國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
13
frequency via Quadratic-Match Sum in order to match with the frequency of the air freight data. The real GDP should have a positive relation to the country’s imports and exports since it represents the production power and people’s consumption power economically for the market.
4. Empirical Models
There are six models used in this study and they are classified into models (1) through (6). Model (1) is the classical Ordinary Least Square (OLS) model with some control variables. Models (2) is the fixed effect model that considers the product category fixed effect (38 categories) and time fixed effect (116 months).8 Models (3) to model (6) are all extended models based on the fixed effect model (2), which consider the specific effects of five main categories (model (3)), the dynamic panel setting by adding the 1-month lagged dependent variable (model (4)), the lead 1-month independent dummy variables for the policy variables from L1 through L4A (model (5)), and the category specific time trend (model (6)), respectively.
In each model, four approaches from (a) through (d) are considered for different dependent variables. In approach (a), (b), and (c), I use the air freight trade volume of TPE, KHH, and the sum of TPE and KHH as the dependent variables, respectively. In approach (d), I use the constructed air freight trade value instead. All models with various approaches are then studied by the imports from China (denoted with I) to capture the trade deflection effect, and exports to the U.S. (denoted with E) to capture the trade creation effect. Hence in what follows, the equation numbering has three digits:
the model code, the approach code, and the import/export code. For instance, model
8 The Hausman test shows that the fixed effect model is more appropriate for the panel data of this study than the random effect model.
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
14
(1.a.E) stands for the OLS model considering the approach of air freight trade volume in terms of exports, and model (2.b.I) for the fixed effect model considering the approach of air freight trade value in terms of imports.
4.1 The OLS Model
As mentioned above, the OLS model considers two approaches for the imports from China (trade deflection) and the exports to the U.S. (trade creation), respectively.
Hence model (1.a.I) and (1.a.E) both share the same form as follows:
(1.a.I) 𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑗𝑗𝑗𝑗= 𝛽𝛽0 + 𝛽𝛽1𝐿𝐿1𝑗𝑗𝑗𝑗+ 𝛽𝛽2𝐿𝐿2𝑗𝑗𝑗𝑗+ 𝛽𝛽3𝐿𝐿3_10𝑗𝑗𝑗𝑗+ 𝛽𝛽4𝐿𝐿3_25𝑗𝑗𝑗𝑗+ 𝛽𝛽5𝐿𝐿4𝑙𝑙𝑗𝑗𝑗𝑗+ 𝛽𝛽6𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑗𝑗+ 𝛽𝛽7𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑗𝑗+ 𝜀𝜀𝑗𝑗𝑗𝑗, (1.a.E) 𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑗𝑗𝑗𝑗= 𝛽𝛽0 + 𝛽𝛽1𝐿𝐿1𝑗𝑗𝑗𝑗+ 𝛽𝛽2𝐿𝐿2𝑗𝑗𝑗𝑗+ 𝛽𝛽3𝐿𝐿3_10𝑗𝑗𝑗𝑗+ 𝛽𝛽4𝐿𝐿3_25𝑗𝑗𝑗𝑗+
𝛽𝛽5𝐿𝐿4𝑙𝑙𝑗𝑗𝑗𝑗+ 𝛽𝛽6𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑗𝑗+ 𝛽𝛽7𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑗𝑗+ 𝜀𝜀𝑗𝑗𝑗𝑗,
where 𝑙𝑙𝑙𝑙𝑖𝑖𝑗𝑗 denotes the trade volume (or quantity) through the channel of both airports for air freight category 𝑗𝑗 in month 𝑡𝑡 (notice that “𝑙𝑙” represents Aggregation) for the imports from China in (1.a.I) and the exports to the U.S. in (1.a.E), respectively. The independent variables 𝐿𝐿1𝑖𝑖𝑗𝑗, 𝐿𝐿2𝑖𝑖𝑗𝑗, 𝐿𝐿3_10𝑖𝑖𝑗𝑗, 𝐿𝐿3_25𝑖𝑖𝑗𝑗 and 𝐿𝐿4𝑙𝑙𝑖𝑖𝑗𝑗 all denote the policy dummy variables for the five times of tariff rate increase in four waves. For each of them, the dummy variable equals 1 when the Trump administration implemented extra tariff rates against China for category 𝑗𝑗 in month 𝑡𝑡, and equals 0 otherwise. For the control variables: 𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑗𝑗 is the real effective exchange rate of Taiwan NTD against the USD in month 𝑡𝑡; 𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑗𝑗 and 𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑗𝑗 are the monthly real GDP at time 𝑡𝑡 for the U.S. and Taiwan, respectively, according to the gravity model. And, 𝜀𝜀𝑗𝑗𝑗𝑗 is the error term.
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
15
The OLS models for the approach of air freight trade value through both airports for air freight category 𝑗𝑗 in month 𝑡𝑡 for and the imports from China (trade deflection) the exports to the U.S. (trade creation), i.e., models (1.b.I) and (1.b.E), are as follows:
(1.b.I) 𝑙𝑙𝑙𝑙𝑙𝑙𝑗𝑗𝑗𝑗 = 𝛽𝛽0+ 𝛽𝛽1𝐿𝐿1𝑗𝑗𝑗𝑗+ 𝛽𝛽2𝐿𝐿2𝑗𝑗𝑗𝑗+ 𝛽𝛽3𝐿𝐿3_10𝑗𝑗𝑗𝑗+ 𝛽𝛽4𝐿𝐿3_25𝑗𝑗𝑗𝑗+ 𝛽𝛽5𝐿𝐿4𝑙𝑙𝑗𝑗𝑗𝑗+ 𝛽𝛽6𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑗𝑗+ 𝛽𝛽7𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑗𝑗+ 𝜀𝜀𝑗𝑗𝑗𝑗,
(1.b.E) 𝑙𝑙𝑙𝑙𝑙𝑙𝑗𝑗𝑗𝑗 = 𝛽𝛽0+ 𝛽𝛽1𝐿𝐿1𝑗𝑗𝑗𝑗+ 𝛽𝛽2𝐿𝐿2𝑗𝑗𝑗𝑗+ 𝛽𝛽3𝐿𝐿3_10𝑗𝑗𝑗𝑗+ 𝛽𝛽4𝐿𝐿3_25𝑗𝑗𝑗𝑗+ 𝛽𝛽5𝐿𝐿4𝑙𝑙𝑗𝑗𝑗𝑗+ 𝛽𝛽6𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑗𝑗+ 𝛽𝛽7𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑗𝑗+ 𝜀𝜀𝑗𝑗𝑗𝑗,
where 𝑙𝑙𝑖𝑖𝑗𝑗 denotes the trade value through both airports for air freight category 𝑗𝑗 in month 𝑡𝑡 (notice that “𝑙𝑙 ” represents trade value) for the imports from China (trade deflection) in (1.b.I) and exports to the U.S. (trade creation) in (1.b.E), respectively.
Notice that the trade value here is calculated by multiplying the trade volume and the average unit price for category 𝑗𝑗 in month 𝑡𝑡.
4.2 The Econometric Issues
In this subsection, several tests will be conducted to test the correlation, collinearity, stationary, autocorrelation and heteroskedasticity of the data.
Firstly, the correlation test on the variables shows a significantly high correlation coefficient of 0.7253 between the first list (L1) and the second list (L2). High correlation issues may result from the following problems: 1) close implementation time that there is only a month of separation. 2) almost the same combination of goods is targeted, mainly electronics, machines, and transportation. 3) the same rate of tax (25%) is imposed. 4) missing data corresponding to the time of tariffs imposition. Thus, the main independent variable of L2, will be dropped because of the smaller scale. On the other hand, the real effective exchange rate (𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑗𝑗) and real GDP of Taiwan and
‧
the U.S. (𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑗𝑗& 𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑗𝑗) are significantly correlated with correlation coefficients of 0.7712 and 0.7153 respectively. Although the result shows a high relation (close to 0.75) for these variables, in order to control the constant value and the size of the market, these three variables will still be considered. Consequently, the dummy variable of L2 will be dropped but the REER and real GDP will be included in the models.
Secondly, the study also checks for the collinearity issue for the models since it will make the estimation of coefficients and standard error to become unstable and volatile. However, the issue of collinearity is not found in the models.
Thirdly, for the reliability of the estimation, this study tests the stationary issue to check the existence of the trend in the dataset. I conduct the Fisher-type unit-root test.
Because some values of the data in a certain period are missing, the multiple panels data could not meet the requirement of the augmented Dickey-Fuller unit-root test (ADF test) and Harris-Tzavalis unit-root test (HT test), where strongly balanced panel data are needed. Data with non-stationary implies that the direct estimator may cause spurious regression problems. This study then finds no existence of the non-stationary issues through the models. Therefore, the solution of using the first difference is not required.
Then, Table 5 shows the result of the Wooldridge test for autoregressive correlation.
Under the confidence level of 5%, the results reject the null hypothesis that there is no