Discrete choice experiment (DCE) is a quantitative technique to investigate individual preferences (choice behaviour). The DCE has a solid foundation in random utility theory [21,22] and includes a Nobel prize-winning econometric approach [23]. Within a DCE, subjects are presented with several choice sets (see Fig 1 for a choice set example). In each choice set subjects are asked to choose between two or more alternatives. The alternatives are described by its characteristics (attributes) [24]. Those attributes are further specified by variants of that attribute (attribute levels). It is assumed that the subject’s preference for an alternative is determined by the levels of those attributes [24]. Resulting choices reveal an underlying utility function. The stated preferences allow the utility calculation of all alternatives (i.e. all permutations of attribute levels, including alternatives that are not presented to respondents). The DCE technique is mainstream in marketing, transport and environmental economics, where it is used to predict individual and collective choices. Its application in healthcare and public health has grown exponentially as the method is easy to apply and appears efficient at first glance [25,26]. The DCE approach combines (i) consumer theory [27] and (ii) random utility theory [28], which both assume that an individual acts rationally and always chooses the alternative with the highest satisfaction (i.e., utility), as well as (iii) experimental design theory and (iv) econometric analysis [26]. See Louviere et al. [29]; Hensher et al. [30]; Rose and Bliemer [31]; Lancsar and Louviere [32]; and Ryan et al. [33] for further details on conducting a DCE and its theoretical issues.

We followed recent guidelines for DCE practice [32,34]. A focus group study and literature [1,35] were used to obtain insights into relevant attributes and attribute levels to be included in the DCE regarding protective behaviour of citizens to transport accidents involving hazardous materials. For the focus group study, we conducted two focus group discussions with the general population living in the direct vicinity of a large waterway in the Netherlands (six and seven participants, respectively). A topic list based on the Protective Action Decision Model [35] was created to structure the discussions on protective behaviour of citizens regarding transport accidents involving hazardous materials. As a result, three transport accident attributes were selected: odour perception, smoke/vapour perception, and the proportion of people in the environment that are leaving at their own discretion. We aimed at selecting a sufficient wide range of attribute levels that could be realistic in the near future and were plausible and understandable for respondents (Table 1). In this research ammonia (toxic) and propane gas (explosive/flammable) were selected as hazardous materials because ammonia and propane (odorized by mercaptan) are transported across the river in large quantities.

The combination of the attributes and levels (Table 1) resulted in 60 possible/hypothetical transport accident scenarios. As fatigue of filling out a lengthy questionnaire prohibited presenting all scenarios to a single individual (i.e., a full factorial design), a subset of scenarios was required (i.e., a fractional factorial design) [36]. For the final selection of the scenarios to be presented in the questionnaire (i.e., the scenarios in which the respondent was asked to choose between three discrete protective behaviour options: staying, seeking shelter, or escaping (see Fig 1, for an example of a so-called choice set)), we generated a D-efficient DCE design [36, 37] using Ngene software (version 1.1.1, http://www.choice-metrics.com/). In order to generate such an efficient design, prior values for the behavioural coefficients of each of the attributes in the utility function were needed. To this end we conducted a pilot study (N = 40) first to collect prior information and to obtain insights in the feasibility and validity of the questionnaire and DCE data. Based on the pilot study (data not shown) we concluded that the target group (general population) was able to understand the DCE, could manage the length of the questionnaire, did not become confused by any unlikely combination of attribute levels, and had no difficulties in completing 12 choice sets of the DCE task.

The choice results from the pilot study (data not shown) were used to estimate behavioural coefficients to each of the attributes, which were used to locate an even more efficient DCE design for the main study (i.e., a design that had smaller standard errors [31,38], thereby increasing the reliability of the estimation of the behavioural coefficients). With this design, resulting in 24 scenarios, we were able to estimate all scenario specific main effects (i.e., to determine what influence the transport accident attributes had on citizens’ protective behaviour (staying, seeking shelter, or escaping)). Presenting a single individual with 24 choice sets was expected to result in a low return rate for the questionnaire and partially completed questionnaires [39]. Therefore, in order to reduce the burden on respondents and to take the pilot results into account, the 24 choice sets were blocked [30] into two sets of 12 choice tasks each. Each questionnaire started with detailed information about the visible accident (i.e. setting the scene) and the protective behaviour alternatives (S1 File). The main part of each questionnaire—randomly assigned to respondents—comprised a block of 12 choice sets. In each choice set, subjects were asked to consider the specific transport accident scenario as realistic and to choose the alternative that was most appealing to them: staying, seeking shelter, or escaping (see Fig 1 for an example of a choice set). The questionnaire further contained questions on background variables (e.g. age, gender, educational level, marital status) and a question assessing experienced difficulty of the questionnaire (5-point scale) (S1 File; the complete questionnaire is available from the authors on request).

A representative sample of 1,994 subjects aged 19–64 years was randomly recruited using the population registry of two cities (Vlissingen and Terneuzen) in the direct vicinity of a large waterway (Westerschelde) in the Southwest of the Netherlands. This population registry contains 31,907 subjects aged 19–64 years. Hence, our representative sample, agrees with a quota of 6.2%. We used age, gender and civil status as variables to guarantee representativeness. Using a rule of thumb as suggested by Orme [40], we strived to include at least 600 respondents. Subjects received an invitation letter and questionnaire by mail, and could return the questionnaire in a postage-paid envelope that was included in the mailing package. Due to practical reasons and positive past experiences, the survey was administered with the return of a mail survey as the only option presented for its completion. Two reminders were sent 1.5 and three weeks later in case of non-response. The information to participants explained that by filling out the questionnaire informed consent was given (S2 File). Under the Dutch law for the agreement on medical treatment, questionnaire surveys are not subject to approval by an institutional ethics committee. However, the Law for Protection of Personal Data requires informed consent and also procedures for the protection of personal privacy. These procedures are laid done in the Code of Conduct for Medical research (at www.federa.org), established by the Council of the Federation of Medical Scientific Societies. In this research project we have strictly adhered to these procedures and the data were analysed anonymously (in more detail, the data were anonymized prior to we–as authors—receiving them, and we–as authors—did not have access to participant names and addresses).

The DCE contained three dependent variables for each choice set. The DCE was analysed by taking each choice (i.e., protective behaviour options staying, seeking shelter, or escaping) made in every choice set as an observation. Taking our interest in preference heterogeneity into account as well as our sample size, a mixed logit model or a latent class model were both good alternatives to analyse the choice observations. Using the Nlogit software (http://www.limdep.com/) and taking the best model fit into account, the observations were analysed by a panel mixed logit model in error component form [41], also called error component model. The error component model is a parsimonious specification of the mixed logit model, where preference heterogeneity is modelled for only a specific set of variables. In our error component model, we capture preference heterogeneity for the possible behavior reactions. Hence, the alternative specific constants for types of reaction ‘hide’ and ‘escape’ were classified as random (assuming a normal distribution), while preferences for the other attributes were assumed to be homogeneous across respondents in the sample. The advantage of using an error component model over a mixed logit model with random coefficient for all attributes is that it can capture the most important elements of preference heterogeneity in a rather parsimonious way. The final specifications of the utility functions used were: Vseeking shelter=β0+β1odour_ammonia_weak+β2odour_ammonia_strong+β3odour_mercaptan_weak+β4odour_mercaptan_strong+β5smoke_vapour_ship+β6smoke_vapour_beach+β7leaving_peopleVescaping=β8+β9odour_ammonia_weak+β10odour_ammonia_strong+β11odour_mercaptan_weak+β12odour_mercaptan_strong+β13smoke_vapour_ship+β14smoke_vapour_beach+β15leaving_peopleVstaying=0 (Eq. 1) where Vrepresent the systematic utility(stated preference)that a respondent hasfor a protective behaviour option; β0,β8represent alternative specific constants for a protective behaviouroption(‘staying’acts as the reference level); β1−7,β9−15represent alternative specific parameter weights(or coefficients)associated with the attribute(level)s:odour perception,smoke/vapourperception,and the proportion of people in the environment that areleaving at their own discretion.

The value of each coefficient represents the importance that respondents assign to a certain attribute or attribute level. However, different attributes utilise different units of measurement. For example, the coefficient ‘the proportion of people in the environment that are leaving at their own discretion’ represents the importance per 10% of people in the environment that are leaving at their own discretion. That is, in a scenario where 50% of people in the environment are leaving at their own discretion, the coefficient should be multiplied by five in order to establish its relative contribution to utility compared to other attributes.

In addition to the parameter estimates, the estimation procedure also allows for tests of statistical significance. Statistical significance of a coefficient (p-value ≤0.05) indicates that respondents considered the attribute (or attribute level) important in making their choices (i.e., protective behaviour) in the DCE. The sign of the coefficient reflects whether the attribute (level) has a positive or negative effect on the specific stated protective behaviour option (utility). A priori we expected all attributes to be important, and that the attribute ‘the proportion of people in the environment that are leaving at their own discretion’ and stronger ammonia/mercaptan odour levels would have a positive effect on leaving the transport accident area. Additionally, sensitivity analyses were conducted to test whether specific respondent characteristics (e.g., age, gender, educational level) had a significant influence on the protective behaviour option made.

We also calculated the choice probabilities (the mean protective behaviour) to provide a way to convey DCE results to policy makers that are more easily understandable. We calculated the choice probabilities for a mild looking transport accident (i.e., a visible transport accident involving hazardous materials on a populated waterway without any odour perception or smoke/vapour perception, and where 0% of the people in the environment will leave the location) as well as for a severe looking transport accident (i.e., a visible transport accident involving hazardous materials on a populated waterway with a strong ammonia odour perception, smoke/vapour perception towards the beach/quay, and where 80% of the people in the environment will leave the location). The mild and severe looking accidents–representing the bookends of the spectrum for hazardous accidents—were chosen to help the reader in understanding the usefulness of such an analysis, and to show how these different accidents lead to a different protective behavior (stay, hide, or escape) of citizens. The choice probabilities were simulated based on the panel error component coefficients using 1,000 pseudo-random draws from the probability density functions estimated for the random error components.

The response to the questionnaire was 44% (881/1,994). Respondents had a mean age of 47 years (SD = 12), about half of the respondents were male, 26% had a high educational level, and they lived predominantly together with a partner (Table 2).

Weak odours did not influence respondents’ protective behaviour to seek shelter more compared to ‘no odour perception’ (Table 3; p = 0.58 and p = 0.10 for weak ammonia and mercaptan, respectively). However, strong odours made respondents seek shelter more (p<0.001), with ammonia having a stronger effect (coefficient values of 1.19 and 0.59 for strong ammonia and strong mercaptan odour perception, respectively; Table 3). For escape, which was much more preferred than seeking shelter (coefficient values of 1.90 and -0.42 respectively), any odour seemed to be important and very much odour made respondents want to escape. Smoke did not seem relevant for respondents to seek shelter or escape more, unless it came close to them. The transport accident attribute ‘proportion of people that are leaving’ proved to influence respondent’s protective behaviour to seek shelter and escape more (Table 3). The positive sign of this coefficient showed that the more people in the environment were seeking shelter or leaving at their own discretion, the more willing the respondent was to leave the transport accident area as well. The estimated standard deviations for the protective behaviour options ‘seeking shelter’ and ‘escaping’ were significant, which indicated stated preference heterogeneity among subjects for these protective behaviour options.

The sensitivity analysis showed that two characteristics of subjects had a significant influence on the stated preference for seeking shelter and escaping (see S1 Table). That is, males were less willing to seek shelter than females, and elderly people were more willing to escape than younger people.

In case of a mild looking transport accident involving hazardous materials at a river in a populated area, based on our DCE results the predicted probability that a subject would stay, seek shelter or escape was 62%, 13%, and 25%, respectively. In case of a severe looking transport accident these probabilities were 1%, 16%, and 83%, respectively.

Fig 2 shows–keeping all other attribute levels equal—that a perception of a strong ammonia odour had a substantial influence on the predicted probability that a respondent will seek shelter or escape more (9% and 40% more, respectively) compared to ‘no odour perception’. That is, the proportion of people that preferred staying instead of leaving decreased from 62% to 13% (i.e., 62% minus 9% more seeking shelter minus 40% more escaping). The proportion of people that preferred staying instead of leaving decreased from 62% to 16% in case of a strong mercaptan perception (i.e., 62% minus 3% more seeking shelter minus 43% more escaping). If smoke/vapour came close to individuals or half of the people in the environment were leaving, individuals that preferred staying decreased from 62% to 30% (i.e., 62% minus 4% more seeking shelter minus 28% more escaping) and 47% (i.e., 62% minus 3% more seeking shelter minus 12% more escaping), respectively.