1 Introduction

China’s regional industrial policy has gone through remarkable changes during the past decades. Until the late 1970s, based on heavy-industry-oriented development strategy and the consideration of national security, most manufacturing enterprises are designed to locate dispersedly across regions. This mandatory fragmentation of domestic market leads to remarkable similarities in regions’ economic structure. Since 1980s, policymakers shifted their attention to the promotion of high-tech industries, aiming to upgrade traditional manufacturing and rejuvenate export by making full use science and high technology. In 1988, the State Council of China officially approved “Torch Program” as a national guideline for accelerating the development of high-tech industries. Following central government’s plan, local governments are committed to develop high-tech industries through various approaches such as high-tech-oriented export policies, subsides to high-tech-related activities and establishment of high-tech industrial bases. In addition to these promotion policies, local protectionism also plays an important role in the geographic distribution of industries. Recent studies on China’s manufacturing in the post-reform period have provided evidence of the negative impact of local protectionism on regional concentration of industries (Li et al. 2003; Bai et al. 2004; Lu and Tao 2009; Ge 2009). China’s reform on central planned economy has led to both administrative and fiscal decentralization (Zheng 2007; Golley 2002). This, in turn, gives local governments not only more authorities but also more incentives to pursuit more tax revenues by protecting local manufacturing (Poncet 2003).

This paper focuses exclusively on China’s high-tech industries, which have obtained great concerns and supports from both central and local governments in recent decades, and attempts to address two related empirical issues. First, we calculate the dissimilarity indices to investigate trends and changes in high-tech industrial specialization and regional concentration during the period 1996–2005. We find both specialization and concentration increase significantly at nation level. In particular, we find a widening gap of industrial structure between coastal and inland regions, while the whole high-tech industry sector becomes increasingly concentrated in coastal regions.

The second issue addressed in this paper concerns the role of various economic policies in the distribution of high-tech industries across regions. Using a dynamic panel approach, we are able to evaluate the dynamic impact of economic policies on specialization and concentration while controlling for other driving forces proposed by theories (e.g., knowledge resources and transport costs). In addition, we choose system Generalized Method of Moments method (system GMM), which allows us to solve the endogeneity of lagged variables and to obtain unbiased and consistent estimation results. The estimation result indicates that the implementation of high-tech-oriented export policy and subsidy for science and high technology activities have positive effects, whereas local governments’ protection for local high-tech industries decreases the degrees of specialization and concentration.

The reminder of this paper is organized as follows: in Sect. 2, we briefly review the literature on the determinants of specialization and regional concentration of industries from both theoretical and empirical perspectives. In Sect. 3, we describe the measures and data applied in this paper. Then, we investigate the degrees of industrial specialization and regional concentration of high-tech industries. In Sect. 4, using a dynamic panel approach, we empirically examine the impact of economic policies on specialization by controlling for other driving forces proposed by theories. The conclusion and some policy suggestions are proposed in Sect. 5.

2 A brief literature review: economic policies, specialization and concentration

Substantial theoretical studies have provided explanations for the driving forces that impact the geographical distribution of economic activities. Neo-classical trade theory suggests that industrial specialization depends on comparative advantage that is determined by cross-region disparities in the relative abundance of factor endowments (Ohlin 1935). In a model of two regions (North and South), two tradable goods (food and machinery) and two factors of production (high-skilled labor and low-skilled labor), suppose that North is relatively well endowed with high-skilled labor, and South with low-skilled labor. Production of both goods requires both factors, while the production of machinery is relatively high-skilled intensive. In a closed economy, North can produce machinery more efficiently than South. This results in relatively low prices for machinery in North and food in South. With bilateral trade that incurs no transport costs, prices will be equalized, which results in a higher price for machinery in North and a higher price for food in South. Consequently, North and South will have incentive to specialize in the production of machinery and food, respectively. The factor abundance model thus accounts for regional patterns of specialization as a result of regional disparities in endowments, such as natural resources, technologies and human resources.

Contrary to the limited role for geography in neo-classical trade theory, in New Economic Geography (NEG) models, the location of production emerges endogenously. Krugman (1991) develops a two-region, two-good model involving labor mobility, plant-level scale economies and transport costs. Monopolistic firms tend to locate their production in large markets so as to save on transport costs. Firms’ locational choice in turn impacts local market condition. Goods tend to be cheaper in the region with more firms since consumers in this region will import a narrower range of products and thus avoid more of the trade costs (Baldwin et al. 2003). These two aspects (known as backward and forward linkages) work together to shape the distribution of production activities into a manufacturing core and an agricultural periphery. In the framework of NEG models, the location of manufacturing is highly associated with the level of transport costs. In general, further integration characterized by lower transport costs will lead to higher level of regional concentration of manufacturing.

One key issue discussed frequently in related empirical studies is to study the trends and determinants of specialization and concentration. Kim (1995) investigates the long-run trends in US regional manufacturing structure over the period 1860–1987. He finds that industrial specialization rose significantly from 1860 while fell substantially and continuously since the 1930s. Meanwhile, he finds that trends and changes in regional concentration of manufacturing are consistent with specialization. Manufacturing became more concentrated (dispersed) as regions become more specialized (despecialized) during the given period. However, in the case of European countries, Midelfart-Knarvik et al. (2000) find an increasing trend in industrial specialization but a decreasing trend in regional concentration of EU manufacturing from the early 1980s onwards. Aiginger and Davies (2004) explain this divergence using the entropy index methodology. They find that the increasing specialization has been offset by faster growth in the smaller counties (with respect to economic activity) since 1980s. In turn, industries have become less concentrated across countries.

Notably, the impact of various economic policies on industrial geography has also been investigated broadly in recent studies from both theoretical and empirical perspectives. Forslid and Midelfart (2005) employ a model with vertically linked firms to examine the design of industrial policy in a high-wage economy hosting an industrial cluster of vertically linked industries. They find that government’s decision to tax or subsidize the industries in the cluster highly depends on the level of integration. Li et al. (2003) develop a two-region model to examine how and when interregional trade protection might arise. They find that domestic fiscal decentralization, particularly tax reform, together with high external trade protection, causes interregional protection. As for empirical studies, Holmes (1998) examines the impact of pro-business and anti-business policies on industrial location in United States. He finds a large and abrupt increase in manufacturing activity emerges when one crosses a state border from an anti-business state into a pro-business state. Head et al. (1999) study the location choice of Japanese manufacturing establishments and evaluate the effectiveness of US state promotional policies. States that offered foreign trade zones, job-created subsides and low taxes are found to attract more foreign investment. Devereux et al. (2007) evaluate the impacts of discretionary government grants on location of new plants in United Kingdom. They find the effectiveness of grants increases as agglomeration externalities, measured by the number of other plants in that location in the same industry.

There is also a rich body of empirical literature on industrial agglomeration and specialization in China. Young (2000) investigates the specialization with respect to five sectors in China between 1978 and 1997 and finds an increasing trend of convergence in economic structure. Consequently, the reform and opening up in 1978 results in more severe segmentation of domestic market. Consistent with Young’s findings, using data of inter-regional trade flows, Poncet (2003) shows not only that the extent of regional integration, measured by the inter-provincial trade flow intensity is low, but also that it has decreased between 1987 and 1997.

On the contrary, using data from the input-output table among provinces, Naughton (2003) finds that inter-provincial trade grew more rapidly than either provincial GDP or foreign trade between 1987 and 1992, indicating an increase in national economic integration. Bai et al. (2004) calculate the Hoover coefficient using more disaggregated data of 32 2-digit industries and find specialization in China increased over the period 1985–1997. Consistent with Bai et al. (2004), recent studies have provided more evidence supporting the steadily increasing trends of specialization and concentration in China since 1980s (Wen 2004; Catin et al. 2005).

Despite the controversy on the extent of domestic integration, empirical studies have produced clear consensus on the relationship between local protection and industrial specialization. In general, it is found that industrial agglomeration is lower in industries with greater contributions to local tax revenues and in industries with higher degrees of state ownership, suggesting the important role of local protectionism in obstructing the process of spatial concentration of manufacturing industries (Lu and Tao 2009, Ge 2009).

To our knowledge, excepting local protectionism, the role played by other economic policies in determining the geographical concentration of production activities has been neglected in most literature. By comparison, our study contributes to existing empirical literature in this aspect through a complete investigation into the possible impact and significance of economic policies on industrial specialization and regional concentration of China’s high-tech industries during the period 1996–2005.

3 Measuring specialization and concentration of high-tech industries

3.1 Definitions, measures and data

Industrial specialization focuses on the production structure of a specific region, and refers to the extent to which a given region concentrates its activities in a small number of industries. A region is said to be highly specialized if a few high-tech industries account for a large share of its total production. Regional concentration, on the other hand, is the extent to which production activity in a given high-tech industry is concentrated in only a small number of regions. An industry is said to be highly concentrated if a few regions account for a large share of its total output. These two terms are highly related and reflects the distribution of industries from different perspectives. Nevertheless, whether specialization and concentration changes in parallel or not varies across periods and countries.Footnote 1 For example, EU studies investigating the location of industries find that the two terms go in different directions in 1990s (Brülhart and Torstensson 1996; Amiti 1999; Midelfart-Knarvik et al. 2000; Aiginger and Davies 2004). By contrast, this divergence is not found in the case of China, where both specialization and concentration of manufacturing as a whole shows increasing trends since 1980s (Bai et al. 2004; Lu and Tao 2009).

A number of measures have been constructed to measure the geographic distribution of production activities. We primarily use a dissimilarity measures based on the Krugman specialization index (Krugman 1991) to investigate the degrees of specialization and concentration. For a country with \(J\) geographic units and \(I\) high-tech industries \((J=30, I=5), q_{ij}\) is output of high-tech industry \(i\) for region \(j (i = 1, {\ldots }, I; j = 1, {\ldots }, J)\). \(Q_{j}=\sum _i {q_{ij}}\) is total output of all high-tech industries in region \(j\). \(Q_{i} = \sum _j {q_{ij} }\) is total output of all regions in industry \(i\). \(Q=\sum _i {Q_j } =\sum _j {Q_j } \) is the national total output in all high-tech industries.

The dissimilarity index of specialization DIS\(_{j,k}\) compares the differences in industrial structure of region \(j\) and region \(k\) and is defined as:

$$\begin{aligned} {\text{ DIS}}_{j,k} =\frac{1}{I}\sum _i {\left| {\frac{q_{ij} }{Q_j}-\frac{q_{ik} }{Q_k }} \right|} \end{aligned}$$
(1)

Summing up the dissimilarity indices, we get the specialization index of a specific region \(j\) as:

$$\begin{aligned} {\text{ SPEC}}_j =\frac{1}{J}\sum _j {{\text{ DIS}}_{j,k} } \end{aligned}$$
(2)

The SPEC\(_{j}\) index reflects the average level of industrial specialization of region \(j\). It takes value zero if region \(j\) has an industrial structure identical to the other regions in a country, and takes a maximum value of \(2/I\) if region \(j\) has no industries identical to the rest of the country.

The degree of regional concentration is measured similarly by the dissimilarity index of concentration DIC\(_{i}\) as follows:

$$\begin{aligned} {\text{ DIC}}_i =\frac{1}{J}\sum _j {\left| {\frac{q_{ij} }{Q_i }-\frac{Q_j }{Q}} \right|} \end{aligned}$$
(3)

The DIC\(_{i}\) index measures the degree of regional concentration by summing up the absolute differences between regional and national shares in total output in high-tech industry \(i\). The index equals zero if high-tech industry \(i\) is distributed evenly across regions and increases if industry \(i\) become increasingly concentrated in a few regions.

The calculation of dissimilarity indices requires data on output across a set of regions and industries. For this paper, we choose 5 high-tech industries defined by the Catalog for High-technology Industrial Statistics Classification (2002) as follows:Footnote 2 Aircraft and Spacecraft (A&S); Electronic and Telecommunication Equipment (E&T); Computers and Office Equipment (C&O); Pharmaceutical Manufacturing (P&M); Medical Equipment and Meters (M&M). According to current administrative division of China, we choose 30 provinces as our geographical units and divide them into three coastal regions and four inland regions: Northern Coast, Middle Coast, Southern Coast, Northern Inland, Middle Inland, Southern Inland and Far West Inland.Footnote 3 The data of high-tech industries are drawn from the National Bureau of Statistics (NBS), China Statistics Yearbook on High Technology Industry (2002, 2003 and 2008), which provide data on 5 high-tech industries for 30 provinces over the period of 1996–2005.

3.2 Industrial specialization and regional concentration of high-tech industries

In this section, we start to investigate the geographic distribution of high-tech industries and primarily focus on three aspects: the similarity of high-tech industrial structure between regions; the degrees of industrial specialization and regional concentration of high-tech industries; and the relation between trends and changes in industrial specialization and regional concentration.

Tables 1 and 2 reports the dissimilarity indices DIS\(_{j,k}\) for seven regions, respectively. The national average DIS\(_{j,k}\) goes up significantly from 0.101 in 1996 to 0.132 in 2005, indicating that regions become more specialized compared with other regions during the given period. The last column reports the specialization index SPEC\(_{j}\) of each region, where Southern Coast is the most specialized region and Southern Inland is the least specialized region over years.

Table 1 Dissimilarity indices of specialization in 1996
Full size table
Table 2 Dissimilarity indices of specialization in 2005
Full size table

In addition, the dissimilarity indices show that three costal regions have more similar industrial structures. Specifically, the DIS\(_{j,k}\) index between Middle Coast and Southern Coast goes down from 0.087 in 1996 to 0.036 in 2005, indicating a convergence of high-tech industrial structure over years. On the other hand, the Inland average DIS\(_{j,k}\) index increases from 0.067 in 1996 to 0.082 in 2005,Footnote 4 which is considerably lower than the national average during the same period, suggesting that inland regions are also more similar to each other in their industrial structures. On the contrary, the DIS\(_{j,k}\) indices between three costal regions and four inland regions increase significantly over years. Therefore, the increase in specialization indices at nation level is mainly caused by the widening differences between the high-tech industrial structures of coastal and inland regions.

The increasing differences in regional industrial structure could also be reflected by the changes in regional concentration of industries. Table 3 reports the concentration indices DIC\(_{i}\) for each high-tech industry form 1996–2005. The degree of concentration at individual industry varies significantly over the period. A&S manufacturing is the most concentrated high-tech industry with the largest concentration indices far greater than other industries, while E&E manufacturing is the most dispersed high-tech industry with consistently lowest indices. The concentration indices of P&M manufacturing, which was not very concentrated in 1996, have increased significantly over years. On the contrary, initially highly concentrated industries, such as C&O and M&M, have become increasingly dispersed during the given period.

Table 3 Regional concentration of high-tech industry
Full size table

The last row of Table 3 reports the concentration indices of high-tech industry sector as a whole, calculated as the average of indices of each high-tech industry for each year. The concentration index for high-tech industry sector increases from 0.627 in 1996 to 0.707 in 2005, indicating an increase in the level of regional concentration of high-tech industry sector in China. Table 4 provides more information on regions’ shares in high-tech industry sector. In 1996, 75.5 % of China’s high-tech industry sector was located in coastal regions; this share continued to increase and reaches 89.6 % in 2005. Therefore, despite the variations in individual high-tech industry, at the nation level, high-tech industry sector as a whole, become increasingly concentrated in coastal regions as regions become increasingly specialized.

Table 4 Regions’ shares in high-tech industries in 1996 and 2005
Full size table

In summary, in this section, we calculate the dissimilarity indices to measure the degrees of specialization and concentration of China’s high-tech industries during the period 1996–2005. We find the degree of industrial specialization increases significantly while the differences in industrial structures between coastal and inland regions become larger. Furthermore, the whole high-tech industry sector becomes more concentrated in coastal regions. The similar increasing trend of specialization and concentration is consistent with previous studies, investigating the distribution of production activities in China (Bai et al. 2004; Lu and Tao 2009) but inconsistent with those studies in the case of European manufacturing (Brülhart and Torstensson 1996; Midelfart-Knarvik et al. 2000; Aiginger and Davies 2004). The direction of changes in specialization and concentration lies with the changes in the size of industries and regions in different periods. Specifically, EU studies show that smaller member states increase their shares in total manufacturing over decades. That is, smaller member states tend to grow more rapidly than the larger states (Aiginger and Davies 2004). The opposite case happens in China. As shown in Table 4, small regions has experienced a sharp drop in their shares in high-tech industry sector, while large regions increase their shares in high-tech industry sector from 52.0 % in 1996 to 72.3 % in 2005. Moreover, the shares of large regions in large and small industries also increase significantly over years. In this sense, the increase in specialization and concentration indeed reflects the same fact that high-tech industries become increasingly concentrated in some large coastal regions in China. Therefore, we primarily focus on industrial specialization of regions in the following empirical analysis.

4 Empirical analysis

4.1 Econometric specification and method

Our empirical study uses a panel data set of 5 high-tech industries and 30 provinces during the period 1996–2005. Moreover, the utilization of the dynamic panel approach allows us to control for not only the time-invariant individual heterogeneity but also the lagged effects of variables.

We start by estimating a simple static panel model in the following specification:

$$\begin{aligned} {\text{ SPEC}}_{jt} =\alpha +\beta X_{jt} +f_j +\varepsilon _{jt} \end{aligned}$$
(4)

where SPEC\(_{jt}\) is the specialization indices calculated in Sect. 3, which measures the degree of specialization of region \(j\) in year \(t\). \(X_{jt}\) is the vector of independent variables. \(\alpha \) and \(\beta \) denote the constant and coefficient vectors, respectively. \( f_{j}\) stands for the time-invariant fixed effect (e.g., local reserves of mineral resources, regional cultural backgrounds and so on). The error term \(\varepsilon _{jt}\) is assumed uncorrelated with the vector \(X_{jt}\) so that all independent variables are strictly exogenous.

The static specification of Eq. (4) indicates that high-tech industrial structure adjusts instantaneously after the implementation of certain economic policy. However, in reality, adjustment often progresses slowly and highly depends on its previous pattern (Bai et al. 2004). Therefore, a dynamic panel data model is constructed to explore the potential lagged effects of both dependent and independent variables as follows:

$$\begin{aligned} {\text{ SPEC}}_{jt} =\alpha +\delta {\text{ SPEC}}_{j,t-1} +\gamma _0 X1_{jt} +\gamma _1 X1_{j,t-1} +\theta X2_{jt} +f_j +\varepsilon _{jt} \end{aligned}$$
(5)

where SPEC\(_{j,t-1}\) is the lagged dependent variable, which measures the lagged impact of specialization in previous year \(t-1\). Other independent variables are divided into two groups. In the first group, \(X1_{jt}\) and \(X1_{j,t-1}\) are the vectors of contemporaneous and lagged economic policy variables, respectively. In the second group, \(X2_{jt}\) is the vector of other controlled variables. \(\delta , \gamma _{0}, \gamma _{1}\) and \(\theta \) are coefficient vectors. We furthermore assume that the error term \(\varepsilon _{jt}\) is not autocorrelated.

Due to the presence of individual effects, SPEC\(_{j,t-1}\) is positively correlated with the error term. As a result, the inclusion of lagged dependent variable SPEC\(_{j,t-1}\) as regressor could induce the biased estimation results of ordinary least squares (OLS) or other common regression methods for panel data set (Bond 2002). We thus adopt first-difference Generalized Method of Moments (first GMM) developed by Arellano and Bond (1991) to solve this problem. First, the individual effect \(f_{j}\) is eliminated by the first-differencing transformation of Eq. (5). Second, the lagged SPEC\(_{j,t-l} (l=2,3,{\ldots })\) are used as valid instruments for the difference term \(\Delta {\text{ SPEC}}_{j,t-1}\).

Although the first-difference GMM estimator is consistent according to our initial assumption, it might perform poorly as we attempt to explore the times series properties of individual series. The instruments available for the equation in first-differences tend to be weak when the individual series are highly persistent (Bond 2002). In such cases, the system GMM estimator developed by Arellano and Bover (1995) and Blundell and Bond (2000) provides better estimation results with smaller bias and greater precision. Specifically, the lagged first-difference \(\Delta {\text{ SPEC}}_{j,t-1}\) are treated as valid instruments for \({\text{ SPEC}}_{j,t-1}\) in the level Eq. (5). Moreover, other independent variables \(X1_{jt}\) and \(X2_{jt}\), which are assumed to be exogenous, serve as their own instruments, indicating the complete time series \((X1_{j,1}, X1_{j,2},{\ldots },X1_{j,t})\) and \((X2_{j,1}, X2_{j,2},{\ldots },X2_{j,t})\) are valid instrumental variables.

4.2 Specification of independent variables

4.2.1 Control for economic policy

High-tech-oriented export policy. The extent and time periods of opening up vary significantly across regions. Eastern coastal regions where reform and opening up were initially performed have gradually become the center of high-tech industrial production since 1980s. After further acceleration of economic opening to inland in 1990s, middle and western regions also implement high-tech export-oriented policy based on the successful experiences of eastern region. Consequently, those export-oriented high-tech industries, such as C&O and E&T manufacturing, grow rapidly and play an important role in transforming regional structure. The importance of high-tech-oriented export policy is measured by the variable openness to export, which is defined as the share of regional export in regional high-tech industrial output.

Subsidy for science and technology (S&T) activities.

Local governments tend to promote high-tech industries and high-tech enterprises through various subsidies. Nevertheless, more subsidies form governments do not indicate higher level of industrial specialization. Indeed, the impact of subsidy on regional industrial structure is highly related to the allocation of subsidies across high-tech industries. We construct a variable, subsidy for S&T, which is defined as the index of subsidy allocation \({\text{ ISA}}_{j}\) as a proxy to measure the impact of local government’s subsidy on regional high-tech industrial structure:

$$\begin{aligned} {\text{ ISA}}_j =\sum _{i=1}^I {\left| {\frac{{\text{ subsidy}}_{ ij} }{{\text{ subsidy}}_j }-\frac{1}{I}} \right|} \end{aligned}$$
(6)

For a given region \(j, {\text{ subsidy}}_{ij}\) is high-tech industry \(i\)’s share in region \(j\)’s total subsidies, \({\text{ subsidy}}_{j}\). In an extreme case, if government equally allocates subsidies across industries, each high-tech industry gets \(1/I\) of total subsidies. Thus, the ISA\(_{j}\) index measures the differences between actual allocation ratio and equal allocation ratio for each high-tech industry. Significant differences indicate high level of concentration of subsidies in few high-tech industries.

State-level high-tech industrial base (SHIB) policy.

In 1980s, the emerging high-tech industrial cluster represented by Silicon Valley has received a great deal of attention throughout the world. Over the same time period, the State Council of China officially approved the first SHIB in Jiangsu province. By 2005, 113 high-tech industrial bases have been established across 17 provinces in China. These industrial bases facilitate regional concentration of high-tech industries in two ways. First, some knowledge such as tacit knowledge cannot be formalized or written down. As a result, knowledge spillovers are expected to be more localized within the geographic scope of industrial base. Second, although State Council approves the establishment of high-tech industrial base, local governments are responsible for their overall administration and guidance. Consequently, various preferential policies are often implemented to attract high-tech enterprises. Therefore, high-tech enterprises tend to concentrate in SHIBs to benefit from knowledge spillovers and various local preferential policies. Regions with better developed SHIBs will be expected to attract more high-tech enterprises to concentrate in their SHIBs. We thus construct a time-vary dummy variable to measure the existence of SHIBs in one province. The variable takes a value of 1 if one province has at least one SHIB in any given year.

Local protectionism.

Local protectionism is in variety of forms. For example, according to current tax law in China, qualified high-tech enterprises can enjoy 15 % reduction in income tax rate. In most cases, as the identification of high-tech enterprise is under the administration of local government, local high-tech enterprises tend to pass the identification more easily with the protection of local governments. Unlike the trade barriers among countries in the context of international trade, it is very difficult to measure local governments’ protection for local high-tech enterprises within a country directly. We thus turn to consider the outcome of protectionism. Specifically, local high-tech enterprises could gain surplus profits due to various forms of local protectionism that significantly improves their products’ competitiveness in local markets. We employ a variable local profit ratio, which is defined as the percentage of profit in total sales to measure the impact of local protectionism on industrial specialization. Since higher profit ratio indicates higher local protectionism, we would expect the variable has a negative effect on specialization.

4.2.2 Controls for others

Although we mainly focus on the impact of economic policy on industrial specialization, it is still necessary to control for other related determinants of industrial specialization proposed by theories.

Knowledge resources.

Most existing literatures have confirmed the impact of fixed regional resource endowments on industrial specialization. For example, due to the high dependence of some raw materials, most extractive industries are highly concentrated in regions with abundant coals or oils. In contrast to those general manufacturing, high-tech industry is characterized as knowledge intensive, indicating that regions have more knowledge resources would be more specialized in high-tech industries. As it is difficult to measure the invisible knowledge flows empirically, we turn to focus on the carriers and transmitters of knowledge. The most effective way to transfer knowledge is by face-to-face communication of high-skilled labors. Therefore, we construct a variable high-skilled labor intensity, which is defined as the share of scientists and engineers in regional total employee weighted by national average, as a proxy for the impact of regional comparative advantages of knowledge resources.

Local transportation conditions.

Since 1980s, both central and local governments have launched large-scale constructions of local infrastructure. The construction of railways and highways might have significant impact on the geographic distribution of high-tech industries since it greatly reduce the transport costs. High-tech enterprises prefer to locate in regions with better provision of transport network so as to benefit from lower local transaction costs and more convenient connection with other regions. We employ a variable transportation per capita to measure the impact of local transportation conditions on specialization.

Table 5 Definitions and statistics of variables
Full size table

Table 5 summarizes definitions and statistics of variables. The data of SHIBs are from Torch High Technology Industry Development Center, Ministry of Science and Technology of China. The data of regional population and length of railways and highways are from (NBS) China Statistical Yearbook (1997–2006); data of other variables are from (NBS) China Statistics Yearbook on High Technology Industry (2002, 2003 and 2008).

4.3 Estimation results

Table 6 reports the estimation results of the static model of industrial specialization with three estimation methods: pooled OLS, fixed effects and random effects. Due to the remarkable differences among estimation methods, we perform two tests to determine the proper specification for the static model. The F test accepts the significance of individual-specific effects at 1 % level, confirming the validity of the specification of fixed effects model. Moreover, the corresponding p value of Hausman statistic is smaller than 1 %, rejecting the null hypothesis that the individual-specific effects are uncorrelated with the independent variables. Taken together, the specification of the fixed effects model is more preferred than the pooled OLS and random effects model.

Table 6 Estimation results of static model of industrial specialization
Full size table

Focusing on the estimation results of fixed effects model, openness to export and the SHIB dummy variable have positive and significant coefficients at 1 and 5 % level, respectively, which suggests high-tech-oriented export policy and the establishment of SHIB have positive impacts on industrial specialization. While the coefficient of subsidy for S&T variable is positive but insignificant, providing only weak evidence supporting the positive effect of this policy. Furthermore, the coefficient of local profit ratio is negative but insignificant. Therefore, we cannot argue confidently that local governments’ protection for local high-tech enterprises has a negative impact on industrial specialization. As for other independent variables, the estimation result indicates that both knowledge resources and transportation condition have positive and significant impacts on specialization.

To test the robustness of the specification of the fixed effects model, we loosen the initial assumption of the error term \(\varepsilon _{jt}\) and allow it follow the first-order autocorrelation as follows:

$$\begin{aligned} \varepsilon _{jt} =\varphi \varepsilon _{j,t-1} +\mu _{jt} \end{aligned}$$
(7)

The last column of Table 6 reports the estimation results of the fixed effects model with AR (1) disturbance. It is worth noting that the estimation result varies significantly as we control for the potential first-order autocorrelation of error terms. Compared with the fixed effects model, the coefficients of all economic policy variables become smaller. More importantly, the inclusion of the first-order autocorrelated error term in the static model allows us to partially account for the lagged responses of industrial specialization. The positive and significant coefficient (0.794) of AR(1) confirms the existence of first-order autocorrelation of the error term, suggesting that the specification of the static model is incorrectly specified. Therefore, we turn to investigate the dynamic specifications that yield more consistent estimation results.

We estimate the dynamic model of industrial specialization with first-difference and system GMM methods. Taking into account that too many instrumental variables might weaken the Sargan test by overfitting the endogenous variable, we include one lag of dependent variable and independent variables of economic policy in the specification of dynamic model. Table 7 reports the estimation results of first-difference GMM and system GMM methods in the first and second columns, respectively. For each estimation method, the Sargan test does not reject the null hypothesis that the overidentifying restrictions are valid; while the Arellano–Bond autocorrelation test indicates that there is no evidence of second-order serial correlation.

Table 7 Estimation results of dynamic model of industrial specialization
Full size table

To assess the bias of first-difference GMM and system GMM estimators, we first investigate the stationary properties of each time series by estimating a simple AR (1) model for all independent variables (Estimation results and specification in details are presented in Appendix). According to the estimation results of the AR (1) model, all variables are found to be highly persistent. Second, we estimate the dynamic model with pooled OLS and least squares dummy variables (LSDV) method, respectively (see column three and four in Table 6). As Roodman (2009) suggested, given that the lagged explainable variable SPEC\(_{j,t-1}\) is positively correlated with the error term, the coefficient of \(\mathrm{SPEC}_{j,t-1}\) is biased upwards in the OLS estimation but is biased downwards in the LSDV estimation. Therefore, a proper estimate of the true parameter should lie within the range of 0.704–0.929. By comparison, the coefficient of SPEC\(_{j,t-1}\) in the first-difference GMM estimations (0.546) is far below the lower limit (0.704). This result is in line with Blundell and Bond (2000), confirming that the instruments available for the equation in the first-differences are likely to be weak when the individual series exhibit strong persistence. By contrast, the coefficient of SPEC\(_{j,t-1}\) in the system GMM estimation lies suitably within the range defined by OLS and LSDV estimation. We thus prefer the estimation results obtained from the system GMM methods.

According to the estimation results of the system GMM method, the coefficient of lagged SPEC\(_{j,t-1}\) is positive and statistically significant at the 1 % level. As we expected, the positive impact of previous industrial specialization indicates that the adjustment of high-tech industrial structure is a slow process and highly depended on its historical pattern. This result is consistent with Bai et al. (2004) in the case of specialization in China’s manufacturing.

As for economic policy variables, openness to export has a positive and significant coefficient, which confirms the impact of economic opening and high-tech export-oriented policy on industrial specialization and concentration. Our result is also consistent with the findings in Ge (2009), further confirming that export-oriented high-tech industries have higher degree of concentration in China. Regions with more openness to foreign market would attract more FDIs and MNCs to invest in those export-oriented high-tech industries, and thus tend to have higher shares in total high-tech industries.

The contemporaneous coefficient of local government’s subsidy for S&T activities is positive and significant, indicating that local governments prefer to allocate subsidies disproportionately across high-tech industries. Consequently, the share of the promoted high-tech industry in total industrial output increases rapidly, thereby increasing of the degree of industrial specialization and concentration.

The estimation result reveals that the establishment of SHIBs has positive but insignificant coefficient, which gives us confidence to question whether these SHIBs could successfully attract high-tech enterprises or not. Our finding is in line with the study of Zhao et al. (2008), who argue that although local governments consider SHIBs as one of the most important instrument for promoting local high-tech industries, the actual development of SHIBs is still in its infancy stage, with the scale of most SHIBs being quite small. Moreover, the efficiency of SHIBs could also be weakened by other developed zones, which were established prior to SHIBs yet provide roughly similar preferential policies, or by those high-tech enterprises located outside SHIBs but still be eligible to enjoy the preferential policies for SHIBs.

Local profit ratio variable has a negative and highly significant coefficient, suggesting that local protection for high-tech industries has a negative effect on industrial specialization. This result provides strong evidence supporting our hypothesis that local high-tech enterprises tend to obtain higher profits under the protection of local governments. As a result, protection deteriorates segmentation of domestic high-tech product market, eventually impeding industrial specialization and concentration of high-tech industries.

In addition to the contemporaneous impact of economic policy, Table 8 summarizes the short-run effect, which is measured by the coefficient of \(\gamma _{0}\) and the long-run effect, which is measured by \((\gamma _{0}+\gamma _{1})/ (1-\delta )\) for each economic policy variable (The dummy SHIB variable is not reported here due to its insignificance). According to these calculation results, the implementation of economic opening and high-tech-oriented export policy, subsidy for S&T activities have positive long-term effects. On the contrary, local protectionism has a negative long-term impact on specialization.

Table 8 Short-run and long-run effect of economic policy
Full size table

As for other controlled variables, we find that high-skilled labor intensity variable has positive and significant coefficient at 1 % level, which confirms our previous hypothesis that regions with higher high-skilled labor intensity tend to have comparative advantage of knowledge resources. Consequently, such comparative advantage might facilitate regional concentration of high-tech industry. We do not find a significant impact of local transportation on high-tech industrial specialization. The coefficient of transportation per capita is positive but insignificant. Considering the time period of our empirical study, regional disparities in transportation network have been greatly diminished due to the large-scale construction of infrastructure in inland regions over years. Moreover, the less dependency of high-tech industry on natural resources endowments also indicates that the variation of transport costs might have little impact on regional concentration of high-tech industries.

Taken together, these estimation results not only confirm the role of economic policies in determining the geographic distribution of high-tech production activities, but also provide an explanation for the increasing trends in regional concentration of high-tech industries, as described in Sect. 3. Although regions become more specialized over years, the significantly negative effect of local protectionism seriously obstructs the diffusion of high-tech industrial production activities from coastal provinces to inland provinces. On the other hand, most industrial transfer occurs within coastal regions, leading to the convergence of industrial structure between middle and southern coastal regions over years. Consequently, inland regions could not benefit from the rapid growth of high-tech industrial sector in coastal regions, while the disparities between coastal and inland regions have become increasingly larger.

5 Conclusions

This paper empirically examines the dynamic impact of economic policies on industrial specialization and regional concentration of China’s high-tech industries during the period 1996–2005. It is found that the degrees of industrial specialization and regional concentration, measured by the dissimilarity indices, increase steadily during the period of 1996–2005, while high-tech industry sector becomes increasingly concentrated in costal regions. It is also found that the implementation of high-technology-oriented export policy and subsidy for science and high-technology activities encourage specialization and concentration, whereas local governments’ protection for local high-tech enterprises results in convergence in regional industrial structure and obstructs regional concentration of high-tech industries. The estimation result is robust not only to the use of various estimation techniques, but also to the control for other factors proposed by theories such as knowledge resources and transport costs.

In addition, our study provides some policy implications for both central and local governments of China. First, central government should further accelerate economic openness in inland regions with various macro-level preferential policies so as to facilitate the transfer of some export-oriented high-tech industries from eastern coastal region to inland regions. Second, the significant negative impact of local protectionism suggests that further reducing entrance barriers and encouraging domestic market unification would play a more important role in stimulating industrial specialization. However, the lagged effects of previous specialization pattern indicate that inland governments require more effort and time to catch up with their counterparts in eastern coastal regions.

Our study is a preliminary exploration on the impact of economic policies on industrial specialization and regional concentration of high-tech industries. The specification of dynamic panel model confirms our initial assumption that both industrial relocation and the impact of economic policies are dynamic processes. Nevertheless, unlike the high-tech industries that obtain direct promotion form both central and local governments in recent years, the impact of economic policies on manufacturing in general might vary significantly across regions. Therefore, it would be worthwhile to extend our study on the dynamics of industrial specialization and concentration of manufacturing with more disaggregated industrial data in future research.