estimates so that they can ‘run’ decision-making models in their heads. Such estimates are necessary, but not sufficient, for effective decisions. Accurate estimates alone cannot tell people what actions are possible or what goals are worth pursuing. They might not even show what risks are worth worrying about—insofar as there may be nothing to do about large risks, while small risks might be expeditiously handled (Fischhoff et al. 1983). Nonetheless, some notion of risk size is needed to begin focusing one’s attention.
In one early attempt to evaluate lay estimates of the size of risks, Lichtenstein et al. (1978) asked people to estimate the number of deaths in the United States from 30 causes (e.g. botulism, tornadoes, motor vehicle accidents). The ‘people’ in this study were members of the League of Women Voters and their spouses. Generally, the people in the studies described here have been paid for participation, sometimes drawn from university populations (students and staff) and sometimes recruited through civic groups (e.g. garden clubs, parent–teacher associations, bowling leagues). As a result, they are typically older (and perhaps better motivated) than the proverbial college sophomores of some psychological research. These groups have been found to differ more in what they think than in how they think, i.e. their respective experiences have created larger differences in specific beliefs than in thought processes. Fuller treatment of sampling issues must await another opportunity.
Lichtenstein et al. (1978) used two different response modes, allowing a check for the consistency of responses. One task presented pairs of causes; subjects chose the more frequent and then estimated the ratio of frequencies. The second task asked subjects to estimate the number of deaths in an average year. These subjects were told the answer for one cause, as an anchor, providing an order-of-magnitude feeling for the kinds of answers that were expected; a pretest had found that many subjects lacked a good idea of how many people live or die in the United States in an average year. The study reached several conclusions which have been borne out by subsequent studies (Vlek and Stallen 1981). These concerned internal consistency, anchoring bias, compression, availability bias, and miscalibration of confidence judgements.
Estimates of relative frequency were quite consistent both within and across response mode. Thus, people seemed to have a moderately well-articulated internal risk scale, which they were able to express even in the unfamiliar response modes used in these studies—in life, they had probably never been asked any question as explicit as these quantitative estimates of risk.
Direct estimates were influenced by the anchor that the investigators provided. Subjects told that 50 000 people die annually from vehicle accidents produced estimates two to five times higher than did subjects told that 1000 die from electrocution. Thus, people seem to have less of a feel for absolute frequency, rendering them sensitive to the implicit cues in how questions are asked (Poulton 1989; Hurd 1999; Schwarz 1999).
Subjects’ estimates showed less dispersion than did the statistical estimates. While the statistical estimates varied over six orders of magnitude, the typical subject’s estimates ranged over three or four. In this case, the result was overestimation of small frequencies and underestimation of large ones. However, the anchoring bias suggests that this overall pattern might have changed with different procedures, making the compression of estimates the more fundamental result. For example, using an even lower anchor (e.g. the average annual toll of botulism deaths) would have reduced the overestimation of small frequencies and increased the underestimation of large ones. If these responses reflect subjects’ actual feeling for the relative size of different risks, then people may have difficulty appreciating the enormous range in the frequencies of life’s risks. That would not be surprising, considering how rare it is for media reports or health-care professionals to provide explicit quantitative estimates.
At each level of statistical frequency, some causes of death (e.g. homicide, tornadoes, flood) consistently received higher estimates than others. Additional analyses showed these to be causes that are disproportionately visible (e.g. as reported in the news media, as experienced in subjects’ lives). This bias seemed to reflect a special case of a general tendency to estimate the frequency of events by the ease with which they are remembered or imagined—while failing to realize what a fallible index such availability is (Tversky and Kahneman 1973; Kahneman et al. 1982). These results are consistent with those in experimental psychology, showing that people are generally quite proficient at tracking how frequently events are observed, but not so good at detecting systematic biases in those observations (Hasher and Zacks 1984; Koriat 1993).
Miscalibration of confidence judgements
In a subsequent study (Fischhoff et al. 1977), subjects were asked how confident they were in their ability to choose the more frequent in a pair of causes of death (e.g. tornado, asthma). They tended to be overconfident. For example, they chose correctly only 75 per cent of the time when they were 90 per cent confident of having done so. This result is a special case of the general tendency to be inadequately sensitive to the extent of one’s knowledge (Lichtenstein et al. 1982; Yates 1989).
Figure 1 shows typical results from such a calibration test. In this case, subjects expressed their confidence in having chosen the correct answer to two alternative questions regarding health risks (e.g. alcohol is (a) a depressant or (b) a stimulant). In the figure, each point reflects the proportion of correct responses associated with answers assigned a particular probability of being correct. Thus, in the lowest curve, subjects were correct about 70 per cent of the time when 100 per cent certain of being correct. The two upper curves reflect a group of middle-class adults and some of their adolescent children, recruited through school organizations. As in other studies of cognitive ability, the judgemental processes of these groups are quite similar (Keating 1988). The third curve reflects at-risk teenagers, recruited from group homes and treatment centres. They knew less about these risk issues, but were just as confident; indeed, over 40 per cent of their responses indicated complete confidence in the associated answer. One possible explanation of this greater overconfidence is that their personal experiences with risks create an illusion of understanding, leading them to feel inappropriately like experts. A second is that the high-risk teenagers have less ability to think critically about the bases of their beliefs or less willingness to do so. Effective decision-making requires not just having knowledge, but also recognizing the limits to one’s understanding.
Fig. 1 Calibration curves for adults (top, white: n = 45), not-at-risk teenagers (middle, dark: n = 43), and at-risk teenagers (bottom, white: n = 45). Each point indicates the proportion of correct answers among those in which subjects expressed a particular confidence level. The size of each circle indicates the percentage of answers held with that degree of confidence. (Source: Quadrel 1990.)
Response mode problems
One recurrent obstacle to assessing or improving laypeople’s estimates of risk is reliance on verbal quantifiers for both communicating and eliciting risk estimates. It is difficult for people to know what experts mean when a risk is described as ‘very likely’ or ‘rare’. It is equally difficult for experts to evaluate lay perceptions expressed in those terms. Such terms mean different things to different people and even to the same person in different contexts (e.g. likely to be fatal versus likely to rain, rare disease versus rare Cubs baseball championship). Such ambiguity has been found even within communities of professionals, such as doctors and intelligence officers (Lichtenstein and Newman 1967; Beyth-Marom 1982; Merz et al. 1991). The criticality of such ambiguity depends on how the estimates are used. Sometimes, inferred probabilities of 1 and 10 per cent will lead to the same choice; sometimes they will not. The variability of interpretations should increase with the diversity of individuals and decisions.
Table 1 shows the results of asking a fairly homogeneous group of subjects (undergraduates at an American Ivy League college) to judge seven risks in both quantitative and qualitative terms, with reported statistical rates. The quantitative estimates used a response scale that explicitly offered probabilities as low as 0.01 per cent (or 1 in 10 000). The qualitative scale used typical labels (converted to interval-scale equivalents in the data analyses: 1, very unlikely; 2, unlikely; 3, somewhat unlikely; 4, somewhat likely; 5, likely; 6, very likely). Comparing the two response scales revealed a non-linear relationship between the two. Specifically, the median probabilities (column 1) associated with the median qualitative estimates (column 2) were as follows: very unlikely, 0.01 per cent; unlikely, 0.5 per cent; somewhat unlikely, 5 per cent; somewhat likely, 25 per cent; likely, 60 per cent; very likely, 96 per cent. Budescu and Wallsten (1995) review the evidence on the predictability of such equivalences across tasks, and what they reveal about experiences with uncertain events.
Table 1 Risk estimates elicited with numerical and verbal response modes compared with statistical risk estimates (‘Please estimate your personal risk to the following events in the next 3 years.’)
Lichtenstein et al. (1978) provided anchors in order to give subjects a feeling for how to answer. The anchors should have improved subjects’ performance by drawing responses to the correct range, within which subjects were drawn to higher or lower values depending on the value of the anchor that they received. Most of the study’s conclusions were relatively insensitive to these anchoring effects—except for the critical question of how much people overestimate or underestimate the risks that they face. That depended, in part, on how the question was asked.
A study by Fischhoff and MacGregor (1983) provides another example of the dangers of relying on a single response mode. They used four different response modes to ask about the chances of dying, conditional on being afflicted with each of various maladies (in the United States).
How many people die out of each 100 000 who contract influenza?
How many people died out of the 80 million who caught influenza last year?
For each person who dies of influenza, how many have it and survive?
Eight hundred people died of influenza last year. How many survived?
Again, there was strong ordinal consistency across response modes, while absolute estimates varied over one to two orders of magnitude. A follow-up study looked for independent evidence of the relative suitability of these different response modes. It found that subjects liked one format much less than the others. They were also least able to remember statistics reported in that format. This was also the format that produced the most discrepant results—estimating the number of survivors for each person who succumbed to a problem.
Estimating the accuracy of risk estimates requires not only an appropriate response mode, but also a standard against which responses can be compared. The studies described above asked about population risks in situations where credible statistical estimates were available to the investigators. People’s performance might be different (and more difficult to evaluate) with risks whose magnitude is less readily calculated. Furthermore, for many decisions, people’s understanding of population risks is less relevant than their understanding of personal risks. Unfortunately, personalized risk statistics are usually difficult to come by.
In order to circumvent these problems, some investigators have asked subjects to judge whether they are more or less at risk than others in (more or less) similar circumstances (Svenson 1981; Weinstein 1989; Weinstein and Klein 1996). They find that most people in most situations see themselves as facing less risk than average others; that could, of course, be true for only half a population. A variety of processes could account for such a bias, including both cognitive ones (e.g. the greater availability of the precautions that one takes—than those taken by others) and motivational ones (e.g. wishful thinking). Such an optimism bias could prompt unwanted risk-taking (e.g. because warnings seem more applicable to other people than to oneself). In a direct test of this hypothesis, Quadrel et al. (1993) asked subjects to judge the probability of various misfortunes befalling them and several comparison individuals (a close friend, an acquaintance, a parent, a child).
The events involved ‘a death or injury requiring hospitalization over the next 5 years’, from sources like vehicle accidents, drug addiction, and explosions. Subjects were most likely to see each person’s risks similarly, meaning that perceived invulnerability was the exception, rather than the rule. However, when they did make a distinction, they were twice as likely to see themselves as the person facing less risk. This perception of relative personal invulnerability was particularly large for risks seen as under some personal control. Here too, adults and adolescents responded similarly, despite the common belief that teenagers take risks, in part, because of a unique perception of invulnerability (Elkind 1967). A more complicated account is needed to explain the risks that teenagers take and avoid (Schulenberg et al. 1997; Arnet 1999).
A log-linear response mode
Figure 2 shows the response mode used in this study. It uses a linear scale for probabilities between 1 and 100 per cent and six cycles of a logarithmic scale for smaller ones. The scale was explained to subjects in groups and introduced with a few examples having obvious and extreme values (e.g. being hit by lightning, catching a cold), in order to help subjects to understand it. Quadrel et al. (1993) found similar scale usage, not to mention similar beliefs (as noted above), among groups of middle-class adults, their high-school children, and high-risk teenagers recruited from group homes and treatment centres. The statistical risk estimates of Table 1 were elicited with a variant of this scale. Comparison between columns 1 and 4 allows a quantitative assessment of the accuracy of the risk estimates. Formal analysis might show whether errors of this magnitude would be large enough to influence decisions relying on them (see discussion of value-of-information analysis below).
Fig. 2 Log–linear response scale for eliciting probability assessments, facilitating the expression of very small probabilities. (Source: Quadrel et al. 1993.)
In Quadrel et al.’s administration, the instructions were read aloud to circumvent any problems. However, subsequent studies have just handed out the scales with minimal instructions (Fischhoff and Bruine de Bruin 1999). In a questionnaire study with a random national sample of women, Woloshin et al. (1998) found that a log–linear scale was at least as reliable (and well accepted) as four competing response modes (two verbal and two numerical). The limiting factor on the use of such scales may not be the subjects’ ability to understand quantitative probabilities, but the clarity of the instructions and their willingness to work. People prefer to receive numerical probabilities, but would rather provide verbal ones (which require less effort and make a weaker commitment) (Erev and Cohen 1990).
The critical question in considering the usefulness of such a response mode is whether the additional information that it provides compensates for the extra demands that it imposes on subjects. For example, does it help to see that the young adults in Table 1 moderately underestimate their risk of herpes infection (over the next 3 years), while moderately overestimating their risk of HIV (although recognizing that it is still very small)? Is it helpful to know that Quadrel et al.’s subjects assigned probabilities of less than 1 in 10 million about 10 per cent of the time and probabilities of less than 1 in 10 000 about one-third of the time. Using responses to quantitative scales, Viscusi (1992) has argued that the public-health establishment has more than succeeded in convincing adults of the risks of smoking. If that is the case, then educational efforts should be focused on teenagers or on the risks of addiction. With similar procedures, Black et al. (1995) concluded that women in the United States overestimate the risk of breast cancer. If so, then campaigns designed to increase awareness might need to be reconsidered.
The response distributions for these last two studies revealed an anomaly often found with studies using open-ended probability scales (e.g. ‘use a number between 0 per cent, meaning impossible, and 100 per cent, meaning certain’). A disproportionate share of responses lay at ’50’, whereas most responses were much lower. Many of these responses seemed to mean ’50–50,’ rather than the numeric probability. Such ‘blips’ seem particularly common when respondents do not know what to say or do not want to think about the event in question—as might be the case with many risks (Fischhoff and Bruine de Bruin 1999). If taken at face value, these 50s inflate group estimates of small risks. If taken as evidence of respondents’ discomfort with the question (or epistemic uncertainty), they provide another side to how people deal (or avoid dealing) with risks. Bruine de Bruin (1999) provides further examples of such blips, including ones from technical experts, as well as procedures for recalibrating response distributions that include such expressions of epistemic uncertainty (Gärdenfors and Sahlin 1982).
Realizing the potential of precise response scales requires applying them to precisely defined events. Some years ago, the United States National Weather Service considered abandoning quantitative probability-of-precipitation forecasts because of reported consumer confusion. Murphy et al. (1980) discovered that the confusion was actually about the event being predicted. For example, recipients were uncertain whether a 70 per cent chance of rain referred to the portion of the area that would receive rain, the percentage of time that it would rain, the chance of measurable rain at any spot in the area, or that chance at the weather station. (It is the last.) Event ambiguity is treated further below (and by Fischhoff (1994) and Fischhoff et al. (1999)). Where the magnitude of risk perceptions matters, we prefer to use quantitative response scales with well-defined events.
These studies provide measures of risk perceptions, if one assumes that people define ‘risk’ as ‘probability of death’. However, observation of scientific practice shows that, even among professionals, ‘risk’ can have a variety of meanings (Crouch and Wilson 1982; Fischhoff et al. 1984; Royal Society 1992; National Research Council 1996). For some experts, ‘risk’ equals expected loss of life expectancy; for others, it is expected probability of premature fatality; for still others, it is total number of deaths or deaths per person exposed or per hour of exposure, or loss of ability to work (Starr 1969; Inhaber 1979; Wilson 1979; Kammen and Hassenzahl 1999).
Unwitting use of different definitions can lead to controversy and confusion, insofar as the relative riskiness of different jobs, avocations, technologies, and diseases depends on the choice of definition. Although often left to the conventions of technical experts, the choice of definition is a political/ethical decision that can significantly affect a society’s allocation of resources. For example, hazards producing deaths by injury become relatively ‘riskier’ if one counts the total years lost rather than weighting all deaths equally. That measure of risk places a greater premium on deaths among young people, because more years of life are lost with them. Some of the apparent disagreement between experts and laypeople regarding the magnitude of risks in society seems due to differing definitions of ‘risk’ (Slovic et al. 1979; Vlek and Stallen 1980; Fischhoff et al. 1983; National Research Council 1996, 1999).
At times, ‘risk’ is used as a discrete rather than a continuous descriptor, i.e. an activity or technology is described as being a risk or not being a risk. Conversely, it might be described as being ‘safe’ or not. Such a shorthand expression naturally conveys rather little information, beyond the summary judgement regarding where the hazard in question falls relative to some threshold. Without further detailed study (using methods like those described here) one could not know what the individual or institution using the phrase meant. The phrase might refer to a general de minimis level (reflecting risks that can confidently be treated as not worth worrying about), the invocation of a precautionary principle (reflecting an appraisal of the probability distribution over possible risk levels), or the result of a cost–benefit summary (such that the risk is negligible in the context of the other consequences). Even when people talk in zero-risk terms, they may mean something different than a technical specialist using the same words. When those specialists wish to communicate regarding safety, they bear a particular burden to ensure that their terminology matches that of their audience. Otherwise, their public may feel that a social contract was violated, when they discover what the experts really meant.
One early study asked experts and laypeople to estimate the ‘risk of death’ faced by society as a whole from 30 activities and technologies (Slovic et al. 1979). The experts’ judgements were much more highly correlated with statistical estimates of average-year fatalities than were the laypeople’s estimates. When laypeople were asked to estimate average-year fatalities, they responded much like the experts. However, when laypeople estimated ‘risk of death’, they also seemed to consider (what they saw as) the catastrophic potential of the technology (i.e. its ability to cause large numbers of deaths in non-average years). Thus, experts and laypeople agreed about routine death tolls (for which scientific estimates are relatively uncontroversial) and disagreed about the possibility of anomalies (for which the science is typically much weaker). This seemingly reasonable pattern would be obscured by the casual observation that experts and laypeople disagree about ‘risk’ or by the assumption that any disagreement means that the experts are right and the laypeople are wrong.
Sensing that there was something special about catastrophic potential, some risk experts have suggested that social policy pay special attention to the regulation of hazards carrying that kind of threat. However, one experimental study found that people did not care more for losing many lives in a single incident than for losing the same number of lives in separate incidents (Slovic et al. 1984). Rather, catastrophic potential worries people because a technology posing such threats may prove to be out of control, despite its promoters’ promises. Such ‘surprise potential’ is strongly correlated with ‘catastrophic potential’ in people’s judgements; the same is presumably true in scientific estimates (Funtowicz and Ravetz 1990).
When accidents involving large numbers of fatalities are easy to imagine, catastrophic potential can be rated high because of availability, even when estimates of average-year fatalities are relatively low, as was the case for nuclear power in this study.
However, the two features represent rather different ethical bases for distinguishing among risks.
Dimensions of risk
Uncertainty and catastrophic potential are not the only dimensions of risk that might influence how they are judged. Much research and speculation has been devoted to these features of risk (Lowrance 1976; Slovic et al. 1980, 1985; Green and Brown 1981; Vlek and Stallen 1981; Cole and Withey 1982), with the set of proposed features running to several dozen (Jenni 1997). This is an unwieldy number of features for a descriptive theory of risk perceptions, a prescriptive guide to risk decisions, or a scheme for predicting public responses to new hazards or hazard-reduction schemes. As a result, various empirical studies have attempted both to test these speculations and to reduce the number of considerations. Most have elaborated on a correlation scheme offered by Fischhoff et al. (1978). In it, members of a liberal civic organization rated 30 environmental hazards on nine hypothesized aspects of risk. Factor analysis reduced the mean ratings of nine aspects to two ‘dimensions’, which accounted for 78 per cent of the variance. Similar patterns were found with students, members of a conservative civic organization, members of a liberal women’s organization, and technical risk assessors. Figure 3 shows the factor scores for 30 hazards within the common factor space for these four groups.
Fig. 3 Location of 30 hazards within the two-factor space obtained from League of Women Voters, student, active club, and expert groups. Respondents evaluated each activity or technology on each of nine features. Ratings were subjected to principal components factor analysis, with a varimax rotation. Connecting lines join or enclose the loci of four group points for each hazard. Open circles represent data from the expert group. Unattached points represent groups that fall within the triangle created by the other three groups. (Source: Slovic et al. 1985.)
Hazards at the high end of the vertical factor (e.g. food colouring, pesticides) tended to be new, unknown, involuntary, and delayed in their effects. High (right-side) scores on the horizontal factor (e.g. nuclear power, commercial aviation) mean that consequences are seen as certain to be fatal, and to affect large numbers of people, should something go wrong. The vertical factor was labelled unknown risk and the horizontal factor dread risk. They might be seen as capturing the cognitive and emotional bases of people’s concern respectively.
Other studies, employing variants on this ‘psychometric paradigm’, have yielded results that are similar in many respects. For example, despite changes in elicitation mode, scaling techniques, items rated, and subject population, two or three dimensions have proved adequate. Where a third dimension emerges, it typically refers to the absolute number of lives exposed to the threat in present or future generations; catastrophic potential has been used as a label. The position of particular technologies in this space proves to be highly robust. Moreover, that position is correlated strongly with various attitudes, including the desired stringency of regulation. Such analyses of mean responses are most suitable for predicting aggregate responses to hazards. The international and intercultural comparison of such risk spaces has proven to be a fruitful area, with a standard methodology revealing local differences (and similarities) (Kuyper and Vlek 1984; Ënglander et al. 1986; Goszczynska et al. 1991; Karpowicz-Lazreg and Mullet 1993; Vaughan 1993; Jianguang 1994; Rohrmann 1994).
The multidimensional character of risk means that hazards that are similar in many ways may still evoke quite different responses. This fact is neglected in appeals to accept one risk because one accepts another risk that is similar to it in some ways (Fischhoff et al. 1981; Crouch and Wilson 1982). The most ambitious of these appeals present elaborate lists of hazards, exposure to which is adjusted so that they pose equivalent statistical risks (e.g. consuming one tablespoonful of peanut butter and living for 50 years at the boundary of a nuclear power plant both create a one-in-a-million risk of premature death). Recognizing that such comparisons are often perceived as self-serving, the Chemical Manufacturers Association commissioned a guide to risk comparisons (Covello et al. 1988), which presents many such lists, along with the attached caution (in capital letters): ‘Warning! Use of data in this table for risk comparison purposes can damage your credibility’. The guide also offers advice on how to make risk comparisons, if one feels the compulsion, along with examples of more and less acceptable comparisons. Although the advice was derived logically from risk-perception research, it was not tested empirically. In such a test, we found little correlation between the authors’ predicted degree of acceptability and that judged by several diverse groups of subjects (Roth et al. 1990; MacGregor 1991).
One possible reason for the failure of these predictions is that the manual’s authors knew too much (from their own previous research) to produce truly unacceptable comparisons. More important than identifying the specific reasons for this failure is the general cautionary message: because we all have experience in dealing with risks, it is tempting to assume that others share our intuitions. Often, they do not. Effective risk communication requires careful empirical research. A poor risk communication can cause more public health (and economic) damage than the risks that it attempts to describe. One should no more release an untested communication than an untested medical device. The need for research is further magnified when one crosses cultural or national boundaries.
Over the past decade, many risk professionals have recognized the need to have risk priorities, while respecting differences in the definition of risk (Davies 1996). The result has been various forms of risk-ranking exercises, in which groups of citizens debate which risks merit the greatest attention. Participants are allowed (even encouraged) to disagree about which consequences matter. However, staff work attempts to provide a common credible basis for how large and likely those consequences are. Perhaps the most ambitious efforts has been conducted by the United States Environmental Protection Agency (1993), which has promoted some 50 regional, state, and national exercises. The US National Institutes of Health has convened a director-level Council of Public Representatives, in order to help scientific and lay communities understand one another’s priorities (Institute of Medicine 1998).
Scientific estimates of the magnitude of a risk require detailed specification of the conditions under which it is to be observed. For example, a fertility counsellor estimating a woman’s risk of an unplanned pregnancy would consider the frequency and timing of intercourse, the kinds of contraceptive used (and the diligence of their application), her physiological condition (and that of her partner), and so on. If laypeople are to make accurate assessments, they require the same level of detail. That is true whether they are estimating risks for their own sake or for the benefit of an investigator studying risk perceptions.
When investigators omit necessary details, they create adverse conditions for subjects. In order to respond correctly, subjects must first guess the question and then know the answer. Consider, for example, the question: ‘What is the probability of pregnancy with unprotected sex?’ A well-informed subject who understood this to mean a single exposure would be seen as underestimating the risk—by an investigator who intended the question to mean multiple exposures.
Such ambiguous ‘events’ are common in surveys of public risk perceptions. For example, a National Center for Health Statistics survey (Wilson and Thornberry 1987) question asked: ‘How likely do you think it is that a person will get the AIDS virus from sharing plates, forks, or glasses with someone who had AIDS?’ Fischhoff (1989b) asked a relatively homogeneous group of subjects to answer this question, and then to say what they thought was meant regarding the amount and kind of sharing that it implied. For their responses to be interpretable, subjects must spontaneously assign the same value to each missing detail and investigators must guess what value subjects have chosen. These subjects generally agreed about the kind of sharing (82 per cent interpreted it as sharing during a meal), but not about the frequency (a single occasion, 39 per cent; several occasions 20 per cent; routinely, 28 per cent; uncertain, 12 per cent). Thus, these subjects were answering different questions, rendering their responses ambiguous. In this case, the response mode was also ambiguous (very likely, unlikely, and so on), so that even precise questions would have revealed little. A survey question about the risks of sexual transmission evoked similar disagreement.
Interestingly, all the subjects who reported uncertainty about the frequency and intensity of sharing (or of sexual activity) still made likelihood judgements. If people are willing to respond to survey questions that they do not understand, any relationship between their reported beliefs and behaviours would tend to be blurred. That could, in turn, lead an observer to think, for example, that ‘information does not work with teenagers’, insofar as their actions seem unrelated to their beliefs. If so, that would be a special case of the general tendency for poor measurement to reduce the power of research designs. An important role of the National Center for Health Statistics study, one of an annual series, is to guide national policy on HIV/AIDS. No one has studied what readers of the National Center for Health Statistics survey’s results believed about subjects’ interpretations of its question. However, if they misunderstood subjects’ beliefs, then they may have produced ineffective and misdirected communications.
Aside from their methodological importance, the details that subjects infer can be substantively interesting. People’s intuitive theories of risk are revealed in the variables that they note and the values that they supply. In a systematic evaluation of these theories, Quadrel (1990) asked adolescents to think aloud as they estimated the probability of several deliberately ambiguous events (e.g. having an accident after drinking and driving, contracting AIDS through sex).
These subjects typically wondered (and made assumptions) about a large number of features. In this sense, subjects arguably showed more sophistication than the investigators who created the surveys from which these simplistic questions were taken or adapted. Generally, these subjects were interested in variables that could figure in scientific risk analyses, i.e. they wanted relevant information that had been denied them by the investigators. However, there were some interesting exceptions. Although subjects wanted to know the ‘dose’ involved with most risks, they did not ask about the amount of sex in a question about the risks of pregnancy or in another question about the risks of HIV transmission. They seemed to believe that an individual either is or is not susceptible to the risk, regardless of the amount of the exposure. In other cases, subjects asked about variables with a less clear connection to risk level (e.g. how well members of the couple knew one another).
In a follow-up study, Quadrel (1990) presented richly specified event descriptions to teenagers drawn from the same populations (school organizations and substance abuse treatment homes). Subjects initially estimated the probability of a risky outcome on the basis of some 12 details. Then, they were asked how knowing each of three additional details would change their estimates. One of those details had been identified as relevant by subjects in the preceding study; two had not. Subjects in this study were much more sensitive to changes in the relevant details than to changes in the irrelevant ones. Thus, at least in these studies, teenagers did not balk at making quantitative judgements regarding complex stimuli. When they did so, they revealed consistent intuitive theories in rather different tasks.
Studies integrating structured and open-ended methods are increasingly being used to get around the limitations of conventional surveys for eliciting beliefs regarding complex or unfamiliar topics (Fischhoff 1991; Schwarz 1996, 1999). These procedures assume that more reactive measurement is needed, if respondents are to understand such topics. They attempt to enrich, rather than bias, responses by offering a neutral mix of competing perspectives and prompts to think more deeply. These methods can also reveal the intuitive theories that respondents invoke, as they construe tasks in personally meaningful ways (McIntyre and West 1992; Gregory et al. 1993; Beattie et al. 1998; Fischhoff et al. 1999; Payne et al. 1999). Although groups are sometimes used as a forum for airing issues, the responses that ‘count’ in these studies are ones made in private by individual participants or in public through collective agreement (as in the risk-ranking exercises). As such they differ from the focus groups popular in market research, in which survey questions, commercial messages, political postures, or consumer products are discussed by groups of laypeople. Focus groups can be very productive in generating otherwise unanticipated perspectives. However, they create rather different situations than those faced by individuals or meaningful groups trying to make sense out of a question, message, or product. Merton (1987), who initially devised focus groups (and, before them, focused interviews) as a technique for uncovering possible hypotheses, has discouraged their use for testing (even those) hypotheses.
Cumulative risk—a case in point
As knowledge accumulates about people’s intuitive theories of risk, it will become easier to predict which details subjects know and ignore, as well as which omissions they will notice and rectify. In time, it might become possible to infer the answers to questions that are asked from ones that are not—as well as the inferences that people make from risks that are described explicitly to risks that are not. The invulnerability results reported above show the need for empirical research to discipline extrapolations from one setting to another. Asking people about the risks to other people like themselves is not the same as asking them about their personal risk. Nor can it be assumed that hearing about others’ risk levels will lead people to draw personal conclusions.
One common, and seemingly natural, extrapolation is across settings differing in the number of exposures to a risk. Telling people the risk from a single exposure should allow them to infer the risk from the number of exposures they expect to face; asking subjects what risk they expect from one number of exposures should allow one to infer what they expect from other numbers. Unfortunately, for both research and communication, teenagers’ insensitivity to the amount of intercourse in determining the risks of pregnancy or HIV transmission proves to be a special case of a general problem. Several reviews (Cvetkovich et al. 1975; Morrison 1985) have concluded that between one-third and one-half of sexually active adolescents explain their not using contraceptives with variants of, ‘I thought I (or my partner) couldn’t get pregnant’. A safe exposure or two might confirm that belief, discouraging behaviour that increased long-term risk.
In another study, Shaklee and Fischhoff (1990) found that adults greatly underestimated the rate at which the risk of contraceptive failure accumulates through repeated exposure—even after eliminating (from the data analysis) the 40 per cent or so of subjects who saw no relationship between risk and exposure. One corollary of this bias is not realizing the extent to which seemingly small differences in annual failure rates (the statistic that is typically reported) can lead to large differences in cumulative risk. Bar-Hillel (1974) and Cohen and Hansel (1958) found underaccumulation in simple clearly described gambles. Wagenaar and Sagaria (1975) found it in estimating cumulative environmental degradation.
After providing practice with a response mode facilitating the expression of small probabilities, Linville et al. (1993) asked college students to estimate the probability of HIV transmission from a man to a woman as the result of 1, 10, or 100 cases of protected sex. For one contact, the median estimate was 0.10, a remarkably high value compared with public health estimates (Fineberg 1988; Kaplan 1989). For 100 contacts, the median estimate was 0.25, which is a more reasonable value; however, it is also quite inconsistent with the single-exposure estimate. Assuming their independence, 100 exposures should provide a near-certainty of transmission. Very different pictures of people’s risk perceptions would emerge if a study asked just one of these questions. Conversely, risk communicators could achieve quite different effects if they chose to relate the risk of just one exposure or just 100. Communicators might create confusion if they chose to communicate both risks, leaving recipients to reconcile the seeming inconsistency.
Mental models of risk decisions
Each of these studies brought one element of a decision to subjects’ attention. A more comprehensive research strategy asks respondents to judge each element in a standard representation of their decision-making situation. Perhaps the most common of such models have an expectancy-value form (Feather 1982). In them, decisions are assumed to be determined by a multiplicative combination of the rated likelihood and (un)desirability of various prespecified consequences. Health belief and theory of reasoned action models fall into this general category. For example, Bauman (1980) had seventh graders evaluate 54 possible consequences of using marijuana, in terms of their importance, likelihood, and valence (i.e. whether each is positive or negative). A ‘utility structure index’, computed from these three judgements, predicted about 20 per cent of the variance in subjects’ reported marijuana usage. Related studies have had similar success in predicting other teenage risk behaviours.
The experience of these studies resembles that of earlier studies of ‘clinical judgement’, which successfully predicted expert decision-making with multiple regression models applied to experts’ ratings of standard variables. Initially, investigators interpreted the regression coefficients as reflecting the weights that people give to different concerns (Hoffman 1960; Hammond et al. 1964; Goldberg 1968). However, formal analyses eventually showed that many weighting schemes would produce similar predictions, as long as they contained the same variables (or correlated surrogates) (Wilks 1938; Dawes and Corrigan 1974). The good news in this result is that any linear combination of relevant variables will have some predictive success. The bad news is that it can be very difficult to distinguish alternative models, in terms of their relative accuracy as descriptions of decision-making processes. Thus, linear models can have considerable practical value in predicting choices, while still having limited ability to clarify how choices are made (Camerer 1981; Dawes et al. 1989). As a result, linear models provide a sort of cognitive task analysis, identifying the kinds of factors that might be involved in people’s choices. Other procedures are needed to clarify the finer structure of how decisions are made.
One such procedure was used by Beyth-Marom et al. (1993), who asked teenagers drawn from low-risk settings (e.g. sports teams, service clubs) and their parents to produce possible consequences of several decisions (e.g. deciding to smoke marijuana which was passed around at a party). Some subjects were asked to consider the act of accepting the offer, while others considered the consequences of rejecting it, in order to see whether these formally complementary options would elicit complementary perceptions. In almost all respects, the teenagers and parents responded quite similarly. On average, they produced about six consequences, with a somewhat higher number for accepting the risky offer than for rejecting it (suggesting that the thought of doing something is more evocative than the thought of not doing it). Respondents produced many more bad than good consequences of doing the focal behaviour, but fairly equal numbers for not doing it; thus, avoiding the risk was not as attractive as accepting it was unattractive. Most of the consequences that respondents mentioned were social reactions and personal effects. The social reactions of peers were particularly salient as consequences of rejecting the risk behaviour (e.g. more subjects said ‘They will laugh at me if I decline the offer’ than ‘They will like me if I accept’). The thought of doing a behaviour once and of doing it regularly evoked somewhat different consequences. For example, the social reactions of peers were mentioned more frequently as consequences of ‘accepting an offer to smoke marijuana at a party’, while decreased mental function was mentioned more frequently for ‘using marijuana’. These open-ended questions produced quite different consequences from the ones that appeared in earlier studies which required respondents to evaluate each item in a fixed list (e.g. the proportion of positive consequences was lower here). It would seem difficult to understand lay perceptions, or to improve them, without understanding such details, which seem to require an open-ended approach to emerge. For example, one might waste time and credibility trying to bring adolescents to adult’s level of awareness about consequences, something that they already seem to have.
Another study further weakened the degree of imposed task structure by letting teenagers choose three recent difficult decisions in their lives, to be described in their own terms (Fischhoff 1996a). These descriptions were coded in terms of their content (what’s on teenagers’ minds) and structure (how those issues are formulated). Figure 4 shows a moderately well-structured choice about drinking and driving. None of the decisions that the 105 teenagers chose dealt with drinking and driving, although quite a few dealt with drinking. For those decisions that were mentioned, few had an option structure as complicated as that in the simple decision tree of Fig. 4. Rather, most were described in terms of a single option (e.g. whether to go to a party where alcohol would be served).
Fig. 4 Decision tree for whether to take or decline a ride from friends who have been drinking. (Source: Fischhoff and Quadrel 1991.)
In a two-option decision, as in Fig. 4, the consequences of the alternative option are logically implied. However, that need not mean that they are intuitively obvious to the decision-maker. Indeed, Beyth-Marom et al. (1993) found that the consequences produced for engaging in a risky behaviour were not the mirror image of the consequences of rejecting that opportunity. This asymmetry is also seen in experimental results showing that foregone benefits of decisions, or their opportunity costs are much less visible than their direct costs (Kahneman et al. 1991; Thaler 1991). The differential visibility of such consequences can, in turn, be associated with ineffective decision-making. For example, the direct risks of vaccinating one’s child loom disproportionately large, relative to the indirect risks of not vaccinating (Harding and Eiser 1984; Ritov and Baron 1990).
We believe that a mix of structured and unstructured studies is needed to piece together a full account of lay decisions—as a prelude to predicting or aiding them. Normative decision theory provides a conceptual framework for determining which topics to study. Descriptive decision theory provides methodological and theoretical tools for pursuing that study. All are imperfect. However, in combination, they can begin to provide the sort of complex descriptions that people’s decisions about complex topics deserve. Attention to methodological detail is always critical. Decision variables will explain little if they are measured poorly. People whose behaviour seems unpredictable may lose the respect of observers and, thereby, become the target of manipulation—by others who conclude that this is the only way to get them to behave responsibly (for their own good). In this way, imprecise science (not to mention reliance on anecdotal observation) can undermine civil society.
Mental models of risk processes
Role of mental models
As noted above, people often have flawed intuitive theories of how risks accumulate, not realizing how risks mount up through repeated exposure—and perhaps neglecting the long-term perspective altogether. Such research can improve the communication of quantitative probabilities. Those probabilities are of greatest direct use to individuals who face well-formulated decisions in which quantitative estimates of a health risk (or benefit) play clearly defined roles. For example, a couple explicitly planning their family size need to know the probability of success and of side-effects for whichever contraceptive strategies they will consider. Or, a home-owner poised to decide whether to test for radon needs quantitative estimates of the cost and accuracy of tests, the health risks of different radon levels, the cost and efficacy of ways to mitigate radon problems, and so on (Svenson and Fischhoff 1985).
Often, however, people are not poised to decide anything. Rather, they just want to know what the risk is and how it works. Such substantive knowledge is essential for following an issue in the news media, participating in public discussions, feeling competent to make decisions, and generating options among which to decide. In these situations, people’s objective is to have intuitive theories that correspond to the main elements of the reigning scientific theories (emphasizing those features relevant to control strategies).
The term ‘mental model’ is often applied to intuitive theories that are sufficiently well elaborated to generate predictions in diverse circumstances (Galotti 1989). Mental models have a long history in psychology (Craik 1943; Johnson-Laird 1983; Oden 1987), having been used in such diverse settings as uncovering how people understand physical processes (Gentner and Stevens 1983), international tensions (Means and Voss 1985), complex equipment (Rouse and Morris 1986), energy conservation (Kempton 1987), psychological interactions (Furnham 1988), and the effects of drugs (Jungermann et al. 1988; Slovic et al. 1989).
If these mental models contain critical ‘bugs’, they can lead to erroneous conclusions, even among people who are otherwise well informed. For example, not knowing that repeated sex increases the associated risks could undermine much other knowledge. Bostrom et al. (1992) found that many people know that radon is a colourless odourless radio-active gas. Unfortunately, people also associate radio-activity with permanent contamination. However, this property of (widely publicized) high-level waste is not shared by radon. Not realizing that the relevant radon byproducts have short half-lives, home-owners might not even bother to test (believing that there was nothing that they could do, should a problem be detected). They might also not appreciate the risk in minute concentrations, which release their energy quickly.
Eliciting mental models
In principle, the best way to detect such misconceptions would be to capture people’s entire mental model on a topic. Doing so would also identify those correct beliefs upon which communications could be built (and which should be reinforced). The critical methodological threat to capturing mental models is reactivity—changing responses as a result of the elicitation procedure. One wants neither to induce nor to dispel misconceptions, either through leading questions or subtle hints. The interview should neither preclude the expression of unanticipated beliefs nor inadvertently steer subjects around topics (Ericsson and Simon 1980; Galotti 1989; Hendrickx 1991).
Bostrom et al. (1992) offer one possible compromise strategy, which has been used for a variety of risks, including HIV/AIDS, other sexually transmitted diseases, vehicle insurance, mammography, Lyme disease, paint stripper, Cryptosporidium, and nuclear energy sources in space (Kempton 1991; Maharik and Fischhoff 1992; Fischhoff 1996b, 1999b; Fischhoff et al. 1998; Morgan et al., in press). Their interview protocol begins with very open-ended questions, asking subjects what they know about a topic, then prompting them to consider exposure, effects, and mitigation issues. These basic categories seemed so essential that mentioning them would correct an oversight, rather than introduce a foreign concept. Subjects are asked to elaborate on every topic that they mention, and then to elaborate on those elaborations. Once these minimally structured tasks are exhausted, subjects sort a large stack of diverse photographs according to whether each seems related to the topic, explaining their reasoning as they go. When previously unmentioned beliefs appear at this stage, they are likely to represent latent portions of people’s mental models—the sort that might emerge in everyday life if people had cause to consider specific features of their own radon situation. For example, when shown a picture of supermarket produce counter, some respondents told us that plants might become contaminated by taking up radon from the air or soil. Some also inferred that their houseplants would not be so healthy if they had a radon problem.
Once transcribed, interviews are coded into an expert model of the risk. This is a directed network, or influence diagram (Howard 1989; Burns and Clemen 1993), showing the different factors affecting the magnitude of the risk. The expert model is created by iteratively pooling the knowledge of a diverse group of experts, using appropriate elicitation procedures (Fischhoff 1989a; Morgan and Henrion 1990; Kammen and Hassenzahl 1999). It might be thought of as an expert’s mental model, although it would be impressive for any single expert to have such comprehensive knowledge (e.g. about the factors involved in both lung clearance and building materials emissions). Moreover, laypeople can also be sources of expertise (e.g. about their own behaviour (useful for estimating exposure patterns), about side-effects that have yet to be established in the scientific literature, or about how well equipment actually works).
Figure 5 shows a portion of our influence diagram for radon, focused on reducing the risks in a house with a crawl space. An arrow between nodes indicates that the value of the variable at its head depends on the value of the variable at its tail. Thus, for example, the lungs’ particle clearance rate depends on an individual’s smoking history. Influence diagrams are convenient ways to display the functional relationships among variables. Their structure allows, in principle, the substitution of quantitative estimates of these relationships and to compute risk levels. Influence diagrams can also be mapped into decision trees, showing the relevance of various facts for decision-making (which can, in turn, provide guidance on the critical question of which are most worth communicating).
Fig. 5 Expert influence diagram for health effects of radon in a home with a crawl space. This diagram was used as a standard and as an organizing device to characterize the content of lay mental models. (Source: Morgan et al. 1992.)
Such a model provides a template for characterizing lay mental models in communication-relevant terms. Once mapped into the expert model, lay beliefs can be analysed in terms of their appropriateness, specificity (i.e. level of detail), and focus. For most risks, beliefs can be categorized as pertaining to exposure processes, effects processes (i.e. health and physiology), and mitigation behaviours—the basic components of risk analysis. Other beliefs provide background information, which influences interpreting many of the relations in the diagram (e.g. radon is a gas). In evaluating appropriateness, we characterized beliefs as accurate, erroneous, peripheral (correct, but not relevant), or indiscriminate (too imprecise to be evaluated). Bostrom et al. (1992) found that most subjects knew that radon concentrates indoors (92 per cent mentioned), is detectable with a test kit (96 per cent), is a gas (88 per cent), and comes from underground (83 per cent). Most knew that radon causes cancer (63 per cent). However, many also believed erroneously that radon affects plants (58 per cent), contaminates blood (38 per cent), and causes breast cancer (29 per cent). Only two subjects (8 per cent) mentioned that radon decays. (Subjects were drawn from civic groups in the Pittsburgh area, which had a moderate degree of radon publicity.) The robustness of these beliefs was examined (and generally confirmed) in subsequent studies using more easily administered structured questionnaires derived from the open-ended interviews.
The first step in designing communications is to select the information that they should contain. In many existing communications, this choice seems arbitrary, reflecting some expert’s notion of ‘what people ought to know’. Poorly chosen information can have several negative consequences: it can waste recipients’ time, it can be seen as wasting their time (indicating insensitivity to their situation), it can take up the place (in the media or school) that could be filled with pertinent information (imposing an opportunity cost), and it can lead them to misunderstand the extent of their knowledge. In addition, recipients may be judged unduly harshly if they are uninterested in information that seems irrelevant to them, but has been deemed significant by the experts. The authors of the Institute of Medicine’s important report Confronting AIDS (1986) despaired after a survey showed that only 41 per cent of the public knew that AIDS was caused by a virus. Yet, one might ask what role that information could play in any practical decision (as well as what those subjects who answered correctly meant by ‘a virus’).
The information in a communication should reflect a systematic theoretical perspective, capable of being applied objectively. Below, three candidates are listed for such a perspective, suggested by the research cited above.
Mental model analysis
Communications could attempt to convey a comprehensive picture of the processes creating (and controlling) a risk. Bridging the gap between lay mental models and expert models would require adding missing concepts, correcting mistakes, strengthening correct beliefs, and de-emphasizing peripheral ones. Following the mental model procedure outlined above has several potential advantages: (a) it allows the emergence of lay beliefs that never would have occurred to an expert (e.g. plants are sensitive to radon concentrations); (b) it reduces the chances of omitting critical concepts, by disciplining the experts to define their universe of expertise in terms of the influence diagram; (c) it reduces the clutter created by peripheral information that is routinely included in messages, without much thought to its role; (d) it increases the chances of revealing the terms in which laypeople express their beliefs.
Communications could attempt to give recipients the appropriate degree of confidence in their beliefs. They would focus on cases where people confidently hold incorrect beliefs that could lead to inappropriate actions or lack the confidence in correct beliefs needed to act on them. For example, only 45 per cent of the high-risk teenagers in Quadrel’s (1990) study knew that having a beer would affect their driving as much as drinking a shot of vodka. However, they were, on average, very confident in their (usually wrong) answers. For this particular question, the adults were just as overconfident as the high-risk youth, whereas the low-risk teenagers judged their chances of a correct response more realistically. Such local misconceptions or ‘bugs’ can undermine otherwise correct beliefs, and hence deserve focused attention in communications.
Those who provide information have an obligation to communicate how much confidence should be placed in it. For example, Fortney (1988) reported the results of a meta-analysis on all the then available studies of the health effects of oral contraceptives. She concluded, with great confidence, that the effect of contraceptive pill use was somewhere between increasing a woman’s life expectancy by 4 days and decreasing it by 80 days (for a non-smoker, using it throughout her reproductive career). Fortney could also say that it was highly unlikely that this forecast would change; the existing research base was so large that no conceivable single additional study could materially change the conclusions. Such an explicit estimation of uncertainty is much more valuable than any verbal summary. Unfortunately, individuals are all too likely to be left guessing at the definitiveness of the studies reported in a typical popular account. Reporting results responsibly is a continuing problem for the scientific community.
Communications could attempt to provide the pieces of information having the largest possible impact on pending decisions. Value-of-information analysis is the general term for techniques determining the sensitivity of decisions to different information (Raiffa 1968).
Merz (1991) applied value-of-information analysis to a well-specified medical decision—whether to undergo carotid endarterectomy. Both this procedure, which involves scraping out an artery leading to the head, and its alternatives have a variety of possible positive and negative effects. These effects have been the topic of extensive research, providing quantitative risk estimates of varying precision. Merz created a simulated population of patients, varying in their physical conditions and preferences for different health states. He found that knowing about a few, but only a few, of the possible side-effects would change the preferred decision for a significant portion of patients. He argued that communications should focus on these few side-effects; doing so would make better use of patients’ limited attention than a laundry lists of possibilities (although none of those should be hidden). He also argued that his procedure could provide an objective criterion for identifying the information that must be transmitted (and understood) in order to ensure medical informed consent.
Between the time that Merz (1991) submitted his dissertation and its defence, the results of a major clinical trial were released. Incorporating them in his model made little difference to its conclusions (Merz et al. 1993), i.e. from this perspective, information produced by the trial had little practical importance for determining the advisability of the surgery. This is not to say that the study did not contribute to the understanding of fundamental physiological processes, or that it might not have produced other results that would have been more useful to patients. However, the results give pause for thought regarding the allocation of research resources. Thus, value-of-information analysis can be used for prioritizing the scientific information to be collected as well as that to be transmitted (Fischhoff 2000). For example, it has been applied to the testing of chemicals for carcinogenicity (National Research Council 1983; Lave and Omenn 1986).
The choice among these approaches would depend on, among other things, how much time is available for communication, how well the decisions are formulated, and what scientific risk information exists. For example, value-of-information analysis might be particularly useful for identifying the focal facts for public service announcements. Calibration analysis may be used to identify surprising facts, of the sort that might both grab recipients’ attention and change their behaviour. A mental model analysis might be more suited for the preparation of explanatory brochures or curricula.
Once information has been selected, it must be presented in a comprehensible way. That means taking into account the terms that recipients use for understanding individual concepts and the mental models that they use for integrating those concepts. It also means building on the results of research on text comprehension. That research shows, for example, (a) that comprehension improves when text has a clear structure and especially when that structure conforms to recipients’ intuitive representation of a topic, (b) that critical information is more likely to be remembered when it appears at the highest level of a clear hierarchy, and (c) that readers benefit from ‘adjunct aids’, such as highlighting, advance organizers (showing what to expect), and summaries. Such aids might be better than full text for understanding, retaining, and being able to look up information. Fuller treatment can be found in sources such as Reder (1985), Kintsch (1986), Garnham (1987), Ericsson (1988), and Schriver (1989).
In a given application, several formats may meet these general constraints. Atman et al. (1994) created two brochures, using clear but different structures for explaining the risks of radon. One was organized around a decision tree, showing the options facing home-owners, the probabilities of possible consequences, and the associated costs or benefits. The second was organized around a directed network, representing, in effect, a simplified version of the expert model partially depicted in Fig. 5. Both brochures were compared with the widely distributed Citizen’s Guide to Radon (US Environmental Protection Agency 1989), which was built around a question-and-answer format with little attempt to summarize or impose a general structure. All three brochures substantially increased readers’ understanding of the material presented in them. However, the structured brochures did better (and similar) jobs of enabling readers to make inferences about issues that were not mentioned explicitly and to give advice to others who had not read the material. To the Environmental Protection Agency’s great credit, its brochure was much more extensively evaluated than the vast majority of public health communications—although without the benefit of these procedures from applied cognitive psychology (Desvousges et al. 1989; Smith et al. 1995).
Effective risk communications can help people to reduce their health risks, or to obtain greater benefits in return for those risks that they take. Ineffective communications can not only fail to do so, but also incur opportunity costs, in the sense of occupying the place (in recipients’ lives and society’s functions) that could be taken up by more effective communications. Even worse, misdirected communications can prompt wrong decisions by omitting key information or failing to contradict misconceptions, create confusion by prompting inappropriate assumptions or emphasizing irrelevant information, and provoke conflict by eroding recipients’ faith in the communicator. By causing undue alarm or complacency, poor communications can have greater public health impact than the risks that they attempt to describe. Because communicators’ intuitions about recipients’ risk perceptions cannot be trusted, there is no substitute for empirical validation (Fischhoff et al. 1983; Fischhoff 1987; Slovic 1987; National Research Council 1989).
The most ambitious evaluations ask whether recipients follow the recommendations given in the communication (Lau et al. 1980; Weinstein 1987). However, that standard requires recipients not only to understand the message, but also to accept it as relevant to their personal circumstances. For example, home-owners without the resources to address radon problems might both understand and ignore a communication advocating testing, and women might hear quite clearly what actions an ‘expert’ recommends for reducing their risk of sexual assault, yet reject the political agenda underlying that advice (Fischhoff 1992). Judging a programme’s effectiveness according to its behavioural effects requires great confidence that one knows what is right for others.
A more modest, but ethically simpler, evaluation criterion is to ensure that recipients have understood what a message was trying to say. That necessary condition might prove sufficient if the recommended action is obviously appropriate—once one knows the facts. Unfortunately, formal evaluations seem to be remarkably rare among the myriad of warning labels, health claims and advisories, public service announcements, operating instructions, and so on encountered in everyday life and work (Laughery et al. 1994).
Evaluating what people take away from a communication faces the same methodological challenges as measuring their ambient risk perceptions. The evaluator wants to avoid changing people’s beliefs through cues embedded in how questions and answers are phrased, restricting the expression of non-expert beliefs, or suppressing the expression of inconsistent beliefs (across questions).
For example, in the course of evaluating its radon risk communications, the Environmental Protection Agency (Desvousges et al. 1989) posed the following question: ‘What kinds of problems are high levels of radon exposure likely to cause? (a) Minor skin problems; (b) eye irritations; (c) lung cancer.’ This question seems to risk inflating subjects’ apparent level of understanding in several ways. Subjects who know only that radon causes cancer might deduce that it causes lung cancer. The words ‘minor’ and ‘irritation’ might imply that these are not the effects of ‘high levels’ (of anything). Moreover, there is no way for subjects to express other misconceptions, such as that radon causes breast cancer or other lung problems (which emerged with some frequency in open-ended interviews) (Bostrom et al. 1992).
Table 2 summarizes approaches to reader-based evaluation. In principle, open-ended interviews provide the best way to reduce such threats. Performing them to the standards of scientific publication is labour intensive. It involves conducting, transcribing, and coding interviews, with suitable reliability checks (in addition to the effort of producing an expert model and determining explicit communication goals). The stakes riding on many risk communications should justify that investment, considering the costs of dissemination and of the ensuing ineffective choices. Realistically, the necessary time and financial resources will not always be available. In some cases, a few open-ended one-on-one interviews might still provide valuable stepping stones to structured tests suitable for mass administration. Those quizzes will cover the critical topics in the expert model, express questions in terms familiar to subjects, and estimate the prevalence of misconceptions. Even if systematic study is impractical, one-on-one interviews, using think-aloud protocols, can be administered quickly, purely for their heuristic value. It is depressing how often this rudimentary precaution is not taken.
Table 2 Data collection options for reader-based evaluations of risk communications
The format of a message conveys priorities. Figure 6 shows the expected inhalation dose of methylene chloride, for individuals performing a common paint-stripping task, in a medium-sized room with moderate ventilation. (Details of the model in general and of this specific application are given by Riley et al. (1998).) Each curve assumes that users read the first five items on the labels of six different products (taken from the shelves of a local hardware store), and then follow those directions perfectly. Clearly, product B does a much better job than the others for users with these reading habits. That advantage vanishes for users who read only the text emphasized on a label or only the directions (with all labels performing equally poorly). For those (rare?) users who read everything, there is virtually no exposure at all with product A, the labels for B, D, and E produce exposures like those for B in the figure, and products C and F perform no better. Reducing risks to such products requires understanding users’ behaviour, reading patterns, and mental models (determining how far they follow the instructions that they read).
Fig. 6 Inhalation dose of methylene chloride for users of paint strippers who read and follow the first five points on the labels of six products.
Understanding risk perception and risk communication is a complicated business, perhaps as complicated as assessing the magnitude of the risks being considered. A chapter of this length can, at best, indicate the dimensions of this complexity and the directions of plausible solutions. In this treatment, we have emphasized methodological issues because we believe that these topics often seem deceptively simple. Because we ask questions in everyday life, eliciting the beliefs of others may seem straightforward; because we talk every day, it may seem simple to communicate about health risks. Unfortunately, there are many pitfalls to such amateurism, some of which emerge in the research described here. Hints at these problems can be found in those occasions in life where we have misunderstood or been misunderstood, particularly when discussing unfamiliar topics with strangers.
Research on these topics is fortunate in being able to draw on well-developed literatures in such areas as cognitive, health, and social psychology, psycholinguistics, psychophysics, and behavioural decision theory. It is unfortunate in having to face the particularly rigorous demands of assessing and improving complex beliefs about health risks. These often involve unfamiliar topics, surrounded by unusual kinds of uncertainty, for which individuals and groups lack stable vocabularies. Health-risk decisions also raise difficult and potentially threatening trade-offs. Even the most carefully prepared and evaluated communications may not be able to eliminate the anxiety and frustration that such decisions create. However, systematic preparation can keep communications from adding to the problem. At some point in complex decisions, we ‘throw up our hands’ and go with what seems right. Good risk communications can help people go further into the problem before that happens.
Health-risk decisions are not just about cognitive processes and coolly weighed information. Emotions play a role, as do social processes. Nonetheless, it is important to get the cognitive part right, lest people’s ability to think their way to decisions be underestimated and underserved. For the resolution of risk issues to hinge on the light of information, rather than the heat of controversy, those managing the risk bear special responsibility to behave in a credible way. If one examines the risks that hit the headlines, it has often been the case that the authorities were slow to realize, or at least acknowledge, that they might be managing a risk that warranted public scrutiny (Fischhoff 1995). Table 3 shows one characterization of the stages that authorities may go through as they gradually (and perhaps begrudgingly) attempt to satisfy their public’s desire to understand, and perhaps participate in the control of, a risk. The energies that the public sees iauthorities investing in generating openness partially determine the credibility of the communications that they receive.
Table 3 Developmental stages in risk management
Table 4 offers one specification of the conditions that risk specialists must meet in order to secure public trust. It is patterned after the conditions that experts must meet in order to secure one another’s trust. It includes both scientific and social conditions, concerning, respectively, the content and the conduct of science. In each domain, there are conditions associated with both each specific case and the general process of analysing risk issues.
Table 4 Conditions for public trust in risk analysis
* Preparation of this chapter was supported by the National Institute of Alcohol Abuse and Alcoholism, the National Institute of Allergies and Infectious Disease, and the National Science Foundation Center for Integrated Assessment of the Human Dimensions of Global Change.
Arnet, J.(1999). Adolescent storm and stress, reconsidered. American Psychologist, 54, 317–26.
Atman, C.J., Bostrom, A., Fischhoff, B., and Morgan, M.G.(1994). Designing risk communications: completing and correcting mental models of hazardous processes, Part 1. Risk Analysis, 14, 779–88.
Bar-Hillel, M.(1974). Similarity and probability. Organizational Behavior and Human Performance, 11, 277–82.
Bauman, K.E.(1980). Predicting adolescent drug use: utility structure and marijuana. Praeger, New York.
Beattie, J., Covey, J., Dolan, P., et al.(1998). On the contingent valuation of safety and the safety of contingent valuation. Journal of Risk and Uncertainty, 17, 5–26.
Beyth-Marom, R.(1982). How probable is probable? Numerical translation of verbal probability expressions. Journal of Forecasting, 1, 257–69.
Beyth-Marom, R., Austin, L., Fischhoff, B., Palmgren, C., and Quadrel, M.J.(1993). Perceived consequences of risky behaviors. Developmental Psychology, 29, 549–63.
Black, W.C., Nease, R.F., and Tosteson, A.N.A.(1995). Perceptions of breast cancer risk and screening effectiveness in women younger than 50 years of age. Journal of the National Cancer Institute, 8, 720–31.
Bostrom, A., Fischhoff, B., and Morgan, M.G.(1992). Characterizing mental models of hazardous processes: A methodology and an application to radon. Journal of Social Issues, 48, 85–100.
Bostrom, A., Atman, C.J., Fischhoff, B., and Morgan, M.G.(1994). Evaluating risk comunications: completing and correcting mental models of hazardous processes. Part 2. Risk Analysis, 14, 789–98.
Bruine de Bruin, W.(1999). People’s understanding of probability: ‘It’s a fifty-fifty chance’. PhD Dissertation, Carnegie–Mellon University.
Budescu, D.F. and Wallsten, T.S.(1995). Processing linguistic probabilities: General principles and empirical evidence. In Decision making from the perspective of cognitive psychology (ed. J.R. Busemeyer, R. Hastie, and D.L. Medin). Academic Press, New York.
Burns, W.J. and Clemen, R.T.(1993). Covariance structure models and influence diagrams. Management Science, 39, 816–34.
Camerer, C.(1981). General conditions for the success of bootstrapping models. Organizational Behavior and Human Performance, 27, 411–22.
Cohen, J. and Hansel, C.E.M.(1958). The nature of decisions in gambling. Acta Psychologica, 13, 357–70.
Cole, G. and Withey, S.(1982). Perspectives in risk perceptions. Risk Analysis, 1, 143–63.
Covello, V.T., Sandman, P.M., and Slovic, P.(1988). Risk communication, risk statistics, and risk comparisons: A manual for plant managers. Chemical Manufacturers Association, Washington, DC.
Craik, K.(1943). The nature of explanation. Cambridge University Press.
Crouch, E.A.C. and Wilson, R.(1982). Risk/benefit analysis. Ballinger, Cambridge, MA.
Cvetkovich, G., Grote, B., Bjorseth, A., and Sarkissian, J.(1975). On the psychology of adolescents’ use of contraceptives. Journal of Sex Research, 11, 256–70.
Davies, J.C.(ed.) (1996). Comparing environmental risks. Resources for the Future, Washington, DC.
Dawes, R.M. and Corrigan, B.(1974). Linear models in decision making. Psychological Bulletin, 81, 95–106.
Dawes, R.M., Faust, D., and Meehl, P.(1989). Clinical versus actuarial judgment. Science, 243, 1668–74.
Desvousges, W.H., Smith, V.K., and Rink III H.H.(1989). Communicating radon risk effectively: radon testing in Maryland. EPA 230-03-89-408. US Environmental Protection Agency, Washington, DC.
Douglas, M.(1992). Risk and blame. Routledge, London.
Elkind, D.(1967). Egocentrism in adolescence. Child Development, 38, 1025–39.
Ënglander, T., Farago, K., Slovic, P., and Fischhoff, B.(1986). A comparative analysis of risk perception in Hungary and the United States. Social Behavior, 1, 55–6.
Erev, I. and Cohen, B.L.(1990). Verbal versus numerical probabilities: efficiency, biases and the preference paradox. Organizational Behavior and Human Decision Processes, 45, 1–18.
Ericsson, K.A.(1988). Concurrent verbal reports on text comprehension: a review. Text, 8, 295–325.
Ericsson, K.A. and Simon, H.A.(1980). Verbal reports as data. Psychological Review, 87, 215–51.
Feather, N.(ed.) (1982). Expectancy, incentive and action. Erlbaum, Hillsdale, NJ.
Fineberg, H.V.(1988). Education to prevent AIDS. Science, 239, 592–6.
Fischhoff, B.(1987). Treating the public with risk communications: A public health perspective. Science, Technology, and Human Values, 12, 13–19.
Fischhoff, B.(1989a). Eliciting knowledge for analytical representation. IEEE Transactions on Systems, Man and Cybernetics, 13, 448–61.
Fischhoff, B.(1989b). Making decisions about AIDS. In Primary prevention of AIDS (ed. V. Mays, G. Albee, and S. Schneider), pp. 168–205. Sage, Newbury Park, CA.
Fischhoff, B.(1990). Psychology and public policy: tool or tool maker? American Psychologist, 45, 57–63.
Fischhoff, B.(1991). Value elicitation: is there anything in there? American Psychologist, 46, 835–47.
Fischhoff, B.(1992). Giving advice: decision theory perspectives on sexual assault. American Psychologist, 47, 577–88.
Fischhoff, B.(1994). What forecasts (seem to) mean. International Journal of Forecasting, 10, 387–403.
Fischhoff, B.(1995). Risk perception and communication unplugged: twenty years of process. Risk Analysis, 15, 137–45.
Fischhoff, B.(1996a). The real world: what good is it? Organizational Behavior and Human Decision Processes, 65, 232–48.
Fischhoff, B.(1996b). Public values in risk research. Annals of the American Academy of Political and Social Science, 545, 75–84.
Fischhoff, B.(1999). Why (cancer) risk communication can be hard. Journal of the National Cancer Institute Monographs, 25, 7–13.
Fischhoff, B.(2000). Scientific management of science? Policy Sciences, 29, 1–15.
Fischhoff, B. and Bruine de Bruin, W.(1999). Fifty/fifty = 50? Journal of Behavioral Decision Making, 12, 149–63.
Fischhoff, B. and MacGregor, D.(1983). Judged lethality: how much people seem to know depends upon how they are asked. Risk Analysis, 3, 229–36.
Fischhoff, B. and Quadrel, M.J.(1991). Adolescent alcohol decisions. Alcohol Health and Research World, 15, 43–51.
Fischhoff, B., Slovic, P., and Lichtenstein, S.(1977). Knowing with certainty: the appropriateness of extreme confidence. Journal of Experimental Psychology: Human Perception and Performance, 3, 552–64.
Fischhoff, B., Slovic, P., Lichtenstein, S., Read, S., and Combs, B.(1978). How safe is safe enough? A psychometric study of attitudes towards technological risks and benefits. Policy Sciences, 8, 127–52.
Fischhoff, B., Lichtenstein, S., Slovic, P., Derby, S.L., and Keeney, R.L.(1981). Acceptable risk. Cambridge University Press, New York.
Fischhoff, B., Slovic, P., and Lichtenstein, S.(1983). The ‘public’ vs. the ‘experts’: Perceived vs. actual disagreement about the risks of nuclear power. In Analysis of actual vs. perceived risks (ed. V. Covello, G. Flamm, J. Rodericks, and R. Tardiff), pp. 235–49. Plenum Press, New York.
Fischhoff, B., Watson, S., and Hope, C.(1984). Defining risk. Policy Sciences, 17, 123–39.
Fischhoff, B., Downs, J., and Bruine de Bruin, W.(1998). Adolescent vulnerability: A framework for behavioral interventions. Applied and Preventive Psychology, 7, 77–94.
Fischhoff, B., Welch, N., and Frederick, S.(1999). Construal processes in preference elicitation. Journal of Risk and Uncertainty, 19, 139–64.
Fortney, J.(1988). Contraception: a life long perspective. In Dying for love. National Council for International Health, Washington, DC.
Funtowicz, S.O. and Ravetz, J.R.(1990). Uncertainty and quality in science for policy. Kluwer, Boston, MA.
Furnham, A.F.(1988). Lay theories. Pergamon, Oxford.
Galotti, K.M.(1989). Approaches to studying formal and everyday reasoning. Psychological Bulletin, 105, 331–51.
Gärdenfors, P. and Sahlin, N.E.(1982). Unreliable probabilities, risk taking and decision making. Synthese, 53, 361–86.
Garnham, A.(1987). Mental models as representations of discourse and text. Halsted Press, New York.
Gentner, D. and Stevens, A.L.(ed.) (1983). Mental models. Erlbaum, Hillsdale, NJ.
Goldberg, L.R.(1968). Simple models or simple processes? American Psychologist, 23, 483–96.
Goszczynska, M., Tyszka, T., and Slovic, P.(1991). Risk perception in Poland: A comparison with three other countries. Journal of Behavioral Decision Making, 4, 179–93.
Green, C.H. and Brown, R.A.(1981). The perception and acceptability of risk. Duncan of Jordanstone School of Architecture, Dundee.
Gregory, R., Lichtenstein, S., and Slovic, P.(1993) Valuing environmental resources: A constructive approach. Journal of Risk and Uncertainty, 7, 177–97.
Hammond, K.R., Hursch, C.J., and Toddy, F.J.(1964). Analyzing the components of clinical inference. Psychological Review, 71, 438–56.
Harding, C.M. and Eiser, J.R.(1984). Characterizing the perceived risks and benefits of some health issues. Risk Analysis, 4, 131–41.
Hasher, L. and Zacks, R.T.(1984). Automatic processing of fundamental information. American Psychologist, 39, 1372–88.
Heimer, C.A.(1988). Social structure, psychology, and the estimation of risk. Annual Review of Sociology, 14, 491–519.
Hendrickx, L.C.W.P.(1991). How versus how often: the role of scenario information and frequency information in risk judgment and risky decision making. Doctoral Dissertation, Rijksuniversiteit Groningen, The Netherlands.
Hoffman, P.J.(1960). Paramorphic models representation of clinical judgment. Psychological Bulletin, 57, 116–31.
Howard, R.A.(1989). Knowledge maps. Management Science, 35, 903–22.
Hurd, M.D.(1999). Anchoring an acquiesence bias in measuring assets in household surveys. Journal of Risk and Uncertainty, 19, 111–36.
Inhaber, H.(1979). Risk with energy from conventional and non-conventional sources. Science, 203, 718–23.
Institute of Medicine(1986). Confronting AIDS. National Academy Press, Washington, DC.
Institute of Medicine(1998). Scientific opportunities and public needs. National Academy Press, Washington, DC.
Jasanoff, S.(1986). Risk management and political culture. Russell Sage Foundation, New York.
Jenni, K.(1997). Attributes for risk evaluation. PhD Dissertation, Carnegie–Mellon University.
Jianguang, Z.(1994). Environmental hazards in the Chinese public’s eyes. Risk Analysis, 14, 163–9.
Johnson-Laird, P.N.(1983). Mental models. Cambridge University Press, New York.
Jungermann, H., Schutz, H., and Thüring, M.(1988). Mental models in risk assessment: Informing people about drugs. Risk Analysis, 8, 147–55.
Kahneman, D., Slovic, P., and Tversky, A.(ed.) (1982). Judgment under uncertainty: heuristics and biases. Cambridge University Press, New York.
Kahneman, D., Knetsch, J.L., and Thaler, R.H.(1991). The endowment effect, loss aversion,and status quo bias. Journal of Economic Perspectives, 5, 193–206.
Kammen, D. and Hassenzahl, D.(1999). Should we risk it? Princeton University Press.
Kaplan, E.H.(1989). What are the risks of risky sex? Operations Research, 37, 198–209.
Karpowicz-Lazreg, C. and Mullet, E.(1993). Societal risk as seen by the French public. Risk Analysis, 13, 253–8.
Keating, D.P.(1988). Cognitive processes in adolescence. Ontario Institute for Studies in Education, Toronto.
Kempton, W.(1987). Variation in folk models and consequent behavior. American Behavioral Scientist, 31, 203–18.
Kempton, W.(1991). Lay perspectives on global climate change. Global Environmental Change, 1, 183–208.
Kintsch, W.(1986). Learning from text. Cognition and Instruction, 3, 87–108.
Koriat, A.(1993). How do we know that we know? Psychological Review, 100, 609–39.
Krimsky, S. and Golding, D.(ed.) (1992). Theories of risk. Praeger, New York.
Krimsky, S. and Plough, A.(1988). Environmental hazards. Auburn House, Dover, MA.
Kuyper, H. and Vlek, C.(1984). Contrasting risk judgements among interest groups. Acta Psychologica, 56, 205–18.
Lau, R., Kaine, R., Berry, S., Ware, J., and Roy, D.(1980). Channeling health: a review of the evaluation of televised health campaigns. Health Education Quarterly, 7, 56–89.
Laughery, K.S., Wogalter, M.S., and Young, S.L.(ed.) (1994). Human factors perspectives on warnings. Human Factors and Ergonomics Society, Santa Monica, CA.
Lave, L.B. and Omenn, G.S.(1986). Costs effectiveness of short-term tests for carcinogenicity. Nature, 324, 29–39.
Leiss, W. and Chociolko, C.(1994). Risk and responsibility. McGill–Queens University Press, Montreal and Kingston.
Lichtenstein, S. and Newman, J.R.(1967). Empirical scaling of common verbal phrases associated with numerical probabilities. Psychonomic Science, 9, 563–4.
Lichtenstein, S., Slovic, P., Fischhoff, B., Layman, M., and Combs, B.(1978). Judged frequency of lethal events. Journal of Experimental Psychology: Human Learning and Memory, 4, 551–78.
Lichtenstein, S., Fischhoff, B., and Phillips, L.D.(1982). Calibration of probabilities: state of the art to 1980. InJudgment under uncertainty: heuristics and biases (ed. D. Kahneman, P. Slovic, and A. Tversky), pp. 306–39. Cambridge University Press, New York.
Linville, P.W., Fischer, G.W., and Fischhoff, B.(1993). AIDS risk perceptions and decision biases. In The social psychology of HIV infection (ed. J.B. Pryor and G.D. Reeder), pp. 5–38. Erlbaum, Hillsdale, NJ.
Lowrance, W.(1976). Of acceptable risk. Kaufmann, San Francisco, CA.
MacGregor, D.(1991). Worry over technological activities and life concerns. Risk Analysis, 11, 315–25.
McIntyre, S. and West, P.(1992). What does the phrase ‘safer sex’ mean to you? Understanding among Glaswegian 18 year olds in 1990. AIDS, 7, 121–6.
Maharik, M. and Fischhoff, B.(1992). The risks of nuclear energy sources in space: some activists’ perceptions. Risk Analysis, 12, 383–92.
Means, M.L. and Voss, J.F.(1985). Star Wars: a developmental study of expert and novice knowledge studies. Journal of Memory and Language, 24, 746–57.
Merton, R.F.(1987). The focussed interview and focus groups. Public Opinion Quarterly, 51, 550–66.
Merz, J.F.(1991). Toward a standard of disclosure for medical informed consent: Development and demonstration of a decision—analytic methodology. PhD dissertation, Carnegie–Mellon University.
Merz, J., Druzdzel, M., and Mazur, D.J.(1991). Verbal expressions of probability in informed consent litigation. Medical Decision Making, 11, 273–81.
Merz, J., Fischhoff, B., Mazur, D.J., and Fischbeck, P.S.(1993). Decision-analytic approach to developing standards of disclosure for medical informed consent. Journal of Toxics and Liability, 15, 191–215.
Morgan, M.G. and Henrion, M.(1990). Uncertainty. Cambridge University Press, New York.
Morgan, M.G., Fischhoff, B., Bostrom, A., Lave, L., and Atman, C.J.(1992). Communicating risk to the public. Environmental Science and Technology, 26, 2048–56.
Morgan, M.G., Fischhoff, B., Bostrom, A., and Atman, C. Risk communication: the mental model approach. Cambridge University Press, New York, in press.
Morrison, D.M.(1985). Adolescent contraceptive behavior: a review. Psychological Bulletin, 98, 538–68.
Murphy, A.H., Lichtenstein, S., Fischhoff, B., and Winkler, R.L.(1980). Misinterpretations of precipitation probability forecasts. Bulletin of the American Meteorological Society, 61, 695–701.
National Research Council(1983). Research needs for human factors. National Academy Press, Washington, DC.
National Research Council(1989). Improving risk communication. National Academy Press, Washington, DC.
National Research Council(1996). Understanding risk: Informing decisions in a democratic society. National Academy Press, Washington, DC.
National Research Council(1999). Toward environmental justice. National Academy Press, Washington, DC.
Oden, G.C.(1987). Concept, knowledge, and thought. Annual Review of Psychology, 38, 203–27.
Okrent, D. and Pidgeon, N.(ed.) (1998). Risk perception versus risk analysis (Special Issue). Reliability Engineering and System Safety, 58.
Otway, H.J. and Wynne, B.(1989). Risk communication: paradigm and paradox. Risk Analysis, 9, 141–5.
Payne, J.W., Bettman, J.R., and Schkade, D.A.(1999). Measuring constructed preferences: towards a building code. Journal of Risk and Uncertainty. 19, 243–70.
Poulton, E.C.(1989). Bias in quantifying judgment. Erlbaum, Hillsdale, NJ.
Quadrel, M.J.(1990). Elicitation of adolescents’ risk perceptions: qualitative and quantitative dimensions. PhD dissertation, Carnegie–Mellon University.
Quadrel, M.J., Fischhoff, B., and Davis, W.(1993). Adolescent (in)vulnerability. American Psychologist, 48, 102–16.
Raiffa, H.(1968). Decision analysis: introductory lectures on choices under uncertainty. Addison-Wesley, Reading, MA.
Reder, L.M.(1985). Techniques available to author, teacher, and reader to improve retention of main ideas of a chapter. In Thinking and learning skills. Vol. 2,Research and open questions (ed. S.F. Chipman, J.W. Segal, and R. Glaser), pp. 37–64. Erlbaum, Hillsdale, NJ.
Riley, D., Fischoff, B., Small, M., and Fishbeck, P.S.(2001). Evaluating the effectiveness of risk-reduction strategies for chemical products. Risk Analysis, 21, 357–69.
Ritov, I. and Baron, J.(1990). Status quo and omission bias. Reluctance to vaccinate. Journal of Behavioral Decision Making, 3, 263–77.
Rohrmann, B.(1994). Risk perception of different societal groups: Australian findings and cross-national comparisons. Australian Journal of Psychology, 46, 15–163.
Roth, E., Morgan, G., Fischhoff, B., Lave, L., and Bostrom, A.(1990). What do we know about making risk comparisons? Risk Analysis, 10, 375–87.
Rouse, W.B. and Morris N.M.(1986). On looking into the black box: prospects and limits in the search for mental models. Psychological Bulletin, 100, 399–463.
Royal Society(1992). Risk analysis, perception and management. Royal Society, London.
Schriver, K.A.(1989). Evaluating text quality: the continuum from text-focused to reader-focused methods. IEEE Transactions on Professional Communication, 32, 238–55.
Schulenberg, J., Maggs, J., and Hurnelmans, K.(ed.) (1997). Health risks and developmental transaction during adolescence. Cambridge University Press, New York.
Schwarz, N.(1996). Cognition and communication: judgmental biases, research methods and the logic of conversation. Erlbaum, Hillsdale, NJ.
Schwarz, N.(1999). Self reports. American Psychologist, 54, 93–105.
Shaklee, H. and Fischhoff, B.(1990). The psychology of contraceptive surprises: judging the cumulative risk of contraceptive failure. Journal of Applied Psychology, 20, 385–403.
Slovic, P.(1987). Perceptions of risk. Science,236, 280–5.
Slovic, P., Fischhoff, B., and Lichtenstein, S.(1979). Rating the risks. Environment, 21, 14–20, 36–9.
Slovic, P., Fischhoff, B., and Lichtenstein, S.(1980). Facts and fears: understanding perceived risk. InSocietal risk assessment: How safe is safe enough? (ed. R. Schwing and W.A. Albers Jr), pp. 181–214. Plenum Press, New York.
Slovic, P., Fischhoff, B., and Lichtenstein, S.(1984). Behavioral decision theory perspectives on risk and safety. Acta Psychologica, 56, 183–203.
Slovic, P., Fischhoff, B., and Lichtenstein, S.(1985). Characterizing perceived risk. InPerilous progress: Technology as hazard (ed. R.W. Kates, C. Hohenemser, and J. Kasperson), pp. 91–123. Westview, Boulder, CO.
Slovic, P., Kraus, N.N., Lappe, H., Letzel, H., and Malmfors, T.(1989). Risk perception of prescription drugs: report on a survey in Sweden. Pharmaceutical Medicine, 4, 43–65.
Smith, V.K., Desvousages W.H., and Payne, J.W.(1995). Do risk information programs promote mitigating behavior? Journal of Risk and Uncertainty, 10, 203–21.
Stallen, P.J.M. and Tomas, A.(1988). Public concerns about industrial hazards. Risk Analysis, 8, 235–45.
Starr, C.(1969). Societal benefit versus technological risk. Science, 165, 1232–8.
Svenson, O.(1981). Are we all less risky and more skillful than our fellow drivers? Acta Psychologica, 47, 143–8.
Svenson, O. and Fischhoff, B.(1985). Levels of nvironmental decisions. Journal of Environmental Psychology, 5, 55–68.
Thaler, R.(1991). Quasi-rational economics. Russell Sage Foundation, New York.
Tversky, A. and Kahneman, D.(1973). Availability: A heuristic for judging frequency and probability. Cognitive Psychology, 4, 207–32.
US Environmental Protection Agency(1989). A citizen’s guide to radon. US Environmental Protection Agency, Washington, DC.
US Environmental Protection Agency(1993). A guidebook to comparing risks and setting environmental priorities. US Environmental Protection Agency, Washington, DC.
Vaughan, E.(1993). Individual and cultural differences in adaptation to environmental risks. American Psychologist, 48, 1–8.
Viscusi, K.(1992). Smoking: Making the risky decision. Oxford University Press, New York.
Vlek, C. and Stallen, P.J.(1980). Rational and personal aspects of risk. Acta Psychologica, 45, 273–300.
Vlek, C. and Stallen, P.J.(1981). Judging risks and benefits in the small and in the large. Organizational Behavior and Human Performance,28, 235–71.
Wagenaar, W.A. and Sagaria, S.D.(1975). Misperception of exponential growth. Perception and Psychophysics, 18, 416–22.
Weinstein, N.(1987). Taking care: understanding and encouraging self-protective behavior. Cambridge University Press, New York.
Weinstein, N.D.(1989). Effects of personal experience on self-protective behavior. Psychological Bulletin, 105, 31–50.
Weinstein, N.D. and Klein, W.M.(1996). Unrealistic optimism: present and future. Journal of Social and Clinical Psychology, 15, 1–8.
Wilks, S.S.(1938). Weighting systems for linear functions of correlated variable where there is no dependent variable. Psychometrika, 8, 23–40.
Wilson, R.(1979). Analyzing the risks of everyday life. Technology Review, 81, 40–6.
Wilson, R.W. and Thornberry, O.T.(1987). Knowledge and attitudes about AIDS: provisional data from the National Health Interview Survey, August 10–30 1987. Advance Data, No. 146.
Woloshin, S., Schwartz, L.M., Byram, S., Fischhoff, B., and Welch, H.G.(1998). Scales for assessing perceptions of event probability: a validation study. Medical Decision Making, 14, 490–503.
Yates, J.F.(1989). Judgment and decision making. Prentice Hall, Englewood Cliffs, NJ.
Yates, J.F.(ed.) (1992). Risk taking. Wiley, Chichester.