Exploring Preference Heterogeneity and Acceptability for Forest Restoration Policies: Latent Class Choice Modeling and Principal Component Analysis

Abstract

The restoration of forest ecosystems damaged by wildfires and pest outbreaks has become increasingly urgent. However, the public-good nature of forests, the involvement of diverse stakeholders, and the spatial variability of degradation present significant challenges to effective policy design. In particular, previous studies have largely examined these threats in isolation, and few have provided integrated economic analyses of their combined impacts. This gap underscores the need to better understand heterogeneous public preferences and their implications for restoration policy. To address this, we conducted a discrete choice experiment (DCE) with 1021 Korean citizens and applied a two-stage analytical framework combining principal component analysis (PCA) and latent class choice modeling (LCM). Five distinct preference segments were identified, each exhibiting substantial variation in willingness to pay (WTP) for restoration attributes. Policy simulations further revealed that public acceptance declines sharply at higher cost levels, highlighting the importance of setting realistic financial thresholds for broad support. While visual materials, consequentiality checks, and cheap talk scripts were employed to mitigate hypothetical bias, the limitations of external validity and potential sampling biases should be acknowledged. Our findings provide empirical evidence for tailoring restoration policies to different stakeholder groups, while also stressing the financial and institutional constraints of implementation. In particular, the results suggest that cost thresholds, citizen engagement, and awareness-raising strategies must be carefully balanced to ensure both effectiveness and public acceptance. Taken together, these insights contribute to evidence-based forest policymaking that is both economically efficient and socially acceptable, while recognizing the context-specific limitations of the Korean case and the need for comparative studies across countries.

1. Introduction

Since the Korean War, South Korea (hereafter Korea) has been globally recognized as a leading nation in forest restoration [1,2]. However, the quality of its forest ecosystems has deteriorated significantly in recent years due to large-scale forest fires, landslides, and the spread of invasive species [3,4,5,6]. These ecological disturbances are expected to intensify further with climate change [7,8,9]. According to the FAO (2020), forests continue to face alarming rates of degradation and deforestation due to anthropogenic pressures—such as fires, invasive species, and logging—that undermine their health, vitality, and ability to provide essential ecosystem services [10,11]. Without active restoration, achieving sustainable forest management (SFM) and meeting the Sustainable Development Goals (SDGs) will become increasingly difficult, while the public benefits provided by forests will likely diminish [12]. Although Korea urgently needs to restore deforested and degraded forests, such efforts require substantial financial resources and long-term commitment in alignment with the United Nations’ Decade on Ecosystem Restoration (2021–2030).

A key challenge in designing and implementing effective restoration policies lies in the considerable heterogeneity of public preferences, values, and attitudes, and while previous research has examined wildfires in terms of prevention, suppression, and restoration costs [13,14,15] (as well as pest outbreaks in relation to timber losses, control expenditures, and forest health impacts [16,17,18]), very few studies have jointly analyzed these two critical drivers of ecosystem degradation in an economic context [19,20,21]. Despite its importance, the segmentation of environmental behavior has been relatively underexplored in forest-related research compared to other sectors like health or transportation [22,23,24]. By explicitly addressing this gap, the present study contributes to the literature by jointly analyzing preferences for restoration under wildfire and pest scenarios, thereby providing a more comprehensive understanding of how the public values ecosystem recovery when facing multiple, interacting threats.

To reduce the uncertainty of forest restoration policy, therefore, the discrete choice experiment (DCE) provides valuable tools to uncover this heterogeneity and to design and implement forest restoration policies effectively. It is essential to account for the latent environmental preferences of citizens and various stakeholders [19,20]. If environmental preferences are treated as uniform, it can lead to policy failures and public resistance, yet overly fine segmentation can also be inefficient. Grouping stakeholders based on latent attitudes and behaviors offers a practical balance in helping to design targeted and acceptable forest restoration strategies [20,21]. Moreover, as economic valuation studies focusing on forest fires, pests, and ecological risks are still rare, estimating willingness to pay (WTP) is critical for evaluating the feasibility and efficiency of such policies [15,20,23,24].

The purpose of our study is to identify heterogeneous preferences, estimate WTP, and assess acceptability for forest restoration policies using a two-stage modeling framework combining latent class choice modeling (LCM) and principal component analysis (PCA) [25]. First, we estimated latent classes through LCM; then, PCA was applied to further explore group-level environmental behaviors and WTP patterns. This combined approach enabled a deeper understanding of how citizens’ and stakeholders’ preferences influence support for forest restoration. Furthermore, we used each of the estimated LCM and PCA classes simultaneously to cross-check the distribution of WTP by groups and attributes in indicating their environmental preferences. The remainder of our study is structured as follows. The next section sets the stage with a review of econometric model specification and literature. Section 3 presents our empirical study, while the estimation results are discussed in Section 4. This paper then concludes with a summary in Section 5. Additionally, the appendix is at the back of this paper.

2. Discrete Choice Framework and Model Specification

2.1. Evidence on Preference Heterogeneity and Environmental Risks

To explore heterogeneity in environmental preferences, this study employed both principal component analysis (PCA) and latent class choice modeling (LCM). PCA is a straightforward statistical technique that reduces dimensionality while preserving the maximum possible variation in the dataset [25,26], and it has been widely applied to uncover latent attitudinal structures across fields such as environment, health, and transportation [15,27,28,29,30,31,32,33,34,35,36,37,38,39]. For example, Rhead et al. (2018) used PCA to classify environmental concerns into four ecological groups [32], while Chen et al. (2019) identified three distinct preference classes for urban river restoration using LCM [14,33]. Other applications have compared different econometric models—such as LCM, mixed logit, and random parameter logit—in analyzing heterogeneous preferences for recreation or environmental management [15,34]. These studies consistently demonstrate that unobserved heterogeneity in environmental preferences can be meaningfully captured through latent class segmentation, providing important insights for policy design [31,32,33,34,35,36,37,38].

2.2. Choice Experiment and Model Specification

The DCE we applied is based on Lancaster’s theory of consumers [39], which holds that consumer utility is based on the characteristics of a good rather than on the product or service itself. It is a theory of the consumer based on the economic rationality of utility maximization advocated by McFadden’s random utility theory [40], which assumes that randomly selected individual consumers are economically rational and utility-maximizing decision makers concerning the given alternatives. Under this random utility theory, the utility derived from a choice decision in choice modeling is composed of an observable part, such as the choice attributes that reflect the respondents’ preferences in the presented choice set and an unobservable component (the stochastic error component). The underlying utility function applied in our study follows the assumptions of the basic model of choice modeling. The DCE is a very efficient methodology for capturing a wide range of latent preferences, which consist of a variety of attributes and levels that can define environmental goods or environmental services (although modeling a wide range of social preferences, in general, poses practical difficulties).

In a fresh approach to explore these latent environmental preferences, we used a combined LCM approach in a stepwise fashion [41,42,43], linking the latent class model classification of PCA with the LCM approach. In general, multinomial logit models assume that all respondents have the same preference parameters, but in practice, each respondent’s preference parameters might be different. To overcome this limitation, we applied the combined LCM model, which assumes a discrete distribution of preference coefficients, where the choice behavior of individuals is based on observable attributes and latent heterogeneity due to variation in attribute factors that are theoretically unobserved by the researcher [15]. The advantage of an LCM is that it does not require any special assumptions about the shape of the parameter distributions for individual consumers. Instead, individuals are essentially partitioned into finite identifiable classes, and the partitioning/grouping process is stochastic based on the selected environmental attributes and levels. Environmental preferences are assumed to be heterogeneous across classes and homogeneous within classes. The remaining advantage is that the heterogeneity of environmental preferences can be inferred, and the identifiers that characterize the environmental characteristics of the group can be uniformly identified in the model framework. As a result, we demonstrated that the heterogeneity of environmental preferences in specific DCE environmental data can be better described by the LCM model. We experimentally analyzed the distribution of latent environmental preferences and WTP by considering those grouped by the PCA in addition to the procedures applied in many of these studies [25]. These flexible advantages make latent class models more popular in various fields of consumer theory, market analysis, and the environment [44].

Given this situation, we used two statistical methodologies in our study: LCM and PCA. In this section, we mainly describe the theoretical background of LCM. To understand these latent and complicated attributes behind the numbers and choices in our DCE data, this LCM, which is a type of latent class model that uses a probabilistic class allocation model, consists of two sub-models as the first part of a class-specific choice model and the last part of a class membership model [45]. The former part assigns the probability of choosing each alternative [46], and the latter contains the likelihood of an individual $n$ belonging to a certain class, given that individual $n$ belongs to a certain class $q$ [47,48,49,50]. In general, the class membership model is defined as logit, where the utility of an individual $n$ belonging to class $q$ can be expressed as Equations (4) and (5) [51].

The class-specific choice model specified with the MNL model returns the probability that individual $n$ chooses a specific alternative $i$ on choice task $t$, given that respondents belong to a certain class $q$. To begin, the utility of individual $n$ derives from choosing an alternative $i$ choosing task $t$, given that respondents belong to the class $q$, and it is formulated as Equation (1):

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

(1)

where ${ V}_{nit|q}$ is the observed part of the utility; ${\epsilon }_{nit}$ is an independent and normally distributed Extreme Type I error term over individuals, alternatives, and classes; $X_{nit}$ is the vector of observed attributes of alternative $i$ and normally includes an alternative-specific constant ($ASC$) that captures the systematic utility of omitted variables [52]; and ${\beta }_q$ is the vector of unknown parameters that need to be estimated.

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

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

(2)

The probability of the sequence of $T_n$ choices made by an individual $n$, denoted by $y_n$, is the product of the probability associated with each chosen alternative in their sequence:

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

(3)

The probability that individual $n$ belongs to class $q$ given their characteristics is represented by the following logit formula:

$$P(q)={\displaystyle \frac{e^{\left({\alpha }_q\right)}}{{\sum }_{q=1}^Q{e}^{\left({\alpha }_q\right)}}},$$

(4)

where ${\alpha }_1$ = 0 for identification reasons.

Therefore, the overall choice probability, which is weighted across all classes by the class membership probability $P(q$), of observing $y_n$ is formulated as Equation (5):

$$P(y_n| X_n, \beta , Q)={\sum }_{q=1}^{Q}P\left(q\right)P\left(y_n\right|X_n, {\beta }_q)={\sum }_{q=1}^Q{\displaystyle \frac{e^{\left({\alpha }_q\right)}}{{\sum }_{q=1}^Q{e}^{\left({\alpha }_q\right)}}}{\prod }_{t=1}^T\left(p\right(y_{nit}|X_{nit}, {\beta }_q)).$$

(5)

Finally, the overall log-likelihood of all individuals $N$ for our LCM is as follows:

$$ln L={\sum }_{n=1}^{N}lnP\left(y_n\right| X_n, \beta , Q).$$

(6)

We employed this LCM formulation throughout our study. Our results provide a better model fit, with stronger explanatory power than the single model, and they capture more realistic environmental choice behaviors by incorporating environmental attitudes and behaviors. This may explain why our more complex model yields both higher and lower WTP estimates.

3. Empirical Study

3.1. Attributes, Experimental Design, and Scenario

The research questionnaire was divided into three sections. The first section employed a five-point Likert scale to measure the environmental attitudes, choices, and behaviors of the public [53,54,55], as well as to elicit their hidden specific environmental attitudes and final latent consequences because such factors are becoming increasingly important to address ecological problems in research [32]. Second, respondents read our specific environmental scenario, and descriptions of the attributes and levels were shown as an example of a choice experiment. Respondents were asked to complete a series of ten choice tasks depicting the current forest status (SQ) and two preferred forest management policies (alternative1 and alternative2) that would restore the degraded and deforested forests from the damage incurred by the two environmental threats of forest pests and diseases and forest fires. After selecting the preferred choice sets, respondents answered follow-up questions, such as identifying protesters, two types of consequentiality [56], prioritizing attributes, awareness of environmental threats to ecosystem services, and socioeconomic questions. The attributes and levels used in the choice experiment are outlined in

Table 1. The attributes and levels of a sustainable forest management scenario.

	Levels	Level 1	Level 2	Level 3	Level 4
Attributes
Forest fires		Extreme (Red) *	High level (Orange)	Moderate level (Yellow)	Low level (Blue)
		(Very high, 85↑)	(High, 66~85)	(Moderate, 51~65)	(Low, 51↓)
Forest pests and diseases		Extreme (Red) *	High level (Orange)	Moderate level (Yellow)	Low level (Blue)
Restriction on forest-related outdoor activities		No restrictions *	Slight restrictions	Moderate restrictions	Strong restrictions
		(zero)	(0~1 million)	(1~2 million)	(2~5 million)
Biodiversity loss (number of pine trees damaged)		Very large losses *	Large losses	Moderate losses	Slight losses
		(−300~−400 k trees)	(−200~−300 k trees)	(−100~−200 k trees)	(0~−100 k trees)
Forest recreation fund		Low *	Slight increase	Moderate increase	Substantial increase
		KRW 0~KRW 10,000	KRW 10,100~KRW 25,000	KRW 25,100~KRW 35,000	KRW 35,1000~KRW 50,000

Note: * It represents the status quo (SQ) reference; adapted from Jeon (2025) [58], Frontiers in Behavioral Economics, 4, 1596416. https://doi.org/10.3389/frbhe.2025.1596416, CC BY 4.0. ↑: More than 85, ↓: less than 51.

. Visual images were applied to present each attribute level to improve our questionnaire comprehension, and a cheap talk question was added to prevent respondents from overestimating their WTP and to remind them of the level of their disposable income before selecting the main choice tasks [23,57].

Our questionnaire and hypothetical scenario were carefully designed to capture the specific conditions of Korean forests, with a particular emphasis placed on environmental threats to ecosystem services and sustainable forest management (SFM) [59]. In discrete choice experiment (DCE) studies, the definition of attributes and levels plays a central role in shaping the experimental design. Thus, it is crucial to align the selected attributes with both theoretical considerations and practical assumptions about forest policy. An excessive number of attributes or levels can confuse respondents and may reduce the statistical efficiency of the design [23]. To address this, we considered current forest policies, fire occurrence patterns, and pest and disease control systems, and we also adjusted the attributes and levels to reflect biodiversity loss and the impacts of outdoor recreational activities. The cost attribute for forest restoration was specified across a range of payment intervals appropriate for the respondent population [60]. The following section provides details of the policy scenario developed for this study, including the final attributes and levels, the expert consultation process, the pilot survey, and the overall questionnaire design.

Considering the situation above, we selected five major attributes through a relevant literature review [60] and from expert advice from the National Institute of Forest Science (NIFoS) in Korea. To begin with, we examined a ‘forest fire’ threat that has the most significant negative impacts on ecosystem services, both at national and global levels, and we then selected the direct and critical attributes arising from them as attributes and variables. Forest fire risk is a growing factor that is simultaneously and continuously connected with successful growing stumpage stock. The damaged area and costs in forestry are becoming larger, even though the frequency of forest fires varies slightly from year to year [61,62]. For instance, the area affected by fire was 24,797 ha in 2022, but it was 2920 ha in 2020; hence, it has been steadily increasing over the years. The frequency and damage of forest fires are rapidly increasing and the Statistical Yearbook of the KFS (2021) reveals that the main causes of them include the following: ‘carelessness of visitors’ at 38.1%, ‘weed burning’ at 6.3%, ‘trash burning’ at 7.7%, ‘cigarette littering’ at 9.7%, ‘visitors to ancestral graves’ at 4.6%, ‘children’s carelessness’ at 0.9%, and at ‘building fire’ 5.7% of the recent forest fire incidents. Meanwhile, 26.9% of incidents are attributed to other minor causes [17]. Following the statistical yearbook of the KFS (Law No. 19115), the scheme of the KFS provides a four-level forest fire alert system [0 (lower risks) →100 (higher risk)] in accordance with the data analysis results of the NIFoS forest fire forecasting system (forestfire.nifos.go.kr): ‘blue level’ (under 50), ‘yellow level’ (51~65), ‘orange level’ (66~85), and ‘red level’ (higher than 86).

Secondly, regarding ‘forest pests and diseases’, a case of pine wilt diseases (PWDs) were unintentionally introduced from Japan and recognized as a global threat to the forestry sector, and they are also viewed as a major threat to healthy ecosystems in forests and to the provision of ecosystem services [7]. More than 12 million pine trees infected with PWDs have been urgently removed throughout all nations with a huge cost to prevent and control the spread of PWDs with bestial injection vaccination since 1988, and outbreaks of forest pest damage are about 35 million hectares of forests annually over the globe [9]. To combat these threats, the KFS has already implemented ‘the forest disease and pest outbreak index and alert system’ with four levels: ‘interest’ (blue, lower than 50), ‘caution’ (yellow, 51~65), ‘serious’ (orange, 66~85), and ‘very serious’ (red, higher than 86). These levels were established from the results of periodic surveillance and were based on the scale of occurrence, the speed of spread, and the degree of damage in accordance with Article No. 4 of the Forest Pest Control Regulation.

The third attribute was ‘restrictions on forest-related outdoor activities (or forest access restrictions)’to control and manage these two risky factors, but forest outdoor activities are one of the favored ecosystem services in Korea [3,35]. An average of 15 million tourists have visited annually for camping, hiking, healing, and education in forests [3]. As previously discussed, the majority of forest fires in Korea are caused by human activity, so we set up a scenario in which restricting outdoor activities related to forests would reduce the risk. Specifically, the assumption for the restriction attribute was that there are currently no restrictions; hence, the potential for forest fires and pests and diseases to spread is currently at high risk. Therefore, the direction of forest management policies in our research scenario was set to restrict these two environmental risks. Regarding the key information, there are more than 170 recreational sites within national, public, and private forests across the country, and the number of visitors has continuously increased from 4 million in the 2000s to more than 15 million in the 2020s, which is about 15 million tourists that visit the facilities on average.

Our DCE design put these attributes onto four levels: zero restriction (no restriction), 1~1 million people restriction, 1~2 million people restriction, and 2~5 million people restriction. The fourth attribute was ‘biodiversity loss’ in forests, which includes the number of pine trees infected by PWDs as an impact directly related to and caused by the above two risk factors, where the causal pathogen is the pine wood nematode (PWN), Bursaphelenchus xylophilus [7]. In Korea, pine trees are one of the main trees infected by Bursaphelenchus xylophilus, which is a PWD transmitted by vector pine sawyer beetles in the genus Monochamus. It has also been supposed that once they are infected by PWDs, all pine trees must be chopped down or the suspected pine trees by infection are removed all together (https://www.apsnet.org (accessed on 15 August 2023)). Furthermore, pine trees hold a special place in Korean history as an evergreen tree with a long lifespan, and they are the most favorite trees with an iconic importance and a long historical background. It is also an economically important natural resource as timber worldwide. Based on the KFS statistics of PWD damage, the possible levels of pine trees damaged by two risky attributes are distributed across four levels: 0~100,000 trees, 100,000~200,000 trees, 200,000~300,000 trees, and 300,000~400,0000 trees.

The last monetary attribute, ‘forest restoration cost’, which charts the cost to restore deforested and degraded forests, can also be called a forest restoration fund. This fund is a tax instrument for a one-time payoff of one year per household because forest management and restoration in Korea is mainly funded by the government’s budget. To be specific, it is classified into 11 categories, including KRW 0~KRW 50,000 KRW (USD 38.29) by the unit of KRW 5000 KRW (USD 3.8). For reference, a citizen per year paid about KRW 6700 (USD 5.13) in taxes for the forest sector (KFS) in 2020.

To set our realistic forest restoration policy scenario to reflect the reality of Korea’s forests by including the situation thus far, we firstly set it such that the forest ecosystem is constantly threatened by forest fires and pests and diseases. Consequently, pine tree populations have been increasingly damaged, and both tourists and forest users have been adversely affected. To address this, the scenario assumes a restoration of the degraded forests. Table 1 summarizes the attributes and levels included in the questionnaire, each illustrated with visual materials. Respondents were presented with ten choice tasks, each containing three alternatives (two policy options and the status quo), as shown in the second section of the survey. After completing each choice set, participants answered follow-up questions. To ensure clarity and realism, the survey design emphasized tangible and relatable choice settings. Accordingly, graphics and pictorial representations were employed to convey alternatives more effectively than words or numbers alone.

While the use of visual images, cheap talk scripts, and realistic policy attributes enhances the comprehensibility and plausibility of the survey design, it is important to acknowledge the potential limitations of external validity. As with most stated preference methods, our hypothetical choice setting may not fully replicate actual decision-making behavior in real-world forest policy contexts. Respondents could be subject to hypothetical bias—overstating their willingness to pay in a survey setting compared to actual financial commitments—or social desirability bias, whereby answers reflect socially approved attitudes rather than true preferences. These limitations cannot be entirely eliminated; however, several measures were adopted to mitigate them. First, we incorporated visual representations of attribute levels and scenarios to increase realism and cognitive engagement. Second, a cheap talk script was provided to remind respondents of their budget constraints and the importance of realistic responses [63]. Third, we employed follow-up questions on consequentiality and protest responses to detect potential bias and to interpret responses with caution. Taken together, these approaches strengthen the validity of the results, but the findings should still be interpreted with awareness of the limitations inherent in hypothetical choice experiments.

For the design of the experimental choice set for our research, we used an efficient design method to design a total of 200 choice sets and iterated them through the design process to ensure that each attribute was as evenly distributed as possible, and the D-error of our design was 0.101, the A-error was 0.192, and the S-error was 9.67 [64]. In the DCE survey, we asked our respondents to choose between two improved alternatives (Alternatives A and B) and the current situation of the degraded forest (which stated zero-additional cost for no increase in restoration funding). Finally, 10 choice sets from among them were randomly assigned to each respondent (see

Table 2. An example of a choice set.

	Alternatives	Improved Situation		Current Situation
Attributes and Levels		Alternative 1	Alternative 2	Alternative 3 (Status Quo)
Forest fires		(Blue)	(Orange)	(Red)
Forest pest and diseases		(Orange)	(Blue)	(Red)
Restrictions on forest-related outdoor activities		(3~5 million restriction)	(Under 1 million restriction)	(No restrictions)
Biodiversity loss (no. of pine trees infected)		(−100 k~−200 k pine tree loss)	(−200 k~−300 k pine tree loss)	(−300 k~−400 k pine tree loss)
Forest restoration costs		(KRW 10,000)	(KRW 25,000)	0 KRW 0

Note: See Figure A1 (Ethics approval of the questionnaire survey).

).

3.2. Results

3.2.1. The Sample Demographics of the Respondents

A stratified random sampling strategy was employed to ensure that the survey reflected the demographic structure of Korean adults being aged 20 years and above. Fieldwork was carried out nationwide from August to September 2023 through face-to-face interviews conducted by a professional survey company (ST Innovation). Sampling quotas were applied across all major administrative regions to secure balanced representation in terms of gender, age, and income. The survey was administered using tablets, allowing for real-time data entry and quality control during the discrete choice experiment, which assessed public preferences for forest restoration under threats to ecosystem services and sustainable forest management (SFM).

Table A1. Demographic characteristics of the respondents (N = 1021).

Variables	Description of Variables	Frequencies (Ratio)	Mean	Standard Deviation
Gender	1. Male	452 (44.3%)	1.56	0.49
	2. Female	569 (55.7%)
Age	1. 20 ~ 29	178 (17.4%)	2.6	1.16
	2. 30 ~ 39	352 (34.5%)
	3. 40 ~ 49	269 (26.3%)
	4. 50 ~ 59	140 (13.7%)
	5. Older than 60	82 (8.0%)
Education	1. Primary school or below	3 (0.3%)	3.96	0.57
	2. Secondary school	7 (0.7%)
	3. High school	146 (14.3%)
	4. University	732 (71.7%)
	5. Postgraduate or above	133 (13.0%)
Employment status	1. Students	43 (4.2%)	2.13	1.59
	2. Employed full time	613 (60.0%)
	3. Housekeeper	173 (16.9%)
	4. Public officer	67 (6.6%)
	5. Unemployment	64 (6.3%)
	6. Others	61 (6.0%)
Monthly income level (in USD)	1. Below KRW 2,000,000 (USD 1538)	112 (11.0%)	3.23	1.34
	2. KRW 2,000,000~KRW 2,900,000 (USD 1538~USD 2231)	225 (22.0%)
	3. KRW 3,000,000~KRW 3,900,000 (USD 2308~USD 3000)	265 (26.0%)
	4. KRW 4,000,000~KRW 4,900,000 (USD 3077~USD 3769)	155 (15.2%)
	5. More than KRW 5,000,000 (USD 3846)	264 (25.9%)

summarizes the demographic characteristics of the 1021 participants, which yielded 10,210 valid choice responses. In total, 1021 individuals participated, producing 10,210 valid observations. Among them, 569 (55.7%) were female, while 621 (60.8%) were between the ages of 30 and 49. More than 70% of respondents held at least a bachelor’s degree, and approximately 60% were employed full time. Around 40% reported monthly household incomes above USD 3001. Although residents from metropolitan regions were slightly overrepresented, responses were obtained from all areas: Seoul (33.8%), Incheon/Gyeonggi (28.4%), Busan/Ulsan/Gyeongnam (13.7%), Daegu/Gyeongbuk (9.8%), Chungcheong (5.6%), Jeolla (5.0%), Gangwon (2.9%), and Jeju (0.8%).

3.2.2. Results from the PCA-Based Classification of Latent Environmental Attitudes and Behaviors

Following the IBM guidelines for PCA clustering [65], the results were interpreted to classify groups of environmental attitudes. Accurately identifying public attitudes and behaviors is essential for evaluating the acceptability of forest restoration policies given the close connection between environmental issues and the general population. To capture these attitudes among Korean citizens, 20 items were selected from Milfont et al. (2010) [55], and these were measured on a five-point Likert scale ranging from “strongly disagree” to “strongly agree” (see Appendix B,

Table A4. The respondents’ environmental attitude and behavior results.

1. 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 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
2. 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
3. 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 all right 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

). This section examines the reliability, validity, and model fit of the data, and it also further analyzes how the derived latent variables and classes influenced WTP for sustainable forest management policies through PCA and latent class choice modeling.

As an initial step, PCA was conducted on the responses to the environmental attitude and behavior items using IBM SPSS Statistics 28 (licensed). Prior to extracting factors, internal consistency and reliability were assessed. Cronbach’s alpha for the 20 items was 0.83, exceeding the commonly accepted threshold of 0.70. Sampling adequacy was confirmed by a Kaiser–Meyer–Olkin (KMO) value of 0.86, while Bartlett’s test of sphericity yielded a Chi-Square statistic of approximately 6251.30 (df = 190, p < 0.001). These results demonstrate that the dataset is well suited for PCA. Overall, the test statistics indicate that the structural model of environmental attitudes and behaviors provides a robust framework to explain both support for SFM and the actual WTP for forest ecosystem services.

According to Appendix A,

Table A2. The total variance of the environmental attitudes and behaviors.

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

, the problem seemed amenable to a five-factor solution as the 58% of variance explained decreased sharply in the five components; in addition, there was less than 1 eigenvalues and the factor loadings appeared in the columns for Factors 1~5. To summarize the main responses in

Table 3. The results of the environmental class ratio when subjected to PCA.

	Component 1	Component 2	Component 3	Component 4	Component 5
Question items	A3_6, A3_4, A2_2, A2_1, A3_5, A3_3	A1_3, A1_2, A1_4, A1_1	A2_4, A2_3, A2_5, A1_5	A1_7, A3_7, A1_8, A1_6	A3_1, A3_2
Ratio (N = 1021)	150 (14.7%)	138 (13.5%)	206 (20.2%)	281 (27.5%)	246 (24.1%)
Group characteristics	Human self-benefit- centered group	Environmental activity preferred group	Ecosystem sensitivity awareness group	Active environmental protection group	Passive environmental protection group

Note: The full list of 20 question items is provided in Appendix Table A4 and Table A6 (coding Q1–Q20 corresponds to the questionnaire items used in the PCA analysis); adapted from Jeon (2025) [58], Frontiers in Behavioral Economics, 4, 1596416. https://doi.org/10.3389/frbhe.2025.1596416, CC BY 4.0.

, the respondents (Component 1, 14.7%) tended to prioritize their convenience and economic pursuit even though they were aware of the environment and the importance of nature (called this the human self-benefit-centered group). On the other hand, for the respondents who the contribution of ecosystem services to stress relief, well-being, and happiness in nature (Component 2) was considered very important we categorized them as the environmental activity preferred group (13.5%). Third, the respondents of Component 3 recognized that the planet Earth and natural ecosystems are at risk were called the ecosystem sensitivity awareness group (20.2%). Fourthly, respondents of Component 4 showed that natural problems can be solved and mitigated by scientific and technological innovation and were named the active environmental protection group (27.5%). The items regarding saving electricity and controlling heating temperatures were also understood as another component (Component 5), and these items also reflected positive attitudes towards sustainable forest management (called the passive environmental protection group (24.1%)) [21,53,65,66].

Appendix B,

Table A5. The PCA results with IBM SPSS 24 (28.0.1.1).

Variables	Components
	1	2	3	4	5
A3_6	0.782	−0.018	−0.108	0.189	−0.148
A3_4	0.752	0.021	−0.155	0.095	−0.096
A2_2	0.656	0.130	0.017	−0.074	0.457
A2_1	0.610	0.109	0.084	−0.047	0.492
A3_5	0.593	0.033	0.259	0.022	0.026
A3_3	0.545	0.085	−0.195	0.360	−0.203
A1_3	0.025	0.758	0.296	0.078	0.064
A1_2	0.052	0.756	0.199	0.202	0.143
A1_4	0.087	0.727	0.033	0.188	0.004
A1_1	0.096	0.689	0.051	0.238	0.102
A2_4	−0.061	0.159	0.806	0.087	0.094
A2_3	−0.073	0.188	0.724	0.167	0.143
A2_5	0.078	0.104	0.655	0.273	0.079
A1_5	−0.089	0.461	0.462	−0.073	0.230
A1_7	0.098	0.164	0.168	0.725	0.090
A3_7	0.213	0.123	0.050	0.704	0.102
A1_8	−0.081	0.344	0.183	0.624	0.195
A1_6	0.092	0.170	0.330	0.485	0.113
A3_1	−0.090	0.096	0.187	0.148	0.698
A3_2	−0.017	0.158	0.140	0.315	0.689

Notes: Extraction method—principal component analysis; rotation method—varimax with Kaiser normalization, where rotation converged in six iterations. Boldface indicates the position of each component group.

and

Table A6. Class-specific respondent characteristics.

	Class 1 (n = 150)		Class 2 (n = 138)		Class 3 (n = 206)		Class 4 (n = 281)		Class 5 (n = 246)
	Mean	Standard Deviation	Mean	Standard Deviation	Mean	Standard Deviation	Mean	Standard Deviation	Mean	Standard Deviation
Socioeconomic background
Gender	1.52	0.50	1.51	0.50	1.46	0.49	1.58	0.49	1.67	0.47
Age	37.63	8.77	41.49	12.45	36.75	11.07	42.26	12.26	42.69	11.31
Income	3.11	1.20	3.03	1.33	3.33	1.36	3.30	1.34	3.26	1.38
Education	3.95	0.41	3.75	0.62	3.99	0.56	4.01	0.58	4.02	0.58
Environmental attitudes and behaviors
A1_1	4.26	0.54	3.12	0.70	2.85	1.16	3.58	0.67	2.39	1.73
A1_2	4.68	0.49	3.21	0.61	3.48	0.96	3.90	0.33	3.91	0.90
A1_3	4.56	0.52	3.27	0.66	3.58	0.97	3.97	0.45	4.37	0.58
A1_4	4.36	0.67	3.17	0.70	3.05	1.17	3.65	0.65	4.40	0.58
A1_5	4.47	0.60	3.40	0.63	3.90	1.06	4.05	0.51	3.96	0.89
A1_6	4.47	0.57	3.22	0.62	3.08	1.18	3.74	0.73	4.71	0.50
A1_7	4.49	0.55	3.03	0.71	2.97	1.16	3.63	0.76	4.04	1.04
A1_8	4.50	0.51	3.07	0.65	3.00	1.10	3.72	0.58	3.72	0.98
A2_1	4.23	0.69	3.29	0.55	3.36	1.01	3.55	0.72	4.16	0.70
A2_2	4.37	0.83	3.31	0.60	3.45	1.03	3.52	0.71	3.35	1.07
A2_3	4.46	0.68	3.30	0.60	3.74	1.04	4.00	0.53	3.28	1.07
A2_4	4.38	0.65	3.28	0.59	3.68	1.06	3.93	0.48	4.68	0.46
A2_5	4.50	0.57	3.19	0.57	3.34	1.05	3.82	0.53	4.62	0.54
A3_1	4.35	0.53	3.46	0.56	3.64	1.00	3.95	0.48	4.25	0.88
A3_2	4.58	0.55	3.49	0.60	3.48	1.00	3.89	0.47	4.37	0.73
A3_3	4.44	0.78	3.22	0.68	2.74	1.25	2.89	1.00	4.28	0.75
A3_4	4.11	0.98	3.18	0.56	3.19	0.95	3.18	0.79	2.11	1.08
A3_5	4.43	0.57	3.36	0.55	3.59	1.01	3.50	0.87	2.18	0.81
A3_6	4.36	0.71	3.21	0.54	3.04	1.09	2.98	0.88	3.24	1.44
A3_7	4.13	1.16	2.86	0.71	2.54	1.07	3.31	0.69	1.91	0.79
Follow-up questions: WTP resistance, attribute ranking, consequentiality, questionnaire relevant questions, others
D1_1	3.48	1.23	3.40	0.73	3.21	1.09	3.36	0.77	3.25	0.88
D1 _2	3.65	1.14	3.40	0.55	3.49	0.85	3.46	0.65	3.17	1.04
D1_3	3.13	1.35	2.97	0.72	2.87	1.03	2.69	0.88	3.59	0.70
D1_4	3.13	1.33	3.04	0.66	2.67	1.12	2.63	0.92	2.39	0.91
D1_5	3.23	1.26	3.03	0.74	3.18	1.05	2.91	0.89	2.22	0.98
D1_6	3.31	1.34	3.20	0.68	3.10	1.12	2.93	0.84	2.54	0.99
D1_7	3.50	1.17	3.40	0.73	3.54	0.99	3.43	0.84	2.74	1.08
D1_8	3.50	1.21	3.34	0.69	3.43	0.96	3.22	0.81	3.33	1.07
D1_9	2.79	1.44	2.93	0.66	2.73	1.10	2.43	0.86	3.03	1.05
D2_1	2.00	1.16	1.70	1.13	2.05	1.29	1.91	1.36	1.78	0.70
D2_2	2.65	1.19	2.75	1.13	2.54	1.22	2.68	1.13	1.98	1.38
D2_3	2.91	1.28	3.08	1.13	2.88	1.16	3.03	1.20	2.64	1.15
D2_4	3.44	1.19	3.30	1.20	3.30	1.19	3.30	1.13	2.91	1.21
D2_5	4.00	1.34	4.17	1.21	4.23	1.14	4.09	1.25	3.29	1.14
D3_1	4.11	0.68	3.18	0.65	3.04	1.00	3.48	0.66	4.18	1.15
D3_2	4.19	0.72	3.10	0.62	3.08	0.99	3.36	0.73	3.35	0.89
D4_1	4.16	0.68	3.28	0.63	3.47	0.94	3.52	0.60	3.34	0.93
D4_2	4.25	0.77	3.34	0.64	3.52	0.79	3.56	0.62	3.71	0.73
D4_3	4.24	0.77	3.30	0.63	3.42	0.86	3.63	0.65	3.79	0.69
D4_4	4.04	0.86	3.26	0.68	3.33	0.98	3.34	0.75	3.88	0.78
D4_5	4.28	0.72	3.34	0.58	3.33	0.93	3.52	0.63	3.40	0.98

presents the background characteristics of the respondents in the five latent classes, and it also summarizes the mean and standard deviation values of the socioeconomic variables, environmental attitudes and behaviors, and the follow-up questions for each of the five classes.

3.2.3. Results of Classifying Latent Environmental Groups Using LCM Based on Environmental Attitudes and Behaviors

By applying the LCMs, we aimed to address the limitation in logit models, which typically rely on a strong IID (independent and identically distributed) assumption. This assumption constrains the ability to adequately explain variations in preferences. If only this basic model is applied to environmental issues where various stakeholders are involved, the results and models lead to inaccurate inferences and environmental policies. It is, therefore, essential to analyze their diverse environmental preferences in greater detail and to develop and implement policies that align with their latent preferences using LCMs.

In this study, the probabilistic latent class model was applied to identify the five latent classes of the groups (log-likelihood −10,877.06) (AIC 21,766.13) at Class I model→−9580.23 (AIC 19,186.46) at Class II model→−9412.63 (AIC 18,865.28) at Class III model→−9173.36 (AIC 18,400.74) at Class IV model→−9079.56 (AIC 18,277.13) at Class V model, as shown in

Table 4. The results of the attribute estimates by latent class when considering heterogeneity in environmental preference using LCM.

		Model I		Model II		Model III		Model IV		Model V
	Attributes	Estimate	t-Value	Estimate	t-Value	Estimate	t-Value	Estimate	t-Value	Estimate	t-Value
Latent Class 1	Forest fire.class1	0.231	2.065 **	1.756	4.887 ***	1.838	6.180 ***	0.681	3.145 ***	0.049	0.245
	Forest pests and diseases.class1	0.159	1.318	0.579	1.583	0.598	1.696 *	0.414	1.618	−0.131	−0.584
	Forest access restrictions.class1	−0.001	-0.100	0.032	0.986	0.046	1.356	0.008	0.327	−0.013	−0.640
	Forest Biodiversity Loss.class1	0.001	0.106	0.002	0.038	0.010	0.265	0.023	0.787	0.023	0.871
	Cost (forest restoration fund).class1	−0.017	−16.371 ***	−0.028	−8.859 ***	−0.021	−8.809 ***	−0.012	−6.347 ***	0.002	0.855
	SQ_asc.class1	−0.978	−18.888 ***	0.337	2.288 **	0.569	4.785 ***	−0.356	−3.485 ***	−3.323	−14.693 ***
Latent Class 2	Forest fire.class2	-	-	−0.083	−0.636	−0.054	−0.370	−0.060	−0.374	−1.106	−3.008 ***
	Forest pest and disseas.class2	-	-	0.059	0.414	−0.017	−0.109	−0.087	−0.500	0.143	0.243
	Forest access restrictions.class2	-	-	−0.012	−0.934	−0.007	−0.450	−0.012	−0.714	−0.030	−0.606
	Forest Biodiversity Loss.class2	-	-	−0.004	−0.226	0.021	1.112	0.006	0.271	−0.180	−2.650 ***
	Cost (forest restoration fund).class2	-	-	−0.017	−13.291 ***	0.000	0.185	−0.003	−1.623	−0.097	−14.813 ***
	SQ_asc.class2	-	-	−2.195	−25.873 ***	−1.854	−24.012 ***	−3.294	−19.979 ***	−4.859	−17.722 ***
Latent Class 3	Forest fire.class3	-	-	-	-	−0.466	−1.203	4.284	4.685 ***	0.203	0.179
	Forest pests and diseases.class3	-	-	-	-	0.478	1.010	−0.617	−0.663	0.935	1.111
	Forest access restrictions.class3	-	-	-	-	−0.064	−1.652 *	0.098	1.040	−0.153	−1.808 *
	Forest Biodiversity Loss.class3	-	-	-	-	−0.082	−1.554	0.037	0.334	−0.118	−1.315
	Cost (forest restoration fund).class3	-	-	-	-	−0.101	−16.722 ***	−0.047	−5.525 ***	−0.176	−7.831 ***
	SQ_asc.class3	-	-	-	-	−4.014	−18.200 ***	1.465	4.805 ***	−3.822	−7.772 ***
Latent Class 4	Forest fire.class4	-	-	-	-	-	-	−0.949	−1.949 *	0.573	2.301 **
	Forest pests and diseases.class4	-	-	-	-	-	-	0.418	0.726	0.325	1.289
	Forest access restrictions.class4	-	-	-	-	-	-	−0.055	−1.193	0.010	0.398
	Forest Biodiversity Loss.class4	-	-	-	-	-	-	−0.106	−1.572	0.034	1.108
	Cost (forest restoration fund).class4	-	-	-	-	-	-	−0.120	−13.881 ***	−0.004	−1.612
	SQ_asc.class4	-	-	-	-	-	-	−4.729	−15.728 ***	−0.246	−1.956 *
Latent Class 5	Forest fire.class5	-	-	-	-	-	-	-	-	4.381	4.325 ***
	Forest pests and diseases.class5	-	-	-	-	-	-	-	-	-0.832	−0.955
	Forest access restrictions.class5	-	-	-	-	-	-	-	-	0.060	0.658
	Forest Biodiversity Loss.class5	-	-	-	-	-	-	-	-	0.030	0.262
	Cost (forest restoration fund).class5	-	-	-	-	-	-	-	-	−0.037	−3.780 ***
	SQ_asc.class5	-	-	-	-	-	-	-	-	1.688	4.466 ***
Class membership probability	Constant.class2	-	-	0.975	11.071 ***	0.745	8.809 ***	0.351	3.936 ***	−0.865	−6.101 ***
	Constant.class3	-	-	-	-	−0.290	−2.305 **	−0.978	−8.034 ***	−.782	−9.643 ***
	Constant.class4	-	-	-	-	-	-	−0.729	−5.025 ***	−0.278	−2.697 ***
	Constant.class5	-	-	-	-	-	-	-	-	−1.194	−10.069 ***
Model fit criteria (Log-Likelihood, AIC)		−10,877.06, 21,766.13		−9580.23, 19,186.46		−9412.63, 18,865.28		−9173.36, 18,400.74		−9079.56, 18,227.13

***: p < 0.001, **: p < 0.05, and *: p < 0.1.

. Additionally, the results show that, although there was some overlap in the environmental behavior among the groups, there was a strong tendency for the groups to have different environmental attitudes and preference heterogeneities. This was only possible to establish through the analysis of these LCMs [30,47,67]. In addition, the statistically significant variables in the table were different for each class and attribute and were important factors that distinguished the characteristics of the groups. This is consistent with the results of the previous analysis of environmental attitudes when subjected to PCA. Therefore, it is necessary to establish forest restoration policies that reflect environmental preferences as accurately as possible in terms of degraded forest restoration policies, forest recreation use patterns, forest biodiversity, and environmental threats.

The results of our estimated LCM models were as follows. In addition to estimating the best fit for the LCMs for our discrete choice data based on the information criteria loglikelihood values, we also found that the best goodness-of-fit measures of the model estimation were obtained when there were five types of LCMs (Table 4). This is relatively consistent with the results of the previous analysis of environmental attitude data when using PCA, as discussed in Section 3.2.2. We based our findings on the five LCMs, focusing primarily on the statistically significant attributes in each model. Before presenting the results, it is important to highlight that the class membership coefficients for the five latent environmental classes (constant.class2~class5) were all highly statistically significant. This indicates that the latent characteristics of the five groups were clearly present among our research subjects, aligning with the previous PCA findings and strongly supporting the existence of heterogeneous environmental preferences [68,69].

First, from an estimated consequence of LCMs (the first latent class in Class I), all attributes showed appropriate signs, as was expected in our research hypothesis. Since the first SQ_asc in Class I was negative as the coefficient of the status quo, it represented a situation where the degraded and deforested forests stayed as such, which is a scenario that is negative for our utility. On the other hand, this also addresses the fact that the restoration of degraded forests seems to positively affect the utility. Regarding the results of the remaining main attribute variables, the cost attribute forest restoration fund was negative and was very statistically significant at the 99% level, which means that, as the cost values increased, the probability decreased. Next, the attribute coefficient of forest fires showed a positive sign, which means the respondents had a WTP to lower the fire levels from the red level (status quo) to the blue safe level. The coefficient of the forest pest and disease attribute was also represented as positive but statistically insignificant, which means that it represented a positive WTP to release environmental threat. Even if it was statistically insignificant, the coefficient of forest outdoor-related restriction was negative, which means that the effect of our visitors’ restrictions in forests was negative on our utility. Lastly, a positive WTP for reducing forest biodiversity loss was estimated. The Class V model, with a log-likelihood value of −9079.56, outperformed expectations; thus, we selected the Class V model to explain the latent environmental preferences, to analyze behavior patterns, and to estimate the final WTP.

In this V model, the respondent results for Latent Class 1 showed a very interesting feature: only the ASC coefficient, which indicates the current degradation of the forest, was statistically significant, while the other main attribute variables were statistically insignificant. This can be interpreted as the respondents in this group saying, ‘I dislike the current degraded forest, and I don’t care about the other important attributes’. Hence, this group was characterized as being cost-insensitive (an insignificant positive coefficient for the cost attribute) [14] and not caring about other attributes; thus, they were labeled ‘disengaged’. In the next class, Latent Class 2, the coefficient of both forest fire and biodiversity loss attributes was paradoxically estimated with the opposite sign to that expected and statistically significant at 5% and 10%. This means that we needed to explain the tendency to disfavor improvement of the degraded forests in environmental threats. This can be summarized as a group that tends to be resistant or opposed to paying forest restoration funds as a tax instrument. According to the follow-up questions, 24% of the respondents chose the SQ alternative, which was due to WTP resistance (with the main reasons being resistance to the creation of a forest restoration fund, the burden of paying taxes, and government accountability for environmental problems). Therefore, this class was labeled a paradoxical group with cost sensitivity. In the third latent group, Latent Class 3, only the restriction on outdoor-related activities attribute (restrn.class3) had a statistically significant and negative value. This means that the restriction of the activities by forest management policy had a negative impact on our utility. This result suggests that they are a group that is relatively sensitive to and enjoy being in the forests. Therefore, this class was categorized as the forest recreational active group. In Latent Class 4, this group considered only the coefficient of the forest fire attribute, which was statistically significant at 5%, while the coefficient of cost was insensitive, with the appropriate expected sign shown. This group was named the forest policy passive group because they did not exhibit any distinct group characteristics. The last latent class, Latent Class 5, was aware that the attribute variable (SQ_ASC.class 5) was positive, which means that the currently degraded forest condition was preferred. Given the sign and magnitude of these coefficients in Group 5, this suggests that they were sensitive to the cost of forest disasters and showed WTP enough to mitigate for forest fire risks, but the choice of the status quo reflected a strong resistance to paying a forest restoration fund tax instrument. Therefore, this was a potential group where forest restoration costs and forest fire shocks had a large impact on our utility but also reflected resistance to paying taxes and to government environmental policies. This group was called the cost-sensitive resistor group.

Appendix A,

Table A3. Marginal WTP estimation result by each type of class model (I~V) using LCMs.

	Attribute	Model I	Model II	Model III	Model IV	Model V
Latent Class 1	Forest fire.class1	KRW 13,431.31 ** (USD 10.3)	KRW 61,933.74 *** (USD 47.4)	KRW 86,045.00 *** (USD 65.9)	KRW 54,939.13 *** (USD 42.1)	KRW −23,576.50 (USD −18.1)
	Forest pests and diseases.class1	KRW 9204.96 (USD 7.0)	KRW 20,441.79 (USD 15.7)	KRW 27,981.10 * (USD 21.4)	KRW 33,365.09 (USD 25.6)	KRW 62,244.18 (USD 47.7)
	Forest access restrictions.class1	KRW −66.86 (USD −0.1)	KRW 1126.24 (USD 0.9)	KRW 2144.13 (USD 1.6)	KRW 668.24 (USD 0.5)	KRW 6272.59 (USD 4.8)
	Forest Biodiversity loss.class1	KRW 86.52 (USD 0.1)	KRW 53.87 (USD 0.0)	KRW 487.49 (USD 0.4)	KRW 1855.59 (USD 1.4)	KRW −10,773.68 (USD −8.3)
Latent Class 2	Forest fire.class2	-	KRW −4790.48 (USD -3.7)	KRW 171,195.40 (USD 131.1)	KRW −21,250.62 (USD −16.3)	KRW −11,412.56 *** (USD −8.7)
	Forest pest and disease.class2	-	KRW 3403.32 (USD 2.6)	KRW 55,180.60 (USD 42.3)	KRW −30,880.53 (USD −23.6)	KRW 1479.51 (USD 1.1)
	Forest access restrictions.class2	-	KRW −723.44 (USD -0.6)	KRW 22,304.35 (USD 17.1)	KRW −4322.33 (USD −3.3)	KRW −311.10 (USD −0.2)
	Forest Biodiversity loss.class2	-	KRW −215.01 (USD −0.2)	KRW −65,910.20 (USD −50.5)	KRW 1956.75 (USD 1.5)	KRW −1859.05 (USD −1.4)
Latent Class 3	Forest fire.class3	-	-	KRW −4621.50 (USD −3.5)	KRW 90,643.93 *** (USD 69.4)	KRW 1154.55 (USD 0.9)
	Forest pests and diseases.class3	-	-	KRW 4740.00 (USD 3.6)	KRW −13,063.63 (USD −10.0)	KRW 5318.51 (USD 4.1)
	Forest access restrictions.class3	-	-	KRW −632.18 * (USD −0.5)	KRW 2066.44 (USD 1.6)	KRW −871.41 * (USD −0.7)
	Forest Biodiversity loss.class3	-	-	KRW −816.95 (USD −0.6)	KRW 785.07 (USD 0.6)	KRW −670.35 (USD 0.5)
Latent Class 4	Forest fire.class4	-	-	-	KRW −7933.60 * (USD −6.1)	KRW 135,919.37 ** (USD 104.1)
	Forest pests and diseases.class4	-	-	-	KRW 3497.82 (USD 2.7)	KRW 77,053.85 (USD 59.0)
	Forest access restrictions.class4	-	-	-	KRW −458.98 (USD −0.4)	KRW 2311.65 (USD 1.8)
	Forest Biodiversity loss.class4	-	-	-	KRW −888.75 (USD −0.7)	KRW 7976.62 (USD 6.1)
Latent Class 5	Forest fire.class5	-	-	-	-	KRW 119,550.87 *** (USD 91.6)
	Forest pests and diseases.class5	-	-	-	-	KRW −22,697.46 (USD −17.4)
	Forest access restrictions.class5	-	-	-	-	KRW 1625.52 (USD 1.2)
	Forest Biodiversity loss.class5	-	-	-	-	KRW 809.03 (USD 0.6)

***: p < 0.001, **: p < 0.05, and *: p < 0.1.

shows the details regarding marginal WTP from LCM, and we established five types of LCMs (model I~V) to analyze the economic value of the environmental attributes. We used LCM because we wanted to estimate the economic value of each attribute in our dataset using the best-fit models. In general, if all respondents had the same environmental behaviors, then the coefficients of an attribute tended to be the same WTP in each class and matched the expected sign of the researcher. However, environmental preference and behaviors are separated by each class, so when these environmental behaviors were divided into different groups, the sign of each attribute coefficient was characterized by a different sign. In other words, the sign of the attribute and the magnitude of the coefficient were logically explained and inferred according to the characteristics, environmental attitudes, and environmental preferences of the group. Looking at the WTP magnitudes of the statistically significant coefficients in each latent class, first, the coefficient of forest fire attribute was statistically significant across the board, with magnitudes estimated to range from KRW 13,431.31 (USD 10.29)~KRW 135,919 (USD 104.09). On the other hand, the negative values of the forest fire attribute were distributed from KRW -7933.60 (USD −6.08) (fourth latent class in the Class IV model) to KRW −11,412.56 (USD 8.74) (second latent class in Class V model). For the forest pests and disease attribute, it was estimated to be KRW 27,981.1 (USD 21.43) (fifth latent class of the Class I model). Next, the forest outdoor activity restriction attribute was only significant in the third latent class of the Class III model, with an estimated value of KRW −632.18 (USD −0.48). Finally, for the economic WTP of the forest biodiversity attribute, no class or latent class showed a statistically significant value. From these results, we can interpret and use the coefficient values that were prominent in each class of models to infer the characteristics of the latent class group.

In this section, we synthesize the results of applying LCM and PCA to our experimental choice data to obtain common probability from each PCA and Model V, from which we can estimate the WTP for the main attributes, as shown in Appendix A, Table A3. The results of estimating the WTP by the attributes using the results in

Table 5. Probability results of the PCA and LCM classes.

Classification	Latent Classes
	LCM 1	LCM 2	LCM 3	LCM 4	LCM 5
PCA1	0.0	1.0	41.7	53.2	4.1
PCA2	16.6	43.4	13.9	25.4	0.6
PCA3	19.9	0.1	37.4	42.7	0.0
PCA4	0.0	94.0	0.0	4.0	2.1
PCA5	0.0	45.4	33.4	0.4	20.8
Average	7.3	36.8	25.3	25.1	5.5

and the marginal WTP in Appendix A, Table A3 that is, the WTP for each attribute was estimated using a latent class model (the class-specific WTP values are also presented in

Table 6. WTP estimation results by class.

Attributes	Class 1	Class 2	Class 3	Class 4	Class 5	WTP
Forest fire	KRW−23,576.5 (USD −18.06)	KRW−11,412.6 (USD −8.74)	KRW 1154.6 (USD 0.88)	KRW 135,919.4 (USD 104.09)	KRW 119,550.9 (USD 91.55)	KRW 77,529.2 (USD 59.37)
Forest pests and diseases	KRW 62,244.2 (USD 47.67)	KRW 1479.5 (USD 1.13)	KRW 5318.5 (USD 4.07)	KRW 77,053.9 (USD 59.01)	KRW −22,697.5 (USD −17.38)	KRW 31,165.0 (USD 23.87)
Forest access restrictions	KRW 6272.6 (USD 4.80)	KRW −311.1 (USD -0.24)	KRW −871.4 (USD -0.67)	KRW 2311.7 (USD 1.77)	KRW 1625.5 (USD 1.24)	KRW 1905.7 (USD 1.46)
Forest biodiversity loss	KRW −10,773.7 (USD −8.25)	KRW −1859.1 (USD −1.42)	KRW −670.4 (USD −0.51)	KRW 7976.6 (USD 6.11)	KRW 809.0 (USD 0.62)	KRW −1414.5 (USD −1.08)
WTP	KRW 34,166.6 (USD 26.17)	KRW −12,103.2 (USD −9.27)	KRW 4931.3 (USD 3.78)	KRW 223,261.5 (USD 170.98)	KRW 99,288.0 (USD 76.04)	KRW 109,185.5 (USD 83.62)

Note: we calculated the confidence interval at 95% of the WTP and gained almost the same value at two decimal places.

). The average WTP for each attribute was calculated by taking the weighted average of the WTP values from each class based on the class probabilities (as shown in Table 5). The average WTP for each attribute was calculated using the following procedure. First, the class-specific WTP values were multiplied by the probability of belonging to each class, and then these values were summed. For example, the average WTP for the forest fire attribute was calculated as follows: $Average WTP=\sum _{i=1}^n({WTP}_i\ast P_i)$, where ${WTP}_i$ represents the WTP value for class $i$ and $P_i$ is the probability of being in class $i$. For example, the average WTP for the forest fire attribute was calculated as follows: average WTP fire = [KRW −23,576.5 * 0.0% + (KRW −11,412.6 * 1.0%) + KRW 1154.6 * 41.7% + KRW 135,919 * 53.2% + KRW 119,550.9 * 4.1%]. This resulted in an average WTP of KRW 77,529.22. This version maintains a formal tone and provides a clear explanation of the process, making it more appropriate for inclusion in a research paper. The WTP for the forest fire risk attribute was KRW 77,529.2 (USD 59.37), for the forest pests and diseases it was KRW 31,165.1 (USD 23.87), for the forest access restriction attribute it was KRW 1905.7 (USD 1.46), and for the forest biodiversity loss attribute it was KRW −1414.5 (USD −1.08). Considering the magnitude of WTP, it can be concluded that the public considers the forest fire and forest pest attributes to be important in restoring degraded forests and for providing sustainable forest ecosystem services. On the other hand, the remaining two attributes forest access restriction and forest biodiversity loss seem to have less economic significance.

3.2.4. Simulation Results for Forest Restoration Policy Acceptance Using LCM

To evaluate the implications of the estimated LCM, we simulated the overall acceptability of forest restoration policies under a realistic scenario. The estimated utility coefficients for the five latent classes—each with six choice parameters—and four class membership constants were applied. In this simulation, the policy alternative was defined as a forest restoration program with the most ecologically favorable attributes allowed under real-world constraints. Specifically, the scenario assumed the lowest fire risk (Level 4), the lowest pest and disease risk (Level 4), the minimal biodiversity loss (0), and the minimal but necessary access restriction (KRW 1 million). The restoration cost attribute was set at KRW 20,000, reflecting a reasonable financial burden for respondents. The alternative-specific constant (ASC) was set to 1 to represent the policy alternative, while the status quo was assigned ASC = 0 with the worst environmental conditions and zero cost (

Table 7. Class-specific policy acceptance under the realistic scenario.

Classes	$\mathbf{Class} \mathbf{Probability} ({\mathit{\pi }}_{\mathit{c}}$)	$\mathbf{Policy} \mathbf{Choice} \mathbf{Probability} ({\mathit{P}}_{\mathit{c}}$)	$\mathbf{Weighted} \mathbf{Probability} ({\mathit{\pi }}_{\mathit{c}}$$\times {\mathit{P}}_{\mathit{c}}$
Class 1	37.74%	3.48%	1.31%
Class 2	15.89%	0.06%	0.01%
Class 3	6.35%	0.01%	≈0.00%
Class 4	28.58%	40.12%	11.47%
Class 5	11.44%	66.66%	7.62%

Notes: All other attributes were fixed to represent an environmentally optimal scenario with minimum required access restriction. And Table 7 presents the results obtained from the author’s own simulation experiment.

).

To calculate the overall acceptability of this scenario, we applied a weighted probability aggregation method using the class membership shares and the class-specific choice probabilities. The formula is expressed as follows:

$$P_{total}={\sum }_{c=1}^5{\pi }_c·P_c,$$

where

-

$P_{total}$ is the total policy acceptance rate across the population;

-

${\pi }_c$ is the probability of belonging to Latent Class C;

-

$P_c$ is the multinomial logit-based choice probability for the policy in Class C.

Each $P_c$ is a conditional probability within its respective class; therefore, the sum $\sum P_c$ can exceed 1. The actual overall acceptance must be computed as a weighted average using the posterior class membership probabilities ${\pi }_c$, ensuring that the final total acceptance rate remains between 0 and 1.

These results show that the total weighted and predicted policy acceptance rate under the given scenario was approximately 20.4%. Most of the support stemmed from Class 4 and Class 5, who exhibited higher sensitivity to positive ecological outcomes and were relatively tolerant to moderate restoration costs. In contrast, Classes 2 and 3, which show strong cost aversion, contributed minimally to the overall acceptance. These findings suggest that balancing ecological restoration benefits with moderate financial and behavioral burdens is key to maximizing public support for forest restoration policies.

In addition, to evaluate the policy implications of the estimated LCM, we conducted a simulation of the overall policy acceptance rate under various forest restoration cost scenarios. In this exercise, we applied the estimated coefficients for the five latent classes (each with six utility parameters and four class membership constants) obtained from our LCM. The goal was to identify a realistic policy alternative that maximizes social acceptance while maintaining necessary constraints for implementation. We fixed certain attributes to represent an ideal forest restoration program under real-world constraints. Specifically, the fire risk level was set at the safest level (Level 4: Blue), pest and disease impacts were minimized (Level 4), forest access restriction was fixed at the lowest feasible level (KRW 1 million, representing minimal but necessary restrictions), and biodiversity loss was assumed to be negligible. The alternative-specific constant (ASC) was set to 1 to indicate the active policy alternative, while the status quo alternative was represented with the worst-case environmental conditions and zero cost. The cost attribute varied systematically from KRW 10,000 to KRW 50,000 in increments of KRW 5000. For each cost level, we computed the utility for the policy and status quo alternatives for each latent class, calculated the class-specific choice probabilities using the multinomial logit formula, and then aggregated these using the class membership probabilities. This provided the expected overall policy acceptance rate for each cost level.

The simulation results, as shown in Figure 1 (Left), reveal that the policy acceptance rate decreased monotonically with increasing costs. At KRW 10,000, the acceptance rate was approximately 40%, while at KRW 50,000, it dropped to below 30%. This highlighted the sensitivity of the population—particularly cost-averse latent classes—to even modest increases in forest restoration fund contributions. The optimal policy design under realistic constraints, therefore, lies in the cost range of KRW 10,000 to KRW 20,000, where acceptance is maximized without compromising ecological feasibility. And the results in Figure 1 (right) present the density of the individual WTP, indicating that WTP declines as the payment amount increases, regardless of class membership.

4. Discussion

This study investigated the heterogeneity in public preferences [27,70,71,72] and the acceptability of forest restoration policies in Korea, aiming to support the efficient design and implementation of strategies for restoring degraded forests and ecosystems. A choice experiment survey was conducted with 1021 Korean citizens aged 20 and older, yielding 10,210 individual choice observations. Based on this dataset, five distinct latent classes were identified using latent class modeling (LCM) and principal component analysis (PCA) through such a segmentation approach (latent environmental classes): an acceptance group of environmental policies, a support group for environmental policy, a resistance group to environmental policy, an indifferent group to environmental policies, and a group for seeking environmental technological solutions to environmental problems.

As summarized in Table 1, the attributes included restoration cost, biodiversity loss, restrictions on forest access, forest fires, and pest outbreaks. The LCM results enabled estimation of marginal willingness to pay (WTP) for each class and showed that, even for identical attributes, preference directions and WTP values differed substantially across classes. For instance, some groups displayed strong sensitivity to restoration cost, while others prioritized biodiversity loss or access restrictions. This pattern highlights substantial variation in the environmental values and attitudes among the public, and it is consistent with broader evidence from environmental choice experiments in Europe and North America [20,73,74]. Rather than generalizing such results, they are better understood as suggestive trends that may highlight directions of potential interest for future research, but they also lack sufficient robustness for direct policy application [75].

To address this heterogeneity, restoration strategies should reflect the environmental preferences of different segments of the population. For instance, highly engaged citizens could be mobilized through participatory programs, such as tree planting and environmental campaigns, while less engaged groups may benefit more from awareness-raising efforts and targeted education. These approaches align with previous studies showing that tailoring engagement strategies to audience characteristics can increase the effectiveness of forest conservation programs [76,77]. Given that wildfires and pest outbreaks are often linked to human behavior, preventive education and proactive communication are essential, particularly during high-risk periods.

Our findings provide a richer explanation for degraded forest restoration management than the general assumption of homogeneous environmental preferences. The empirical application of an integrated latent choice model in the management of forest restoration after degradation due to environmental threats greatly contributes to our understanding of restoration choice behavior. This approach is consistent with studies advocating for hybrid or integrated models to capture complex behavioral drivers in environmental valuation [15,78]. Such a model would also be useful to apply to other complex environmental issues involving multiple stakeholders across diverse contexts.

Simulations of policy acceptance under varying cost levels, directly linked to the restoration cost attribute in Table 1 and the experimental structure illustrated in Table 2, revealed that acceptance exceeded 40% for policies costing less than KRW 10,000, which dropped to approximately 30% for costs between KRW 10,000 and KRW 35,000 and fell again to around 25% for costs of KRW 50,000. These findings mirror patterns in other stated preference studies, where public acceptance of environmental levies declines steeply beyond low-to-moderate payment levels [79]. This underscores the importance of setting realistic cost thresholds to ensure broad public support.

Finally, given that most wildfires in Korea are caused by human negligence, forest fire prevention strategies may benefit from the integration of behavioral economics approaches. Methods such as nudging, social norm messaging, emotional framing, and default settings could help influence public behavior and promote responsible forest use. This suggestion aligns with recent experimental studies demonstrating the effectiveness of behavioral interventions in encouraging pro-environmental behaviors [80,81]. Incorporating these insights may offer an innovative and effective direction for both wildfire prevention and broader forest policy development in Korea.

Our policy recommendations thus suggest tailoring restoration strategies to the different preference segments identified in this study and the attributes, as summarized in Table 1, ranging from participatory programs for highly engaged citizens to awareness-raising campaigns for less engaged groups. However, the feasibility of these strategies is contingent upon financial and institutional constraints. For example, local governments may face budgetary limitations in allocating resources for restoration and education programs, while the Korea Forest Service may encounter jurisdictional or administrative barriers in expanding policy implementation. Moreover, successful enforcement of restoration measures requires the cooperation of local residents, whose support cannot be assumed without appropriate incentives or trust-building mechanisms. Therefore, while our recommendations offer differentiated strategies, their implementation should be carefully evaluated within the broader financial and institutional context to ensure both effectiveness and cost-efficiency.

5. Conclusions

This study, which considered the involvement of diverse stakeholders and the localized and variable nature of forest disasters, demonstrates the importance of accounting for latent environmental preferences and policy acceptance when designing forest restoration policies. By identifying distinct preference groups and their willingness to pay, the findings offer valuable insights for tailoring policies to improve acceptance and effectiveness. There is considerable preference heterogeneity with a group of latent stakeholders assigning relatively higher importance to the factors they perceive. This heterogeneity underlines the need to measure a wide range of outcome dimensions when evaluating forest restoration policy, including latent environmental classes. We conclude that multidimensional measures have further informed our understanding of environmental risks and ecosystem services.

While this study provides useful policy implications, it also has several limitations. First, multicollinearity among attributes may influence the estimation results, and future studies should consider strategies to address this issue. Second, increasing the sample size would enhance the robustness of simulations and enable more detailed subgroup analysis. Differences in the preferences between urban and rural residents or between those with and without experience of forest-related disasters warrant further investigation. We will analyze how the attribute coefficients change and how the overall explanatory power of the model can be improved by simultaneously adding environmental attitudes and socioeconomic variables in a future paper using the hybrid choice model [68]. In addition, considering forest fire and pests and disease attributes, the impact on the utility may be similar, in which case we want to test which of the two attributes respondents are more responsive to, or vice versa (attribute of non-attendance) [58].

In addition, the findings are context-specific and reflect the Korean case, which limits the generalizability of the results to other cultural and institutional settings. The sample also slightly overrepresents respondents from metropolitan areas, which may bias estimates of preference heterogeneity by under- or over-emphasizing certain regional attitudes. These limitations highlight the need for caution when extrapolating our results and suggest directions for future research. In particular, cross-country comparative studies and more balanced sampling strategies will be important to validate and extend the findings.

Overall, our results suggest that differentiated strategies are needed to align restoration costs with public acceptance levels and to design engagement programs that reflect heterogeneous environmental preferences. These policy implications complement the limitations noted above and point toward practical avenues for more effective forest governance.