1 Introduction

The Madrid train bombings of 2004, the London Underground bombings of 2005, and the September 11, 2001, attacks on New York City are just a few examples of how transportation infrastructure serves as one of the highest profile, most salient targets for modern terrorism (e.g., see Bloom 1998; Johnston and Nath 2004; Johnston 2004). Of course, given the multitude of adversary methods and targets, defending transportation infrastructure is a challenging and critical problem from a mitigation/interdiction standpoint. Yet transportation terror poses an equally grave challenge for public risk perception and risk communication policy; transportation-based terror attacks have the potential to disrupt normal travel behavior and patterns, and can bring about sizable economic losses and human casualties (Alexander 2004; Alexander and Alexander 2002; Howie 2009). One often-mentioned yet highly relevant consequence of the September 11 terror attacks was an estimated increase of 344 traffic deaths in late 2001 due to a public aversion to air travel (Blalock et al. 2009), a finding that underscores the need to effectively understand how the public assesses risk in the aftermath of transportation-based terror attacks.

A great deal of empirical work has specifically examined the economic and psychological consequences of the September 11 terror attacks (see Fairbrother et al. 2003; Galea et al. 2003; Marshall et al. 2007). Yet considerably less research has focused on public reactions to other forms of transportation-based terror. Thus, the primary aim of the present study is to model respondents’ affective, cognitive, and intended behavioral reactions to various types of transportation terror in order to (1) better characterize the predictors of risk perception and travel behavior change in the wake of transportation-targeted terror attacks, and (2) understand how the covariation among risk perception variables is moderated by specific features of the attacks.

1.1 Terrorism’s changing landscape

As mentioned above, much of the research on transportation-based terrorism has focused on the aftermath of the September 11 attacks, which were carried out through the hijackings of commercial airlines. Yet the landscape of transportation terror looks much different than it did in 2001; an in-depth analysis of successful terror attacks from 1982 to 2011 (defined as those with two or more casualties) has revealed opposing trends for aviation- and railroad-related terror casualties, with aviation casualties decreasing over time (593 from 1982 to 1991 vs. 0 from 2002 to 2011) and rail casualties increasing over time (0 from 1982 to 1991 vs. 442 from 2002 to 2011; Barnett 2015). Other researchers have similarly pointed out that terror has increasingly moved toward soft-target attacks, such as open urban spaces that are difficult to protect (Coaffee 2009; Hoffman 2003). The consequences of these ‘soft-target’ terror attacks can be quite formidable in the context of transportation; a report by the Mineta Transportation Institute found that, while 20% of all terror attacks induce fatalities, that rate is approximately 37% for attacks on public transportation systems (Jenkins 2001). Cyberterrorism is another especially pressing concern for transportation systems. An attack of this nature could compromise critical components of transportation infrastructure such as train signaling systems (Chen et al. 2014), autonomous vehicles (Petit and Shladover 2015), and air traffic control systems (Abeyratne 2011). These attacks could be designed to cause collisions or malfunctions that could lead, not just to economic or privacy-related consequences, but to substantial loss-of-life. As such, there is a crucial need to defend infrastructure targets against malicious cyberattacks (Ismail et al. 2014; Cerrudo and Spaniel 2015), and to understand the behavioral and economic consequences of this kind of threat.

1.2 Overview of the study

Given the threat that soft-target transportation terror poses, the central focus of our study is twofold. First, we seek to construct a path model that identifies which aspects of respondents’ reactions to transportation attacks (including cognitive, affective, and intended behavioral change variables) are most predictive of each other. This approach allows a rich view of the public reactions that accompany terror attacks by broadly focussing on multiple dependent variables of interest. Similar path-analytic approaches appear in other empirical risk perception work (e.g., Kim et al. 2008; Rundmo 2000; Siegrist 2000) and have previously been used to model terror reactions (see Bux and Coyne 2009). We seek to build on these past works by applying path analysis to the novel domain of soft-target transportation terror, and include a wider array of psychological constructs than is often used in other risk perception path models.

Second, we seek to understand how our constructed model differs based on specific features of the terror attack at hand. By employing an experimental, scenario-based paradigm (in which respondents react to hypothetical, yet realistic, terror scenarios), we directly manipulate specific attack features to model their impact on (1) overall levels of risk perception, and (2) the patterns of covariation between the dependent variables. Past work on transportation terror has made great contributions by directly modeling real-world responses to terror events (such as the 2005 London Underground bombings, see Bux and Coyne 2009, Rubin et al. 2007), but were unable to directly test the kinds of manipulations that our experimental paradigm allows. Thus, our overarching research goal is to contribute to a well-validated model of the public’s response to soft-target transportation terror that can accommodate the moderating effects of various attack features.

1.2.1 Dependent variables of interest

Three of our public response variables pertain to respondents’ pre-standing beliefs and attitudes, measured before the administration of the terror attack scenarios: (1) terrorism risk sensitivity, (2) general risk sensitivity, and (3) trust in the US Department of Homeland Security (DHS). ‘Terrorism risk sensitivity’ refers to the degree to which respondents feel that terrorism poses a risk to their well-being, while ‘general risk sensitivity’ refers to the degree to which respondents feel that their well-being is threatened by hazards not necessarily related to terrorism (such as economic downturns and pandemics). The rationale for these variables’ inclusion comes from Sjöberg (2000), who found that risk perception toward a specific hazard depended on participants’ pre-standing risk attitudes toward the hazards and on their general levels of risk sensitivity; furthermore, other studies have supported the role of terrorism risk sensitivity in predicting terror-related reactions (see Lee and Lemyre 2009). Trust in the DHS serves as a proxy construct for respondents’ trust in the government’s ability to appropriately prevent and respond to terrorist attacks. On a general level, institutional trust has long been studied as a relevant variable in studies of risk perception (Kim et al. 2008; Siegrist 2000; Peters et al. 1997; van der Weerd et al. 2011). Given that this study utilized a US-based sample, we selected the DHS as the target institution for the trust construct due to its role as the primary overseer of transportation security in the USA.

We also focus on four self-reported facets of participants’ post-scenario reactions: fear, anger, perceived risk, and intent to curtail travel related to the target of the terror attack scenario. Respondents’ intent to curtail travel was included as a measure of how respondents’ reactions to the hypothetical terror attacks would translate into real-world behaviors. Fear, anger, and perceived risk were included because lay perceptions of risk are dependent on both emotional (fear/anger) and cognitive (risk assessment) appraisals of threat (Iyer et al. 2014; Loewenstein et al. 2001). Fear and anger, while related, have consistently been shown to differentially correlate with risk perception (fear as a risk inducer and anger as a risk reducer; Han et al. 2007; Lerner and Keltner 2001; Lerner et al. 2003). Anger has also been shown to lead to risk-seeking decision behavior, while fear has been shown to lead to risk-averse decision behavior (Lerner and Keltner 2001). Even beyond the domain of risk perception and decision making, there is evidence that anger motivates approach behaviors, while fear/anxiety motivates avoidance behaviors (Carver and Harmon-Jones 2009), and that the two emotions differ in the levels of information processing that they promote (Nabi 2002). Thus, while the two emotions are undoubtedly related on a psychological level, we chose to treat them separately for modeling purposes, to acknowledge the possibility that they produce contrasting or independent effects on other variables of interest. While many more constructs could be included in the model, the aforementioned variables span the emotional, cognitive, and behavioral facets of respondents’ reactions to the terror scenarios, and allow us to specify how such reactions relate to pre-standing attitudes about risk and trust in government.

1.2.2 Justification for path model structure

While path models can communicate information about the strength of dependencies between variables, it is up to the researchers to theoretically justify the structure imposed on the model in the first place (as many path models can be fit to the same dataset). Three of the dependent variables (trust in DHS, terrorism risk sensitivity, and general risk sensitivity) were measured before administration of the terror scenarios, and have been previously implicated as potential causes of risk judgments (Lee et al. 2010; Sjöberg 2000; Kim et al. 2008; Siegrist 2000; Peters et al. 1997; van der Weerd et al. 2011). Thus, each of these constructs was specified as exogenous potential causes of each of the four post-scenario public reactions (note that our scenario-based experimental design ensures that these variables are not affected by respondents’ reactions to the terror scenarios, as they were measured first). Regarding the post-scenario variables, multiple studies have reported that state affect plays an important role in the formation of risk perceptions (see Loewenstein et al. 2001; Lavine et al. 1998; Slovic and Peters 2006), with several studies confirming that induction of negative emotions can exaggerate cognitive evaluations of risk (Finucane et al. 2000; Johnson and Tversky 1983; Lerner et al. 2003). In general, there is strong empirical support that emotions impact cognitive judgments, and that emotional reactions to stimuli ‘need not be cognitively mediated’ (Loewenstein et al. 2001, p. 10). Thus, both fear and anger were specified as potential determinants of risk perceptions (a cognitive appraisal of the risk posed by future terror attacks). We then specified respondents’ anticipated fear, anger, and risk perceptions as potential causes of their intent to curtail their travel, given that both affect and risk perception have been previously validated as influencing future travel behavior (Reisinger and Mavondo 2005; Sönmez and Graefe 1998a, b).

Thus, the initial path model was a saturated model that initially estimated the strength of all construct interrelationships. We should note that such a model is not synonymous with the assumption that every path will yield a significant coefficient; rather, excluding any path a priori would be considered a strong causal assumption (assuming a path coefficient of zero), for which we did not have sufficient theoretical justification. Rather, we begin with a fully saturated model and allow the analyses to inform which paths are relevant and which are negligible.

1.2.3 Experimental manipulations

To test whether the model of participants’ responses holds or varies across situations, we administered fictional terror scenarios that differed across two independent variables: the attack method (in terms of cyber vs. non-cyber) and the attack location (airport vs. public transportation). Of course, soft-target transportation attacks can vary across a much larger event space than is represented here; the purpose of the present experiment was not to exhaustively study every potential variation in how transportation-based terrorism is realized, but to focus on two specific characteristics of transportation attacks (which have not been previously studied experimentally) and provide a detailed analysis of how these characteristics may influence public reactions (or the patterns of covariation between them).

1.2.4 Attack method: cyber versus non-cyber terror attacks

The distinction between cyber and non-cyber terror attacks has not yet been studied within the context of transportation systems. Cyberattacks are popularly known for the threat they pose to infrastructure systems and securely held data, but within the context of transportation systems, cyberattacks have the potential to cause substantial loss-of-life (through the hacking or disruption of automated transportation systems, such as trains). Yet there is no empirical work on how the public might react to a fatality-inducing cyberattack on transportation networks, as compared to more ‘traditional’ terror methods that cause damage on a similar scale. Thus, we constructed scenarios of terror attacks on public ground transit that either involve traditional methods (improvised explosive devices, or ‘IED’s’) or cyber methods, and examine how the scenario administered impacts respondents’ reactions.

1.2.5 Attack location: aviation versus public ground transportation

Within the USA, the threat of terrorism has long been tied to aviation, beginning with plane hijackings in the 1960s and 1970s. Some of the recent high-profile terror attacks outside the US, such as those in London, Paris, Brussels, and Istanbul, have taken place at airports, yet are more similar to other soft-target attacks on public spaces than to the airplane hijackings of September 11, 2001. To date, no studies have examined whether an attack’s mere association with aviation triggers a unique public response, even if it does not involve breaching aviation security systems. Thus, we also examine respondents’ reactions to terror scenarios involving the use of an improvised explosive device (IED) on either an airport (aviation related) or a public ground transportation system (non-aviation related), to investigate whether transport attack target (aviation vs. public ground transit) affects public responses in terms of affect, risk perception, and behavioral intentions to curtail travel.

1.2.6 Hypotheses

Regarding the overall structure of the path model across scenario conditions, we expect respondents’ self-reported fear to positively predict their self-reported levels of risk perception and intent to curtail travel, while we expect self-reported anger to negatively predict these variables (consistent with the differential effects that fear and anger have on risk perceptions). We also expect trust in the DHS (a proxy construct for government trust) to negatively predict participants’ self-reported fear, anger, risk assessment, and intent to curtail travel, while we expect terrorism risk sensitivity and general risk sensitivity to positively predict these variables. However, we make no specific predictions about which of the post-scenario response variables (fear, anger, risk perception, or intent to curtail travel) these measures will most strongly predict.

Regarding the comparison between cyber-based and non-cyber-based ground attack scenarios, we hypothesize that the path model for the cyber scenario will produce higher path coefficients between general risk sensitivity and the post-scenario reaction variables (fear, anger, risk perception, and intent to curtail travel), and lower path coefficients between these variables and terrorism risk sensitivity; we expect the public to be less familiar with cyberterrorism compared to more traditional terror methods, and would therefore expect attitudes toward ‘risk in general’ (not specific to terrorism) to better predict their reactions than their attitudes toward terrorism (which they may not clearly associate with cyberattacks).

Regarding the comparison between the airport and public transit attacks, we hypothesize that intent to alter travel behavior will be less correlated with fear and risk assessment in the airport than the non-airport scenario, since air travel is typically harder to substitute for other means of transport than is the case for public ground transportation; thus, travel decisions may be more heavily influenced by the unique benefits of the travel option and less on the perceptions of risk or fear of another terror attack. Lastly, we expect terrorism risk sensitivity to be more predictive of fear, anger, and risk perceptions in the airport scenario than in the public transit scenario, given the salient association between aviation and terrorism in the USA (due to the September 11, 2001, attacks).

While our primary analytical focus is on the path model’s estimates across scenarios, it is also worth analyzing the mean-level impacts of attack method (cyber vs. IED use) and location (airport vs. public transit). We expect the train scenario (cyberattack on public transit) to induce higher levels of risk assessment, fear, and intent to curtail travel than the bus scenario (IED attack on public transit), due to the novel nature of the cyberattack method. We make no specific hypotheses regarding the airport/transit effect on fear, risk perception, and intent to curtail travel; the public might react more adversely to airport-located attacks due to aviation’s longstanding association with terrorism in the west, but also might react less severely due to the relatively short amounts of time most people spend in airports.

2 Methods

2.1 Participants

A total sample size of n = 512 participants were recruited via a national panel established by Decision Research (University of Oregon) to investigate transportation-related risk perceptions. This panel was established in 2008 with funds from the National Science Foundation. Those interested in participation were asked to register with Decision Research, and were subsequently selected based on demographic characteristics (a quota sample). Prior to the administration of this experiment, our respondents had participated in other panel studies that involved other experimental manipulations. One subgroup (n = 152) had viewed a factual video describing various aspects of transportation security 40 days prior to the current study, while a larger subgroup (n = 335) read about a hypothetical aviation-related terror attack and provided reactions to that attack roughly 2 weeks prior to the current study. There was considerable delay between these past manipulations and this study, and these subgroups were statistically unrelated to the three experimental manipulations we employed (χ2(4) = .48, p = .98). All participants were paid $10 for participation. It is worth noting that, because the sample was entirely US-based, care should be taken in extrapolating the results to other populations or cultures.

Participants were assigned to one of three groups, in which they read one of three mock news stories involving a terror attack: (1) the use of an improvised explosive device (IED) on a public bus, (2) a cyber attack resulting in the crash of a public train, or (3) the use of an improvised explosive device (IED) at an airport passenger terminal. After removing participants who failed an attention check question, the final sample size used for analyses was n = 430.

Table 1 shows the percentage breakdown of each scenario condition by gender and education. Chi-square tests for significance found that scenario condition was not significantly associated with gender (χ2(2) = 4.37, p = .11) nor education (χ2(12) = 10.37, p = .58). The median ages for the bus (public transit/IED), train (public transit/cyber), and airport (airport/IED) conditions were 43.5 (IQR = 22.0), 42.0 (IQR = 17.25), and 43.0 (IQR = 23.0) years, respectively (age did not significantly differ among scenario conditions, F = .502, p = .605). The final sample sizes by groups were nbus = 144, ntrain = 128, and nair = 158.

Table 1 Respondent sex and education level by attack scenario
Full size table

2.2 Procedures

Participants began the experiment by responding to self-report measures of trust in the DHS, terrorism risk sensitivity, and generalized risk sensitivity. They were then randomly assigned to read one of three terror attack scenarios, each of which described an attack on a transportation target within the city of Los Angeles. The scenarios were formatted as mock news stories, and described one of three events: the detonation of an IED on a city bus in a crowded downtown location, a Metro train crash caused by a cyberattack, and the detonation of an IED in the passenger terminal at LAX airport. The bus scenario served as the baseline condition, since it described a non-cyber terror attack on a ground public transportation system. The effect of attack target (aviation vs. public transit) was assessed by comparing results from the bus and airport IED attack scenarios, while the effect of attack mode (cyber vs. IED detonation) was assessed by comparing the results from the bus and train ground attack scenarios. Thus, the design allowed for the evaluation of both independent variables using only three scenarios. Each scenario took roughly 2–3 min to read, and was accompanied with relevant pictures and visual aids (descriptions of each scenario appear in Table 2). After reading the scenario, respondents completed the measures of fear, anger, risk assessment, and intent to curtail travel.

Table 2 Descriptions of each attack scenario
Full size table

2.3 Measures

2.3.1 Trust in DHS

Participants reported the degree to which they were confident in the Department of Homeland Security’s ability to prevent and mitigate the adverse consequences of terrorist attacks. The Department of Homeland Security was chosen as the specific government agency of interest because (1) its mission largely consists of preventing acts of terrorism, and (2) it is responsible for preventing terrorist attacks on transportation systems. Respondents indicated the degree to which they agreed with each of the following three statements, on a 1 (strongly disagree)–7 (strongly agree) scale:

  • DHS is effective in preventing terrorist attacks

  • DHS would be effective in minimizing harm from a terrorist attack

  • DHS is capable of dealing with national crises

2.3.2 Terrorism risk sensitivity

Respondents provided the degree to which they believed their quality of life was threatened by the following six terrorism-related threats on a scale from 1 (No Risk) to 7 (Extremely High Risk):

  • A terrorist biological attack (like an anthrax attack)

  • A terrorist attack against a commercial airplane

  • A terrorist bombing of a tourist location

  • A terrorist radiological attack (conventional bomb combined with radiological material)

  • Cyber terrorist attacks

  • Any type of terrorist attack

2.3.3 Generalized risk sensitivity

Participants also reported the degree to which they believed their quality of life was threatened by various non-terror-related societal risks. Participants rated the following nine hazards on a scale from 1 (No Risk) to 7 (Extremely High Risk):

  • Pandemic diseases (like H1N1 flu)

  • Natural disasters (like earthquakes, hurricanes, tornados and flooding)

  • Technological accidents (like oil spills, nuclear accidents, chemical fires)

  • Bacterial contamination in food

  • National financial crisis

  • High unemployment rate

  • Housing market

  • Low rates of return on Wall Street

  • Chronic gridlock in Congress

2.3.4 Fear and anger (post-scenario)

After reading their assigned terror attack scenario, participants reported the degree to which the type of attack described would make them feel fearful and angry. Responses for each of the two items ranged from 1 (not at all fearful/angry) to 4 (very fearful/angry).

2.3.5 Risk perception (post-scenario)

After reading their assigned terror scenario, participants indicated their perception of the risk posed by similar attacks in the USA in the future. Respondents answered the following two items on a scale from 1 (not at all/no change in my perception of risk) to 5 (very much/great increase in my perception of risk):

  • How much do you feel you would be at risk in the next 30 days?

  • To what extent would the attack be a sign that the risk to the affected transportation system was greater than what you had thought?

2.3.6 Travel behavior adjustment (post-scenario)

Lastly, respondents indicated whether the terror attack they read about would cause them to alter their travel plans. Participants indicated whether they would maintain (coded as 0) or change (coded as 1) their plans for the following four hypothetical trips. A composite score (from 0 to 4) was obtained by adding the four dichotomous codes.

  • Travel for business/your job

  • Travel for a special event (e.g., wedding and graduation)

  • Travel to see family

  • Travel for leisure

2.3.7 Analysis

The data did not conform to the multivariate normality assumption required for traditional Structural Equation Modeling (this largely stemmed from a lack of granularity in the fear, anger, and travel behavior adjustment constructs). Thus, the data were analyzed using Partial Least Squares (PLS) path modeling: a technique that attempts to maximize the variance in the endogenous variables explained by the predictors. It employs an algorithm that identifies the item loadings (for the manifest variables) that maximize the covariances between the latent constructs (which are calculated as weighted sums of their manifest variables), and thus relies on no distributional assumptions. Likewise, inference over model parameters is conducted via nonparametric bootstrapping.

PLS modeling inherently focuses on maximizing the predictive capability of a path model, rather than reproducing a sample covariance matrix—and due to the compositional nature of the ‘latent’ constructs, it does not account for measurement error. As such, variables are not to be interpreted as necessarily reflecting respondents’ underlying psychological constructs, but simply the types of attitudes and reactions that tend to covary in response to terror attacks. While such results should not be taken as analogous to ideographic (within-individual) mental processes, they can shed light on nomothetic/group processes, and answer questions about who, in the general public, reacts most severely to different types of terror attacks (in terms of their accompanying attitudes and behaviors).

3 Results

3.1 Comparison of construct scores

Table 3 presents the means and standard deviations of respondents’ standardized fear, anger, risk perception, and travel behavior adjustment scores. Due to non-normality in the data, significance tests were computed with nonparametric Wilcoxon–Mann–Whitney tests, and three of the between-subject comparisons reached significance. On average, respondents who viewed the train scenario (cyberattack on public transit) had significantly higher levels of risk perception than those who viewed the bus scenario (IED detonation on public transit; d = 0.39, p < .001), while those who viewed the airport IED scenario had lower levels of risk perception than the bus IED scenario (d = 0.40, p < .001). This suggests that participants perceived higher levels of risk due to the threat of a cyberattack than a more traditional explosives attack on ground transport, but perceived less future risk due to an explosives attack at an airport location than an explosives attack on public ground transit. While travel change behavior did significantly differ between the airport and bus scenarios, this could likely be due to the greater ease with which bus trips can be canceled/replaced, relative to air travel.

Table 3 Standardized construct score means (standard deviations) and effect sizes for differences
Full size table

3.2 Measurement model validation

Separate path models were estimated for the bus, train, and airport scenario conditions to verify that (1) item loadings were acceptably high for all three groups, and that (2) constructs were formed similarly between the conditions. Two items were removed from analyses due to high cross-loadings with other constructs in all scenario conditions (the ‘cyber terrorist attacks’ item for terrorism risk sensitivity loaded highly on general risk sensitivity, and the ‘pandemic diseases’ item for general risk sensitivity loaded highly onto terrorism risk sensitivity). Table 4 presents the remaining item loadings for each of the three scenarios. Almost all item loadings exceed .70 (with the exception of two in the train condition), and all constructs demonstrated high composite reliability (> .80, as measured by Dillon–Goldstein’s rho). In terms of discriminant validity, all loadings exceeded their cross-loadings on other constructs, and the Average Variance Extracted (AVE) for each latent variable (the proportion of indicator variance explained by the construct) exceeded all squared path loadings with other variables (this criterion ensures that each latent variable shares more variance with its indicators than with the other latent variables in the model). Thus, the data met all recommended criteria for establishing convergent and discriminant construct validity (see Hair et al. 2016 for an overview of PLS model validation).

Table 4 Item loadings for each scenario condition
Full size table

After the convergent and discriminant validity of the original measurement model was confirmed, the path model was tested for compositional measurement invariance between the three scenario groups. Because Partial Least Squares modeling calculates latent variables as weighted sums of the manifest variables, this type of analysis ensures that the latent variables are constructed in the same manner across groups, and establishing this type of measurement invariance is necessary to compare path coefficients between the groups. The tests used 1000 replications to randomly redistribute observations between two scenarios to construct an empirical distribution for the correlation between composite scores (which should equal 1.00 under measurement invariance); a more detailed description of the procedure can be found in Henseler et al. (2016). We conducted this test for each latent variable across the pairwise group comparisons of interest: the bus scenario versus the airport scenario (testing the effect of attack location) and the bus scenario versus the train scenario (testing the effect of attack method), with p values ranging from .18 to .99, confirming no significant differences in how the constructs were composed among the three scenarios.

3.3 Structural model evaluation

With compositional measurement invariance confirmed, we examined whether the models’ path coefficients significantly differed based on attack target (aviation vs. ground transit) or mode of ground attack (cyber vs. IED). Table 5 displays the estimated unstandardized path coefficients for all three scenario conditions. Nonparametric bootstrapping was used to compute p values for the differences between the three scenario models’ path coefficients (over 1000 iterations), though no path coefficients’ differences approached significance (p values ranged from 0.45 to 0.99).

Table 5 Unstandardized path coefficients for scenario path models
Full size table

Thus, data from all conditions were pooled into a single path analysis to make consistent cross-scenario inferences about intercorrelations between the dependent variables. For this whole-sample path model, path coefficients were retained based on their p value and f2 effect size (conceptually, the degree to which a path increases its target variable’s R2 value); paths with p values larger than .05 or f2 values smaller than .02 (considered a small effect size; Cohen 1992) were removed from the model, resulting in the final path model displayed in Fig. 1.

Fig. 1
figure 1

Final path model, collapsed across scenario conditions, with R2 values for all endogenous constructs. All nonsignificant/negligible paths removed (all p’s < .05)

Full size image

The pruned path model suggests that respondents’ tendency to avoid travel in the aftermath of a terror attack is well-predicted by their self-reported fear, anger, and risk perception, with the effect of fear being partially mediated by risk perception. Overall, respondents’ self-reported fear plays a stronger role in the path model than anger (as evidenced by larger path coefficients), and fear and anger impacted intent to alter travel behavior in opposing directions. Terrorism risk sensitivity and generalized risk sensitivity predicted post-scenario fear and anger, respectively, while trust in the DHS was removed from the path model due to the lack of reliable path coefficients.

4 Discussion

The overarching goal of this study was to examine the dependencies between respondents’ attitudes, affect, risk perceptions, and behavior intentions in the context of soft-target transportation terror, and how such dependencies differed across attack scenarios. Contrary to our hypotheses, there were no significant moderating effects of attack method (cyber vs. IED) nor attack location (airport vs. public transit); responses to each of the three scenarios were thus modeled by a similar set of path coefficients. From a practical standpoint, this suggest that those who react most strongly to a transportation terror attack (such as those who severely alter their travel behavior) would generally share the same characteristics (i.e., higher levels of terrorism risk sensitivity, fearful reaction to the attacks), regardless of the specific features of the attack.

Some of our original hypotheses regarding the structure of the ‘combined’ path model (pooled across scenario conditions) were supported. Self-reported fear was positively predictive of risk assessment and the intent to alter travel behavior (as hypothesized), while anger was negatively predictive of travel behavior intent (though not of risk assessment). These results confirm past findings that fear and anger can have differential effects on risk-related variables (e.g., the finding that fear and anger are positively and negatively predictive of risk perception, respectively; Lerner et al. 2003). Thus, even though fear and anger were positively correlated in our sample (r = 0.45), the fact that fear had a positive effect on risk perception (while anger did not) and that the two emotions predicted travel behavior in opposite directions seems to justify their treatment as separate constructs for modeling purposes. Terrorism risk sensitivity was positively predictive of fear, as was partially hypothesized (we did not specify which post-scenario reaction variable it would correlate with); however, general risk sensitivity was, interestingly, only predictive of anger (resulting in a negative total effect on intent to alter travel behavior). While these two risk sensitivity constructs are undoubtedly correlated in the population (as they were in our sample), their differing associations with fear, anger, and perceived risk suggest a multifaceted view of the concept of ‘risk sensitivity’ that warrants further psychometric investigation. Lastly, the finding that trust in the DHS demonstrated no reliable predictive value in the model also warrants further investigation, given that past studies have found institutional trust to be an important correlate of risk-related behaviors.

Lastly, though not central to our original research aims, there were notable differences in self-reported risk perception between the attack scenarios. In line with our hypothesis, participants expressed higher levels of perceived risk in the train attack scenario (cyberattack on public transit) as opposed to the bus scenario (IED detonation on public transit), though the same did not hold for self-reported levels of fear. Respondents also indicated lower levels of risk assessment in response to the airport scenario as opposed to the bus scenario, which might be due to the infrequency with which most people visit airports, compared to using public ground transportation. Intent to curtail travel followed the same trend, with greater tendency to change travel plans in the train scenario than the bus scenario (d = 0.15, though not significant) and lower in the airport scenario than the bus scenario (d = 0.21). However, these results were not accompanied by corresponding differences in fear or anger; furthermore, the significant difference in travel behavior changes between the ground and aviation scenarios should be interpreted with caution, since there are far fewer feasible travel alternatives to flying than to using public ground transportation. Thus, we find suggestive evidence that the prospect of a lethal cyberattack was more troublesome for respondents than a more ‘conventional’ explosives attack, and that the prospect of an airport attack was less so.

The path model results and mean-level impacts of attack features suggest interesting avenues for further research on soft-target terrorism. It is worth exploring whether the effects of attack method and location on risk perception are restricted to our specific stimuli (cyber methods, airport location), or generalize to other attack features. Increased risk perception in the cyberattack scenario might be due to the novelty of lethal cyberterrorism, in which case, other novel terror attack methods might induce similar effects. Similarly, the attenuated risk perception responses in the airport condition (compared with the bus scenario) might be due to the relative infrequency with which most people visit airports, in which case, post-attack risk perception might be linked to an individual’s personal experience with the attack venue. Regarding the path model, future studies might explore the role of government trust in terrorism risk perception (which was not a significant predictor of any risk-related variables), and whether different facets of risk sensitivity map onto different affective states (such as fear vs. anger) in other domains. Thus, while some of our original hypotheses were not supported, we argue that our data can help researchers work toward a more fully integrated model of public terror response, especially in the increasingly prevalent domain of soft-target terrorism.

Among this study’s limitations was our survey’s reliance on single-item construct measures for fear and anger. We of course acknowledge the statistical and methodological issues that arise from single-item construct measures, and interpret our results with according caution; however, multiple studies have found that the predictive validity of single-item measures can rival that of brief multi-item measures (Bergkvist and Rossiter 2007; Drolet and Morrison 2001), even when measuring affect (Larsen et al. 2009). We should also note that one of our main dependent variables of interest, intended travel behavior change, might reflect variance in individuals’ life circumstances in addition to their reactions to the fictional terror attacks. Based on residence location (e.g., rural vs. urban) or work patterns (commute patterns, vacation time), some individuals may be more able to alter their travel behavior in response to terror-related worry than others, and future studies of this variable should take such information into account. Another more conceptual limitation of this experiment is its reliance on fictional scenarios rather than actual terror attacks. Thus, there is a need for field investigations to study real-time reactions to terror attacks (e.g., Lin et al. 2017). Yet experiments such as this one (and others that rely on the use of hypothetical terror scenarios, e.g., Burns and Slovic 2009; Cui et al. 2016; Rosoff et al. 2012, 2013) can be used to generate hypotheses about public reactions, which can then be validated in field studies of behavioral responses to actual terror events. We believe the insights uncovered in our model can inform such hypotheses, and may serve other risk perception researchers in attempt to more accurately model public responses to terrorism.