Explaining the urban premium in Chinese cities and the role of place-based policies

Abstract

This paper relies on the empirical framework introduced in Combes et al. (2012) to address the following main questions: (a) what are the relative contributions of agglomeration and selection forces expected to drive the urban productivity premium previously observed in Chinese cities, and (b) to what extent does the industrial parks and zones (IPZs) program, a popular place-based policy, simultaneously influence selection and agglomeration mechanisms? The main findings are as follows. First, both agglomeration and selection forces are observed in larger, denser Chinese cities, indicating that earlier studies that failed to take into account selection likely overestimate the effect of agglomeration economies. Second, after taking into account non-random site selection based on matching, the IPZs program intensifies both agglomeration and selection forces, although the results depend strongly on who administers the program. The empirical findings highlight a theoretical connection between state intervention and explaining the observed urban premium in a transitioning economy context.

Keywords

Agglomeration China place-based policy selection TFP urban premium

Introduction

It is well known that cities offer a number of productivity advantages for firms not available elsewhere, leading to the notion of the “urban premium” (Behrens et al., 2014). Conventional agglomeration theory contends that a number of co-existing factors give rise to agglomeration economies, such as enhanced learning within cities, the sharing of infrastructure and markets, and improved matching of firms and workers (Duranton and Puga, 2004). Despite these positive factors associated with urban agglomerations, a number of negative externalities also exist, including hyper-competition, high rents, and congestion, making the net effects of agglomeration theoretically uncertain (Glaeser, 1998).

In the empirical literature, however, most studies tend to find significantly higher productivity levels in larger and more dense cities (Ciccone and Hall, 1996; Eriksson, 2011; Knoben et al., 2016).¹ Recent meta-works indicate that a doubling of city density is associated with a 4–7% increase in productivity (Melo et al., 2009; Melo et al., 2017). Related studies also show that firms in larger cities exhibit higher export intensity (Malmberg et al., 2000), innovation performance (Moreno et al., 2005), and job mobility (Eriksson et al., 2008). These better performance outcomes observed in the empirical literature are typically interpreted as evidence of agglomeration economies.

Alternative explanations exist, however, that could potentially explain the higher observed performance outcomes in larger cities. A small but growing number of studies, for instance, emphasize the importance of distinguishing between true agglomeration economies and alternative explanations – that is, spatial sorting and selection (Ottaviano, 2011).² Spatial sorting and selection are two processes that can either, in combination or independently, lead to larger aggregate productivity levels even without facilitating agglomeration economies. Ignoring potential sorting or selection effects, which is commonplace in the literature, may result in over-estimating the true role of agglomeration economies.

To tackle this issue, Combes et al. (2012) introduce a novel semi-parametric quantile framework (known as the CDGPR and referred to as such hereafter) capable of distinguishing between selection effects and agglomeration economies. Based on the French context, the authors find empirical support for only agglomeration economies but no selection effects. In other country contexts, however, Arimoto et al. (2014) find evidence for both agglomeration and selection effects in Japan’s silk-reeling industry, while Accetturo et al. (2018) find only evidence of agglomeration economies in the case of Italy.

Following Combes et al. (2012), this paper employs the CDGPR framework to study the underlying sources of the productivity premium in Chinese cities, distinguishing between agglomeration and selection forces. Departing from existing studies, this paper further applies the CDGPR framework to consider what is the role of state-intervention policy in simultaneously influencing the selection and agglomeration mechanisms. The main intervention policy of interest studied in this paper is the industrial parks and zones (IPZs) program, a popular place-based policy in China.³

The empirical framework reveals two main findings. First, we find that the urban premium in China is driven by both agglomeration and selection forces, indicating that prior studies likely overestimate the agglomeration effect as a result of ignoring the role of selection. Second, after taking into account non-random site selection of the IPZs program based on matching techniques, we find that the IPZs program affects both agglomeration and selection forces. The magnitude of the effects depends strongly, however, on who administers the program: IPZs created by local city officials reduce the selection effect in larger cities but also exhibit lower regional competitiveness overall compared to their counterparts created by higher-level officials.

These findings contribute to the agglomeration and development literature in the following ways. First, we apply the CDGPR framework to distinguish between agglomeration and selection forces in a large, transitioning economy context and compare the results with previous findings. A common theme linking these previous studies, however, is that they are limited to only studying advanced market economy contexts (Accetturo et al., 2018; Arimoto et al., 2014; Combes et al., 2012).⁴

Relatedly, the existing agglomeration studies applied to China, and other transitioning economies, tend to find a positive relationship between urban agglomerations and various economic performance measures, that is, productivity, wages, innovation, and stock returns (Au and Henderson, 2006; Combes et al., 2015; Firth et al., 2017; Wang et al., 2010). These observed positive performance outcomes related to urban agglomerations are interpreted as accruing due to genuine agglomeration economies. Yet, the results shown in the present paper indicate that the size of the agglomeration effects are over-estimated by the existing studies, since they ignore the role of selection effects.

As its second main contribution to the literature, this paper relies on the original CDGPR framework to better understand how spatial policies potentially influence the selection and agglomeration forces within a unified framework. In the original CDGPR approach, Combes et al. (2012) assume that the local productivity cut-off point depends solely on local market size. In the current paper, this assumption is relaxed in order to consider the role of place-based policies, in addition to city density, in influencing selection and agglomeration mechanisms. To this end, our framework is capable of testing to what extent place-based policies may contribute together with market size to shape productivity distributions at the local level.

Lastly, the findings from this paper are also relevant to the place-based policy literature. Ongoing debates exist regarding the effectiveness of spatially based policies to increase firm and regional performance (Barca et al., 2012).⁵ Several empirical studies of place-based policies in China suggest that place-based policies improve the performance of targeted firms and regions (Howell, 2019; Koster et al., 2018). Yet, existing studies fail to capture to what extent the positive effects are driven by the exit of poor-performing firms from targeted areas, that is selection.

The structure of this paper is as follows. Section 2 introduces the main theoretical framework of the paper, followed by a description of China’s IPZs in Section 3. Section 4 introduces the data, variable development procedures, and the econometric approach. Section 5 presents the main results based on the original CDGPR model and compares them to the existing findings based on advanced market economy contexts. Section 6 reports the results from the CDGPR model showing the effects of state intervention on selection and agglomeration forces. Section 7 concludes.

Theory and CDGPR model

To distinguish between selection and agglomeration forces, Combes et al. (2012) nest a firm selection model introduced by Melitz and Ottaviano (2008) into a theoretical model of agglomeration economies (Fujita and Ogawa, 1982). The authors’ theoretical framework (referred to as CDGPR) identifies three different forces expected to alter the shape of the productivity distribution: agglomeration (A), dilation (D), and selection (S). The theoretical model implies that agglomeration influences the shape of the distribution in larger regions by positively affecting the productivity of all firms, while dilation makes it less compressed by raising more relatively the productivity of the firms at the right tail of the distribution. By contrast, selection makes the shape of the productivity distribution in larger regions more left-truncated due to tougher selection, that is, a higher minimum productivity level required for firm survival (refer to Combes et al. (2012) for the theoretical proof of their model).

The main implication of CDGPR shows that if cities are ranked in terms of population ( $N_{1} > N_{2} > \dots > N_{R - 1} > N_{R}$ ), then the agglomeration and dilation effects are stronger for larger locations ( $A_{1} > A_{2} > \dots > A_{R - 1} > A_{R}$ and $D_{1} > D_{2} > \dots > D_{R - 1} > D_{R}$ ) and the selection effects are stronger in larger locations ( ${\bar{h}}_{1} > {\bar{h}}_{2} > \dots > {\bar{h}}_{R - 1} > {\bar{h}}_{R}$ ).⁶ Agglomeration shifts the distribution of log productivity rightward by A_i, and dilation asymmetrically increases log productivity of the more productive firms by D_i. Selection makes it more left-truncated due to the presence of tougher selection, whereby a share S_i of entrants with lower productivity values are eliminated.

Model extensions

Combes et al. (2012) introduce their original CDGPR framework by assuming that the local productivity cut-off point depends solely on local market size. More recent studies that adapt the CDGPR framework consider additional mechanisms that may influence the local cut-off threshold in addition to market size. Accetturo et al. (2018), for instance, extend the theoretical model by considering how differences in market access, along with heterogenous market size, influences selection and agglomeration forces.

Similarly, we also relax the assumption in the original CDGPR model that the local productivity cut-off point only depends on the local market size. Rather, we consider the role of place-based policies, in addition to density, in influencing selection and agglomeration mechanisms. That is, we expect that the intensity of the selection and agglomeration effects will differ in areas depending on the share of firms affected by place-based policies, even when the local market size is the same in the two areas.

Our extension is based in part on the well known fact that the role of state intervention is capable of influencing both agglomeration and selection mechanism. State intervention in the form of place-based policies is expected to lead to a rightward shift in the productivity distribution in larger cities. Support for place-based policies is rooted in agglomeration theories that show how the geographical concentration of firms can lead to productivity-enhancing improvements, resulting in a more efficient allocation of resources and greater overall output.⁷

Place-based policies are also expected to influence the extent to which underperforming firms get selected out of the market. In Seidel and von Ehrlich (2011), for instance, the authors develop a new economic geography model that takes into account firm heterogeneity and distortions in the financial market. In their model, low-productivity firms are precluded from external finance, leading to their premature exit from the market. Yet, a policy intervention aimed at promoting financial-market development effectively helped to mitigate these negative selection effects, allowing lower-performing firms to stay in the market longer.

China’s IPZs program

The IPZs program in China has led to the creation of more than 1568 national-level and provincial-level industrial parks and zones, accounting for around 10% of China’s total GDP, despite occupying only 0.1% of its total land (Zheng et al., 2017). The main goals and objectives of the IPZs program aim to achieve four set targets: attracting and using foreign capital, establishing Sino–foreign joint ventures, increasing exports, and promoting a market-driven economy (Wang, 2013). Existing empirical studies typically find that China’s IPZs program led to an improvement across a variety of economic indicators including wages, FDI, employment, output, capital, the overall number of firms, and TFP growth.⁸

Not all of the IPZs in China are administered at the same administrative level, however. Rather they tend to be administered at the local city level or by provincial/central levels of government. Compared to the ones approved by local city governments, IPZs approved by upper levels of government enjoy more favorable policies, such as lower-interest-rate loans, larger tax cuts, cheaper land prices and utility costs, and so forth (Zheng et al., 2017). These higher-level IPZs are noted as typically being well organized, well planned, and generating highly efficient gains (Zeng, 2011).

Following the initial success of IPZs designated by national and provincial government officials, many local governments also began to imitate similar policies to establish their own IPZs (Zeng, 2011). While the goals of the locally established IPZs were to attract investments or protect local enterprises, they were administered without appropriate assessment and planning. As a result, IPZs implemented by local government levels led to a large misallocation of resources, which, combined with “race-to-the-bottom” fiscal incentives, led to the declining efficiency of zone programs. For these reasons, it is likely that the effects of place-based policies on selection and agglomeration forces will vary depending on who implements them.

Empirical framework

Hypotheses

Existing agglomeration studies in China show positive effects of agglomeration in larger cities (Au and Henderson, 2006; Howell et al., 2018; Lu and Tao, 2009). Howell et al. (2016) further show that the effects of agglomeration economies are more pronounced for better-performing firms. For these reasons, we expect to observe a rightward shift in the distribution of log productivity (A), and asymmetrical increases in log productivity for better-performing firms (D) in China’s larger cities.

We also expect to observe a left truncation in the distribution of log productivity (S) in China’s larger cities. This assumption is based on the fact that in the late 1990s and early 2000s, large-scale dismantling of state-owned enterprises led to the massive creation of new firms. Around the same time period, a flood of foreign direct investments was allowed to enter into the Chinese economy, typically being directed to larger cities along the coast. Both the mass creation of new entrepreneurial firms combined with the surge of foreign capital led to intense competition in the economy.

Hypothesis 1: In larger cities, a positive and statistically significant effect exists on agglomeration (A), dilation (D), and selection (S).

Existing studies generally find that regional targeting increases firm and regional performance in China (Howell, 2019; Zheng et al., 2017), which supports the notion that spatial policies promote agglomeration economies. It is therefore expected that larger cities with an IPZ will exhibit a rightward shift in the distribution of log TFP relative to their non-IPZ counterparts. We also expect IPZs to influence the selection mechanism. China relies heavily on industrial policies to shelter domestic firms from foreign competition to promote industrial development and competitiveness within Chinese cities (Haley and Haley, 2013). The effects of state intervention, proxied by IPZs, are therefore expected to help mitigate the selection effect for underperforming firms, implying a reduction in the left truncation of log TFP distribution.

Hypothesis 2: State intervention in larger cities increases agglomeration forces (A and D) but reduces selection effect (S).

Data

This study utilizes the Annual Report of Industrial Enterprise Statistics (ASIF) compiled by the State Statistical Bureau of China for the years 1998 to 2007. The data covers all firms with an annual turnover of approximately 5m RMB (approx. US$600,000), which accounts for 95% of industrial output in China. The firm data contains an unusually extensive list of variables, including information on location, employment, total sales, gross output, geographic location, industry affiliation, financial structure, exports, and so forth.

In addition to the ASIF dataset, we also rely on city statistical yearbooks to obtain the employment information and area size of each city prefecture over the time period. These two variables are used to calculate city size, a proxy for agglomeration size, which takes into account the significance of collective resources and the size of the local labor market. City size is calculated for each city, $Density = \frac{N_{r}}{A_{r}}$ , where N_r is the size of the working population in city r, and A_r is the area of the city (km²). The density variable is then matched to the ASIF data by year and city code.

While labor productivity is generally the most widely used measure of firm productivity, it does not take into account capital intensity. This is a key disadvantage in the case of China where the share of labor earnings in GDP accounts for less than one-half of Chinese manufacturing. Thanks to the detailed ASIF data on firm-level information, we prefer to use total factor productivity (TFP) to measure firm productivity. To correct for simultaneity and endogeneity issues, TFP estimates for Chinese firms are derived using the Ackerberg et al. (2015) approach (ACF), which relies on using intermediate inputs to proxy for unobserved productivity instead of investments. A key advantage of the ACF approach is that it includes fixed effects to take into account firms’ (unmeasured) productivity advantages that persist over time. The construction of additional variables required to estimate TFP, for example a firm’s value added, capital stock and investment, follows closely the outline in Brandt et al. (2012).

Identification of IPZs

Following Howell (2019), we identify whether or not a firm is covered by a new IPZ using information related to the firm’s address, which is then used to obtain TFP distributions for treatment and control firms, respectively, in smaller versus larger cities. Scanning several variables in the ASIF database related to firms’ location (i.e. province, city, county, town or street and doorplate number, sub-district office, neighborhood committee, address, and street), we carry out a text-based analysis in order to help identify whether a firm is covered by an IPZ. Specifically, 16 keywords are used in the text analysis to indicate whether or not a firm is inside of an IPZ exists: kaifa, gaoxin, jingkai, jingji, yuanqu, baoshui, bianjing, kejiyuan, chuangyeyuan, huojuyuan, huojuqu, gongyeyuan, chanyeyuan, gongyequ, gongyexiaoqu, and chukoujiagong.

In addition, postal information is also gathered for the official location of IPZs based on development zone policy reports. To reduce potential measurement bias, this information is used to compare whether the firms identified by the keyword search are indeed located within an area that is officially designated as an IPZ. Overall, less than 2% of firms are misclassified as being within an IPZ when, according to official policy reports, there is no IPZ. The treatment group thus refers only to the firms that are located within postal addresses that contain an IPZ, according to official policy reports, and that contain at least one of the 16 keywords along one of the six firm-identifier variables.

Figure 1 reports the number and share of firms covered by an IPZ and whether that IPZ is implemented by central/provincial governments or by local city government. Over the time period, the share of covered firms increases from around 6% in 1998 to slightly more than 20% in 2009. Higher-share firms tend to be covered by IPZs administered by central and provincial governments – around 13% in 2009 compared to 8% firms covered by IPZs administered by local city governments.

Figure 1.

Number and share of firms located inside of an industrial park and zone (IPZ).

Econometric estimation

The methodology introduced by Combes et al. (2012) is used to simultaneously measure the intensity of selection and agglomeration effects. Since the baseline cumulative of log productivity, $\tilde{F}$ , is not observed, it is not possible to use it to estimate parameters A_i, D_i, and S_i. Combes et al. (2012) get around this issue by comparing the distribution of log productivity across two cities of different size i and j to difference out $\tilde{F}$ (see their paper for the proof).

The novelty of their approach is that it uses non-parametric quantile techniques that rely only on the empirical cumulative distribution of log productivities, F, in each location. Specifically, the authors examine how much of the difference in the productivity distributions of larger cities (>200,000) and smaller cities (<200,000) can be explained by relative agglomeration (rightward shift and dilation of the distribution) and selection (left truncation of the distribution). Moreover, their approach also measures the extent to which the distribution is right-skewed, or dilation, to take into account that agglomeration more relatively benefits the most productive firms.

Based on their model, the estimation strategy is as follows:

F_{i} (ϕ) = \max [0, \frac{F_{j} (\frac{ϕ - A}{D}) - S}{1 - S}], i f S_{i} > S_{j}

(1)

F_{j} (ϕ) = \max [0, \frac{F_{i} (D ϕ + A) - \frac{- S}{1 - S}}{1 - \frac{- S}{1 - S}}], i f S_{j} > S_{i}

(2)

where

D = \frac{D_{i}}{D_{j}}, A = A_{i} - D A_{j}

, and

S = \frac{S_{i} - S_{j}}{1 - S_{j}}

By combining the above two relations, and using the change of variable $u \to r_{s} (u)$ , where $r_{s} (u) = \max (0, \frac{- S}{1 - S)} + [1 - \max (0, \frac{- S}{1 - S})] u$ , the key relationship that can be exploited to fit the model to the data is as follows:

λ_{i} (R_{s} (u)) = D λ_{j} (S + (1 - S) r_{s} (u)) + A

(3)

Equation (3) implies how the quantiles of the productivity distribution in larger cities relate to those in smaller cities through relative parameters A, D, and S. To compare the two distributions symmetrically, Combes et al. (2012) derive the following equality condition:

{\tilde{m}}_{θ} (u) = λ_{j} ({\tilde{r}}_{S} (u)) - \frac{1}{D} λ_{i} (\frac{{\tilde{r}}_{S} (u) - S}{1 - S}) + \frac{A}{D^{'}}

(4)

where

θ = (A, D, S)

and

{\tilde{r}}_{S} (u) = \max (0, S) + [1 - \max (0, S)] u

. Using this condition, Combes et al.’s (2012) estimator

\hat{θ} = (\hat{A}, \hat{D}, \hat{S})

, which is the estimator used in this paper, is described as follows:

\hat{θ} = \arg \min M (θ), where M (θ) = \int_{0}^{1} {[{\hat{m}}_{θ} (u)]}^{2} d u + \int_{0}^{1} {[{\tilde{\hat{m}}}_{θ} (u)]}^{2} d u

(5)

where

\tilde{\hat{m}} (u)

is the empirical counterpart of

\tilde{m} (u)

, obtained by replacing the true quantiles, λ_i and λ_j, by their estimators.

A measure of goodness of fit $R^{2} = 1 - \frac{M (A, S, D)}{M (0, 1, 0)}$ can be derived from the optimization problem in equation (5), which assesses what share of the mean-squared quantile differences between the cluster and non-cluster distributions is accounted for by the estimated set of model parameters.

As in the original CDGPR model, TFP is aggregated over the 1998–2007 time period. Because of the minimum sales threshold enforced on the ASIF data, approximately 30% of firms report information for only one time period. These are the firms that enter and exit the sample in the same year as a result of hovering around the minimum sales threshold. To reduce concerns about firm dynamics (firm exit), only firms that report information for two or more consecutive years are included in the estimation of aggregate city-level TFP.

By aggregating TFP over the time period, the CDGPR model treats the agglomeration and selection forces as being fixed and static over time. This is a key shortcoming of the CDGPR model. It is well known that the relationship between city size and agglomeration effects may exhibit a non-linear, dynamic pattern driven, in part, by the attenuating effects of negative externalities (i.e. congestion costs) on the agglomeration effect over time (Brülhart and Mathys, 2008; Dai et al., 2017). Such dynamic interactions are important for future research but lie outside the purview of CDGPR model and current paper.

A second shortcoming of the CDGPR model is that it focuses only on a selection force, in addition to the agglomeration force, but ignores the potential spatial sorting effect. In their theoretical contribution, Behrens et al. (2014) show that the spatial sorting of higher-skilled workers will lead to larger aggregate productivity levels even in the absence of agglomeration economies. Empirical work further finds that after controlling for worker sorting, it tends to reduce the estimated advantages of city density (Combes and Gobillon, 2015) and the dynamic gains from agglomeration (D’Costa and Overman, 2014).

Despite these two shortcomings, the CDGPR model offers a powerful approach to identify the role of selection versus agglomeration forces and to further explore the question of how IPZs influence selection and agglomeration mechanisms. As summarized in Accetturo et al. (2018), the main benefits of the CDGPR methodology is that it is (a) driven entirely in theory and allows a simultaneous assessment of selection and agglomeration effects; (b) non-parametric, meaning it does not impose parametric assumptions about the underlying distribution of firm-productivity levels; and (c) based on a comparison of all the quantiles of the two distributions, not only of specific percentiles, thereby improving robustness and efficiency of parameter estimation.

Place-based policies, selection, and agglomeration

In this section, the original model by Combes et al. (2012) presented above is used to take into account the effect of IPZs on selection and agglomeration forces. A key empirical concern relates to the non-random assignment of place-based policies, which makes identification particularly difficult. That is, place-based policies often target depressed or low-growth areas, making any comparisons between treated and control areas biased. To correct for non-random policy assignment, propensity score matching (PSM) is employed to ensure the comparability between prefectural cities with and without a new IPZ. The identification strategy is briefly outlined below.

To perform matching, a list of covariates is linked to the probability that a particular firm becomes covered by a new IPZ in order to identify the most appropriate control group. The main covariates include the firm’s share of equity possessed by a state-owned entity, capital intensity, sales growth, and entry rates, as well as city average-employment density and export intensity. All time-variant explanatory variables are lagged by one year in order to mitigate simultaneity concerns. In addition, the estimation also includes a dummy for whether or not a firm receives FDI, a set of firm-size dummies, as well as a set of region, industry, and year dummies. See Appendix A for the matching procedure and results.

After constructing the balanced data, separate sub-samples are created for treated groups (i.e. firms covered by an IPZ) and control groups (i.e. firms not covered by an IPZ) and based on whether they are respectively located in smaller versus larger cities. The same estimation procedure, outlined above, is applied to obtain parameters for A, S, and D, and in this case, these parameters are compared across treated versus non-treated subsamples. The standard errors for the estimated parameters are obtained by 200 bootstrap iterations, including re-estimating firms’ productivities.

Figure 2 compares the TFP distribution using kernel densities for firms located in smaller versus larger cities, using the original, non-matched sample in Panel A and the matched sample in Panel B. In Panel A, the TFP distribution of firms in larger cities is right-shifted, suggesting positive agglomeration effects, especially for higher-performing firms suggesting positive dilation effects. Moreover, the left truncation of the TFP distribution observed in larger cities is indicative of selection effects, for example a larger share of lower-performing firms are indeed selected out of larger cities.

Figure 2.

Agglomeration, selection and TFP distribution for whole sample (Panel A) and matched sample (Panel B).

The relationships observed using the matched sample in Panel B are similar to the ones observed using the full sample in Panel A, suggesting that the matching process did not distort underlying economic relationships. These results offer some suggestive, albeit tentative, evidence suggesting that selection and agglomeration forces are at play in the Chinese context. The next section applies the semi-parametric method introduced by Combes et al. (2012) to further distinguish between selection and agglomeration effects and the potential moderating role of place-based policies.

Baseline results

In this section, the main goal is to demonstrate the role of selection and agglomeration in China and compare the outcomes to findings from the existing literature. To facilitate such comparisons, the estimation procedure to obtain the parameters for A, S, and D follows closely the original one introduced by Combes et al. (2012). Specifically, the standard errors for the estimated parameters are obtained by 200 bootstrap iterations, including re-estimating firms’ productivities. TFP at the firm level is also averaged across years, setting $T F P_{i} = \frac{1}{T} \sum_{t = 1}^{T} T F P_{i t}$ , where T denotes the number of years the firm is observed.

Table 1 presents the baseline results from comparing TFP distributions for firms located in smaller versus larger cities. A small city is defined as having an employment-density value that is below the mean value across all cities, otherwise it is defined as a larger city. The estimate for relative agglomeration (A) is positive and significantly statistically different from zero. This result indicates a rightward shift in the productivity distribution (or a higher mean productivity) for larger cities compared to smaller cities. The estimate for relative dilation (D) is greater than 1 and is significantly statistically different than 1, thus indicating that the productivity distribution is more dilated in larger cities. The estimate for relative selection (S) is positive and significantly statistically different than zero, revealing that the productivity distribution is more truncated in the left tail for firms located in larger cities.

Table 1.

Results: Estimates of agglomeration (A), dilation (D), and selection (S).

		Constrained specifications
	Baseline	No selection	No dilation
Smaller vs. larger cities	(1)	(2)	(3)
Panel A: Whole sample
A (relative agglomeration)	${0.125}^{* * *}$ $(0.013)$	${0.163}^{* * *}$ $(0.017)$	${0.079}^{* * *}$ $(0.008)$
D (relative dilation)	${1.159}^{* * *}$ $(0.040)$	${1.022}^{* * *}$ $(0.035)$	……
S (relative selection)	${0.023}^{* * *}$ $(0.007)$	……	${0.031}^{* * *}$ $(0.009)$
Total no. firms	539,919	539,919	539,919
No. firms in smaller cities	295,619	295,619	295,619
No. firms in larger cities	244,300	244,300	244,300
Pseudo-R²	0.979	0.913	0.921
Panel B: PSM
A (relative agglomeration)	${0.121}^{* * *}$ $(0.016)$	${0.169}^{* * *}$ $(0.019)$	${0.072}^{* * *}$ $(0.010)$
D (relative dilation)	${1.153}^{* * *}$ $(0.039)$	${1.017}^{* * *}$ $(0.031)$	……
S (relative selection)	${0.021}^{* * *}$ $(0.006)$	……	${0.028}^{* * *}$ $(0.008)$
Total no. firms	470,393	470,393	470,393
No. firms in smaller cities	279,856	279,856	279,856
No. firms in larger cities	190,537	190,537	190,537
Pseudo-R²	0.982	0.917	0.941

PSM: propensity score matching

Notes: $^{*}$ p <0.1; $^{* *}$ p <0.05; $^{* * *}$ p <0.01. Panel A reports the results obtained from estimating models on the whole sample, while Panel B reports the results using the PSM method to balance the data. In column (1), the baseline model results are reported. In column (2), the baseline model is constrained to ignore selection effects (S=0). In column (3), the baseline model is constrained to ignore dilation effects (D=1).

Following Combes et al. (2012), the model is re-estimated using a set of constrained specifications, where (S) is set to zero (no selection) in column (2) and (D) is set to unity (no dilation) in column (3). In column (2), the size of the estimate on agglomeration increases to 0.222 in the constrained case S = 0. The smaller pseudo-R² implies that ignoring (S) leads to misspecification. In column (3), compared to the unconstrained model, the size of the estimate on relative agglomeration decreases in size to 0.064, while the size of the estimate for relative selection increases to 0.041. In Panel B, the results obtained using the matched data remain qualitatively the same as for the whole sample, suggesting that the results are not sensitive to the matching strategy.

In support of Hypothesis 1, the findings indicate that the higher observed productivity of firms in larger cities is driven by both agglomeration effects, which asymmetrically benefit the more productive firms, as well as selection effects. When dilation is constrained to unity, the selection effect arises more strongly, indicating that tougher competition leads to a stronger selection effect for firms located in clusters. Failure to take into account this selection effect will lead to an overestimation of the effect of agglomeration economies. The baseline results on the agglomeration and dilation effects are similar to the initial findings found by Combes et al. (2012). Unlike the results presented by Combes et al. (2012) in the French case, however, selection effects play an important role in explaining the productivity premium in Chinese cities.

Alternate specifications and sensitivity checks

As a sensitivity check, Table 2 re-estimates the parameters for (A), (S), and (D) using several alternative specifications. Column (1) checks whether the baseline findings are sensitive to spatial scale by looking at employment densities at a higher spatial unit, the province level. The results generally confirm the initial findings, although the agglomeration and selection effects are both less pronounced. Column (2) restricts the sample of firms to only the ones that have a total number of employees below the mean number of workers across all firms. The main concern here is that due to the minimum sales threshold of 5m RMB, the ASIF data only includes “above-scale” firms that are likely to benefit from economies of scale. After re-estimating the parameters on the restricted sample, the results confirm that agglomeration, selection, and dilation effects occur in comparatively smaller-sized firms, and that the baseline findings are not merely driven by economies of scale.

Table 2.

Alternate specifications for estimating A, D, and S.

	Alternative spatial unit	Scale economies	Proximity to port
Smaller vs. larger cities (provinces)	(1)	(2)	(3)
A (relative agglomeration)	${0.082}^{* * *}$ $(0.012)$	${0.148}^{* * *}$ $(0.017)$	${0.188}^{* * *}$ $(0.024)$
D (relative dilation)	${1.095}^{* * *}$ $(0.033)$	${1.202}^{* * *}$ $(0.041)$	${1.188}^{* * *}$ $(0.048)$
S (relative selection)	${0.010}^{* * *}$ $(0.003)$	${0.032}^{* * *}$ $(0.006)$	${0.039}^{* * *}$ $(0.010)$
Total no. firms	470,393	351,160	101,226
No. firms in smaller cities (provinces)	289,856	194,993	52,102
No. firms in larger cities (provinces)	180,537	156,167	49,124
Pseudo-R²	0.962	0.942	0.872

Notes: $^{*}$ p <0.1; $^{* *}$ p <0.05; $^{* * *}$ p <0.01. Results are obtained using the matched sample of firms. In column (1), employment densities are computed for the provincial level and firms are split into smaller versus larger provinces based on mean values, in order to take into account different spatial unit of analysis. In columns (2) and (3), the spatial unit returns back to the city level. In column (2), the sample is restricted to only firms that have a number of employees that is below the mean value across all firms. In column (3), the sample is restricted to only cities that are within 100 km to the nearest port.

Finally, column (3) considers access to foreign market by restricting the sample to include only cities that are within 100 km from the nearest port. As found in Accetturo et al. (2018), the results here show that better access to foreign markets indicates lower transportation costs and a stronger selection mechanism. This is because lower transportation costs mean better opportunity for profit, which increases the number of operating firms. The results are similar to the initial findings, although the selection and agglomeration effects are considerably stronger as expected due to the entrance of a large number of firms, particularly firms that target the foreign market.

Combes et al. (2012) show that the size of the estimates on A, S, and D vary considerably across specific industries, particularly for high-tech industries. Taking into account sectoral heterogeneity, Table 3 splits firms into two groupings based on the technological intensity of the firm’s industry. If a firm is an industry that has a research-and-development (R&D) intensity (defined as the ratio of R&D expenditures divided by the total sales revenue) value that is at or below the mean value across all four-digit industries, it is classified as being in a lower-tech industry, otherwise it is classified as being in a higher-tech industry. Column (1) corresponds to the grouping of firms in lower-tech industries, while column (2) corresponds to the grouping of firms in higher-tech industries.

Table 3.

Estimates of agglomeration (A), dilation (D), and selection (S), high-tech versus low-tech industries.

	Lower-tech industries	Higher-tech industries	Difference
Smaller vs. larger cities	(1)	(2)	(3)
A (relative agglomeration)	${0.094}^{* * *}$ $(0.020)$	${0.178}^{* * *}$ $(0.039)$	${0.084}^{* * *}$
D (relative dilation)	${1.101}^{* * *}$ $(0.022)$	${1.214}^{* * *}$ $(0.051)$	${0.113}^{* * *}$
S (relative selection)	${0.011}^{* * *}$ $(0.003)$	${0.034}^{* * *}$ $(0.008)$	${0.023}^{* * *}$
Total no. firms	272,081	198,312
No. firms in smaller cities	173,424	106,432
No. firms in larger cities	98,657	91,880
Pseudo-R²	0.964	0.989

Notes: $^{*}$ p <0.1; $^{* *}$ p <0.05; $^{* * *}$ p <0.01. Results are obtained using the matched sample of firms. In column (1), the sample is restricted to industries that have a research-and-development (R&D) intensity (defined as the ratio of R&D expenditures divided by the total sales revenue) value that is at or below the mean value across all four-digits industries. In column (2), the sample is restricted to industries that have an R&D intensity value that is above the average value across all four-digits industries.

The results show that the size of the coefficients on (A), (D), and (S) are larger in column (2) compared to column (1). As indicated by column (3), the size of the differences is highly statistically significant. The results thus indicate the important role of the technological landscape, suggesting that agglomeration effects as well as selection effects are more pronounced for firms in higher versus lower tech industries.

One explanation for this outcome is that firms in higher-tech industries face a faster-pace environment, whereby firms’ own technological stock can become quickly obsolete. In such environments, firms face increasing pressure to look outside their boundaries and acquire external knowledge inputs from other nearby firms, research labs, and universities to keep pace with the fast-changing technological landscape. At the same time, firms that are unsuccessful in the external search are likely to be selected out much more quickly than in slower-paced industries.

Extended model application: IPZs, agglomeration, and selection

To reveal the effects of IPZs on agglomeration and selection forces, PSM-CDGPR models are re-estimated separately for treated and non-treated locations, using the matched sample that corrects for non-random policy assignment. The results are presented in Table 4. In column (1), the matched firm data is restricted to only the control-group cities without a new IPZ, while column (2) restricts the matched data to only the treated cities with a new IPZ.

Table 4.

Effects of the industrial parks and zones (IPZs) program on the estimates of A, D, and S.

	Control group (IPZ=0)	Treatment group (IPZ=1)	Difference
Smaller vs. larger cities	(1)	(2)	(3)
A (relative agglomeration)	${0.075}^{* * *}$ $(0.014)$	${0.195}^{* * *}$ $(0.019)$	${0.12}^{* * *}$
D (relative dilation)	${1.142}^{* * *}$ $(0.044)$	${1.176}^{* * *}$ $(0.049)$	${0.034}^{* * *}$
S (relative selection)	${0.016}^{* * *}$ $(0.007)$	${0.061}^{* * *}$ $(0.008)$	${0.045}^{* * *}$
Total no. firms	385,102	85,291
No. firms in smaller cities	229,837	50,019
No. firms in larger cities	155,265	35,272
Pseudo-R²	0.988	0.978

Notes: $^{*}$ p <0.1; $^{* *}$ p <0.05; $^{* * *}$ p <0.01. The values refer to the main CDGPR model results applied to the matched sample based on propensity score matching. In column (1), the sample is restricted to only firms that are located outside of an IPZ. In column (2), the sample is restricted to only firms located inside an IPZ.

The results show that the coefficients on selection, agglomeration, and dilation are positive and highly statistically significant for firms in larger cities, irrespective of whether or not they are located in an IPZ. A comparison of the size of the coefficients, however, reveals that the effects of selection and agglomeration are more pronounced in column (2), and the difference in size is statistically significant as shown in column (3). The results suggest that agglomeration and selection forces exist for firms located in larger cities and that IPZs enhance both mechanisms.

These findings provide partial support for Hypothesis 2, at least with respect to the finding that IPZs enhance the agglomeration effects. Counter to expectations, however, IPZs are also found to enhance selection effects, meaning a higher left truncation in the log TFP distribution. One potential explanation for this finding is that firms located within an IPZ area are selected out of the market due to a fixed amount of subsidies applied to the targeted areas, inducing some rationing behavior.

Additional robustness checks

Despite their popularity, it is acknowledged that propensity scores are criticized in the recent literature, since matching only takes place for observables (King and Nielsen, 2016). In light of this concern, an alternative matching technique – Coarsened Exact Matching (CEM) – is employed as a robustness check. CEM is shown to yield less biased estimates of the causal effect across different sample sizes compared to PSM (Iacus et al., 2012). The CEM-CDGPR model results are presented in Panel A of Table 5. The qualitatively similar results help to confirm that the main findings presented above are not driven by the chosen matching procedure.

Table 5.

Alternative specifications and checks for robustness.

	Control group (IDZ=0)	Treatment group (IDZ=1)	Difference
Smaller vs. larger cities	(1)	(2)	(3)
Panel A: CEM
A (relative agglomeration)	${0.073}^{* * *}$ $(0.015)$	${0.199}^{* * *}$ $(0.021)$	${0.126}^{* * *}$
D (relative dilation)	${1.139}^{* * *}$ $(0.048)$	${1.171}^{* * *}$ $(0.052)$	${0.032}^{* * *}$
S (relative selection)	${0.020}^{* * *}$ $(0.008)$	${0.057}^{* * *}$ $(0.0010)$	${0.037}^{* * *}$
Total no. firms	341,832	79,128
No. firms in smaller cities	213,712	41,281
No. firms in larger cities	128,120	37,847
Pseudo-R²	0.981	0.972
Panel B: PSM + additional covariates
A (relative agglomeration)	${0.073}^{* * *}$ $(0.015)$	${0.199}^{* * *}$ $(0.021)$	${0.126}^{* * *}$
D (relative dilation)	${1.139}^{* * *}$ $(0.048)$	${1.171}^{* * *}$ $(0.052)$	${0.032}^{* * *}$
S (relative selection)	${0.020}^{* * *}$ $(0.008)$	${0.057}^{* * *}$ $(0.0010)$	${0.037}^{* * *}$
Total no. firms	341,832	79,128
No. firms in smaller cities	213,712	41,281
No. firms in larger cities	128,120	37,847
Pseudo-R2	0.981	0.972
Panel C: PSM + balanced sample
A (relative agglomeration)	${0.073}^{* * *}$ $(0.015)$	${0.199}^{* * *}$ $(0.021)$	${0.126}^{* * *}$
D (relative dilation)	${1.139}^{* * *}$ $(0.048)$	${1.171}^{* * *}$ $(0.052)$	${0.032}^{* * *}$
S (relative selection)	${0.020}^{* * *}$ $(0.008)$	${0.057}^{* * *}$ $(0.0010)$	${0.037}^{* * *}$
Total no. firms	341,832	79,128
No. firms in smaller cities	213,712	41,281
No. firms in larger cities	128,120	37,847
Pseudo-R²	0.981	0.972

CEM: coarsened exact matching; PSM: propensity score matching

Notes: $^{*}$ p <0.1; $^{* *}$ p <0.05; $^{* * *}$ p <0.01. Panel A presents the CDGPR model results using an alternative matching procedure based on CEM. To check whether potential unobservable covariates could be confounding the main results, Panel B reports the CDGPR model results after including the following two additional covariates into the matching procedure: (a) average wage in logarithm at the firm level and (b) Donaldson and Hornbeck’s (2016) measure of market access constructed at the city level. To check whether firm dynamics may be biasing the main results, Panel C re-runs the original PSM procedure applied to only the balanced sample of firms. In column (1), the sample is restricted to only firms that are located outside of an industrial park and zone (IPZ). In column (2), the sample is restricted to only firms located inside an IPZ.

The main findings may be biased due to the presence of confounding factors not included in the matching procedure. As a robustness check, we re-estimate the PSM-CDGPR model after including new measures of human capital and market access into the matching procedure. In the absence of having information on workers’ educational attainment or work experience, we rely on the firm’s average wages in logarithm as a crude measure of human capital. We further construct a city-level proxy of market access based on the measure introduced in Donaldson and Hornbeck (2016). The results are presented in Panel B of Table 5. The main findings remain qualitatively the same as the original ones, suggesting that our results may not be biased due to unobservable confounding factors.

Another potential issue is that the firm’s TFP is averaged over the whole time period as imposed by the static CDGPR framework. This creates concerns about measurement bias due to firm dynamics related to firms entering and exiting our sample at different time periods. To help reduce this concern, we re-estimate the PSM-CDGPR on only the balanced sample of firms. We report the results in Panel C of Table 5. The results remain qualitatively the same as before, suggesting that firm dynamics are not driving the original findings.

Central/provincial versus local IPZs

In this subsection, we consider whether the IPZ effects on selection and agglomeration depend on whether central/provincial authorities versus local city officials established the IPZ. In Panel A of Table 6, the results for IPZs established by central/provincial authorities reflect the overall findings. By contrast, the results in Panel B reveal different implications on agglomeration and selection forces for IPZs established by local authorities. Specifically, the size of the coefficients on (A), (D), and (S) are smaller in column (2) than in column (1), and the size of the difference is statistically significant as indicated in column (3). The findings thus suggest that place-based policies administered by local city officials mitigate the selection effects, but this comes at the expense of also reducing agglomeration forces.

Table 6.

Heterogeneous effects of the industrial parks and zones (IPZs) program on the estimates of A, D, and S.

	Control group (IDZ=0)	Treatment group (IDZ=1)	Difference
Smaller vs. larger cities	(1)	(2)	(3)
Panel A: Central/provincial IPZs
A (relative agglomeration)	${0.075}^{* * *}$ $(0.014)$	${0.248}^{* * *}$ $(0.026)$	${0.173}^{* * *}$
D (relative dilation)	${1.142}^{* * *}$ $(0.044)$	${1.209}^{* * *}$ $(0.059)$	${0.067}^{* * *}$
S (relative selection)	${0.016}^{* * *}$ $(0.007)$	${0.047}^{* * *}$ $(0.014)$	${0.031}^{* * *}$
Total no. firms	385,102	69,079
No. firms in smaller cities	229,837	40,210
No. firms in larger cities	155,265	28,869
Pseudo-R²	0.988	0.981
Panel B: Local city IPZs
A (relative agglomeration)	${0.075}^{* * *}$ $(0.014)$	${0.056}^{* * *}$ $(0.019)$	$- {0.019}^{* * *}$
D (relative dilation)	${1.142}^{* * *}$ $(0.044)$	${1.124}^{* * *}$ $(0.049)$	$- {0.018}^{* * *}$
S (relative selection)	${0.016}^{* * *}$ $(0.007)$	${0.009}^{* * *}$ $(0.002)$	$- {0.007}^{* * *}$
Total no. firms	385,102	16,212
No. firms in smaller cities	229,837	9,809
No. firms in larger cities	155,265	6,403
Pseudo-R²	0.988	0.975

Notes: $^{*}$ p <0.1; $^{* *}$ p <0.05; $^{* * *}$ p <0.01. Results are obtained using the matched sample of firms. In column (1), the sample is restricted to only firms that are located outside of an IPZ. In column (2) of Panel A, the sample is restricted to only firms located inside of an IPZ created by central and provincial authorities, while column (2) of Panel B is restricted to only firms located inside of an IPZ created by local city officials.

Despite attempts to mimic the success of zones implemented by the central government in the early 1990s, the results tend to confirm that local place-based policies were poorly implemented by city officials, as shown by the reduction in agglomeration forces. Specifically, the creation of local IPZs led to the undesirable outcome of misallocating resources, driven at least partially by enabling a higher share of lower-performing firms to remain in the market. A distortion of the selection effect and other common explanations, such as race-to-the-bottom fiscal policies (Zeng, 2011), are likely to contribute together to the declining efficiency of locally implemented IPZs programs.

Mechanism: Comparing heterogenous payoffs for firms inside versus outside IPZs

A key question that emerges from the results above is why would local officials mitigate the selection mechanism, that is, protect a higher share of lower-performing firms from premature exit, at the expense of reducing their competitiveness, as seen by lower agglomeration effects and lower dilation effects? A straightforward economic rationale is that local officials may anticipate a particularly high payoff, betting that by enabling lower-performing firms to remain in the market, they will eventually be able to benefit from agglomeration economies, for example knowledge spillovers. In turn, these lower-performing firms will be motivated to upgrade their products and technological capabilities and, in aggregate, increase regional competitiveness over time.

It is not possible, however, to test the rationale of local policymakers using Combes et al.’s (2012) methodology, due to its analysis at the aggregated city level. In addition, the creation of IPZs varies both in time and in space; averaging firm-level TFP over the years, as is done above and by Combes et al. (2012), eliminates this information, which might be helpful for estimating the effect of place-based policies.

Therefore, we employ a subsequent micro-based analysis that can better exploit all of the firm-level information available. The main interest here is to examine the effects of new IPZs on firm-level TFP, especially for lower-performing firms. A positive effect of locally created IPZs observed at the bottom end of the TFP distribution would indicate that lower-performing firms indeed benefit from agglomeration economies, thereby helping to explain why local policymakers would keep a large share of lower-performing firms in the market.

To carry out the analysis, we apply a quantile differences-in-differences (QDID) estimator to the matched sample of firms to study the heterogenous effects of place-based policies along the TFP distribution at the firm-level. In the QDID model, the treatment effects are estimated for separate quantiles while including industry and region-fixed effects. To avoid firm dynamics such as firm entry/exit, only incumbent firms are included in the analysis.

A shortcoming of the QDID model is that it only gives the differences-in-differences value using the same quantiles in the treatment and control group. Alternative approaches are also used as robustness checks, such as the changes-in-changes (CIC) model by Athey and Imbens (2006). In addition, the recentered influence function (RIF) method of Firpo et al. (2009) is also used to estimate the unconditional TFP distribution. See Appendix B for a discussion of these approaches.

The results are reported in Table 7. In Panel A, the treatment group is defined as incumbent firms located within a city with a new IPZ created by central and provincial authorities, and the coefficients are respectively obtained from the PSM-QDID, PSM-CIC, and PSM-RIF-DID approaches. The results reveal that the coefficients are positive at the bottom end of the distribution and monotonically increase moving rightward along the quantiles. Moreover, the size of the coefficients at each estimated quantile is significant in both a statistical and economic sense. The results suggest that IPZs established by central and provincial authorities led to a significant productivity gain for the “average” firm, and this is true irrespective of the estimation strategy.

Table 7.

Distributional effects of the industrial parks and zones (IPZs) program on firm productivity.

	Quantile treatment effect
	.1	.25	.5	.75	.9	ATE
Panel A: IPZ $_{Cent / Prov}$
PSM-QDID	${0.200}^{* * *}$ $(0.028)$	${0.231}^{* * *}$ $(0.039)$	${0.283}^{* * *}$ $(0.047)$	${0.336}^{* * *}$ $(0.052)$	${0.342}^{* * *}$ $(0.063)$	${0.275}^{* * *}$ $(0.044)$
PSM-CIC	${0.112}^{* *}$ $(0.036)$	${0.144}^{* * *}$ $(0.041)$	${0.191}^{* * *}$ $(0.050)$	${0.248}^{* * *}$ $(0.062)$	${0.238}^{* * *}$ $(0.076)$	${0.195}^{* * *}$ $(0.047)$
PSM-RIF-DID	${0.101}^{* *}$ $(0.031)$	${0.129}^{* * *}$ $(0.048)$	${0.173}^{* * *}$ $(0.059)$	${0.203}^{* * *}$ $(0.072)$	${0.229}^{* * *}$ $(0.093)$	${0.165}^{* * *}$ $(0.054)$
Panel B: IPZ _Local
PSM-QDID	${0.065}^{* * *}$ $(0.015)$	${0.037}^{* * *}$ $(0.010)$	$- {0.022}^{* * *}$ $(0.006)$	$- {0.045}^{* * *}$ $(0.011)$	$- {0.070}^{* * *}$ $(0.021)$	$- {0.016}^{* * *}$ $(0.005)$
PSM-CIC	${0.025}^{* * *}$ $(0.006)$	${0.013}^{* * *}$ $(0.003)$	$- {0.012}^{* * *}$ $(0.003)$	$- {0.035}^{* * *}$ $(0.008)$	$- {0.047}^{* * *}$ $(0.015)$	$- {0.011}^{* * *}$ $(0.002)$
PSM-RIF-DID	${0.009}^{* * *}$ $(0.003)$	${0.006}^{* * *}$ $(0.002)$	$- {0.003}^{* * *}$ $(0.001)$	$- {0.013}^{* * *}$ $(0.003)$	$- {0.019}^{* * *}$ $(0.04)$	$- {0.004}^{* * *}$ $(0.001)$
No. firms	470,393	470,393	470,393	470,393	470,393	470,393
Industry FE	Yes	Yes	Yes	Yes	Yes	Yes
Region FE	Yes	Yes	Yes	Yes	Yes	Yes

Notes: Matching techniques are respectively combined with QDID estimator, the CIC estimator, and the RIF-DID estimator. Separate models are estimated for IPZs created by provincial and central authorities (Panel A) and for IPZs created by local city officials (Panel B). Robust standard errors, reported in parentheses, are obtained with 500 bootstrap repetitions. Average Treatment effect (ATE); Fixed-effects (FE).

In Panel B, by contrast, the results show that at the bottom part of the TFP distribution, the coefficients are positive and statistically significant but turn negative at the upper quantiles. An inspection of the coefficients further suggests that the size of the positive effects for lower-performing firms is trivial and insignificant in an economic sense, especially when using the CIC and RIF-DID models. Unlike in Panel A, the results suggest that locally implemented IPZs led to a reduction in productivity on the “average” firm, irrespective of the estimation strategy.

The findings thus fail to provide any empirical support for locally designated place-based policies to distort selection penalties in agglomerations. Specifically, spatially targeting “losers” does not bring about significantly higher payoffs in terms of productivity gains via agglomeration economies for these lower-performing firms. Rather, the implications bring about the undesirable effects of reducing city competitiveness via the reduction in agglomeration benefits for the average firms and higher-performing firms, as shown above using Combes et al.’s (2012) approach. These results bring into question the policy rationale and point to possibly other factors besides higher payoffs – such as improper planning and race-to-the-bottom policies – which force some local officials to geographically target clustered firms.

Conclusion

There is increasing attention in the literature on distinguishing between genuine agglomeration economies versus alternative explanations like selection and sorting (Arimoto et al., 2014; Combes et al., 2012), albeit studies on transitioning economy contexts are sparse. Moreover, it was not known until now to what extent state-intervention policies can potentially influence selection and agglomeration forces typically found in large urban agglomerations. This paper contributes to both of these gaps in the literature by testing the semi-parametric quantile framework introduced by Combes et al. (2012) and here applied to the Chinese context. Emphasis is placed on exploring selection and agglomeration forces in Chinese cities, as well as the role of state intervention in influencing these two sources of productivity advantages.

The main results from this paper show that the higher productivity observed in large Chinese urban agglomerations is shaped by both selection and agglomeration forces. Whereas in the French context, Combes et al. (2012) only find evidence of agglomeration economies but not selection. One potential explanation for the contrasting finding is that China’s market reforms led to mass entry of new privately owned firms in the late 1990s, combined with an opening up of reforms that led to a flood of FDI, and in turn, created a hyper-competitive climate responsible for creating the observed selection effects. The results also show that agglomeration forces play a much stronger role than selection in shaping productivity distribution, although previous studies that ignore the role of selection likely overestimate the importance of agglomeration economies in China.

The results further show that state intervention via the IPZs program affects both the selection and agglomeration forces. After correcting for the non-random assignment issues, the main results reveal that the IPZs program increases both the left truncation of the TFP distribution (selection effect) and the rightward shift in the TFP distribution (agglomeration effect). The findings depend critically, however, on who administers the IPZ: if created by local city officials, IPZs are found to reduce both the selection and agglomeration effects. A key question that naturally arises is why do local officials create IPZs despite their negative effects on agglomeration forces?

One economic rationale we test is that local policymakers anticipate that lower-performing firms are able to benefit from a particularly high payoff via agglomeration economies as a result of being able to stay in the market longer. Subsequent analyses reveal, however, that locally implemented IPZs not only fail to significantly motivate lower-performing firms to benefit from agglomeration economies but also cause higher-performing firms to decrease productivity. These results suggest that the reason for local officials to spatially target lower-performing firms is related to other factors besides a higher payoff. Such factors may be the consequence of poor planning practices that led to a non-trivial misallocation of resources and race-to-the-bottom fiscal incentives.

Footnotes

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship and/or publication of this article: The authors The author(s) acknowledge generous support from the School of Economics at Peking University and the Natural Science Foundation of China (No. 71603009 and No. 71603010).

Notes

References

Accetturo

Di Giacinto

Micucci

, et al. (2018) Geography, productivity and trade: Does selection explain why some locations are more productive than others? Journal of Regional Science (Online View) 58(5): 949–979.

Ackerberg

Caves

Frazer

(2015) Identification properties of recent production function estimators. Econometrica 83(6): 2411–2451.

Arimoto

Nakajima

Okazaki

(2014) Sources of productivity improvement in industrial clusters: The case of the prewar Japanese silk-reeling industry. Regional Science and Urban Economics 46: 27–41.

Athey

Imbens

(2006) Identification and inference in nonlinear differencein-differences models. Econometrica 74(2): 431–497.

Henderson

(2006) Are Chinese cities too small? Review of Economic Studies 73: 549–576.

Baldwin

Okubo

(2006) Heterogeneous firms, agglomeration and economic geography: Spatial selection and sorting. Journal of Economic Geography 6: 323–346.

Barca

McCann

Rodriguez-Pose

(2012) The case for regional development intervention: Place-based vs place-neutral approaches. Journal of Regional Science 52(1): 134–152.

Behrens

Duranton

Robert-Nicoud

(2014) Productive cities: Sorting, selection, and agglomeration. Journal of Political Economy 122(3): 507–553.

Brandt

Van Viesebroeck

Zhang

(2012) Creative accounting or creative destruction? Firm-level productivity growth in Chinese manufacturing. Journal of Development Economics 2: 339–351.

10.

Brülhart

Mathys

(2008) Sectoral agglomeration economies in a panel of European regions. Regional Science and Urban Economics 38(4): 348–362.

11.

Chen

Xiang

(2016) Geography versus policy: Exploring how location matters in placed-based policies using a natural experiment in China. Unpublished Manuscript.

12.

Ciccone

Hall

(1996) Productivity and the density of economic activity. American Economic Review 86: 54–70.

13.

Combes

Gobillon

(2015) The empirics of agglomeration economies. In: Duranton

Henderson

Strange

(eds) Handbook of Urban and Regional Economics, 5th ed. Amsterdam: Elsevier.

14.

Combes

Démurger

(2015) Migration externalities in Chinese cities. European Economic Review 76: 152–167.

15.

Combes

Duranton

Gobillon

, et al. (2012) The productivity advantages of large cities: Distinguishing agglomeration from firm selection. Econometrica 80(6): 2543–2594.

16.

Dai

(2017) How urbanization economies impact tfp of r&d performers: Evidence from china. Sustainability 9(10): 1766.

17.

D’Costa

Overman

(2014) The urban wage growth premium: Sorting or learning? Regional Science and Urban Economics 48: 168–179.

18.

Donaldson

Hornbeck

(2016) Railroads and American economic growth: A “market access” approach. The Quarterly Journal of Economics 131(2): 799–858.

19.

Duranton

Puga

(2004) Microfoundations of urban agglomeration economies. In: Henderson

Thisse

(eds) Handbook of Regional and Urban Economics, 4th ed. Amsterdam: Elsevier.

20.

Duranton

Martin

Mayer

, et al. (2011) The Economics of Clusters. Oxford: Oxford University Press.

21.

Eriksson

(2011) Localized spillovers and knowledge flows: How does proximity influence the performance of plants? Economic Geography 87(2): 127–152.

22.

Eriksson

Lindgren

Malmberg

(2008) Agglomeration mobility: Effects of localisation, urbanisation, and scale on job changes. Environment and Planning A 40(10): 2419–2434.

23.

Firpo

Fortin

Lemieux

(2009) Unconditional quantile regressions. Econometrica 77(3): 953–973.

24.

Firth

Shan

(2017) Do agglomeration economies affect the local comovement of stock returns? Evidence from China. Urban Studies 54(5): 1142–1161.

25.

Frick

Rodríguez-Pose

Wong

(2018) Toward economically dynamic special economic zones in emerging countries. Economic Geography 95: 1–35.

26.

Fujita

Ogawa

(1982) Multiple equilibria and structural transition of non-monocentric urban configurations. Regional Science and Urban Economics 12: 161–196.

27.

Glaeser

(1998) Are cities dying? Journal of Economic Perspectives 12(2): 139–160.

28.

Haley

(2013) Subsidies to Chinese Industry: State Capitalism, Business Strategy, and Trade Policy. Oxford: Oxford University Press.

29.

Howell

(2019) Heterogeneous impacts of China’s economic and development zone program. Journal of Regional Science (Forthcoming) https://doi.org/10.1111/jors.12465.

30.

Howell

Rudai

, et al. (2016) Technological relatedness and asymmetrical firm productivity gains under market reforms in China. Cambridge Journal of Regions, Economy and Society 9(3): 1–17.

31.

Howell

Rudai

, et al. (2018) Agglomeration, (Un)-related variety and new firm survival in China: Do local subsidies matter? Papers in Regional Science 97: 485–500.

32.

Iacus

King

Porro

(2012) Causal inference without balance checking: Coarsened exact matching. Political analysis 20(1): 1–24.

33.

King G and Richard N (2019) Why propensity scores should not be used for matching. Political Analysis 27(4): 435–454.

34.

Knoben

Arikan

van Oort

, et al. (2016) Agglomeration and firm performance: One firm’s medicine is another firm’s poison. Environment and Planning A 48(1): 132–153.

35.

Koster

Cheng

Gerritse

, et al. (2018) Place-based policies, firm productivity and displacement effects: Evidence from Shenzhen, China. Journal of Regional Science 59(2): 187–213.

36.

Tao

(2009) Trends and determinants of China’s industrial agglomeration. Journal of Urban Economics 65(2): 167–180.

37.

Wang

Zhu

(2019) Place-based policies, creation and agglomeration economies: Evidence from China’s economic zone program. American Economic Journal: Economic Policy 11(3): 325--360.

38.

Malmberg

Lundequist

(2000) Agglomeration and firm performance: Economies of scale, localisation, and urbanisation among Swedish export firms. Environment and Planning A 32(2): 305–321.

39.

Melitz

Ottaviano

(2008) Market size, trade, and productivity. Review of Economic Studies 75(1): 295–316.

40.

Melo

Graham

, and Noland

(2009) A meta-analysis of estimates of urban agglomeration economies. Regional Science and Urban Economics 39: 332–342.

41.

Melo

Graham

Levinson

, et al. (2017) Agglomeration, accessibility and productivity: Evidence for large metropolitan areas in the US. Urban Studies 54(1): 179–195.

42.

Moreno

Paci

Usai

(2005) Spatial spillovers and innovation activity in European regions. Environment and planning A 37(10): 1793–1812.

43.

Ottaviano

(2011) ‘New’ new economic geography: Firm heterogeneity and agglomeration economies. Journal of Economic Geography 11: 231–240.

44.

Rosenthal

Strange

(2004) Evidence on the Nature and Sources of Agglomeration Economies. In: Henderson

Thisse

(eds) Handbook of Urban and Regional Economics, Volume 4, chapter 49. London: Elsevier.

45.

Seidel

von Ehrlich

(2011) Agglomeration and credit constraints. ETH Zurich Working Paper, pp.1–17.

46.

Wang

(2013) The economic impact of special economic zones: Evidence from Chinese municipalities. Journal of Development Economics 101(1): 133–147.

47.

Wang

Lin

GCS

(2010) Industrial clustering and technological innovation in China: New evidence from the ICT industry in Shenzhen. Environment and Planning A 42(8): 1987–2010.

48.

Zeng

(2011) How Do Special Economic Zones and Industrial Clusters Drive China’s Rapid Development? World Bank Policy Research Working Paper No. 5583.

49.

Zheng

Sun

, et al. (2017) The birth of edge cities in China: Measuring the effects of industrial parks policy. Journal of Urban Economics 100: 80–103.