Examining the Effects of Built Environments and Individual Characteristics on Commuting Time under Spatial Heterogeneity: An Empirical Study in China Using HLM

Abstract

A large number of studies have provided evidence regarding the factors that influence commuting time. However, few studies have explored such effects in the context of considering spatial heterogeneity across cities, which limits the generalizability of the findings. This study addresses this gap by utilizing a dataset of 113 cities in China across the years 2014, 2016, 2018, and 2020. A two-level hierarchical linear model (HLM) was developed to explore the combined effects of city-level and individual-level factors on commuting time by constructing a nested “city-individual” relationship. The results show that (1) built environments at the city level significantly impact commuting time; (2) a non-linear association between population density and commuting time (U-shaped relationship) was identified, as well as between the number of buses and commuting time (inverted U-shaped relationship); (3) the urban construction land area and road area per capita exert negative effects on commuting time; (4) the impacts of individuals’ jobs–housing balance, travel allowances, and education on commuting time vary across cities. These findings might contribute to optimizing the design of a built environment, addressing the challenge posed by longer commuting times, and providing a better understanding of the effects of individuals’ characteristics on commuting time while considering the inherent differences across cities.

1. Introduction

Commuting time holds a significant role in people’s daily routines, representing the duration spent traveling between home and work in the context of daily time use [1]. In the context of rapid urbanization in developing countries, commuting time is increasing [2]. However, excessive commuting time can have significant economic, social, and environmental consequences. For example, longer commuting time hampers per capita income growth and reduces productivity [3]. It also impedes social interactions among residents, leading to a decline in social and economic vitality [4]. Moreover, commuting accounts for a large proportion of urban transportation demand and is closely linked to urban carbon emissions [5]. Additionally, it has detrimental effects on both physical and mental health, as well as the well-being of individuals [6,7]. Therefore, it is crucial to investigate the factors that influence commuting time in order to mitigate its continual increase.

A large number of studies have provided evidence of the impact of various factors on commuting time, with particular attention on built environments and individual characteristics [8,9,10,11]. For example, researchers have demonstrated that built environments, such as a built-up area, population density, and land-use mix, significantly impact an individual’s commuting time [12,13]. However, few studies have examined such effects in the context of spatial heterogeneity [14,15,16,17].

Spatial heterogeneity refers to the varying impact of the same factors across different spatial scales or geographical locations [13]. It is crucial to consider spatial heterogeneity when examining the association between built environments as well as individual characteristics and commuting time, as this relationship may vary from city to city. Failing to account for spatial heterogeneity may lead to ecological fallacy, where the global association between two variables derived from all regions may contradict the local association between the same variables derived from individual regions [18]. As suggested by scholars in the field, the association between environmental attributes and commuting behavior is not globally consistent but rather varies and can even reverse across different regions [13,18]. Therefore, when it comes to comprehensive study across multiple regions, it is necessary to control for the interference brought by such spatial heterogeneity. This approach provides accurate and context-specific insights into the factors influencing commuting time while also ensuring the provision of a scientific basis for urban planning and construction.

This study aims to explore the factors impacting commuting time in the context of considering the spatial heterogeneity among cities. A hierarchical linear model (HLM) is used to capture such effects by constructing nested relationships between cities and individuals. Specifically, this study addresses the following two key research questions: (1) how do built environments at the city level affect individuals’ commuting time? (2) How do individual characteristics affect commuting times after controlling for spatial heterogeneity across cities? A 4-year dataset of 113 cities in China is employed to provide empirical insights into the above questions.

The remainder of this article is structured as follows. Section 2 presents a literature review on the association between built environments and individual characteristics with commuting time. Section 3 describes the data, methods, and variables used in this study. Section 4 describes the model results. Section 5 discusses the findings and shortcomings of the current study. In Section 6, we present the conclusions and corresponding recommendations and raise some questions for further research.

2. Literature Review

2.1. Impacts of Built Environments on Commuting Time

A built environment is a human-crafted environment that emerges through artificial design modifications to natural conditions, primarily related to patterns of land use, urban design, and transportation systems [19]. It constitutes a multiple constructs concept, which has been described by a variety of interconnected spatial variables [20]. An increasing number of studies have examined the correlation between built environments and commuting time [15,21,22,23]. The typical elements of built environments are outlined in “5D” framework, which include density, diversity, design, destination accessibility, and distance to transit [24].

Density expresses the concentration of population, housing, employment, buildings, activities, and other elements within a spatial unit [25]. Scholars have identified two different linear relationships between density and commuting time. Some scholars have suggested that under certain conditions, a high population density and a high residential density are associated with shorter commuting time [26]. However, others have revealed that an increase in urban population density often leads to higher traffic congestion, which, in turn, is expected to increase commuting time [10,14,27]. The non-linear relationship between density and commuting time has also raised concerns [22,28]. For example, higher employment density may reduce commuting times for transit commuters living in non-downtown urban areas and new town areas, but it may increase commuting times for those living in downtown areas [22].

Diversity reflects the degree of mixed land use within an area. Highly mixed land uses are believed to promote proximity between residences and workplaces, thus having positive impacts on reducing commuting time [8,13]. However, it has also been argued that increasing the degree of mixed land use in residential areas tends to reduce the chance of driving and, thus, increases commuting time in general [12].

The concept of design consists of various aspects, such as structural design, road network features, and community design, which distinguish between pedestrian-oriented and motorized-oriented environments [25]. The impact of urban spatial structure and road network characteristics on commuting time has been a focal point of research. Studies have found that commuting time varies according to the nature and stage of development of the polycentric structure [29,30]. Additionally, a higher road area per capita has been found to contribute to a reduction in commuting time [14,31].

Destination accessibility expresses the ability to reach activities or locations by modes of travel or the number of destinations that can be reached in a given travel time [25]. Several studies have highlighted the association between accessibility and commuting time [9,32]. It has been observed that shorter commuting times are associated with higher accessibility to public transit or private vehicles [10,23]. However, it is worth noting that higher public transit accessibility has also been found to be significantly linked to longer commuting time for those who use non-motorized vehicles [15].

Distance to transit refers mainly to the distance to the nearest public transit station [25]. Research has consistently shown that living closer to a public transit station is associated with shorter commuting time [33,34]. Furthermore, the improvement of a rail transit system has been found to have positive effects on reducing commuting time to employment centers [35].

In addition, other built environments and their interrelated spatial variables, such as urban size and urbanization rate, have also been identified as influential factors affecting commuting time. For example, research has shown that commuting time tends to increase with urban size [1,16,36]. Similarly, the larger the built-up area, the longer the commuting time for residents [31]. Additionally, urbanization can induce various environmental transformations [37], potentially resulting in prolonged commuting times for residents in cities with high urbanization rates compared to those in cities with low urbanization rates [38]. Overall, the previous findings highlight the nuanced relationship between built environments and commuting time.

2.2. Impact of Individual Characteristics on Commuting Time

The typical individual characteristics affecting commuting time include an individual’s jobs–housing relationship, work characteristics, and socio-demographic attributes. Specifically, individual jobs–housing relationships consist of elements such as jobs–housing balance, commuting distance, commuting mode, etc. Research has consistently shown that jobs–housing imbalances are the most important determinant for longer commuting time [39]. In particular, based on the data from a survey of employees at their workplaces in the sub-centers of Beijing, a study has revealed that the jobs–housing balance has a more significant impact on workers’ commuting times than socio-demographic characteristics [40]. Commuting distance is found to be roughly proportional to commuting time [32,36]. For commuting modes, some literature has demonstrated that commuting by public transit tends to take longer in terms of travel time than private car travel [10,14,17]. This discrepancy is attributed to factors such as the spatial and temporal arrangement of services (e.g., bus schedules and predetermined routes and stops), the interchanges between different travel modes, as well as the ease of walking access to and from public transit stations [41]. However, it is important to note that the benefits of private car commuting have gradually diminished as escalating private car ownership has led to persistent traffic congestion and, subsequently, lower travel speeds [41].

Work characteristics, such as working hours and allowances, are additional factors that can affect commuting time. Research has shown that there is a positive correlation between working hours and commuting time [42]. Moreover, scholars have also found an inverted U-shaped relationship between commuting time and work duration [1,43]. Furthermore, the financial incentives provided by employers may influence commuters’ behavior [44,45]. For example, employer-provided allowance, such as travel allowance or parking fees, can significantly reduce solo commuting trips by more than 30% [44].

Socio-demographic attributes play a significant role in determining commuting time among individuals, as evidenced by a substantial body of literature. According to the household responsibility hypothesis, women who bear the bulk of household maintenance duties (e.g., child care) tend to choose shorter commuting times [27,46]. Age has been found to exhibit varying effects on commuting time. It was found that older workers are more likely to have longer commuting times than those under age 25 [42]. At the same time, other studies have revealed that younger commuters (under 35 years old) were less sensitive to the cost of time and tended to have longer commutes [47]. Moreover, it has also been found that there is a non-linear relationship between age and commuting time, with middle-aged commuters more inclined to have the longest commuting time [48]. Education level has also been found to be positively associated with commuting time [17,49]. Occupations exhibit distinct effects. Individuals who work in the private sector are more likely to accept longer commuting times due to the concentration of their jobs in specific locations compared to their more evenly distributed residential areas [50], and compared to low-skilled professional employees, mid- to high-level professional employees are more prone to have higher commuting times [17]. Income has been recognized as an important factor affecting commuting time. A study conducted using data from Brazil between 1992 and 2009 has found that workers in the poorest 10% of the population spend, on average, 20% more time commuting compared to the wealthiest 10% [51]. This can be attributed to low-income individuals facing spatial constraints in terms of accessing a workplace and their need for affordable accommodation [52]. However, other scholars have made different observations, revealing that higher-income residents in China have longer commuting times due to the fact that they are more likely to drive to workplaces [14,53]. Additionally, household attributes, such as married status and having children, are considered potential factors influencing commuting time [50].

To summarize, while previous studies have examined the relationship between built environments, individual characteristics, and commuting times, few have shed light on their impacts within the context of spatial heterogeneity across cities. Overlooking spatial heterogeneity across cities can lead to inconsistent parameter estimation, as a single linear model only “averagely” captures the global relationship rather than the local part of the relationship [54]. In recent years, there has been a growing recognition of the importance of spatial heterogeneity in commuting studies, with findings indicating that the effects of factors vary across different spatial units [55,56]. Furthermore, the majority of the literature focuses on commuting issues within individual cities, drawing conclusions based on the unique characteristics of those specific cities. As a result, the insights gained may not directly apply to the broader challenges faced by a wider range of cities. In order to address these gaps, this study develops a two-level HLM to explore the factors affecting commuting time while considering spatial heterogeneity across 113 cities in China. Particularly, we focus on the effect of built environments on commuting times by employing a dataset of 113 cities in China across the years 2014, 2016, 2018, and 2020.

3. Methods

3.1. Data Collection

The commuting behavior data used in this study were obtained from the China Family Panel Studies (CFPS) survey dataset (http://www.isss.pku.edu.cn/cfps/, accessed on 20 June 2022). The CFPS is designed to capture and track various aspects of social, economic, demographic, educational, and health changes in China by collecting data at the individual, household, and community levels, making it a valuable resource for academic research and public policy analysis. The CFPS data are collected on a biennial basis. They cover 25 provincial administrative areas across the country, providing a representative sample of the population.

This study utilizes the dataset spanning four years (2014, 2016, 2018, and 2020), which was carefully selected by excluding the non-working sample and those with missing data. Ultimately, the dataset employed in this study consists of 12,100 valid samples obtained from 3025 individuals across 113 cities in China. These 113 cities belong to 25 provinces in China (accounting for 73.53% of all provinces), covering the eastern, central, and western regions of China. These cities exhibit significant variations in terms of their characteristics. Firstly, the 113 cities are located in various economic regions of China, and at different levels of economic development. China’s classification into four economic regions (the eastern, central, western, and northeastern economic regions) offers a systematic framework that delineates the socio-economic disparities across the country and aids in formulating region-specific developmental strategies [57]. In the current study, the cities are distributed as follows: a total of 37 cities are in the eastern economic region, 35 in the central economic region, 24 in the western economic region, and 17 in the northeastern economic region. Moreover, these cities exhibit varying GDP levels, ranging from the highest per capita GDP of CNY 165,851 in Wuxi City to the lowest at CNY 17,430 in Dingxi City in 2020 [58]. Second, these cities have distinct urbanization rates, with the highest being 92.15% in Dongguan City and the lowest at 36.18% in Longnan City in 2020 [58]. Lastly, there are disparities in public transportation and infrastructure development among the 113 cities. For instance, in 2020, Beijing had 23,948 buses, while Puer City had 98 [58], and the road area per capita varied from 4.76 square meters in Shanghai to 35.47 square meters in Xuancheng City [59]. This diverse representation allows for a comprehensive exploration of commuting dynamics.

Figure 1 illustrates the distribution of the 12,100 samples across 25 provinces included in the study. Among them, Henan Province has the largest sample size, while Guizhou Province has the smallest sample size (Figure 1).

In contrast to the cross-sectional data, the current study uses panel data to enhance the modeling of residents’ travel behavior and achieve more accurate predictions of individual outcomes by mixing inter-individual differences and intra-individual dynamics [60,61].

Table 1. Socio-demographic characteristics of individuals (n = 3025 in each year).

Attributes	Distribution	Percent (%)
		2014	2016	2018	2020
Gender	Male	54.64	54.64	54.64	54.64
	Female	45.36	45.36	45.36	45.36
Age	25 and below	10.91	5.82	2.84	0.93
	26–35	27.37	27.31	25.06	22.28
	36–50	45.29	44.17	44.86	42.55
	51–65	15.44	21.19	24.99	30.25
	65 and above	0.99	1.52	2.25	4.00
Education	Elementary school and below	54.02	53.62	34.18	32.93
	Middle school	23.14	23.27	27.37	27.04
	High school/technical school	13.92	14.15	16.69	17.12
	Associate degree/bachelor’s degree and above	8.93	8.96	21.75	22.91
Marital Status	Married	87.17	89.02	90.28	90.38
	Other status	12.83	10.98	9.72	9.62
Occupation	Working in private enterprise/self-employed	40.69	40.56	41.36	40.36
	Others	59.31	59.44	58.64	59.64
Perceived income level	Very low	14.41	16.23	8.56	8.86
	Low	26.12	28.46	19.50	18.21
	Middle	51.47	47.34	54.58	56.56
	High	6.94	6.21	12.07	11.37
	Very high	1.06	1.75	5.29	4.99

presents the socio-demographic characteristics of individuals across the years 2014, 2016, 2018, and 2020. Notably, there exist slight variations in the socio-demographic characteristics across these years, with the most recent year, 2020, exhibiting the following key features. The proportion of males to females (54.64% vs. 45.36%) of the respondents was close. Young and middle-aged individuals between the ages of 26 and 50 made up the largest proportion of respondents (64.83%). Only about 23% of the respondents had a higher level of education (an associate degree or a bachelor’s degree). The majority of respondents were married (90.38%). In terms of occupation, 40.36% of respondents worked in private enterprises or were self-employed. More than half of the respondents perceived their income to be at the middle level in their respective cities (56.56%).

3.2. Construction of Hierarchical Linear Model (HLM)

Traditional statistical models (e.g., linear regression) often treat the spatial variables derived from aggregated surveys and the individual attributes from non-aggregated surveys in the same way. However, this approach can lead to unrealistic statistical relationships. To overcome these limitations, the hierarchical linear model (HLM) is employed, which effectively handles stratified data by linking the macro-level and micro-level components [62]. The HLM is particularly suitable for analyzing nested data, as it identifies the regression relationship between factors and output variables at different levels and captures interactions across nested data structures. In addition, unlike a single linear model, which assumes spatial homogeneity in the influence of the factors, HLM considers inter-group variations and provides a means to account for spatial heterogeneity [63,64].

Therefore, this paper utilizes the HLM to construct the nested relationship between cities and individuals, aiming to understand the influence of various factors on commuting time within the context of the spatial heterogeneity between cities. As shown in Figure 2, this study postulates that commuting time is influenced by both city-level and individual-level variables.

Specifically, a two-level HLM model has been constructed as follows:

Level 1: Y_ij = β_0j + β_1jX_ij + r_ij

(1)

Level 2: β_0j = γ₀₀ + γ₀₁H_j + μ_0j

(2)

β_1j = γ₁₀ + μ_1j

(3)

Level 1 and Level 2 represent the individual level and city level, respectively. i represents individual-level units (i = 1, 2, …, 12,100), and j represents city-level units (j = 1, 2, …, 452). Y_ij represents the commuting time of sample i in the j group, X_ij represents individual variables (including jobs–housing relationship, working characteristics, and socio-demographic attributes) of sample i in group j. The intercept β_0j represents the average commuting time in group j. β_1j represents the regression coefficients associated with each individual variable predictor, which explains the variation in commuting time across all samples in group j, and r_ij is the random error of sample i in group j. H_j represents the variables at the city level in group j. γ₀₀ is the grand mean of commuting time across all groups; γ₀₁ represents the regression coefficients of H_j and refers to the fixed slopes of the city-level variables, explaining the effect of city-level factors on commuting time; μ_0j is the error term that measures the variances in β_0j. γ₁₀ is the grand mean of the slopes of X_ij across all city groups; μ_1j is the error term that measures the variances of β_1j.

When the Level 2 equations are substituted into the Level 1 equation, we get a mixed model. The results of the mixed model include fixed effects and random effects [65]. The fixed effect refers to the effect of factors influencing the average slope or intercept, whereas the random effect refers to the reliability of the unexplained variability in the slopes and intercepts [66]. That is, the fixed effects present the effects of variables (individual characteristics, built environments, etc.) on commuting time without considering the spatial heterogeneity across cities, whereas the random effects indicate the impact of individual characteristics on commuting time varying across cities. If the random effect of a variable is significant, it means that its slope is affected by the predictors of the city level; therefore, its impact on commuting time is different across cities.

3.3. Variables

In this paper, commuting time is taken as the dependent variable, denoting the one-way duration respondents spend on commutes. These data are derived from the CFPS dataset. As indicated in

Table 2. Commuting time distribution (n = 12,100).

Variables	Distribution	Percent (%)
Commuting time (One-way; mins)	≤15	58.11
	16–30	29.12
	31–45	4.88
	46–60	5.55
	>60	2.35

, the majority of the sample (87.23%) had a commuting time of less than 30 min, while 7.9% had a commuting time of over 45 min, and 2.35% spent more than 60 min commuting. In China, a commuting time of 45 min or less is generally considered as an important indicator of a good quality of life for urban residents, while a commuting time exceeding 60 min is considered extreme commuting [67]. The 2022 Report on Commuting Monitor for Chinese Major Cities has revealed that about 24% of commuters in 44 major Chinese cities had not yet achieved a commuting time within 45 min, and 13% of commuters had a commuting time exceeding 60 min based on the analysis of big data from internet maps and mobile phones [67]. Although the proportion of samples with a commuting time of over 45 min in the present study was lower than it of the above-mentioned report, the issue of commuting time still requires significant attention.

Independent variables include Level 1 and 2 variables (Figure 2). Their detailed variables and definitions are provided in

Table 3. Definition of independent variables.

Variables	Definition
Level 2: Variables in city level
Urban Built Environment Factors
PD	Population density of a city each year (1000 persons/km2)
PD2	Square of population density of a city each year
CLA	Area of urban construction land of a city each year (km2) (in the model, it is divided by 1000 to reduce the differences between the data ranges of the initial variables to diminish the bias of the results.)
PCRA	Road area per capita in a city each year (m2)
Bus	The number of buses in a city each year (100 vehicles)
Bus2	The square of the number of buses in a city each year (in the model, it is divided by 1000 to reduce the differences between the data ranges of the initial variables to diminish the bias of the results)
RTS	The number of rail transit stations in a city each year
Taxi	The number of registered taxis in a city each year (1000 vehicles)
UR	Urbanization rate: refers to the percentage of urban resident population in the total population each year
Economic Attributes
GDPPC	GDP per capita of a city each year (RMB 10,000)
Level 1: Variables in individual level
Jobs–housing Relationship
Jobs–housing balance	Jobs–housing balance of individuals each year, with values assigned according to the location of an individual’s residence and workplace. 1 = an individual’s residence and workplace are located within one town (jobs–housing balance); 0 = an individual’s residence and workplace are located in different towns (jobs–housing imbalance)
Working Characteristics
Weekly working hours	The specific working hours per week.
Travel allowance	Travel allowance each year. 1 = employer provides travel allowances; 0 = otherwise
Socio-demographic Attributes
Gender	1 = male; 0 = female
Age	Age in each year
Age2	The square of an individual’s age in each year (in the model, it is divided by 1000 to reduce the differences between the data ranges of the initial variables to diminish the bias of the results)
Education	Education level in each year. 1 = elementary school and below; 2 = middle school; 3 = high school/technical school; 4 = associate degree/bachelor’s degree and above
Marital status	Marital status in each year. 1 = married; 0 = other status
Occupation	Occupation in each year. 1 = working in private enterprise/self-employed; 0 = others
Perceived income level	Respondents’ subjective evaluation of one’s income level in their cities. 1 = very low; 2 = low; 3 = average; 4 = high; 5 = very high

. Specifically, this study selected a set of nine extensively recognized urban built environment indicators as the independent variables, alongside an economic indicator serving as a control variable at Level 2, as expounded in Table 3. The selection process was underpinned by the guiding principles of the established “5Ds” framework, the availability of data, and insights gleaned from existing research [14,21,31]. The comprehensive inclusion of these 10 indicators was predicated upon their demonstrable significance in molding commuting behavior [7,16,26,38]. The data of these 10 variables were obtained from the statistical yearbooks of each province or city.

Individual variables at Level 1 spanning the domains of jobs–housing relationship, working characteristics, and socio-demographic attributes were selected, guided by both the data comprehensiveness within the CFPS dataset and the existing scholarly insights [39,40,42,43]. Notably, the relationship between jobs–housing balance and urban commuting has garnered substantial research attention [15,40,68]. Work-related characteristics, such as weekly working hours and travel allowance, have frequently been used for interpreting commuting time [1,43,44,45]. Meanwhile, socio-demographic attributes have consistently emerged as the pivotal determinants influencing commuting time [9,42,50]. These individual variables were obtained from the CFPS dataset.

4. Results

This section presents the effects of city characteristics and individual variables on commuting time. All of the models were conducted by using HLM 6.08. The parameters of the HLM are estimated by using the maximum likelihood estimation method. The model results include fixed and random effects. The significance level is set at p = 0.05.

First, a null model was created using only the dependent variable to assess the suitability of the data for HLM analysis. Then, the variables from Level 1 were added to build a random coefficient model. This model was then refined by removing insignificant variables to create a parsimonious model, allowing us to examine the relationship between individual characteristics and commuting time. The hypotheses tested include the variation in the built environment across cities (random intercept) and the variation in the relationship between individual characteristics and commuting time across cities (random slope). This helped us to confirm the correlation between the dependent and the explanatory variables under the HLM. Finally, the variables from Level 2 were introduced to create a full model. Similarly, a parsimonious model was derived by eliminating the insignificant variables, enabling us to explore the impact of the explanatory variables (at Level 2) on the dependent variable. The aim is to examine the combined effects of built environments and individual characteristics on commuting time under the same assumptions [69].

4.1. Construction and Results of Null Model

The null model, which has no predictor variables at either the individual level or the city level, was first constructed to examine whether individual commuting time would vary significantly at the city level. The equation of the null model is as follows:

Individual level:

Y_ij = β_0j + r_ij

(4)

City level:

β_0j = γ₀₀ + μ_0j

(5)

Based on the null model, the coefficient of intra-class correlation (ICC) was obtained to assess the appropriateness of using a multi-level model [70]. The equation of the ICC is as follows:

ICC = σ²_inter-group/(σ²_inter-group + σ²_within-group)

(6)

ICC is an ANOVA-based correlation measure. It represents the ratio of the between-group variance to the total variation and helps determine the extent to which the differences between groups influence the dependent variable. An ICC value below 0.059 is considered a low intra-group correlation, 0.059 to 0.138 is a moderate intra-group correlation, and an ICC value of 0.138 or higher indicates a high intra-group correlation [71,72].

Table 4. Results of null model.

Fixed Effect	Coefficient	SE	DF	T-Ratio	p-Value
Average integration level	18.072	0.324	451	55.804	0.000
Random Effect	Variance	SD	DF	Chi-square	p-Value
City-level effect	27.763	5.269	451	1640.294	0.000
Individual-level effect	385.994	19.647

reveals the significant findings for both the fixed and random effects at the 0.1% level, indicating the differences in the average commuting time between city groups and the significant variations in commuting time among individuals between these groups. Furthermore, the within-group variance of commuting time is 385.994, which represents the degree of difference in the commuting time of the samples within the groups of the cities. This indicates that there is considerable variation in the commuting time among the samples within each city. The inter-group variance of commuting time is 27.763, which represents the degree of difference in commuting time between the groups of the cities. This suggests that there are distinct variations in commuting time across cities. Therefore, the ICC value is 0.067, indicating that 6.7% of the variation in commuting time can be attributed to city-level factors. Given that the ICC in this study is within the moderate range, it suggests that the differences between city groups significantly impact commuting time. Consequently, the data used in this study are appropriate for conducting HLM analysis, as the variability at the city level is meaningful and justifies the consideration of multi-level effects.

4.2. Construction and Results of Random-Coefficient Model

The random coefficient model was constructed to estimate the random effects, which captured the individual differences in the outcomes and rates of change, and was used to conduct inferential tests on the reliability of these individual differences [66]. Only the individual-level variables from Level 1 were added to the empty model. The two-level random coefficient HLM equation is as follows:

Level 1 (Individual level):

Y_ij = β_0j + β_1j × Jobs–housing balance + β₂ × Weekly working hours + β_3j ×
Travel allowance + β_4j × Male+β_5j × Age + β_6j × Age² + β_7j × Education + β₈ ×
Married + β_9j × Working in private enterprise/self-employed + β_10j ×
Perceived income level + r_ij

(7)

Level 2 (City level):

β_0j = γ₀₀ + μ_0j;
 β_1j = γ₁₀ + μ_1j;
…    
  β_10j = γ₁₀₀ + μ_10j

(8)

In the random coefficients regression, commuting time (Y_ij) is predicted by 10 observation variables at Level 1. There are no predictors at Level 2, but the city variable is used to group individuals. The coefficient μ_1j…μ_10j, μ_0j of variables at Level 1 in the intercept of Y_ij, and r_ij are modeled as the random effects of the city variables. The coefficient γ₁₀…γ₁₀₀ and the γ₀₀ of the variables at Level 1 in the intercept of Y_ij are treated as fixed effects. This model explores whether the effect of city variables discovered in the null model may be attributed, in part, to the fact that different individuals are grouped under different city variables [73]. In general, constructing the Level 2 model aims to explore the predictors of the intercept and slope parameters, with the Level 1 coefficients (i.e., intercepts and slopes) serving as the dependent variables [66].

As shown in

Table 5. Results of random coefficient model.

Variables	Fixed Effect		Random Effect
	Coefficient	SE	Variance	SD
Jobs–housing Relationship
Jobs–housing balance	−12.097 ***	0.459	33.853 ***	5.818
Working Characteristics
Weekly working hours	−0.047 ***	0.010	0.006	0.078
Travel allowance	3.352 ***	0.882	48.862 *	6.990
Socio-demographic Attributes
Male	2.079 ***	0.315	1.744	1.320
Age	0.324 ***	0.089	0.049	0.221
Age2	−3.144 **	0.996	4.210	2.051
Education	0.806 **	0.235	6.428 ***	2.535
Married	−1.946 ***	0.526	4.841	2.200
Working in private enterprise/ self-employed	2.597 ***	0.342	4.200	2.049
Intercept	24.134 ***	0.701	58.227 ***	7.631

Note: the continuous variables in Level 1 are centered based on the group mean. * p < 0.05. ** p < 0.01. *** p < 0.001.

, nine individual-level variables have significant impacts on commuting time. The intercept term at the individual level is significant at the 0.1% level, indicating the need to include predictor variables at the city level to further explain the variation in commuting time. Furthermore, the random effects of jobs–housing balance, travel allowance, and education are also significant. This suggests that we need to add predictor variables at the city level to further explain the variation in regression coefficients of these factors.

4.3. Construction and Results of Full Model

Based on the results of the above random coefficient model, we proceed to construct a full model by adding predictor variables at the city level. The full model is a mixed effects model that incorporates both the individual-level and city-level variables. Its purpose is to examine the significance of the joint effect of these variables on the dependent variable and to assess the extent to which they explain the variance. The two-level full HLM equation is as follows:

Level 1 (Individual level):

Y_ij = β_0j + β_1j × Jobs–housing balance + β₂ × Weekly working hours + β_3j ×
Travel allowance + β_4j × Male + β_5j × Age +β_6j × Age² + β_7j × Education + β₈ ×
Married + β_9j × Working in private enterprise/self-employed + r_ij

(9)

Level 2 (City level):

β_0j = γ₀₀ + γ₀₁ × PD + γ₀₂ × PD² + γ₀₃ × CLA + γ₀₄ × PCRA + γ₀₅ × Bus + γ₀₆ × Bus² + γ₀₇ ×
RTS + γ₀₈ × Taxi+ + γ₀₉ × UR + γ₀₁₀ × GDPPC +μ_0j;
β_1j = γ₁₀ + μ_1j;
…     
 β_10j = γ₁₀₀ + μ_10j

(10)

The results of the full model are shown in

Table 6. Results of the full model.

Variables	Fixed Effect		Random Effect
	Coefficient	SE	Variance	SD
Level 2 (city level) 1
Urban Built Environment Factors
PD	−4.721 ***	0.851
PD2	1.233 ***	0.208
CLA	−5.629 ***	1.075
PCRA	−0.103 **	0.036
Bus	0.045 **	0.015
Bus2	−0.338 ***	0.083
Taxi	0.352 ***	0.060
UR	8.783***	1.848
Level 1 (individual level) 2
Jobs–housing Relationship
Jobs–housing balance	−12.321 ***	0.469	36.845 ***	6.070
Working Characteristics
Weekly working hours	−0.049 ***	0.010	0.006	0.080
Travel allowance	2.965 **	0.870	39.341 **	6.272
Socio-demographic Attributes
Male	2.114 ***	0.316	3.186	1.785
Age	0.326 ***	0.088	0.042	0.205
Age2	−3.166 **	0.999	3.511	1.874
Education	0.863 **	0.239	7.012 ***	2.648
Married	−1.863 **	0.532	5.776	2.403
Working in private enterprise/ self-employed	2.517 ***	0.346	4.860	2.205
Intercept	24.031 ***	0.691	47.972 *	6.926

Note: * p < 0.05. ** p < 0.01. *** p < 0.001. ¹ Variables at Level 2 are centered based on the grand mean. ² Continuous variables at Level 1 are centered based on the group mean.

.

4.3.1. Impact of Built Environments on Commuting Time

As shown in Table 6, the population density of a city has a significant negative effect on commuting time, whereas its squared term has a significant positive effect on commuting time. This implies that there is a U-shaped relationship between population density and commuting time. Specifically, commuting time tends to decrease as population density increases up to a certain threshold, but beyond that threshold, it begins to increase. This result is not surprising. In cities with high population densities, there is often a positive effect known as the agglomeration effect, which improves travel accessibility and shortens commuting times. However, when population densities become excessively high, it is prone to increase congestion and ultimately lead to longer commuting times [22,28].

Urban construction land area and road area per capita have a significant negative impact on commuting time. This implies that cities with larger urban construction land areas and more extensive road areas tend to have shorter commuting times. The expansion of urban construction land areas is often associated with the development of polycentric cities, where employment locations are decentralized. This spatial arrangement may reduce residents’ commuting time [74,75]. In addition, an increase in per capita road area can somewhat alleviate traffic congestion [76], resulting in reduced commuting time.

The number of buses has a significant positive impact on commuting time, whereas its square has a significant negative influence on commuting time. In other words, the number of buses has an inverted U-shaped relationship with commuting time. Within a certain threshold, fewer buses are associated with shorter commute times. Probably, this is because individuals living in cities with fewer buses tend to choose jobs adjacent to their homes, thus reducing commuting time [14]. Once the number of buses reaches a certain threshold, individuals’ commuting times increase accordingly. This increase is often a result of traffic congestion rather than an improvement in the bus system [31]. However, when it is beyond the threshold, a higher number of buses can help alleviate traffic congestion and minimize commuting times [10,23]. The number of taxis has a significant positive effect on commuting time, indicating that a higher number of taxis is associated with longer commuting times. The increase in vehicles, including taxis, is often linked to urban traffic congestion, which can result in extended commute durations.

Urbanization rate has a significant positive effect on commuting time. This suggests that residents in cities with higher urbanization rates tend to have longer commuting times. As a previous study explained, when the urbanization rate surpasses a certain level (about 50%), the advantages derived from enhanced transportation facilities can be outweighed by continued urban expansion and population growth, potentially leading to an increase in commuting times [14].

4.3.2. Impact of Individual Characteristics on Commuting Time

As shown in Table 6, jobs–housing balance has a negative impact on commuting time, which affirms the positive contribution of jobs–housing balance in reducing commuting time [15,40,68]. Weekly working hours are negatively correlated with commuting time, as individuals with longer weekly working hours often prioritize achieving a balance between commuting time and working time [1]. Those who have a travel allowance are more likely to have a longer commuting time. Commuting behavior is more responsive to financial incentives from employers when compared to other types of travel [44,45]. People with travel allowances may choose to live further away from their workplaces to benefit from lower housing costs or a better residential environment, but this choice often results in longer commuting times [77].

Regarding socio-demographic attributes, it is observed that males generally have longer commuting times compared to women. This difference can be attributed to women often taking on more household tasks, resulting in shorter commuting times [46,78]. Age also plays a role, having a positive influence on commuting time, but the relationship is non-linear. There is an inverted U-shaped relationship between age and commuting time, indicating that middle-aged people tend to have longer commuting times compared to younger and older people [48]. This could be attributed to the higher life and work pressures experienced by middle-aged people, which may lead them to extend the scope of their job search and potentially lead to longer commuting times. Furthermore, factors such as higher education, an unmarried status, and employment in the private sector or self-employment are associated with longer commuting times, aligning with previous research findings [17,49,50].

4.3.3. The Spatial Heterogeneity of the Impact of Individual Variables on Commuting Times

The significant random intercept in Table 6 confirms that the average commuting time varies across cities, indicating spatial heterogeneity. Furthermore, the significant random slopes for variables such as jobs–housing balance, travel allowances, and education suggest that the impact of these variables on commuting time is influenced by city-level factors. This suggests that the relationship between these variables and commuting time varies across different cities. This highlights the spatial heterogeneity across cities in the impact of individual characteristics on commuting time. Conducting ordinary regression analysis without considering this spatial heterogeneity between cities can lead to biased and inaccurate results. Therefore, it is crucial to account for spatial variations when studying the relationship between these variables and commuting time.

5. Discussion

The findings emphasize the significant role of the built environment at the city level in determining commuting time while considering the spatial heterogeneity among cities. The U-shaped relationship between population density and commuting time aligns with the findings of previous studies [22,28], suggesting that an optimal population density of a city can contribute to shorter commuting times. While our findings indicate a negative impact of increased urban construction land area on commuting time, it is important to note that a larger urban size does not necessarily lead to better outcomes. The uncontrolled expansion of urban construction land area can result in various negative consequences, including traffic congestion, environmental pollution, and housing difficulties. However, it is worth mentioning that large urban construction land areas are often accompanied by a polycentric spatial structure and improved transportation facilities [79,80]. These factors, rather than urban construction land itself, may have positive effects on reducing commuting time. Therefore, the focus should be on considering the broader context and associated factors when analyzing the impact of urban construction land area on commuting time. The results also indicate that increased road area per capita can lead to shorter commuting times, which is consistent with the results of previous studies [14,21,31]. Similarly, from the perspective of saving land use, it does not signify that the larger the road area of a city the better, but rather that regions lacking road facilities should be remedied so as to minimize commuting time [81,82].

Our study incorporates variables, such as the number of buses and rail transit stations, to characterize the public transit system. The result further contributes to the existing literature by identifying a non-linear relationship between the number of buses and commuting time. Specifically, our findings reveal an inverted U-shaped association, implying a promising avenue for facilitating reduced commuting times by substantially increasing the number of buses. Research has indicated that bus services play a significant role in alleviating traffic congestion by promoting a shift in commuting modes away from cars [83]. Additionally, the introduction of new buses to alleviate overcrowding could be a viable approach [84,85]. These strategies hold the potential to enhance the appeal of bus commuting, curbing private car usage, mitigating traffic congestion, and ultimately leading to a reduction in commuting durations. Moreover, this finding adds nuance to previous studies that have presented differing views on the topic. While some scholars suggest that more buses may reduce commuting time [7], others argue that more buses are associated with longer commuting time [21]. Therefore, it is imperative to adopt a nuanced view of the dual effects of buses. While additional buses may reduce traffic congestion, the start-stop operation at bus stops may have a negative impact on traffic [83]. In such cases, the development of diverse street transit options as a supplement to the conventional bus system emerges as a viable strategy for enhancing accessibility and effectively reducing commuting times. While this paper excludes other street transit modes, such as bus rapid transit (BRT), due to their non-mainstream status in Chinese cities, the potential influence of such urban diversified transit options on commuting behavior should not be dismissed. For example, BRT has been observed to affect residents’ commuting preferences for public transit due to its cost-effectiveness, service capacity, affordability, and relative flexibility [86]. Anticipating and exploring the role of emerging public transportation modes in shaping commuting time remains an intriguing avenue for future inquiry. When taken together, constructing a multi-modal public transportation system that supports efficient connections between nodes may enhance resident commuting experiences and promote sustainable urban development. The intricate interplay between this multi-modal system, private car utilization, traffic congestion, and their combined impact on commuting behavior warrants further exploration and investigation.

In terms of the factors at the individual level, our findings indicate that jobs–housing balance may reduce commuting time, sharing the same viewpoint with previous studies [15,39,40,68,82]. Maintaining a balance between jobs and housing can help alleviate the rising commuting times and traffic congestion in urban areas. Therefore, it is crucial for governments to implement policies and urban planning strategies that promote jobs–housing balance. For example, one effective strategy for megacities is to build a polycentric structure.

The relationship between income-related variables and commuting time is not significant in our study, which differs from the significant findings reported in many previous studies [14,15,17,52]. Unlike the direct use of income data, we utilized the perceived income level as an income-related variable to explore its association with commuting time, as the information on income was incomplete in the dataset. Although the level of perceived income better implies an individual’s perception of income inequality and socio-economic status, which is recognized as an important indicator for understanding human behavior [87,88], it may not be significant, possibly because both the actual and perceived income levels should be considered together in the model.

In addition, individuals with travel allowances tend to accept longer commuting times, as these travel allowances could be considered as compensation for the additional time spent commuting. This finding is supported by previous studies [44,45,77]. The relationship between age and commuting time follows an inverted U-shaped relationship, confirming the viewpoint of Levinson (1998) that middle-aged individuals tend to have the longest commuting times, while older and younger people are more sensitive to commuting times [48]. The effects of individual characteristics shed light on individual commuting needs and, therefore, might help guide the more appropriate allocation of resources. For example, governments should consider the commuting needs of vulnerable groups and implement differentiated urban management policies to promote urban equity.

Previous studies have examined the influence of built environments and individual characteristics on commuting time, but few have considered spatial heterogeneity across cities. In our study, we developed a two-level HLM to explore the combined effects of these factors on commuting time while accounting for the inherent differences between cities. This approach is supported by the findings of other scholars who also highlight the importance of considering spatial heterogeneity in studying commuting factors [13,18,55,56]. Further, we have found that the effects of jobs–housing balance, travel allowances, and education on commuting time vary across cities. This implies that the influence of these three individual-level factors on commuting time varies across cities due to the impact of city-level factors. This has significant implications for urban planning and policy-making. Governments should tailor their strategies regarding jobs–housing balance to reduce commuting time. Additionally, it is important for policymakers to adopt differentiated policy instruments to meet the commuting needs of individuals with travel allowances and varying education levels across different cities. Our study highlights the importance of context-specific approaches in formulating and implementing urban planning and management policies [89,90].

The study has several limitations. First, the data used in this study are not inherently for the purpose of serving how the investigated factors impact commuting behavior, which inevitably resulted in the exclusion of certain variables such as commuting distance and mode. These omissions may prevent a comprehensive understanding of the nuanced variations in commuting times. For instance, previous research has revealed that a single transport mode (e.g., car or bicycling) is negatively correlated with commuting time, while multiple transport modes (e.g., car and metro) are positively correlated with commuting time [40]. Second, due to data limitations, we also did not have access to the spatial details of individuals’ residential and employment locations. Consequently, the built environment is considered at the broader city level, not at the neighborhood level. Numerous earlier studies have revealed the impacts of neighborhood-level built environments on commuting time, capturing the subtle differences in their effects [13,22]. For example, for active commuters in the internal regions of Nanjing, China, commuting time is affected mostly by the land use mix at the working end [13]. Nevertheless, the city-level built environments serve as crucial indicators when formulating urban planning [91], playing a pivotal role in shaping commuting behavior within a broader context. Furthermore, this study strives to unravel the factors impacting commuting time in the context of spatial heterogeneity across cities, and city-level built environments emerge as more proper for illustrating the distinctions between cities when compared to neighborhood-level considerations. Nonetheless, the exploration of multi-level built environment effects on commuting times within the context of complex nested relationships between cities and neighborhoods warrants further investigation. Third, although our case cities cover small, medium, and large cities, the majority of them are small and medium cities. Although the cities in China primarily consist of small and medium cities, the typicality and diversity of the case cities still need to be improved.

6. Conclusions

This study explores the influences of factors at both the city level and individual level on commuting time in the context of considering spatial heterogeneity across cities. This investigation was conducted using the CFPS dataset spanning the years 2014, 2016, 2018, and 2020, encompassing 113 cities at different scales in China, and was achieved through constructing a bi-level HLM. The cities in China are in a phase of rapid urbanization and motorization, with increasing commute times. It is, therefore, important to formulate optimization strategies in urban planning and management to minimize commuting time. Our results provide insights into the design of built environments from the perspective of optimizing commute times and shed light on understanding the effects of individuals’ characteristics on commute time while considering the inherent differences between cities. The findings of this study can inform the development of strategies aimed at improving commuting efficiency and satisfying individuals’ commuting needs.

There are three key findings in this study. First, a non-linear association between population density and commuting time (U-shaped relationship) is identified, as well as between the number of buses and commuting time (inverted U-shaped relationship). These findings highlight the importance of determining an optimal population density and embracing a nuanced view of the dual effects of increasing the number of buses in order to minimize commuting times. It is crucial for future planning policies to consider these non-linear relationships when optimizing urban population density and the number of buses. Second, enhancing urban construction land area and road area per capita may diminish commuting time. Under the premise of controlling land sprawl, future urban planning should develop a polycentric urban structure in cities with a larger construction land area and improve the road facilities in underdeveloped areas to enhance commuting convenience. Lastly, the impacts of jobs–housing balance, travel allowances, and education on commuting time vary across cities. These findings highlight the need to consider the individual-level factors in planning and management policies after controlling for spatial heterogeneity. Specifically, it is crucial for cities to customize optimization measures for their unique built environments in order to promote jobs–housing balance, meet the diverse commuting needs of individuals, and ultimately reduce commuting time. The above findings provide feasible ideas to address the challenges of increasing commuting time, and they foster sustainable urban planning and development.

Future research should strive to incorporate a wider array of representative variables to refine the estimation of factors influencing commuting time, and this exploration should extend to encompass the intricate inter-relationships between cities, neighborhoods, and individuals. Moreover, a deeper understanding of how different built environments interact, including their synergy with broader urban development indicators, is essential for comprehending their impact on commuting behavior. Furthermore, the development of methodologies capable of capturing and predicting the influence of changes in city-level built environments on residents’ commuting time over longer timeframes is of significant value.