Analytical approach and estimation models

Our analytical approach is adapted from that applied in Limao and Venables (2001) and applies a gravity model to predict bilateral trade and FDI flows by each pair of GMS members.⁷ Estimation parameters of particular interest are the responses of trade and FDI to various transport cost factors including cross-border road infrastructure, and the determinants of investments in cross-border transport infrastructure. Accordingly, the empirical analysis centers around three functional relationships: a trade equation, an FDI equation, and a cross-border road infrastructure equation.

1. Trade equation:

• X_ij : exports of country i to country j via land
• Y_i , Y_j : vector of fixed or predetermined characteristics of country i (j) related to trade such as distance, economic size (GDP), population, land area, domestic road infrastructure, and similar variables routinely used in gravity model estimates.
• F_ij : country i’s foreign direct investment from country j.
• R_i , R_j : vector of variables measuring border area and general domestic road infrastructure of country i (j ).
• ?_ij : other factors not accounted for (model error).

The trade equation incorporates standard variables used in gravity models plus variables of particular interest in this research (i.e., measures of cross border and domestic road infrastructure, and FDI from the trading partners). Other factors seen as important in driving levels of bilateral trade, which are elements in vectors Y_i and Y_j) are tariff rates, inflation rates, and a broad characterization of the export/import environment in the countries. A principal aim in the analysis is to quantify the “incremental effect” of cross-border road infrastructure on trade relative to the effect of domestic road infrastructure. Trade is envisioned to be a function of both the quality of road infrastructure generally in each country and of road infrastructure in border areas in particular. Both road indicators are seen as being relevant to determining the flows of goods and raw materials between the countries since they form part of the transport network used to connect markets. In the next subsection of the paper we discuss our expectations regarding the signs of estimation coefficients, while further details concerning the definition, measurement, and sources of data used are left to the notes to Table 1 [ PDF 111KB | 4 page(s) ] and the Appendixes [ PDF 176.4KB | 19 page(s) ].

The value of exports between pairs of GMS countries is of interest both because of our interest in exploring the empirical relationship between cross-border infrastructure and trade, and because trade levels are broadly indicative of transportation costs. While reliable information on overland transport costs is generally unavailable for GMS countries, examination of overland trade flows can provide some insight into changing overland transport costs in the GMS.

2. FDI equation:

• F_ij : country i's foreign direct investment inflow from country j
• Y_i , Y_j : vector of characteristics of country i and j (same as in trade equation)
• z_i : vector of characteristics related to country i’s investment climate
• R_i , R_j : vector of variables measuring border area and general domestic road infrastructure of country i (j).
• X_ij : exports of country i to country j via land
• e_ij : other factors not accounted for (model error).

The FDI equation specifies capital flows as being determined by several factors that also appear in the trade equation (e.g. economy size and resources, inflation rate, tariff rates). Of particular interest is the relative contribution of general road infrastructure and road infrastructure in the border area. In addition, FDI is viewed as being influenced by the volume of trade and the FDI and trade environment in the FDI-recipient country.

3. Cross border infrastructure equation:

• Road_ij : measure of the stock and quality of country i’s road infrastructure in the border area with country j.
• Y_i , Y_j : characteristics of country i and j (same as in transport cost equation).
• z_i : vector of characteristics related to country i’s trade and investment climate
• R_i ,R_j : vectors of variables measuring general domestic road infrastructure of country i/j.
• X_ij : exports of country i to country j via land
• ?_ij : other factors not accounted for (model error).

Lastly, we define the cross-border infrastructure equation wherein structural characteristics of the country, the investment climate, the level of trade and FDI flows, and the quality of roads in the country in general are related to cross-border road infrastructure. The main reason for including this equation in our analysis is to examine the possibility of reverse causality in the construction of cross-border infrastructure, i.e., this equation allows for the possibility that the construction of roads in border areas is a response to—rather than a cause of—trade and FDI flows.

Dataset, estimation model, and estimation procedures

Our dataset is formed from a cross-sectional time series of data available for GMS member economies for the period of 1981-2003. Observations in the dataset are defined at the country-pair level over time. In all, 30 country pairs can be formed across the ⁶ GMS member countries (i.e., Cambodia-Lao PDR, Cambodia-Myanmar,…, Yunnan (PRC)-Thailand, Yunnan (PRC)-Viet Nam). Descriptive statistics from the dataset along with details on the data sources and definitions of variables are summarized in Table 1 [ PDF 111KB | 4 page(s) ]. Because the resulting dataset captures the value of variables for the country-pair over time, Table 1 [ PDF 111KB | 4 page(s) ] presents the number of observations of each variable and country-pair over time (years). Nonetheless, due to the small number of GMS countries and relatively short time period for which most data are available for some GMS countries, our analysis faced challenges in model estimation. For example, data at the start of our panel is available for only a few GMS countries because some of the poorer GMS countries suffered major military conflicts in the 1970s and were only establishing or recovering their national statistical capacity in the early 1980’s.

The Appendix provides detailed explanations of key variables and of the sources of data. Two key concepts are cross border infrastructure and domestic road infrastructure. For the former we use as a proxy the road density in the provinces that share a border with a GMS neighbor. Where there is more than one such provinces we take an average. For the latter we use the average road density of all provinces in a country that do not share a border with a GMS neighbor. Limitations in available data representing transport costs in the GMS made us forgo the estimation of the determinants of transport costs (as in Limao and Venables, op. cit.), so instead we estimate the trade and FDI equations with road infrastructure being one of the explanatory variables. Also, quantification of indirect economic impacts that come through trade and FDI is judged premature and is deferred until a more rigorous structure of the trade-FDI nexus can be modeled and supported by improved data.⁷

Following the general functional relationships defined above, our estimation models define total exports, FDI, and investments in cross-border infrastructure (Xij) from country i to country j in time t as:

and the following signs are hypothesized for the estimation parameters:

In logarithmic form, we have:

Country GDP is considered a key variable in the base gravity model, and larger economies are expected to engage in greater trade. Trade is viewed as being positively affected by the economic mass of the trading partners and negatively affected by the distance between them. Other factors also act against the ‘gravity like’ forces of economy size. Geographic area and population size are factors expected to reduce trade orientation by increasing the size of the domestic market and making economic activity more inwardly oriented. Additional variables, such as indicators of cultural affinity and sharing contiguous borders are usually added to empirical gravity models. Using this as base model, we can add variables for cross-border road infrastructure and FDI to consider the effect of these two variables on trade flows—controlling for the standard variables treated in the gravity model—and providing our basis for estimating the trade equation outlined above.

Models are estimated using the Generalized Least Squares (GLS) Random Effects estimator for cross sectional time series data. We forego detailed discussion of technical details pertaining to the estimation procedure except to note that estimation coefficients reflect a weighted average of the cross-sectional and time-series association between the dependent and independent variables included, and the weighting is defined by the estimation parameter theta—which is reported for our panel estimates.⁸

The overall statistical significance of the estimation models is tested using a Wald Chi-square test, while the need for the random effects estimator as opposed to treating the cross-sectional time-series data simply as a cross-section and applying regular GLS is tested through a Breusch and Pagan Langranian Multiplier test (technical details are also in Green, 2003). The Wald Chi-square test indicates the probability of a false rejection of the null hypotheses that the model has no explanatory power over the dependent variable. The statistical significance of estimation parameters is tested using a test that is functionally equivalent to a standard t-test applied in Ordinary Least Squares (OLS) and GLS regressions. Estimation coefficients can be interpreted as elasticities following the standard treatment of log-linear regressions.

We also estimate our models for single years of data using standard GLS estimation. However, cross-sectional estimates using single years of our data offer a clearly inferior estimation approach as they do not take advantage of the panel data's capacity to trace the impact of changes in cross-border road infrastructure over time. In addition, cross-sectional estimates face severe sample size constraints. Nonetheless, they can provide insight into the evolution of the relationship between our dependent and explanatory variables over time.

Download this Discussion Paper [ PDF 309.9KB| 35 pages ].

Comment(s)

There are [0] comment(s) for this entry. Post a comment.