of the bilateral pairs of countries report differing level for the trade between them, the approach of Kox and Lejour (2005) is followed to determine a ranking of countries in terms of their reporting reliability. After the duplicate bilateral pairs and missing data are removed, there remain approximately 1400 bilateral country pairs reporting imports of total services for at least one of the years in the sample (1999-2001). Pooled across the three years there are just over 3800 observations. The number of country pairs and observations for the disaggregated sector data are lower as countries are less likely to provide full information on the four sub-categories of services.
GDP, GDP per capita and Population
Data on GDP, GDP per capita (both in US dollars) and population variables are from the World Development Indicators database (World Bank, 2005).26 Some regressions were implemented using GDP at purchasing power parity but this was not found to significantly change the results.
Distances are measured as the distance between capital cities of each country in kilometres and are from the Great Circle Distance programme.27
The dummy for adjacency takes the value of one if the two countries have a common border and zero otherwise. In the dataset there are 47 country pairs that share a common border. The EU dummy takes a value of one if the two countries are members of the European Union. The language dummy takes the value of one if the two countries share a common (official) language and zero otherwise. In total there are 174 pairs of countries that share a common language in the data.
Given the possible importance of a shared language to services trade, some simulations were implemented with an extended language variable that incorporated data on the prevalence of non-official languages and the degree to which they are spoken in countries (Clair et al., 2004). It is continuous variable that ranks the degree of similarity of the languages of trading partners (or the linguistic distance between them). This had little influence on regression results compared to the use of the standard variable. This may reflect the fact the sample of countries covered in the model are primarily OECD countries, the majority of which have one dominant language in use. The extended language variable does not capture significant variation that is not already captured in the standard dummy variable method.
The summary statistics for each of the explanatory variables are shown in table 1. To examine the possibility of multicollinearity amongst the explanatory variables, the spearman rank correlation coefficients are shown in table 2. As there are no two variables with a correlation greater than 0.37, there is no evidence that multicollinearity is a major issue with these variables
26 Belgium and Luxembourg are treated as single country in OECD (2003), so the sum is used for GDP per capita and population variables. Available at: http://www.wcrl.ars.usda.gov/cec/moregen.htm. 27