Multiscale Effects of Hedonic Attributes on Airbnb Listing Prices Based on MGWR: A Case Study of Beijing, China

Abstract

An investigation into the pricing mechanism of Airbnb is crucial for achieving the sustainable development of sharing economy accommodations and has great academic and practical significance. The existing pricing studies on sharing economy accommodation have identified a set of important factors impacting prices based on the hedonic price model. However, the spatial scale of the impact of various hedonic attributes on Airbnb listing prices is not yet clear. This study takes Beijing, China, as a case study; develops a conceptual framework that incorporates four categories of hedonic attributes; and uses a spatial heterogeneity perspective to investigate the multiscale spatial effects of various attributes on the prices of Airbnb listings. Our findings revealed the following: (1) The explanatory power of different categories of attributes towards listing prices varies from high to low, as follows: functional attributes, locational attributes, reputational attributes, and host status attributes, among which the functional attributes are the most important determinants of Airbnb listing prices. (2) There are multiscale, spatially heterogeneous relationships between Airbnb listing attributes and prices. Specifically, the functional attribute variables have local influencing scales, the reputation attribute variables have regional scales, and the variables of host status and locational attributes have global scales. (3) Compared with ordinary least squares (OLS) regression and geographically weighted regression (GWR), multiscale geographic weighted regression (MGWR) improves overall modelling ability by introducing multiple scales and is better suited to illuminating the hedonic pricing of sharing economy accommodations. This study provides new insights into the spatially varied relationships between listing attributes and Airbnb listing prices, which can deepen our understanding of sharing economy accommodation and help hosts formulate location-based pricing strategies.

1. Introduction

Sharing economy accommodation, such as Airbnb, emerged in the context of sustainable development and flourished due to the innovation of platform technology [1]. It enables hosts to obtain economic benefits by transferring the use rights of idle houses and encourages consumers to participate in the economic activities of local communities. In addition, it can help solve employment problems and promotes the sustainable development of society [2,3]. As a result, sharing economy accommodation is closely related to sustainability and is a more sustainable alternative to traditional accommodation [4,5,6,7].

Pricing is widely acknowledged to be one of the most critical factors determining the sustainable development of the accommodation industry [8]. Reasonable pricing is the key to hosts’ attainment of profits and is also an important indicator governing consumers’ propensity to choose sharing economy accommodation [9]. However, the complexity and heterogeneity of Airbnb listings make pricing challenging [10]. In addition, in the context of COVID-19, the development of the sharing economy accommodation industry is faced with challenges such as increasing uncertainty, reduced demand for accommodation, and increasing difficulty with respect to making a profit, which bring about higher requirements for the pricing strategies employed by hosts [11,12]. Therefore, clarifying the price mechanism of sharing economy accommodation has important practical significance for hosts’ management and the sustainable development of the industry.

As in many other spatial processes, the process that generates listing prices may also have potential spatial nonstationarity [13,14,15,16]. First, the relationship between the same influencing factor and listing prices may have spatial heterogeneity. For example, the distance to the city center might not have the same effect on listing prices across an entire study area. In urban areas, the closer to the city center, the higher the listing price; in suburbs, however, it seems that the farther away from the city center, the higher the listing price. In addition, there may be differences with respect to the influencing scales of various factors on listing prices. Taking Los Angeles as an example, the relationship between the number of bedrooms and listing prices only remains stable within a very small spatial scale, and it changes dramatically when we move beyond this scale. In comparison, factors such as the number of reviews and the distance to tourist destinations have relatively stable impacts on listing prices across the city, so they have global influencing scales [14].

The abovementioned subtle disparities lead to the relationships between the influencing factors and listing prices showing spatial heterogeneity and multiscale effects in space. This heterogeneity and scale difference exist not only within cities but also between different cities [14]. Beijing, a megacity and the capital of China, has an active sharing economy accommodation market and can be a representative case. Therefore, based on the Beijing Airbnb listing data in 2021, we conducted research on the pricing mechanism of sharing economy accommodation in Beijing.

This study aims to explore the multiscale effects of hedonic attributes on Airbnb listing prices, and attempts to answer the following questions: Which attributes affect the Airbnb listing prices in Beijing? Do the relationships between various attributes and listing prices have spatial heterogeneity? If so, at what spatial scales do they vary?

To answer the above questions, we formulated two main hypotheses.

H1: 

In Beijing, China, Airbnb listing prices are spatially dependent.

H2: 

There are multiscale, spatially heterogeneous relationships between Airbnb listing attributes and prices.

The remainder of this paper is organized as follows: Section 2 summarizes the relevant literature. Section 3 provides the conceptual framework, data, and methods. Section 4 illustrates the results. Section 5 provides the conclusion and discussion. Section 6 puts forward the theoretical and practical implications and points out the limitations and future research directions.

2. Literature Review

2.1. The Spatial Heterogeneity and Multiscale Effects in the Hedonic Price of Airbnb

Like houses, Airbnb listings are a type of differentiated product characterized by durability, heterogeneity, and spatial fixity [10,14,17,18]. The hedonic price model, derived from Lancaster’s consumer theory and Rosen’s implicit market model [19,20,21], is a major scientific method for constructing price indices for differentiated products [22]. The hedonic price model assumes that the products are typically sold as a package of attributes, and the price of a product can be viewed as a function of the product’s measurable, utility-based attributes, or characteristics. Since Ridker and Henning applied the hedonic price theory to a housing market analysis in 1967 [23], and after more than 50 years of practice and development, the hedonic price model has developed into one of the most widely used models in research into real estate, hotels, and other fields [24,25,26,27,28].

Although Airbnb is seen as a more sustainable alternative to traditional accommodation such as hotels, both belong to accommodation services, so their pricing mechanisms have theoretical commonality. According to existing research on hedonic pricing, hotel property characteristics, star ratings, hotel services, ancillary facilities, location, and external market factors are the most significant determinant elements of hotel prices [29,30,31]. Inspired by hotel-pricing literature, many scholars have analyzed the hedonic prices of Airbnb listings, and these studies have also confirmed the applicability of the hedonic price model in the field of sharing economy accommodation [10,13,14,15,16,17,32,33,34,35].

Some early research studied the hedonic prices of listings based on ordinary least squares (OLS) regression [10,17,32]. OLS regression can construct efficient and robust explanatory equations when the relationships between the predictor variables and response variable are consistent over space. However, the relationship between Airbnb listing attributes and prices may vary across space; that is, the relationship is spatially non-stationary and there is spatial heterogeneity. According to consumer theory, this spatial heterogeneity may be due to consumers in different regions having different needs and preferences when it comes to Airbnb listings, so the impact of the attributes on prices also has spatial heterogeneity. Existing studies have generally verified this spatial heterogeneity [14,15,16,36,37].

As a local regression model, classical geographically weighted regression (GWR) is one of the most widely used methods for exploring potential spatial heterogeneity [38,39]. The significant contribution of GWR is that it allows relationships between predictor variables and response variables to vary across space [40,41,42]. Some studies have found that compared with OLS regression, GWR is better suited for exploring spatially varying relationships between Airbnb listing attributes and prices [13,15,36]. However, GWR also has limitations, as it assumes that all relationships vary at the same spatial scale. This assumption ignores scale differences in the spatial variation of different relationships, since the relationships of different listing attributes and listing prices may vary at different scales (locally, regionally, or globally) [43]. By introducing the multiscale concept, the multiscale geographic weighted regression (MGWR) model proposed by Fotheringham et al. (2017) can better reflect the scale-related difference between spatial relationships [44]. The most notable improvement offered by MGWR over GWR is that it allows each relationship to vary at a unique spatial scale and is, therefore, better suited for the multiscale analysis of spatially heterogeneous processes [44,45,46,47,48].

However, in the field of hedonic pricing with respect to sharing economy accommodation, related research is very limited. To the best of our knowledge, currently, only Hong and Yoo (2020) [14] have used MGWR to model Airbnb prices in New York and Los Angeles, and they drew the following conclusions: MGWR is superior to OLS and GWR, there is heterogeneity in the relationships between listing attributes and prices, and the influencing scales of attributes differ not only within cities but also between cities [14]. Taking Beijing, China, as a case study can expand our understanding of the Airbnb market-pricing mechanism from a heterogeneous perspective, compensate for the lack of related research, and provide international insights into the pricing mechanisms of different cities.

2.2. Price Determinants of Airbnb Prices

Previous studies have examined the hedonic prices of sharing economy accommodation and identified a range of important attributes that influence price. Scholars have classified these hedonic attributes in different ways. Chen and Xie (2017) classified the attributes as intrinsic factors (functionality and host efforts) and extrinsic factors (consumer reviews and competition) [10]. Hong and Yoo (2020) discussed the influencing factors from four perspectives: listing functions, reputation, hosts, and location [14]. Wang and Rasouli (2022) classified the attributes into three categories: structure, reputation, and location [34]. Following the classification method of Hong and Yoo (2020), we divide the listing attributes into the following four categories: functional attributes, host status attributes, reputational attributes, and locational attributes.

Functional attributes reflect the internal characteristics of an Airbnb listing and have been incorporated into nearly all studies on hedonic price. Commonly used functional variables include accommodation capacity, the number of bathrooms, the number of beds, the number of facilities, etc. [4,6,28,29,30,31]. Numerous studies have revealed that functional attributes related to accommodation demand, especially the number of bedrooms and accommodation capacity, have positive impacts on listing prices [10,14,16,17,49]. For example, Gyódi and Nawaro (2021) found price increases of 6.6–25.6% for every additional bedroom [16].

Reputation, as a key attribute of an Airbnb listing, is a significant criterion in the customer’s choice of accommodation [50]. Reputation-related variables, such as review numbers and ratings, have been enumerated in almost all Airbnb price determinant models [10,14,15,16,17,32,33,34,35,36,49,51,52,53]. Concerning the number of reviews, the results are mixed: in some cases, no significant effect on price was found or it was only partially significant [15,52]; in other cases, the effect was significant and negative [17,32,34,36,49,52,53]. Likewise, there is no consensus in the literature concerning the impact of ratings on listing prices, as Zhang et al. (2017) found that ratings had a negative impact on listing prices [36] while Wang and Nicolau (2017) and Deboosere et al. (2019) found that the higher the rating, the higher the price [32,49]. Additionally, Hong and Yoo (2020) showed that in most listings in New York and Los Angeles, the impact of ratings on prices was not statistically significant [14].

Host heterogeneity is also an important factor affecting listing prices. Airbnb is a special marketplace wherein heterogeneous sellers sell differentiated products. That is to say, not only are listings heterogeneous, but hosts are also heterogeneous, and different hosts may make different pricing decisions due to their different motivations and professional abilities [18]. Therefore, most studies incorporate the host’s status as an important attribute in the model. Commonly used variables include the possession of a Superhost badge [10,14,17,32,51] and professional host status (i.e., a host who manages more than two listings) [15,17,36,53]. Existing studies show that host attributes have a mixed impact on listing prices. Taking professional host status as an example, many scholars have found that professional hosts can better implement dynamic pricing, thereby creating higher revenue [15,17,53,54]. For example, Voltes-Dorta and Inchausti-Sintes (2020) found that professional hosts charge approximately 9% more than non-professional hosts [15]. However, some scholars have come to different conclusions; for example, Gibbs et al. (2018) found that among five Canadian cities, professional hosting only resulted in a slight premium in Montreal, and in the other four cities there was no price difference between professional and non-professional hosts [17]. In Hong Kong and New York City, it has been found that professional hosts charge less than non-professional hosts [53,54].

The most important attribute is location, which has been shown to play a key role in the hedonic pricing of Airbnb listings [10,14,15,16,17,32,36,49,51,55,56]. Typically, locations are represented in terms of accessibility and local market competition. First, accessibility is further divided into destination accessibility and traffic accessibility. The commonly used variables in this regard include the distance to popular destinations such as a city center [15,32,51], a local city hall [17], a convention center [36], a coastline [16], one or more popular attractions [14,55], and the number of tourist attractions nearby [51], as well as the distance to transportation facilities such as bus stations [51], subway stations [16,49], or railway stations [51]. Existing studies have confirmed a positive correlation between destination accessibility and listing prices [17,36,51,56]; that is, the closer to the destination, the higher the listing price. For example, Gibbs et al. (2018) found that for every 5 km decrease in the distance to a city hall, prices increased by between 4.0% and 20.7% [17]. Voltes-Dorta and Inchausti-Sintes (2021) found that in Bristol, the number of tourist attractions around the listing is significantly positively correlated with the price [51]. However, consistent conclusions have not been drawn regarding the impact of transport accessibility. Although transportation accessibility is generally positively correlated with price, Deboosere and Kerrigan (2019) and York and Gyódi K (2021) found that in New York City and Rome, listing prices near subway stations are lower [16,49]. Voltes-Dorta and Sánchez-Medina (2020) also found that in Bristol, the closer a listing is to a bus station, the lower the price per guest for apartment-type listings [15]. In addition, a few studies have investigated the effect of market competition on listing prices using the number of nearby hotels or listings as variables [10,14,51]. The impact of market competition on prices usually follows traditional pricing mechanisms; for example, Chen and Xie (2017) and Voltes-Dorta and Inchausti-Sintes (2021) found that the number of local competitors negatively affects listing prices, i.e., greater supply lowers prices due to higher competition in the local market [10,51]. However, Hong and Yoo (2020) found that the influence of the number of nearby listings was completely insignificant in terms of listing prices in Los Angeles but showed a positive effect in the New York area [14].

3. Study Area, Conceptual Framework, and Methods

3.1. Study Area and Data Sources

Different Airbnb listing types may have different pricing mechanisms because they target different groups of guests with different value preferences [15,57]; therefore, in this study, we take entire apartments as a research object, representing the most common listing type. Beijing, the capital of China and the most famous tourist destination, has a mature online sharing economy accommodation market. Considering the distribution of apartment-type listings, our research was limited to the main urban area within the sixth ring road of Beijing as the case study. The total study area is about 2267 km² and covers the entire central urban area and some of the suburban area (Figure 1).

The sample listing data originate from Inside Airbnb, http://insideairbnb.com/ (accessed on 26 January 2021), a third-party website that compiles data from publicly accessible information on Airbnb.com. Considering the calculation efficiency of MGWR and the availability of listings, we deleted listings with abnormal prices, those unavailable within 30 days, and those with 0 reviews, resulting in 1499 retained listings. Then, we mapped the spatial coordinates of the samples using ArcGIS for further analysis. In addition, the top 20 most popular tourism destinations in Beijing were obtained from CTrip, https://www.ctrip.com/ (accessed on 4 May 2021), one of the most famous tourism websites in China. Data on POIs such as subway stations and sports and leisure facilities were captured through GaoDe Map, https://www.amap.com (accessed on 4 May 2021). GaoDe is one of the top two map service providers in China, incorporates 70 million POIs, covers more than 300 cities in China, and provides an open platform for helping clients obtain their desired location services. In this study, we mainly used the GaoDe polygon searching API to obtain POI data on Beijing.

3.2. Conceptual Framework and Variables

As heterogeneous products, different Airbnb listings have different attributes, and the utility and satisfaction of consumers are also different. Therefore, it is necessary to use the hedonic price model to accurately evaluate the series of attributes that Airbnb listings incorporate. In this study, we propose a conceptual framework for the hedonic pricing of Airbnb listings, which includes four aspects: functional attributes, reputational attributes, host status attributes, and locational attributes (Figure 2). Initially, we selected 16 variables, including maximum number of guests accommodated; number of bedrooms; number of facilities; superhosts and professional hosts; reviews and ratings; the number of restaurants, scenic spots, shopping malls, cultural facilities, and sports and leisure services within 1 km of the listing; and the distance to the CBD, popular attractions, and the nearest subway station. After combining the results of the stepwise regression and theoretical importance, 10 variables were retained. The selection of variables is shown in

Table 1. Model variables.

Category	Variable	Definition	Unit	Source
Dependent variable	price	The daily price of the listing	CNY	Inside Airbnb
Function	accommodates	Maximum number of guests accommodated	no.	Inside Airbnb
	bedrooms	The number of bedrooms	no.	Inside Airbnb
Reputation	reviews	The number of reviews	no.	Inside Airbnb
	ratings	Review scores for overall rating	no.	Inside Airbnb
Host status	superhost	A dummy variable indicating whether a host is a superhost: 1—superhost and 0—not uperhost	-	Inside Airbnb
	professional-host	A dummy variable indicating whether a host has more than two listings: 1—professional host and 0—non-professional host	-	Inside Airbnb
Location	num-sports	The number of sports and leisure service POIs within a 1 km range	no.	GaoDe
	num-hotels	The number of accommodation service POIs within a 1 km range	no.	GaoDe
	dist-hotspots	Sum of distance to 20 popular attractions in Beijing	km	CTrip, GaoDe
	dist-metro	Distance to the nearest subway station	km	GaoDe

.

The hedonic price model determines the hedonic price through a regression equation, so the choice of the specific form of the model function will directly affect the analytical results. The most commonly used functional forms include linear, logarithmic, and semi-logarithmic models. There has been debate concerning which functional form is optimal. Consistent with Wang and Nicolau (2016), Deboosere et al. (2019), and Hong and Yoo (2020), we adopt a linear functional form to describe the hedonic price model of shared accommodation and use the daily price of the listing as the dependent variable [14,32,37]. We also investigated alternate dependent variables, such as the price per room, but we abandoned these variables due to poor performance. The details of the other variables are as follows:

Firstly, in order to test the impact of functional attributes on prices, we created two ordinal explanatory variables: accommodates and bedrooms, which represent the maximum number of guests that the listing can accommodate and its number of bedrooms, respectively.

Secondly, in order to test the impact of reputational attributes on prices, we created two ordinal explanatory variables, reviews and ratings, which represent the number of reviews and review scores of listings, respectively. It is worth mentioning that the mean rating was 96.55, which is very high and supports previous research results on Airbnb’s rating system [33].

Thirdly, in order to test the impact of host status attributes on prices, we created two dummy variables: superhost and professional host. Among them, superhost indicates whether the host has a Superhost badge. A Superhost badge is an award offered by Airbnb to hosts in recognition of their consistently excellent services and guest experiences. In our sample, 57.71% of hosts had a Superhost badge. Professional host is used to measure the professionalism of the host. Following the classification method of Gibbs et al. (2018) and Magno et al. (2018), we define a host who has two or more listings as a professional host [17,58]. In this case, 79.3% of the hosts showed professional characteristics.

Lastly, we chose four variables, num-sports, num-hotels, dis-hotspots, and dis-metro, to represent the locational attributes of listings. num-sports and num-hotels represent the number of sports and leisure service POIs and accommodation service POIs within 1 km of the listing, respectively, and dis-hotspots and dis-metro are two distance variables that were measured based on Euclidean distances. In particular, dis-hotspots measures the distance to multiple destinations. Based on CTrip’s destination rankings, we selected 20 popular attractions in Beijing, assigned weights to each attraction according to the number of reviews, and then weighted them to calculate the total distance from the list to these 20 attractions. dis-metro represents the distance to the nearest subway station, and the average distance is less than 1 km, which means that traffic accessibility may be an important attribute.

In order to avoid model distortion or inaccurate estimation due to the possible strong correlation between variables, the variance inflation factor (VIF) of each variable was calculated to test the multicollinearity of the variables used in the model. The VIFs of all variables were lower than 3.5 and were thus lower than the empirical value of 7.5, indicating that there was no multicollinearity problem in the model and that all of the variables could be treated as independent. The descriptive statistics of each variable are shown in

Table 2. Variable-descriptive statistical analysis.

Variable	Min	Max	Mean	SD	VIF
price	65	3983	374.214	277.859	-
accommodates	1	16	2.785	1.842	2.853
bedrooms	1	11	1.275	0.710	2.848
reviews	10	251	30.017	27.025	1.054
ratings	74	100	96.550	3.738	1.287
superhost	0	1	0.570	0.495	1.290
professional host	0	1	0.790	0.407	1.051
num-sports	0	342	76.90	59.793	3.178
num-hotels	0	273	70.51	60.035	3.439
dist-hotspots	9.443	10.500	9.880	0.263	1.315
dist-metro	0.036	7.338	0.835	0.727	1.090

Note: In order to obtain comparable bandwidths, all variables were standardized using z-value standardization.

.

3.3. Methods

• Ordinary Least Squares

Following the technology roadmap provided by Comber et al. (2020) and Oshan et al. (2020) [59,60], we apply an OLS model to examine the global results and to provide a baseline against which the GWR and MGWR models can be compared. In the OLS model, relationships are assumed to be constant throughout the study region and can be denoted by:

$${\mathrm{y}}_{\mathrm{i}}={\mathsf{\beta }}_0+{\displaystyle \sum }_{\mathrm{k}}{\mathsf{\beta }}_{\mathrm{k}}{\mathrm{x}}_{\mathrm{ik}}+{\mathsf{\epsilon }}_{\mathrm{i}}$$

(1)

where ${\mathrm{y}}_{\mathrm{i}}$ is the observation of the dependent variable at location i, x_ik is the kth variable at location i, ${\mathsf{\beta }}_0$ is the estimated intercept, ${\mathsf{\beta }}_{\mathrm{k}}$ is the corresponding coefficient, ${\mathsf{\epsilon }}_{\mathrm{i}}$ is the error, i = {1, 2, 3, …, 1499}, and k = {1, 2, 3, …, 10}.

• Geographical Weighted Regression

Compared to the OLS regression model, GWR relaxes the spatial stationarity assumption inherent in global models and permits relationships to vary spatially at the same scale. Specifically, the GWR model includes spatially variable parameters based on the coordinates of each observation, $\left({\mathrm{u}}_{\mathrm{i}},{\mathrm{v}}_{\mathrm{i}}\right)$, and is denoted by:

$${\mathrm{y}}_{\mathrm{i}}={\mathsf{\beta }}_{0\left({\mathrm{u}}_{\mathrm{i}},{\mathrm{v}}_{\mathrm{i}}\right)}+{\displaystyle \sum }_{\mathrm{k}}{\mathsf{\beta }}_{\mathrm{k}\left({\mathrm{u}}_{\mathrm{i}},{\mathrm{v}}_{\mathrm{i}}\right)}{\mathrm{x}}_{\mathrm{ik}}+{\mathsf{\epsilon }}_{\mathrm{i}}$$

(2)

where ${\mathsf{\beta }}_{0\left({\mathrm{u}}_{\mathrm{i}},{\mathrm{v}}_{\mathrm{i}}\right)}$ is the estimated intercept, ${\mathsf{\beta }}_{\mathrm{k}\left({\mathrm{u}}_{\mathrm{i}},{\mathrm{v}}_{\mathrm{i}}\right)}$ is the coefficient of local variable k at location i, ${\mathsf{\epsilon }}_{\mathrm{i}}$ is the error, i = {1, 2, 3,…, 1499}, and k = {1, 2, 3, …, 10}.

• Multiscale Geographical Weighted Regression

MGWR is suggested as the default local model specification when studying processes’ spatial heterogeneity and scale [44]. MGWR relaxes the “same spatial scale” assumption of GWR and assigns specific bandwidths to each variable [43]. It is denoted by:

$${\mathrm{y}}_{\mathrm{i}}={\mathsf{\beta }}_{{\mathrm{bw}}_0\left({\mathrm{u}}_{\mathrm{i}},{\mathrm{v}}_{\mathrm{i}}\right)}+{\displaystyle \sum }_{\mathrm{k}}{\mathsf{\beta }}_{{\mathrm{bw}}_{\mathrm{k}}\left({\mathrm{u}}_{\mathrm{i}},{\mathrm{v}}_{\mathrm{i}}\right)}{\mathrm{x}}_{\mathrm{ik}}+{\mathsf{\epsilon }}_{\mathrm{i}}$$

(3)

The meanings of the parameters as the same as those in GWR, and ${\mathsf{\beta }}_{{\mathrm{bw}}_{\mathrm{k}}}$ represents the bandwidth of the k-th variable.

In this paper, we apply MGWR-2.2 software, which was developed by the Spatial Analysis Research Center (SPARC) of Arizona State University to estimate the OLS, GWR, and MGWR, and the maps were made with ArcGIS10.8.

4. Results and Analysis

4.1. Spatial Pattern Analysis of Airbnb Listing Prices

Before constructing the GWR and MGWR models, it was necessary to conduct a spatial dependence test on the listing price in order to decide whether to use the spatial econometric model. Spatial dependence is usually described by spatial autocorrelation, and the global Moran’s I index is an appropriate test statistic [39,61]. Based on ArcGIS software, we calculated the global Moran’s I index of the analyzed listing prices. The result shows that at the significance level of 1%, Moran’s I index is 0.0612, the p-value is 0.00036, and its normal statistic Z value is 3.567447, thereby exceeding the critical value of 2.58, which indicates that there is significant spatial autocorrelation in the listing prices; thus, the spatial econometric model can be used for analysis [41]. However, the global Moran’s I index cannot reveal specific regional heterogeneity. Therefore, we further drew a LISA clustering diagram of the listing prices to analyze their local clustering characteristics (Figure 3).

It can be seen from Figure 3 that the high–high agglomeration areas of listing prices in Beijing are mainly located near the Beijing CBD. The low–low agglomeration areas are mainly distributed outside the third ring road. The high–low agglomeration areas are mainly distributed along the edges of the low–low agglomeration areas, and the low–high agglomeration areas are mainly distributed near the high–high agglomeration areas. Through the above analysis, hypothesis H1 has been verified; that is, in Beijing, China, Airbnb listing prices are spatially dependent.

4.2. Model Comparison

First, we compared the three models using the following four indicators: R², adjusted R², corrected Akaike information criterion (AICc), and sum of residual squares (RSS), and the results are shown in

Table 3. R2, adjusted R², AICc, and RSS of the three models.

Models	R2	Adjusted R2	AICc	RSS
OLS	0.443	0.439	3401.834	835.414
GWR	0.618	0.577	3132.628	573.231
MGWR	0.733	0.69	2757.505	400.085

. The AICc and RSS of MGWR are significantly smaller than those of the OLS model. As a general rule of thumb, when the difference between the AICc of two models exceeds 3, the model with the smaller AICc value has a better fitting effect. In this case, the adjusted R² of the MGWR model is 0.690, which is about 25.1% higher than that of the OLS model; that is, the MGWR model explains 25.1% more variable data than the OLS model, which, again, shows that the goodness-of-fit of the GWR model is significantly higher than that of the OLS model. Similarly, it can also be shown that the MGWR model is superior to the GWR model.

In addition, to analyze the level of residual spatial autocorrelation, we evaluated Moran’s I and its significance for each model, as shown in

Table 4. Residual spatial autocorrelation of three models.

Models	Moran’s I	p-Value	z-Score	Pattern
OLS	0.067	0.000	3.900	Clustered
GWR	0.042	0.015	2.442	Clustered
MGWR	0.014	0.400	0.848	Random

. It can be seen that both the residuals of the OLS model and the GWR model exhibit significant spatial clustering, and only the MGWR residuals are randomly distributed without spatial autocorrelation. The above results show that the spatial non-stationarity in the relationships between the listing attributes and prices cannot be ignored. Compared with the GWR model, the MGWR model delivers a better model fit and reduced residual spatial autocorrelation; therefore, it is more suitable for this study.

4.3. Analysis of Optimized Bandwidths and Scale

Bandwidth is a key concept in the GWR and MGWR models, providing a measure of the “influencing scale” of spatial processes [60]. Compared with GWR, the most obvious advantage of MGWR is that it not only considers the differences in geographical location when estimating the regression coefficient but also the influencing scales of listing attributes on prices through varying bandwidths.

In Figure 4, the blue and orange bars represent the bandwidths for each variable generated by the MGWR model and the standard errors of the parameter estimates, respectively. The black dotted and solid lines represent the average value of all bandwidths obtained by MGWR and the single bandwidth generated by GWR, respectively. We can find that compared with the single optimal bandwidth of 290 generated by GWR, the results of MGWR show that the bandwidths of each property attribute are quite different:

accommodations < Intercept < bedrooms < ratings < reviews < dist_metro < professionalhost = num_sport < num_hotelt < superhost = dist_hotspots.

As we hypothesized in H2, there are multiscale, spatially heterogeneous relationships between Airbnb listing attributes and prices. In general, functional attributes have local scales, reputational attributes have regional scales, and host status and locational attributes have global scales. The bandwidth of accommodations, the intercept, and bedrooms are 45, 49, and 80, respectively, indicating that the relationships between these variables and prices have obvious spatial non-heterogeneity, and their parameter estimates show large spatial differences. The bandwidths of the two reputational attributes of ratings and reviews are 355 and 849, indicating that although the relationships between the reputational attributes and listing prices show spatial non-stationarity, they only differ at regional scales. The bandwidths of the other host status and locational attribute variables (superhost, professional host, num-sports, num-hotels, dist-hotspots, and dist-metro) are close to the sample size of 1499, indicating that they are global variables, and their impact on listing prices is basically stable over the entire study area.

4.4. Analysis of Parameter Estimates

4.4.1. OLS Results

As a baseline model, the OLS model helps us understand the relationship between these price determinants and listing prices across the study area.

Table 5. Summary statistics for OLS parameter estimates.

Variable	Coefficient	SE	t-Value	p-Value
accommodates	0.344 ***	0.033	10.526	0
bedrooms	0.341 ***	0.033	10.442	0
reviews	−0.049 *	0.020	−2.479	0.013
ratings	0.070 ***	0.022	3.178	0.001
superhost	−0.048 *	0.022	−2.190	0.028
professional host	0.028	0.020	1.404	0.160
num-sports	0.221 ***	0.035	6.406	0
num-hotels	−0.093 **	0.036	−2.598	0.009
dist-hotspots	−0.107 ***	0.022	−4.820	0
dist-metro	0.065 ***	0.020	3.231	0.001
Adjusted R2	0.443

Note: ***, **, and * represent significance at the levels of 0.001, 0.01, and 0.05, respectively.

shows the OLS regression results. We found that, at the global scale, variables such as accommodates, bedrooms, ratings, num-sports, and dist-metro have positive influences on listing prices. Reviews, superhost, num-hotels, and dist-hotspots have negative effects on listing prices, and professional host has a positive but nonsignificant effect on listing prices. From the absolute value of the coefficient, the two functional attribute variables, accommodates and bedrooms, have the greatest impact on listing prices, which verifies previous conclusions [10,17,32,56,62].

4.4.2. MGWR Results

The GWR and MGWR models demonstrate superior performance because they account for the spatially varying effects of variables. Since the MGWR model is more effective than the GWR model and can represent the scale-related differences of spatial relationships, here, we only focus on the local estimation of MGWR. The statistical description of each coefficient of MGWR is shown in

Table 6. Summary statistics for MGWR parameter estimates.

Variable	Bandwidth	Mean	Min	Median	Max	Ratio of Significance (%)
intercept	49	0.027	−0.696	0.035	0.833	21.7
accommodates	45	0.355	−1.314	0.298	1.942	46.5
bedrooms	80	0.343	−0.213	0.261	1.148	46.0
reviews	355	0.031	−0.089	0.035	0.143	14.3
ratings	849	−0.038	−0.134	−0.027	0.010	33.0
superhost	1498	−0.038	−0.046	−0.038	−0.03	61.6
professional host	1482	0.022	0.013	0.022	0.030	0.0
num-sports	1482	0.179	0.171	0.180	0.183	100.0
num-hotels	1494	0.036	0.024	0.036	0.038	0.0
dist-hotspots	1498	−0.154	−0.162	−0.154	−0.143	0.0
dist-metro	1004	0.038	−0.012	0.027	0.157	14.3

. The second column of Table 6 shows the bandwidth for each variable; the 3rd–6th columns show the mean, minimum, median, and maximum of the local parameter estimates for each variable, respectively. The 7th column indicates the proportion of significant coefficients (p ≤ 0.05). We find that in some or all Airbnb listings, 7 out of 10 variables are significant as drivers of price, namely, accommodates, bedrooms, reviews, ratings, superhost, num-sports, and dist-metro. Compared with the OLS model, the significances of num-hotels and dist-hotspots show different results. The possible reason for this is that, as a global regression method, OLS generates an equation with fixed coefficients across the study area, which may oversimplify the hedonic price of Airbnb listings or lead to spurious regressions.

Figure 5 shows the significant regression coefficient distribution of the MGWR model. The dotted line in the middle of the violin represents the median value, the upper and lower solid lines represent the inter-quartile range, and the width represents the frequency of the regression coefficient. The wider sections of the violin plot represent a higher probability that the regression coefficient will take on the given value; the thinner sections represent a lower probability.

From the violin plot, it can be seen that the significance parameter estimates for most of the variables show results consistent with the OLS model, where bedrooms, ratings, num-sports, and dist-metro have positive influences on listing prices, while reviews and superhost are negatively related to listing prices. However, intercept and accommodates show changing correlations from negative to positive, and intercept in particular shows an obvious two-level differentiation state. It is worth noting that the variable accommodations has a small number of negative estimated parameters, which means that in some places, accommodations has a negative association with listing prices. This result differs from that of the OLS model and needs to be further explored by visualizing the estimated parameters. It can also be seen that the explanatory power of the different categories of the attributes with respect to the listing prices varies from high to low, as follows: functional attributes, locational attributes, reputational attributes, and host status attributes, among which the functional attributes are the most important determinants of Airbnb listing prices.

Next, we selected the variables that passed the significance test (p ≤ 0.05) and combined them with the visualization function of ArcGIS software to analyze the spatial patterns of the estimated parameters of each variable (Figure 6).

• Spatially varying estimates of intercept

In this study, the estimated parameters of intercept showed significant spatial heterogeneity (Figure 6a). In the Shuangjing Subdistrict, there is a negative relationship between intercept and price, while in the Yansha Commercial Area, there is a positive relationship between intercept and price. Although both areas have locational advantages, the community environment and spatial quality are quite different. Shuangjing Subdistrict was built long ago, its houses are old, and its community environment is relatively poor. On the contrary, Yansha Commercial Area, as one of the most mature foreign-related business and living areas in Beijing, has complete supporting facilities and a comfortable community environment. Thorough MGWR, the intercept was found to be statistically non-zero after normalizing the variables, and the spatial heterogeneity of the parameter estimates identified hot spots of high and low listing prices [60]. Additionally, the local intercepts reflect the influence of different locations on the dependent variable when other independent variables are determined [46,60,62]. We controlled the location factors such as the distance to popular attractions and the nearest subway station; therefore, intercept can reflect the influence of other location factors that are not included in the model. For example, studies have shown that streetscape features have a certain impact on Airbnb listing prices [34], and it is necessary to incorporate these factors into multi-scale analysis in the future.

• Spatially varying estimates of accommodates

As one of the functional attributes of a listing, accommodates is significantly related to the listing price, and its impact shows the greatest spatial difference. The estimated parameters range from −1.314 to 1.942, with a mean of 0.589. Judging by the absolute value of the coefficient, its influence is strong. As shown in Figure 6b, in 94% of the statistically significant samples, the accommodates variable positively affects the listing price; the more guests the listing can accommodate, the higher the price, which is consistent with previous research results [17,32,55]. However, the results also found that accommodates negatively affects listing prices near Beijing Capital Airport. The possible reason is that, compared with other regions, small-capacity listings near the airport are more popular and there is insufficient demand for large-capacity listings, resulting in lower prices. The bandwidth of accommodates is 45, accounting for 3.3% of the sample size, which is a local scale, much lower than the influencing scales of the other variables. On average, this scale is close to the street scale of Beijing. This shows that in Beijing, Airbnb prices are very sensitive to the accommodates variable and that there is obvious spatial heterogeneity in the relationship between accommodates and listing price, which remains stable only in the street range. Once this scale is exceeded, the coefficients change dramatically.

• Spatially varying estimates of bedrooms

It can be seen from Figure 6c that the estimated parameters of bedrooms are significant and all positive, varying from 0.237 to 1.148, with a mean of 0.607. This shows that the greater the number of bedrooms, the higher the price. The reason is that although Airbnb represents a new type of business model, it is essentially an accommodation service, and the basic facilities and functions of the listing are still the determining factors of consumers’ willingness to pay and the price of the listing. Therefore, the number of bedrooms has a significant positive effect on the price of listings, which is consistent with existing research results [10,14,15,17,49]. The bandwidth of bedrooms is 80. This is very small, indicating that bedrooms affects listing prices at a local scale and exhibits large spatial heterogeneity.

• Spatially varying estimates of reviews

As shown in Figure 6d, about 14.3% of the listings’ estimated reviews parameters are significant, with a mean of −0.077. This result is consistent with the conclusions of some other scholars who indicated a negative relationship between the number of reviews and the listing price [17,32,34,36,52,53,61]. The possible reason for this negative relationship might be that, for low-priced listings, more reviews precipitate greater information transparency; therefore, it is harder for hosts to overprice, or perhaps more reviews simply indicate a higher demand for cheaper listings [32]. Additionally, for high-priced listings, researchers have suggested that they often receive fewer bookings and fewer reviews [63,64]. As a result, it seems that the more reviews there are, the lower the listing price. However, in 85.7% of the listings, there was no significant correlation between the listing price and the number of reviews, which is consistent with the research conducted by Voltes-Dorta and Sánchez-Medina (2020) [15]. Combined with the absolute values of the coefficients, we find that the number of reviews has a limited impact on listing prices and is not a major factor with respect to hosts’ determination of prices. Additionally, the bandwidth of reviews is 849, which is a regional scale, indicating that there is a certain degree of spatial heterogeneity in the relationship between reviews and prices, but it can remain stable over a relatively large area.

• Spatially varying estimates of ratings

The estimated parameters of ratings are only statistically significant in 33% of the listings we examined and are mainly concentrated around Beijing’s CBD. In this area, the ratings variable positively affects listing prices, and the coefficients range from 0.074 to 0.143, with a mean of 0.109 (Figure 6e). This positive effect is consistent with the research conclusions of Wang and Nicolau (2017) and Deboosere et al. (2019) [32,49]. However, in most of the listings, the impact of ratings on prices was not statistically significant, which is consistent with the research conclusions of some scholars [14,65]. The possible reason for this is that compared with dissatisfied guests, satisfied guests are more likely to write reviews; thus, extremely high ratings lose their informational value [65]. The bandwidth of ratings is 355, indicating that ratings affect listing prices at a regional scale.

• Spatially varying estimates of superhost

As is shown in Figure 6f, the estimated parameters of superhosts are statistically significant in 61.6% of the listings we examined, varying from −0.046 to −0.037 and with a mean value of −0.041. This shows that Superhost status has a negative effect on the price of most listings, although the effects are relatively weak. Hosts are often considered to be a key component of the accommodation-sharing market, and having a Superhost badge means that a host is very experienced and has good credibility, which is a reflection of the host’s professionalism and personal brand. Previous studies have shown that Superhost status can bring a significant price premium [14,32]. However, our study came to the opposite conclusion. This could be because this data set was acquired on 26 January 2021. Due to COVID-19, the tourism accommodation market was in a downturn. To avoid losses from vacant properties, superhosts may charge lower prices [17]. The bandwidth of superhost is 1498, which is a global scale, indicating that the relationship between superhost and listing prices remains stable throughout the study area.

• Spatially varying estimates of num-sports

Another indicator that has a greater impact on Airbnb listing prices is num-sports. Figure 6g shows that among all the listings, the estimated parameters of num-sport are significantly positive, ranging from 0.171 to 0.183. As one of the most important elements of urban tourism [66], sports and leisure services positively affect listing prices. The more sports and leisure services there are around the listing, the higher the listing price. This is because, first of all, the more sports and leisure services there are around the listing, the greater the utility and satisfaction consumers can receive from the listing. In addition, the more sports and leisure services there are, the more potential consumers there are, which has a positive impact on the price, in line with consumer theory and the law of supply and demand. This is consistent with the conclusions of house price research [67,68]. For example, Raymond and Love (2000) pointed out that private clubs, swimming pools, gymnasiums, and various sports amenities tend to drive up house prices [68]. Judging from the absolute value of the coefficients, the impact of num-sport on listing prices is only lower than that of accommodates and num-bedrooms, indicating that the number of sports and leisure services around listings is another important source of consumer utility, and consumers are willing to pay higher prices for abundant sports and leisure services. The bandwidth of num-sports is 1482, which is a global scale, indicating that the impact of the number of surrounding sports and leisure services on listing prices is basically the same over the entire study area.

• Spatially varying estimates of dist-metro

As shown in Figure 6h, the estimated parameters of dist-metro are statistically significant only in the northwestern part outside the fourth ring road. This could be because the subway stations are densely distributed in the main urban area, and with the popularization of online ride-sharing and shared bicycles, travel modes are gradually becoming diversified and tourists’ dependence on the subway is constantly decreasing. In this study, the distance to the nearest subway station had a positive impact on listing price, and the Huitian area was the most obviously affected area. This positive effect is similar to that observed by Deboosere et al. (2019) in New York and Gyódi and Nawaro (2021) in Rome [16,49]. The likely reason for this is that Airbnb’s guests value the traveling time or convenience more than the distance to a subway station. Taking Tiantongyuan North Station—which is known for its traffic congestion—as an example, the out-of-station waiting time can be up to 15 min, and the maximum queue length can exceed 3000 passengers, which brings significant travel costs [69]. In addition, the congested bus shuttles and the disturbing traffic noise also worsen guests’ experiences in the surrounding listings. Therefore, it is difficult for hosts to gain enough of a competitive advantage from this factor to warrant a price premium, resulting in lower listing prices in areas closer to subway stations.

5. Conclusions and Discussion

Based on Beijing’s Airbnb listing data and POI data from 2021, we constructed classic OLS and GWR models and an emerging MGWR model to analyze the multiscale effects of different hedonic attributes influencing Airbnb listing prices and came to the following conclusions:

(1)

In Beijing, China, Airbnb listing prices are spatially dependent.

(2)

The explanatory power of different types of attributes for listing price varies from high to low, as follows: functional attributes, locational attributes, reputational attributes, and host status attributes. Functional attributes are the most important determinants of the listing price of an Airbnb accommodation, which means that functional attributes are the main source of consumer utility.

(3)

There are multiscale, spatially heterogeneous relationships between Airbnb listing attributes and prices. The variables of functional attributes have relatively small bandwidths (accommodates 45; bedrooms 80), which shows that functional attributes affect listing prices at local scales and their relationships show obvious spatial heterogeneity. Second, the variables of reputational attributes have larger bandwidths (reviews 355; ratings 849), which can be viewed as regional scales. This means that although the impacts of reputational attributes on listing prices are also spatially heterogeneous, the coefficients remained stable regionally. Finally, the bandwidths of the host status attribute and locational attributes are close to the sample size (superhost 1498, professional host 1482, num_sport 1482, and dist-metro 1004), which are global scales, meaning that the effects of these variables on prices remain stable across the entire study area with negligible spatial variation.

These conclusions support the previous hypotheses (H1 and H2).

(4)

From a methodological perspective, this study empirically proves the validity of the MGWR model. Compared with OLS regression and GWR, MGWR improves overall modelling ability by introducing the concept of multiscale analysis, and this is consistent with the results of previous studies [14].

The most important contribution of this study is its deeper exploration of the heterogeneous relationships between hedonic attributes and Airbnb listing prices. Compared with Hong and Yoo (2020)’s study, we find that in the three cities , New York, Los Angeles, and Beijing, the functional attributes all affect listing prices at local scales, while locational attributes all affect prices at global scales. However, the influencing scales of the host status attributes and reputational attributes are different in these cities. This shows that the influencing scales of pricing variables in different cities have both commonalities and differences. From this perspective, each city’s hedonic pricing research has unique value and importance.

6. Theoretical and Practical Implications

To the best of our knowledge, this is the first multiscale analysis of the spatial heterogeneity in the pricing mechanism of China’s sharing economy accommodation market. We developed a conceptual framework that incorporates four categories of hedonic attributes—functional attributes, host status attributes, reputational attributes, and locational attributes—and used the perspective of spatial heterogeneity to investigate the multiscale spatial effects of various hedonic attributes on Airbnb listing prices. This constitutes a theoretical contribution to the study of hedonic pricing in sharing economy accommodation and related topics.

Our study also makes methodological contributions. By comparing the OLS, GWR, and MGWR models, it was found that the MGWR model provides better explanatory power and more reliable results. This empirically demonstrates the effectiveness of the MGWR approach, which can serve as a better option for in-depth research on the hedonic prices of sharing economy accommodation.

Our conclusions are expected to help Airbnb hosts develop location-based pricing strategies, which have important practical significance for hosts’ management and the sustainable development of the industry. Firstly, regardless of the location, hosts should prioritize the basic lodging needs of their guests, and they should set their prices based on the prices of nearby listings with similar functional attributes. In addition, hosts can increase the price if there are more sports and leisure facilities near the listing. They can also adjust listing descriptions to more effectively highlight advantages, such as advantages regarding traveling times or convenience instead of the distance to the subway station. Lastly, potential hosts who plan to enter the sharing economy accommodation market should avoid investing in large-capacity listings near Beijing Capital Airport.

However, this study also has limitations. First of all, our research sample was limited to Beijing, China. Due to the differences in price mechanisms in different cities, these research results may not be suitable for generalization to other cities. Second, the sample data in this study are cross-sectional data that do not consider the time factor; thus, future research should collect listing prices over time and use panel data to improve the robustness of results. Third, the listing price of sharing economy accommodation is the result of the combined effects of multiple factors. In the future, socioeconomic characteristics and more locational attributes should be taken into consideration to expand and deepen our understanding of the pricing of Airbnb listings.