First of all, I ran the system GMM(Generalized Method of Moments) regression with all independent variables and found GDP per capital and budgetary revenue per capital are not significant different from zero according to t test statistics. Then I excluded these two variables and ran the system GMM again. This result showed all the rest variables are significant and all time dummies made effects to explained variable.
Secondly, I ran the difference GMM using all the rest variables. One instrument variable (Dummy1995) was dropped because of collinearity. And Arellano-Bond test for AR(2) in first differences (z = 0.45 < P value 0.653) shows that I cannot reject the null hypothesis. This means there are serial correlations among random error term in difference model.
Thirdly, in order to test whether instrument variables overfit endogenous variables, I use only instruments variables to do regress (see Appendix 2). The Hausman test statistics is 13.40 (P=0.0199), based on the null hypothesis cannot be rejected. This means the systematic coefficients difference between regression with and without instrument variables. In other words, instrument variables introducing to this model help to get unbiased estimator by controlling the endogenous problems.
Consequently, I choose system GMM estimator with instrument variables as best result in this model and report the system GMM result, difference GMM result and IVRE result in table 1.Estimations are done using the STATA statistical software package.
Table 1. Estimation Results for the Determinants of Infrastructure Spending
using a GMM Model with dynamic panel data
| (1) System GMM | (2) System GMM | (3) Dif.GMM | (4) IVRE | |
| Base infrastructure
spending per capital (loginfrapc) |
-.189358**
(.028503) |
-.1818695**
(.0273606) |
-.2934077**
(.0434367) |
|
| FDI per capital
(logfdipc) |
.0204186**
(.0096699) |
.0228581**
(.0095565) |
.000578
(.0224727) |
-.0253611**
(.0092131) |
| Administrative expenditure per person
(logadmpp) |
.0687069*
(.0415533) |
.0632908**
(.0293476) |
.0886512
(.0748527) |
-.1062419**
(.0348083) |
| Urban and Rural Saving per capital
(logsavepc) |
.0713513**
(.0313059) |
.0914837**
(.0299464) |
.1086967**
(.0307424) |
.0606777**
(.0203213) |
| GDP per capital
(loggdppc) |
.0051987
(.0512733) |
|||
| Budgetary revenue per capital
(logrevpc) |
.0070749
(.0299651) |
|||
| Dummy 1995 | .3698738
(.3377216) |
.4851577**
(.2230368) |
||
| Dummy 1999 | -.1133635**
(.0299295) |
-.1180968**
(.0234622) |
-.0694978*
(.0389091) |
.1372388**
(.0336464) |
| Dummy 2002 | .3213466**
(.0294483) |
.3152924**
(.044391) |
.3683505**
(.0405553) |
.291532**
(.033832) |
| Sargan test statistic | 0.003 | 0.018 | 0.024 | |
| Error term
AR(1) serial correlation test AR(2) serial correlation test |
0.000 0.958 |
0.000 0.986 |
0.000 0.653 |
|
| Number of observations | 330 | 330 | 300 | 330 |
| Number of Instruments | 198 | 198 | 167 |
Note: (1) Standard Errors are under each coefficients, and * means p<5%, ** means p <1%, vacancy means that coefficient is not significant. (2) In GMM estimation, I only consider time dummies are strictly exogenous and insert them as each single column. Adversely, other instruments variables are considered as not exogenous and are treated lagged values. (3) I cannot reject the null hypothesis in Sargan test is difference GMM and second system GMM. That means instrument variables are uncorrelated to some set of residuals, and therefore they are acceptable, healthy, instruments.
As discussed above, I accept the system GMM result reported in the second column of Table 2 as the test estimator for this case. This model performs generally well with most variables significant at the .01 level. The signs of the coefficients are generally consisted with expectations.
First of all, I want to explain the two variables which are excluded from this model: GDP per capital and budgetary revenue per capital. The motivation that I use the absolute number of GDP rather than increase rate of GDP in the model is to estimate the relationship between different GDP level and infrastructure spending. In other words, theoretically, poorer provinces have relative little resource to increase infrastructure spending. Similarly, the budgetary revenue should have the same effect on infrastructure spending. However, both variables are not significant in the model. This result relates to the financing mechanism of local governments in China. The budgetary revenue and GDP cannot completely demonstrate the financing ability of Chinese local governments because out-budget revenue or out-system revenue account large proportion in total local financial resource. Ping (2006) show that about 50% of revenue in local governments comes from out-budget and out-system revenue based on case study. So in order to attract investment, even poorer provinces would use various methods, such as land leasing, borrowing, or toll fees, to increase infrastructure spending, so that the officials can have better performance.
Secondly, the results show that increase rate of infrastructure spending is negatively with the base infrastructure investment. This indicates infrastructure spending has obvious catch-up effect which means the increasing rate of infrastructure spending in provinces with less base infrastructures grow faster (see Chart 1). Generally, the western area is less developed than eastern and middle parts. This result also implies the fact that political centralization and yardstick competition promote less developed provinces playing actively role in infrastructure spending.
Eastern- Western- Middle-
Chart 1 increasing rate of infrastructure spending in different areas of China
Source: Zhang Jun etc. (2007)
Thirdly, FDI per capital play a significantly positive role in determining the increase of infrastructure spending. This verifies my observations related to infrastructure spending decision making. As the most important yardstick competition measurement, FDI has significant positive effect on increasing rate of infrastructure spending. The better infrastructure can attract more foreign investment, and investors will require further infrastructure improvements after they invest in certain area. To this point, FDI plays the “vote by feet” role in China and incentive more infrastructures imputes.
Fourthly, the results of my panel data also show the administration fee per person has positive effect on the increase rate of infrastructure spending. The expanding administration fee reflects the expanding scale of local governments. Actually, the infrastructure investments are to large extent promoted by governments imputes in China. So it is clearly that expanding governments are willing to increase infrastructure spending so that officials would have better performance.
Fifthly, the positive coefficient of urban and rural saving per capital shows that local financial resource has positive effects on infrastructure spending. It is easier to understand this atmosphere if we consider the infrastructure financing channels of local governments. Although the local governments are illegal to issue municipal bonds, they can get long-term infrastructure loans from policy-oriented banks or commercial banks through various forms. So the local financial resource provided fund to finance local infrastructure spending.
Finally, signs of three time dummies, though some not significantly, are consistent with expectations. This demonstrates that selected time dummies can control some macro disturbances in the model. Whereas the Eastern Asian Financial Crisis in 1999 negatively influent the increasing rate of infrastructure spending, other events all play positive role in determining infrastructure investment. This result can logically stand.
The empirical result from estimation undertaken in this study highlights the unique effects of fiscal decentralization and political centralization on local infrastructure increasing rate. Except base infrastructure spending, FDI per capital, government scale and local financial resource all play significantly positive role in increasing infrastructure spending. The results conflicted classical public expenditure theory in Tiebout Model come from the incentive of fiscal decentralization with centralized political system as assumptions. For example, FDI instead of residents votes by feet to impulse local governments to increase infrastructure spending; officials are willing to expand the government scale and care little about the risk of borrowing to finance infrastructures because it will help them win the yardstick competition and get promotion.
The most contribution of this study is to help people understand the determine factors of increasing infrastructure spending under unique fiscal and political framework in China.
Two direct policy implications can be drawn as follows. Firstly, tax sharing reform in 1994 put pressure on local governments to expand infrastructure spending and unchanged fiscal decentralization gives incentives to local officials to increase infrastructure spending. Both sides can explain the reasons why the infrastructure spending increase dramatically during this period. Secondly, in order to improve the performance of infrastructure projects, we can not only focus on fiscal reform, such as increasing transparency of budget procedure, and the political reform cannot be ignored, such as promotion evaluation in bureaucracy system.
Yang He, PhD in economics
Associate Professor
School of Taxation, Central University of China