Monocentric or Polycentric? the Urban Spatial Structure of Employment in Beijing

The spatial structure of Beijing has changed dramatically since the reforms of the late 1970s. It is not clear, however, whether these changes have been sufficient to transform the city's monocentric spatial structure into a polycentric one. This paper uses 2010 enterprise registered data to investigate the spatial distribution of employment in Beijing. Using a customized grid to increase the spatial resolution of our results, we explore the city's employment density distribution and investigate potential employment subcenters. This leads to several findings. First, Beijing still has strong monocentric characteristics; second, the city has a very large employment center rather than a small central business district; third, five subcenters are identified, including four in the suburbs; and fourth, a polycentric model that includes these subcenters possesses more explanatory power than a simple monocentric model, but by only four percentage. We conclude that the spatial structure of Beijing is still quite monocentric, but may be in transition to a polycentric pattern.


Introduction
China is urbanizing at an unprecedented rate, with many cities experiencing rapid population expansion [1]. Beijing, the country's capital, has been one of its fastest-growing cities over the past few decades [2][3][4]. According to census data, Beijing's population grew from 13.6 million in 2000 to more than 19.6 million in 2010, an increase of 44.5%. The aggregation of population in the city center has caused many problems, including traffic congestion [5,6], air pollution [7][8][9] and a lack of affordable housing [10,11].
These problems are often blamed on the capital's spatial structure, which critics argue is overly monocentric and should be decentralized. In fact, Beijing has been making efforts to establish a polycentric urban spatial structure by implementing a series of urban plans and related policies, targeting the decentralization of both employment and population [12,13]. Since the economic reforms of the late 1970s, Beijing has issued three comprehensive city plans, with planning periods of 1981-2000, 1991-2010 and 2004-2020 [14,15]. Each of these plans has included the formation of a multi-center development pattern among its goals. For example, the 1991-2010 Beijing Comprehensive Plan outlined a polycentric development pattern that it called "decentralized cluster style (Fensan Zutuan Shi)". In this proposal, the urban area would consist of the existing old city and ten edge clusters, each surrounded by a greenbelt (the first greenbelt) that would control urban sprawl and improve environmental quality. With the development of urban land and housing markets since the early 1990s, however, Beijing's urban area expanded quickly, encroaching on the greenbelt. By the early 2000s, the greenbelt had all been developed, and the old city merged completely with the edge clusters [16]. The 2004-2020 Beijing Comprehensive Plan followed its predecessor's principle of decentralized development, proposing a three-tier development structure composed of the central city, the new towns and the common towns. The central city referred to the old city and the ten edge clusters from the 1991-2010 plan; the new towns included eleven towns where suburban district and county governments were located [17]; and the common towns referred to other designated towns. Between the central city and the new towns, a second greenbelt was planned with the same purpose as the first.
The city's spatial structure and these efforts to control it have attracted broad interest from researchers and policy makers. Many studies have focused on the spatial distribution of Beijing's residential population, concluding that, in general, the city has shown a trend toward polycentricity and suburbanization [18][19][20][21][22][23]. However, just because a city's residential population has suburbanized does not necessarily mean that its urban spatial structure has also changed to a polycentric form, especially if employment is still concentrated in the city center. The spatial distribution of employment plays an important role in the structuring of urban spaces. Indeed, some urban economic theories begin from the spatial distribution of employment and determine population distribution from distance to the employment center, transport costs and other factors [24][25][26][27]. Thus, a polycentric city is usually defined by the presence of one or more employment subcenters outside of the central business district [28].
Unfortunately, few studies have examined the spatial distribution of employment in Chinese cities, probably due to the difficulty of data collection. China has conducted six population censuses, in 1953, 1964, 1982, 1990, 2000 and 2010, and thanks to the country's hukou (household registration) system, annual population data are available for most cities. In contrast, the country has conducted only two economic surveys of limited scope, in 2004 and 2008. What is more, data collected in these surveys are of doubtful quality due to the lack of unified classification standards among administrative departments at different levels [29].
These few studies have yielded no consensus. Some researchers suggest that Beijing is still dominated by a monocentric urban spatial structure [30,31], while others conclude that Beijing has already entered a polycentric era [31,32]. For example, Sun et al. [32] use employment data from the secondary and tertiary industries to identify five employment subcenters: Yangfangdian Street, Zhongguancun Street and Shangdi Street in Haidian District, Heping Street in Chaoyang District and Yingfeng Street in Fangshan District. Due to the limitations of existing employment data, these studies were mainly based at the town level, which leads to relatively coarse results [33].
This paper seeks to deepen the understanding of Beijing's urban spatial structure by using a better set of employment data: the 2010 enterprise registered data from the Beijing Industry and Commerce Bureau. We customize a 1.5 km × 1.5 km grid as the research unit and follow the two-stage method proposed by McMillen [28] to identify employment subcenters. Finally, we establish a polycentric model to explain the employment density distribution. Our findings will supplement the current understanding of the city's spatial structure and lay the foundation for better urban development, planning and policy-making in the future.
The paper is organized as follows: Section 2 describes the research area, data and method, while Section 3 provides a general description of employment density in Beijing in 2010. Section 4 evaluates a monocentric model, identifies Beijing's employment subcenters and examines the subcenters' ability to explain the entire employment density distribution of the city. Conclusions are presented in Section 5.

Research Area
Beijing has a total administrative area of approximately 16,410 km 2 . The city's northwest is a primarily mountainous area of about 10,072 km 2 and is protected for ecological reasons, while the southeast is made up of plains of about 6338 km 2 ( Figure 1). The main urban area is located on this plain, with the Ming and Qing Dynasty Imperial Palace (the Forbidden City) at its center. Since 1949, a system of ring roads has been developed around the Forbidden City, with the city's expansion requiring the construction of the 2nd, 3rd, 4th, 5th and 6th Ring Roads in turn [34]. With Tiananmen as the center, the average radius of these five ring roads is about 4 km, 7 km, 10 km, 15 km and 25 km, respectively. Eight radial expressways have also been constructed to link the central city and suburban areas ( Figure 2). The city's subway system is a similar ring-and-radial network; Subway Lines 2 and 10 are loop lines, while Lines 13 and 4, as well as the Changping, Fangshan, Yizhuang, Shunyi, Batong and Airport lines are all radial lines.  We define a research area of 4429 km 2 , which includes the area within the 6th Ring Road and a 10-km buffer zone beyond it ( Figure 2). We selected this research area based on several considerations. First, in view of the subway system's extent and assumptions about the commuting range of private cars, this area represents a reasonable commuting scope and can be regarded as an integrated labor market. Indeed, according to bus-pass data, the majority of transportation flow occurs within this area [35]. Second, the research area includes not only the central city, but also seven new towns, which have closely integrated with it [36] (Figure 3). Third, this area accommodates most of Beijing's population and firms, accounting for about 80.9% and 74.5%, respectively. Fourth, if the research area were expanded further, many mountainous regions with weak links to the central city would be included.

Data
The employment data are from the Beijing Industry and Commerce Bureau's 2010 list of registered enterprises and institutions. The registered enterprises include state-owned, private, collective, foreign-invested, foreign-owned and other types of enterprise, while the institutions include government agencies, institutions and social organizations. Collected information includes enterprise and institutional names, addresses, turnover, registered capital, the number of their employees, and so on. We use ArcGIS and Google Maps to locate the enterprise address, giving us the spatial location of all 411,300 enterprises [37].
The Beijing Industry and Commerce Bureau's dataset includes 14.91 million employees, the sectoral composition of which is in general accordance with the published data of the Beijing Statistical Yearbook [38] (Table 1). This dataset covers a wide range of industries; business services, technological exchange and promotion services and wholesale trade are the top three by employment, accounting for 6.3%, 5.9% and 5.8% of the total employment, respectively ( Table 2).
One shortcoming of our data is that the employment figures for some large enterprises, such as the China National Tobacco Corporation, the China Network Communications Group Corporation and Sony (China) Limited, include employees working not only in Beijing, but also elsewhere in China. If these companies are included in the study, they might bias the spatial distribution of employment in the research area. Therefore, we eliminated 560 companies of this kind from the dataset, representing 0.14% of the total number of enterprises in the research area [39]. Due to their relatively small numbers and lack of spatial concentration, excluding these enterprises has little effect on our results. Because of these removals, the dataset used in this study contains 410,800 enterprises and 13.51 million employees.   [42] use an average of 3 square miles as their research unit. In light of these scales, we defined a customized grid as our research unit.
Choosing the appropriate grid size is critical and can affect the calculated employment density. If the grid is too small, there will be many units with zero employment density, such as parks or residential blocks, and many others with extremely high employment density, such as blocks of high-rise office buildings. If the grid is too large, employment concentrations may be offset by neighboring parks, and centers of employment may be hard to identify [28].
We tested grids with cell sizes of 1 km × 1 km, 1.5 km × 1.5 km and 2 km × 2 km [43]. As shown in Table 3, the research area includes 4266 cells of 1 km × 1 km with 24.9% of them being null, 1883 cells of 1.5 km× 1.5 km with 9.2% of them being null and 1030 cells of 2 km× 2 km with 6.5% of them being null (Table 3). Considering both the number of samples and the percentage of null-value units, we chose to use the 1.5 km × 1.5 km grid as our research unit.

Identifying Subcenters
A number of methods have been used to identify subcenters in metropolitan areas. Among them, the most direct relies on the researcher's knowledge and observations, but this is highly subjective. Building on observations and experience, some researchers use a threshold value-setting method for identifying employment subcenters. For example, Giuliano et al. [42] define a subcenter as a group of contiguous tracts that each have at least 10 employees per acre and that together have at least 10,000 employees. This method is also subjective, since the identification of subcenters is sensitive to the choice of the threshold. For example, McMillen et al. [44] use a threshold of 20 people/acre, with total employees exceeding 20,000.
McDonald [45] proposed a more objective method for identifying employment subcenters, defining a subcenter as a cluster of significant positive residuals from a simple regression of employment density on the distance from the central business district. McDonald et al. [41] subsequently used the same method to identify employment subcenters in Chicago, with reasonable results. McMillen [28] argued that a subcenter should also have a significant effect on overall employment in the city and proposed a two-stage method, which first identifies candidate subcenters and then estimates their explanatory power. This two-stage method is better at eliminating subjective influences and can easily be employed by other researchers [28,40].
We adopt a similar two-stage method. In the first stage, we follow McDonald's [46] method of identifying potential subcenters. First, we establish a monocentric model. The classical monocentric model assumes that all of a city's jobs are concentrated in its central business district [24,26,27]. However, this assumption cannot fit reality. When investigating the spatial distribution of employment, many studies assume that employment distribution is similar to population distribution, but more concentrated in the city center [47]. To capture this dynamic, a negative exponential model has been widely used.
D(s) is employment density; s is the distance to Tiananmen, often regarded as Beijing's center [48]; α is employment density at distance zero and β (β < 0) is density gradient; u is the error term.
Second, since the negative exponential model is nonlinear, we use its logarithmic form to estimate it by the ordinary least squares method (OLS).
Here, c = lnα, and other symbols are the same as in model Equation (1). Using this monocentric model, the residuals between real and predicted values can be calculated, and those observations with residuals significant at the 95 percent confidence level are chosen as potential subcenters.
Third, to deal with the potential problem that a unit surrounded by very low employment density units could be falsely identified as a candidate subcenter, we follow the threshold method of Muñiz et al. [49], which uses the average employment density of the research area as the threshold for determining potential employment subcenters.
In the second stage, we build a polycentric model to test whether our candidate subcenters help explain the research area's overall employment density distribution. Heikkila et al. [50] propose three different hypotheses about the role of subcenters, corresponding to three different polycentric models. As potential subcenters have different industries and different functions [51], the hypothesis that benefits from all subcenters being completely complementary is most suitable for our study. Based on an exponential function, a polycentric model that added control variables can be expressed as: where ( , , ) i sub X D s s is the density of a research unit, s and i sub s are the distance to Tiananmen and the potential employment subcenter i, respectively; X is control variables, including the distance to the nearest subway station, the highway and the airport; a0, ai(i = 1, 2…) and δ are parameters to be estimated; u is an error term. A log-transformed version of model Equation (3) is also adopted here, for the same reason. Thus, our polycentric model is as follows: Here, c = lnα0 and other symbols are the same with model Equation (3). However, many studies suggest that the employment density declines more precipitously when moving away from subcenters than from the main city center; therefore, the inverse distances to the employment subcenters are often employed as interpretation variables [40,46], as shown in model Equation (5). This can also solve the problem of potential collinearity between the distance variables. In our research, model Equation (5) is adopted to estimate the effect of subcenters on overall employment density distribution.

Effect of the Potential Subcenter on Local Employment Density
To examine the effect of potential subcenters on local employment density, we estimate regressions with the distance to Tiananmen and the distances to each potential subcenter as the explanative variables and confine the observations to certain scopes (5, 10 and 15 km radius) to the potential subcenters shown in model Equation (6).
Control variables are excluded in the above model Equation (6) since distances to subway stations and highways are in general highly correlated with the distance to the subcenters, and the distance to the airport has little effect on the employment density when the observations are limited to a local scope.

Employment Density in Beijing
The average employment density in the research area is 3125 employees/km 2 . There is a great deal of variation around that mean, however, from just two employees/km 2 up to a high of 98,746 employees/km 2 in one grid cell on Jianguomen Street. The spatial distribution of employment displays four major characteristics as follows.

Density Declines from the Center to the Suburb
The research area is divided by the five ring roads into six concentric zones, numbered from I, in the center, to VI, on the periphery (Figure 4). Employment density varies among these ring zones, decreasing dramatically as one moves outward from the center. The employment densities of the six zones, from center to periphery, are 29,248 employee/km 2 , 31,547 employee/km 2 , 22,571 employee/km 2 , 6897 employee/km 2 , 1447 employee/km 2 and 563 employee/km 2 ( Table 4). From the innermost to the outermost ring zone, the employment density drops by about 98%. The highest employment density is not observed in the innermost ring zone, but in the next zone out, Zone II, which is not predicted by the classic monocentric model. Several factors may contribute. First, historic-site preservation has prevented economic concentration. The 2nd Ring Road was built at the site of the old city wall and contains within it the city's major sites of historical and cultural importance, including the Ming and Qing Dynasty Imperial Palace, Tiananmen (which together cover about 1.2 km 2 ), the Temple of Heaven and historic quadrangles and lanes (hutong). Preservation of the old city has long been an important part of city planning, and many important sites have been protected from modern development as a result. Second, development and construction within the 2nd Ring Road are subject to strict planning regulations, such as building height caps and density limits, which have influenced employment density. Third, the city has relocated populations and manufacturing from the center to suburban areas to reduce pressure on the central city.

A Vast Employment Center
It is also worth noting that the 4th Ring Road seems to be a boundary outside which the employment density plummets, while the employment densities of the areas within the 4th Ring Road (Zones I, II and III) are relatively similar. What is more, the area within the 4th Ring Road accounts for only 6.8% of the research area, but it hosts almost 60% of its employment (Table 5). It may be more appropriate to think of a large employment center of 300 km 2 than a small central business district, as assumed in theory ( Figure 5).

Figure 5.
Beijing's employment density distribution at different scales.

The Influence of Transportation
When the city's major roads are added to the employment density map, it appears that employment density is closely associated with major transit roads, extending radially along the expressways ( Figure 5). For example, higher employment density can be observed along the east-west Subway Line 1 (Figures 1 and 5), the Beijing-Tibetan expressway to the northwest, the Beijing-Shijiazhuang expressway to the southwest and the Beijing-Kaifeng expressway to the south (Figures 2 and 5). We established a simple OLS regression model to further examine the effects of the highways and subway on employment density: where ( , , ) highway subway D s s s is employment density, is the distance between a given unit and Tiananmen, shighway is the distance from the unit to the nearest highway, ssubway is the distance from the unit to the nearest subway station, c is a constant and u is a random error term. According to the regression results, both highways and subway stations have a significant positive impact on employment density at a 99% confidence level (Table 6).

Employment Density Varies by Direction
We also examined directional variation in employment density ( Figure 6). We select nine directions with potential employment subcenters, as well as six directions with no specific high-employment unit in the suburbs [52] and choose grids intercepted with each direction to draw the employment density profile for each direction (Figure 7). s Figure 6. Directions of employment density profile.
We can make several observations based on Figure 7. First, employment density seems to decline differently along different headings. Second, despite the overall downward trend from the center to the suburbs, in most directions, the employment density first increases and then decreases. Third, the employment density increases again in the remote suburbs in several directions, suggesting the existence of employment subcenters. Fourth, in some directions, the employment density drops very dramatically beyond a certain distance to the city center, producing a cliff-shaped profile. This suggests that employment may gather within a certain area.
These observations suggest that the spatial structure of employment shows both monocentric and polycentric characteristics. Employment density decreases from about 30,000 employees/km 2 within the 3rd Ring Road to 563 employees/km 2 in the outermost ring zone, suggesting a monocentric structure. At the same time, peaks of employment density in certain suburban areas suggest a polycentric pattern. We investigate this further in the next section.  Table 7 reports the results of model Equation (2). The estimation adopts White's [53] method to calculate standard errors and the covariance matrix to eliminate a potential heteroscedasticity problem. We find that both the model and the coefficient of the distance to Tiananmen are statistically significant at the 99% confidence level (Table 7). For every kilometer that the distance to Tiananmen increases, employment density decreases by 14.6%. The adjusted R 2 is 0.333, indicating that this single distance variable explains 33.3% of the variation in employment density. In other words, the monocentric model works well [54] and the distance to Tiananmen is significant in explaining employment distribution in the research area.

Results from the Monocentric Model
While the monocentric model is useful in identifying employment subcenters, we wanted to clarify the scope at which it works. Accordingly, we varied the limits of the research area (Table 7). When limiting it to within the 6th Ring Road and the 5th Ring Road, the monocentric model is significant; when limiting it to within the 4th Ring Road, the monocentric model is still significant, but the explanatory power is greatly reduced; when limiting it to within the 3rd Ring Road, the monocentric model is no longer significant (Table 7). This is consistent with the direct observation from Figure 5 that, within the 3rd Ring Road and even within the 4th Ring Road, employment density is much higher, with no apparent decline away from the center. Furthermore, the explanatory power of the monocentric model is strongest within the 5th Ring Road and declines as the research area expands, which may be a result of employment subcenters in the suburbs.

Subcenter Identification
Based on the monocentric model, we conducted a residual analysis and identified 31 research units where the residual is significantly greater than zero (at the 95% confidence level) as employment subcenter candidates. Seven of them have lower employment density than the overall average and are removed from consideration [55]. For the remaining 24 units, when the units are very close to each other (within 3 km [56]), we selected the unit with the highest employment density to represent them. This gave us nine potential sub-centers located in Shangdi Street, Haidian Street, Shahe Town, Changping Urban District (including South Street and North Street), Renhe Street, Gongchen Street, Xinhua Street, Zhangjiawan Town and Longquan Street. We use the names of these streets and towns to name the nine potential subcenters in the following discussion.

Effects of Potential Employment Subcenters on Overall Employment Density
The results of model Equation (4) are shown on the left side of Table 8. After some tests, we include two dummy variables of the distance to the subway station and highway and one continuous variable of the distance to the airport into the regression function to control other factors' influences. The distances to Tiananmen, Haidian Street, Changping Urban District, Renhe Street and Gongchen Street have a significant impact on employment density. The distances to Shangdi Streets, Longquan Street and Xinhua Street are also significant, but the coefficient signs are positive, which is not consistent with the expectations. One possible problem of this is collinearity. For example, Shangdi Street, Haidian Street and Shahe Street are in the same direction, and the distance between them is short, so the distances to these three potential subcenters are highly correlated [57]. Another reason is that the effect of proximity to a subcenter declines rapidly with distance, as mentioned in Section 2.3.2.  To solve these problems, model Equation (5) has been estimated, and the result is shown in Table 8

Effects of Potential Subcenters on Local Employment Density
The results demonstrate that most models are significant at the 95% confidence level, with a few exceptions, including the models within 5 km of Renhe Street and Shahe Town and models within 5 km and 10 km of Zhangjiawan Town (Table 9). All coefficient signs of the subcenters are as expected.
Each of the five subcenters identified in Section 4.2.1 has a strong influence on local employment density within certain distances. Changping Urban District and Xinhua Street significantly affect employment density within a radius of 5 km, 10 km and 15 km. Haidian Street and Gongchen Street are significant within a radius of 5 km and 10 km, but their influence is unclear when the radius is expanded to 15 km. Renhe Street is somewhat unique, given that its effects are significant at 10 km and 15 km, but not significant within a radius of 5 km. It could be that employment density does not decline significantly within 5 km of Renhe Street. For the other potential subcenters, Shangdi Street and Longquan Street influence employment density within a radius of 5-10 km, while Shahe Town and Zhangjiawan Street have little effect on local employment density.
According to these results, the nine potential employment subcenters can be divided into three types. The first includes Haidian Street, Xinhua Street, Changping Urban District, Gongchen Street and Renhe Street. These five candidates clearly influence local and overall employment distribution. The second type includes Shangdi Street and Longquan Street, which influence only local employment distribution. The third type includes Shahe Town and Zhangjiawan Street, which have high employment density, but little effect on employment distribution. According to McMillen [28], a subcenter should have a significant effect on overall employment. We therefore identify the first group of candidates as Beijing's employment subcenters (Figure 8).

Polycentric Model
Based on the five identified subcenters, we use model Equation (5) to explain the overall employment density distribution. The regression results are shown in Table 10. To compare the polycentric and monocentric models, a monocentric model with added control variables has also been conducted here.
The polycentric model and the coefficients of all independent variables are significant at a 99% confidence level. A partial F-test suggests that the five subcenter distance variables together are significantly more than zero, indicating that these identified subcenters influence the overall employment distribution. However, adding the distance to the five subcenters improves the entire model's explanatory power by only four percent: the adjusted R 2 of the polycentric model is 0.434, compared to the monocentric model's adjusted R 2 of 0.393.

Discussion and Conclusions
Although many cities in developed countries have polycentric or even dispersed employment patterns [41,[58][59][60][61][62], the situation in Beijing appears to be different. First, the city still has very strong monocentric characteristics. The single factor of a cell's distance to Tiananmen explains 33.3% of its employment density, a figure that increases further when peripheral cells are excluded (34.9% inside the 6th Ring Road, 45.0% inside the 5th Ring Road). Second, the city has a very large employment center, which is centered at Tiananmen and generally overlaps with the 4th Ring Road. We call this center "large" because it has an area of about 300 km 2 (6.8% of the research area) and accounts for 58.7% of employment, whereas other cities' downtown areas can often be treated as points. Third, despite the identification of five sub-centers, it is difficult to conclude that the city has achieved a polycentric structure, as the polycentric model improves the explanatory power of the overall employment distribution by only 4% over the monocentric one [63,64]. Based on these findings, we conclude that the spatial structure of Beijing is still monocentric, but may be in transition to a polycentric pattern.
Beijing has tried to foster subcenters and to achieve a decentralized spatial development pattern to a certain degree; however, our study suggests that the success of these efforts has been limited. The monocentric model still explains the spatial structure of the city's employment very well, and adding the identified subcenters improves its explanatory power by only four percentage points. At the same time, though four of the five identified subcenters-Xinhua Street, Changping Urban District, Gongchen Street and Renhe Street-are located in the planned new towns, the other three new suburb towns have not achieved development as expected. We speculate that several factors contribute to Beijing's strongly monocentric character and the formation of its large employment center. First, the central city enjoys better infrastructure than other areas and hosts most of the central and city government departments, schools, hospitals, museums, public facilities and parks. Second, the ring-and-radials road structure plays a very important role in facilitating the concentration of people and economic activities in the central city. Third, and more important, the power of agglomeration remains decisive. As a result, new development still prefers locations close to the central city.
With these somewhat disappointing results in mind, several inferences may be drawn from the findings of this study. First, the monocentric characteristics of employment distribution suggest that the agglomeration economy still plays a dominate role in Beijing, and so planning interventions and heavy public investment in decentralized development may result in a loss of economic efficiency [65]. We feel that planning tools are useful in guiding and regulating development, but market needs and the efficiency of public investment also deserve careful considerations. Given the fact that the city center still accommodates most of the employment, the government should also pay attention to further improving the infrastructure and facilities of existing developed areas, thus making the city a more attractive and livable place. Second, given that the employment is still concentrated in the city center based on the findings of this paper while the population has already redistributed to the suburban area according to the literature suggests the possibility of a more imbalanced job-housing spatial distribution and longer commuting distances, which also deserves more attention when making plans and policies.