Telecommunications infrastructure and regional income convergence in China: panel data approaches

Transcription

1 Ann Reg Sci (2008) 42: DOI /s ORIGINAL PAPER Telecommunications infrastructure and regional income convergence in China: panel data approaches Lei Ding Kingsley E. Haynes Yanchun Liu Received: 20 April 2007 / Accepted: 6 November 2007 / Published online: 11 January 2008 Springer-Verlag 2008 Abstract The conditional convergence framework constitutes the theoretical basis for different dynamic panel data approaches. But models with different specifications or estimated by different methods may have very different results. This study empirically tests the results by applying different panel data approaches to the study of telecommunications infrastructure in regional economic growth across China. Specifically, the pattern of regional economic growth across 29 regions in China from 1986 to 2002 is examined. The results suggest the system GMM estimation is more likely to produce consistent and efficient estimates than OLS and fixed-effect estimation. Findings indicate a significant and positive relationship between telecommunications infrastructure and regional economic growth in China and the empirical results from different estimations suggest robust results for this particular assessment. JEL Classification H54 O47 O53 The authors wish to acknowledge the insightful and helpful comments of two anonymous referees that significantly raised the quality of this study. Any errors or misinterpretation are the responsibility of the author. L. Ding Center for Community Capital, University of North Carolina, Chapel Hill, NC 27599, USA Lei_Ding@unc.edu K. E. Haynes (B) Y. Liu School of Public Policy 3C6, George Mason University, 4400 University Dr. 2C9, Fairfax, VA 22030, USA khaynes@gmu.edu; mmcclure@gmu.edu Y. Liu yliua@gmu.edu

2 844 L. Ding et al. 1 Introduction In recent years neoclassical economics has witnessed an expansion of various economic growth theories and models. Two mainstream theoretical and empirical branches of the literature dominate current studies. One is the classical Solow aggregate production function approach (e.g., Solow 1956), and the other resides in the conditional convergence discussion stressing identification of factors that lead to economic growth (Barro and Sala-i-Martin 1991). In general, the convergence hypothesis in the neoclassical growth model manifests itself by reflecting that poor economies tend to grow faster than rich economies in terms of per capital income or product (Solow 1956). It arises from the assumption that the diminishing returns to capital in richer regions directs the flow ofliu capital to poorer regions for higher returns (diminishing returns to capital), a mechanism boosting growth in the poorer regions and achieving convergence. Barro and Sala-i-Martin (1991) finds that the convergence holds among groups of countries with certain characteristics in common and among regions within a country that have certain commonalities. This approach allows testing for conditional convergence by adding to a Solow equation a set of variables reflecting differences in the steady-state equilibrium. Barro and Sala-i-Martin (1991, 1992) work finds that the convergence holds among groups of countries with certain characteristics in common and among regions within a country that have certain commonalities. The conditional convergence approach avoids causality and endogeneity issues by identifying the determinative effects of initial per capita income and a series of other conditional variables (e.g. prior infrastructure investments) upon per capita GDP growth. However, the crucial assumption in the conditional convergence method is the existence of identical aggregate production functions for all the countries or regions being examined. In other words, the unobservable country effects are omitted or considered insignificant under the Barro approach. Islam (1995) challenges that assumption by allowing for differences in production functions across regions. He establishes a dynamic panel data model and compares his result with that under the static crosssection framework used by Mankiw et al. (1992). Specifically, Islam (1995) proposes to include the unobservable country effects in the equation and estimates them over time and space using a panel dataset. In this study, two similar dynamic model based on panel data are employed to identify determinants of economic growth, especially the role of telecommunications infrastructure in regional economic growth, in China. Many early panel data studies rely on different versions of the Least Squares with Dummy Variables (LSDV) estimation method (e.g., Demurger 2001). While this method has some advantage over simple ordinary least square (OLS) estimation because of its ability to control for the omitted variable bias, the LSDV estimator does not explicitly address the serial autocorrelation and heteroscedasticity within individual observations. More importantly, the endogeneity problem cannot be solved completely in a one equation model. Recent developments in estimation methods, such as first difference GMM estimator proposed by Caselli et al. (1996) and the system GMM estimator developed by Blundell and Bond (1998) addressed these problems explicitly. This paper, instead of probing theoretical conflicts among different approaches and estimation methods, attempts to examine the role of one conditional variable

3 Telecommunications infrastructure and regional income 845 telecommunications infrastructure in regional economic growth across 29 provinces in China from 1986 to 2002 with empirical implementation of different approaches. The study contributes to understanding of the controversy around the role of infrastructure in regional economic dynamics. It has been widely recognized that public spending or investment in physical infrastructure can facilitate market transactions and generate positive externalities among businesses, and consequently boost regional economic growth. Studies focusing on examining effects of telecommunications infrastructure on regional economic growth abound in current literature (e.g., Saunders et al. 1994; Yilmaz et al. 2001). However, the fact that most of these studies focus on economically developed regions or countries merits this paper for analysis in the context of a developing economy like China. Another contribution of this paper to the literature lies in the empirical finding of this paper based on different panel data approaches and different estimation methods, which assess robustness that is not part of early studies. This paper is organized as follows: Sect. 2 reviews the literature on the relationship between telecommunications infrastructure and economic growth and on China s income convergence. Section 3 illustrates the dynamic panel data approaches, the estimation methods, and the data. Section 4 discusses the analysis results and the final section concludes the paper. 2 Literature Provision of public infrastructure is probably the most common and possibly, under specific circumstances, the most effective means by which governments promote economic growth. Quantitative assessment of measuring links between infrastructure and regional economic growth primarily follow production function approaches (e.g., Cobb-Douglas) and use applications of aggregated national time series data to investigate the relationship between public infrastructure capital and aggregated output in the private sector (e.g., Aschauer 1989). However, other studies challenge the research findings of infrastructure s positive and significant effects on regional economic growth because the direction of causality and crowding-out effect of public spending were largely ignored in early studies (Holtz-Eakin and Schwartz 1995). The source of funding for public infrastructure must not crowd out private capital borrowing and the public infrastructure created by investment must add new needed capacity to be an efficient contribution to regional economic activity (Haynes et al. 2006). Much of this has been reviewed generally in Haynes (2006) and technically in Fedderke and Bogetic (2006). A quick review of the literature reveals several mechanisms demonstrating how telecommunications infrastructure supports economic growth. First, telecommunications infrastructure can be a direct input in the production process. A number of cross-section studies on infrastructure belong to this group (e.g., Yilmaz et al. 2001). Second, telecommunications infrastructure can increase productivity of other productive inputs. Saunders et al. (1994) reviews earlier empirical studies and indicates high-quality telecommunications services are able to boost growth in other sectors through lower transaction costs, larger integrated markets, better market information and faster information diffusion. Third, telecommunications infrastructure can attract

4 846 L. Ding et al. resources from outside regions, serving as a catalyst for regional economic growth. The catalyst role of telecommunications infrastructure has gained substantiation by both empirical and theoretical studies (Mody 1997; Sun et al. 2002). More important, telecommunications infrastructure capital stands in some contrast with other types of infrastructure capital: telecommunications has experienced radical technical and productivity change; it has attracted large amounts of investment capital from both the public and private sectors; and its rapid diffusion has been propelled by sharply reduced costs and increased capacity (Nadiri and Nandi 2003). As a result, telecommunications infrastructure has the potential to lead leapfrogging development in the developing countries. Bruce (1989) and Singh (1999) indicate the potential role of the telecommunications sector in accelerating the pace of development in developing countries. Many policy makers and theoreticians propose that information technologies, especially telecommunications, can help developing countries accelerate their pace of development or telescope the stages of growth. A vast literature on telecommunications infrastructure investment and economic growth exists and identifies telecommunications service availability as a crucial element in the accumulation of factors boosting economic growth at both the regional and sector specific level (e.g., Hardy 1980; Cronin et al. 1991; Cohen 1992; Yilmaz et al. 2001; Datta and Agarwal 2004). However, as Ding and Haynes (2006) summarize, most the existing literature suffer the endogeneity problem of the telecommunications variable, ignore the rapid development in mobile communications, and lack consideration of the spatial dependence problem. Spatial implications are inherent in the development of telecommunications infrastructure due to the significant impact of telecommunications service availability on interregional economic activities. New network neighborhoods may possibly form with advancements in telecommunications technologies and the relatively diminishing advantages of some forms of geographic proximity. This geographical imbalance should be even more important in analyses conducted upon developing and newly integrating economies like China, which exhibit dramatic location-sensitive differences in terms of regional development. Theoretically, better availability of traditional telecommunications infrastructure and more advanced telecommunication technologies liberate economic activities from geographical restraints, and allow them to decentralize from the core to the periphery while maintaining necessary connections (Abler 1970; Stough and Paelinck 1988). The regional disparity of telecommunications infrastructure endowments has been studied extensively. The imbalanced telecommunications technologies penetration among populations and regions in developed economies has been examined by studies focusing on the topics of universal service and the digital divide (Dinc et al. 1998; Duwadi 2003; Norris 2000). China s economic growth and income convergence has also been studied extensively (e.g., Chen and Fleisher 1996; Jian et al. 1996; Mody and Wang 1997; Raiser 1998; Demurger 2001; Weeks and Yao 2003). Chen and Fleisher (1996)usesbothan extended Solow growth model with cross-sectional and panel data to investigate per capita production across China s provinces from 1978 to Weeks and Yao (2003) employ dynamic panel data approaches to investigate income convergence for the prereform and reform periods. One contribution of their study is that they tried different estimation methods and their results suggest a well-specified system GMM estimators

5 Telecommunications infrastructure and regional income 847 produce the best results. None of these studies considers the impact of infrastructure on growth but a few others do. Mody and Wang (1997) uses the panel data on the output of 23 industrial sectors for seven of China s coastal regions during the second half of the 1980s. They find that both transport, measured by road network length, and telecommunications facilities have been the engines of growth during this short time period from 1985 to Using panel data for a sample of 24 Chinese provinces from 1985 to 1998, Demurger (2001) finds that the development of telecommunications in rural areas helped reduce the burden of isolation and had a positive impact on economic growth. In his estimation, the introduction of the number of telephones per capita as a proxy for telecommunications development, confirmed telecommunications significant impact on growth performance and corroborated Mody and Wang (1997) results on the positive effect of telecommunications on growth. 3 Methods and data 3.1 Methodology Barro and Sala-i-Martin (1991, 1992) seminal work gives rise to the conditional convergence concept. Simply put, the conditional convergence happens when partial correlation between growth in income over time and its initial level is negative. This approach allows testing for conditional convergence by adding to a Solow equation a set of variables reflecting differences in the steady-state equilibrium. According to Barro and Sala-i-Martin (1991) and in the telecommunications context, the static approach in this study would have the following specification for regions in China: n 1 GRTH i = α 0 + β 1 Ln(GDP) i,0 + β j X i,0 + β n TEL i,0 + ε i (1) where i indexes provinces in China; GRTH i represents the annual growth rate of real GDP per capita during the study period for province i; Ln(GDP) i,0 represents the initial level of real GDP per capita in logarithm form for province i; TEL i,0 represents a measure of initial telecommunications infrastructure in province i in China; and X i,0 contains a set of variables accounting for production factors and other conditioning variables at the beginning of the study period for province i. Corresponding to the inherent weakness in satisfying the strict assumption of identical aggregate production functions across regions in the cross section model, Islam (1995) advocates and implements a panel data approach with fixed rather than random effects in order to allow for the differences of the unobservable individualities, or country effects. So, the use of panel data approaches mitigates some of the problems for cross-sectional studies, such as the very small number of observations and unrealistic assumption that all the countries have identical production functions. This analysis employs an extended version of the Barro approach, i.e., to assign the dynamic characteristic to this model by elongating the single-year dataset into a panel dataset. This framework has been used in some recent studies (Chen and Fleisher 1996; j=2

6 848 L. Ding et al. Demurger 2001). It has the following specification: GRTH it = α 0 + β 1 Ln(GDP) i,t 1 + n β j X it + β n TEL it + µ i + η t + ν it (2) j=2 where t indexes year; GRTH it represents the annual growth rate of real GDP per capita for province i in year t.µ i and η t are province- and time-specific parameters, respectively. The former denotes productivity level differences among provinces and takes into account unmeasured characteristics of regions, including economic reforms, natural resources, or geographical location differences. The latter is introduced to control for time-specific effects which include the rate of technological change and temporary shocks or policy changes that might have affected all regions at the same time. This is particularly important during the 1990s to China since economic reforms and growth accelerated enormously. As shown in Islam (1995) and employed in Weeks and Yao (2003), sometimes the model specified in Eq. (2) can be rewritten as: Ln(GDP) i,t = α 0 + β 1 Ln(GDP) i,t 1 n + β j X it + β n TEL it + µ i + η t + ν it (3) j=2 Since there is a lagged dependent variable on the left-hand side of the equation, this model is a dynamic panel data approach. This approach is advantageous in its ability to allow for the differences across countries or regions, as mentioned above. Datta and Agarwal (2004) used a panel data approach similar to Islam (1995) by further introducing a lagged growth variable on the right hand side of the equation: GRTH it = α 0 + γ GRTH i,t 1 + β 1 Ln(GDP) i,t 1 + n β j X it j=2 + β n TEL it + µ i + η t + ν it (4) Equation (2) thus nests the more general type model into Eq. (4). In Eq. (4), a lagged growth rate and a lagged GDP per capita variable are introduced at the same time, which leads in fact to a second order finite difference dynamic model. With unobservable bias controlled or corrected, this model is able to capture the short-run autoregressive behavior by adding the lagged dependent variable as an independent variable. Both Model A and B are dynamic panel data approaches. The dynamic panel data approach draws skepticism itself in that serial autocorrelation and business cycle effects are inevitably introduced when more than one observation on each economy are added (Mankiw 1995). So it is important to discuss different estimation methods of the panel data approaches before we specify the model. There are many methods that can be used to estimate the growth models outlined above, such as the OLS estimation, LSDV estimation, maximum likelihood estimation (MLE), minimum distance estimation (MD), and different GMM estimations. The OLS estimation of the panel data does not consider the unobserved time and

7 Telecommunications infrastructure and regional income 849 regional effects. As a result, the OLS estimation suffers from a positive correlation between the lagged dependent variable and the error term. Because of the positive correlation between the lagged dependent variable and the fixed effects (µ i ), OLS estimates of the convergence rates tend to be biased upwards and may be inconsistent (Roodman 2007). The LSDV estimator is based on the fixed-effects assumption. In contrast to fixedeffect assumption, random effects assume non-correlation between the individual observations and the exogenous variables in the model. This assumption is unsuitable for this application due to the built-in correlation between these individual observations and the exogenous variables in this specification. By controlling for the omitted variable bias and the endogeneity problem, Caselli et al. (1996) asserts that the panel data approach is better than the cross-section regression approach. However, LSDV does not eliminate dynamic panel bias since here the lagged dependent variable is negatively correlated with the error term, biasing its coefficient estimate downward (Roodman 2007). Interestingly, as demonstrated by Sevestre and Trognon (1996), generally the OLS and LSDV estimator can provide a bound for the true value of the coefficient of the lagged dependent variable. Good estimates of the true parameter should therefore lie in the range between these values or at least close to it. For the estimators that eliminate the individual-effect term through first differencing such as the first difference GMM estimator proposed by Caselli et al. (1996), it does not matter whether the group effect is fixed or random. The system GMM estimator developed by Blundell and Bond (1998) makes further improvement over the first difference GMM estimator by addressing the weak instruments problem, the problem of heteroscedasticity and autocorrelation within individuals, and the endogeneity problem explicitly. This general estimator is designed for situations with small T, large N panels, where there are few time periods and many individuals and the independent variables are not required to be strictly exogenous. In addition, the performance of the system GMM can be tested by the identification of an estimation bound for the parameters estimated by the OLS and LSDV estimator as mentioned above. Some empirical studies have confirmed this (Nerlove 1999; Weeks and Yao 2003). A new command was recently introduced into STATA, which allows us use of this method (Roodman 2007). With a small T and large N panel, this study uses all three estimation methods, OLS, LSDV, and system GMM for both panel data models outlined above. 3.2 Data The dataset in this analysis is constructed for the 29 regions of China for the 17-year period from 1986 to The 29 regions out of the 31 provinces, autonomous regions and municipalities of China are included, except Tibet Autonomous Region and the Chongqing Municipality. Tibet is excluded due to missing data, while Chongqing municipality was established in 1997 and hence is incapable of being included for the whole data period. Appropriate adjustments on the Chongqing municipality data are made so that it can be included as part of its original parent province Sichuan.

8 850 L. Ding et al. The initial year of 1986 is selected because a review of the economic development history of China reveals that before 1985 the size of telecommunications infrastructure in China was small and insignificant compared to other physical infrastructure such as transport and electricity. The effects of telecommunications infrastructure on regional economic growth before 1986 are reasonably estimated to be marginal or non-existent. Another issue of the time dimension for the panel dataset in this study is related to the length of sub-periods. Islam (1995) uses 5-year intervals in constructing his dataset in order to eliminate the possible business cycle effects. However, since the overall economic growth in China after mid-1980s shows a continuously upward trend, and by data plotting it is easy to see that neither a dramatic economic downturn nor a smooth downward trend exists in the 17 years of data. Thus, the panel dataset used in this study is based on annual data following Weeks and Yao (2003) and Chen and Fleisher (1996), and the benefit of this treatment is that more observations provide more information for the regression analysis. The conditional variables are selected along the guidelines of the growth literature. The growth variables in early literature range from traditional economic variables such as physical labor and capital inputs, to a broader scope of economic variables such as human capital, public capital, R&D investment and regional inequalities, even social capital, religion, institutions, and political variables (Levine and Renelt 1992; Sala-i -Martin 1997). Based on economic inputs, productivity and infrastructure considerations as well as the availability of data, this research selected fixed investment, foreign direct investment, employment, human capital, population growth, urbanization level, share of industrial output by state-owned enterprises, transportation, as well as telecommunications infrastructure for the initial analysis (Table 1). The aggregate provincial economic data of real GDP per capita, growth rates, population, fixed investment, employment, urbanization, transport density, foreign direct investment, state-owned enterprise outputs, total industrial output and number of telephones for different provinces, autonomous regions and cities from 1986 to 2002 is provided by the Comprehensive Statistical Data and materials from 50 years of New China (NSB 1999), Statistics on Investment in Fixed Assets of China (NSB 2002) and then updated with China Statistics Yearbook from 1999 to Average years of schooling before 1998 for the human capital variable is referenced to Yu (2001) and the data on human capital after 1998 is updated based on data from the China Statistics Yearbook from 1999 to It is noteworthy that this analysis adopts the measurement of the telecommunications infrastructure by using tele-density, i.e., the number of telephone sets per 100 inhabitants including both fixed line and mobile phones (MII 2004). Most previous studies measure the telecommunications infrastructure with number of main lines (Savage et al. 2003; Hardy 1980; Demurger 2001). An examination of the circumstances and facts in China in terms of telecommunications infrastructure development reveals that the rapid deployment of mobile communications technologies has caused the mobile service subscribers to outnumber fixed line subscribers since 2003 (Ding and Haynes 2007). Bias would be inevitable if the old measurement of fixed lines for telecommunications infrastructure were used. Therefore, the number of all (fixed and mobile) telephones per 100 inhabitants is appropriate for this study.

9 Telecommunications infrastructure and regional income 851 Table 1 Definition of variables, mean values, and expected signs Variable Definition Mean Expected sign GROWTH GROWTH t 1 LGDP t 1 INV FDI POPGROW EMPLOY TELE HC URBAN SOE TRANS Annual growth rate of real GDP per capita One-year lag of growth rate of real GDP per capita log value of real GDP per capita in 1995 RMB Share of fixed investment in GDP Share of foreign direct investment divided by total fixed investment Annual population growth rate Share of total employment to total population Number of telephones per capita Human capital measured by the average years of schooling for the population aged 6 and above Share of urban population to total population. Share of industrial output by state-owned enterprises in total industrial output Transportation density as measured by the length of rail, highway, and waterway networks per square kilometer Note: based on authors calculation; mean values are presented for selected years; expected signs are listed for all independent variables After including all the conditional variables selected for this study, Eqs. (2) and (4) can be reformulated as: and GRTH it = α 0 + β 1 Ln(GDP) i,t 1 + β 2 INV it + β 3 FDI it + β 4 POP it + β 5 EMP it + β 6 HC it + β 7 URBAN it + β 8 SOE it + β 9 TRANS it + β 10 TEL it + µ i + η t + ν it (5) GRTH it = α 0 + γ GRTH i,t 1 + β 1 Ln(GDP) i,t 1 + β 2 INV it + β 3 FDI it + β 4 POP it + β 5 EMP it + β 6 HC it + β 7 URBAN it + β 8 SOE it + β 9 TRANS it + β 10 TEL it + µ i + η t + ν it (6)

10 852 L. Ding et al. Both Eqs. (5) and (6) are used in this analysis to test for the role of telecommunications infrastructure in regional economic growth across 29 provinces from 1986 to For simplification we call Eq. (5) Model A and Eq. (6) Model B. Definition of the variables, mean values in representative years (1986, 1995 and 2001), and the expected signs are listed in Table 1. The expected positive lagged growth rate of per capita real GDP follows research findings of Datta and Agarwal (2004). Barro and Sala-i-Martin (1991) explains the expected negative relationship between growth rate of real GDP per capita and the lagged real GDP per capita (i.e., the conditional convergence discussion). The standardization of real GDP by population determines the negative effects of population growth rate on the growth rate of real GDP per capita. The positive signs of employment, fixed assets investment, foreign direct investment and tele-density are consistent with past growth studies (as indicated in the Literature section). In this analysis, the positive contribution of telecommunications infrastructure to regional economic growth is hypothesized in the context of regional economic development in China. Therefore, a positive sign is expected. 4 Results and discussion In the initial analysis, we put all the independent variables in the regressions. Table 3 summarizes the results from the fixed-effect estimation (LSDV) and the system GMM estimation for both models. Because of the obvious flaws in OLS estimation, its results would only be included for the final models to help provide a range for the true convergence speed. Diagnostics tests show that there is significant first-order autocorrelation and heteroskedasticity within regions for this panel dataset. The Hausman test (see Table 2) for the difference between the fixed effects and random effects tests is significant, supporting the adoption of fixed effects approach in this study. In other words, the Hausman test indicates the fixed effect model is preferred over the random effect model. Consequently the LSDV estimation for the dynamic models should be a two-way fixed effect approach, with 28 region dummy variables and 16 time dummy variables added. As Table 3 shows, the lagged growth variable has been positive and significant for Model B. The fixed investment variable (INV), foreign investment (FDI), and population growth (POP) have been consistently significant across different models and different estimation methods. Although the lagged GDP per capita variable (LGDP t 1 ) is negative and significant in most cases, the coefficient is insignificant for the system GMM estimation of Model A and only slightly significant for the system GMM estimation of Model B. The telecommunications infrastructure variable is only significant in the fixed-effect models. All other variables are insignificant in all models. We decided to exclude the four variables (HC, URBAN, SOE, TRANS) from further analysis for two reasons. First, these four variables are not significant in any of the estimations in Table 3. Second, some of these variables may be highly correlated with other independent variables such as the lagged GDP per capita variable and the telecom variable. So we have reduced models and we used three estimations methods

11 Telecommunications infrastructure and regional income 853 Table 2 Results of Hausman test Coefficients (b) fixed (B) (b-b) Difference Standard error GRTH t Ln(GDP) t INV FDI POP EMP HC URBAN SOE TRANS TEL Chi-square = Probability > chi-square = Table 3 Initial regression results Variables Model A Model B LSDV System GMM LSDV System GMM GRTH t (6.84) 0.561(6.42) LGDP t ( 5.56) 0.019( 1.14) 0.099( 6.12) 0.018( 1.99) INV 0.128(5.1) 0.150(5.56) 0.087(3.53) 0.063(3.4) FDI 0.111(3.47) 0.121(2.05) 0.077(2.48) 0.073(2.9) POP 0.791( 4.94) 0.740( 4.54) 0.812( 5.32) 0.832( 6.38) EMP 0.105(1.94) 0.058(0.87) 0.076(1.47) 0.008( 0.25) TEL 0.069(2.66) 0.026(0.77) 0.055(2.21) 0.031(1.06) HC 0.001(0.12) 0.003(0.26) 0.001(0.11) 0.008( 1.27) URBAN 0.027(0.27) 0.001(0.02) 0.109(1.11) 0.069(1.7) SOE 0.001(0.22) 0.015( 1.06) 0.000(0.05) 0.006(0.87) TRANS (0.71) 0.008( 0.25) 0.005(0.14) 0.008( 0.42) Intercept 0.698(5.50) (1.24) 0.689(5.71) 0.136(2.28) R Square Note: Results are based on data of 29 regions for a 17-year period in China. Model A is specified in Eq. (5) and Model B further adds a lagged growth variable in Model A (see Eq. 6). For Model B, diagnostics tests results include: Hausman test, χ 2 (11) = 1,069, p < 0.00; Wooldridge test for autocorrelation F(1, 28) = 63.9, p < 0.001; heteroskedasticity test: χ 2 (28) = 171.7, p < Diagnostics tests results for Model A is very similar to those of Model B and thus not listed here. Asteriks indicate significant levels. One (***); five (**) and ten (*) percent to estimate the final models. Furthermore, as previously literature suggests, income levels, urbanization rate, and educational levels have significant impacts on penetration rate of fixed-line telephony in China (Hu and Zhou 2002). In the reduced system

12 854 L. Ding et al. Table 4 Moran s I tests on Model B for spatial error dependency Year OLS LSDV System GMM Moran s I p-value Moran s I p-value Moran s I p-value Moran s I statistics for Model A are not listed to save some space and the results are quite similar *** Significant at 1% level ** Significant at 5% level * Significant at 10% level GMM models, we addressed the endogeneity problem explicitly by treating the telecom variable as an endogenous variable, which is determined by the lagged GDP per capita variable (LGDP t 1 ), and some exdogenous variables including human capital (HC), urbanization level (URBAN), and transportation density (TRANS). The results of the reduced models are listed in Table 5. Before we interpret the estimation results, we need to assess the spatial dependence issue. The presence of spatial dependence may lead to misspecification in that un-resolved spatial dependence in the error terms of regional econometric models, due to the omitted variables, may be related to the connectivity of neighboring regions (Kelejian and Robinson 1997; Yilmaz et al. 2001). A properly specified regional econometric model sorts to assure reduction or elimination of spatial dependence. To check spatial error dependence, Moran s I tests are carried out on residual values of Model A and Model B for each year. Table 4 shows that Moran s I statistics for Model B. For the LSDV estimation, the Moran s I statistics are generally insignificant and is only slightly significant in one year. For the OLS estimation, the Moran s I statistics for the residue values are significant in four of the 17 years. The results for the system GMM estimation is similar with the Moran s I s significant in three years. Thus, it may be reasonable to conclude that the spatial dependence issue is not a major concern since

13 Telecommunications infrastructure and regional income 855 all the models reduce the spatial dependence to a generally acceptable level. Of course, it seems LSDV estimation does a better job in eliminating the spatial dependence than other estimation methods. Table 5 shows that the results of the reduced models, estimated by OLS, LSDV, and system GMM. As discussed previously, since the system GMM estimator is able to control for individual-specific heteroscedasticity and autocorrelation problems and omitted variable bias, in addition to eliminating endogeneity bias, the discussion of results is primarily based on the system GMM models. The OLS regressions are included to help identify the range for the true value of the coefficient of the lagged dependent variable. Based on the regression results, the credible range for the coefficient of the lagged GDP per capita variable (LGDP t 1 ) should be (from LSDV) to (from OLS) for both Model A and Model B. The estimated coefficient on LGDP t 1 is for Model A and for Model B. Accordingly, the convergence speeds (implied λ) are 1.1% for Model B and 1.6% for Model A, which are significantly slower than the convergence speed estimated based on LSDV (about 10%). If translated into the time needed for the output level to move halfway between its initial- and steady levels (half-life), it would be 43 years for Model A and 63 years for Model B. Though the speed of convergence among regions in China is not the major concern in this study, it is still useful to compare their results with the results from the previous studies. Table 6 compares the convergence speed and the implied lambda from this study with several recent China income convergence studies. A comparative analysis between findings in this study and some prior studies on China indicates that while the results of this study are consistent with some early studies, this study appears to suggest a little longer period for economic convergence among regions in China. The convergence speed is a little slower than some early studies. This may be a result of the selection of time period from mid-1980s and these years witness a deteriorating regional inequality in China. The authors question inclusion of pre-1980s years in analysis due to low quality of data, different economic institutions, and overall change in the openness to global markets. As Table 5 shows, the results of most models suggest a positive impact of telecommunications infrastructure on regional economic growth in China. The only exceptions are the OLS estimation and the system GMM estimation for Model A but the sign of the coefficient is also positive. Considering the fact that we have considered the endogeneity problem in our system GMM estimation, we are relatively confident about the conclusion that telecommunications infrastructure significantly contributes to regional economic growth in China. Model B predicts that the real GDP per capita will increase by 0.4 percent for an increase in teledensity by 10 persons per 100 inhabitants. Generally, the models show significant telecommunications infrastructure variables with positive signs, confirming that telecommunications infrastructure does contribute positively to regional economic growth in China. Results of both models seem to indicate that China s regional growth rates are positively related to fixed investments, employment, foreign direct investment, and negatively related to population growth. As expected, Model B also shows a positive relationship between the growth rates and lagged growth rates.

14 856 L. Ding et al. Table 5 Determinants of regional growth Variables Model A Model B OLS LSDV System GMM OLS LSDV System GMM GRTHt (10.04) 0.291(6.83) 0.565(7.71) LGDPt ( 1.12) 0.096( 5.98) 0.016( 2.08) 0.005( 1.32) 0.096( 6.24) 0.011( 2.92) INV 0.043(2.36) 0.129(5.38) 0.107(3.29) (1.59) 0.082(3.42) 0.054(2.73) FDI 0.130(6.95) 0.109(3.49) 0.135(2.72) 0.079(4.47) 0.081(2.67) 0.068(2.83) POP 0.791( 4.82) 0.799( 5.02) 0.577( 4.21) 0.776( 5.21) 0.815( 5.38) 0.681( 6.01) EMP 0.061(2.17) 0.108(2.06) 0.138(2.42) 0.027(1.03) 0.071(1.41) 0.037(1.34) TEL 0.030(1.42) 0.061(2.61) 0.029(0.98) 0.036(1.85) 0.053(2.39) 0.041(1.65) Intercept 0.065(1.97) 0.713(5.94) 0.082(1.45) 0.036(1.22) 0.703(6.15) 0.045(1.94) Implied λ 0.5% 10.1% 1.6% 0.5% 10.1% 1.1% Half-life (years) R Square Notes: t-statistics in parentheses *** Significant at 1% level ** Significant at 5% level * Significant at 10% level

15 Telecommunications infrastructure and regional income 857 Table 6 Comparison of convergence speed of recent studies Study Chen and Fleisher (1996) Gundlach (1997) Weeks and Yao (2003) Ding and Haynes (2007) Sample 25 provinces 29 provinces 28 provinces 29 provinces Dependent GDP per capita Output per GDP per capita GDP per capita variable worker Study period Implied λ 1.6% 2.2% 2.23% % Half-life (years) Methods Panel data, OLS Cross-sectional System GMM System GMM 5 Conclusion This paper implements different panel data approaches derived from Barro s conditional convergence framework and different estimation methods in testing of the role of telecommunications infrastructure on long run regional economic growth using a sample of 29 provinces of China from 1986 to These approaches estimate the relationship between regional economic growth and initial economic conditions, fixed investment, employment, population growth, foreign direct investment, and telecommunications infrastructure. The results of this paper confirm early studies that the system GMM estimation is more likely to produce consistent and efficient estimates. Since OLS does not control for unobserved characteristics and LSDV estimation generally overestimates the convergence speed, the system GMM is a preferred estimation method. According to findings from the dynamic models, endowments in telecommunications infrastructure across different regions in China during the 17 years do contribute significantly to economic performance and growth. This result generally holds across different models and even after controlling for the endogeneity problem. The confirmation on the contributive role of telecommunications infrastructure on regional economic growth in China is implicative for policy making in the context of developing countries. The findings in this paper generally support the investment induced growth point of view and consequently supports investment policy in the telecommunications and infrastructure sectors in lagging regions of developing countries. This is consistent with the knowledge of the current literature on the relationship between the infrastructure and economic growth. Therefore, the results of this study tend to support the proportion that, ceteris paribus, facilitating telecommunications infrastructure is significant for assisting economic growth in less developed regions of developing countries where such infrastructure is poorly developed. Another major contribution of this paper lies in its provision of additional evidence for validation of the conditional convergence discussion. The tests on the regions in China show that controlling for other factors, regions with lower levels of initial real GDP per capita tend to grow with faster rates than those with higher levels of initial real GDP per capita in China. Particularly in the context of general rapid economic growth in China, it will take about years to eliminate half of the gap between the

16 858 L. Ding et al. lagging and leading regions. Regional economic policymakers in China can benefit from this study in assessing their policy agendas. Appendix Table 7 Time and regional fixed effects Model A-LSDV Model B-LSDV Coefficient p-value Coefficient p-value yr yr yr yr yr yr yr yr yr yr yr yr yr yr yr yr Beijing Fujian Gansu Guangdong Guangxi Guizhou Hainan Hebei Heilongj Henan Hubei Hunan Inner Mo Jiangsu Jiangxi Jilin Liaoning

17 Telecommunications infrastructure and regional income 859 Table 7 continued Model A-LSDV Model B-LSDV Coefficient p-value Coefficient p-value Ningxia Qinghai Shaanxi Shandong Shanghai Shanxi Sichuan Tianjin Xinjiang Yunnan Zhejiang Note: See the results for other variables in Table 5 *** Significant at 1% level ** Significant at 5% level * Significant at 10% level References Abler RF (1970) What makes cities important. Bell Telephone Mag 49:10 15 Aschauer DA (1989) Is public expenditure productive? J Monetary Econ 23: Barro RJ (1991) Economic growth in a cross section of countries. Q J Econ 106(2): Barro RJ, Sala-i-Martin X (1991) Convergence across states and regions. Brooking Pap Econ Act 100: Barro RJ, Sala-i-Martin X (1992) Convergence. J Pol Econ 100(2): Blundell R, Bond S (1998) Initial conditions and moment restrictions in dynamic panel data models. J Econom 87(1): Bruce R (1989) Options and development in the telecommunications sector. In: Wellenius B, Nulty TE, Stern RD (eds) Restructuring and managing the telecommunications sector: a world bank symposium. The World Bank, Washington Caselli F, Esquivel G, Lefort F (1996) Reopening the convergence debate: a new look at cross-country growth empirics. J Econ Growth 1(3): Chen J, Fleisher BM (1996) Regional income inequality and economic growth in China. J Comp Econ 22: Cohen R (1992) The impact of broadband communication on the U.S. economy and on competitiveness. Economic Strategy Institute, Washington, DC Cronin FJ, Parker EB, Colleran EK, Gold MA (1991) Telecommunications infrastructure and economic growth: an analysis of causality. Telecomm Policy 15(6): Datta A, Agarwal S (2004) Telecommunications and economic growth: a panel data approach. Appl Econ 36: Demurger S (2001) Infrastructure development and economic growth: an explanation for regional disparities in China? J Comp Econ 29(1): Dinc M, Haynes KE, Stough RR, Yilmaz S (1998) Regional universal telecommunication service provisions in the U.S.: efficiency versus penetration. Telecomm Policy 22(6): Ding L, Haynes KE (2006) The role of infrastructure in regional economic growth: the case of telecommunications in China. Aust J Reg Stud 12(3): Ding L, Haynes KE (2007) Technology, innovation and latecomer strategies: Evidence from the mobile handset manufacturing sector in China. In: Johansson K, Stough R (eds) Innovations and entrepreneurship in functional regions (forthcoming)

18 860 L. Ding et al. Duwadi K (2003) Telecommunications investment, economic growth, and universal service in a global economy. In: Regional and industry analysis branch papers. Federal Communications Commission, Washington, DC Fedderke JW, Bogetic JW (2006) Infrastructure and growth in South Africa: direct and indirect productivity impacts of 19 infrastructure measures. World Bank Policy Research Working Paper 3989, August Accessed 3 Jan 2007 Hardy AP (1980) The role of the telephone in economic development. Telecomm Policy 4(4): Haynes KE, Yilmaz S, Dinc M (2006) Regional telecommunications investment impacts and efficiency considerations. In: The emerging digital economy. Springer, Berlin, pp Haynes KE (2006) Infrastructure: the glue of mega-cities. Report sponsored by the Hague, Netherlands, Mega-cities Foundation. Accessed 3 Jan 2007 Holtz-Eakin D, Schwartz AE (1995) Infrastructure in a structural model of economic growth. Reg Sci Urban Econ 25(2): Hu A, Zhou S (2002) How can China cope with the enlarging digital divide? (in Chinese) China Ind Econ 3:21 30 Islam N (1995) Growth empirics: a panel data approach. Q J Econ 110: Jian T, Sachs JD, Warner AM (1996) Trends in regional inequality in China. National Bureau of Economic Research Working Paper: 5412 Kelejian HH, Robinson DP (1997) Infrastructure productivity estimation and its underlying econometric specifications: a sensitivity analysis. Pap Reg Sci 76: Levine R, Renelt D (1992) A sensitivity analysis of cross-country growth regressions. Am Econ Rev 82: Mankiw NG, Romer D, Weil DN (1992) A contribution to the empirics of economic growth. Q J Econ 107: Mankiw NG (1995) The growth of nations. Brookings Pap Econ Act 1: , Washington, DC Ministry of Information Industry (MII) (2004) Statistical information (in Chinese). Accessed 3 oct 2006 Mody A (1997) Infrastructure strategies in East Asia: the untold story. World Bank, Washington, DC Mody A, Wang FY (1997) Explaining industrial growth in coastal china: economic reforms...and what else? World Bank Econ Rev 11(2): NSB (National Statistical Bureau) (1999) Comprehensive statistical data and materials on 50 years of New China. China Statistics Press, Beijing NSB (National Statistical Bureau) (2002) Statistics on investment in fixed assets of China ( ). China Statistics Press, Beijing Nadiri MI, Nandi B (2003) Telecommunications infrastructure and economic development. Working Paper Nerlove M (1999) Properties of alternative estimators of dynamic panel models: an empirical analysis of cross-country data for the study of economic growth. In: Hsiao C, Lahiri K, Lee LF, Pesaran MH (eds) Analysis of panels and limited dependent variable models. Cambridge University Press, Cambridge, pp Norris P (2000) The global divide: information poverty and internet access worldwide. International Political Science Association, Quebec City Raiser M (1998) Subsidising inequality: economic reforms, fiscal transfers and convergence across Chinese provinces. J Dev Stud 34(3):1 26 Roodman D (2007) How to do xtabond2: an introduction to difference and system GMM in Stata, Center for Global Development working paper Sala-i -Martin X (1997) I just ran two million regressions. Am Econ Rev 87(2): Saunders R, Warford J, Wellenius B (1994) Telecommunications and economic development. The Johns Hopkins University Press, Baltimore Savage SJ, Schlottman A, Wimmer BS (2003) Telecommunications investment, liberalization and economic growth. Related Publication Sevestre P, Trognon A (1996) Dynamic linear models. In: Matyas L, Sevestre P (eds) The econometrics of panel data: a handbook of the theory with applications, 2nd edn. Advanced Studies in Theoretical and Applied Econometrics, vol 33. Kluwer, Dordrecht, pp Singh JP (1999) Leapfrogging development? The political economy of telecommunications restructuring. State University of New York Press, Albany Solow RM (1956) A contribution to the theory of economic growth. Q J Econ 70(1):65 94 Stough S, Paelinck J (1998) Impact of telecommunication technology on metropolitan spatial form in the European community. Regional Science Vienna Congress