**4. Model Specifications**

Following previous studies (Aristotelous 2001; Chit et al. 2010), the model used for estimating the effects of exchange rate volatility and exports is specified as follows:

$$
\ln \text{EN}\_{it}^{m} = a\_{\text{cl}} + \beta\_1 \ln \text{GDP}\_{it} + \beta\_2 \ln \text{RE} \text{XR}\_{it} + \beta\_3 \ln \text{VOL}\_{it} + \varepsilon\_{it} \tag{1}
$$

where EX*mit* denotes the real export value in thousands of US dollars of the manufacturing sector as well as its 10 subsectors *m* at time *t* from Vietnam to its export partners *i*. GDP*it* represents the real Gross Domestic Product (GDP) in a foreign partner country *i* of Vietnam, deflated by the GDP deflator. The real bilateral exchange rates (REXR*it*) between Vietnam and its counterparts are measured by multiplying the relative price and the bilateral exchange rates, which are indirectly derived from US-based currency. The relative price is the ratio of the consumer price index (CPI) of export partners to the CPI of Vietnam.<sup>1</sup> Therefore, an increase in the value of the real exchange rate indicates a depreciation of Vietnam's currency. Finally, our variable of interest is the volatility of the bilateral exchange rate (VOL*it*), measured by the GARCH model.

The first step is to check whether all variables of interest are stationary. We used three panel unit root tests, including those of IPS (Im et al. 2003), Maddala and Wu (1999), and Choi (2001). Unlike other types of panel stationary tests, these tests allow data to be unbalanced. Next, the long-run relationship among these variables was checked using the cointegration tests introduced by Pedroni (1999, 2001), together with long-run estimations based on Panel Dynamic Ordinary Least Squares (DOLS). We used DOLS because it is asymptotically unbiased, normally distributed, and controls for the problem of endogeneity. Finally, to investigate the influence of the exchange rate on the growth of exports during the surveyed period, the equation was transformed in terms of an error correction model (ECM):

<sup>1</sup> We would like to thank an anonymous referee for his/her suggestion to use the real bilateral exchange rate. Many of the countries in the sample had varying inflation rates relative to Vietnam.

$$\begin{split} \Delta ln \text{EX}\_{it}^{\text{m}} &= a\_i + \beta\_1 \Delta ln \text{EX}\_{it-1}^{\text{m}} + \sum\_{j=0}^{n} \gamma\_j \Delta ln \text{GDP}\_{it-j} + \sum\_{j=0}^{n} \theta\_j \Delta ln \text{RE} \text{XR}\_{it-j} \\ &+ \sum\_{j=0}^{n} \theta\_j \Delta ln \text{VOL}\_{it-j} + \varrho \text{EC}\_{it-1} + \varepsilon\_{it} \end{split} \tag{2}$$

where Δ represents the difference between variables after taking their logarithm. EC*it*−<sup>1</sup> is a lagged error term that is derived by estimating Equation (1).

It is expected that becoming a member of the World Trade Organization (WTO) and the global financial crisis are events that had clear effects on export performance in Vietnam, and they were taken into consideration as well. Nguyen (2016) asserted that the WTO accession was the turning point in Vietnam's trade policy, thus potentially impacting its export performance. The author used Chow breaking tests to detect either a structural change and or regime change between the manufacturing and non-manufacturing sectors. The results reveal that, according to the model, there was a structural change beginning in 2007. In this sense, we employed the dummy variable *DWTO*, which is given a value of 1 as of 2007, when Vietnam officially entered the WTO. Another dummy *DCrisis* was assigned the same unit for 2009, the year of the global financial crisis. Thus, the following equation was used for the estimation:

$$\begin{split} \Delta ln \text{EX}\_{it}^{m} &= \alpha\_{i} + \beta\_{1} \Delta ln \text{EX}\_{it-1}^{m} + \sum\_{j=0}^{n} \gamma\_{j} \Delta ln \text{GDP}\_{it-j} + \sum\_{j=0}^{n} \theta\_{j} \Delta ln \text{RE} \text{XR}\_{it-j} \\ &+ \sum\_{j=0}^{n} \theta\_{j} \Delta ln \text{VOL}\_{it-j} + \beta\_{4} D\_{\text{Crisis}} + \beta\_{5} D\_{\text{WTO}} + q \text{EC}\_{it-1} + \varepsilon\_{it} \end{split} \tag{3}$$

Annual data over the 2000–2015 period were used in this study. The real foreign GDP, deflated by the GDP deflator, originated from World Bank Indicators, while the values of exports were from Organization and Economic Co-operation and Development (OECD) statistics. Although the Standard Industrial Classification (SIC) Codes classify the manufacturing sector into 22 subsectors, the OECD classification groups them into 10 major ones. As such, we tend to use these 10 subsectors of manufacturing exports because of the data collection. Also, some sectors do not engage in exporting in Vietnam, so using 10 subsectors reduces the problem of missing data. The bilateral exchange rate and the consumer price index (CPI) were taken from International Financial Statistics (IFS). It should be noted that the GARCH model requires high-frequency data to ensure accuracy. Thus, we adopted the monthly bilateral exchange rate to estimate the volatility. To convert monthly volatility to annual data, we averaged the volatility of the relevant year. Table 2 summarizes all of the data in the study.



All variables are in logarithm term, except for volatility measures.

To analyze the impact of geographical characteristics on the link between exchange rate volatility and export performance of the manufacturing sector and its 10 subsectors, we not only used the whole sample for the estimation, but also applied all of the above steps to three different regions (Asia, Europe, and America).

As previously discussed, there are diverse volatility measures used in empirical studies. However, in this study, for a particular nation, we applied the General Autoregressive Conditional Heteroscedasticity (GARCH) model to measure exchange rate volatility. The GARCH model includes two equations: (i) the mean equation and (ii) the conditional variance equation. With the condition that the log difference of an exchange rate series follows the random walk model, the GARCH model is suitable for the measurement of volatility. For GARCH(1,1), the two equations were constructed as follows:

$$\begin{array}{c} \boldsymbol{e}\_{t}^{i} = \boldsymbol{\mu}\_{0} + \boldsymbol{\mu}\_{1} \boldsymbol{e}\_{t-1}^{i} + \boldsymbol{\mu}\_{t'}^{i} \text{ where } \boldsymbol{\mu}\_{t}^{i} \sim \mathcal{N}(0, \boldsymbol{h}\_{t}^{i}), \text{ and} \\ \text{VOL}\_{\text{GARCH}} = \boldsymbol{h}\_{t}^{i} = \boldsymbol{\beta}\_{0} + \boldsymbol{\beta}\_{1} \boldsymbol{\mu}\_{t-1}^{i} + \boldsymbol{\beta}\_{2} \boldsymbol{h}\_{t-1}^{i}. \end{array}$$

The conditional variance equation of GARCH(1,1) consists of a constant *β*0, an ARCH term *<sup>μ</sup>i*2*t*−<sup>1</sup> and a GARCH term *<sup>h</sup>it*−1. We utilize the monthly data into the GARCH model and the monthly volatility of exchange rates is the conditional variance.

It is vitally important to adopt the appropriate GARCH model for estimating exchange rate volatility. Nishimura and Hirayama (2013) propose three steps in estimating GARCH-based volatility. The procedure begins with checking the appearance of ARCH effects by using ARCH-LM heteroscedasticity test, and then selecting the length of the optimal lag using Akaike information criterion (AIC) in the mean equation. Next, the second is to estimate the mean and variance equation simultaneously, then determining the appropriate model of ARCH and GARCH terms with the minimum value of Schwarz's Bayesian Information Criterion (SIC). Finally, the Ljung-Box tests are performed on the standardized residuals and standardized residuals squared. The optimal model is determined if these Ljung-Box tests can reject the null hypothesis of no autocorrelation. Although a few studies have attempted to use various lag lengths in the GARCH model (Asteriou et al. 2016), empirical evidence has confirmed that the GARCH(1,1) model is the most appropriate measure of exchange rate volatility (Chit et al. 2010; Erdem et al. 2010). In a recent investigation by Vieira and MacDonald (2016), the use of GARCH(1,1) appeared to predominate among various types of ARCH models for measuring volatility, as it was found in up to 75 out of 106 ARCH series. In addition, Hansen and Lunde (2005) asserted that the GARCH(1,1) model was superior to other complicated GARCH models when they took 330 ARCH-type specifications into consideration. In this sense, the GARCH(1,1) was utilized for the volatility measurement.
