-
Our research area includes 31 provinces in China, except for Hong Kong, Macao, and Taiwan. City extents comprise our geographic unit of analysis because they typically represent the spatial scale at which policies are formulated for the attraction and retention of talent, and factors influencing talent often vary significantly across cities (Nie and Liu, 2018; Gu et al., 2019a). Based on individual-level records from population censuses and sample surveys, the stock of talent (i.e., the raw number of highly educated people) is aggregated to the city level. A total of 309 cities are chosen for our basic dataset. Cities in China include prefecture cities and autonomous prefectures. When we did regression analysis, we excluded data on autonomous prefectures because of their poor quality and number of missing entries. Then, we eliminate samples of missing data and with no neighboring city. Finally, we used a total of 233 cities out of 309 units for regressions. The selected 233 cities were representative since they covered cities of different sizes, administrative levels, and geographical regions. China’s economic geography is divided into four economic-geography regions: the Eastern, Central, Western, and Northeastern regions (Fig. 1) (Zhou et al., 2019).
-
FollowingYang et al. (2014), we used the concentration index (CI) method to describe the spatial concentration of talent across cities. The concentration index can be expressed in terms of the proportion (%) of talent in 1% of a country’s land area. It is represented as:
$$ {CI}_{i}=\frac{{P}_{i}/{P}_{n}\times 100{\text{%}} }{{A}_{i}/{A}_{n}\times 100{\text{%}} }=\frac{{P}_{i}/{A}_{i}}{{P}_{n}/{A}_{n}} $$ (1) where
$ {CI}_{i} $ denotes the concentration index of city$ i $ ;$ {P}_{i} $ is the talent stock of city$ i $ ;$ {A}_{i} $ is the land area of city$ i $ ;$ {A}_{n} $ is the total land area of China;$ {P}_{n} $ is China’s total stock of talent. -
Measuring the degree of spatial autocorrelation is vital for the understanding of the distribution of talent in China. Moran’s I is one of the most widely used methods for spatial dependence of geographical attributes between neighboring regions (Zhou et al., 2019). The formula of the Moran’s I is:
$$ I=\left(X{'}WX\right)/\left(X{'}X\right) $$ (2) where I is the Moran’s I for the density of talent of cities.
$ X $ denotes the vector of observations,$ X{'} $ is the transpose of$ X $ , and$ W $ is the standardized spatial weight matrix under queen criteria. The queen spatial weight matrix has been widely applied in defining the spatial relationship of China’s cities (Hong and Sun, 2011; Gu et al., 2019a). The value of Moran’s I ranges from –1 to 1, with positive values for positive spatial autocorrelation and negative values for negative spatial autocorrelation. -
Coefficient of variation (CV) can measure the degree of differences in the distribution of talent of cities. CV is calculated as follows:
$$ CV=\dfrac{1}{\overline {P}}\sqrt{\frac{\displaystyle\sum _{i=1}^{n}{({P}_{i}-\overline {P})}^{2}}{n-1}} $$ (3) where CV is the value of the coefficient of variation.
$ \overline {P} $ denotes the average stock of talent across cities; n is the number of cities. -
To model migration flows and identify the patterns underpinning their key drivers, we use a negative binomial model (NBM) framework which is appropriate to model count data with overdispersion (Cameron and Trivedi, 2013): our dependent variables are counts of talent, i.e., number of people with a college degree or above. The NBM introduces a parameter α that measures overdispersion in the data. For our dependent variables, the variance was significantly larger than the mean, so that we applied a fixed-effects overdispersion NBPM proposed by Hausman et al. (1984) to model panel data taking the density of:
$$ {{Pr}}\left({y}_{it}|{\mu }_{it},{\rm{\alpha }}\right)=\frac{\varGamma \left({\alpha }^{-1}+{y}_{it}\right)}{\varGamma \left({\alpha }^{-1}\right)\varGamma \left({y}_{it}+1\right)}{\left(\frac{{\alpha }^{-1}}{{\alpha }^{-1}+{\mu }_{it}}\right)}^{{\alpha }^{-1}}{\left(\frac{{\mu }_{it}}{{\alpha }^{-1}+{\mu }_{it}}\right)}^{{y}_{it}} $$ (4) $$ E\left({y}_{it}|{x}_{it}\right)={\mu }_{it}={\rm{exp}}\left({x}_{it}{'}\beta \right) $$ (5) where Γ is the Gamma integral;
$ {\mu }_{it} $ equals$ E\left({y}_{it}|{x}_{it}\right) $ ; α is the dispersion parameter of the Gamma distribution. When this parameter tends to 0, the NBM becomes a Poisson Model;$ {y}_{it} $ represents the talent stock of city i at time t;$ {x}_{it} $ represents a vector of independent variables of the city i at time t;$ \beta $ represents the vector of estimates capturing the relationship between the independent and dependent variables. The model also incorporates a parameter for individual-specific fixed effects (Cameron and Trivedi, 2013).To correct for spatial autocorrelation in our regression estimates, we use eigenvector spatial filtering (ESF). ESF is implemented as follows: first, we constructed an n-by-n binary spatial weight matrix W under queen criteria using GeoDa to represent the connectivity of Chinese cities (n denotes the total number of cites). The second step is to construct a transformed spatial weight matrix. Given a matrix M = I–AA’/n, where I is an n-by-n identity matrix, and A is an n-by-1 vector of 1s. M can center the spatial weight matrix W by MWM. Third, we calculated the eigenvalues and eigenvectors of the matrix MWM:
$$ MWM=E\Lambda E' $$ (6) where E is the matrix of eigenvectors decomposed by the transformed matrix, and
$ \varLambda $ is a diagonal matrix with corresponding eigenvalues.Getis and Griffith (2002) demonstrated that the eigenvectors in the ESF method are orthogonal and uncorrelated, and each eigenvector corresponds to a unique pattern of spatial autocorrelation. Additionally, each individual eigenvector in E can be linked to a Moran’s I:
$$ {I}_{j}=\frac{n}{{A}{'}WA}\frac{{e}_{j}{'}{{M}}W{{M}}{e}_{j}}{{e}_{j}{'}{{M}}{e}_{j}} $$ (7) where n is the number of cities. Ij is the Moran’s I for the jth eigenvector ej. Griffith (2003) suggested identifying with a critical value of the corresponding eigenvalues indicating a specific minimum spatial autocorrelation level (e.g., Moran’s I = 0.25). After the selection of the candidate eigenvectors, a subset of eigenvectors can be chosen with the Akaike information criterion (AIC, Chun, 2011). This process minimizes the estimation error without reducing too much the degree of freedom of the model.
For space-time panel data analysis (e.g., the NBPM), eigenvectors need to be concatenated T times to match the total number of space-time observations (Chun, 2011). Because the spatial structure of the data is invariant over time, linear mixed models (LMM) and generalized linear mixed models (GLMM) are suggested to apply to account for spatial and temporal autocorrelation with ESF technique (Chun, 2011; 2014). Fortunately, the fixed effects NBPM specification proposed by Hausman et al. (1984) can accommodate the time-invariant spatial autocorrelation represented by the selected eigenvectors (Gu et al., 2019b). After adding eigenvectors that are used to control the effect of spatial autocorrelation, our model specification can be defined by:
$$ E\left({y}_{it}|{x}_{it},{e}_{it}\right)={\mu }_{it}={\rm{exp}}({x}_{it}'\beta +{e}_{it}'\gamma) $$ (8) where eit denotes the selected eigenvectors of city i at time t, and
$ \gamma $ is the vector of estimators. -
Following existing studies (Liu and Shen, 2014; Nie and Liu, 2018; Gu et al., 2019a; 2020a; Qi et al., 2020), we defined talent: individuals with a college degree or above. The rationale of using academic qualifications to define talent is that academic qualifications provide an accurate representation of human capital embedded in the labor force. In most cases, regional development policies of cities often use academic qualifications to define and divide talent (Rowe 2013).
We used data from the National Bureau of Statistics of China (2001; 2006; 2011; 2016). The data extracted from the surveys conducted every five years between 2000 and 2015 can support a four-year panel data analysis. In 2000, there were 45.6 million highly educated individuals in China, comprising 3.67% of the total population. By 2005, this number had increased by 30 million to comprise 75.7 million, and by 74 million to represent 120.1 million in 2010. In 2015, the pool of Chinese talent comprised 151.1 million and accounted for 10.99% of the total population. For the independent variables, we used to explain the spatial distribution of talent in China. We draw on data from 1999, 2004, 2009, and 2014 reported by the National Bureau of Statistics of China (2000; 2005; 2010; 2015). As in previous studies (Liu and Shen, 2014; Gu et al., 2019a), we included independent variables to capture two broad sets of factors: economic and amenity factors, and also introduced a range of control variables. Three variables were used to capture differences in economic opportunities across cities: gross domestic product (GDP), the average number of urban employed staff and workers per 10 000 habitants (EMPLOY), and the proportion of tertiary industry output in GDP (INDUS). Expectedly, a higher level of GDP and a larger number of urban employees per 10 000 provides adequate and diverse job opportunities for the region, which may attract more talent to settle in (Nie and Liu, 2018). A higher rate of tertiary industry outputs in GDP means a more optimized industrial structure, which may also conduce the stock of talent (Gu et al., 2019a).
Four variables were used to capture differences in public sector amenities: including the ratio of total expenditure on science, technology and education in total fiscal expenditure (STEEXPEND), the ratio of fiscal expenditure to revenue (SPEND), the number of primary school teachers per 10 000 students (PRIEDU) and the number of doctors per 10 000 people (MEDICAL). Expectedly, better provision of public services will increase the stock of talent (Glaeser et al., 2001). We also included two variables to account for differences in natural amenities: green coverage rate (GERRN), and sulfur dioxide emissions (SO2). Research has also shown the vital role of natural amenities in the decision-making of talent (Knapp and Gravest, 1989; Partridge, 2010).
Additionally, we introduced four other control variables which may help to explain the distribution of talent: the number of college students per 10 000 habitants (UNISTU), population sizes of the city (POP), per capita fixed-asset investment (FAI), and population density (DENS). The dependent variable of the study is the raw number of talent, thus it is essential to control the effect of the population size as well as the economic density and agglomeration effect. The number of college students per 10 000 habitants was used to control the supply of talent, and FAI was considered as the proxy for urban construction investment. A description of variables is provided in Table 1.
Table 1. Description and expected effects of variables
Type Variable Description Expected effect Number Mean Dependent variables TALENT Number of people with a college degree or above of each city in 2000, 2005, 2010, 2015 932 359602.1577 Economic opportunity variables GDP Gross GDP of each city in 1999, 2004, 2009 and 2014 / (10000 yuan (RMB)) + 932 15.6558 EMPLOY Average number of urban employed staff and workers per 10000 habitants in
1999, 2004, 2009, and 2014+ 927 6.8535 INDUS The proportion of tertiary industry to GDP of each city in 1999, 2004,
2009 and 2014 / %+ 931 36.4876 Amenity variables STEEXPEND The proportion of per capita science, technology and education expenditure
to the financial expenditure of each city in 1999, 2004, 2009 and 2014 / %+ 932 18.1333 SPEND The ratio of per capita financial expenditure to per capita fiscal revenue of
each city in 1999, 2004, 2009 and 2014 / %+ 932 214.6244 PRIEDU Number of primary school teachers per 10000 primary school students of
each city in 1999, 2004, 2009 and 2014+ 931 6.2942 GREEN Greening rate of each city in 1999, 2004, 2009 and 2014 / % + 929 35.2725 SO2 Emissions of industrial sulfur dioxide of each city in
1999, 2004, 2009 and 2014 / t– 924 8.6791 MEDICAL Number of doctors per 10000 people of each city in 1999, 2004, 2009 and 2014 + 932 2.8203 Control variables UNISTU Number of college students per 10000 people of each city in
1999, 2004, 2009, and 2014+ 932 3.6455 POP Population sizes of each city in 1999, 2004, 2009, and 2014 / (10000 persons) + 932 5.8624 FAI Per capita fixed assets investment of each city in
1999, 2004, 2009 and 2014 / (10000 yuan (RMB))NS 931 8.9678 DENS Population density of each city in 1999, 2004, 2009 and 2014 / (persons/km2) + 916 6.4728 Notes: ‘+’ denotes a positive expected effect, ‘–’ denotes a negative expected effect, ‘NS’ represents an unsure expected effect; All independent variables are in natural logarithm expect for INDUS, STEEXPEND, SPEND, and GREEN -
This section determines the extent of spatial concentration of talent across cities in China. We divided 309 cities into three types of areas based on their CI score: intensively-distributed areas (CI ≥ 2), evenly-distributed areas (0.5 < CI < 2), and sparsely-distributed areas (CI ≤ 0.5). We analyzed the distribution of talent across these areas and the results are reported in Table 2. We found that talent tended to concentrate on a small number of intensively-distributed areas with an increasing density from 32.28 persons/km2 in 2000 to 99.97 persons/km2 in 2015. Just over 10% of China’s land area clustered about 70% of the highly educated population during the 2000–2015 period.
Table 2. Distribution characteristics of talent from 2000 to 2015
Area Talent proportion / % Land proportion / % Density / (persons/km2) 2000 2005 2010 2015 2000 2005 2010 2015 2000 2005 2010 2015 Intensively-distributed area 66.96 66.59 70.65 65.16 11.60 12.09 12.30 11.67 32.38 46.03 85.09 99.97 Evenly-distributed area 28.03 27.16 23.40 28.84 26.36 25.94 23.64 27.94 5.97 8.75 14.67 18.48 Sparsely-distributed area 5.01 6.25 5.28 6.00 62.03 61.97 64.06 60.39 0.45 0.84 1.22 1.78 Most of the highly educated population has tended to concentrate in large urban agglomerations and provincial capitals, particularly in the eastern coastal region of the country (Fig. 2). This region encompasses intensively-distributed, economically developed areas, including first-tier cities such as Beijing and Shanghai, capital cities in the central and eastern regions such as Zhengzhou, Taiyuan, Wuhan, and Nanjing, and the cities in Pearl River Delta urban agglomerations such as Dongguan, Zhuhai and Zhongshan.
Figure 2. The concentration index of spatial patterns of talent at the city-level of China, 2000–2015. Hong Kong, Macao and Taiwan of China are not included
Evenly-distributed areas were mainly located in the central-eastern areas of China and have seen a fluctuation in the share of the highly educated population, with a slight rise in 2015. Yet, these areas displayed a significant increase in the density of talent. Sparsely-distributed areas emerged persistently in western and north-west parts of the country. They occupied the majority of the land area (more than 60%) but have consistently contained only less than 6% of China’s highly educated population though they recorded a small increase in density reflecting a reduction in land area.
We also mapped the Hu line and found that the majority of intensively and evenly-distributed areas were located in the southeast of the Hu line (Hu, 1935). Averagely, 94.7% of China’s highly educated population was distributed in the southeast of the Hu line (covered 44.1% of the country’s total land areas) between 2000 and 2015. Yet, the northwest region only had 5.2% of the highly educated population and 55.9% land areas. This finding was consistent with the research on the Hu line’s segregation on the distribution of China’s total population (Wang et al., 2019).
Besides, significant spatial autocorrelation was detected in talent distribution. We computed the Moran’s I for the density of talent of cities between 2000 and 2015 in China and observed that this value has increased from 0.030 (P < 0.05) in 2000 to 0.073 (P < 0.001) in 2015. This indicated that the density of talent of a city was closely related to that of its nearby cities and that this spatial dependence degree has strengthened over time. Also, the distribution of talent showed spatial inequality. The CV is 1.855 in 2000 to 2.073 in 2005 and down to 1.862 in 2015, indicating a persistent pattern of high spatial concentration in the density of highly educated population across cities, with little variation over time. The increasing uneven trend in the first ten years was closely related to a significant difference in college enrollment expansion in particular cities, including first-tier cities (e.g., Beijing and Shanghai) and provincial capitals (e.g., Wuhan and Chengdu) after the introduction of the proposed the ‘Action Plan for Education Revitalization in the 21st Century’. This plan has resulted in a variation in the local supply of educated people across cities. The slight decrease in the concentration of talent between 2010 and 2015 may reflect recent national initiatives of regional development to reduce spatial inequalities (e.g., the ‘National Medium- and Long-Term Talent Development Plan (2010–2020)’ and the ‘National New-Type Urbanization Plan (2014–2020)’.
Underpinning this national pattern of spatial concentration, stark regional differences in the distribution of talent exists. To explore this, we obtained the average density of talent in China’s four economic-geography regions from 2000 to 2015 (Fig. 3). The eastern region consistently recorded the highest density of talent, displaying a significant increase from 20 persons/km2 in 2000 to over 70 persons/km2 in 2015. The western region reported the lowest and most marginal density level remaining below 10 people/km2 since 2000. The results further showed that after the implementation of the ‘China Western Development’ in 1999, a larger proportion of talent moved away from the central, northeastern, and western parts of China to the eastern part, and this trend has been more evident over time.
-
To control time-fixed effects, we constructed three time dummy variables in all the models. We tested for strict multi-collinearity of models with the VIF test under the specification of pooled ordinary least squares (OLS). The VIF values of each variable were all less than 4, indicating that there was no strict multi-collinearity problem. We also calculated the covariance matrix for all variables and found that the pair-wise correlation coefficients did not exceed 0.7. Furthermore, including time-lagged independent variables in our models helps reducing potential endogeneity caused by reverse causation. Additionally, the use of ESF effectively reduces the effect of spatial autocorrelation in residuals and further alleviates any potential endogeneity (Getis and Griffith, 2002; Gu et al., 2019b).
To test for spatial autocorrelation, we calculated a panel-type standardized Moran’s I for talent stock in 233 cities during 2000–2015. The Moran’s I for this test was 0.125 (P < 0.01), indicating statistically significant and positive spatial autocorrelation in the spatial pattern of talent in cities. To reduce the effect of the spatial autocorrelation, it would be good to employ the ESF specification to perform our regression analysis. All models in the study were constructed under the specification of the NBPM. The model results are reported in Table 3. Model 1 only includes three economic variables while Model 2 includes six amenity variables. Model 3 incorporates economic and amenity variables together, followed by a more general model (Model 4) considering all the control variables. Model 5 is an ESF NBPM, where selected eigenvectors are added to control the effect of the autocorrelation. Ultimately, the results of Model 6 using the bootstrap standard errors serve as the robustness check of our main findings.
Table 3. Results from the negative binomial panel models and eigenvector spatial filtering negative binomial panel models
Variables (1) NBPM TALENT (2) NBPM TALENT (3) NBPM TALENT (4) NBPM TALENT (5) ESF NBPM TALENT (6) ESF NBPM TALENT GDP 0.2547*** 0.2487*** 0.0975*** 0.1345** 0.1345* (0.0422) (0.0430) (0.0588) (0.0582) (0.0837) EMPLOY 0.0654* 0.0952** 0.1809*** 0.1726*** 0.1726*** (0.0391) (0.0442) (0.0476) (0.0486) (0.0650) INDUS 0.0006 –0.0009 –0.0036 –0.0025 –0.0025 (0.0021) (0.0022) (0.0024) (0.0023) (0.0026) STEEXPEND 0.0107*** 0.0113*** 0.0065** 0.0061** 0.0061** (0.0030) (0.0030) (0.0031) (0.0031) (0.0028) SPEND –0.0002 –0.0001 –0.0001 –0.0001 –0.0001 (0.0001) (0.0001) (0.0001) (0.0001) (0.0001) PRIEDU 0.1341** 0.1729*** 0.1795*** 0.1931*** 0.1931** (0.0641) (0.0652) (0.0629) (0.0619) (0.0778) GREEN 0.0026*** 0.0026*** 0.0029*** 0.0028*** 0.0028*** (0.0006) (0.0006) (0.0006) (0.0006) (0.0009) SO2 0.0192 0.0090 0.0035 0.0057 0.0057 (0.0128) (0.0126) (0.0035) (0.0127) (0.0123) MEDICAL –0.0345 –0.0546 –0.0333 –0.0501 –0.0501 (0.0494) (0.0521) (0.0512) (0.0515) (0.0631) UNISTU 0.0124* 0.0138** 0.0138** (0.0060) (0.0062) (0.0058) POP 0.4222*** 0.3603*** 0.3603*** (0.0922) (0.0916) (0.1168) FAI –0.0016 0.0007 0.0007 (0.0314) (0.0317) (0.0373) DENS –0.0287 –0.0344 –0.0344 (0.0241) (0.0234) (0.0244) year2005 0.2242*** 0.1393 0.0639 0.1717 0.1260 0.1260 (0.0388) (0.1112) (0.1110) (0.1274) (0.1249) (0.1359) year2010 0.5735*** 0.6362*** 0.3745*** 0.6162*** 0.5449*** 0.5449*** (0.0619) (0.1142) (0.1233) (0.1454) (0.1437) (0.1703) year2015 0.6587*** 0.9533*** 0.4807*** 0.7406*** 0.6459*** 0.6459*** (0.0941) (0.1125) (0.1393) (0.1684) (0.1669) (0.2078) CONSTANT –1.9983*** 1.2035*** –3.1975*** –3.6825*** –3.8318*** –3.8318*** (0.6704) (0.4178) (0.8282) (0.8451) (0.8376) (1.1247) N 926 921 915 898 898 898 Eigenvectors no no no no yes yes AIC 16675 16570 16390 15976 15970 15970 Log likelihood –8330.4301 –8275.0203 –8182.1567 –7970.8279 –7959.6980 –7959.6980 Notes: *** represents P < 0.01, ** represents P < 0.05, * represents P < 0.1; Standard errors are in parentheses for Models 1 to 5, and bootstrap standard errors are in parentheses for Model 6. Meanings of variables see Table 1 -
This section begins with the results of each model. Results from Model 1 showed statistically significant coefficients of GDP (GDP) and average urban employed staff and workers per 10 000 habitants (EMPLOY), indicating the positive impacts of economic development level and job opportunities. However, the coefficient of INDUS was insignificant. Results from Model 2 revealed significant coefficients for scientific technology and education expenditure proportion (STEEXPEND), the number of primary school teachers per 10 000 students (PRIEDU), and regional greening rate (GREEN), indicating the roles of urban public services and natural comforts in driving the distribution of talent. However, SPEND and SO2 seemed to have no relationship with the stock of talent. This was in part because the expenditure-income ratio represents both the ability of public services and the local financial burden, which may lead to an ambiguous effect. Likewise, emissions of industrial sulfur dioxide can measure both environmental pollution and local economic level to some degree. Also, we found that the effect of local medical services was insignificant.
Results from Model 3 showed similar magnitudes of coefficients for economic-related variables and amenity variables as Models 1 and 2, which, to some extent, verified the robustness of the impact of the twofold factors on local talent stock. After we controlled the economic variables, the effect of SO2 was still insignificant, implying that Chinese talent did not consider air pollution when they decided whether to settle in. Also, they would not take medical services into account.
Model 4 entered four control variables but presented very similar results for economic variables and amenity variables as Models 1–3, despite minor changes in their coefficients and significance. In line with our expectations, the number of college students per 10 000 people (UNISTU), and the local population scale (POP) were positively related to the stock of talent, verifying the impact of talent supply and city size. However, the coefficients of population density (DENS) and per capita fixed assets investment (FAI) were not statistically significant in Model 4.
In Model 5, we further added the selected set of eigenvectors as the proxies for spatial autocorrelation and found that the log-likelihood of the Model 5 increased compared to Model 4, together with a decline in its AIC value. This indicated that the ESF NBPM had a better model fitting, which outperformed the models without eigenvectors. Likewise, both of the individual and time fixed effects were controlled. Results from Model 5 unveiled a significant effect of economic opportunities and amenities on talent distribution. The regional economic development level played a prominent role. GDP reflects higher economic benefits for local communities and population in terms of income and consumption spillover effects, generating further consumption in the local economy as well as job creation. The results indicated that an average of a 1% increase in urban GDP led to a 0.1345% increase in the stock of talent between 2000 and 2015. The role of employment chances was also significant, with a 1% increase in the number of average urban staff and workers per 10 000 resulting in a rise in the local talent stock of 0.1726%. The industrial structure of the local economy did not seem to play an important role in influencing the pool of local talent, which was partly due to the differences in the industrial structure were small across most cities.
Specific urban amenities were also significantly associated with the spatial distribution of talent. In contemporary China, a large percentage of talent tends to migrate with their family and a primary concern is an education for their children (Gu et al., 2020b). The primary education context of destination cities is thus important in attracting talent. Our results showed that if the number of primary school teachers per 10 000 students (PRIEDU) increased by 1%, the pool of talent expanded by 0.1931%. Other urban amenities also played a role but to a lesser extent. The ratio of per capita science, technology, and education expenditure to financial expenditure (STEEXPEND) emerged as a key factor representing the importance placed by the local government on urban technology and education development. Our results indicated that a 1% increase in this ratio resulted in a rise in the stock of talent of the city of 0.0061%. Besides, the rate of urban greening (GREEN) displayed a positive relationship with talent stock. Model 5 suggested that a 1% increase in urban greening rate led to a 0.0023% increase in the talent stock of the city.
Out of our expectation, other amenity variables were insignificant in Model 5, including the proportion of per capita fiscal expenditure to fiscal revenue (SPEND), sulfur dioxide emissions (SO2), and the number of doctors per 10 000 (MEDICAL), which implied that the relationship between urban amenities and the distribution of talent was not much clear. This might be partly because there were various manners in our model that measured urban amenities. Fortunately, the three significant variables of amenities had a strong relationship between the distribution of talent across the models, which showed the crucial role of amenities, especially for public services and greening rate.
-
We observed the robustness of our model results in the following two aspects. First, we adopted a strategy for adding variables by step. We found that key variables of economic opportunities and amenities remained significant when we gradually put control variables and eigenvectors into the model. Also, when only entering economic or amenity variables, the key variables were still significant. This has confirmed the robust relationship between economic opportunities, amenities, and the stock of talent in the city. Beyond that, we constructed a model with a bootstrap technique (resampling 400 times) to compute the robust standard errors (Model 6) and regressed the dependent variable on the same set of variables. The results reported in Table 3 revealed that the impact of economic and amenity variables was still robust. In summary, through econometric analysis, we found that urban GDP, average urban employed staff and workers per 10 000, the ratio of per capita science, technology, and education expenditure, the number of middle school teachers per 10 000 students, and urban greening rate had a robust and critical relationship between talent distribution in Chinese cities during the fifteen years.
Geography of Talent in China During 2000–2015: An Eigenvector Spatial Filtering Negative Binomial Approach
-
Abstract: The increase in China’s skilled labor force has drawn much attention from policymakers, national and international firms and media. Understanding how educated talent locates and re-locates across the country can guide future policy discussions of equality, firm localization and service allocation. Prior studies have tended to adopt a static cross-national approach providing valuable insights into the relative importance of economic and amenity differentials driving the distribution of talent in China. Yet, few adopt longitudinal analysis to examine the temporal dynamics in the stregnth of existing associations. Recently released official statistical data now enables space-time analysis of the geographic distribution of talent and its determinants in China. Using four-year city-level data from national population censuses and 1% population sample surveys conducted every five years between 2000 and 2015, we examine the spatial patterns of talent across Chinese cities and their underpinning drivers evolve over time. Results reveal that the spatial distribution of talent in China is persistently unequal and spatially concentrated between 2000 and 2015. It also shows gradually strengthened and significantly positive spatial autocorrelation in the distribution of talent. An eigenvector spatial filtering negative binomial panel is employed to model the spatial determinants of talent distribution. Results indicate the influences of both economic opportunities and urban amenities, particularly urban public services and greening rate, on the distribution of talent. These results highlight that urban economic- and amenity-related factors have simultaneously driven China’s talent’s settlement patterns over the first fifteen years of the 21st century.
-
Key words:
- talent distribution /
- determinants /
- eigenvector spatial filtering /
- panel data analysis /
- China
-
Table 1. Description and expected effects of variables
Type Variable Description Expected effect Number Mean Dependent variables TALENT Number of people with a college degree or above of each city in 2000, 2005, 2010, 2015 932 359602.1577 Economic opportunity variables GDP Gross GDP of each city in 1999, 2004, 2009 and 2014 / (10000 yuan (RMB)) + 932 15.6558 EMPLOY Average number of urban employed staff and workers per 10000 habitants in
1999, 2004, 2009, and 2014+ 927 6.8535 INDUS The proportion of tertiary industry to GDP of each city in 1999, 2004,
2009 and 2014 / %+ 931 36.4876 Amenity variables STEEXPEND The proportion of per capita science, technology and education expenditure
to the financial expenditure of each city in 1999, 2004, 2009 and 2014 / %+ 932 18.1333 SPEND The ratio of per capita financial expenditure to per capita fiscal revenue of
each city in 1999, 2004, 2009 and 2014 / %+ 932 214.6244 PRIEDU Number of primary school teachers per 10000 primary school students of
each city in 1999, 2004, 2009 and 2014+ 931 6.2942 GREEN Greening rate of each city in 1999, 2004, 2009 and 2014 / % + 929 35.2725 SO2 Emissions of industrial sulfur dioxide of each city in
1999, 2004, 2009 and 2014 / t– 924 8.6791 MEDICAL Number of doctors per 10000 people of each city in 1999, 2004, 2009 and 2014 + 932 2.8203 Control variables UNISTU Number of college students per 10000 people of each city in
1999, 2004, 2009, and 2014+ 932 3.6455 POP Population sizes of each city in 1999, 2004, 2009, and 2014 / (10000 persons) + 932 5.8624 FAI Per capita fixed assets investment of each city in
1999, 2004, 2009 and 2014 / (10000 yuan (RMB))NS 931 8.9678 DENS Population density of each city in 1999, 2004, 2009 and 2014 / (persons/km2) + 916 6.4728 Notes: ‘+’ denotes a positive expected effect, ‘–’ denotes a negative expected effect, ‘NS’ represents an unsure expected effect; All independent variables are in natural logarithm expect for INDUS, STEEXPEND, SPEND, and GREEN Table 2. Distribution characteristics of talent from 2000 to 2015
Area Talent proportion / % Land proportion / % Density / (persons/km2) 2000 2005 2010 2015 2000 2005 2010 2015 2000 2005 2010 2015 Intensively-distributed area 66.96 66.59 70.65 65.16 11.60 12.09 12.30 11.67 32.38 46.03 85.09 99.97 Evenly-distributed area 28.03 27.16 23.40 28.84 26.36 25.94 23.64 27.94 5.97 8.75 14.67 18.48 Sparsely-distributed area 5.01 6.25 5.28 6.00 62.03 61.97 64.06 60.39 0.45 0.84 1.22 1.78 Table 3. Results from the negative binomial panel models and eigenvector spatial filtering negative binomial panel models
Variables (1) NBPM TALENT (2) NBPM TALENT (3) NBPM TALENT (4) NBPM TALENT (5) ESF NBPM TALENT (6) ESF NBPM TALENT GDP 0.2547*** 0.2487*** 0.0975*** 0.1345** 0.1345* (0.0422) (0.0430) (0.0588) (0.0582) (0.0837) EMPLOY 0.0654* 0.0952** 0.1809*** 0.1726*** 0.1726*** (0.0391) (0.0442) (0.0476) (0.0486) (0.0650) INDUS 0.0006 –0.0009 –0.0036 –0.0025 –0.0025 (0.0021) (0.0022) (0.0024) (0.0023) (0.0026) STEEXPEND 0.0107*** 0.0113*** 0.0065** 0.0061** 0.0061** (0.0030) (0.0030) (0.0031) (0.0031) (0.0028) SPEND –0.0002 –0.0001 –0.0001 –0.0001 –0.0001 (0.0001) (0.0001) (0.0001) (0.0001) (0.0001) PRIEDU 0.1341** 0.1729*** 0.1795*** 0.1931*** 0.1931** (0.0641) (0.0652) (0.0629) (0.0619) (0.0778) GREEN 0.0026*** 0.0026*** 0.0029*** 0.0028*** 0.0028*** (0.0006) (0.0006) (0.0006) (0.0006) (0.0009) SO2 0.0192 0.0090 0.0035 0.0057 0.0057 (0.0128) (0.0126) (0.0035) (0.0127) (0.0123) MEDICAL –0.0345 –0.0546 –0.0333 –0.0501 –0.0501 (0.0494) (0.0521) (0.0512) (0.0515) (0.0631) UNISTU 0.0124* 0.0138** 0.0138** (0.0060) (0.0062) (0.0058) POP 0.4222*** 0.3603*** 0.3603*** (0.0922) (0.0916) (0.1168) FAI –0.0016 0.0007 0.0007 (0.0314) (0.0317) (0.0373) DENS –0.0287 –0.0344 –0.0344 (0.0241) (0.0234) (0.0244) year2005 0.2242*** 0.1393 0.0639 0.1717 0.1260 0.1260 (0.0388) (0.1112) (0.1110) (0.1274) (0.1249) (0.1359) year2010 0.5735*** 0.6362*** 0.3745*** 0.6162*** 0.5449*** 0.5449*** (0.0619) (0.1142) (0.1233) (0.1454) (0.1437) (0.1703) year2015 0.6587*** 0.9533*** 0.4807*** 0.7406*** 0.6459*** 0.6459*** (0.0941) (0.1125) (0.1393) (0.1684) (0.1669) (0.2078) CONSTANT –1.9983*** 1.2035*** –3.1975*** –3.6825*** –3.8318*** –3.8318*** (0.6704) (0.4178) (0.8282) (0.8451) (0.8376) (1.1247) N 926 921 915 898 898 898 Eigenvectors no no no no yes yes AIC 16675 16570 16390 15976 15970 15970 Log likelihood –8330.4301 –8275.0203 –8182.1567 –7970.8279 –7959.6980 –7959.6980 Notes: *** represents P < 0.01, ** represents P < 0.05, * represents P < 0.1; Standard errors are in parentheses for Models 1 to 5, and bootstrap standard errors are in parentheses for Model 6. Meanings of variables see Table 1 -
[1] Anselin L, 1988. Spatial Econometrics: Methods and Models. Dordrecht: Kluwer. [2] Arntz M, 2010. What attracts human capital? Understanding the skill composition of interregional job matches in Germany. Regional Studies, 44(4): 423–441. doi: 10.1080/00343400802663532 [3] Breznitz D, Murphree M, 2011. Run of the Red Queen: Government, Innovation, Globalization, and Economic Growth in China. London: Yale University Press. [4] Cameron A C, Trivedi P K, 2013. Regression Analysis of Count Data (Vol. 53). Cambridge: Cambridge University Press. [5] Casado-Díaz J M, Martínez-Bernabéu L, Rowe F, 2017. An evolutionary approach to the delimitation of labour market areas: an empirical application for Chile. Spatial Economic Analysis, 12(4): 379–403. doi: 10.1080/17421772.2017.1273541 [6] Chun Y, Griffith D A, 2011. Modeling network autocorrelation in space–time migration flow data: an eigenvector spatial filtering approach. Annals of the Association of American Geographers, 101(3): 523–536. doi: 10.1080/00045608.2011.561070 [7] Chun Y, 2014. Analyzing space-time crime incidents using eigenvector spatial filtering: an application to vehicle burglary. Geographical Analysis, 46(2): 165–184. doi: 10.1111/gean.12034 [8] Chi G Q, Zhu J, 2008. Spatial regression models for demographic analysis. Population Research and Policy Review, 27(1): 17–42. doi: 10.1007/s11113-007-9051-8 [9] Clark T N, Lloyd R, Wong K K et al., 2002. Amenities drive urban growth. Journal of urban affairs, 24(5): 493–515. doi: 10.1111/1467-9906.00134 [10] Deng Wei, Zhang Shaoyao, Zhou Peng et al., 2020. Spatiotemporal characteristics of rural labor migration in China: evidence from the migration stability under new-type urbanization. Chinese Geographical Science, 30(5): 749–764. doi: 10.1007/s11769-020-1147-7 [11] Faggian A, Corcoran J, Rowe F, 2016. Evaluating the effects of Australian policy changes on human capital: the role of a graduate visa scheme. Environment and Planning C: Government and Policy, 34(1): 151–170. doi: 10.1177/0263774X-15614755 [12] Faggian A, Corcoran J, Rowe F, 2017. Special issue on youth and graduate migration. The Annals of Regional Science, 59(3): 571–575. doi: 10.1007/s00168-017-0845-2 [13] Florida R, 2002. The economic geography of talent. Annals of the Association of American Geographers, 92(4): 743–755. doi: 10.1111/1467-8306.00314 [14] Florida R, Mellander C, Qian H, 2012. China’s development disconnect. Environment and Planning A, 44(3): 628–648. doi: 10.1068/a44284 [15] Getis A, Griffith D A, 2002. Comparative spatial filtering in regression analysis. Geographical Analysis, 34(2): 130–140. doi: 10.1111/j.1538-4632.2002.tb01080.x [16] Glaeser E L, Kolko J, Saiz A, 2001. Consumer city. Journal of Economic Geography, 1(1): 27–50. doi: 10.1093/jeg/1.1.27 [17] Graves P E, 1976. A reexamination of migration, economic opportunity, and the quality of life. Journal of Regional Science, 16(1): 107–112. doi: 10.1111/j.1467-9787.1976.tb00954.x [18] Griffith D A, 2003. Spatial Autocorrelation and Spatial Filtering: Gaining Understanding Through Theory and Scientific Visualization. Berlin: Springer Science & Business Media. [19] Gu H Y, Meng X, Shen T Y et al., 2019a. China’s highly educated talents in 2015: Patterns, determinants and spatial spillover effects. Applied Spatial Analysis and Policy, 1–18. doi: 10.1007/s12061-019-09322-6 [20] Gu H Y, Liu Z L, Shen T Y et al., 2019b. Modelling interprovincial migration in China from 1995 to 2015 based on an eigenvector spatial filtering negative binomial model. Population, Space and Place, 25(8): e2253. [21] Gu H Y, Jie Y Y, Li Z T et al., 2020a. What drives migrants to settle in Chinese cities: a panel data analysis. Applied Spatial Analysis and Policy, 1–18. doi: 10.1007/s12061-020-09358-z [22] Gu H Y, Liu Z L, Shen T Y, 2020b. Spatial pattern and determinants of migrant workers’ interprovincial hukou transfer intention in China: evidence from a national migrant population dynamic monitoring survey in 2016. Population, Space and Place, 26(2): e2250. doi: 10.1002/psp.2250 [23] Gu H Y, Ling Y K, Shen T Y et al., 2020c. How does rural homestead influence the hukou transfer intention of rural-urban migrants in China? Habitat International, 105: 102267. doi: 10.1016/j.habitatint.2020.102267 [24] Harris J R, Todaro M P, 1970. Migration, unemployment & development: a two-sector analysis. American Economic Review, 60(1): 126–142. [25] Hausman J, Hall B H, Griliches Z, 1984. Econometric models for count data with an application to the patents-R&D relationship. Econometrica, 52(4): 909–938. doi: 10.2307/1911191 [26] Henderson, J V, 2007. Understanding knowledge spillovers. Regional Science and Urban Economics, 37(4): 497–508. doi: 10.1016/j.regsciurbeco.2006.11.010 [27] Hong E, Sun L X, 2011. Foreign direct investment and total factor productivity in China: a spatial dynamic panel analysis. Oxford Bulletin of Economics and Statistics, 73(6): 771–791. doi: 10.1111/j.1468-0084.2011.00672.x [28] Hu Huanyong, 1935. The distribution of population in China, with statistics and maps. Acta Geographica Sinica, 2(2): 33–74. (in Chinese) [29] Knapp T A, Gravest P E, 1989. On the role of amenities in models of migration and regional development. Journal of regional science, 29(1): 71–87. doi: 10.1111/j.1467-9787.1989.tb01223.x [30] Lao X, Gu H Y, 2020. Unveiling various spatial patterns of determinants of hukou transfer intentions in China: a multi-scale geographically weighted regression approach. Growth and Change, 1–17. doi: 10.1111/grow.12429 [31] Lewis W A, 1954. Economic development with unlimited supplies of labour. The Manchester School, 22(2): 139–191. doi: 10.1111/j.1467-9957.1954.tb00021.x [32] Li Hongsheng, Wang Yingjie, Han Jiafu, 2012. Origin distribution patterns and floating population modeling: Yiwu City as a destination. Chinese Geographical Science, 22(3): 367–380. doi: 10.1007/s11769-012-0534-0 [33] Li Q, Xie Z, 2020. Spatial distribution of floating young talents and influencing factors of their settlement intention-based on the dynamic monitoring data of the national floating population in 2017. Economic Geography, 40(9): 27–34. (in Chinese) [34] Liu Y, Shen J F, 2014. Jobs or amenities? Location choices of interprovincial skilled migrants in China, 2000−2005. Population, Space and Place, 20(7): 592–605. doi: 10.1002/psp.1803 [35] Liu Y, Shen J F, 2017. Modelling skilled and less-skilled interregional migrations in China, 2000−2005. Population, Space and Place, 23(4): e2027. doi: 10.1002/psp.2027 [36] Liu Y, Shen J F, Xu W et al., 2017. From school to university to work: migration of highly educated youths in China. The Annals of Regional Science, 59(3): 651–676. doi: 10.1007/s00168-016-0753-x [37] Liu Y, Xu W, 2017. Destination choices of permanent and temporary migrants in China, 1985−2005. Population, Space and Place, 23(1): e1963. doi: 10.1002/psp.1963 [38] Lucas Jr R E, 1988. On the mechanics of economic development. Journal of Monetary Economics, 22(1): 3–42. doi: 10.1016/0304-3932(88)90168-7 [39] Miguélez E, Moreno R, Suriñach J, 2010. Inventors on the move: tracing inventors’ mobility and its spatial distribution. Papers in Regional Science, 89(2): 251–274. doi: 10.1111/j.1435-5957.2010.00280.x [40] National Bureau of Statistics of China, 2001. Tabulation on the 2000 Population Census of the People’s Republic of China by County. Beijing: China Statistics Press. [41] National Bureau of Statistics of China, 2006. Tabulation on the 2005 1% National Population Sample Survey of the People’s Republic of China. Beijing: China Statistics Press. [42] National Bureau of Statistics of China, 2011. Tabulation on the 2010 Population Census of the People’s Republic of China by County. Beijing: China Statistics Press. [43] National Bureau of Statistics of China, 2016. Tabulation on the 2015 1% National Population Sample Survey of the People’s Republic of China. Beijing: China Statistics Press. [44] National Bureau of Statistics of China, 2000, 2005, 2010, 2015. China City Statistical Yearbook in 1999, 2004, 2009, 2014. Beijing: China Statistics Press. [45] Nie Jingxin, Liu Helin, 2018. Spatial pattern and the resulting characteristics of talent Flows in China. Scientia Geographica Sinica, 38(12): 1979–1987. (in Chinese) [46] Partridge M D, 2010. The duelling models: NEG vs amenity migration in explaining US engines of growth. Papers in Regional Science, 89(3): 513–536. doi: 10.1111/j.1435-5957.2010.00315.x [47] Qian H F, 2010. Talent, creativity and regional economic performance: the case of China. The Annals of Regional Science, 45(1): 133–156. doi: 10.1007/s00168-008-0282-3 [48] Qi Honggang, Qi Wei, Liu Shenghe, 2020. Talents concentration in the Guangdong-Hong Kong-Macao Greater Bay Area, China: Evolution pattern and driving factors. Geographical Research, 39(9): 2000–2014. (in Chinese) [49] Rodríguez-Pose A, Tselios V, 2011. Mapping the European regional educational distribution. European Urban and Regional Studies, 18(4): 358–374. doi: 10.1177/0969776411399345 [50] Rowe F, 2013. The geography and determinants of regional human capital in eight Latin American and Caribbean countries. Regional Problems and Policies in Latin America (pp. 379–405). Berlin, Heidelberg: Springer [51] Rowe F, 2017. The CHilean Internal Migration (CHIM) database: temporally consistent spatial data for the analysis of human mobility. Region, 4(3): 1–6. doi: 10.18335/region.v4i3.198 [52] Rowe F, Corcoran J, Faggian A, 2013. Mobility patterns of overseas human capital in Australia: the role of a ‘new’graduate visa scheme and rural development policy. Australian Geographer, 44(2): 177–195. doi: 10.1080/00049182.2013.789589 [53] Scott A J, 2010. Jobs or amenities? Destination choices of migrant engineers in the USA. Papers in Regional Science, 89(1): 43–63. doi: 10.1111/j.1435-5957.2009.00263.x [54] Shenzhen Municipal Government, 2019. Opinions on the implementation of the ‘Peacock Plan’ for the introduction of high-level overseas talents. Available at: http://www.gd.gov.cn/zwgk/zcfgk/content/post_2531436.html [55] Sjaastad L A, 1962. The costs and returns of human migration. Journal of Political Economy, 70(5, Part 2): 80–93. Available at: https://www.jstor.org/stable/1829105 [56] Song H S, Zhang M, Wang R Q, 2016. Amenities and spatial talent distribution: evidence from the Chinese IT industry. Cambridge Journal of Regions, Economy and Society, 9(3): 517–533. doi: 10.1093/cjres/rsw017 [57] State Council of China, 2010. The National Medium and Long-Term Talent Development Plan (2010−2020). Available at: http://www.gov.cn/jrzg/2010-06/06/content_1621777.htm. (in Chinese) [58] Tang A Z R, Rowe F, Corcoran J et al., 2014. Where are the overseas graduates staying on? Overseas graduate migration and rural attachment in Australia. Applied Geography, 53: 66–76. doi: 10.1016/j.apgeog.2014.06.004 [59] Todaro M P, 1969. A model of labor migration and urban unemployment in less developed countries. The American Economic Review, 59(1): 138–148. [60] Wang Fahui, Liu Cuiling, Xu Yaping, 2019. Analyzing population density disparity in china with GIS-automated regionalization: the Hu line revisited. Chinese Geographical Science, 29(4): 541–552. doi: 10.1007/s11769-019-1054-y [61] Woodward D, Figueiredo O, Guimaraes P, 2006. Beyond the Silicon Valley: university R&D and high-technology location. Journal of Urban Economics, 60(1): 15–32. doi: 10.1016/j.jue.2006.01.002 [62] Xu Z B, Ouyang A J, 2018. The factors influencing China’s population distribution and spatial heterogeneity: a prefectural-level analysis using geographically weighted regression. Applied Spatial Analysis and Policy, 11(3): 465–480. doi: 10.1007/s12061-017-9224-8 [63] Yang Y Z, Feng Z M, Wang L et al., 2014. Research on the suitability of population distribution at the provincial scale in China. Journal of Geographical Sciences, 24(5): 889–906. doi: 10.1007/s11442-014-1127-x [64] Yu Z L, Zhang H, Tao Z L et al., 2019. Amenities, economic opportunities and patterns of migration at the city level in China. Asian and Pacific Migration Journal, 28(1): 3–27. doi: 10.1177/0117196819832054 [65] Zhou L, Tian L, Gao Y et al., 2019. How did industrial land supply respond to transitions in state strategy? An analysis of prefecture-level cities in China from 2007 to 2016. Land Use Policy, 87: 104009. doi: 10.1016/j.landusepol.2019.05.028 [66] Zeng J Y, Liu Y, Wang R Y et al., 2019. Absorptive capacity and regional innovation in China: an analysis of patent applications, 2000−2015. Applied Spatial Analysis and Policy, 12(4): 1031–1049. doi: 10.1007/s12061-019-09300-y