1 Introduction

Stated preference surveys are used to measure the value of public goods to society. Despite their broad application in environmental economics and other areas, a remaining concern is the validity and reliability of stated preference value estimates. One of several factors shown to matter for stated preferences and potentially affecting the validity and reliability of the value estimates is the type and amount of information provided to survey respondents before preference elicitation (Blomquist and Whitehead 1998; Munro and Hanley 2001; Johnston et al. 2017). While the importance of different types of information provision is well established, less is known about the mechanisms underlying the frequently found effects of information on stated preferences. Based on an empirical study with varying information about the policy context of the valuation scenario, this paper investigates two new potential pathways of the information-induced effects on stated preferences.

How changes in provided information affect stated preferences has been researched for many years (e.g., Hoevenagel and van der Linden 1993; Munro and Hanley 2001; Czajkowski et al. 2016). The literature reports varying magnitudes and directions of such information effects. Importantly, the investigations differ in the type of information studied, which may contribute to differences in magnitudes, directions and underlying pathways of the effects. Our study focuses on descriptive information about the valuation scenario intended to increase respondents’ knowledge about the evaluated good, by specifying the current state or the content, extent, expected outcomes, policy context or mechanism of the proposed policy change. We refer to text conveying such information in a survey as information scripts throughout the paper. Our empirical analysis is based on information about the policy context of the valuation scenario in particular. We do not consider other types of survey scripts investigated in the stated preference literature that change respondents’ knowledge about economic incentives underlying their responses or provide any other technical information related to the preference elicitation task.Footnote 1

We aim to investigate two new potential pathways how the provision of information about the policy context of the valuation scenario may affect stated preferences. First, it may change respondents’ perceptions of the survey consequentiality, which, in turn, can matter for stated preferences. A survey is consequential when there is a positive probability that the survey results will influence the decision of policy makers regarding the provision of the considered good and the collection of the payment related to this provision (Carson and Groves 2007; Johnston et al. 2017). Second, the information may affect respondents’ perceptions of the credibility of the scenario considered for the valuation, and this in turn can influence stated preferences. Johnston et al. (2017) note that the valuation scenario should be seen as credible by respondents in order to derive valid value estimates from stated preference surveys. If the provision of information about the policy context of the valuation scenario affects perceptions of consequentiality and credibility, and the shifts in the perceptions in turn affect stated preferences, these two pathways may explain parts of the effects of the information provision on stated preferences.

This study utilizes data from a discrete choice experiment survey conducted in April and May 2019 on a representative sample of 1276 residents of the city of Bremen in Germany. The survey elicits respondents’ preferences towards extending urban green spaces as a climate change adaptation measure. To address the research question, we design two randomly assigned survey versions that differ with respect to the information displayed before eliciting preferences. Both versions provide necessary information for understanding the valuation scenario. One version, in addition, presents extended information about the scenario’s policy context. We estimate mixed logit and hybrid mixed logit models to analyze how the additional information affects stated preferences, how it shifts consequentiality and credibility perceptions, and how these perceptions matter for stated preferences.

Our study uses a new approach for analyzing how changes in the information about the valuation scenario can affect stated preferences, by applying the hybrid mixed logit framework to gain a new perspective on the pathways of information effects. Our results can deliver important insights for stated preference practitioners in constructing balanced information scripts. Understanding mechanisms that lead information effects to emerge is essential for correctly designing stated preference studies and for obtaining valid value estimates to support public decision making. Information provided on the policy context of the valuation scenario could plausibly affect perceptions of consequentiality and credibility. Findings on the importance of these pathways in explaining information effects can guide decisions on what information to include. Moreover, the study contributes to the line of research on the validity of stated preference methods. It provides additional evidence on the role of consequentiality perceptions and the little studied credibility perceptions for willingness to pay for a public policy project. Results may inform stated preference practitioners on whether and how these desirable characteristics of a valuation survey can be strengthened with specific information scripts.

Scripts in stated preference surveys involve many types of information, which can matter for the consequentiality and credibility perceptions. Although our empirical case study focuses on the sensitivity of stated preferences to variations in the information about the policy context of the valuation scenario, the studied pathways of consequentiality and credibility are likely also relevant for explaining effects on stated preferences induced by other types of information. While the role of survey scripts for perceived consequentiality has been given extensive attention in the recent stated preference literature, perceived credibility has been less scrutinised. For instance, survey scripts intended to strengthen consequentiality by informing that the results will be available to decision-makers (as examined, e.g., in Lloyd-Smith et al. (2019) and Zheng et al. (2021)) may affect credibility perceptions. The pathways of consequentiality and credibility may play a role in the effects of information going beyond the types of scripts in the focus of this paper.

2 Information Effects, Consequentiality and Credibility

For many years, stated preference reseachers have studied the effects of information on stated preferences, including information about the current state or the content, extent, expected outcomes, policy context or mechanism of the proposed policy change of the valuation scenario. The majority of the studies find that extending such information increases the value estimates for a range of environmental goods like forest and moorland management, water quality, and species conservation (e.g., Hoevenagel and van der Linden 1993; Munro and Hanley 2001; Bateman and Mawby 2004; Rambonilaza and Brahic 2016; Vanermen et al. 2021). Yet, the empirical evidence is not consistent. A few studies report no or very limited effects of such additional information (e.g., Berrens et al. 2004; MacMillan et al. 2006; Needham et al. 2018). Varying the way how the information is provided is also shown to affect stated preferences. Ajzen et al. (1996) find that an information script using stronger arguments has a larger effect on willingness to pay, Czajkowski et al. (2016) observe a difference in stated preferences between a less and more positively framed information script, and in the study by Yang and Hobbs (2020) a narrative rather than a scientific information script leads to larger willingness to pay estimates.

The literature does not provide a consistent explanation for the effects of extended information about the valuation scenario on stated preferences. The effects can emerge through various pathways. Munro and Hanley (2001) argue that respondents rationally adapt their decisions after learning about the benefits from the environmental good, which leads to higher value estimates. Similarly, Bergstrom et al. (1990) develop a theoretical model showing that information about the benefits derived from the environmental good can influence willingness to pay by altering the perceived marginal utility from the good. This is in line with the results obtained by Hoehn and Randall (2002) and Hasselström and Håkansson (2014) that the additional information affects only those respondents for whom the information is new, and it aligns with the findings that the effect of the additional information is particularly strong for goods that are unfamiliar (Bateman and Mawby 2004) and of high personal relevance (Ajzen et al. 1996).

As a different explanation, Hoevenagel and van der Linden (1993) suggest that respondents use availability heuristics, as specified by Tversky and Kahneman (1974), to simplify the preference elicitation task: respondents assign greater importance to attributes with more extensive information. Other authors explain information effects with directional context effects: the survey methodology literature finds that questions displayed earlier in a questionnaire may provide an interpretative framework influencing responses to questions asked further in the survey (Tourangeau and Rasinski 1988; Tourangeau et al. 2000; Moore 2002; Dillman 2011). In the context of stated preferences, Pouta (2004) and Liebe et al. (2016) show that asking attitudinal questions before eliciting preferences can increase the value estimates. Information scripts may provide a similar interpretative framework. Finally, experimenter demand effects could explain the influence of information scripts. This concept refers to respondents adapting their behavior according to cues about how the experimenter may expect them to behave (Quidt et al. 2018). Information scripts could constitute such a cue for respondents about expected responses in a preference elicitation task.

We argue there could be other possible pathways leading to effects of information about the valuation scenario, which, to the best of our knowledge, have not been addressed in the stated preference literature so far. One of such pathways is that information scripts may affect respondents’ perceptions of the survey consequentiality, which, in turn, can matter for stated preferences. Consequentiality has been identified as a necessary condition for truthful preference disclosure and, therefore, is required for valid value estimates (e.g., Vossler et al. 2012; Carson et al. 2014). In a consequential survey, a respondent believes that the survey results may influence the decision of policy makers regarding the provision of the considered good and the collection of the payment related to this provision (Carson and Groves 2007; Johnston et al. 2017). Most previous studies find that stated preference value estimates increase with the strength of the consequentiality belief (Herriges et al. 2010; Vossler and Watson 2013; Groothuis et al. 2017; Vossler et al. 2018). However, some studies find no relationship (Broadbent 2012; Oehlmann and Meyerhoff 2017) or observe value estimates to decrease with stronger consequentiality beliefs (Vossler et al. 2012). If information scripts influence consequentiality perceptions (as some consequentiality scripts are shown to do; e.g., Oehlmann and Meyerhoff 2017), the effects of information scripts on stated preferences could be partly explained by this indirect pathway of shifted consequentiality perceptions.

Another unexplored pathway could be through respondents’ perceptions of the credibility of a valuation scenario considered in the survey. These perceptions are likely influenced by the information scripts discussed here and at the same time they may affect stated preferences. Credibility of the valuation scenario is a desired feature of a survey and important for valid value estimates (Flores and Strong 2007; Johnston et al. 2017). Kataria et al. (2012) investigate to what extent respondents believe in the provided information about the current state and the policy change. They find that respondents who view the information as more credible have a higher willingness to pay for improving river water quality. In contrast, Vasquez et al. (2019) do not find significantly different willingness to pay among respondents differing in the degree of how credible they perceive the presented current state of water quality. If information scripts influence perceived credibility and these perceptions affect stated preferences, this pathway could also partly explain effects of the information on stated preferences.

Our study contributes to the reviewed literature in several ways. First, it delivers new insights into the frequently studied effects of the provision of additional information about the valuation scenario on stated preferences. Besides investigating how the commonly found information effects (e.g., Hoevenagel and van der Linden 1993; Munro and Hanley 2001; Bateman and Mawby 2004; Rambonilaza and Brahic 2016; Vanermen et al. 2021) transfer to urban green and information about the policy context of climate change adaptation, our study can also improve understanding of the mechanism of the effects by analyzing the two potential pathways of shifts in perceived consequentiality and credibility. To the best of our knowledge, no previous research has investigated the roles of consequentiality and credibility perceptions for explaining effects of information scripts on stated preferences. Second, our study extends the literature on survey consequentiality, which has been frequently examined in recent years because of its importance for the validity of value estimates (e.g., Herriges et al. 2010; Vossler et al. 2012; Vossler and Watson 2013; Carson et al. 2014; Vossler et al. 2018). However, to the best of our knowledge, our study is the first to test whether information scripts can strengthen perceived consequentiality. Third, our investigation provides new insights into the role of perceived credibility of the valuation scenario. Although similarly important for the validity of value estimates, there have been few studies so far devoted to this question (Kataria et al. 2012; Vasquez et al. 2019). Our results may help to clarify their inconsistent findings on the relation between perceived credibility and willingness to pay, and our study is the first to test whether information scripts can strengthen perceived credibility.

3 Empirical Data

The survey was conducted in the German city of Bremen and implemented online. It elicited preferences of residents towards extending urban green spaces as a climate change adaptation measure. The value estimates were used in a cost-benefit analysis of the climate change adaptation strategy for the city. The survey provided respondents with detailed information about the considered policy scenario and the attributes characterizing it. Half of respondents received additional contextual information about the climate change adaptation strategy of the city. After the presentation of the information, respondents participated in a discrete choice experiment. Follow-up questions queried about respondents’ perceptions about the survey consequentiality and the credibility of the policy scenario. The questionnaire ended with socio-demographic questions.

3.1 Discrete Choice Experiment

The survey considered a citywide policy aimed at extending urban green spaces. The policy was characterized by five attributes, as presented in Table 1. The final selection of the attributes, their levels and definitions were guided by expert insights from the city administration collected in two workshops. The selection was tested in an online pretest with 115 respondents recruited from an online panel.

Table 1 Discrete choice experiment attributes and their levels
Full size table

The attributes included the number of street trees per 100 meters, the share of green areas in the city’s total area, the share of extensive green roofs (i.e., those with thin substratum), the share of intensive green roofs (i.e., those with thick substratum) and an annual cost per individual. Each of the four non-cost attributes took one of three possible levels. For each attribute, one of the levels corresponded to the current average level in the city. The two other levels represented extensions compared to the current situation. The status quo levels were derived from geographic information system data and were verified by experts from Bremen’s public administration. The monetary attribute was defined as a compulsory yearly payment for every resident of the city that the city would collect and spend exclusively on the development and maintenance of the urban green.

The attributes were explained in detail on separate screens of the survey prior to the discrete choice experiment. Respondents were informed about the current average levels in Bremen for each non-monetary attribute and these levels were labelled as “As today” in the discrete choice experiment.

The discrete choice experiment consisted of a sequence of nine choice tasks. Every task included two policy alternatives and a status quo alternative, out of which respondents were asked to choose their most preferred option. The right-hand side alternative was always the status quo labelled as Current state, with all non-cost attribute levels set to the current city average levels and no cost. The two policy alternatives involved changes to the current state and were named as Combination A and Combination B. The cost for the hypothetical alternatives ranged between 5 and 400 Euro. Figure 1 shows an example choice task.

Fig. 1
figure 1

An example choice task

Full size image

The design involved 36 choice tasks split into four blocks. Each respondent was presented with a randomly assigned block of nine choice tasks. The design was created with the Stata module dcreate, using the modified Fedorov algorithm to maximize the D-efficiency for dummy coded attributes in a multinominal logit model (Hole 2017). Priors from the pretest were used. The design was compared to various other designs and tested via simulation in terms of efficiency and bias.

3.2 Information Script Treatments

Prior to the discrete choice experiment, all respondents were provided with information about the good to be valued and its characteristics. The information included a description of the five attributes, current levels of the attributes in Bremen and expected effects of the proposed changes in the attribute levels on the cityscape, leisure usability, biodiversity, water retention and heat mitigation. Respondents were further reminded via a standard script that the city administration would implement the policy and collect the payments depending on the survey outcome.

The survey involved two randomly assigned treatments that differed in whether additional information about the policy context of the valuation scenario was provided or not. Half of respondents were assigned to the No Script sample and did not see any additional information. The other half of respondents were assigned to the Script sample and were shown an additional script before the choice tasks about the policy context of the proposed policy change, namely describing the city’s climate change adaptation strategy, expected impacts of climate change and that the urban green attributes are part of this strategy as measures the city of Bremen could apply. The script reads as follows:Footnote 2

The Senate of Bremen adopted the climate change adaptation strategy for Bremen in April 2018. The strategy document explains the consequences of climate change for the city of Bremen. Strong rain, river and storm floods will become more likely. The strategy document predicts a rising risk of flooding with property damages, such as flooded basements and underground garages. According to the strategy document, heat waves will also become more likely. These can reduce your productivity and strain your cardiovascular system. The climate change adaptation strategy mentions several measures which the city of Bremen could apply. The first part of this survey focuses on some of these measures.

Only respondents in the Script sample were explicitly told that the considered urban green measures were part of the city’s climate change adaptation strategy. The respondents in the No Script sample instead only were displayed this short introduction:

No Script: The first part of this survey focuses on possible urban green measures for the city of Bremen.

Other than the two differences in the survey script explained above, the survey questionnaires used in the treatments were identical. When presenting the hypotheses, econometric approach and results below, we refer to the additional information provided to the Script sample respondents as the information script.

3.3 Elicitation of Perceptions of Consequentiality and Credibility

Information about perceived consequentiality and credibility was collected after the discrete choice experiment. Although related, consequentiality and credibility are distinct concepts. Consequentiality refers to the survey responses having actual consequences by influencing final decisions of policy makers. Credibility refers to the proposed changes being realistic, that is, whether it is plausible that they could be implemented in general, independent of any potential influence of the survey. This distinction is confirmed by the respondents’ answers: the two consequentiality variables are not strongly correlated with the four credibility variables, with Spearman rank correlation coefficients ranging from 0.12 to 0.29.

Recent works suggest that consequentiality perceptions could be elicited with the use of questions that differentiate between respondents’ perceptions towards policy consequentiality and payment consequentiality (e.g., Vossler et al. 2018; Zawojska et al. 2019a, b). Following the definition in Johnston et al. (2017), policy consequentiality can be understood as a positive probability that survey responses will influence decisions related to the outcome in question; payment consequentiality corresponds to a positive probability that the payment for the considered project will be collected if the project is implemented.

We used two questions targeting the measurement of perceived policy consequentiality and perceived payment consequentiality, respectively: “To what degree do you believe that your responses will affect which measures will be implemented in the city of Bremen?”; “To what degree do you believe that your responses will affect whether you will have to pay the additional cost if the measures are implemented?” The Likert response scale included six levels labelled as “I strongly believe,” “I rather believe,” “I neither believe, nor do not believe,” “I rather do not believe,” “I do not believe at all,” and “I do not know.”

Similarly to Kataria et al. (2012), we collected data on perceived credibility of the policy change, by querying respondents about the likelihood that the proposed changes could be realized. For each of the four non-cost attributes we asked: “How likely do you think it is that the proposed extent of the changes can actually be realized?” The Likert response scale included six levels labelled, respectively, as “very likely,” “rather likely,” “neither likely nor unlikely,” “rather unlikely,” “very unlikely,” and “I do not know.”

3.4 Survey Implementation

The selection of the attributes, their levels and descriptions were guided by the expertise of administration officials of the city of Bremen. Two workshops with representatives of the city administration took place in June and October 2018 and additional interviews were conducted in follow-up bilateral meetings and phone calls. The questionnaire was then developed under consultation with representatives from the general population through personal interviews. A pilot study with 115 respondents was used to assess and improve the questionnaire and the experimental design.

The final survey was implemented online as Computer-Assisted Web Interviews (CAWI) and consisted of 52 screens. The survey was administered by a professional public opinion polling agency from April to May 2019. 1276 residents of Bremen and adjacent districts completed the questionnaire. Participants were recruited with two modes. Firstly, 1011 respondents belonged to a panel that the polling agency recruited offline without the possibility of self-enrollment. Secondly, to increase the sample size, 5000 letters with a link to the online survey were sent via mail to a random sample of home addresses provided by the city of Bremen. 265 additional respondents were recruited this way. For the empirical analysis, we excluded 98 respondents who responded “I do not know” to at least one of the questions on consequentiality and credibility perceptions. Therefore, our investigation below is based on a sample of 1178 respondents.

Table 2 shows socio-demographic characteristics of the respondents, separated into the Script and No script samples. Differences in these characteristics between the two samples are negligible. A t-test for the difference in means for age and \(\chi ^2\)-tests for differences in shares for the remaining characteristics indicate no statistically significant differences between the treatment groups.

Table 2 Socio-demographic characteristics
Full size table

4 Empirical Strategy

4.1 Hypotheses

Based on the literature reviewed in Sect. 2, we derive the following hypotheses for our empirical analysis:

  1. 1.

    Respondents who see the additional information script exhibit increased willingness to pay for extending urban green spaces.

Numerous studies find that additional information about the valuation scenario increases the willingness to pay for a range of environmental goods like forest and moorland management, water quality, and species conservation (e.g., Hoevenagel and van der Linden 1993; Munro and Hanley 2001; Bateman and Mawby 2004; Rambonilaza and Brahic 2016; Vanermen et al. 2021). We expect a similar effect in the context of urban green and climate change adaptation. The script was designed to meet the criteria that have been identified by previous studies as factors for particularly strong information effects. Namely, the extension of urban green and the adaptation to climate change in the city of their residence is likely of high personal relevance to respondents (Ajzen et al. 1996). This was reinforced by using the second person form in the script. Although the climate adaptation strategy in Bremen was adopted one year before the survey, part of the information provided is likely new to most respondents (Hasselström and Håkansson 2014), as only 6% indicated in the survey that they knew the strategy and its content. Also, the script avoids a scientific style (Yang and Hobbs 2020).

  1. 2.

    Respondents who see the additional information script indicate stronger perceived survey consequentiality.

Some studies find that targeted scripts describing potential consequences of the survey outcome can strengthen perceptions of survey consequentiality (e.g., Oehlmann and Meyerhoff 2017). However, opposite results are also reported (e.g., Czajkowski et al. 2017). The information script in this study was designed to potentially strengthen perceived consequentiality by presenting the climate change adaptation strategy for Bremen and explaining that the urban green attributes are mentioned as measures that the city of Bremen can apply. This emphasizes to respondents that their decisions in the choice experiment concern the implementation of an actual strategy of the city, and therefore may strengthen the perception that the results of the survey will influence decisions of policy makers.

  1. 3.

    Respondents who see the additional information script indicate stronger perceived credibility of the valuation scenario.

To the best of our knowledge, there has been no research yet on whether survey scripts can affect the perceived credibility of the valuations scenario. The information script in this study was designed to potentially strengthen perceived credibility by explaining that the urban green attributes are measures listed in the climate change adaptation strategy. Respondents may find the proposed policy of extending urban green more plausible and realistic when learning that it is part of a strategy developed by expert policy makers.

  1. 4.

    Respondents who have stronger perceptions of the survey consequentiality exhibit larger willingness to pay.

Most previous studies find that willingness to pay increases with the strength of the consequentiality perceptions (e.g., Herriges et al. 2010; Vossler and Watson 2013; Groothuis et al. 2017; Vossler et al. 2018). Therefore, we expect to find the same relationship in our study.

  1. 5.

    Respondents who have stronger perceptions of the credibility of the valuation scenario exhibit larger willingness to pay.

We are aware of only two studies investigating the relationship between perceptions of scenario credibility and willingness to pay. Kataria et al. (2012) find that strong credibility perceptions correlate with larger willingness to pay. In contrast, Vasquez et al. (2019) do not find a significant relationship.

4.2 Econometric Approach

We estimate two models to answer the research question whether effects of information scripts on stated preferences can arise because of information-induced changes in perceptions about the survey consequentiality and the scenario credibility. Model I is a mixed logit model (Greene 2011) and uses only the collected data on stated preferences. Model II is a hybrid choice model and uses the data on both the stated preferences and the perceptions, also known as an integrated choice and latent variable approach (Ben-Akiva et al. 2002). Model I is equivalent to the discrete choice component of Model II with a null latent variable vector. For this reason, the description below focuses on a more general, hybrid choice modelling framework.

A hybrid choice model is a flexible tool that allows for including unobservable characteristics of individuals, such as perceptions about consequentiality and credibility, into a random utility framework. The unobservable perceptions enter the model indirectly as latent variables, since direct inclusion of self-reported measures of the perceptions in choice models may lead to econometric issues, such as a measurement error. Hybrid choice models are being increasingly used to analyze choice behavior of individuals in the area of environmental economics (e.g., Faccioli et al. 2020; Abate et al. 2020) and elsewhere (e.g., Thorhauge et al. 2019; Schmid and Axhausen 2019; Albaladejo and Diaz-Delfa 2020; Golebiowska et al. 2020).

However, hybrid choice models come at a cost. The models are more difficult to estimate and bear a risk of overfitting. Especially models estimated on small sample sizes are highly problematic. In their simulation study, Vij and Walker (2016) could show that hybrid models are performing well with sample sizes of 500 respondents or more. In many cases, model fit in hybrid models does not improve over simpler models while the potential biases are not reduced (Vij and Walker 2016; Mariel and Meyerhoff 2016). Mariel and Meyerhoff (2016) and Vij and Walker (2016) suggest using such models only when the focus is on behavioral insights rather than prediction or model fit.

In our case, we use a hybrid choice model because (a) we have a large sample size allowing us to reliably recover parameters, and (b) our main focus is on understanding the indirect effects of information provision, i.e., the focus is on what hybrid choice models are good at. We employ the hybrid choice mixed logit approach as proposed by Czajkowski et al. (2017), which combines the standard mixed logit model (Greene 2011) with the multiple indicators and multiple causes model (Jöreskog and Goldberger 1975). We use the hybrid choice model to find a relationship between the information treatment, respondents’ unobservable perceptions about the consequentiality and credibility, and their preferences towards the considered green climate change adaptation measures in Bremen.

The hybrid choice model employed in our study involves three components: a discrete choice model part, structural equations and measurement equations. The components are estimated simultaneously and linked by latent variables that are used to capture unobservable perceptions of respondents. We consider two latent variables: the first one is assumed to explain Likert-scale responses to the four questions querying about respondents’ perceptions about credibility of the considered policy scenario attributes; and the second one is assumed to explain Likert-scale answers regarding respondents’ perceptions about the two aspects of the survey consequentiality (i.e., policy and payment consequentiality). For ease of representation, we henceforth refer to these variables as latent credibility and latent consequentiality, respectively. In the measurement equations, correlations between the latent variables, which capture directly unobservable perceptions of the respondents, and indicator measures of the underlying unobservable perceptions are modelled. For each latent variable, there is one structural equation. In the structural equations, the information treatment variable is used to explain potential variation in the respective latent variable. The discrete choice component is a mixed logit model that additionally includes the information treatment, latent credibility and latent consequentiality to explain possible shifts in means of the preference parameters. Each of these three components of the hybrid choice model is presented in detail below.

4.2.1 Discrete Choice Component

The discrete choice component depicts the decision process of respondents when making a selection in the discrete choice experiment. Modelling of preferences disclosed through such choices is typically based on a random utility framework (McFadden 1974). According to the framework, the utility of individual \(i\) from selecting alternative (policy scenario) \(j\) in choice task \(t\), \(U_{ijt}(\cdot )\), depends on observed characteristics of the policy, including non-monetary choice task attributes, \(X_{ijt}\), and a monetary attribute, cost \(C_{ijt}\), and on unobserved idiosyncrasies represented by a stochastic component \(e_{ijt}\). Formally, it can be written as

$$\begin{aligned} U_{ijt}(\cdot )= \boldsymbol{\beta}_{{\varvec{i}}}^{'} {{\varvec{X}}}_{{\varvec{ijt}}} - \alpha _{i}C_{ijt} +e_{ijt} \end{aligned}$$
(1)

where \(\boldsymbol{\beta}_{{\varvec{i}}}\) is a vector of individual-specific preference parameters (i.e., marginal utilities from the policy attributes) and \(\alpha _{i}\) is a cost parameter representing marginal utility of income. All marginal utility parameters are individual-specific, as suggested by indexing over \(i\). This allows for heterogeneous preferences among respondents, leading to the mixed logit specification. Instead of estimating the marginal utility parameters separately for every respondent, we follow the standard practice and assume that the parameters are from a multivariate distribution, and allow for non-zero correlation of the parameters (Train 2009).

The underlying model in (1) may be estimated in either preference space or willingness to pay space (Train and Weeks 2005). Both specifications are behaviorally equivalent, but when estimating the model in willingness to pay space, preference parameters can be readily interpreted as willingness to pay amounts. We employ this approach. To derive the willingness to pay space model, we first divide all arguments in (1) by the logit scale parameter \(\theta _i\) to get

$$\begin{aligned} U_{ijt}(\cdot )= \boldsymbol{\gamma}_{{\varvec{i}}}' {{\varvec{X}}}_{{\varvec{ijt}}} - \lambda _{i}C_{ijt} +\epsilon _{ijt} \end{aligned}$$
(2)

where \(\gamma _i=\frac{\beta _i}{\theta _i}\) is a vector of preference-space coefficients on non-monetary policy attributes, \(-\lambda =\frac{-\alpha _i}{\theta _i}\) is a preference space coefficient on the policy cost, and \(\epsilon _{ijt}\) is an error term with an i.i.d. type I extreme value distribution and constant variance \(var(\epsilon _{ijt})=\frac{\pi ^2}{6}\) (Scarpa et al. 2008; Train and Weeks 2005).

Marginal willingness to pay values for changes in the non-monetary policy attributes can be calculated as a ratio of the coefficients on these non-monetary attributes and the cost coefficient, that is, as \(\omega _i=\frac{\gamma _i}{\lambda _i}=\frac{\beta _i}{\alpha _i}\). We thus reformulate (2) to obtain the willingness to pay space specification (Train and Weeks 2005),

$$\begin{aligned} U_{ijt}(\cdot )= \lambda _i\left[ \left( \frac{\boldsymbol{\gamma}_{{\varvec{i}}}}{\boldsymbol{\lambda}_{{\varvec{i}}}}'\right) {{\varvec{X}}}_{{{\varvec{ijt}}}}-C_{ijt}\right] +\epsilon _{ijt} = \lambda _i[{\boldsymbol{\omega}_{{\varvec{i}}}'{{\varvec{X}}}_{{{\varvec{ijt}}}}}-C_{ijt}] +\epsilon _{ijt} \end{aligned}$$
(3)

The elements of vector \(\boldsymbol{\omega}_{{\varvec{i}}}\) are random coefficients assumed to be normally distributed. To ensure a positive marginal utility of income, we follow the standard practice and define \(\lambda _i=e^{\nu _i}\), where \({\nu _i}\) is the underlying latent normal factor that specifies the lognormally distributed cost coefficient (Scarpa et al. 2008; Thiene and Scarpa 2009).

In order to investigate the relation of consequentiality and credibility perceptions with stated preferences, we extend (3) to allow the random coefficients to be a function of individual-specific latent variables, denoted by vector \({{\varvec{LV}}}_{{\varvec{i}}}\) (i.e., latent credibility and latent consequentiality), and information treatment variable \(S_i\). We hence specify the vector of willingness to pay parameters as

$$\begin{aligned} \boldsymbol{\omega}_{{\varvec{i}}}=\boldsymbol{\omega}^*_{{\varvec{i}}}+\boldsymbol{\delta}'{{\varvec{LV}}}_{{\varvec{i}}}+\varrho S_i \end{aligned}$$
(4)

where \(\boldsymbol{\omega}^*_{{\varvec{i}}}\) has a multivariate normal distribution with a set of means and a covariance matrix to be estimated; \(\sigma\) and \(\varrho\) are vectors of parameters to be estimated and \(S_i\) is a binary variable that takes a value of one if respondent \(i\) faced the additional information script and zero otherwise.

Following the same notation, we redefine the cost coefficient as

$$\begin{aligned} \lambda _i=e^{\nu _i+\boldsymbol{\psi}'{{\varvec{LV}}}_{{\varvec{i}}}+\tau S_i} \end{aligned}$$
(5)

with parameters \(\psi\) and \(\tau\) be estimated.

4.2.2 Measurement Component

Latent variables in hybrid choice models are used to capture individual characteristics that are not directly observable and cannot be objectively measured. Instead of exact measures, the models rely on the use of indicators of the unobservable characteristics, which are expected to be correlated with the latent variables. Hence, in the measurement equations, the two latent variables are used to explain respective indicators on the perceived survey consequentiality and the perceived policy scenario credibility. Given the discrete and ordinal nature of the indicators (Likert-scale responses), we specify the measurement equations as ordered probit regressions. Formally, this relationship can be represented as

$$\begin{aligned} {{\varvec{I}}}^{{\varvec{*}}}_{{\varvec{i}}}={{\varvec{LV}}}_{{\varvec{i}}\boldsymbol{\varGamma}}+\boldsymbol{\eta}_{{\varvec{i}}} \end{aligned}$$
(6)

where \({\varvec{I}}^{{\varvec{*}}}_{{\varvec{i}}}\) is a vector of the indicator variables (i.e., measures of the unobservable perceptions), \(\boldsymbol{\varGamma}\) is a matrix of coefficients to be estimated and \(\boldsymbol{\eta}_{{\varvec{i}}}\) denotes a vector of error terms assumed to come from a multivariate normal distribution with zero means and an identity covariance matrix. The dependent variables in vector \({{\varvec{I}}}^{{\varvec{*}}}_{{\varvec{i}}}\) on the left-hand side of (6) are characterized by five ordered levels, corresponding to different levels of agreement with Likert-scale questions, which can be written as

$$\begin{aligned} \begin{array}{lcr} {\varvec{I}}_{{\varvec{i}}}=1 &{} if &{}\boldsymbol{\mu}_\mathbf{0}< {{\varvec{I}}}^{{\varvec{*}}}_{{\varvec{i}}} \le \boldsymbol{\mu}_\mathbf{1} \\ {\varvec{I}}_{{\varvec{i}}}=2 &{} if &{}\boldsymbol{\mu}_\mathbf{1}< {{\varvec{I}}}^{{\varvec{*}}}_{{\varvec{i}}} \le \boldsymbol{\mu}_\mathbf{2} \\ {\varvec{I}}_{{\varvec{i}}}=3 &{} if &{}\boldsymbol{\mu}_\mathbf{2}< {{\varvec{I}}}^{{\varvec{*}}}_{{\varvec{i}}} \le \boldsymbol{\mu}_\mathbf{3} \\ {\varvec{I}}_{{\varvec{i}}}=4 &{} if &{}\boldsymbol{\mu}_\mathbf{3}< {{\varvec{I}}}^{{\varvec{*}}}_{{\varvec{i}}} \le \boldsymbol{\mu}_\mathbf{4} \\ {\varvec{I}}_{{\varvec{i}}}=5 &{} if &{}\boldsymbol{\mu}_\mathbf{4}< {{\varvec{I}}}^{{\varvec{*}}}_{{\varvec{i}}} \le \boldsymbol{\mu}_\mathbf{5} \\ \end{array} \end{aligned}$$
(7)

where the numbers from 1 to 5 correspond to levels from strongly disagree to strongly agree, respectively, and \(\boldsymbol{\mu}_\mathbf{0},\boldsymbol{\mu}_\mathbf{1},\boldsymbol{\mu}_\mathbf{2},\boldsymbol{\mu}_\mathbf{3},\boldsymbol{\mu}_\mathbf{4},\boldsymbol{\mu}_\mathbf{5}\) are vectors of threshold parameters, with each element of vector \(\boldsymbol{\mu}_\mathbf{0}\) being equal to \(-\infty\) and each element of \(\boldsymbol{\mu}_\mathbf{5}\)being equal to \(+\infty\). Hence, \(\boldsymbol{\mu}_\mathbf{1}\),\(\boldsymbol{\mu}_\mathbf{2}\), \(\boldsymbol{\mu}_\mathbf{3}\) and \(\boldsymbol{\mu}_\mathbf{4}\) need to be estimated.

To facilitate interpretation, the means of the latent variables are normalized to zero, and to assure identification, their variances are normalized to one (Daly et al. 2012; Raveau et al. 2012). As a result, all latent variables have the same scale and, therefore, their relative importance can easily be assessed.

4.2.3 Structural Component

To estimate the effect of the information script on the unobservable perceptions, we include a structural component in our hybrid choice model, where latent variables \({{\varvec{LV}}}_{{\varvec{i}}}\) are explained by information treatment variable \(S_i\). This relationship can be described by the following equation

$$\begin{aligned} {{{\varvec{LV}}}}_{{\varvec{i}}}=S_i\boldsymbol{\varPsi}+\xi _i \end{aligned}$$
(8)

with a vector of coefficients \(\boldsymbol{\varPsi}\) to be estimated and error terms \(\xi _i\).

4.3 Model Estimation

All components of the hybrid choice model are estimated jointly with the maximum simulated likelihood method. Similarly, the mixed logit model is estimated using this method. It is known that results obtained with the method can be sensitive to the selection of starting values, optimization techniques and convergence criteria (Czajkowski et al. 2017). To make sure our findings are robust, we have estimated various specifications of the models and employed various starting values. As a large number of draws is recommended for lowering the simulation error and increasing the probability of unraveling identification issues (Czajkowski and Budziński 2019), we use 10,000 Sobol draws with a random linear scramble and a random digital shift in the final models shown in the paper. The models are estimated in Matlab using a package for modelling discrete choice experiment data, which is available at https://github.com/czaj/DCE under CC BY 4.0 license.

5 Results

5.1 Mixed Logit Model

To investigate the effect of the information script on stated preferences without taking into account the studied pathways (hypothesis 1), we estimate the mixed logit model in willingness to pay space (Model I). Results are provided in column “Model I” in Table 3. Coefficients of the non-cost attributes in the part “Means” show the No Script sample’s estimated average marginal willingness to pay values per year in euro for an increase in the attributes by one tree per 100 meter or by one percentage point. Respondents in the No Script sample are willing to pay on average 29 euros per year for an increased number of street trees in Bremen by one additional tree per 100 meters of street, 23 euros per year for the extension of green areas by one percentage point of the total city area, 2 euros per year for one additional extensive green roof per 100 roofs, and 12 euros per year for one additional intensive green roof per 100 roofs. The status quo coefficient is negative, suggesting that respondents are on average willing to pay 20 euros for implementing one of the two combinations of urban green measures rather than maintaining the status quo, holding all attributes constant. All estimated standard deviations for non-cost parameters are statistically significant and larger than the estimated means (part “Standard deviations” in Table 3). This indicates substantial preference heterogeneity with a considerable share of respondents exhibiting negative willingness to pay.

The coefficients in the part “Script interactions” allow us to evaluate the information effect—that is, whether the means in the Script sample differ from the means in the No Script sample. For the non-cost attributes, all interaction coefficients are positive and statistically significant at least on the 10% level. That is, respondents who see the information script are willing to pay more for increases in these attributes than those who do not see it. The interaction with the status quo is negative: respondents who see the information script are willing to pay more for the implementation of one of the two combinations of urban green measures rather than maintaining the status quo, holding all attributes constant. This difference between the two samples is highly significant (p value \(<0.001\)). This finding is reflected in the frequency with which respondents chose the status quo alternative. Respondents who do not see the script choose the status quo alternative in 28.1% of choices. Respondents who see the script choose the status quo alternative only in 20.7% of choices. These results confirm our hypothesis 1 and are in line with the majority of the literature that finds increased value estimates upon provision of additional information about the valuation scenario (Munro and Hanley 2001; Bateman and Mawby 2004).

5.2 Hybrid Choice Model

To investigate whether part of the information effect identified in Model I can be attributed to a shift in perceptions of the survey consequentiality and the scenario credibility, we analyse the data with a hybrid choice framework (Model II). The results for the discrete choice component of the model are shown in the last column of Table 3; the results of the measurement equations are reported in Table 4 and Table 5; and the results of the structural equations are provided in Table 6. The bottom part of Table 3 contains in addition the diagnostics statistics for the entire hybrid choice model, jointly for the three components. Note that most diagnostic indicators cannot be directly compared between Model I and Model II because for Model II they are based on an extended set of equations and data.

Table 3 Results of mixed logit (Model I) and the mixed-logit component of hybrid choice (Model II)
Full size table

In the measurement equations (Table 4), one latent variable (LV1) is used to explain responses to the four questions on the credibility of each attribute. All four coefficients are positive and statistically significant. This implies that stronger stated credibility in all questions consistently corresponds to higher values of the latent variable. We refer to this variable as latent credibility. The second latent variable (LV2) is used to explain responses to the questions on perceived payment and policy consequentiality (Table 5). The significantly positive coefficient estimates in these measurement equations suggest that stronger stated consequentiality in both questions consistently corresponds to higher values of the latent variable. We refer to this variable as latent consequentiality.

Table 4 Model II—measurement equations for credibility (ordered probit)
Full size table
Table 5 Model II—measurement equations for consequentiality (ordered probit)
Full size table

The structural equations (Table 6) estimate the effect of the information script on the latent perception variables. The script has a significantly positive impact on the latent credibility (p value = 0.02). That is, respondents of the Script sample perceive the credibility of the proposed attributes as stronger than respondents in the No Script sample, confirming hypothesis 3. The coefficient of the script effect on the latent consequentiality is also positive, but statistically not significantly different from zero (p value = 0.31). Therefore, we cannot confirm hypthesis 2. These findings make the credibility pathway possible as a potential explanation of the information effect on stated preferences, but do not support a pathway via shifts in consequentiality perceptions.

Table 6 Model II—structural equations
Full size table

Table 3 shows that willingness to pay values and standard deviations in the discrete choice component of Model II are very similar to those obtained in Model I. The interactions of the latent credibility with all non-cost attributes are statistically significantly positive. That is, respondents who perceive the proposed extension in green climate change adaptation measures as more credible are willing to pay more for the extension, confirming hypothesis 5. One standard deviation change in the latent credibility corresponds to a change from the mean willingness to pay in the No Script sample by 1.07 euros for extensive green roofs to 10.35 euros for green areas. The interaction of the latent credibility with the status quo is statistically significantly negative. Respondents who view the extension of the attributes as more credible by one standard deviation in the latent variable are willing to pay 6.24 euros more for one of the two combinations of green climate change adaptation measures than for the status quo, holding all attributes constant.

The interactions of the latent consequentiality with all non-cost attributes are also statistically significantly positive, confirming hypothesis 4. That is, respondents who perceive the survey as more consequential are willing to pay more for the extensions of the green climate change adaptation measures. The size of the consequentiality interactions is similar to the credibility interactions. One standard deviation change in the latent consequentiality corresponds to a change from the mean willingness to pay in the No Script sample by 0.85 euros for extensive green roofs to 10.05 euros for street trees. The interaction of the latent consequentiality with the status quo is statistically significantly negative. Respondents who perceive the survey as more consequential by one standard deviation in the latent variable are willing to pay 6.55 euros more for one of the two combinations of green climate change adaptation measures than for the status quo, holding all attributes constant.

6 Discussion and Conclusions

We analyze two pathways which can potentially explain effects of information about the valuation scenario on stated preferences: information scripts could influence perceptions of the survey consequentiality and the scenario credibility, which in turn may matter for stated preferences. While stated preference studies on information effects consider a wide range of types of information, our empirical research focuses on information about the policy context of the valuation scenario. We use a split sample approach in a discrete choice experiment survey concerning preferences for green climate change adaptation measures in the German city of Bremen. All respondents are provided with necessary information for understanding the valuation scenario, half of the respondents see additional information about the climate adaptation strategy that the measures are part of. We investigate whether differences in stated preferences between the two samples can be explained by the information-induced shifts in consequentiality and credibility perceptions, using a mixed logit model and a hybrid choice model.

Our results confirm that the information effects frequently found for various environmental goods (e.g., Bateman and Mawby 2004; Rambonilaza and Brahic 2016; Vanermen et al. 2021) also emerge in the context of urban green and climate change adaptation. Facing the additional information about the climate adaptation context of the valuation scenario increases willingness to pay for the urban green measures. To investigate whether shifts in the perceptions of consequentiality and credibility may explain this effect, we analyze two parts of the pathways separately. First, our data suggests that the information script strengthens the credibility perceptions, while its effect on consequentiality perceptions is small and not statistically significant. Second, we observe that stronger perceptions of both consequentiality and credibility correspond to larger willingness to pay. These results indicate that a shift in credibility might explain part of the information effect: the information script strengthens perceptions of the credibility of the policy change, which in turn is positively related to willingness to pay for the proposed change. On the other hand, because the effect of the information script on perceptions of the survey consequentiality is small and insignificant, this pathway appears unlikely to explain a meaningful portion of the information effect in our case study. Our results suggest that both perception pathways cannot be responsible for a major part of the information effect: in the hybrid model that controls for the two perception pathways with latent variables, the willingness to pay for all attributes is still significantly larger in the sample seeing the additional information than in the sample that does not.

There are some methodological limitations of our research. The hybrid choice approach prevents measurement errors in modelling responses to the questions on credibility and consequentiality perceptions leading to endogeneity, but it does not protect from other causes of endogeneity. First, the elicited preferences might influence the self-reported perceptions, not the other way around, as we elicit the perceptions after preference elicitation. Recent research has shown that the position of the perception elicitation has an effect on both perceptions and stated preferences (Lloyd-Smith et al. 2019; Zawojska et al. 2019a, b), which indicates potential endogeneity between the two. Second, there might be an omitted variable that influences both self-reported perceptions and stated preferences (Chorus and Kroesen 2014). In these two cases, the direction of the causality between the unobservable perceptions and stated preferences is not clear which may lead to endogeneity issues (Mariel and Meyerhoff 2016; Walker and Ben-Akiva 2002). Given these concerns, some caution is needed towards the results about the credibility perception pathway. However, potential endogeneity does not affect our findings that the information script strengthens credibility perceptions and affects stated preferences, because they are based on the randomized exogenous treatment.

Our study uses a new approach for analyzing how information about the valuation scenario affects stated preferences by applying the hybrid mixed logit framework. This gives rise to several potential future directions of research. First, various other potential pathways might explain the effect of an information script on stated preferences. We only investigate the pathways via consequentiality and credibility perceptions. Further research could examine other pathways such as knowledge on the good to be valued, attitudes towards the good or policy scenario, perceived demand from the experimenter, cognitive availability and trust in institutions.

Second, previous studies by Hoevenagel and van der Linden (1993), Ajzen et al. (1996); Hasselström and Håkansson (2014), Czajkowski et al. (2016) and Yang and Hobbs (2020) show that effects of the information about the valuation scenario on stated preferences can depend on the style, length and valence of the information script. We designed a brief information script that could plausibly affect stated preferences as well as perceptions of the survey consequentiality and scenario credibility. However, designing a script that has an even larger effect on stated preferences could improve the precision of decomposing its effect into several pathways. Information about the policy context, as in our case study, might also lead to a different effect than information about other aspects of the valuation scenario, such as information about the current state or the content, extent, expected outcomes or mechanism of the policy change. It would be valuable to investigate whether the role of different pathways varies for different types of scripts.

Third, previous research shows that the effects of information scripts can also depend on the type of good to be valued. The effects are larger if the good is of high personal relevance to the respondents (Ajzen et al. 1996), which is likely the case for urban green and adaptation to climate change in the respondents’ city of residence. Further, information effects are larger if the good is unfamiliar or non-use values are important (Bateman and Mawby 2004). Green areas, green roofs and street trees in our study have some non-use value, but may be familiar to most of the respondents. A similar information script might cause a larger information effect in the valuation of a less familiar but still personally relevant good, which again would allow a more precise decomposition of the effect into several pathways.

Our results have practical implications for stated preference surveys. The perceived credibility of the scenario is often overlooked, and there is little advice on how it can be reinforced. Our results show that a brief information script on the policy context of the evaluated policy scenario can strengthen credibility perceptions. In our study, the information script improved perceived credibility more than the more frequently studied consequentiality perceptions. Stated preference practitioners should incorporate effective information scripts because a credible valuation scenario is desirable for valid value estimates (Johnston et al. 2017). In addition, follow-up questions on perceived credibility should be incorporated (such as those described in Sect. 3.3.) to assess whether the information provided is sufficient for valid value estimates. Further, our results indicate that strengthening credibility perceptions with information scripts about the valuation scenario may in turn affect stated preferences. This suggests that when including additional information intended to improve credibility, changes in value estimates are a desirable consequence rather than a sign of bias. However, a significant part of the effect of the script on stated preferences cannot be attributed to the pathways via credibility or consequentiality. This demands pretesting to determine whether the survey information about the valuation scenario affects stated preferences via potentially desirable pathways such as learning or stronger perceptions of credibility and consequentiality, or undesirable pathways such as availability heuristics or an experimenter demand effect.