Microsoft Word Cheongaftaeffects doc

Empirical Model and Data 

 

There exist many studies at the aggregate level on the trade impact of PTAs with mixed 

evidence for trade creation or diversion (see Baier and Bergstrand, 2007; Glick and Rose, 

2002; Ghosh and Yamarik, 2004). Adams et al. (2003) provide an extensive review of the 

empirical findings of such aggregate-level studies. They also summarize the results of 

seven studies that consider the trade effects of AFTA, noting that all these studies find 

more trade creation than trade diversion. However, in studies not considered by them, 

two do not find positive results for AFTA (Sharma and Chua, 2000; and Elliott and 

Ikemoto, 2004), while two do find evidence for greater intra-ASEAN trade (Hapsari and 

Mangungsong, 2006; and Bun, Klassen, and Tan, 2007).  

 

To contribute to the empirical literature on AFTA, I use the gravity model (see Anderson, 

1979; and Bergstrand, 1985) for theoretical foundations) as the framework for my 

empirical analysis. The gravity model is an expenditure equation that basically relates 

one country’s import expenditure on another country’s output to the importing and 

exporting countries’ GDPs and trade costs between the two countries. Trade costs capture 

all barriers that impede trade and include all border-related (e.g. tariff) costs, 

transportation and distribution costs, and exchange rate costs. Although the model is 

often estimated with total bilateral trade, the underpinning theory actually explains 

unidirectional trade from one country to another. Furthermore, as Anderson and Wincoop 

(2004) point out, estimating the gravity model at an aggregated level is likely to lead to 

biased estimates because border barriers and elasticities of substitution vary by product.  

In addition, as Medvedev (2006) finds, an estimation of the gravity equation with 

aggregate data and PTA dummy variables may detect little benefit from preferential trade 

liberalization because the PTA may initially exclude a significant number of products and 

have long phase-in periods. With disaggregated data on trade and preference margins, I 

am able to control for unobserved heterogeneity at the import-exporter-product level. So, 

---

 

 

 

6

 

to explain imports (M

ijk

) of country i from country j of product k in year t, I use the 

following estimating equation: 

 

ijkt

t

ijk

jt

it

ASEAN

j

ikt

NONASEAN

j

ikt

ijkt

GDP

GDP

D

Prefmarg

D

Prefmarg

M

ε

α

α

β

β

β

β

β

+

+

+

+

+

=

+

=

+

=

ln

ln

)

1

(

*

ln

)

1

(

*

ln

ln

4

3

2

1

0

 

 

The first two explanatory variables are the natural logarithm of the preference margin 

offered by country i (ln Prefmarg

ikt

) interacted with dummies for nonASEAN trade 

partners and ASEAN trade partners. The preference margin is calculated as the difference 

between the simple average MFN tariff and the simple average CEPT tariff. The set of 

importer (i) countries considered is only ASEAN. The coefficient on the Prefmarg-

nonASEAN interaction will capture trade diversion, if negative, while that on the 

Prefmarg-ASEAN interaction will measure trade creation, if positive.  The difference 

between the coefficients will provide a measure of the net welfare effects of AFTA. The 

next two explanatory variables are those conventionally used in gravity equations, i.e. 

GDP of the importing and exporting countries. To capture all time-invariant effects of 

any pair of trading countries, which includes the distance variable, I include importer-

exporter-product  fixed effects

ix

. Besides distance, these fixed effects capture variables 

such as remoteness, exchange rate regimes, colonial history, common languages, prior 

industry expectations, and other fixed industry-level characteristics. I also include year 

effects.

 

 

As Santos and Senreyro (2006) note, estimation of the log-linear specification of the 

gravity model with OLS in the presence of heteroskedasticity will produce inconsistent 

estimates of the elasticities. This result, which applies generally to log-linearized models, 

obtains because the non-linear transformation of the estimating equation makes the 

conditional expectation of the error depend on higher moments of the data. They 

recommend using the Poisson Pseudo- (or Quasi-) Maximum Likelihood (PQML) 

estimator to solve this problem. Another benefit of this estimator is that zero observations 

of the dependent variable can be included in the estimation to avoid truncation bias. 

Given panel data, I estimate the model with fixed-effects PQML. The PQML estimator is 

perfectly robust to heteroskedasticity if the conditional variance is proportional to the 

conditional mean. However, as this assumption is unlikely to hold, I compute robust 

standard errors as derived by Wooldridge (1999) for the fixed-effects PQML estimator. 

 

The import data is taken from COMTRADE, MFN and preferential tariff rates are 

sourced from the ASEAN Secretariat, and GDP data are from the World Bank’s 

Development Indicators. As tariff-rate data is published at the HS 9-digit level, I compute 

the tariff rates for each HS 6-digit category as the simple average tariff. This is necessary 

as the most disaggregated import data available for ASEAN countries from COMTRADE 

are at the HS 6-digit level.  

 

The sample is limited to the period 2001 to 2003 for three reasons. First, CEPT tariffs 

during this period were predetermined by ASEAN countries in 1998 following the Asian 

Crisis. Therefore, CEPT tariffs in 2001 to 2003 are less likely to be endogenous and 

neither are preferential margins as MFN tariffs did not change from 2001 to 2003. 

---

 

 

 

7

 

Second, data on ASEAN preference margins are available to the public only from 2001 

onwards. Third, there was a significant change in 2002 of the HS trade-classification 

system, which was subsequently implemented in 2003 onwards by ASEAN countries. As 

such, by ignoring data after 2003, I avoid attributing any AFTA effects spuriously to the 

change in trade classification.  In addition, this study does not consider foreign direct 

investment (FDI) and non-tariff trade barriers as explanatory variables although these are 

important factors in determining trade flows. The main reason for their exclusion is data 

availability – which is very rarely at a disaggregated level –  and data quality.