Current Issue

The occurrence of vacancies in single-family houses is influenced by a complex interplay of various factors, including economic, social, physical, topographical, environmental, and locational characteristics. Previous studies have demonstrated that factors influencing housing vacancies arise from various dimensions. These studies have sought to identify the causes of vacancies by categorizing them into levels such as the housing unit itself (including individual houses and parcels), the community, and the urban or regional scale. This includes individual building factors related to vacant houses, such as size, area, building structure, road width, and building age (Baba and Hino, 2019; Morckel, 2013; Nadalin and Igliori, 2017; Yin and Silverman, 2015). Studies examining socioeconomic characteristics have identified population variables as one of the key factors influencing housing vacancies. Most studies have emphasized that demographics are a primary cause of housing vacancies and that their impact varies depending on the regional context (Bassett et al., 2006; Edward and Joseph, 2005; Immergluck, 2016; Mallach, 2017; Morckel, 2013; Nassauer and Raskin, 2014; Page, 2001; Ribant and Chen, 2020; Wilhelmsson et al., 2011). Additionally, factors such as commercial services, surrounding landscape quality, environmental conditions, topography, and location have been examined as influencing the occurrence of housing vacancies. Some studies have indicated that areas with higher land values tend to exhibit higher vacancy rates (Ma et al., 2020). Discussions on environmental characteristics are closely linked to quality of life, emphasizing factors such as infrastructure accessibility, environmental hygiene, and commuting convenience (Chen et al., 2023; Deng and Ma, 2015; Joo and Lee, 2021; Hackman et al., 2019; Terada et al., 2016).

Generally, older, smaller, deteriorated, inexpensive, and poorly maintained houses have a higher probability of becoming vacant (Baba and Shimizu, 2023; Gu et al., 2019; Morckel, 2014; Sargent et al., 1997). Studies highlighting the importance of land size and price in maintaining housing value (Zhao, 2022) and the role of negative externalities from surrounding environments in promoting housing vacancies (Joo et al., 2022) underscore the necessity of proactive measures to address vacancies.

Predicting housing vacancy has long been a critical research topic across urban studies, real estate economics, and spatial planning. Early studies primarily adopted regression-based frameworks such as logistic regression, Ordinary Least Squares (OLS), and Generalized Linear Models (GLM) to examine the relationships between vacancy occurrence and various socioeconomic or physical attributes. These models were often chosen for their interpretability and statistical rigor. However, they operate under the assumptions of linearity, homoscedasticity, and normally distributed residuals, which may not hold in the context of highly skewed or non-normal vacancy data. A fundamental distinction must be made between explanatory and predictive modeling. Shmueli (2010) emphasizes that models designed for explanation are not necessarily optimal for prediction, and vice versa. In this respect, much of the early literature on housing vacancy focused on hypothesis testing and variable interpretation, rather than on optimizing predictive performance.

Recent advances in computational capacity and the availability of large-scale urban data have shifted scholarly focus toward prediction-oriented modeling frameworks, particularly those emerging from the fields of machine learning and nonparametric statistics. Machine learning algorithms such as decision trees, random forests, support vector machines, and artificial neural networks have been increasingly applied to forecast vacancy occurrences or to classify high-risk housing units (Martin et al., 2017; Newman et al., 2016). These models are data-driven and flexible in functional form, thus well-suited for capturing nonlinear and high-order interactions that are difficult to specify a priori.

Several studies distinguish between binary prediction models (i.e., predicting whether a house is vacant) and risk scoring models that generate continuous vacancy risk indices (Appel et al., 2014; Lee and Newman, 2017; Lee et al., 2021). While machine learning approaches have generally shown higher predictive accuracy, they often lack comparative assessments and are criticized for their “black-box” nature, particularly in policy contexts. Amandolia (2021) contributes a rare comparative evaluation of multiple nonparametric methods, underscoring the importance of validating predictive accuracy across models rather than relying on a single technique.

Nevertheless, most existing models−whether parametric or nonparametric−fail to incorporate spatial autocorrelation, despite the inherently geographic nature of housing vacancy. Spatial proximity often exerts a significant influence on vacancy risk due to neighborhood effects, service access, or physical decline. In response to this gap, this study employs Spatial GAMLSS, which integrates spatial random effects into a flexible distributional modeling framework. The theoretical structure and implementation of the Spatial GAMLSS model are detailed in Section 3.2.

This study addresses several key gaps identified in the existing literature on housing vacancy prediction and offers new contributions from both methodological and empirical perspectives.

First, previous studies on housing vacancy have predominantly employed traditional statistical models such as multiple regression or hedonic pricing models. While these approaches are effective for identifying linear associations, they often fail to capture nonlinear effects and spatial autocorrelation, both of which are crucial in understanding the geographic distribution of vacancy risk. This study addresses this methodological limitation by employing the Spatial GAMLSS framework. This model offers a flexible structure capable of estimating vacancy probabilities with higher accuracy by accounting for distributional complexity and spatial dependence−factors overlooked in most prior work.

Second, much of the existing research has relied on either aggregated regional data or small-sample household surveys, which limits the spatial granularity and inferential strength of the analysis. In contrast, this study builds a unit-level dataset of single-family housing, incorporating detailed locational, structural, and environmental attributes. The integration of high-resolution data with a spatially-aware statistical model enhances the explanatory power and predictive reliability, particularly in identifying localized high-risk zones.

Third, prior studies have rarely distinguished between urban and non-urban vacancy mechanisms, despite meaningful differences in demographic trends, land use intensity, and infrastructure accessibility. This study explicitly incorporates an urban–non-urban comparative framework, thereby revealing heterogeneous vacancy mechanisms across spatial contexts. This perspective not only enriches theoretical understanding but also supports the development of more context-sensitive housing policies.

In summary, the study makes three primary contributions: (1) It introduces a novel application of Spatial GAMLSS to vacancy risk prediction, advancing methodological approaches in housing vacancy research; (2) It leverages high-resolution, unit-level data for improved spatial inference; and (3) It incorporates an urban–non-urban comparative lens, enabling more nuanced policy insights for managing vacancy in diverse regional settings.

This section presents the analytical framework employed in this study. A sequential modeling strategy was developed to enhance both distributional flexibility and spatial sensitivity, culminating in a Spatial GAMLSS model with latent structured effects. The methodological procedure consists of five stages (M1–M5), which progressively integrate distributional complexity, nonlinear terms, and spatial random effects. Details of the distributional assumptions, model specification, and evaluation metrics are provided below. To estimate housing vacancy probabilities and identify spatial risk zones, this study adopts the Spatial GAMLSS framework−a flexible semi-parametric regression model capable of addressing the statistical complexities typically found in urban vacancy data. These complexities include distributional skewness, non-linearity, and spatial autocorrelation, which conventional regression approaches often fail to accommodate. Recent empirical applications underscore the strengths of GAMLSS and its spatial extension. Recent methodological overviews highlight that GAMLSS are particularly useful when data exhibit non-normal or bounded characteristics, as they allow modeling the entire distribution—including location, scale, and shape parameters—rather than only the mean. For example, Marmolejo-Ramos et al. (2022) demonstrated through applications in learning analytics how GAMLSS can flexibly fit skewed and bounded data, offering advantages over traditional regression models. Zhang et al. (2015) also utilized a GAMLSS-based model to predict China’s annual maximum precipitation, successfully modeling the nonlinear effects of explanatory variables using a Cubic Spline function. These studies have demonstrated that GAMLSS is a powerful tool for analyzing data with non-normal distributions and strong skewness. The GAMLSS model is increasingly being utilized in the housing sector as well. Florencio et al. (2011) introduced GAMLSS for land price prediction, demonstrating superior predictive performance compared to traditional regression models. Similarly, Cajias (2018) reported that GAMLSS produced more accurate results than Generalized Additive Models (GAM) in rent prediction. However, most of these studies did not explicitly incorporate spatial elements, limiting their ability to fully account for the influence of location and surrounding areas. Spatial GAMLSS extends the traditional GAMLSS model by incorporating spatial autocorrelation, making it advantageous for explaining how housing vacancy probabilities can vary depending on location, even under identical housing conditions. For instance, De Bastiani et al. (2018) utilized a Spatial GAMLSS model to analyze real estate price data and effectively captured the spatial patterns of price distributions across regions. Spatial GAMLSS, which accounts for spatial autocorrelation, enables precise analysis of the impact of geographic location on housing vacancies. This makes it a valuable tool for visualizing and predicting the spatial distribution of vacancy probabilities.

The classical GAMLSS model, developed by Rigby and Stasinopoulos (2005), extends the capabilities of GLM and GAM by allowing the dependent variable to follow any parametric distribution from a broad family (e.g., Poisson, gamma, inverse Gaussian, zero-inflated beta). Unlike traditional models that generally focus on modeling only the mean parameter, GAMLSS enables simultaneous modeling of the location (μ), shape (σ), skewness (ν, τ) through separate link functions. This multi-parameter framework enhances the model’s flexibility to capture heterogeneity in both central tendency and distributional form. In the context of this study, the μ and σ parameters are modeled as smooth additive functions of explanatory variables using penalized B-splines (P-splines), which allow for a flexible, data-driven representation of non-linear effects without imposing strict functional assumptions. Model selection is conducted based on penalized deviance, Akaike Information Criterion (AIC), and generalized pseudo R² to ensure optimal distributional and functional fit.

Crucially, the spatial extension of GAMLSS integrates spatial structure into the estimation process by applying an Intrinsic Conditional Autoregressive (ICAR) prior−equivalent to the Intrinsic Autoregressive (IAR) model−for the spatial random effect linked to the μ parameter. Following the approach by De Bastiani et al. (2018), this spatial formulation assumes that adjacent areal units exhibit correlated latent risks, thereby accounting for unmeasured spatial heterogeneity. This approach is especially relevant when physical, social, or policy factors influencing vacancy risks are spatially clustered but only partially observed in the dataset.

The resulting Spatial GAMLSS model not only improves the accuracy of local vacancy risk estimates but also provides a basis for visualizing the spatial diffusion of vacancy pressures across urban regions. Its ability to flexibly model non-normal distributions, capture non-linear relationships, and incorporate spatial structures makes it a suitable framework for analyzing the spatial patterns and underlying factors of housing vacancy.

In this study, the response variable is modeled as a continuous probability ranging between 0 and 1, representing the likelihood that a given single-family house is vacant. Accordingly, the μ parameter is directly interpreted as the estimated vacancy probability at each observational unit. Although the conceptual outcome is binary (i.e., vacant vs. non-vacant), a continuous probability formulation was adopted to facilitate spatial interpolation and to better capture uncertainty across units with varying characteristics. For the distributional assumption, several candidate distributions−including the Gaussian, gamma, and inverse Gaussian−were empirically evaluated. Based on model performance criteria such as the AIC and generalized R², the Exponential Gaussian distribution was selected as optimal. This distribution was particularly well-suited for the data’s asymmetric structure, accommodating right-skewed probability values while ensuring stability in parameter estimation. In operationalizing the spatial random effects, the spatial adjacency structure was constructed based on Queen contiguity using administrative units (eup, myeon, and dong). The ICAR prior was applied to the μ parameter to account for latent spatial dependence among neighboring areas, enabling more accurate estimation of spatially correlated vacancy risks.

For binary response variables (0/1), the mean parameter (μ) in the GAMLSS framework corresponds to the predicted probability of success under the binomial distribution. Accordingly, μ represents the expected value of the outcome, conditional on the covariates. This interpretation is consistent with that in GLM, where μ directly reflects the probability of a “success” event, such as a housing unit being vacant. In this study, even though the vacancy status is observed as a binary outcome, a continuous probability formulation was adopted for μ to enable probabilistic interpretation, smooth spatial prediction, and incorporation into subsequent risk mapping.

To accurately capture the spatial distribution and statistical characteristics of housing vacancy, this study constructed a five-stage modeling framework (M1–M5). Each stage incrementally expands the analytical scope, beginning with a simple linear regression model and progressively incorporating distributional flexibility, nonlinear covariate effects, multi-parameter estimation, and spatial autocorrelation structures.

In the first stage (M1), the linear regression model assumes a normally distributed response variable Y with mean μ and constant variance σ². The mean is expressed as a linear combination of the design matrix X and coefficient vector β, as shown in Equation (1):

(1)

The second stage (M2) introduces the Generalized Linear Model (GLM), which extends the response distribution to the exponential family. A link function g(•) is applied to the mean μ, yielding the linear predictor η = g(μ). While this enhances distributional flexibility, the relationships between variables remain linear (Equation (2)):

(2)

In the third stage (M3), the GAM is adopted to relax linearity assumptions. Each covariate is modeled with a nonparametric smoothing function s_j(x_j), commonly implemented using P-splines. This enables the model to capture nonlinear covariate effects directly from the data (Equation (3)):

(3)

The fourth stage (M4) adopts the generalized additive models for location, scale, and shape (GAMLSS) framework. Unlike GLM and GAM, which typically model only the mean, GAMLSS enables separate modeling of multiple distribution parameters−mean μ, variance σ, and skewness ν−each with its own link function and smoothing structure. In this study, the exponential Gaussian distribution is used, with identity, log, and log link functions applied to μ, σ, and ν, respectively (Equation (4)):

(4)

In the final stage (M5), spatial effects are integrated into the GAMLSS framework by incorporating a spatial random effect into the mean predictor μ. This accounts for potential spatial autocorrelation among adjacent administrative units. The spatial effect follows an Intrinsic Conditional Autoregressive (ICAR) structure based on a queen contiguity matrix constructed from municipal boundary data. This enables the model to reflect similarities among geographically proximate areas. The spatial random effect model is given by Equation (5):

(5)

The spatial random effect γ is assumed to follow an ICAR prior, and its estimates are computed using the following formulation, where λ = σϵ2 /σγ2 represents the ratio of error variance to spatial signal (Equation (6)):

(6)

All models were fitted using the RS algorithm, and model performance was assessed using generalized R² and the AIC. The generalized R² is defined Equation (7):

(7)

Where L(0) denotes the log-likelihood of the null model (intercept only), and L(fitted) represents the log-likelihood of the fitted model, and n is the number of observations. Higher values of generalized R² indicate better model performance. The AIC is expressed Equation (8):

(8)

Where, l^ is the fitted log-likelihood, df is the effective degrees of freedom, and k is the penalty parameter, set to 2 in the case of AIC.

The response variable (probability of being vacant) follows an Exponential Gaussian distribution with parameters μ, σ, and ν. The link function for the parameters was set as the identity function for mean (μ) and the log function for variance (σ) and skewness (ν). This approach allows each parameter of the model to flexibly vary according to the explanatory variables. The final model can be expressed Equation (9).

(9)

All models were estimated using the RS (residual scoring) algorithm via the GAMLSS package in R. The spatial contiguity matrix was implemented using the spdep package. Model performance was assessed using the Akaike Information Criterion (AIC) and generalized R², and the detailed model comparison results are presented in Section V.

This study employs seven explanatory variables to estimate the probability of housing vacancy: Floor area, Roof structure, Building age, Owner-occupancy status (dummy variable), Distance from administrative boundaries, Distance from water boundaries, and Distance from environmental pollutant emission facilities. These variables were selected based on prior empirical findings and their theoretical relevance to the structural and spatial dynamics of Jinju-si.

Floor area serves as a proxy for the physical scale and functional capacity of the housing unit. Smaller houses tend to exhibit lower market appeal, are more susceptible to functional obsolescence, and are often bypassed in reinvestment decisions, especially in peripheral or aging neighborhoods. Prior research has consistently shown that reduced floor area is positively associated with higher vacancy rates (Clark and Herrin, 1997; Lester et al., 2013).

Roof structure reflects the structural durability and construction quality of the house. Houses with outdated or vulnerable roofing materials (e.g., wood, panels) are more likely to suffer from leakage, fire hazards, or thermal inefficiency. Such homes are frequently excluded from reinvestment cycles and may face prolonged vacancy, particularly in declining neighborhoods (Liu et al., 2024; Yue et al., 2022).

Building age captures the degree of physical deterioration and structural obsolescence. Older buildings often require greater maintenance, may not comply with contemporary building standards, and are more likely to remain vacant, especially in areas with declining population and weak housing demand (Newman, 2013; Lee et al., 2022).

Owner-occupancy status is a key socio-economic indicator reflecting the likelihood of housing management and residential stability. Owner-occupied homes tend to be better maintained and less prone to vacancy, whereas absentee ownership is often associated with disinvestment and higher vacancy risk (Martin et al., 2017; Molloy, 2016).

Distance from administrative boundary measures spatial marginality. Properties located near jurisdictional borders frequently suffer from reduced access to services, limited infrastructure investment, and lower integration into planning frameworks. These “edge effects” have been associated with housing instability and elevated vacancy rates (Henry et al., 2001; Geoghegan et al., 1997).

Distance from water boundary captures the influence of natural landscape features. While waterfront proximity can enhance property values in some metropolitan areas, in many secondary cities like Jinju-si, water-adjacent lands are often underdeveloped, flood-prone, or designated for non-residential use, thus becoming a locational disadvantage (Newman et al., 2016; Yue et al., 2022).

Distance from environmental pollutant emission facilities accounts for the negative externalities of industrial land uses. Residential properties near pollution-emitting facilities (e.g., factories, waste plants) face health and environmental concerns, often leading to reduced desirability and increased vacancy risks. In Jinju-si, these facilities are typically located in industrial zones with adjacent residential areas showing higher vacancy rates (Cozens and Tarca, 2016; Joo and Lee, 2021).

All spatial variables were operationalized using high-resolution GIS data from national sources such as the National Geographic Information Institute and the Ministry of the Interior and Safety. Euclidean distance metrics and buffer zones were applied to quantify spatial exposures.

This study estimates the probability of housing vacancies in single-family houses in Jinju-si and predicts future high-risk vacancy zones through the following four steps.

First, based on existing research and data analysis, potential variables that may influence housing vacancies were identified. The correlations between these variables were examined, and candidate variables for analysis were selected. Through this, the aim was to enhance the accuracy of vacancy probability analysis by considering regional characteristics and the spatial distribution of housing vacancies.

Second, the Spatial GAMLSS model was used to estimate the probability of housing vacancies and analyze the influence of each variable. In this process, spatial autocorrelation and nonlinear relationships were incorporated, allowing for the estimation of housing vacancy probabilities in both urban and non-urban areas. This approach enabled the quantitative assessment of spatial factors influencing housing vacancies in specific areas.

Third, based on the Spatial GAMLSS model, areas with high housing vacancy probabilities were visualized. Specifically, spatial clustering techniques were employed to identify high-risk areas for housing vacancies (hotspots) and stable areas (cold spots), enabling the prediction of future vacancy-prone regions.

Fourth, by comparing high-risk housing vacancy areas classified into urban and non-urban regions, the study identified priority areas for policy intervention. This allowed for the identification of key causes of housing vacancies and the development of region-specific management strategies. As a result, effective policy directions were proposed for the efficient allocation of limited resources and the resolution of housing vacancy issues.

Through the analytical procedures outlined in <Figure 4>, this study aims to provide a comprehensive understanding of the probability and location of housing vacancies in single-family houses in Jinju-si, and to establish a practical policy foundation for addressing the housing vacancy issue.

Figure 4. 
Research flowchart

This study conducted a hotspot analysis in Jinju-si to identify areas and buildings at high risk of housing vacancy in urban and non-urban regions. The hotspot analysis was based on housing vacancy probabilities estimated using the Spatial GAMLSS model, and the results were derived for urban areas, non-urban areas, and the entire region of Jinju-si.

Two types of hotspot analyses were conducted to reflect both meso- and micro-level spatial patterns. First, the spatially estimated vacancy probabilities (μ) were aggregated to the eup–myeon–dong level, and the average values were used to identify area-based risk clusters using the Getis-Ord Gi* statistic (Figures 8(a) and 9(a)). Second, the same μ values at the individual housing unit level were used as input for point-based hotspot analysis (Figures 8(b) and 9(b)), allowing the detection of statistically significant clusters of high-risk properties. This dual-scale analysis facilitates a more comprehensive interpretation of vacancy risk across both regional and building levels.

Figure 8. 
Vacancy risk in urban areas based on hotspot analysis

Figure 9. 
Vacancy risk in non-urban areas based on hotspot analysis

<Figure 8(a)> illustrates the areas at risk of housing vacancy in the urban regions of Jinju-si. The risk of housing vacancy was low in new urban developments and areas near transit hubs, whereas it was high in areas with aging residential and infrastructure facilities, as well as gateway regions on the urban outskirts. This suggests that housing vacancy risk increases in areas with lower residential desirability and declining housing demand. In particular, traditional old city centers and early-developed residential areas exhibited a higher risk of housing vacancies. In contrast, newly developed urban districts, areas urbanized in connection with industrial complexes, and regions with active educational and economic activities showed a lower likelihood of housing vacancies. These findings indicate that residential preferences and the distribution of economic activities within the city significantly influence the risk of housing vacancies.

<Figure 8(b)> visualizes buildings in Jinju-si’s urban areas with a high risk of housing vacancy. High-risk buildings were predominantly located in areas with aging residential environments or on the urban outskirts. These areas often exhibited low residential desirability due to inadequate transportation infrastructure or limited accessibility to the city center. Additionally, regions with declining old commercial districts or mixed urban-rural residential environments showed relatively higher vacancy risks. In contrast, buildings with a low risk of vacancy were identified in areas where recent development has been active or where urban infrastructure is well-established. These areas exhibited high residential desirability and low vacancy rates due to the stable maintenance of residential and commercial functions, as well as well-developed transportation and industrial infrastructure. Notably, regions undergoing new urban development or industrial complex construction were found to have stable housing demand and a low risk of vacancy. Summarizing the results of vacancy risk predictions in the urban areas of Jinju-si, it was found that areas with aging residential environments and poor transportation accessibility had a high risk of housing vacancies. In contrast, regions characterized by active urbanization and economic activities exhibited a lower likelihood of vacancy occurrence. These findings suggest that factors such as the level of urbanization, economic activity, and residential desirability play a significant role in influencing housing vacancy risks.

<Figure 9(a)> illustrates the high-risk areas for housing vacancy in the non-urban regions of Jinju-si. The analysis revealed that vacancy risk was elevated in industrial areas with a manufacturing base and rural areas with low accessibility to urban centers. These regions shared several common economic and social characteristics. First, areas with limited transportation accessibility exhibited a pronounced risk of housing vacancy. Regions lacking connectivity to major transportation networks or located far from urban centers exhibited reduced residential desirability due to limited mobility for residents and lower accessibility to local commercial and living infrastructure. In such areas, the loss of functionality in former transportation hubs, caused by rail relocations or route closures, contributed to higher vacancy rates. Second, areas with weakened economic foundations due to industrial restructuring showed a higher risk of housing vacancies. While regions with active manufacturing industries in the past experienced the development of industrial and agro-industrial complexes, recent declines in manufacturing competitiveness and operational rates have led to economic stagnation. As a result, reduced worker inflows and declining local housing demand have increased the likelihood of housing vacancies. Third, natural constraints were also found to influence housing vacancy. In agriculture-focused regions, frequent flooding and natural disasters often resulted in poor living conditions. These factors contributed to long-term resident outmigration, ultimately leading to an increase in housing vacancies. In contrast, areas with well-connected transportation networks or relatively stable tourism and agricultural bases showed lower vacancy risks. Overall, the risk of housing vacancies in non-urban areas exhibited significant variation depending on industrial and economic structures, accessibility to urban centers, and the level of regional infrastructure.

<Figure 9(b)> presents the spatial distribution of buildings with a high risk of vacancy in the non-urban areas of Jinju-si. The analysis revealed that vacancy risks were predominantly high in regions engaged in agriculture- and manufacturing-based economic activities, as well as in peripheral areas with limited transportation accessibility. Regions with weakened economic foundations exhibited higher risks of housing vacancies. Areas previously characterized by thriving agriculture and manufacturing activities have recently experienced increasing vacancy rates due to industrial restructuring and a decline in economic activity. These regions, often concentrated with agro-industrial or manufacturing complexes, have seen decreased building utilization as industrial competitiveness has eroded and population outflows have intensified. Moreover, areas with limited accessibility to urban centers also demonstrated elevated vacancy risks. These regions often suffer from poor connectivity to major transportation infrastructure or are located at considerable distances from city centers, restricting access to essential residential and commercial amenities. This lack of accessibility diminishes the residential appeal of these areas, acting as a significant factor that exacerbates vacancy risks. In contrast, areas with lower vacancy risks were rural regions characterized by relatively high accessibility to urban centers and well-developed residential and commercial infrastructure. These areas, benefiting from a balanced influence of urbanization, maintained stable economic activities and exhibited reduced vacancy risks. These findings underscore the need for vacancy management policies tailored to the specific characteristics of non-urban areas. In regions experiencing economic stagnation, strategies aimed at industrial revitalization and enhancing residential appeal are essential. Meanwhile, in areas with limited transportation accessibility, policies focused on improving physical connectivity could serve as effective measures to mitigate housing vacancy issues.

This study analyzed the characteristics and key factors of housing vacancy occurrences in urban and non-urban areas of Jinju-si, highlighting significant differences in the spatial patterns and causes of vacancies between these two regions. The findings underscore the necessity for tailored vacancy management strategies that account for regional characteristics. Using a Spatial GAMLSS model, the probabilities of housing vacancy occurrences in urban and non-urban areas were estimated, enabling the effective prediction of high-risk zones and vulnerable properties. This approach provided a detailed understanding of the spatial features and influential factors driving housing vacancy issues. The results offer valuable insights for devising targeted policy responses to address these challenges effectively.

In urban areas, the risk of housing vacancy was higher in old city centers, early-developed residential neighborhoods, and areas with declining commercial activity. This trend can be attributed to the gradual marginalization of older urban districts during the urbanization process, alongside the weakening of commercial and infrastructure functions. Notably, urban fringe areas with limited transportation infrastructure or reduced accessibility also exhibited a higher likelihood of vacancies. In contrast, regions with well-established urban infrastructure, such as new towns and areas near transportation hubs, showed relatively low vacancy risks. These findings highlight the necessity of policy interventions to address spatial imbalances in urban economic activity and housing demand. Strategies such as urban center regeneration, commercial district revitalization, and transit-oriented development are essential to mitigate these disparities. In non-urban areas, the risk of housing vacancy was higher in regions where agricultural and manufacturing-based economic activities have weakened and in peripheral areas with limited accessibility to city centers. This trend appears to be driven by a decline in residential attractiveness due to economic base deterioration caused by industrial structural changes and insufficient connectivity to transportation networks. Conversely, areas with stable agricultural and tourism bases or convenient access to urban centers exhibited relatively low vacancy risks. These findings underscore the need to address housing vacancy issues in non-urban areas through economic base strengthening and infrastructure enhancement. Furthermore, a comparative analysis of vacancy factors between urban and non-urban areas revealed distinct differences. For instance, owner-occupancy status was identified as a significant factor in suppressing housing vacancy in non-urban areas, while it did not exert a notable influence in urban areas. This suggests that residential stability and a sense of community play a critical role in mitigating vacancy risks in non-urban contexts. The distance to water boundaries had contrasting effects depending on regional characteristics. In urban areas, homes closer to water boundaries were associated with higher residential desirability. In contrast, in non-urban areas, proximity to water boundaries increased the likelihood of housing vacancy due to greater exposure to natural disaster risks. These findings highlight the necessity of tailored vacancy management policies that consider regional characteristics and adopt differentiated approaches for urban and non-urban areas.

The housing vacancy probabilities estimated using the Spatial GAMLSS model enabled the visualization of high-risk vacancy areas and individual at-risk buildings. The analysis revealed that in urban areas, vacancy risks were concentrated in aging residential zones and areas with poor transportation accessibility. In non-urban areas, regions with weakened agricultural and manufacturing bases, as well as peripheral locations, exhibited higher vacancy risks. Additionally, risk predictions at the individual building level provided practical evidence for identifying buildings in urgent need of management, facilitating the efficient allocation of resources and prioritization of problem-solving efforts. In conclusion, this study underscores the necessity of adopting differentiated approaches that incorporate regional characteristics and problem-specific factors to address housing vacancy issues in urban and non-urban areas. In urban areas, priorities should include the revitalization of old downtowns, activation of commercial districts, and development of transit-oriented hubs. In non-urban areas, strategies such as revitalizing industrial complexes, improving transportation networks, and strengthening economic foundations are required. Proactive interventions targeting high-risk areas and buildings can prevent the exacerbation of housing vacancy issues. Furthermore, region-specific policies can promote efficient resource utilization and effective problem-solving tailored to the unique needs of each area.

This study conducted a comparative analysis of housing vacancy characteristics in urban and non-urban areas of Jinju-si using a Spatial GAMLSS model and hotspot analysis. It identified the key factors and spatial distributions associated with housing vacancies and effectively predicted high-risk areas and individual properties prone to vacancy. By analyzing the differentiated factors driving housing vacancies in urban and non-urban areas, the study highlighted the necessity of formulating tailored policies based on regional characteristics. Furthermore, it provided policy-makers with actionable data that can serve as practical decision-making tools. In urban areas, the risk of housing vacancies was found to be higher in declining commercial districts and aging residential neighborhoods, whereas in non-urban areas, limited transportation accessibility and weakened economic foundations emerged as key factors. These findings underscore the necessity for differentiated vacancy management strategies tailored to the specific contexts of urban and non-urban areas. They also provide critical baseline data for developing detailed intervention plans targeting high-risk areas and individual properties prone to vacancy.

Nevertheless, this study has four key limitations. First, it is based on cross-sectional data and does not account for temporal trends in housing vacancies. Incorporating longitudinal data through time-series analysis could enable a deeper understanding of the dynamic changes in housing vacancies and provide a more comprehensive evaluation of the effects of policy interventions over time. Second, while this study focuses on the regional characteristics of Jinju-si, the findings may not be directly applicable to other areas due to varying local contexts. Comparative studies across diverse regions are necessary to generalize the results. Third, this research primarily emphasizes spatial factors and does not fully account for the interactions with socio-economic variables such as income levels, employment rates, and educational environments. Integrating these factors could provide a clearer understanding of the complex causes behind housing vacancies. Fourth, although the study offers policy implications, it does not evaluate the feasibility or effectiveness of the proposed policies in real-world applications.

To address these limitations, future research should proceed in the following directions. First, it is essential to classify urban areas into more specific types, such as old downtowns, new developments, and commercial centers, as well as non-urban areas into agricultural or tourism-focused zones, to analyze the characteristics and factors of housing vacancies in greater detail. This approach would enable the development of tailored policy recommendations for each subregion. Additionally, utilizing temporal data to examine the dynamic changes in housing vacancy rates and predict long-term vacancy risks is imperative. Such analyses could provide a foundation for designing sustainable housing vacancy management and response strategies. Third, while this study focused primarily on spatial factors, housing vacancy issues are intricately linked to socioeconomic variables such as income levels, educational environments, and demographic structures. For instance, regions with lower income levels may face difficulties in maintaining and repairing homes, increasing the likelihood of vacancies. Similarly, areas with inadequate educational facilities may see reduced residential appeal, leading to population outflows. From a demographic perspective, aging populations and youth migration can be significant drivers of housing vacancies. By integrating socioeconomic factors with spatial considerations, future research could more effectively identify the complex causes of housing vacancies. To achieve this, it would be beneficial to include a broader range of variables−such as household income, educational attainment, population density, age distribution, and employment rates−into the modeling process. Employing dimensionality reduction techniques, such as principal component analysis, or hierarchical modeling approaches could help manage computational complexity while maintaining analytical rigor. This comprehensive approach would provide a clearer understanding of the multifaceted nature of housing vacancy dynamics.

In conclusion, this study compared and analyzed the factors and spatial characteristics of housing vacancies in urban and non-urban areas of Jinju-si, emphasizing the necessity of tailored vacancy management policies based on regional characteristics. Future research should adopt a long-term and integrated approach to identify the root causes of housing vacancies and propose practical and effective policies for urban and regional regeneration. This study serves as a starting point for understanding housing vacancy trends and improving more reliable predictive models.