The empirical analysis of the effect of AfT flows on recipient-countries’ services export share is investigated by taking cue from the empirical literature on the determinants of services exports more generally, and more specifically from the above-mentioned few studies on the effect of AfT flows on recipient-countries’ services trade. Therefore, we consider specifically control variables that are deemed to influence the expected effect of AfT flows on recipient-countries’ services export share through merchandises exports and FDI inflows. These control variables include the real per capita income, the population size, the real exchange rate, trade policy, the financial development depth, the education level and the institutional and governance quality.
The use of the real per capita income (which also acts for the development level) as a control variable aims to account for the effect of the domestic demand structure on services exports (e.g., Linder, 1961). Specifically, the real per capita income reflects the economies of scale, as the latter can induce a higher demand for new services and consequently generate higher services production and exports (e.g., Li et al., 2005; Nyahoho, 2010; Sapir and Lutz, 1981). The population size variable aims to capture countries’ size. As noted by Goswami et al. (2012), the rise in the population size is associated with a higher demand for services, and the expansion of the services sector. This could, in turn, be associated with a rise in services exports. Thus, we expect a rise in the real per capita income to be positively associated with countries’ integration into the world market for services exports. Trade policy liberalization can affect services exports through its pro-competitive effect in domestic markets that increases firms’ productivity (e.g., Melitz 2003), its effect on the expansion of the market sizes that is conducive to economies of scale (e.g., Alesina et al. 2005) and finally through the promotion of knowledge diffusion (e.g., Grossman and Helpman 2015). The services exports effect of trade policy liberalization can also take place indirectly through the effect of trade policy liberalization on exports of goods. In light of the ‘network effect’ of merchandises exports on services exports (e.g., Eichengreen and Gupta, 2013a and Sahoo and Dash, 2014), one can expect that trade policy liberalization that promotes goods exports will result in greater services exports. Overall, we expect greater trade policy liberalization to result in greater integration of countries into the world market for services exports. The real exchange rate variable has been introduced in model (1) in light of the importance of this variable for services export dynamics. The few existing works on the matter have shown that an appreciation of the real exchange rate exerts a negative effect on services exports, which is even more pronounced than its negative effect on goods exports (e.g., Abeysinghe and Yeok, 1998; Baggs et al. 2010; Eichengreen and Gupta, 2013b; Sahoo and Dash, 2014; 2017; Smith, 2004). The education level variable (proxied here by the secondary school enrolment rate) has been included in model (1) because almost all empirical studies on the determinants of services exports have emphasized the importance of an educated workforce for the promotion of services exports (e.g., Eichengreen and Gupta, 2013a, 2013b; Goswami et al., 2012). Similarly, financial development is key for the enhancement of services exports (e.g., Sahoo and Dash, 2017). Finally, institutions matter for export dynamics (e.g., Araujo et al., 2016). In particular, the absence of good institutions prevent firms from taking advantage of new trade opportunities (e.g., Méon and Sekkat, 2008), including in services markets (e.g., Gani and Clemes, 2016; Kimura and Lee, 2006; Sahoo and Dash, 2014, 2017; Cattaneo et al. 2010).
In light of the foregoing, we consider the following baseline model (1):
Log(SERVINT)it=α0+α1 Log(SERVINT)(it-1)+α2Log(AfT)it+α3 Log(GDPC)it+α4 Log(POP)it+α5 Log(REER)it+α6 Log(TP)it+α7 Log(EDU)it+α8 Log(FINDEV)it+α9INSTit+μi+λt +ωit
The subscripts i and t represent respectively the countries’ index and the time-period. Model (1) is estimated using an unbalanced panel dataset of 105 countries (of which 35 LDCs) over the period 2002–2016, based on the availability on data concerning variables in this model. Non-overlapping sub-periods of 3-year average (i.e., 2002–2004; 2005–2007; 2008–2010; 2011–2013; and 2014–2016) have been used for the variables so as to mitigate the effect of business cycles on these variables (this yields 5 sub-periods). The parameters α0 to α9 are to be estimated. μi represent countries’ unobservable time invariant specific effects that can affect countries’ services exports. εit is a well-behaving error-term. λt are time dummies, and represent global shocks affecting simultaneously all countries’ services exports.
The dependent variable “SERVINT” represents for a given country in a given year, the share of services exports in the world services exports. As noted above, for the sake of simplification, we will henceforth refer it as to ‘services exports share’, ‘share of services exports’ or ‘services export integration’. The variable “AfT” stands for the measure of AfT flows, which can be either total AfT flows (“AfTTOT”) or one of its three components described in Section 2, i.e., AfT for economic infrastructure (“AfTINFRA”), AfT for productive capacity (“AfTPROD”), and AfT for trade policy and regulation (“AfTPOL”). All AfT variables are expressed in US Dollar, constant prices 2016. The variables “GDPC”, “POP”, “REER”, “TP”, “EDU”, “FINDEV” and “INST” represent respectively the real per capita income, the population size, the real exchange rate, trade policy, the education level, the financial development depth and the institutional and governance quality. The description of all these variables is provided in Appendix 1, and their related standard descriptive statistics are displayed in Appendix 2. Appendix 3 shows the list of the 105 countries, including that of the 35 LDCs. As it can be noted in model (1), the natural logarithm (“Log”) has been applied to all variables in model (1) (with the exception of the variable “INST” which has negative and positive values) not only to reduce their skewness, but also to obtain coefficient in terms of elasticities. The variables expressed in terms of ratios (i.e., “SERVINT” and “FINDEV”) are not expressed in percentage. Finally, the introduction of the one-period lag of the dependent variable as a regressor in model (1) aims to capture the state dependence nature of the share of services exports, i.e., its likely persistence over time.
We begin the empirical analysis by estimating a static specification of model (1) - i.e., by removing the one-period lag of the dependent variable among regressors - using the fixed effects (“FE”) estimator and the cross-section weighted feasible generalized least squares (FGLS) estimator. The variable “AfT” is measured by the total AfT flows. The results of these estimations are displayed in Table 1. However, all regressors (except for the population size variable) in this static specification of model (1) can be endogenous due notably to the reverse causality from the dependent variable to each of these regressors. Taking for example the AfT variables, one can expect that donors would provide a higher amount of AfT to countries with a lower share of services exports with a view to helping them promote their services exports. Likewise, countries that aim to further integrate into the world market for services exports can further liberalize their trade regime, adopt appropriate exchange rate policies, promote education of their citizen, adopt measures to enhance the development of their domestic financial markets, and promote better institutional and governance quality. Finally, a better integration into the world market for services exports can also influence countries’ development level, proxied by their real per capita income. In addition to the possible endogeneity of these regressors, estimating the static specification of model (1) can raise the omitted variables problem, due in particular to the omission of the one-period lag of the dependent variable. At the same time, the use of the FGLS approach to estimate the dynamic specification of model (1) (i.e., as it stands) can yield biased and inconsistent estimates because of the potential correlation between the one-period lag of the dependent variable and countries’ specific effects. The bias induced by this correlation is particularly due to the nature of our panel data, which contains a small time-period and relatively large cross-section dimension (this is the so-called Nickell bias - see Nickell, 1981). We take into account these endogeneity concerns by estimating the dynamic model (1) using the Generalized Methods of Moments (GMM) estimator, in particular the two-step system GMM estimator of Arellano and Bover (1995) and Blundell and Bond (1998). This estimator is particularly suitable for dynamic panels with a small-time dimension and large cross-section, and particularly where variables, notably the dependent variable exhibits a strong persistence over time. Additionally, this estimator helps handle the aforementioned endogeneity concerns (i.e., the correlation between the one-period lag of the dependent variable and countries’ specific effects, as well as the reverse causality issue). Estimating model (1) by means of the two-step system GMM estimator amounts to estimating a system of equations, where an equation in differences is combined with an equation in levels, and where lagged first differences are used as instruments for the levels equation and lagged levels are used as instruments for the first-difference equation. This estimator has been found to be more efficient than the difference-GMM estimator developed by Arellano and Bond (1991) when variables are persistent over time. This is because using lagged values of the first difference of variables generates weak instruments in the difference-GMM estimator (Arellano and Bover, 1995; Blundell and Bond, 1998). Therefore, the two-step system GMM estimator would be our preferred estimator. Its consistency is evaluated using the standard Arellano-Bond test of first-order serial correlation in the error term (denoted AR(1)) and no second-order autocorrelation in the error term (denoted AR(2)) (whose null hypothesis should not be rejected) as well as the Sargan/Hansen test of over-identifying restrictions (OID), which determines the validity of the instruments used in the estimations. In other words, we should expect that, at the 10% level of statistical significance, the p-value associated with the AR(1) test should be lower than 0.10, and the p-values related to the AR(2) test and to the OID test should be higher than 0.10. Finally, we need to meet the rule of thumb, whereby the number of instruments should not exceed the number of countries used in the regressions, otherwise the above-mentioned tests may lose power (Roodman, 2009). In the regressions based on the two-step system GMM approach, all variables except the population size variable, have been considered as endogenous. To meet the requirements of this estimator, including the rule of thumb described above, we have used in the regressions a maximum of 3 lags of the dependent variable as instruments, and a maximum of 4 lags of endogenous variables as instruments. All regressions are performed using both the total AfT flows variable as well as its three components.
Against this background, we perform the following estimations in the empirical analysis based on the two-step system GMM approach. First, we estimate the dynamic model (1), results of which are reported in Table 2. Second, we investigate whether the effect of AfT flows on services export share in LDCs versus NonLDCs. To that effect, we estimate another specification of model (1) in which we include a dummy variable (“LDC”) - which takes the value 1 for countries in the LDC category, and 0, otherwise - that is interacted with each of the AfT variables, namely total AfT flows and its three components. The outcomes of these estimations are reported in Table 3. Third, we now test hypothesis 1 set out above (i.e., whether the effect of AfT flows on countries’ services exports shares depends on their degree of integration into the world market of merchandises exports (also henceforth referred to as ‘merchandises export integration’). We test this hypothesis by estimating other variants of model (1) that include a variable denoted “GOODINT” - which represents for a given country and in a given year, the share of merchandises exports in world merchandises exports - as well as its interaction with the relevant “AfT” variable (either total AfT flows, or each of its components) in model (1). Results of these estimations are provided in Table 4. Fourth, and lastly, we test hypothesis 2, i.e., whether the effect of AfT flows on countries’ services export shares passes through FDI inflows. To perform this analysis, we estimate different other specifications of model (1) that contain the variable measuring FDI inflows, and denoted “FDICST”, which is interacted with the relevant “AfT” variable. Results of these estimations are reported in Table 5. Note that “FDICST” is the real FDI inflows (constant US$ 2010 prices). It has been computed by first multiplying the ratio of inward FDI flows to GDP by the real GDP (constant 2010 US$) so as to obtain values of FDI inflows in constant 2010 US$ (denoted “FDI”) (see for example Nagel et al. 2015 and Herzer, 2011 who have used the same procedure). Given that the variable “FDI” contains negative and positive values, and is additionally highly skewed, we have transformed it by means of the formula borrowed from Yeyati et al.(2007), which goes as follows: FDICST = signFDI*log(1+FDI), where FDI refers to the absolute value of the variable “FDI”. It is also noteworthy that in the regressions whose outcomes are presented in Tables 4 and 5, both “GOODINT” and “FDICST” variables, as well as their related interaction variables have been considered as endogenous because of the potential reverse causality from the dependent variable to each of these variables.