In this section, we first introduce the variables and their sources in 3.1, and then propose three sets of empirical models in 3.2, and lastly in 3.3 we show the results of panel unit root tests. In each set of model, there are two specifications in which the industry-level export value and export share are used as the dependent variables, respectively. In particular, in Model 1.1 and 1.2, we apply the ordinary least square method. In Model 2.1 and 2.2, we then implement fixed effect panel data analysis. In Model 3.1 and 3.2, we further study the industry-specific effects of the real effective exchange rate.
3.1 Variables and data source
We examine the impact of the RMB revaluation on the export value and share of China’s exports in manufacturing sectors at the industry level over the period of 2001 to 2009. To calculate the key variables, i.e., the industry-level export value (𝐸𝑥𝑣) and export share (𝐸𝑥𝑠), we mainly use the China Customs Database. This dataset includes all transactions in terms of the Harmonized Commodity Description and Coding
System (HS code) reported by the exporters and importers in China within each year.
Notice that since our focus in this study is at the industry level instead of product level, we will have to re-categorize all exporting goods into industry classification following
The International Standard Industrial Classification (ISIC) of All Economic Activities,
which ends up with the industry-level data of export value and share of 27 manufacturing industries.5The sources of other main explanatory variables are as follows. We calculate the real effective exchange rate (𝑅𝐸𝐸𝑅) using the data of 80 main trade partner countries of China from the International Financial Statistics. We obtain and calculate the real GDP of China (𝐶ℎ𝑛𝑌) and the world (𝑊𝑌), represented by the same 80 countries as above, from the National Accounts Main Aggregates Database published by the United Nations. The industry-level variables, including sales per worker (𝑆𝑎𝑙𝑒𝑠𝑒𝑚𝑝), output per worker (𝑂𝑢𝑡𝑝𝑢𝑡𝑒𝑚𝑝) and capital-labor ratio (𝐾𝐿) are calculated by the data from the Chinese Industrial Enterprises Database (CIED) conducted by the National Bureau of Statistics (NBS) of China. The exact definition and calculation of these variables will be illustrated in detail with the empirical model specifications in 3.2.
Table 3: List of Variables
<to be inserted here>
3.2 Econometric model
We investigate the determinants of the industry-level export value and export share in total export value of each industry in China’s manufacturing sector.6 We start with, for industry 𝑗 in year 𝑡, the benchmark model of OLS regression in two versions, “export value” (Model 1.1) and “export share” (Model 1.2):
(1.1)
ln𝐸𝑥𝑣𝑗𝑡 = 𝛼0+ 𝛼1ln𝑅𝐸𝐸𝑅𝑡+ 𝛼2ln𝑊𝑌𝑡−1+ 𝛼3ln𝐶ℎ𝑛𝑌𝑡+ 𝛼4ln𝑃𝑑𝑡𝑦𝑗𝑡+ 𝛼5ln𝐾𝐿𝑗𝑡 + 𝑢𝑗𝑡
5According to China’s Industrial Classification for National Economic Activities, category C (manufacturing sector) contains the industries from code 13 through 42. In addition, originally there are 29 industries in manufacturing sector. The industry of 27 is dropped since the China Customs Database is lack of the records for the years of 2001 through 2005. The industry of 33 is dropped since the China Customs Database is lack of the records for the year of 2008.
6We also study the data at the macro level by regressing China’s overall export value on the real effective exchange rate, world income, and the dummy variables for China’s entering WTO and exchange rate reform. We find the coefficients significant and positive for the former two variables, negative for the reform dummy, and insignificant for the WTO dummy. This result implies that a RMB depreciation and world income increase are associated with a rise in China’s overall export value, which is consistent with the findings of the literature using macro trade data.
(1.2)
𝐸𝑥𝑠𝑗𝑡= 𝛼0+ 𝛼1ln𝑅𝐸𝐸𝑅𝑡+ 𝛼2ln𝐶ℎ𝑛𝑌𝑡+ 𝛼3ln𝑃𝑑𝑡𝑦𝑗𝑡+ 𝛼4ln𝐾𝐿𝑗𝑡+ 𝑢𝑗𝑡;
with the notations explained as follows and all variables taken natural logarithm. For the dependent variable, 𝐸𝑥𝑣𝑗𝑡 is the export value (adjusted with 2010 price of China);
𝐸𝑥𝑠𝑗𝑡 is the export share which is defined as 𝐸𝑥𝑗𝑡⁄𝐸𝑥𝑡, the share of export value of industry 𝑗 in the overall export value of all 27 industries (𝐸𝑥𝑡) in year 𝑡.
However, the sample industries in these two regressions are different. In the regression for export values (Model 1.1, 2.1 and 3.1), we include all 27 industries. In the regression for export shares (Model 1.2, 2.2, and 3.2), we include only 14 industries that was ranked the 10th largest exporting industries in any of the years during the sample period, 2001-2009.7 The reason for the latter is twofold: (1) we want to focus on those industries of highest shares which play the most important role in China’s exporting sector and economic development; (2) we want to avoid the potential high collinearity issue caused by the fact that the sum of the export share of all industries should be equal to one if we consider all industries. The selection of sample industries is also applied to the following regressions.
For the independent variables, the variable 𝑅𝐸𝐸𝑅𝑡 which is obtained from the real effective exchange rate (hereafter REER) in year 𝑡 between the RMB and the currencies of 80 China’s exporting partner countries, with each currency’s weight equal to China’s export value to that country over the total export value in that year. It can be shown as below: weighted sum of real GDPs of 80 China’s main exporting countries in the world in the year 𝑡, using the respective country-specific export share as the weight. According to the gravity model, we expect its coefficient to be positive. The variable 𝐶ℎ𝑛𝑌𝑡 is the
7 The 14 industries include the industry 14, 17, 18, 19, 24, 26, 32, 34, 35, 36, 37, 39, 40 and 41. When running the regression, Stata automatically takes industry 34 as the benchmark industry.
real income of China. We introduce this variable since we think that China’s exporting sector is very likely to be affected by its own economic development course, and we expect its coefficient to be positive.
As for the two industry-level productivity (denoted by ln𝑃𝑑𝑡𝑦𝑗𝑡 in the models above), we consider two measures, 𝑆𝑎𝑙𝑒𝑠𝑒𝑚𝑝𝑗𝑡 and 𝑂𝑢𝑡𝑝𝑢𝑡𝑒𝑚𝑝𝑗𝑡 (also denoted by SPC and OPC), which are the averaged per-worker sales and per-worker output of enterprises of industry 𝑗 in year 𝑡, respectively. We expect their coefficients to be positive because higher productivity is more likely to enhance the competitiveness of an industry. The last variable ln𝐾𝐿𝑗𝑡 represents the average of capital-labor ratio of enterprises of industry 𝑗, whose sign is expected to be positive, given that the development of China’s exporting sector is moving from labor intensity toward capital intensity. Lastly, 𝑢𝑗𝑡 is the error term of each observation for industry 𝑗 in year 𝑡, assumed to follow log-normal distribution.
We further consider the industry fixed effects by proposing Model 2.1 and 2.2 as follows. To explore whether the two events of China’s foreign exchange reform in 2005 and global financial crisis in 2008 affect China’s industry-level export value and share of industries, we include two year dummy variables 𝐸𝑥𝑅𝑒𝑓 and 𝐹𝑐𝑟. Model
where 𝜆𝑗 measures the industry fixed effect for industry 𝑗, regardless of years.
The dummy variable 𝐸𝑥𝑅𝑒𝑓denotes China’s foreign exchange reform in 2005, equal to 1 for 2005-2009 and 0 for previous years. Another dummy variable 𝐹𝑐𝑟
l n𝐸𝑥𝑣𝑗𝑡=𝛼0+ 𝛼1ln𝑅𝐸𝐸𝑅𝑡+ 𝛼2ln𝑊𝑌𝑡−1+ 𝛼3ln𝐶ℎ𝑛𝑌𝑡+ 𝛼4ln𝑃𝑑𝑡𝑦𝑗𝑡+ and 0 otherwise). Thus, the effect of the REER on each industry 𝑗 can be measured by the sum of 𝛼1 and 𝛼7.
3.3 Testing for unit root
The panel dataset used in this study contains both time series and cross sectional information, which is helpful for enhancing the degree of freedom and reducing the estimation bias. To access to such dataset with regression analysis, however, we need to first test whether the variables in question exist the issue of non-stationary by unit root tests, so as to avoid the problem of spurious regression and the corresponding inefficiency results of estimation. We test for panel unit root with three methods including Levin, Lin and Chu test (LLC), Im, Pesaran and Shin test (IPS), and Harris-Tzavalis test (HT) and show the results in Table 4. The results suggest that although five out of eight variables are shown to be stationary with the method of LLC, almost all variables are tested to be non-stationary with other two methods. We therefore take first-order difference for all variables except ln𝑅𝐸𝐸𝑅𝑡 to ensure stationarity.8
Table 4: The Results of Unit Root Tests
<to be inserted here>
To prevent from losing too many observations (we only have nine periods) and given that the P-values are not very high for other two methods, we thus decide not to take difference for this variable, but need to report this as a caveat.