This study uses the discrete choice experiment (DCE) method to identify individuals' preferences (Hoyos, 2010) and to estimate their willingness to pay (WTP) for different characteristics of a good or service (Hanley et al., 1998). This approach is based on the theory of random utility (McFadden, 1974) and relies on the analysis of choices made between several alternatives defined by combinations of attributes. DCE was chosen for its ability to quantify trade-offs between attributes and to incorporate a payment vehicle enabling direct monetary estimation. Its implementation in digital format also facilitates large-scale dissemination, and enables a large and geographically diverse sample. This method has several advantages: theoretical soundness, applicability to non-market goods and the ability to model preference heterogeneity. However, it has certain limitations, including a potential cognitive burden for respondents, questionable rationality assumptions and sensitivity to formulation or fatigue biases. Despite these constraints, DCE remains a benchmark method for preference analysis and the economic evaluation of goods and services.

The experimental design was generated using Ngene software following a D-efficient design approach. The design efficiency was optimised for a conditional logit model using prior parameter values derived from expected signs and magnitudes of the attributes. In particular, negative priors were specified for the price attribute, while positive priors were assumed for environmental and economic benefits, in line with standard expectations in environmental valuation studies. The final design consisted of 16 choice sets, each including two policy alternatives and a “status quo” option. To reduce the cognitive burden on respondents, the choice sets were divided into two blocks so that each respondent completed eight choice tasks. In the design generation process, an additional imbalance penalty was considered to avoid excessive level repetition across alternatives and to improve the statistical properties of the design. The final design ensured sufficient variation in attribute levels across choice tasks, allowing reliable estimation of the preference parameters (detailed below in Sect. 2.2.4 and 2.2.5 “Cost to households – electricity bill”, illustrated in Table A2). The detailed tasks with attributes and their respective values are presented in Table A3.

Prior to the choice tasks, attributes and their associated levels were introduced sequentially to respondents, with explanations provided for each attribute to ensure comprehension. The status quo option corresponded to a conventional floating offshore wind development without eco-engineering measures. The exact wording presented to respondents for the status quo explanation was (translated from French): “Option C means a conventional floating offshore wind development WITHOUT artificial reefs”. To encourage respondents to read the attribute descriptions, the “next” button appeared after only a short delay on the screen presenting the explanations. Finally, both the overall length of the questionnaire and the number of choice cards presented to each respondent were deliberately limited to reduce respondent fatigue. In addition, attribute descriptions were written in simple and non-technical language so that they could be easily understood by a general audience, thereby reducing the risk of attribute non-attendance.

An online national survey was performed through a market company in April 2024. Five French territories (i.e. departments) were sampled, depending on their proximity to the planned development of the floating offshore wind farm. These departments were the following: Aude, Bouches-du-Rhône, Hérault, Morbihan and Pyrénées-Orientales. Four of these departments border the coast of the French Mediterranean Sea (south of France), and one (Morbihan) is on the western coast of French Brittany and bordered by the Atlantic Ocean. The total number of respondents was 306, and sampling was carried out in such a way as to obtain proportions as representative as possible to the number of inhabitants in each department, with 20 respondents from Aude, 114 from Bouches-du-Rhône, 87 from Morbihan, 54 from Hérault and 31 from Pyrénées-Orientales. The sampling did not rely on formal quotas or post-stratification weights; however, recruitment through the survey company ensured a balanced distribution in terms of age and gender across departments. The resulting sample composition was verified to be consistent with INSEE departmental demographic data, providing reasonable confidence in the representativeness of the respondents.

We drew up the territorial identities of the sample departments (Table 1) in relation to the subject of study, taking into account demographic and blue economic statistics, including tourism (the touristic rate being the number of touristic beds divided by the number of residents in the department), fishing and industry, information on ecology (through protected areas) and fishing (share of maritime employment and tonnes of seafood landed). The percentages are expressed as a function of the department level.

Literature and data obtained for the Aude department generally point to limited maritime employment in this territory, with 3.1 % of employees working in the maritime sector (INSEE, 2017) and 1600 t yr⁻¹ of seafood products landed (FranceAgriMer, 2024) for a population of 376 000 persons (INSEE, 2025a). However, commercial tourism is highly structuring the economy of the department, with a high rate of second homes (25.3 % of departmental homes) and a relatively high tourist function rate of 42.82 % (Agence de Développement Touristique, 2024). Finally, the maritime domain is still little preserved, with around 6 % of its surface area under any protection regime (Les sites Natura 2000 dans l'Aude, 2019).

The Bouches-du-Rhône department is one of the most urbanised and densely populated French coastal areas, with more than 2 million residents (INSEE, 2025b) but with limited tourism according to the touristic rate of 18 % and 4.8 % of secondary homes (Observatoire en ligne Provence Tourisme, 2025). The area is also characterised by a strong maritime presence but is rather focused on logistics and industry than fishing, with around 3833 t yr⁻¹ (Ifremer, 2024b). It hosts the second-most important commercial harbour in France (Marseille-FOS). However, there are some real natural gems, such as the “Côte Bleue” Marine Park and the Calanques National Park, which are considered true marine sanctuaries thanks to zones that are protected and have no human impact (diving and fishing). This brings the surface area of protected marine areas (all statuses combined) to around 45 000 ha (Bottin et al., 2020) and almost 10 000 ha fully preserved from anthropic activities.

Hérault is characterised by a dense population of more than 1.2 million residents (INSEE, 2025c) and a single harbour structuring the marine employment that is located at Sète, where the fishing landings are concentrated, accumulating around 7146 t yr⁻¹ (Ifremer, 2024c). The proportion of maritime employment in the departmental activity is around 4.4 %. This department is highly attractive for tourism, especially the seaside tourism, and this activity represents a great part of the local economy, highlighted by the tourist function rate of 83 % and the proportion of secondary residences of 17.8 % (INSEE, 2022; Chiffres clés Tourisme et Loisirs Hérault édition 2024, 2025). Also, with 8500 ha of marine protected areas, this department is in the process of reconciling tourism with the protection of its natural heritage (Bottin et al., 2020).

The Atlantic Ocean borders the Morbihan department, giving it a history that is distinct from the Mediterranean departments. This Breton department has one of the highest maritime employment rates in France, with more than 7 % (Février and Le Guen, 2018) for a population above 760 000 residents (INSEE, 2025d). At the same time, the fishing industry in Morbihan is one of the main sectors in the local economy, with almost 22 000 t yr⁻¹ of seafood products landed each year (Ifremer, 2024d). On top of this, the area is a major drawcard for tourists, thanks to its culture and landscapes, with a high tourist function rate of 85 % and 17.8 % of secondary residences. Another attraction for tourism is the balance between maritime exploitation and preservation in the Gulf of Morbihan, with some 70 000 ha of protected marine areas (DREAL Bretagne, 2023).

Last but not least, the Pyrénées-Orientales department lies midway between mountain ranges and coastlines, making it an attractive location for tourism. The population is modest, with almost 490 000 residents. This demographic profile is reflected in the tourism offer, particularly in the tourist function rate of 132 % (Capacité d'accueil Pyrénées Orientales Tourisme, 2025) and a high rate of 27.7 % of second homes (INSEE, 2025e). Tourism is thus the mainstay of the local economy. Meanwhile, maritime activity is more limited, with a low proportion of maritime employment (3.7 %; INSEE, 2017) and more limited landings than other departments (1501 t yr⁻¹; Ifremer, 2024e). The documentation found estimates around 11 000 ha for the surface of marine protected areas of any status (Bottin et al., 2020; De Paoli et al., 2023).

The questionnaire started with socio-demographic questions: place of residence (zip code), education level, current employment status, and income after taxes and per month (France). The choice experiment followed these questions. Before the series of choices, an introduction was included with the following information:

Electricity mix in France and governmental goals,

The status quo scenario chosen reflects France's current trajectory in offshore wind power: rapid intensive development of wind farms, with no particular requirements beyond the regulatory framework imposed. It corresponds to floating wind farm projects that could be described as “classic”, with no specific eco-engineering measures, apart from the environmental monitoring required before and after commissioning and throughout the farm's service life until decommissioning. This scenario serves as a realistic reference point, consistent with national guidelines, and enables the measurement of preferences for alternatives that incorporate greater ecological ambitions.

The attributes were chosen on the basis of a preliminary study in which respondents expressed their fears and priorities with regard to the development of offshore wind power, whether bottom-fixed or floating (Dubois et al., 2025a). Moreover, literature was considered to scale the levels of chosen attributes (Börger et al., 2015; Dalton et al., 2020; Iwata et al., 2023; Kermagoret et al., 2016; Kim et al., 2019; Klain et al., 2020). The definition of levels for each attribute is based on a synthesis of the scientific literature, empirical data from fisheries, energy reports, and adjustments based on the pre-testing of the questionnaire. The aim was to propose realistic, credible and comprehensible levels for respondents as well as ensuring sufficient variability to capture differentiated preferences.

The material used for the structure (recycled or new steel) is a central environmental indicator. With an emission reduction potential of 1.5 t of CO₂ per tonne of steel (World Steel Association, 2024), recycled steel has a 20 %–25 % lower carbon footprint than new steel (Fennell et al., 2022). France already produces around 40 % of its steel from recycled materials (Ministère de la Transition Écologique, 2024), making this attribute credible, measurable and culturally relevant. It also makes it possible to test citizens' sensitivity to aspects of circularity in energy infrastructures.

The biodiversity attribute was defined on the basis of extensive literature on the effects of both offshore wind farms and artificial reefs. Submerged structures (foundations, cables, floats) promote colonisation by fixed species such as mussels, anemones, algae or soft corals (Andersson and Öhman, 2010; Coolen et al., 2018; Degraer et al., 2021; Dubois et al., 2025a), inducing a local increase in biodiversity. Rates of increase in biodiversity ranging from 10 % to 200 % have been reported depending on the context (Brock and Norris, 1989; Fabi and Fiorentini, 1994), although the range generally adopted in previous DCE varies between 10 % and 60 % (e.g. Klain et al., 2020). To remain within a zone of ecological plausibility and to facilitate understanding for respondents, the following four levels were retained: +10 %, +20 %, +30 % and +40 % increase in marine biodiversity on average throughout the service life of the farm. This increase refers to the increase in species richness “S” (Anon, 2009). The experimental design was inspired by previous work carried out on artificial reefs where the addition of hard substrates has demonstrated strong potential for biological colonisation (Koeck et al., 2014; Komyakova et al., 2021). The structures studied were modelled with a volume of around 320 m³ (Glarou et al., 2020), the optimum size suggested in the literature to maximise ecological effects.

The impact of floating wind turbines on fishing activities was assessed by examining changes in the income of local fishers, an indirect measure that was proved pertinent (Bates and Firestone, 2015; Firestone and Kempton, 2007). Based on studies of fishing yields around artificial reefs (CPUE – catch per unit effort), a link was established between an increase in biomass and biodiversity, combined with a potential increase in catches. A literature review (De Backer and Hostens, 2019; Ramos et al., 2006; Reubens et al., 2013) was used to translate CPUE gains into economic impacts. A 60 % catch-to-revenue conversion was adopted on the basis of existing data (Pan, 2021). This rate was then reduced to take into account operational constraints (closed areas, affected ports, etc.). The estimated impact was refined by cross-referencing wind farm development zones with data from the main fishing ports in the French Gulf of Lion (Ifremer, 2024a). To include differentiated but plausible scenarios, and following the pre-test highlighting the absence of an “extreme” case, the levels retained were +1 %, +5 %, +10 % and +15 % increase in fishing income in the zones concerned and on average throughout the service life of the wind farm.

The last attribute was the payment vehicle for the willingness to pay and represents the monthly extra cost on the electricity bill induced by the integration of eco-engineering structures in wind farms. This cost was estimated by modelling the price of steel structures from computer-aided design (320 m³ total volume, 43.5 m³ steel) and its installation offshore. This amount was then integrated into electricity production costs via an economic simulator (Energy101, 2025). Standard parameters were considered for floating wind farms: a capacity of 1050 MW, a capacity factor of 60 %, a lifespan of 20 years and an interest rate of 6 %. Three consumption profiles were simulated (1 person, 2 people and 4 people, respectively, in a studio, a small apartment or a house), with amounts ranging from EUR 0.40 to 7.76 per month depending on the profile. In addition to these estimates, feedback from the pre-test suggested the inclusion of a higher cost level to capture economic trade-offs. Thus, five levels have been retained: EUR 1, 2, 3, 5 and 10 per month, over a 20-year period and for a household. These values were in line with previous research (Kim et al., 2019; Krueger et al., 2011).

The analysis conducted in this study relies on the random utility theory maximisation approach (McFadden, 1974). When a respondent chooses a scenario for a FOWT development, the respondent is supposed to choose the option that maximises the satisfaction that is derived from the attributes and their levels. The utility function is as follows: 1Unj=βXnj+εnj, where Unj represents the utility that individual n derives from alternative j, Xnj is the vector of observed attributes associated with that alternative, β is the vector of preference parameters to be estimated, and εnj is the stochastic error term capturing unobserved influences on choice behaviour.

In the present study, the attribute vector includes the material used for eco-engineering structures (recycled or new steel), the variation in marine biodiversity (%), the change in local fisheries income (%) and the additional monthly electricity bill (EUR per month). Conditional logit models were estimated separately for each department in order to examine potential territorial differences in preferences. All attributes were specified as alternative-varying.

After estimating the conditional logit model, Wald tests were performed to evaluate linear restrictions on the estimated coefficients (Greene, 2019; Woolridge, 2010; Train, 2009). They were applied to assess whether specific parameters were statistically equal across departments. The Wald test is computed from the estimated coefficients and their covariance matrix, and follows an asymptotic chi-square distribution under the null hypothesis that the tested parameters are equal to zero. It allows one to test whether groups of variables contribute significantly to the explanatory power of the model.

The marginal WTP, also called the implicit price, can be estimated for each of the non-cost attributes as follows, as explained by Hanley et al. (1998), where βc is the coefficient of any of the attributes and βy is the coefficient of the cost attribute (which corresponds to the marginal utility of income): 2WTP=-βcβy.

WTP estimates were therefore derived from the conditional logit specification with a linear cost coefficient. Confidence intervals were computed using the delta method based on the estimated variance-covariance matrix of the model parameters (Train, 2009). All models were estimated in R using the “Apollo” choice modelling package (Hess and Palma, 2019).

A zero-inflated negative binomial (ZINB) regression model was used to analyse the determinants of respondents' tendency to choose the status quo option. The dependent variable(“Number of status quo chosen”) represents the number of times each respondent selected the status quo across the eight choice scenarios performed by the respondent. Preliminary inspection of the distribution revealed a large proportion of zeros, indicating that many respondents never chose the status quo option. This overdispersion and excess of zeros (Cameron and Trivedi, 2013) makes traditional ordinary least squares (OLS) regression unsuitable, as it assumes normally distributed residuals and constant variance. Preliminary OLS models confirmed the lack of fit and heteroscedasticity.

The ZINB model decomposes the data-generating process into two parts (Hilbe, 2011, 2014; Yau et al., 2003): (i) a count model, which predicts the number of status quo choices for respondents capable of choosing it, modelled using a negative binomial distribution; and (ii) a zero-inflation model, which predicts the probability that a respondent never selects the status quo, modelled with a logistic regression. The count part included the following covariates: prior attitude towards floating offshore wind power, stated gender, age, level of education, professional status, monthly revenue, prior exposure to offshore wind power projects (having already seen or heard about offshore wind turbines), environmental attitudes (through the NEP mean score), relationship to the ocean (having a relative working with the ocean), fishing activity (having a relative who is a commercial fisher) and, finally, the distance to the coast in kilometres. To ensure the interpretability of the ZINB coefficients and transparency for reproducibility, the variables included in the ZINB model (Table 2) were coded and scaled as in Table 2.

Model selection was informed by comparisons of Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC) across alternative specifications, including Poisson, negative binomial, zero-inflated Poisson (ZIP) and ZINB models. The ZINB model was selected as the most appropriate due to its superior fit (lowest AIC and BIC in Table 3) and ability to accommodate both overdispersion and excess zeros (Greene, 1994; Hall, 2000).

All analyses were conducted in R (version 4.3) using the “pscl” package (Jackman, 2024) for zero-inflated models. Standard errors and statistical significance were derived from the model summary output, and incidence rate ratios (IRRs) were calculated by exponentiating the coefficients from the count model to aid interpretation.

A panel mixed logit model was estimated as a robustness analysis following the conditional logit specification to account for repeated choice observations per respondent and to explore potential unobserved preference heterogeneity (Tables A5 and A6). In contrast to the conditional logit model that assumes homogeneous preferences across individuals, the mixed logit model allows preference parameters to vary randomly across respondents (Hensher et al., 2005; Train, 2009).

Under this specification, the utility that respondent n derives from alternative j in choice situation t can be expressed as 3Unjt=βnXnjt+εnjt, where Xnjt represents the vector of attributes associated with the alternative and βn is a vector of individual-specific preference parameters. The term εnjt represents the unobserved component of utility, and is assumed to be independently and identically distributed according to a type I extreme value (Gumbel) distribution. These parameters are assumed to follow statistical distributions across the population.

In the present study, the coefficients associated with the non-cost attributes (recycled steel, biodiversity increase and local fisheries revenue growth) as well as the alternative-specific constant for the status quo option were specified as normally distributed random parameters. The cost coefficient was specified as lognormally distributed to ensure that the marginal utility of cost remains negative for all individuals.

The resulting utility specification can be written as 4Unjt=βrecycled,nRecyclednjt+βbiodiv,nBiodiversitynjt+βlocal,nLocalnjt+βcost,nCostnjt+ASCsq,n+εnjt. Additional interaction terms were included to explore behavioural and territorial heterogeneity. In particular, respondents' prior attitudes towards offshore wind power interacted with the ASC to capture systematic differences in the propensity to select the status quo alternative (Lancsar and Louviere, 2008; McFadden, 1974). The attitude score (Likert scale from 1 = very positive to 5 = very negative) was mean centred. Interactions between the recycled steel attribute and the department of residence were also included to explore potential territorial differences in material preferences.

Parameters were estimated using simulated maximum likelihood with 2000 draws. The model was estimated using the Apollo package in R (Hess and Palma, 2019), which allows flexible specification of panel mixed logit models.

The mixed logit model was estimated in preference space. Consequently, the coefficients represent marginal utilities rather than direct willingness-to-pay estimates. This specification is therefore used primarily to assess the robustness of the conditional logit results and to analyse the extent of preference heterogeneity across individuals.