*3.2. Estimating Conditional Logit Model*

Table 3 shows the estimation results for the conditional logit model. Model I is a basic model to which the attributes of the public functions—DM, FC, WM, and financial support—are solely assigned. Model II is an extended model with additional demographic variables because individual characteristics need to be used as control variables based on previous research that found that socioeconomic factors may influence the value estimates [69]. Thus, after analyzing 3618 observed data in both models, the basic model showed acceptable results. LLF was −3515.53 (*p* < 0.001), and the Pseudo R-squared (The Pseudo R-squared statistic, which provides an auxiliary explanation for the model fit, is not high. It is, however, preferable to highlight the figure because it tends to be lower than usual regression analysis. For instance, according to Brau [70], the 0.11 level is acceptable), was approximately 11.0%. Moreover, the price variable (*Bid*) was negatively effective at the 1% significance level, which satisfies the theoretical direction of the coefficient. All levels of the attribute variables (The levels named 'Mid' and 'High' of the three attributes indicate 'the change from the low to medium level' and 'the change from the low level to high level', respectively. Thus, those variables were coded as (1, 0) and (0, 1), where the low level signifies the reference alternative (0, 0)) have direct impacts at the 1% significance level (except *FCMid* significant at the 5% level), which implies that the higher the attribute level, the greater the probability of choosing the proposed options compared to the status quo. That is, the enhancement of each function for the dam can increase its utility for local people.


**Table 3.** Estimates of conditional logit models.

Note (1) Model I ⇒ Model II: <sup>χ</sup><sup>2</sup> (0.05, 7) = 18.44 > 14.07; Note (2) \*\*\*, \*\*, \*: Significance at the 1%, 5%, 10% levels, respectively.

Moreover, seven interaction variables between the alternative specific constant (*ASC*) and the demographic variables (Age, Income per household (unit: million KRW), and Education (years of education) are the continuous variables. Gender (with female = 0, male = 1), Marital status (with single = 0, married = 1), Occupation (unemployed = 0, employed =1), and residential area (Daejeon = 0, Chungnam and Chumgbuk = 1) are

dummy variables) were computed to identify other influencers on the choice probability which cannot be examined by the underlying attributes. Consequently, an extended model fitness was achieved as *LLF* (−3506.31) and *Pseudo-R*<sup>2</sup> (0.113) improved compared to the basic model. The age (*t* = −2.50) and income (*t* = 2.46) variables had a negative and positive influence at the 1% significance level. This implies that lower age and higher income levels mean more choice possibilities for improvement alternatives. However, performing the *Hausman* tests for the independence of irrelevant alternatives meant all *p*-values rejected the null hypothesis that parameter estimates are heterogeneous at the 5% significance level, confirming the independence of irrelevant alternatives (IIA) assumption regarding the independence of error terms had been fulfilled.

#### *3.3. Measuring Climate Change Perceptions and Segmenting Respondents*

The analysis revealed that three factors—cause, countermeasures, and effect—were derived in terms of the level of understanding, as in the theoretical composition. Moreover, the levels of awareness and practice are each a single factor. Bartlett's test of sphericity and the Kaiser–Meyer–Olkin values were statistically significant at the 0.01% level (Appendix A). In addition, the reliability tests showed *Cronbach α* values for all EFA factors to be more than 0.77—the level of understanding about the cause (*α* = 0.776), countermeasures (*α* = 0.795), the effect (*α* = 0.808), and the level of awareness (*α* = 0.817)—except for the level of practice with an α value of 0.667, albeit close to 0.7 [71].

Subsequently, a two-step clustering analysis utilizing five such factors was conducted to distinguish the climate change perception segments from the whole group (see Table 4). Two clusters were derived. Due to an independent-samples *t*-test to reveal the features of the two clusters, it was classified into segments of high levels (*H*) and low levels (*L*) regarding the five factors. The "*H*" and "*L*" clusters were, respectively named as "high involvement" and "low involvement."


**Table 4.** Clustering and *t*-test according to climate change awareness.

Note (1) Statistical mean difference: L < H. Note (2) \*\*\*: Significance at 1% level.

#### *3.4. Estimating Implicit Prices by Cluster*

As shown in Table 5, the coefficients for MWTP calculation of each group were estimated based on the extended model. The conditional logit models of the two groups were compared based on the likelihood ratio test [72,73] to clarify the moderating effect according to climate change perceptions. First, the likelihood ratio test between the two models showed that χ<sup>2</sup> was 65.62. This is larger than the threshold of 30.58 at the 1% significance level with 15 degrees of freedom, proving that the moderating effect of climate change perceptions was effective between the two groups.

Regarding the demographic variable interaction with ASC, age (*p* < 0.05), income (*p* < 0.01), occupation (*p* < 0.10), education (*p* < 0.10), and residence area (*p* < 0.05), variables were solely significant in the high-involvement group, while the effect of age (*p* < 0.10), education (*p* < 0.10), and residence area (*p* < 0.05) was marginally revealed in the lowinvolvement group. Thus, the statistical differences in the determinants between the models were verified. Moreover, concerning the main attributes, there were significant effects in both groups, and the directions of influence were also the same. However, since

there is a limit to the comparison of the variable impacts on the significance or the effect size, slope tests were performed to address the statistical differences of the effects [74].


**Table 5.** Comparison of conditional logit models between clusters.

Note (1) LR test b/w two clusters: χ<sup>2</sup> (0.01, 15) = 65.62 > 30.58; Note (2) Coefficient comparison w/χ<sup>2</sup> (0.01, 1) = 6.63;χ<sup>2</sup> (0.05, 1) = 3.84; Note (3) \*\*\*, \*\*, \*: Significance at 1%, 5%, 10% level.

> According to the main results, an improvement from the low-level to the mediumlevel (χ<sup>2</sup> = 4.17, *p* < 0.05), the low-level to high-level (χ<sup>2</sup> = 7.89, *p* < 0.01) of the DM function, and the low-level to high-level (χ<sup>2</sup> = 5.95, *p* < 0.05) of the WM function showed significant differences. Thus, the influence on the choice probability was found to be greater in the high-involvement group than the low-involvement group, although the coefficients had the same directions. In addition, income (t = 4.33) and occupation (t = −4.65) were statistically re-examined as being significant variables only in the high-involvement group. Regarding the local variable residence area, Daejeon region was significantly associated with the high-involvement group (t = −2.62) while the Chungbuk and Chungnam areas correlated to the low-involvement group (t = 2.31). However, there was no significant difference in the effect of age and education variables between the two groups.

> Table 6 shows the results of analyzing the marginal MWTPs for the two groups (Regarding the level changes of each attribute for the pooled sample ('Low' to 'Medium' and 'Medium' to 'High'), 4829 KRW and 1848 KRW for DM, KRW 2916 and 1271 KRW for FC, and 6058 KRW and 2663 KRW for WM were derived, respectively). The 95% MWTP confidence intervals were estimated using Krinsky and Robb [75]'s Monte Carlo simulation, and the t-statistics were derived based on the delta method [76]. First, regarding the DM function, the increments in the two levels of the high-involvement group were significant with the result. The different between the low to medium level is KRW 6467 (with a 95% confidence range of KRW 4261 to KRW 8673), and the medium to high level indicates a difference of KRW 2121 (with a 95% confidence range of KRW 382 to KRW 3860). Meanwhile, the low-involvement group disclosed KRW 3924 (from an interval of KRW 2139 to KRW 5693) in terms of the improvement from low to medium level. However, an insignificant effect on the change from medium to high level was detected. Regarding the FC function, there was no statistical difference between the two groups on the change from low to medium level (high involvement at KRW 2783 vs. low involvement at KRW 3036), while the medium level did not effectively move into the high level in both groups.


**Table 6.** Estimates of MWTP by cluster.

Note (1) Unit of Marginal WTP: won/year-household. Note (2) \*\*\*, \*\*: Significance at 1%, 5% level.

Moreover, the WM function also exhibited significance at each level in both groups as per the results. This indicates that MWTPs (95% confidence interval) of the high-involvement and low-involvement groups, respectively showed KRW 6471(KRW 4290–KRW 8652) and KRW 5661 (KRW 3898–KRW 7, KRW) regarding the change from medium to high level, as well as KRW 3,728 (KRW 1848–KRW 5609) and KRW 1951 (KRW 342–KRW 3555) regarding the change from medium to high level. However, the slope test results showed the intergroup heterogeneity. Finally, the total amount of MWTPs for the high-involvement group was KRW 21,570 (95% confidence interval: [17,450; 25,691]), and that of the low interest group was KRW 14,569 (95% confidence interval: [11,306; 17,832]).Thus, there a merged MWTP difference of around KRW 7000 a year per household between the two groups exists (in US dollar terms, the converted amount is approximately USD 18.58 and USD 12.61, respectively, with a difference of 6.06USD based on the exchange rate system of the Bank of Korea). The large difference in MWTP between high and low climate change awareness groups is in line with Kim et al. [34]'s study, which found that the Daecheong Dam basin was one of the most damaged areas in the summer of 2020, and that damage caused by climate change could worsen in the future.

Both groups showed the same value intensities in the order of water quality monitoring, drought management, and flood control. It is difficult to secure drinking water supplies in neighboring regions, since it continues to suffer from water quality issues such as non-point pollutant sources flowing from surrounding areas of rivers, which has a significant influence on the supply of various water types, such as daily and agricultural water [77]. Based on the findings of these prior research, the greatest MWTP of the survey respondents' water quality monitoring attribute is considered as accurately reflecting the true situation.

The estimated study values indicated results that are different from prior studies [55–58]. Nevertheless, there is agreement on the utility of reducing damages. Granted, the values were limited to tentative results that further research can rectify. These results, however, show that the economic value of the dam's public functions is regulated by climate change awareness. The estimation results can be differentiated from previous studies since the values of the dam's function attributes, corresponding to climate change, can be derived within a single analysis framework and the utility size between the attributes can be compared.

#### **4. Conclusions**

The purpose of this study was to estimate the economic value of the Daecheong Dam for the public function of responding to climate change. It examined the moderating effect of climate change perceptions on value estimates by applying choice experiments. The study specified three dam function attributes such as drought management, flood control, and water quality monitoring, and subdivided each into three levels to improve the status quo. Survey data from 603 households living in Daejeon, Chungbuk, and Chungnam were collected to perform the choice experiments. Subsequently, two clusters, including highinvolvement and low-involvement groups, were extracted based on the climate-change perception index. According to the main results of comparing the marginal willingness-topay between the two clusters, the attributes and price variable significantly affected the choice probability to benefit from improvements in the rational signs of the coefficients. This result does not violate the independence of irrelevant alternatives assumption. The improvement values of high-involvement and low-involvement groups are, respectively, estimated as KRW 21,570 and KRW 14,572 a year per household.

The findings of this study have the following managerial and policy implications. First, the estimates of the economic value of Daecheong Dam for the public function of responding to climate change are the same in both clusters, and were found to be in the order of water quality monitoring, drought management, and flood control. This can be interpreted as the environmental concerns at the study site being fairly reflected, as serious water bloom has occurred in Lake Daecheong since the rainy season in 2016, and because water pollution still needs to be addressed. Moreover, the implementation of a restricted water supply in the Chungnam area in 2015 raised the awareness of national disasters, leading to the perception that drought prevention is a more urgent problem, and demands an immediate countermeasures, in comparison to flood protection. These results show that public projects, for which the levels of public awareness are sufficiently considered, will have higher reception because the related economic value can vary according to public recognition.

As a result, in order to manage future water resources and establish measures to prevent water disasters in consideration of climate change, it is essential to first identify the increasing flood volume and decreasing dry-water volume due to climate change and to establish policies based on citizens' demands. For example, disaster-prevention urban planning, the designation and administration of natural disaster risk improvement zones, and the provision of safe drinking and living water may all boost the policy's positive efficacy.

This study has also demonstrated that the economic value of the dam's public functions are regulated by climate change awareness, which supports the belief that policies in which the public's propensity to climate change is considered can positively promote public welfare. Values of public roles are found for both groups, regardless of climate change perception, but their degree shows remarkable differences between the groups. In particular, the MWTPs on functional improvements are not significant at some levels, suggesting that people may disagree with the use of tax for strengthening its functions; this implies that the actual importance of the dam's role in climate change perception has not been understood fully. Hence, identifying the specific sub-groups according to the awareness level of climate change and building differential communication strategies for a paradigm shift is necessary to heighten the feasibility of implementing dam improvement and development plans. The results of this study can be differentiated from previous ones in that the values of the dam function corresponding to climate change can be derived within a single analysis framework, and the utility value between attributes can be compared. We believe that the obtained economic values translate into suggestions for fulfilling environmental policy needs.

Although it is meaningful, in that this study provides theoretical and practical implications to identify the core priorities of the dam's public functions from the perspective of the value concept, and thus suggest directions of future dam improvement projects, several limitations need to be addressed in further research. First, this study extracts the three attributes as the role of dams for climate change measures through the extensive literature review, expert surveys, and focus group interviews; however, there has been a limit to

the generalizability of these attributes to all cases even though it is considered reasonable to generalize the results of the study when the operating conditions of Daecheong Dam and the demographic characteristics of residents are under similar conditions. Therefore, it is recommended that more common functions appropriate for general cases will be further explored in subsequent studies, and socioeconomic variables reflecting specific regional characteristics will need to be considered more sensitively if the model extension. Second, the estimated value in this study is limited to the tentative results, not conclusive ones, as the choice experiments are based upon the questionnaire survey due to the stated preferences. In fact, since this drawback is an inevitable vulnerability of stated preference experiments, future researchers can develop a methodology to calculate the values of the dam functions using the revealed preference data. Third, at the level setting, the highest level is the expected value for new technology development, but the problem is that we could not explain exactly what kind of technology is anticipated. Regarding this issue, additional research on new technology development will be needed. Fourth, despite the introduction, attributes, and adequate explanation for each level of the study area (Daecheong Dam), the possibility of a Hypothetical Bias cannot be discounted. The need to solve the Hypothetical Bias using various techniques such as cheap talk [78], certainty follow-up [79] and oath [80] is raised. Additionally, a binary discrete choice question is incentive compatible, but multinomial repeated choice is not. In subsequent studies, a research design based on binary discrete choice is required.

Lastly, the conditional logit model requires a strict assumption that it must be accepted by IIA. Accordingly, in this study, the Hausman test was conducted to verify that the IIA assumption was fulfilled, and the conditional logit model was selected as the final research model. Many studies [81–83] stated that IIA assumption is too dependent on the parameterization of the model. The mixed logit model can be considered as an improved alternative that can describe choice probabilities across a given mixing distribution in an adaptable and flexible manner. Various mixing distributions, such as normal, log-normal, triangular, or uniform, can be used, depending on prior information on the taste variation among the decision makers [84]. In subsequent studies, it is necessary to carefully consider these points and select a research model.

**Author Contributions:** Conceptualization, H.O. and H.L.; methodology, H.O. and H.L.; software, H.O. and S.Y.; validation, H.L.; formal analysis, H.O. and S.Y.; investigation, H.O. and S.Y.; resources, H.O. and S.Y.; data curation, H.O. and S.Y.; writing—original draft preparation, H.O. and S.Y.; writing—review and editing, H.O. and S.Y.; visualization, H.O. and S.Y.; supervision, H.L.; project administration, H.L. All authors have read and agreed to the published version of the manuscript.

**Funding:** This research received no external funding.

**Institutional Review Board Statement:** Not applicable.

**Informed Consent Statement:** Informed consent was obtained from all subjects involved in the study.

**Data Availability Statement:** The data will be made available on request from the corresponding author.

**Conflicts of Interest:** The authors declare no conflict of interest.
