Framing effects in disaster risk communication: the case of coastal erosion in the United States

ABSTRACT Governments have the duty to protect their citizens and to prevent natural hazards from turning into disasters. Therefore, they need to communicate effectively about disaster risk to potentially affected populations. We conducted three online survey experiments among United States residents (total N = 1673) to investigate how satellite images combined with information on past events of coastal erosion affect potential framing effects inherent to disaster risk communication. We found a strong framing effect in a decision scenario about human lives, which was reduced when adding an icon array to show the scenario. However, when validating these findings in a more realistic scenario in the context of coastal erosion, we found no framing effects in a risky choice situation and a goal framing setup. Adding satellite images to the textual description of the scenario made respondents more risk seeking and increased their stated behavioural intention to take preventive measures against coastal erosion. The overall awareness among respondents about the issue of coastal erosion was high, resulting in few risky choices and high levels of stated behavioural intentions. Our findings cast doubt on the strength of framing effects in real world scenarios, especially when people hold strong preferences about the issue at stake.


Introduction
Coastal erosion causes around $500 million per year in coastal property loss in the U.S., including damage to structures and loss of land (NOAA, 2013).To mitigate coastal erosion, the U.S. government spends an average of $150 million every year on beach nourishment and other shoreline erosion control measures (NOAA, 2013).Sea-level-riserelated coastal erosion can be linked to climate change which affects the intensity, duration and frequency of many atmospheric, hydrologic, and biologic hazards (IPCC, 2018(IPCC, , 2021)).
Most of these environmental hazards pose a significant risk of material damage and human harm on affected societies.Governments have an economic interest and a humanitarian duty to prevent these hazards from turning into disasters.To this end, disaster risk management is a systematic approach to identifying, assessing and reducing the risks of disasters.It operates on four stages: (1) prevention and mitigation (i.e.avoiding or limiting adverse impacts of hazards), ( 2) preparedness (i.e.knowledge and capacities to effectively anticipate, respond to and recover from the impacts of hazards), ( 3) response (i.e.emergency services during or in the aftermath of a disaster) and ( 4) recovery (i.e.restoration of disaster-affected communities; definitions adapted from UN, 2009).
For disaster risk management to be effective, communication between authorities and the public is imperative and must occur at all four stages (Faulkner & Ball, 2007).However, the purpose of the communication varies from stage to stage.It is especially different between stages (1) and (2), which comprise the time before the hazard strikes, and stages ( 3) and (4), which cover the time during or after the disaster.This study focuses on stages (1) and ( 2), i.e. prevention and mitigation, and preparedness.
Risk communication has different goals.For instance, Rowan (1991) identifies four common goals: (1) creating awareness about the existence of important phenomena, (2) enhancing understanding of complicated ideas, (3) developing agreement about policy options, and (4) motivating action.In communication pertaining to natural hazards, particular relevance is given to the last goal of motivating action, since adaptive measures taken by individuals who themselves or their property are at risk can significantly decrease material and human losses resulting from a disaster.
Thus, disaster risk communicators often have an interest in motivating behavioural change among the message recipients.Since disasters involve potential human and material losses, the corresponding risk messages are prone to be affected by framing effects.A frame is the way in which a certain situation is depicted, for instance whether the focus is put on what can be gained or on what can be lost.In the case of coastal erosion, an exemplary risk messages is: 'If you build your house within 50 m distance from the coastline, you have a 20% chance of keeping it safe over the next five years.'Alternatively, one could say: 'If you build your house within 50 m distance from the coastline, you have an 80% chance of losing it over the next five years.' The actual information content is identical in both sentences.A rational decision maker should thus not be influenced by the way the information is presented. 1Still, humans perceive the erosion situation differently from the two sentences due to the phenomenon of loss aversion as outlined in the prospect theory developed by Kahneman and Tversky: Humans tend to prefer avoiding a loss over acquiring an equivalent gain.This deviation from the rationality postulate has been called framing effect or framing bias (Kahneman & Tversky, 1979;Tversky & Kahneman, 1981).
Framing effects in risk messages might not be consistent with truly informed decision making (Gigerenzer, 2003;Edwards et al., 2001Edwards et al., , 2002)).Individuals take informed decisions when they use reasoning to process relevant information about the advantages and disadvantages of all possible courses of action taking into account their beliefs (Bekker et al., 1999).Enabling informed decision making in risk situations thus implies presenting the risk information in a way that does not push recipients to favour one option over the other simply by the presentation format.
Different studies have examined ways to reduce framing effects, for example, by making participants list advantages and disadvantages of both choice options (Almashat et al., 2008), by making participants describe the options in their own words (Simon et al., 2004), or by presenting information in visual format instead of as text-only (Garcia-Retamero & Cokely, 2011).These three studies pertain to the sectors of health and social dilemmas, for which framing has been studied extensively (Akl et al., 2011;Gong et al., 2013;Levin et al., 1998).
To the best of our knowledge, only few studies have examined framing effects in the context of disaster risk communication, namely to enhance preparedness for earthquakes in New Zealand (Henrich et al., 2015;McClure & Sibley, 2011;McClure et al., 2009).No study, however, has investigated ways to reduce framing effects in disaster risk communication.It can be debated whether it is actually desirable or not to use framing to motivate action that benefits the society as a whole (see also the related debate around nudging, e.g. in Codagnone et al., 2014;Hukkinen, 2016).In this study, we do not address this normative question, but instead investigate the existence of framing effects in disaster risk communication, as well as potential strategies to reduce such framing effects should they occur.
We argue that framing effects occur if the decision situation is presented as textonly.We expect these framing effects to diminish if a suitable visual is shown alongside the text, since visuals can overshadow the framing contained in the text.To test our arguments, we conducted three online survey experiments in the U.S. (total N = 1673).In experiment 1dealing with a choice about human livesa significant framing effect is present in the text-only survey group but becomes insignificant if an icon array shows the decision situation.In experiments 2 and 3, we applied these findings to disaster risk communication and investigated whether satellite images showing past erosion events can reduce framing effects.Since we did not find framing effects in the text-only groups of these two experiments, we could not formally test whether adding images reduces framing biases.However, we found that respondents who saw the images were both more risk seeking and more willing to take preventive measures than those who saw only a text, but no image.

Theory
In this paper, we investigate two types of valence framing effects, 2 namely risky choice framing and goal framing, as well as potential strategies to reduce them.In a risky choice setting, respondents must choose between two options, a riskless and a twooutcome all-or-nothing option.The outcomes of these options are described either in terms of gains (gain frame) or in terms of losses (loss frame).Tversky and Kahneman (1981) were the first to show that this setup leads to a 'choice reversal' where participants in the gain frame prefer the sure option, whereas they prefer the risky option in the loss frame. 3Most risky choice framing studies present respondents with a hypothetical, rather abstract situation (see Kühberger, 1998 for a review).Therefore, our first hypothesis refers to an abstract scenario resembling those commonly found in the literature: Hypothesis 1a 4 (risky choice framing effectchoice about human lives): Presenting a risky choice about human lives with a focus on the gain makes respondents more risk averse, whereas a focus on the loss makes respondents more risk seeking.
Such framing effects have been found in fields as diverse as health, marketing and economics.This suggests that they are not restricted to specific issue areas, but seem to be induced by the risky nature of the decision situation.Therefore, we argue that a similar effect applies also to more real-world decision situations in the context of natural hazards: Hypothesis 2a (risky choice framing effectnatural hazard situation): Presenting a risky choice in the context of natural hazards with a focus on the gain makes respondents more risk averse, whereas a focus on the loss makes respondents more risk seeking.
In goal framing, the impact of a persuasive message promoting a certain behaviour depends on whether the message stresses either the positive consequences of performing the behaviour or the negative consequences of not performing it (Levin et al., 1998).In an early review of goal framing studies, Levin et al. (1998) find that the loss frame generally has a stronger impact on changing behaviour than the gain frame.They explain this difference by the phenomenon of loss aversion, meaning that people are more likely to perform a behaviour to avoid a loss, than they are to obtain an equivalent gain.
Garcia-Retamero and Cokely (2011), however, present a more nuanced view: Whether the gain or the loss frame is more effective depends on the promoted behaviour.Gain frames are more persuasive if the promoted behaviour is perceived as risk averse, meaning that it involves a low risk of an unpleasant outcome (e.g. it prevents the onset of health problems).If, by contrast, the risk of an unpleasant outcome is high (e.g. it detects a health problem), loss frames are more persuasive.This differentiation follows the same logic as the one for risky choice framing: Gain frames make people more risk averse, loss frames more risk seeking.
In the case of disaster risk communication promoting preventive behaviour, there is a low chance that the behaviour evokes a negative outcome for the actor.We thus hypothesise that behaviour preventing the negative outcome of a natural hazard is perceived as risk averse and that, consequently, the gain frame is more persuasive: Hypothesis 3a (goal framing effect): Presenting disaster risk information promoting preventive behavior with a focus on the positive consequences of adopting the behavior makes respondents more likely to state that they would adopt preventive measures than presenting it with a focus on the negative consequences of not adopting the behavior.
Different studies have examined ways to reduce framing effects (Almashat et al., 2008;Garcia-Retamero & Cokely, 2011;Simon et al., 2004).In this study, we focus on visual aids which have been found to reduce framing biases for both goal framing (Garcia-Retamero & Cokely, 2011) and attribute framing 5 (Gamliel & Kreiner, 2013;Garcia-Retamero & Galesic, 2010).For instance, Garcia-Retamero and Cokely (2011) show this in the health context by adding a bar chart illustrating the probabilities communicated in the information text.In this case, the graphic itself did not contain a frame, but only the accompanying text.They argue that visual aids improve the encoding of the relevant information such that the framing contained in the text gets overshadowed.Accordingly, we argue that visual aids can also help to alleviate framing in risky choices: Hypothesis 1b (reducing risky choice framing bias using icon arrays): Presenting a risky choice about human lives with an icon array reduces 6 the framing bias compared to presenting the same risky choice only as text.
We further argue that appropriate visuals can help to alleviate framing effects also in a real world, natural hazard setting.Predicting the occurrence and extent of natural hazards is a complex problem involving many factors.Therefore, the resulting hazard and risk predictions typically involve a considerable amount of uncertainty (Reynolds & Seeger, 2005).Communicating this uncertainty is challenging, especially in static, probabilistic maps (Thompson et al., 2015).
Instead of presenting uncertain predictions in a map, it is also possible to show past occurrences of the disaster and let respondents infer from this historical information to future risk.This is especially relevant for natural events that occur in a sequential, domino-like manner, as is the case for erosion processes.Saint-Marc et al. (2018) find that mapping past risk phenomena is useful to prepare against future risk.Moreover, depictions of past events are usually easier to produce than probabilistic predictions.
Further, it has been suggested that aerial photographs make it easier to locate and identify features than maps (Haynes et al., 2007).On a map, the cartographer filters certain information that she finds relevant, but that might not be relevant to the reader.With aerial photographs or satellite images, by contrast, no information is filtered and spatial items are in their correct relative scale, allowing the reader to extract the information relevant to her.We accordingly posit that satellite images containing information on past hazard events could alleviate framing effects for both risky choice framing and goal framing: Hypothesis 2b (reducing risky choice framing bias using satellite images showing past events): Presenting a risky choice in the context of natural hazards together with a satellite image showing past events reduces the framing effect compared to presenting the same information only as a text.
Hypothesis 3b (reducing goal framing bias using satellite images showing past events): Presenting disaster risk information aimed at behavioral change together with a satellite image showing past events reduces the framing effect compared to presenting the same information only as a text.
We further contend that it is the actual risk information that decreases the framing effect and not the satellite image itself.We base this claim on the fact that in existing studies using visual aids to alleviate framing effects, the applied visuals contained information on the risky situation (Gamliel & Kreiner, 2013;Garcia-Retamero & Cokely, 2011).These studies did not, however, test whether other visuals not containing explicit risk information would have been equally suited to decrease the framing effects (like a placebo test).We therefore propose the following hypothesis for the risky choice situation: Hypothesis 2c (influence of blank satellite image on risky choice framing): Presenting a risky choice in the context of natural hazards together with a satellite image depicting the situation, but not showing past disaster events, does not reduce the framing effect compared to presenting the same information as text only.

Materials and methods
To empirically test our hypotheses, we implemented three online survey experiments (Table 1) with U.S. residents on the platform Amazon Mechanical Turk (MTurk).Samples of MTurk respondents are often more representative of the general population than inperson convenience samples but are notably younger and more ideologically liberal than the public and pay more attention to tasks than other respondents (Berinsky et al., 2012;Goodman et al., 2013).Table S1 provides a comparison of MTurk sample demographics with other internet samples and face-to-face samples.

Experiment 1: risky choice framing in a decision situation about human lives
In the first experiment, the text group was presented with a written statement specifying a choice situation between a risky option and a sure outcome for a ship sinking in the middle of the ocean (see Figure S1 for the exact formulation of the decision situation, which was taken from Simon et al., 2004).Both options have the same expected value.For the image group, the context of the decision situation was described with the same text as in the text group, but the outcomes of the options were shown as icon arraysvisual representations symbolising passengersinstead of as text (see Figure S2).
Participants were randomly assigned to the text/image group and to the loss/gain frame, respectively, resulting in four groups.The gain frame was created by talking only about passengers being saved, in contrast to the loss frame talking only about passengers dying.Our dependent variable is the risk judgement of participants, i.e. if they are risk-seeking or riskaverse in the given decision situation, measured by which option they chose in the experiment.We included the following control variables in the questionnaire: age, gender, level of education, understanding of the decision situation and risk preference.The survey experiment consisted of a two-stage scenario: a risky choice in the first stage (experiment 2) and a goal framing setting in the second stage (experiment 3; see Figure 1).In experiment 2, each respondent faced a hypothetical decision situation where she inherits a plot of land by the beach on which she wants to build a house to move in with her family.The beach had experienced severe coastal erosion during three recent hurricanes.As the dependent variable, the respondent had to choose between two options for the position of the house: right by the beach, where she could build a large house with a higher chance of being damaged by erosion, or at the rear of the plot, where she could build a smaller house with a lower chance of being damaged. 7The two options were presented either with a gain frame ('keeping your house safe') or with a loss frame ('losing your house').
This decision situation resembles a risky choice, where the high-risk option is to build the house right by the beach, and the low-risk option is to build it at the rear of the plot. 8 In the text group (groups 2.1 and 2.2 in Figure 1), the decision situation was described only as text (see Figure S3).In the first image group (groups 2.3 and 2.4), respondents additionally saw a satellite image of the plot with two options for building the house (see Figure 2).In the second image group (groups 2.5 and 2.6), respondents saw the same satellite image, but including bands indicating the length of the beach eroded during three recent hurricanes 9 (see Figure 3).While the mentioned hurricanes have indeed caused severe erosion on U.S. beaches, the extent of erosion shown in the images is hypothetical, albeit realistic.

Experiment 3: goal framing in the context of natural disaster risk communication
After the respondent chose the location for her house, she was informed that two years after moving there, another hurricane caused widespread erosion, but not on the beach close to her house.This was followed by a list of measures she could take to mitigate erosion and to be prepared for the next erosion event (scenario adapted from McClure et al., 2009 who study the case of earthquake preparation messages; erosion control measures taken from Robertson, 2010).These measures are recommended in either a gain or a loss frame (see Figure S4).The image group of experiment 3 saw the same text as the text group but accompanied by a satellite image including erosion bands  (groups 3.3 and 3.4 in Figure 1).The same images as in Figure 3 were shown, but only the one for the option the respondent had chosen in experiment 2. Respondents were fully randomised between experiment 2 and experiment 3, meaning that some respondents saw the satellite image for the second time, and some for the first.
The dependent variable of experiment 3 was the respondent's attitude toward preparation for a major erosion event.This attitude was operationalised through two questions adopted from McClure et al. (2009): 'How important do you think it is to be well prepared for a major erosion event?' (1 = not at all important, 7 = very important) and 'How likely are you to actually take steps to prepare for a major erosion event?' (1 = very unlikely, 7 = very likely).
To correct for the potential incomparability of ordinal scales among respondents, we recoded the answers using the respondents' assessment of two hypothetical vignettes 10 on the same scale as the self-assessment, following the approach suggested by King et al. (2004) and King and Wand (2007).Recoding was done for both attitude questions.Afterwards, the two variables were averaged to form a measure of general attitude toward preparation (Cronbach's alpha = 0.64).Different covariates (Table S2) and demographic variables (age, gender, level of education and state of residence) were included in the questionnaire.

Experiment 1: risky choice framing in a decision situation about human lives
The total number of participants in experiment 1 was 662 (162-168 respondents per group).Descriptive statistics of demographics and covariates are shown in Table 2. To check the balance between the groups, we perform pair-wise joint orthogonality tests (see Hansen & Bowers, 2008;McKenzie, 2015).The regression of the group assignment on all covariates is shown in Table S3 and provides confidence that the randomisation worked.

Hypothesis 1a (risky choice framing effectchoice about human lives)
In Figure 4, we present the proportion of respondents choosing the risky option in the four groups.In the text group, 42% and 77% of the participants chose the risky option in the gain and loss frame, respectively.Using a 2-sample proportion test, we reject (p = 9.568 • 10 −11 ) the null hypothesis that the proportion of risky choice in the gain frame is greater than or equal to the one in the loss frame.This is consistent with the original results of Tversky and Kahneman (1981).Respondents are more risk averse in the gain frame than in the loss frame.

Hypothesis 1b (reducing risky choice framing bias using icon arrays)
In the gain frame, 42% and 56% of the participants in the text and image group, respectively, chose the risky option.Using a 2-sample proportion test, we reject (p = 0.0056) the null hypothesis that the proportion of risky choice in the text group is greater than or equal to the one in the image group.In the loss frame, 77% and 61% of the participants in the text and image group, respectively, chose the risky option.Using a 2-sample proportion test, we reject (p = 0.0015) the null hypothesis that the proportion of risky choice in the text group is smaller than or equal to the one in the image group.Lastly, we compare the proportion of risky choice in the gain and loss frame for the image group.Using a 2-sample proportion test, we cannot reject (p = 0.308) the null hypothesis that the proportion of risky choice in the gain frame is equal to the one in the loss frame.We thus conclude that presenting a risky choice with an icon array reduces the framing bias and could potentially eliminate it completely.As a robustness check, we perform a logistic regression which confirms that the framing effect is present in the text group but becomes insignificant in the image group (Table S4).

Experiment 2: risky choice framing in the context of natural disaster risk communication
In total, 1011 respondents participated in experiment 2 (164-172 respondents per group) and experiment 3 (249-258 respondents per group).Descriptive statistics of demographics and covariates are shown in Table 3. Balance checks are shown in Table S5 and Table S6 for experiment 2 and 3, respectively, and reveal only slight imbalances, which are to be expected given the large number of covariates tested.

Hypothesis 2a (risky choice framing effectnatural hazard situation)
Figure 5 presents the share of respondents choosing the risky option (i.e.building their house by the beach) in the six groups of experiment 2. In the text group (left panel), 7.2% and 7.3% of the respondents chose the risky option in the gain and loss frame, respectively.Using a 2-sample proportion test, we cannot reject (p = 0.488) the null hypothesis that the proportion of risky choice in the gain frame is greater than or equal to the one in the loss frame.Thus, we have to reject hypothesis 2a that respondents are more risk averse in the gain frame than in the loss frame.

Hypothesis 2b (reducing risky choice framing bias using satellite images showing past events)
Considering the satellite image that showed past erosion information (right panel), 12.3% and 15.2% of the respondents chose the risky option in the gain and loss frame, respectively.Hypothesis 2b states that showing the satellite image with erosion bands would reduce the framing effect that is present in the text group.Since we found no framing effect in the text group, we cannot formally test hypothesis 2b.Still, we test whether respondents who saw the erosion information became more risk seeking in the gain frame, as compared to the text group and the blank image group, as well as whether they got more risk averse in the loss frame.None of these effects is significant (Table 4).Lastly, we test whether the two frames in the group that saw the image with erosion information are equal and cannot reject this null hypothesis (p = 0.432).Adding the erosion information increased the share of risky choice compared to the blank image in both frames but this effect is not statistically significant.

Hypothesis 2c (influence of blank satellite image on risky choice framing)
In the group that saw the satellite image without erosion information (central panel), 9.0% and 14.0% chose the risky option in the gain and loss frame, respectively.Hypothesis 2c states that showing the satellite image would not reduce the framing effect that is present in the text group.Since we found no framing effect in the text group, we cannot test this hypothesis.If the satellite image alone was able to make respondents more risk neutral, we would expect that respondents get more risk seeking in the gain frame and more risk averse in the loss frame, as compared to the text group.Since the hypothesis is that the satellite image cannot reduce the framing effect, we test the opposite null hypotheses (Table 4).We can only reject the null hypothesis that the text-loss frame group is larger or equal than the blank image-loss frame group.Additionally, we test whether the two frames in the blank image group are equal but cannot reject this null hypothesis (p = 0.152).These results show that presenting respondents with a satellite image of the plot significantly increases the share of risky choice in the loss frame, compared to when they only receive textual information.In the gain frame, this effect is also present, but not significant.Visually, Figure 5 suggests that the satellite image introduces a framing effect, even though this effect is not statistically significant, as shown above.Overall, the effect of the satellite image seems to be an increase in risky choice, which is stronger in the loss than in the gain frame.
4.3.Experiment 3: goal framing in the context of natural disaster risk communication 4.3.1.Hypothesis 3a (goal framing effect) and hypothesis 3b (reducing goal framing bias using satellite images showing past events) Figure 6 shows the average attitude towards taking preparatory measures against coastal erosion after recoding the original survey items as described in section 3.3 (the attitude score before recoding is presented in Figure S5).In the text group (left panel), the average attitude in the gain and loss frame was 3.91 and 3.85, respectively.In the image group, the corresponding attitude values were 3.94 and 4.04.
To test our hypotheses 3a and 3b, we run a linear regression of the recoded attitude score on frame and image (Table S7, model 1).Since the framing parameter is insignificant, we reject hypothesis 3a.Since the image parameter is insignificant, we conclude that adding the satellite image does not increase the attitude score for respondents in the gain frame.However, in the loss frame, the attitude score is significantly higher in the image group than in the text group (p-value = 0.023 for Welch two-sample t-test).As there is no framing effect in the text group, we cannot formally test whether satellite images decrease the framing effect (hypothesis 3b).Directly comparing gain and loss frame in the image group, we cannot reject the null hypothesis that the means are equal (p-value = 0.130 for Welch two sample t-test).
We also test the robustness of these results by including those covariates for which we found an imbalance as well as state-level fixed effects into the regression (Table S7, model 2).The effects of frame, image and their interaction remain insignificant.The attitude score was significantly higher if people had already heard about erosion and if they felt it was the personal responsibility of people living by the coast to take protective measures.This corresponds to our expectations (see Table S2).Lastly, we repeat these two regressions with the original attitude scale before recoding (Table S7, models 3 and 4).The results remain largely unchanged.

Discussion
This paper investigates the existence of framing effects as well as possibilities to reduce them using visual aids for both a decision situation about human lives and a more realistic scenario in the context of natural hazards.We argue that framing effects occur if the decision situations are presented as text-only, but that they diminish if a suitable visual is shown alongside the text.
In experiment 1, dealing with a decision situation involving human lives, a significant framing effect occurred in the text-only group, which became insignificant if an icon array showed the decision situation.In experiments 2 and 3, dealing with a more realistic decision situation in the context of natural hazards, we did not find framing effects in the text groups.This is surprising given the strong framing effect found in experiment 1 and since framing effects are well established in the literature (Levin et al., 1998;Piñon & Gambara, 2005).In terms of goal framing, we did not find framing effects although we closely reproduced the research design of McClure et al. (2009) who study earthquake preparation measures in New Zealand and find large differences in attitude scores between gain and loss frames.Several reasons can potentially explain why we did not find framing effects in experiments 2 and 3. First, the scenario of these experiments differs strongly from the scenarios employed in most of the cited studies, which usually involve a choice about human life or death.For instance, Schneider (1992) finds that framing effects are only strong if they involve decisions about human lives, while Amsalem and Zoizner (2020) conclude in their meta-analysis that framing effects are weak under more realistic circumstances and that 'citizens appear to be more competent than some scholars envision them to be.' Second, as Levin et al. (1998) note, choice situations only provide an indirect measure of the effect of framing on information processing.An individual forms her decision based on an interplay of the frame and the specifics of the decision situation.If an individual considers a risk as unacceptable for herself, she will most likely not react strongly to the framing but choose the risk averse option irrespective of the frame.In our case, respondents seemed to consider coastal erosion as a severe risk and were on average rather risk averse, which could explain the dampening of the framing effect.
Third, unlike most classical choice situations, we could not assign quantitative risk levels to the two options.For most natural hazards, it is rarely possible to say whether a specific location is at an 80% or 50% risk of being affected by the hazard in the next X years.This is only possible for hazards where long time series of data are available, such as the risk of flooding for which return periods can be calculated.Therefore, we chose to speak only of 'higher' and 'lower' risk, which might be interpreted differently by respondents.
In the goal framing literature, natural hazards have not been studied with the exception of the work by McClure et al. (2009).Most goal framing studies focus on health or marketing related topics.Following a similar logic as for risky choice framing, it could be that goal framing effects are most prominent for messages related to human life or death.Further, it was found that goal framing effects are much less stable than risky choice framing effects (Levin et al., 2001(Levin et al., , 2002)).In our study, support for taking preventive measures was very high (average score of 6.4 on a 7-point scale), which might lead to ceiling effects.
The second focus of experiments 2 and 3 was the influence of satellite images on framing effects.Since we did not find framing effects in the text groups, we could not evaluate whether images can help to decrease such framing effects.Irrespective of the size of the framing effect, we found an impact of adding the images in both experiments.
In experiment 2, showing the satellite image without erosion information significantly increased the share of risky choice.This might be due to two particular features of the satellite image.First, the image shows a clear line between the houses and the beach, which can be interpreted as a wall protecting the houses from erosion (see the 'levee effect', White, 1945).Second, all other houses in the image are right by the beach, which might make respondents wonder why they should build at the rear of the plot if no one else does so.The image contained additional contextual information that was not transmitted in the text and that might have treated different respondents differently, making the comparison difficult.Adding the erosion bands to the image made respondents even more risk seeking, although not at a statistically significant level.This is counterintuitive given that the erosion bands are so wide that the distance between the houses and the water line is smaller than what has been eroded during any of the three hurricanes shown.In experiment 3, adding the image increased the attitude score, which is in line with our expectation.
The third point to discuss is that, overall, only few respondents chose the risky option of building their house by the sea and most of them stated very high intentions to take preventive measures.The average score in the goal framing experiment is much higher at 6.4 than the one in the McClure et al. (2009) study (around 5.2). 11This is surprising since their study recommended low-investment measures such as buying a first-aid kit or rearranging furniture in the house.Our study, by contrast, promoted measures that involve a significant investment on the side of the respondents, such as taking out an insurance or building a sand fence on the beach.One explanation for the very risk averse outcomes could be that we only assess stated preferences (Murphy et al., 2005;Park et al., 2002).However, this in itself cannot yet explain why we find much higher attitude scores than McClure et al. (2009) who also used only stated preference questions.
Important drivers in risky choices and risk-related behaviour are the respondents' familiarity with the risk and their estimation of its severity (Carlton & Jacobson, 2013;Rosenbaum & Culshaw, 2003).The covariates (see Table 3) suggest that many residents of coastal states are aware of the phenomenon and risk of coastal erosion.The results imply that if people are aware of an environmental hazard and can choose where to build their house, they mostly choose the low-risk option.Conversely, this might imply that if people build their house in a high-risk location even though they are aware of the risk, they are either very risk-seeking or they do not have any other choice than building in this location.
Summing up our findings, we conclude with respect to framing that its influence on risky choices or behavioural intentions might be limited beyond the typical, highly stylised scenarios that have mostly been studied in the literature.With respect to adding satellite images, we conclude that such images contain contextual information that cannot be transferred in textual form and that might significantly alter respondents' choices and behavioural intentions.At the same time, it remains unclear how people interpret information on past occurrences of a disaster if they are added to satellite images.
These findings suggest two main areas of further research.First, it should be investigated more thoroughly to which extent the framing effects that have been established in various hypothetical scenarios are valid in real world decision situations.As soon as the decision situation gets more realistic and tangible for respondents, various factors besides the mere information framing emerge that influence respondents' choices.If people have a very clear preference for one option, linguistic framing alone will hardly be able to shift their choice to the undesired option.How effective frames can be constructed for such settings is an important question, given that, for instance, many climate change measures require rapid and drastic behaviour changes at the individual level.Whether it is ethically acceptable to use such frames to nudge people is a normative question that requires a broader societal debate.Understanding if and how they work, by contrast, is a meaningful academic question with considerable policy relevance.
Second, it will be beneficial for the area of risk communication to study in more detail how people understand and use maps and satellite images that present risk information not in a predictive manner, but only by showing the spatial extent of past risk events.Many climate-related hazards occur in a spatially sequential wayerosion of coasts and riverbanks being one example, desertification or salinisation others.It can be scientifically challenging to predict the probability and extent of such events, and even if such probability information is available, it is even more challenging to communicate it to laypeople (see Visschers et al., 2009 for a review).It is thus worth exploring whether people can more easily understand how imminent and threatening an environmental hazard is when they see where past events have occurred.In any case, it appears that using visuals for risk communication requires careful (and potentially qualitative) pretesting to avoid unintended outcomes.
For policy makers and warning agencies, it is particularly noteworthy that framing effects might be limited in realistic decision situations, especially if people hold strong preferences about the issue at stake.Wherever possible, integrating residents and their understanding of the local context into the disaster management should be preferred over mere top-down risk communication assuming an 'ignorant public'.

Figure 1 .
Figure 1.Flow chart of experiments 2 and 3 (covariates are not included in the graph).

Figure 2 .
Figure 2. Satellite images shown to the first image group of experiment 2.

Figure 3 .
Figure 3. Satellite images shown to the second image group of experiment 2, including information on past erosion events.

Figure 4 .
Figure 4. Share of respondents of experiment 1 choosing the risky option with 95% confidence intervals in the gain and loss frame, respectively, for the group that saw only text (left) and the text accompanied by an icon array (right).

Figure 5 .
Figure 5. Share of respondents of experiment 2 choosing to build their house close to the beach with 95% confidence intervals in the gain and loss frame, respectively, for the group that saw only text (left), text and the satellite image without erosion information (centre) and text and the satellite image with erosion bands (right).

Figure 6 .
Figure 6.Average attitude (recoded with vignettes) of respondents of experiment 3 toward taking preventive measures against coastal erosion with 95% confidence intervals in the gain and loss frame, respectively, for the group that saw only text (left) and text and the satellite image with erosion bands (right).

Table 1 .
Overview of survey experiments.For the second and third experiment, we chose coastal erosion in the U.S. as our natural hazard case.Participation in this experiment was possible only for MTurkers residing in seven coastal U.S. states, namely California, Texas, Florida, Georgia, New Jersey, North Carolina and Virginia.These states are all heavily affected by coastal erosion.By targeting them, we tried to reach MTurkers living close to the coast who are familiar with coastal erosion.

Table 2 .
Descriptive statistics of demographics and covariates for the sample of experiment 1.

Table 3 .
Descriptive statistics of demographics and covariates for the sample of experiments 2 and 3.

Table 4 .
Overview of null hypotheses tested regarding the three hypotheses H2a, H2b and H2c relating to experiment 2.