Valuing Forest Restoration Through Environmental Attitudes: A Hybrid Choice Modelling Approach

Abstract

Forest ecosystems are increasingly degraded by natural disasters and human activities, intensifying the need for large-scale restoration. Because restoration outcomes are long-term, uncertain, and largely non-market, understanding how environmental attitudes relate to public preferences and willingness to pay (WTP) is important for socially acceptable and financially feasible policy design. Using a discrete choice experiment in Korea, this study applies a hybrid choice framework that incorporates latent attitudinal variables into a mixed logit structure, allowing attitudes to interact with preference heterogeneity across restoration attributes. Results show significant heterogeneity in choices and WTP. The model identifies two segments with distinct trade-off patterns: one is more sensitive to risk and payment burden, while the other places relatively greater value on restoration and access-related improvements. Although attitudinal indicators are statistically relevant, segment differentiation is more strongly associated with risk sensitivity and cost aversion than with attitudes alone. Compared with conventional choice models, the latent-attitude specification improves behavioural interpretability and model fit, and yields policy-relevant WTP estimates. Overall, the findings indicate that attitudinal information is complementary to economic and risk-related factors, supporting more targeted and publicly aligned forest-restoration policies.

1. Introduction

Forest disturbances caused by natural disasters and human activities increasingly threaten the sustainable provision of ecosystem services—such as carbon sequestration, biodiversity conservation, water regulation, and soil protection—imposing substantial ecological and economic costs worldwide. For example, wildfires alone caused approximately USD 50 billion in annual damages and released around 6450 megatons of CO₂ in 2021, while the cumulative costs of invasive species have exceeded USD 1.288 trillion over the past five decades, underscoring the need for substantial public investment and behaviourally informed forest management and restoration policies [1,2,3,4].

Individual preferences for environmental policies and ecosystem services are widely recognised to be heterogeneous, reflecting differences in perceived benefits, costs, and risk assessments. Beyond conventional socioeconomic determinants, psychological factors—such as environmental attitudes, values, and perceptions—play a critical role in shaping environmental decision-making [5,6]. In the context of forest restoration, which delivers delayed and uncertain benefits, these attitudinal factors are expected to be particularly influential in determining public support and willingness to pay (WTP).

Despite this recognition, many valuation studies continue to treat attitudinal factors as exogenous to choice behaviour, limiting their ability to explain observed preference heterogeneity. Recent advances in discrete choice modelling highlight the importance of integrating cognitive and psychological elements directly into utility frameworks [7,8]. Hybrid choice models (HCMs), which explicitly link latent attitudes to observed choices, provide a theoretically consistent approach for capturing such behavioural mechanisms.

This study contributes to the growing literature on environmental valuation by incorporating latent environmental attitudes and behaviours into a policy-relevant discrete choice experiment framework. By integrating psychological dimensions alongside established economic theory [9], the proposed model offers an interpretable structure for explaining WTP for forest restoration attributes. This approach enables a deeper understanding of how public responses to forest disturbances—such as wildfires, invasive species, and forest degradation—are shaped not only by observable characteristics but also by underlying environmental attitudes [10,11,12].

In particular, Hybrid choice models (HCMs), including latent-class HCMs and mixed logit models (MXLs), offer more flexibility in integrating attitudes and latent behaviours into decision-making frameworks, which otherwise assume homogeneous preferences across individuals. While MXLs capture unobserved heterogeneity through random parameters, HCMs complement these models by capturing underlying psychological constructs, allowing for a more comprehensive understanding of choice behaviour [7,8,13,14,15,16]. This study, therefore, applies a latent-class hybrid choice model to better reflect the complexity of environmental decision-making. The attitudinal statements used to construct latent variables were selected based on environmental psychology and pro-environmental behaviour literature, with adaptation to the Korean forest-restoration context. Specifically, the items were designed to capture anthropocentric and pro-environmental orientations that may influence trade-offs between disaster-risk reduction, biodiversity outcomes, and payment burden in stated-choice settings.

This study has three objectives: (i) to estimate public preferences and WTP for key forest-restoration attributes under disaster-related risks [17,18,19]; (ii) to examine whether latent environmental attitudes improve the behavioural explanation of preference heterogeneity [20,21,22,23]; and (iii) to compare attitudinal effects with risk and cost sensitivity in segmenting respondents. Accordingly, we address three research questions: RQ1: Which restoration attributes (fire risk, pest/disease risk, forest access restrictions, biodiversity loss, and cost) significantly shape choice behaviour and WTP? RQ2: To what extent do latent attitudinal indicators contribute to class membership in a hybrid choice framework? RQ3: Are class differences primarily attitudinal, or are they more strongly associated with risk sensitivity and cost aversion? By answering these questions, we provide policy-relevant evidence for behaviourally informed forest-restoration design in Korea. Although this article uses the same survey platform as our earlier publication [24], it addresses distinct research questions and a different inferential objective. Its novelty lies in the behavioural segmentation framework, which evaluates how latent attitudes interact with risk and cost sensitivity in shaping class-specific restoration preferences, thereby providing policy-relevant evidence for behaviourally informed forest-restoration design in Korea.

2. Theoretical Background

Environmental concerns reflect latent attitudes and anticipated behaviours toward environmental protection and ecosystem service management [24,25]. These concerns shape how individuals perceive changes in forest ecosystem service provision—such as carbon sequestration, biodiversity conservation, and regulating services—and how they evaluate related policy interventions. Because such preferences are inherently latent and not directly observable, HCMs provide an appropriate econometric framework for integrating psychometric constructs into economic decision-making models [26]. In discrete choice analysis grounded in random utility maximization (RUM) [27,28], utility itself is unobserved and inferred from observed choices. The integrated choice latent variable framework extends this logic by explicitly incorporating attitudinal constructs into the utility and class allocation processes [29,30,31]. This structure allows behavioural heterogeneity to be interpreted not only through observed socioeconomic characteristics but also through latent environmental value orientations.

The Korean context of threatened forest ecosystems and sustainable ecosystem services provides an appropriate empirical setting for applying a discrete choice experiment (DCE) combined with HCM [29]. Forest restoration policies involve complex ecological and socioeconomic trade-offs, and public preferences are often heterogeneous and psychologically mediated. Incorporating latent environmental attitudes enables a more behaviourally grounded segmentation of preferences than conventional models relying solely on observable covariates.

2.1. Latent Class Hybrid Choice Modelling

Building on the theoretical foundations outlined above, this subsection presents the latent-class hybrid choice model based on the integrated choice latent variable framework [30,31,32,33]. This approach embeds unobserved psychometric attitude variables within the environmental choice model, allowing preference heterogeneity to be behaviourally interpreted. Latent attitudes reflect the relative importance individuals assign to alternative attributes and are shaped by personal values and socioeconomic characteristics [30]. The theoretical foundations and extensions of this framework are detailed in Ben-Akiva et al. (2002) [13] and Vij and Walker (2016) [32].

The methodology for our HCM application is grounded in choice modelling [34,35] and based on the random utility model. Our approach aims to provide more realistic and comprehensive explanations and outcomes by incorporating heterogeneous environmental attitudes and behaviours—factors often overlooked in previous economic decision-making models. The first theoretical component of our study is DCM, a highly flexible statistical methodology with diverse applications across various fields, including health [19], transportation mode selection [26,32,36], environmental economics [14,17,20,23,30,37], and agricultural economics [38,39]. Specifically, there is a lack of research applying latent-class HCM in environmental economics [5,6,15,32,40], making this approach rare in our field. The second theoretical component is the economic theory of random utility maximization (RUM) proposed by McFadden (1974) [27] and Luce (1958) [28].

Given this gap in the literature, we apply statistical methodologies, latent-class choice modelling, and PCA. This section focuses on the theoretical background of HCM. To better understand the latent and complex attributes behind the choices in our DCE data, we use HCM, a type of latent-class model that employs a probabilistic class allocation model. HCM consists of two sub-models: the first part is the class-specific choice model, which assigns the probability of choosing each alternative, and the second part is the class membership model, which estimates the likelihood of an individual belonging to a particular class [41].

In a latent-class model, the class-specific choice model specified with the multinomial logit model returns the probability that individual $n$ chooses a specific alternative $i$ on choice task $t$, given that respondents belong to a certain class $q$. In our empirical analysis, we adopt a linear-in-parameters specification, and the utility of individual $n$ is derived from choosing an alternative $i$ in choosing task $t$, given that respondents belong to the class $q$. This is formulated in Equation (1):

$$U_{nit|q}= { V}_{nit|q}+{\epsilon }_{nit|q}= {ASC}_i+{{\beta }^{\prime }}_q{x}_{nit}+ {\epsilon }_{nit|q},$$

(1)

where $U_{int}$ represents the overall utility, ${ V}_{nit|q}$ is the observed part of the conditional indirect utility for individual $n$ choosing alternative $i$ in the choice situation $t$ (alt1, alt2, alt3) conditioned on class $q$, and ${\epsilon }_{nit|q}$ refers to an independently and identically distributed extreme type-I error term over individuals, alternatives, and classes. $x_{nit}$ is the vector of the observed attributes of alternative $i$. It normally includes an alternative specific constant ($ASC$) which captures the systematic utility of omitted variables [40]. ${{\beta }^{\prime }}_q$ is the vector of unknown parameters that need to be estimated, and $x_{nit}$ denotes the explanatory variables, consisting of key attributes and levels related to alternative $i$. For the theoretical background on choice probability [27], see the work of Train (2009) [42] and Mariel et al. (2021) [34]. To estimate the marginal WTP, we usually divide the coefficient of the nonmonetary attributes by the negative of the coefficient of the cost attributes, calculated separately for each latent class [42,43,44]. This implicit price is useful for demonstrating the trade-offs between attributes. Comparing the implicit prices reflected by respondents’ environmental preferences for attribute and level combination choices helps us to understand the relative importance and the relative economic values respondents place on each attribute [45].

The probability associated with individual $n$’s chosen alternative $i$, conditioned on class $q$, is formulated as follows:

$$P(y_{nit}|x_{nit}, \beta ) = {\displaystyle \frac{e^{V_{nit|q}}}{{\sum }_{j=1}^J{e}^{V_{njt}|q}}}.$$

(2)

The probability of the sequence of $T_n$ choices made by individual $n$—denoted by the $y_n$ conditional—being in class $q$ is the product of the probability associated with each chosen alternative in their sequence:

$$P(y_n|x_n, {\beta }_q) = {\prod }_{t=1}^{T}P\left(y_{nit}\right|x_{nit}, \beta ).$$

(3)

The probability that individual $n$ belongs to class $q$, given their characteristics, is represented by $P\left(q_n\right)$. Therefore, the overall choice probability for respondent n, which is weighted across all classes by the class membership probability $P(q_n$) of observing $y_n$, is formulated as Equation (4):

$$P(y_n|x_n, \beta , Q)= {\sum }_{q=1}^{Q}P\left(q_n\right)P\left(y_n\right|x_n, {\beta }_q).$$

(4)

We employ this LCM formulation throughout our study. Our results can provide a better model fit, showing better explanatory power. The complexity of our model, considering environmental attitudes and behaviours through a variety of indicators, enables it to explain more reasonable environmental choice behaviour.

To account for the effect of latent variables such as environmental attitudes and behaviours, using latent HCM, the model incorporates structural equations for the utility and latent (indicators) variables and a measurement equation (choice model). In this model, it can be assumed that the responses to the attitudinal questions are mapped to a latent variable included directly in the class membership expression. Latent constructs are based on attitudinal (environmental) indicators, often employing Likert scales that can be modelled through measurement equations [46]. In our case, we consider two latent variables: one representing anthropocentric views and another representing pro-environmental views. While the latter is typically interpreted as valuing nature for its intrinsic worth, it is important to acknowledge that pro-environmental attitudes may also be motivated by perceived benefits to humans, such as ecosystem services. In both cases, the latent variable, ${LV}_{n_m}$, is given by the following structural equation:

$${LV}_{n_m}= {\eta }_n$$

(5)

where $n_{n_m}$ is a deviate from an independent standard normal distribution, representing the value of the mth latent variable for respondent n. Note that a measurement equation can be described by more than one latent variable. In this work, we assume that each latent variable is described by a single latent variable and that each latent variable explains only a single indicator (in addition to entering the choice probabilities; hence, the m indexing, in our case, covers both the latent variable and the indicator). Responses to our behaviour questions are given on a 5-point Likert scale. Since the response is given on an ordered scale, we need to use an ordered model for the measurement equation [47,48], as explained below. Let $I_{n_m}$ be the mth measurement variable for respondent n, which is assumed as a function of the latent variable ${LV}_{n_m}$ and a random component ${\nu }_n$, as in the following:

$$I_{n_m} = {{\xi }_m}^{\prime }{LV}_{n_m}+{\nu }_{n_m},$$

(6)

where ${\xi }_m$ is a vector of parameters measuring the association between respondent n’s response to the mth indicator and latent variable ${LV}_{n_m}$, and ${\nu }_{n_m}$ is a random disturbance term. Given the presence of the error term, $I_{n_m}$ is unobserved. However, what is observed are responses $i_1$, $i_2$, …, $i_5$, indicated by individuals on a 5-point Likert scale. It is assumed that $i_1$ < $i_2$ < … < $i_5$, and

$$=\left\{\begin{array}{c}{i}_1 if-\infty <I_{n_m} \le {\tau }_{m_1}\\ i_2 if {\tau }_{m_1}<I_{n_m}\le {\tau }_{m_2}\\ i_3 if {\tau }_{m_2}<I_{n_m}<{\tau }_{m_3}\\ i_4 if {\tau }_{m_3}<I_{n_m}<{\tau }_{m_4}\\ i_5 if {\tau }_{m_4}<I_{n_m}<+\infty \end{array}\right.$$

(7)

where ${\tau }_{m_1}$, ${\tau }_{m_2}$, ${\tau }_{m_3}$, and ${\tau }_{m_4}$ are threshold parameters to be estimated through an ordered probability model [48].

The probability of observing the specific answer, $i$, in the mth Likert scale is expressed as follows:

$$P\left(z_{n_m}=i|{\xi }_m, {LV}_{n_m}, \tau \right)=F\left(I_{n_m}\le {\tau }_{m_i} \right)-F\left(I_{n_m}\le {\tau }_{m_{i-1}} \right),$$

(8)

where $F$ is the distribution function (logistic). Then, the probability of observing the respondent n’s response to the $m$th attitudinal item is expressed as follows:

$$\left(z_{n_m}|{\xi }_m, {LV}_{n_m}, {\tau }_m\right)={\prod }_{i=1}^5{P\left(z_{n_m}=i|{\xi }_m, {LV}_{n_m}, \tau \right)}^{z_{n_m^i}}$$

(9)

where $z_{n_m^i}$ is an indicator equal to one if respondent n selected the ith item on the mth attitudinal question. The probability associated with the sequence of the M attitudinal questions can thus be represented by Equation (10):

$$P\left(z_n|\xi , {LV}_n, \tau \right)={\prod }_{m=1}^{M}P\left(z_{n_m}|{\xi }_m, {LV}_n, {\tau }_m\right).$$

(10)

In addition to influencing responses to the indicators, the latent variables also appear in the choice probabilities. In this case, it is assumed that they affect the class membership; therefore, they are introduced in the class membership equation, $P\left(q_n\right)$. Specifically, $P\left(q_n\right)$ is expressed as follows:

$$P\left(q_n\right)={\displaystyle \frac{e(c_q+\sum _{m=1}^M{\delta }_{{q_m}^{\prime }}{LV}_{n_m})}{\sum _{q=1}^{Q}e(c_q+{\sum }_{m=1}^M{\delta }_{{q_m}^{\prime }}{LV}_{n_m})}}.$$

(11)

Therefore, an HCM consists of three model components: the structural equation model, the measurement equations, and the choice model. The joint log-likelihood $\left(LL\right)$ of such a model is expressed as follows:

$$LL={\sum }_{n=1}^{N}ln{\iint }_{-\infty }^{\infty }P\left(y_n\right|x_n, \beta , Q,c, \delta )P\left(z_n|\xi , {LV}_{n_m}, \tau \right)d{LV}_{1}d{LV}_2,$$

(12)

where the choice probability now includes $c$ and $\delta$. Since the expression in Equation (12) does not have a closed form, it is necessary to approximate the log-likelihood (LL) via simulation. The empirical application of the latent HCM and PCA in our case study is detailed in the subsequent sections.

2.2. Selection of Attitudinal Indicators Through PCA

To identify latent environmental attitudes and behaviours among Korean citizens for application in the HCM (Section 2.1), we analysed 20 items adapted from Milfont et al. (2010) [49], rated on a 5-point Likert scale (1 = strongly disagree; 5 = strongly agree), to determine the number of distinct attitudinal dimensions emerging from the data [50]. Indicators were grouped into attitude and behaviour factors, and selected using PCA based on their representativeness (

Table A2. Respondents’ results for environmental attitude and environmental behaviour (5-point Likert scale: 1. strongly disagree; 2. disagree; 3. neither agree nor disagree; 4. agree; 5. strongly agree).

(A1) To what extent do you personally agree with the following statements?
Items	Mean	Mode	Standard Deviation
1. I am the kind of person who loves spending time in wild, untamed wilderness areas.	3.55	4	0.96
2. Being out in nature is a great stress reducer for me.	3.95	4	0.79
3. I have a sense of well-being in the silence of nature.	3.99	4	0.78
4. I find it more interesting to be in the forest looking at trees and birds than in a shopping mall.	3.64	4	0.96
5. Nature is important because of what it can contribute to the pleasure and welfare of humans.	4.15	4	0.81
6. What concerns me most about deforestation is that there will not be enough lumber for future generations.	3.72	4	1.00
7. We should protect the environment for the well-being of plants and animals rather than for the welfare of humans.	3.56	4	1.01
8. Conservation is important even if it lowers peoples’ standard of living.	3.71	4	0.92
(A2) To what extent do you personally agree with the following statements?
Items	Mean	Mode	Standard Deviation
1. Most environmental problems can be solved by applying more and better technology.	3.53	4	0.91
2. Science and technology will eventually solve our problems with pollution, overpopulation, and diminishing resources.	3.54	4	0.95
3. If things continue their present course, we will soon experience a major ecological catastrophe.	4.08	4	0.82
4. The Earth is like a spaceship with very limited room and resources.	4.03	4	0.82
5. The balance of nature is very delicate and easily upset.	3.84	4	0.89
(A3) To what extent do you personally agree with the following statements?
Items	Mean	Mode	Standard Deviation
1. I make sure that during the winter the heating system in my room is not switched on too high.	3.98	4	0.77
2. In my daily life, I try to find ways to conserve water or power.	3.95	4	0.81
3. Even if public transportation was more efficient than it is, I would prefer to drive my car.	2.94	2	1.24
4. It is alright for humans to use nature as a resource for economic purposes.	3.08	3	1.03
5. The question of the environment is secondary to economic growth.	3.57	4	1.07
6. The benefits of modern consumer products are more important than the pollution that results from their production and use.	2.97	2	1.13
7. I would like to join and actively participate in an environmentalist group.	3.20	3	1.02

Note: A1_1, A2_3, and related codes refer to survey items listed in the questionnaire; the full wording of each item is reported in Table A2. Adapted from Jeon and Campbell (2025) [25].

), before being incorporated into the HCM [14,20,36,50]. PCA was conducted using IBM SPSS Statistics version 28 and R (R-4.5.1). Model specification followed psychological evidence distinguishing attitudes and behaviours as separate constructs [51,52,53].

To preserve identification stability and numerical tractability in the HCM, we retained one representative indicator per latent construct (A3_6 for anthropocentric orientation and A1_2 for pro-environmental orientation). This parsimonious specification reduces collinearity among highly correlated Likert items and facilitates convergence in simulated maximum-likelihood estimation. Although single-indicator representation simplifies the multidimensional structure of environmental attitudes, the estimated attitudinal effects should therefore be interpreted as conservative proxies rather than comprehensive psychological constructs.

3. Case Study

Korea provides an appropriate empirical setting for examining public support for forest management policies in the context of increasing environmental risks. National afforestation programs implemented since the 1970s have substantially increased forest cover and timber stocks [54], leading to greater public use of forest-based recreational resources. However, recent changes in climatic conditions and the spread of forest pests have introduced new management challenges that may undermine the long-term provision of ecosystem services. In particular, large-scale wildfire incidents and the expansion of pine wilt disease (PWD) have heightened concerns regarding forest resilience and biodiversity conservation, which have been increasingly addressed in recent forest management policy discussions [55].

These developments have generated policy interest in strengthening restoration and preventive management measures. In this context, evaluating public preferences for alternative conservation strategies requires an analytical framework capable of incorporating behavioural and attitudinal factors. The empirical design of the present study therefore focuses on institutional policy scenarios that reflect potential interventions aimed at mitigating ecological risks while maintaining recreational benefits derived from forest ecosystems. Accordingly, the policy scenarios considered in this study are incorporated into a choice experiment framework and analysed using a hybrid choice modelling approach.

Attributes, Levels, and Scenarios

The choice experiment survey (ethical approval procedures were followed, and per-mission was obtained from the University of Stirling for this research (15 March 2023) aims to estimate diverse WTP and measure different preference for forest management policies in Korea. Considering the situation, we selected five major attributes based on a relevant literature review [56] and the advice of experts from the National In-stitute of Forest Science in Korea. The attributes and levels selected for our survey and for-est restoration scenarios are realistic scenarios based on the currently enacted Korean for-est fire and forest pest management control system and the ongoing forest restoration pol-icy of the Korean Forest Service to fully reflect the realistic forest situation.

According to

Table 1. Attributes and levels of sustainable forest management scenarios.

	Levels	Level 1	Level 2	Level 3	Level 4
Attributes
Forest fire risk		(Red, 85 ↑) *	(Orange, 66~85)	(Yellow, 51~65)	(Blue, 51 ↓)
Forest pests and disease risk		(Red) *	(Orange)	(Yellow)	(Blue)
Restriction on forest-related outdoor activities		(Zero) *	(0~−1 million)	(−1~−2 million)	(−2~−5 million)
Biodiversity loss (number of pine trees damaged)		(−300~−400 k trees) *	(−200~−300 k trees)	(−100~−200 k trees)	(0~−100 k trees)
Forest restoration fund		KRW 0 *	KRW 1~KRW 25,000 (USD 0~USD 19.14)	KRW 25,100~KRW 35,000 (USD 19.22~USD 26.80)	KRW 35,100~KRW 50,000 (USD 26.88~USD 38.29)

Note: * It represents the status quo (SQ) reference; adapted from Jeon and Campbell (2025) [25]. ↑: More than 85, ↓: less than 51. KRW denotes Korean won. The sign convention for the “number of damaged trees” attribute is directional and reflects ecological damage intensity; higher absolute levels indicate greater degradation pressure.

, the first attribute reflects variations in wildfire risk under alternative forest management strategies. In the Korean context, fire incidents are largely associated with human activity and accumulated forest fuel loads in densely stocked stands. As timber volumes have increased over time, the potential severity of wildfire impacts on ecosystem services—particularly recreational and regulating functions—has also intensified. Preventive policy interventions, including access management and early response systems, have therefore been implemented to mitigate fire-related risks. The wildfire risk levels presented in the choice tasks were defined based on the national forest fire alert system reported by the Korea Forest Service [54]. A more detailed description of wildfire risk classification and its ecological implications can be found in previous studies [24,25].

The second attribute captures the potential spread of forest pests and diseases, with particular emphasis on pine wilt disease, which has been identified as a major threat to forest health and ecosystem service provision in Korea [56,57,58]. Monitoring and control measures implemented by forest authorities typically rely on an outbreak alert system that classifies risk levels based on infection rates and spatial diffusion patterns. These management interventions are intended to limit ecological degradation while maintaining forest productivity and recreational value. The pest risk levels presented in the choice tasks were defined in accordance with the national forest disease monitoring framework reported by the Korea Forest Service [54]. Further ecological and policy-related details regarding pest outbreaks and management strategies are documented in previous studies [24,25].

The third attribute represents restrictions on forest-related outdoor activities as a preventive measure to mitigate wildfire and pest-related risks. Since a substantial proportion of forest disturbances in Korea are linked to human activity, temporary limitations on visitor access may contribute to reducing environmental pressures associated with fire outbreaks and pest diffusion. However, such restrictions may also decrease the availability of recreational ecosystem services, thereby creating trade-offs between ecological protection and public use. The restriction levels presented in the choice tasks reflect alternative policy scenarios aimed at managing visitor pressure in forest recreational areas. Additional information regarding visitor trends and access management policies can be found in previous studies [24,25].

The fourth attribute captures biodiversity-related degradation risks using the number of trees affected by pest infestations as a proxy indicator. In the context of Korean forest ecosystems, pest outbreaks may result in significant ecological damage that threatens long-term forest resilience and ecosystem service provision. The levels of tree damage presented in the choice tasks were derived from observed ranges of pest-related forest degradation reported in national monitoring data (accessed on 15 August 2023). While this proxy does not fully represent the multidimensional nature of biodiversity loss, it provides a practical and comprehensible measure for respondents when evaluating policy alternatives. Further ecological and management-related details are provided in previous studies [24,25].

The final attribute corresponds to the financial contribution required to support forest restoration initiatives. Given that forest management and conservation programs in Korea are primarily financed through public expenditure, the payment vehicle is framed as an annual household contribution intended to fund preventive and rehabilitative interventions. The contribution levels presented in the choice tasks range from KRW 0 to KRW 50,000, reflecting realistic policy implementation scenarios. For reference, the average per capita tax contribution allocated to the forest sector was approximately KRW 6700 in 2020. Currency conversion was based on the exchange rate reported by the Bank of Korea (USD 1 = 1305.8 KRW; accessed on 28 August 2023).

The experimental design of the choice sets was developed to ensure statistical efficiency in estimating willingness to pay for alternative forest management policies. Attribute levels were combined using an efficient design procedure to minimize parameter variance in the discrete choice model, following established guidelines for stated preference experimental design [35,41]. Based on prior parameter estimates obtained from a pilot survey of 67 respondents (670 choice observations) [34], a total of 200 choice sets were generated. The resulting design achieved a D-error of 0.101, an A-error of 0.192, and an S-error of 9.67. Each respondent was presented with ten randomly assigned choice tasks comprising two policy alternatives and a status quo option associated with zero additional financial contribution. This design structure was adopted to balance the informational requirements of the model with potential respondent fatigue and learning effects during the survey process.

For the design of the experimental choice set for our research, we iterated the set through the design process to ensure that each attribute is as evenly distributed as possible. In the DCE survey, we asked respondents to choose between two improved alternatives (alternatives A and B) and the current situation (SQ), which promises zero additional cost for implementing no improvements to forest restoration. Finally, 10 choice sets are randomly assigned to each respondent (see

Table A1. An example of a choice set for forest management policy.

	Alternatives	Improved Situation		Current Situation
Attributes and Levels		Alternative 1	Alternative 2	Alternative 3 (Status Quo)
Forest fire risk		(Blue)	(Orange)	(Red)
Forest pests and disease risk		(Orange)	(Blue)	(Red)
Restriction on forest-related outdoor activities		(3~5 million restriction)	(Under 1 million restriction)	(No restriction)
Biodiversity loss (no. of pine trees damaged)		(−100 k~−200 k pine trees loss)	(−200 k~−300 k pine trees loss)	(−300 k~−400 k pine trees loss)
Forest restoration costs		(KRW 10,000)	(KRW 25,000)	KRW Zero

Note: Adapted from Jeon and Campbell (2025) [25].

for an example of a choice set), which balances the concerns of learning and respondent fatigue [45]. The experimental choice set and scenario applied in this study were designed to be as easy as possible for respondents to understand. The status quo option (alt3) referred to the current situation, and the other two choice set options (alt1 and alt2) referred to forest management policies aimed at improving the situation. The attributes and levels within each choice set were also designed to be easy to understand, incorporating visual aids and easy-to-determine trade-offs between attributes and levels. Moreover, it is made clear that the SQ choice set is a zero-cost option. Table A1 in Appendix A is an example of a choice set for experimental scenarios in which three options are presented to each respondent: two choice sets and a status quo option.

4. Results

4.1. Socioeconomic Background

The sampling design employed a stratified random sampling approach to ensure demographic representation of the Korean adult population aged 20 and above. (A pilot study was conducted from 2 to 16 October 2023, using face-to-face surveys with 67 participants (670 choice observations) to test the questionnaire. The survey, administered by a professional firm using tablet devices, included sections on environmental attitudes (based on Milfont and Duckitt [49]), experimental choice sets, and respondent demographics. The pilot results showed that while some attribute coefficients were statistically significant and aligned with expectations, others were not. Feedback indicated that the questionnaire was clear and visually appealing but slightly lengthy. Based on this, redundant questions were removed, and the scenario section was streamlined, allowing the final survey to be completed within 15–20 min.) Respondents were recruited from all major administrative regions in Korea, encompassing both metropolitan and rural areas, with quotas applied to achieve balanced distributions across gender, age, and income levels. Although metropolitan areas (e.g., Seoul and Gyeonggi) were slightly over-represented, the sample included a diverse range of socioeconomic groups to reflect national variation. Data were collected in October 2023 through in-person surveys conducted by a professional survey firm (ST Innovation, Seoul, Republic of Korea), using a tablet-assisted questionnaire to enable real-time data entry and quality control. The dataset is derived from a choice experiment in which 1021 individuals (yielding 10,210 discrete choice responses) selected between two policy alternatives: improved forest management options and a status quo scenario representing deforested and degraded forests affected by environmental threats such as forest fires and pest outbreaks.

Table 2. Summary of statistics.

Variables	Description	Mean	Standard Deviation	Minimum	Maximum
AGE	Age	40.47	11.64	20	81
GEN	Gender: Female (1); Male (0)	0.56	0.49	0	1
EDU	Level of Education	3.96 (University level)	0.57	1	5
INC	Monthly Income	KRW 3,000,000~KRW 3,900,000 (USD 2308~USD 3000)	1.34	1	5

Note: Monthly income was coded on a five-point scale: 1 = [KRW 833,333~ KRW 2,490,000], 2 = [KRW 2,500,000~ KRW 4,160,000], 3 = [KRW 4,170,000~ KRW 5,830,000], 4 = [KRW 5,840,000~ KRW 8,330,000], 5 = [KRW 100,000,000].

shows the descriptive statistics for the respondents (in the questionnaire used in this study, the items were organized as follows: Age data were collected through an open-ended question; gender was categorized as male or female; and education level was classified into five groups—primary school (1), middle school (2), high school (3), university (4), and graduate school (5). Income was measured on a five-point scale. Socio-economic reference data from Korea (KOSIS; kosis.kr) indicate the following: gender ratio—male 49.4%, female 50.6%; average age—44.4 years (20s: 11.4%, 30s: 13.1%, 40s: 15.7%, 50s: 17.1%, 60s and above: 27.3%); education level—53% have attained education beyond high school; and average monthly income—4,390,000 KRW (KOSIS National Statistics Portal). The sample size was sufficient to ensure the statistical value of the main parameters and was larger than most discrete choice experiments in this area. Of the 1021 participants, 569 (55.7%) were female, and 621 (60.8%) were in their 40s. Over 70% of the respondents had a bachelor’s degree, and around 40% of our respondents had a monthly income level of over USD 3001. A survey was collected alongside the choice experiment. In this survey, sociodemographic information on respondents was collected. We also collected the preference data on respondents’ environmental behaviours. These data include environmental attitudes and behaviours and forest management attributes such as forest fires, forest pests and diseases, restriction on forest-related outdoor activities, forest biodiversity loss, and forest restoration costs.

For this debriefing and follow-up questions, the main items in the second half of the questionnaire included questions referring to the rank of attribute prioritization, policy and WTP payment consequentiality [59,60], and the literacy difficulty of the questionnaire. To summarize respondents’ responses to the key questions, a significant proportion of respondents said that “they could not afford it,” were at least somewhat aware of the threats to forest ecosystem services, and were somewhat interested in forests. Most important of all, for the main items related to government and forest policy, respondents emphasized the importance of the government’s role and responsibility in countering environmental problems and showed a moderately favourable response to the tax instrument. Lastly, respondents tended to be positive about the monetary value of forest protection and the difficulty faced in preventing ecosystem service disasters (

Table 3. Debriefing and follow-up questions.

	Mean	Standard Deviation
1. I cannot afford to pay.	3.31	0.99
2. I have a better option nearby where I would go to.	3.51	0.80
3. It is not feasible to stop the spread of forest pests and forest fires.	2.76	1.02
4. I do not want to put a monetary value on protecting forests.	2.67	1.07
5. I do not recognize the risks and threats to forest ecosystem services.	2.94	1.03
6. The payment method is inappropriate.	3.01	1.05
7. I pay enough tax already. It is the government’s responsibility.	3.43	0.98
8. The benefits I receive are not worth my rate increases.	3.28	0.97
9. I am not interested in forests.	2.45	1.05

Notes: The default scale is 1–5.

).

4.2. Result of Attitudinal Indicator Selection by Latent Group

To investigate how environmental attitudes and behaviours related to ecosystem services and forest management policy influenced the WTP for forest restoration, this section describes the tests for the reliability, validity, and fit of the model. It also presents the analysis of how the latent variables and classes derived from the PCA and choice modelling affected the actual WTP behaviour. As a first step, we conducted a PCA with IBM SPSS Statistics version 28 and R (R-4.5.1) on the responses to the environmental attitude and behaviour questions. However, before performing the PCA with principal axis factor analysis, we first determined the internal consistency and reliability in the data. The Cronbach’s alpha coefficient for the 20 indicators shows a good fit of 0.83, exceeding the general criterion (0.7) for reliability. To test sampling adequacy, the Kaiser–Meyer–Olkin (KMO) measure shows a result of 0.86, indicating very appropriate data for PCA. The significance of Bartlett’s test of sphericity is 0.00, with 190 degrees of freedom (chi-square = 6251.3), confirming the suitability of the data for PCA. In our data and model, all test statistics exceeded the critical values, indicating that the structural model of environmental attitudes and behaviours is appropriate for explaining the impact of sustainable forest management and actual WTP for forest ecosystem services.

We conducted a principal component analysis (PCA) on 20 attitudinal items related to environmental values and behaviours. This resulted in two components representing distinct environmental orientations: an anthropocentric attitude (Component 1) and an ecocentric attitude (Component 2). The respondents were classified into two latent attitudinal groups based on their scores on these components (see

Table A3. Total variance attributed to environmental attitudes and behaviours to extract indicators.

Component	Initial Eigenvalues *			Rotation Sums of Squared Loadings
	Total	Percent of Variance	Cumulative Percent	Total	Percent of Variance	Cumulative Percent
1	5.13	25.66	25.66	2.75	13.75	13.75
2	2.83	14.17	39.84	2.69	13.47	27.23
3	1.40	7.03	46.87	2.32	11.61	38.84
4	1.24	6.21	53.09	2.20	11.01	49.86
5	1.03	5.16	58.26	1.67	8.39	58.25

* Note: Extraction sums of squared loadings are the same as the initial eigenvalues.

and

Table A4. Results of component estimation with PCA using IBM SPSS 24 (28.0.1.1).

Environmental Attitude Items	Component
	1	2
A1_2	0.710	−0.089
A1_3	0.665	−0.161
A1_8	0.646	−0.108
A1_1	0.611	0.023
A1_7	0.583	0.101
A2_3	0.577	−0.347
A1_4	0.565	0.017
A2_5	0.556	−0.157
A1_6	0.555	−0.002
A3_2	0.549	−0.094
A2_4	0.547	−0.372
A1_5	0.526	−0.330
A3_7	0.520	0.249
A3_1	0.441	−0.195
A3_6	0.153	0.789
A3_4	0.120	0.759
A3_3	0.183	0.624
A2_2	0.358	0.531
A2_1	0.393	0.466
A3_5	0.299	0.438

Notes: Item numbers of the environmental attitude variables are reported in Table A2. Bold items (A1_2 and A3_6) indicate the selected representative indicators for the pro-environmental and anthropocentric components, respectively, based on their dominant factor loadings. Shaded background colors are used solely for visual clarity to distinguish items primarily associated with Component 1 and Component 2; they do not convey additional statistical meaning. The extraction method is principal component analysis (PCA), and the rotation method is Varimax with Kaiser normalization. Rotation converged in six iterations.

). Coded indicators (e.g., A1_2, A3_6) correspond to specific questionnaire statements. To improve transparency, item wording and code mapping are provided in the Appendix A Table A2 and cross-referenced in the main text. As summarized in

Table 4. The representative indicators of latent environmental attitudes (PCA).

	Component 1	Component 2
Question items classified ↓ Extraction procedures (Group characteristics) ↓ Final indicators	A3_6, A3_4, A2_2, A2_1, A3_5, A3_3	A1_2, A1_3, A1_4, A1_1, A2_4, A2_3, A2_5, A1_5, A1_7, A3_7, A1_8, A1_6, A3_1, A3_2
	Anthropocentric attitude	Ecocentric attitude
	↓	↓
	A3_6	A1_2

Note: Component abbreviations (A1_1, A2_3, and related codes) refer to survey items listed in the questionnaire; the full wording of each item is reported in detail in Table A2 and Table A4. The arrows indicate the stepwise selection process from grouped attitudinal items to the final representative indicators used in the hybrid choice model.

, Component 1 reflects a utilitarian and self-benefit-centred perspective on nature. Its representative indicator is item A3_6, which asks whether “human consumption benefits outweigh environmental pollution.” In the context of Korean forests, this item captures respondents’ general acceptance of trade-offs between economic use—such as timber production, land use, or development—and environmental degradation affecting forest ecosystem services. Approximately 35% of respondents agreed with this statement. In contrast, Component 2 captures pro-environmental and emotionally connected views of nature, best represented by item A1_2, where over 76% of respondents agreed that “enjoyment of nature significantly contributes to stress relief.” To improve interpretability, the full decoding list of all 20 attitudinal items—along with their factor loadings and thematic labels—is shown in Table 4. These items help illustrate how the two components reflect distinct cognitive and affective orientations toward the environment [60,61]. Of the 20 items, approximately 13 loaded strongly onto Component 2 (ecocentric), while the remainder loaded onto Component 1 (anthropocentric). These two components serve as attitudinal constructs and are incorporated into the HCM as covariates to explain preference heterogeneity.

Table 4 presents the two key items (A3_6 and A1_2) used as interaction variables within the choice model to test whether attitudinal segmentation improves the model fit and reveals meaningful preference structures (see also Table A3).

Table 5. Response distribution of A3_6 and A1_2 by component.

	Strongly Disagree	Disagree	Neutral	Agree	Strongly Agree
A3_6 (Ratios)	86 (8.4%)	319 (31.2%)	251 (24.6%)	269 (26.3%)	96 (9.4%)
A1_2 (Ratios)	4 (0.4%)	45 (4.4%)	191 (18.7%)	541 (53.0%)	240 (23.5%)

summarizes the response distribution for these two indicators, and further detail is provided in

Table A5. Class-specific respondent characteristics.

	Class 1		Class 2
	Mean	Standard Deviation	Mean	Standard Deviation
Socioeconomic background
Gender	0.55	0.49	0.57	0.49
Age	40.94	12.34	39.76	10.46
Income	3.24	1.36	3.22	1.30
Education	3.96	0.60	3.97	0.51
Environmental attitude and behaviour indicators
A1_1	3.21	0.96	4.05	0.72
A1_2	3.65	0.78	4.39	0.58
A1_3	3.71	0.77	4.40	0.60
A1_4	3.34	0.95	4.10	0.77
A1_5	3.94	0.85	4.46	0.61
A1_6	3.33	0.98	4.29	0.73
A1_7	3.13	0.93	4.20	0.74
A1_8	3.36	0.90	4.22	0.67
A2_1	3.23	0.87	3.97	0.77
A2_2	3.24	0.88	4.00	0.86
A2_3	3.85	0.85	4.43	0.64
A2_4	3.82	0.86	4.33	0.65
A2_5	3.56	0.87	4.26	0.74
A3_1	3.81	0.80	4.24	0.64
A3_2	3.67	0.80	4.36	0.62
A3_3	2.60	1.06	3.46	1.30
A3_4	2.88	0.89	3.37	1.14
A3_5	3.28	1.00	4.01	1.01
A3_6	2.69	0.97	3.39	1.22
A3_7	2.80	0.84	3.79	0.98

.

4.3. Results of HCM and Mixed Logit Estimation

Table 6. Results of the latent-class HCM and MXL.

	Models	HCM					Mixed Logit Model (MXL)
		Class 1		Class 2
Attributes and Variables		Estimate	Rob.t.rat.	Estimate	Rob.t.rat.		Estimate (mu)	Rob.t.rat	S.D.	Rob.t.rat.
Forest fire risk (FFR)	Fire 2	−0.36	−2.82 **	0.07	1.31	FFR	0.28	1.86 *	2.22	8.31 **
	Fire 3	−0.33	−2.47 **	0.05	0.81
	Fire 4	−0.57	−3.39 **	0.02	0.40
Forest pest and disease risk (FPDR)	Pest 2	−0.09	−0.93	−0.01	−0.16	FPDR	0.24	1.58 *	1.46	4.33 **
	Pest 3	−0.12	−1.17	−0.10	−1.73 *
	Pest 4	−0.24	−2.07 **	0.00	−0.05
Forest-related outdoor restriction (FROR)	Restrn 2	0.16	1.43	0.12	2.30 **	FROR	−0.01	−0.73	−0.12	−3.17 **
	Restrn 3	0.04	0.32	0.12	2.14 **
	Restrn 4	0.24	2.05 **	0.02	0.39
Biodiversity loss (BL)	Biolos 2	−0.02	−0.19	0.10	1.86 *	BL	0.00	0.23	0.18	4.73 **
	Biolos 3	−0.08	−0.70	0.08	1.38
	Biolos 4	−0.06	−0.49	0.03	0.49
Forest restoration fund (FRF)	Cost	−0.26	−5.61 **	−0.16	−7.97 **	FRF	−0.03	−12.25 **	0.04	17.19 **
Alternative specific constant (ASC)	Asc_A	−0.68	−3.08 **	2.13	8.36 **	ASC	−2.40	−17.75 **	2.81	21.01 **
	Asc_B	0.18	0.85	2.46	10.73 **
Class membership	$c_2$	1.66	1.64 *	-	-		-	-	-	-
	${\delta }_{2_1}$(A3_6)	1.19	1.21	-	-		-	-	-	-
	${\delta }_{2_2}$(A1_2)	3.04	1.33	-	-		-	-	-	-
	Measurement component of the HCM model							-
		Estimate	Rob.t.rat. (0)
Environmental attitude indicators	${\xi }_1$(A3_6)	−0.92	−2.51 **		-		-	-	-	-
	${\tau }_{1_1}$(A3_6)	−2.71	−9.87 **		-		-	-	-	-
	${\tau }_{1_2}$(A3_6)	−0.50	−5.33 **		-		-	-	-	-
	${\tau }_{1_3}$(A3_6)	0.69	6.57 **		-		-	-	-	-
	${\tau }_{1_4}$(A3_6)	2.58	9.84 **		-		-	-	-	-
	${\xi }_2$(A1_2)	0.19	1.58 ***		-		-	-	-	-
	${\tau }_{2_1}$(A1_2)	−5.55	−11.08 **		-		-	-	-	-
	${\tau }_{2_2}$(A1_2)	−3.00	−20.26 **		-		-	-	-	-
	${\tau }_{2_3}$(A1_2)	−1.19	−15.87 **		-		-	-	-	-
	${\tau }_{2_4}$(A1_2)	1.19	15.81 **		-		-	-	-	-
Goodness of fit (Log-likelihood, AIC, BIC)	(−9393.14, 24,250.20, 18,316.68)						(−9116.77, 18,257.54, 37,714.77)

Notes: ***, **, and * indicate statistical significance at the 5% and 10% levels, respectively. The simulated log-likelihood is based on 1000 draws using Sobol sequences.

presents the results of two estimated choice models: an HCM estimated using the Apollo package (version 0.2.5) in R (version 4.0.2; Choice Modelling Centre, University of Leeds, Leeds, UK) [62] and an MXL model. Among them, the HCM outperformed the others across fit criteria, achieving better explanatory power (pseudo-R² = 0.17), an AIC of 24,250, and the lowest BIC (18,316). The MXL model improved in terms of the log-likelihood and pseudo-R² (0.14) but resulted in a considerably inflated BIC (37,714), likely due to the model complexity and reduced parsimony. These results highlight the strengths of each model: MXL captures unobserved preference heterogeneity, and HCM incorporates both heterogeneity and latent attitudinal constructs. Rather than viewing the models as mutually exclusive, this study adopts a complementary modelling strategy, using each framework to shed light on different dimensions of forest restoration preferences and to enhance the robustness of policy-relevant insights. In Table 6, a positive coefficient indicates higher utility (and choice probability) for an attribute level relative to the reference level, while a negative coefficient indicates lower utility, ceteris paribus. For the payment attribute, a negative sign is expected and represents disutility from higher contribution levels. Thus, coefficient signs should be interpreted as directions of relative preference in the stated-choice context (not absolute support for regulation), and WTP is derived from the ratio of non-monetary coefficients to the payment coefficient.

The HCM finally reveals two latent classes that exhibit distinct forest policy orientations. Class 1 is characterized by strong aversion to forest fire risk, pest and disease damage, and restoration costs, as reflected in statistically significant negative coefficients (e.g., Fire2: t = −2.82, Fire4: t = −3.39, cost: t = −5.61).

The latent class model identifies two distinct classes with differing preference structures. In Class 1, coefficients for forest fire risk and forest pest and disease risk are statistically significant and negative, indicating higher sensitivity to disaster-related attributes. The cost coefficient is also statistically significant and negative, suggesting strong cost sensitivity. In contrast, Class 2 shows positive and statistically significant coefficients for reductions in forest access restrictions (Restrn2–3: t = 2.30, 2.14) and a moderately significant coefficient for biodiversity protection (Biolos2: t = 1.86). Risk-related attributes are generally not statistically significant in this class.

Attitudinal variables A3_6 and A1_2 are statistically significant predictors of class membership. Individuals who agree with A3_6 are more likely to belong to Class 1, whereas those who endorse A1_2 have a higher probability of belonging to Class 2. The estimated coefficients for the environmental attitude indicators selected through PCA are −0.92 for ξ₁(A3_6) and 0.19 for ξ₂(A1_2). The negative coefficient for A3_6 indicates an inverse relationship between anthropocentric attitudes and overall environmental attitudes, whereas the positive coefficient for A1_2 suggests a positive association with pro-environmental attitudes. The corresponding threshold (τ) parameters indicate ordered responses consistent with these latent constructs.

The MXL model shows statistically significant coefficients for forest fire risk (t = 1.86) and forest pest and disease risk (t = 1.58). The estimated standard deviations for these attributes, as well as for forest access restrictions, are statistically significant, indicating substantial unobserved preference heterogeneity across individuals.

Table 7. WTP estimation results (unit: KRW).

Attributes	MXL (Mean)	Levels	HCM
			Class 1		Class 2
			Mean	95% CI	Mean	95% CI
Forest fire risk	10,287 * (USD 7.9)	Fire2	−14,129 *	−22,199~−6059	4193	−1631~10,018
		Fire3	−12,867 *	−21,013~−4721	2950	−3431~9331
		Fire4	−22,412 *	−32,425~−12,398	1498	−4760~7757
Forest pests and disease risk	8661 * (USD 6.6)	Pest2	−3562	−10,933~3809	−534	−6471~5402
		Pest3	−4657	−12,881~3566	−6249 *	906~408
		Pest4	−9057 *	−17,802~−311	−295	−7262~6671
Forest-related outdoor restriction	−377 (USD 0.3)	Restrn2	6253	−1431~13,938	7191 *	1056~13,325
		Restrn3	1459	−6527~9445	7249 *	857~13,642
		Restrn4	9264 *	810~17,718	1422	−5081~7927
Forest biodiversity loss	146 (USD 0.1)	Biolos2	−669	−8558~7219	5927 *	−416~12,271
		Biolos3	−3000	−11,076~5075	4855	−1672~11,382
		Biolos4	−2198	−10,571~6174	1829	−4818~8476

Notes: In 2023, USD 1 was equal to KRW 1305.8 (https://ecos.bok.or.kr/#/SearchStat; accessed on 28 August 2023); 95% CI: 95% confidence interval; * indicates the value obtained from a statistically significant estimate from Table 6.

reports WTP estimates derived from both the MXL and HCM. In the MXL model, respondents exhibit statistically significant mean WTP for reductions in forest fire risk (KRW 10,287) and forest pest and disease risk (KRW 8661). WTP estimates for recreation access and biodiversity protection are not statistically significant at the population level.

In the HCM, WTP estimates vary across latent classes. Class 1 shows high and statistically significant WTP for fire risk reduction (e.g., Fire2: KRW 14,129; Fire4: KRW 22,412) and moderate WTP for pest and disease risk reduction. Class 2 exhibits statistically significant WTP for relaxing recreational restrictions (Restrn2: KRW 7191; Restrn3: KRW 7249) and moderate positive WTP for biodiversity protection.

Figure 1 illustrates the conditional class membership probabilities for Class 1 in relation to two attitudinal indicators: A3_6 (anthropocentric attitude) and A1_2 (ecocentric attitude). For A3_6, those who disagreed with the anthropocentric statement showed low probabilities of belonging to Class 1 (typically 0.1–0.3), while agreement was associated with slightly higher probabilities (0.3–0.5), though rarely exceeding 0.5. A similar pattern was observed for A1_2, where disagreement corresponded with low probabilities (<0.3), and even among those who agreed or strongly agreed, probabilities remained concentrated between 0.3 and 0.5. These patterns suggest that neither anthropocentric nor ecocentric attitudes strongly predict Class 1 membership. Instead, Class 1 appears to reflect individuals with heightened risk sensitivity and cost aversion, rather than distinct environmental value orientations. While A3_6 and A1_2 offer some explanatory value, they are not definitive predictors of class assignment, highlighting the need to integrate multiple attitudinal and behavioural factors when interpreting latent class structures. To support this interpretation,

Table 8. Summary of conditional class membership by attitude response on A3_6 and A1_2 with full sample.

	Mean		Median		Standard Deviation
	A3_6	A1_2	A3_6	A1_2	A3_6	A1_2
Strongly	0.238	0.388	0.214	0.377	0.058	0.045
Disagree	0.276	0.380	0.247	0.375	0.053	0.076
Neither	0.339	0.361	0.320	0.356	0.058	0.070
Agree	0.373	0.320	0.356	0.323	0.066	0.073
Strongly	0.390	0.294	0.357	0.293	0.067	0.076

provides descriptive comparisons of attitudinal responses across classes, complementing the conditional membership patterns shown in Figure 1 (Detailed results for Class 2 are provided in Appendix B).

To facilitate segment-level interpretation, we report estimates not only for the full sample but also separately for Class 1 (A3_6) and Class 2 (A1_2) in Table 8,

Table 9. Summary of the WTP by attitude response with class 1 (A3_6).

		Strongly Disagree	Disagree	Neither	Agree	Strongly Agree
Forest fire risk	Fire 2	−157.95	−857.31	−2003.01	−2633.02	−2942.86
	Fire 3	−806.02	−1409.72	−2398.70	−2942.53	−3209.99
	Fire 4	−4177.74	−5090.08	−6584.68	−7406.55	−7810.76
Forest pests and disease risk	Pest 2	−1253.04	−1368.51	−1557.68	−1661.71	−1712.87
	Pest 3	−5868.26	−5807.06	−5706.81	−5651.68	−5624.56
	Pest 4	−2375.15	−2709.39	−3256.93	−3558.03	−3706.11
Forest-related outdoor restriction	Restrn 2	6965.14	6928.84	6869.36	6836.66	6820.58
	Restrn 3	5871.86	5650.38	5287.56	5088.04	4989.92
	Restrn 4	3283.45	3582.50	4072.41	4341.81	4474.30
Forest biodiversity loss	Biolos 2	4358.61	4106.49	3693.45	3466.33	3354.63
	Biolos 3	2988.18	2688.09	2196.48	1926.15	1793.20
	Biolos 4	872.19	718.39	466.44	327.89	259.75

Note: In 2023, USD 1 was equal to KRW 1305.8 (https://ecos.bok.or.kr/#/SearchStat; accessed on 28 August 2023).

and

Table 10. Summary of the WTP by attitude response with class 2 (A1_2).

		Strongly Disagree	Disagree	Neither	Agree	Strongly Agree
Forest fire risk	Fire 2	−2908.86	−2750.21	−2406.35	−1663.80	−1189.55
	Fire 3	−3180.64	−3043.69	−2746.87	−2105.89	−1696.51
	Fire 4	−7766.39	−7559.43	−7110.86	−6142.18	−5523.50
Forest pests and disease risk	Pest 2	−1707.25	−1681.06	−1624.28	−1501.68	−1423.37
	Pest 3	−5627.54	−5641.42	−5671.51	−5736.49	−5777.99
	Pest 4	−3689.85	−3614.03	−3449.70	−3094.82	−2868.17
Forest-related outdoor restriction	Restrn 2	6822.34	6830.58	6848.43	6886.97	6911.59
	Restrn 3	5000.69	5050.93	5159.82	5394.98	5545.17
	Restrn 4	4459.76	4391.92	4244.88	3927.36	3724.57
Forest biodiversity loss	Biolos 2	3366.89	3424.08	3548.04	3815.74	3986.71
	Biolos 3	1807.79	1875.87	2023.41	2342.03	2545.53
	Biolos 4	267.23	302.12	377.74	541.03	645.33

Note: In 2023, USD 1 was equal to KRW 1305.8 (https://ecos.bok.or.kr/#/SearchStat; accessed on 28 August 2023).

. This allows direct comparison of preference strength, risk sensitivity, and WTP profiles across latent segments. The class differences shown in Table 8 provide a descriptive snapshot of how attitudinal expressions vary across segments. These tables are included to support the exploratory interpretation of class profiles and to complement the conditional membership analysis.

Table 9 summarizes the WTP estimates derived from the relationship between forest management policies, restoration attributes, and respondents’ environmental attitude responses (A3_6). For the forest fire attribute, WTP is relatively high at the “strongly agree” response level, with the highest WTP being observed at the safest level (Fire 4). In the case of the forest pest and disease attribute, WTP is notably high at the “strongly disagree” response level, peaking at the intermediate level (Pest 3). The forest-related outdoor activity restriction exhibits the highest WTP at the initial level of restriction (Restrn2), followed by a declining trend as restrictions intensify. Lastly, the forest biodiversity loss attribute shows the highest WTP at the “strongly disagree” level, with WTP decreasing as responses shift toward the “strongly agree” level.

Table 10 presents the WTP estimates based on the relationship between forest management restoration attributes and respondents’ environmental attitude responses (A1_2). For the forest fire attribute, WTP is relatively high at the “strongly disagree” response level, with the highest overall WTP observed at the safest stage (Fire 4). Similarly, the forest pests and disease attribute show elevated WTP at the “strongly disagree” response level, peaking at the intermediate stage. The forest-related outdoor activity restriction attribute demonstrates the highest WTP at the initial level of constraints imposed by forest policy, followed by a declining trend; however, this relationship is not strongly linked to respondents’ attitudes. Finally, the forest biodiversity loss attribute exhibits an overall high WTP at the “strongly agree” response level.

The application and analysis results presented in Table 6, Table 7, Table 8, Table 9 and Table 10 and Figure 1 contribute to foundation-based research for advancing sustainable forest restoration policies. By identifying heterogeneous environmental preferences and behaviours among citizens, the findings support a shift beyond traditional market-oriented approaches, such as WTP estimation and feasibility assessments, toward designing more behaviourally informed policy strategies.

5. Discussion

5.1. Contextualizing Preference Heterogeneity in Forest Restoration

The results of this study reveal substantial heterogeneity in public preferences for forest restoration, consistent with a growing body of literature emphasizing that environmental decision-making is shaped by both economic trade-offs and underlying psychological factors [30,52,63,64]. By integrating discrete choice modelling with latent attitudinal structures, this study extends previous applications of hybrid choice models in environmental valuation [13,14,18,65]. The identification of two latent classes highlights the existence of distinct stakeholder segments with fundamentally different orientations toward forest restoration. Such segmentation aligns with prior findings that public support for environmental policies often reflects divergent perceptions of risk, cost, and ecological values rather than uniform preferences [66,67].

5.2. Interpretation of Latent Classes

The first latent class is characterized by heightened sensitivity to disaster-related risks and strong cost aversion. Individuals in this group place significant value on reducing forest fire risk and pest and disease threats, suggesting a preference for policies that prioritize immediate risk mitigation, public safety, and financial efficiency. This pattern is consistent with studies showing that exposure to environmental hazards increases demand for preventive and protective measures [68]. In contrast, the second latent class places greater emphasis on biodiversity conservation and recreational access, reflecting a longer-term, amenity-oriented, and ecologically motivated perspective on forest management. This group appears less reactive to short-term risks but more supportive of policies that enhance ecosystem quality and public access to natural spaces. Similar preference structures have been documented in studies focusing on cultural ecosystem services and nature-based recreation [69,70].

The age distribution is relatively concentrated in the 40–49 group, which may affect external validity. Accordingly, the estimated WTP and class shares should be interpreted as most applicable to populations with similar demographic composition, and future studies should test robustness using more age-balanced samples. Although respondents in their 40s were overrepresented (60.8%), we used stratified quotas by region and gender and benchmarked sample structure against national statistics (KOSIS). This age concentration may affect external validity of WTP levels; thus, our WTP estimates should be interpreted as internally valid for the sampled population, with caution in nationwide extrapolation.

5.3. Role of Environmental Attitudes

The inclusion of environmental attitude indicators (A3_6 and A1_2) in the class membership function provides important behavioural insights. Individuals with stronger anthropocentric orientations are more likely to belong to the risk-averse and cost-sensitive class, whereas those with pro-environmental attitudes tend to fall into the ecologically oriented class. Although these attitudinal variables do not solely determine class membership, they capture latent motivational structures that help explain observed heterogeneity in choice behaviour. This finding reinforces the argument that environmental attitudes act as cognitive filters through which individuals interpret policy trade-offs, influencing their willingness to support restoration measures [53,64]. Incorporating such latent constructs therefore enhances the behavioural realism of choice models in forest policy analysis.

The findings suggest that environmental attitudes function as behavioural filters rather than dominant determinants of valuation. In forest-restoration contexts characterized by high uncertainty and disaster salience, perceived risk and expected private burden can outweigh general pro-environmental dispositions. This helps explain why attitudinal significance coexists with moderate class-separation power and supports integrated modelling strategies that combine psychological and economic drivers.

5.4. Comparing Insights from HCM and MXL

The MXL results further confirm the presence of substantial unobserved preference heterogeneity, particularly with respect to forest fire and pest risk attributes. Significant mean and standard deviation estimates indicate that while average willingness to pay for risk reduction is positive, individual responses vary considerably. This suggests that restoration policies targeting ecological risks may generate uneven public support depending on perceived threat levels. However, unlike the HCM, the MXL framework does not explicitly reveal the attitudinal mechanisms underlying these differences. While MXL is effective for identifying the dispersion of preferences, the HCM offers a more comprehensive interpretive framework by linking choice behaviour to latent environmental values. This distinction is consistent with previous comparative studies highlighting the complementary strengths of these modelling approaches [18,64].

5.5. Policy Implications

Taken together, the findings suggest that forest restoration policies should move beyond uniform, one-size-fits-all approaches. Risk-averse stakeholders may respond more favourably to interventions emphasizing disaster prevention, early warning systems, hazard zoning, and publicly funded restoration programmes. In contrast, ecologically oriented stakeholders may be more supportive of participatory restoration initiatives, biodiversity conservation incentives, and investments in forest-based recreational infrastructure. Moreover, tailoring policy communication strategies—such as risk- and cost-focused messaging for one group and value- and nature-oriented narratives for the other—can enhance public acceptance and policy legitimacy. These insights demonstrate the value of integrating attitudinal heterogeneity into forest governance and restoration planning.

To contextualize economic magnitude, WTP estimates should be interpreted relative to household payment burden and existing forest-related public finance in Korea. In our sample, support for risk reduction is economically meaningful but heterogeneous across segments, implying that a uniform levy may generate uneven acceptance. A differentiated policy mix—combining risk-mitigation investments, targeted communication, and gradual contribution design—would better align fiscal feasibility with public support.

6. Conclusions

This study examines public preferences for forest restoration by integrating discrete choice experiments with hybrid choice modelling. By explicitly accounting for environmental attitudes, the analysis reveals substantial heterogeneity in willingness to pay and preference structures across individual groups.

The results show strong public support for reducing forest fire and pest risks, particularly among risk-averse and cost-sensitive individuals, while biodiversity conservation and recreational access are more highly valued by ecologically oriented stakeholders. These differentiated valuation patterns underscore the importance of segmented and adaptive forest restoration strategies.

Methodologically, this study demonstrates that hybrid choice models provide a richer behavioural foundation than conventional models by linking economic preferences to latent environmental values. From a policy perspective, incorporating attitudinal heterogeneity can improve the effectiveness, efficiency, and social responsiveness of forest restoration initiatives.

Several limitations should be acknowledged. Attitudinal variables were incorporated only in the class membership component, stated preference data may involve hypothetical bias, and the country-specific nature of the dataset may limit generalizability. Future research could address these limitations by incorporating attitudes directly into utility functions, applying longitudinal designs, and extending the analysis to comparative international contexts.

Overall, this study contributes to both the methodological advancement of environmental valuation and the practical design of differentiated, attitude-informed forest restoration policies capable of addressing the complexity of contemporary forest governance.

Because both analyses are based on the same survey platform, direct comparisons should be interpreted as complementary evidence rather than independent replication. Our added contribution is a distinct behavioural mechanism analysis that prioritizes class-specific trade-offs and policy targeting over average-effect reporting alone.