Identifying the robust economic, geographical and political determinants of FDI: an Extreme Bounds Analysis

Following standard practice in conducting EBA, we estimate Eq. (1). For each country i, and each specific regression jk (where \({j}\in [1,{M}], {k}\in [1,{K}]\) as specified below), we have:where \(\left( {\frac{\hbox {FDI}}{{Y}}} \right) _{i} \) is FDI inflows as a percentage of GDP into country i. The explanatory variables on the right-hand side are divided in three groups: The first is n standard (‘core’) explanatory variables that are included in every single regression (in addition to a constant) denoted \(\mathbf{X}_{{it}} =\left( {X_{1it} } \quad {X_{2it} } \quad \cdots \quad {X_{nit} } \right) \), where, \(\varvec{\upbeta }_{{jk}} =\left( {\beta _{1k} }\quad {\beta _{2k} } \quad \cdots \quad {\beta _{nk} } \quad \right) ^{{\prime }}\). Following Levine and Renelt (1992), we use a set of exactly three core variables, \(\mathbf{X}_{{it}} \), that are always kept in the equation. The second is \({I}_{{kit}} \), which is the kth variable of interest whose robustness we are testing and is a single variable selected from the set of variables \(\mathbf{Z}_{{it}} \), where the latter is a Kx1 vector containing all of the possible determinants of FDI that are not included in \(\mathbf{X}_{{it}}\). Following Leamer (1983), we consider all of the remaining variables in \({\mathbf{Z}}_{{it}} \) (one at a time and each in turn) as \({I}_{{kit}} \). \({\mathbf{Z}}_{{it}}\) is identified from a wide range of past studies as including potentially important candidate determinants (beyond \(\mathbf{X}_{{it}} )\) that need to be controlled for in FDI regressions. The third is \({\mathbf{Z}}_{{jit}}^{\varvec{k}}\), which is a 3x1 vector of exactly three additional control variables chosen from the pool of possible (non-core) explanatory variables, \({\mathbf{Z}}_{{it}} \), that do not include \({I}_{{kit}}\). For each k, all the possible combinations of the remaining \(K-1\) variables in the predetermined pool of variables \({\mathbf{Z}}_{{it}} \) is considered; there are \(M\left[ {=\frac{\left( {{K}-1} \right) !}{\left( {{K}-1-3} \right) !\times 3!}} \right] \) such combinations. Further, \(j = 1,2,{\ldots },M\), where j denotes the jth estimated combination of the variables: the jth model. The robustness of each variable of interest, \({I}_{{kit}}\), is tested while controlling for \(\mathbf{X}_{{it}} \) and all the possible combinations \({\mathbf{Z}}_{{jit}}^{\varvec{k}}\). Exactly three variables are included in \({\mathbf{Z}}_{{jit}}^{\varvec{k}}\), partly to follow Sala-i-Martin’s (1997) original methodology and also to avoid the perception of data mining or selective reporting of results.¹ There are M possible combinations for each of the \(k = 1,2,{\ldots },K\) variables of interest, giving a total of \(M\times K\) possible regressions. Finally, \({\upvarepsilon }_{{it}} \) is an error term. The aim is to investigate the effects on the statistical significance of \({\upgamma }_{{jk}}\), the coefficient on the kth variable of interest, when varying the combinations of three variables included in \({\mathbf{Z}}_{{jit}}^{\varvec{k}}\). The \(\left( {{j}=1,2,\ldots ,M} \right) \) estimated coefficients for each \({I}_{{kit}} \left( {\hat{\upgamma }_{{jk}} } \right) \) and \(\mathbf{X}_{{it}} \left( {\hat{\varvec{\upbeta }}_{{jk}} } \right) \)are recorded. The standard deviation of these M coefficient estimates is calculated for each \({I}_{{kit}} \) and is denoted as \(\hat{\upsigma }_{k}\).

Sala-i-Martin’s (1997) version of EBA, which we follow here, is based on the fraction of the cumulative distribution function (CDF) of \(\hat{\upgamma }_{{jk}}\) that lies to the right of zero.² If this fraction is sufficiently large (small) for a positive (negative) relationship, \({I}_{{kit}}\) is regarded as robust: specifically, if more than 90 % (less than 10 %) of the CDF for \(\hat{\upgamma } _{{jk}} \) is above zero, \({I}_{{kit}} \) is robust. We apply two variants of Sala-i-Martin’s (1997) EBA, being the normal and non-normal CDF methods. Sala-i-Martin’s method involves the calculation of a CDF for each variable of interest, \({I}_{{kit}} \), using the \(\left( {{j}=1,2,\ldots ,M} \right) \) estimated coefficients, \(\hat{\upgamma } _{{jk}} \), and estimated coefficient variances, \(\hat{\upsigma } _{{jk}}^2\). Using these values the mean of \(\hat{\upgamma } _{{jk}} \) is constructed as the average of each of the M \(\hat{\upgamma } _{{jk}} \), thus, \(\bar{\upgamma }_{k} =\sum _{{j}=1}^{M} \frac{1}{{M}}\hat{\upgamma } _{{jk}}\).³ Similarly, the average of the coefficient variances are \(\bar{\upsigma } _{k}^2 =\mathop \sum _{{j}=1}^{M} \frac{1}{{M}}\hat{\upsigma } _{{jk}}^2\).⁴ Assuming the \({\upgamma }_{{jk}} \) have a standard normal distribution across the M models, the CDF is evaluated at zero as \({\Phi }\left( {0|\bar{\upgamma } _{k} ,\bar{\upsigma } _{k}^2 } \right) \), where \({\Phi }\) denotes the cumulative density based on the standard normal distribution. Finally, the \(\hbox {CDF}(0)\) statistics indicates the larger of the areas under the density function either side of zero [hence \(0.5\le \hbox {CDF}(0)\le 1\)], that is:According to Sala-i-Martin (1997), if the \({\upgamma }_{{jk}}\) are not normally distributed across the M models for any particular k, \(\hbox {CDF}(0)\) can be calculated using the individual CDFs for each of the M regressions. The CDF for the jth regression is denoted as: \({F}_{j} \left( {\left. 0 \right| \hat{\gamma } _{{jk}},\hat{\sigma }_{{jk}}^2 } \right) \) where:The aggregate “non-normal” CDF, denoted \(\hbox {CDF}(0)^{{*}}\), is calculated as the average of the \(\left( {{j}=1,2,\ldots ,M} \right) \) individual CDFs (3), thus:Variables are regarded as robust when both CDFs are at least 0.90. The degree of robustness is assigned as follows: robust at the 1 % level when \(\hbox {CDF}(0)\ge 0.99\) or \(\hbox {CDF}(0)^{{*}}\ge 0.99\) (which is denoted with ***), robust at the 5 % level when either \(\hbox {CDF}\ge 0.95\) (**), robust at the 10 % level when both \(\hbox {CDF}\ge 0.90\) (*).⁵ A variable is regarded as a “fragile” determinant of FDI otherwise.⁶

We carry out two applications of the EBA procedure: firstly using only economic variables (results reported in Sect. 3.1) and then augmenting the dataset with the inclusion of political and geographic variables (Sect. 3.2). In the first of these applications, we test all possible variables considered in both \(\mathbf{X}_{{it}} \) and \({\mathbf{Z}}_{{it}} \) for robustness. In all applications, we report the results based on Sala-i-Martin’s (1997) method assuming both normal and non-normal CDFs (cf. 2 and 4 respectively).

We consider 58 potential economic, political and geographical explanatory factors.⁷ The definitions of the variables used are given in Table 1. Data were constructed from a number of sources, including World Development Indicators 2006 (World Bank 2006). The political and institutional variables are obtained from the International Country Risk Guide (ICRG), and we construct the geographical dummy variables. Our sample is an unbalanced annual panel dataset for 168 economies over the period 1970–2006.

Our first application of EBA that considers only economic determinants employ the fixed-effects estimator in all regressions; this is so as it is more likely to ensure consistent estimates than the random-effects estimator.⁸ However, the random-effects estimator is employed in our second EBA application that incorporates economic, geographical and political variables because some of these variables are perfectly collinear with the (cross-sectional) fixed-effects.

A potential problem for our estimates is endogeneity, which causes OLS estimators to be biased and inconsistent. We identify three potential determinants as being the most likely to be endogenously determined with FDI as the current account balance (% of GDP-CAB), GDP growth (GDPG) and per-capita GDP (GDPP).⁹ We therefore treat CAB, GDPG and GDPP as potentially endogenous in our EBA applications because the costs of incorrectly treating exogenous variables as endogenous are much lower than incorrectly assuming endogenous variables are exogenous. This means that these three variables are excluded from \(\mathbf{X}_{{it}} \) and \({\mathbf{Z}}_{{it}} \) in all EBA applications and are only considered as \({I}_{{kit}} \) variables. Hence, the only inference that could be affected by endogeneity bias is when these covariates are considered as the variable of interest.

The 30 potential economic determinants of FDI that we consider in our first EBA application are listed in the left-hand side of Table 1. The following three core variables, \(\mathbf{X}_{{it}} \), that are always kept in the equation are: openness (denoted OPEN), inflation (INFL), and tax on trade (TTRADE). These core variables are chosen because they have been shown to be robustly linked to FDI in previous empirical work (as well as in our initial experiments), and we do not expect them to be endogenous. All of the remaining 27 economic determinants are considered as the variable of interest, \({I}_{{kit}} \); however, only 24 of these are included in \({\mathbf{Z}}_{{jit}}^{\varvec{k}}\) because we are seeking to minimise the impact of any endogeneity bias that the current account balance (CAB), GDP growth (GDPG) and per-capita GDP (GDPP) variables may cause.¹⁰

Tables 2, 3, 4, 5 summarise the results of our first EBA application. The first column reports the variable of interest under consideration. For each \({I}_{{kit}} \) variable four sets of EBA statistics are reported: one set for the \({I}_{{kit}} \) variable (reported in Table 5) and one set for each of the 3 core variables: OPEN (Table 2), INFL (Table 3) and TTRADE (Table 4).¹¹ Some models cannot be estimated due to insufficient observations, and this causes some variation in the number of regressions run for the different \({I}_{{kit}}\).

The column headed “AVG coeft” gives the variable’s coefficient averaged over the number of regressions used in the EBA application. Sala-i-Martin’s (1997) non-normal CDF [denoted \(\hbox {CDF}(0)^{{*}}\)] and normal CDF [\(\hbox {CDF}(0)\)] statistics are also reported in the tables. Bold emphasis indicates that a variable is robust based upon Sala-i-Martin’s criteria. For a variable to be robust, it must have a CDF of at least 0.90 according to both normal and non-normal criteria (the normal and non-normal CDF broadly yield the same inference); otherwise, the variable is said to be fragile.

Table 2 indicates that the core variable, Trade openness index (OPEN), is robust in 26 out of 27 sets of EBA results (the exception is when TIMEB is the variable of interest). This result is consistent with many previous studies that found openness towards trade to be a significant determinant of FDI (e.g. Chakrabarti 2001; Moosa and Cardak 2006). In all 27 cases, OPEN has an average coefficient sign (“AVG coeft”) that is positive which is consistent with theoretical expectations. From Table 3, the INFL core variable is robust in only one (RATIOT) of the 27 EBA sets and is a fragile determinant for the remaining 26 \({I}_{{kit}}\). The TTRADE core variable is robust in only one (CGD) of the 27 EBA sets and is a fragile determinant otherwise, see Table 4. This is considered as strong evidence against TTRADE and INFL being robust determinants of FDI.

From Table 5 we see that eight non-core variables are unambiguously robust determinants of FDI because both of their CDFs exceed 0.90. These are current account balance (CAB), GDP growth rate (GDPG), GDP per capita (GDPP), highest marginal corporate tax rate (HMTAXCOR), outgoing FDI (FDIO), tertiary and secondary school enrolment (RATIOT and RATIOS, resp.) and Government final expenditure (GFE). The negative and robust coefficient of HMTAXCOR suggests that high corporate taxes in the host country will have a robust negative effect on FDI, in line with the finding of Becker and Fuest (2012). Government expenditure as a proportion of GDP (GFE) is also robust and has, on average, a negative sign, validating the critics of government (see e.g. Mitchell 2005; Sinn 1995). The tertiary enrolment ratio (RATIOT) and secondary enrolment ratio (RATIOS) are both found to be robust determinants of FDI with generally positive coefficients, which are consistent with previous literature and implies that education attracts FDI.¹² FDI outflows (FDIO) is another robust determinant of FDI inflows. This could be because the multinational corporations (MNCs) of developing countries may be both the recipients of incoming FDI and originators of outgoing FDI; thus, FDI and FDIO may be positively correlated. Both GDPG and GDPP robustly and positively affect FDI. As these are measures of future prospects and market size and demand, this finding is consistent with theoretical expectations; however, we cannot rule out bi-directional causality. Additionally, we find that the current account balance (CAB) affects FDI, the negative sign is consistent with theory; however, bi-directional or even reverse causality cannot be ruled out. In sum, all of the nine robust variables (OPEN, GFE, FDIO, RATIOS, RATIOT, CAB, GDPG, GDPP, HMTAXCOR) have theoretically plausible (average) coefficient signs. However, we treat the finding of robustness for the three potentially endogenous variables CAB, GDPG and GDPP with caution and hesitate to conclude that our results offer strong support for it.

Apart from these nine variables, all of the other variables in our first EBA application are fragile. Comparing our findings with previous applications of EBA to FDI provides interesting insights. Moosa and Cardak (2006) found telephone mainlines to be robust, whereas we find it to be fragile in this application; however, we do find it to be a robust determinant in the next application when we consider both geopolitical and economic variables. Further, Moosa and Cardak found GDP growth and tertiary education enrolments to be fragile, while we find these variables to be robust. Chakrabarti (2001) found openness to be robust as we do, though not GDP growth.

In our second EBA application, we include OPEN, GFE and RATIOS as our core variables following the results of our first EBA. OPEN is chosen because it is the only core variable from our first EBA application that is robust. Since the other two core variables (INFL and TTRADE) are not robust in our first EBA application, we seek two different core variables; those should, firstly, be robust with an average coefficient sign that is consistent with theoretical expectations in the first EBA application and that have the highest value for \(\hbox {CDF}(0)^{{*}}\); secondly, they must not be amongst the 3 potentially endogenous variables. The three variables with the highest values for \(\hbox {CDF}(0)^{{*}}\) are GFE \(\left( {\hbox {CDF}(0)^{{*}}=1.00} \right) \), RATIOS (0.95) and CAB (0.95). Since we regard CAB as potentially endogenous, we select the other 2 as core variables, along with OPEN, to be employed in our second EBA application.

We add 28 geographical and political variables (described on the right-hand side of Table 1) to the economic variables to be considered in the second EBA application, allowing us to test the robustness of an extended set of variables. The geopolitical variables are not included in the core set of variables, \(\mathbf{X}_{{it}} \), or the set of three \({\mathbf{Z}}_{{jit}}^{\varvec{k}} \) variables (to help avoid multicollinearity); however, they are all considered (in turn) as the variable of interest, \({I}_{{kit}}\). All of the economic variables (except the three core variables) are considered (in turn) as \({I}_{{kit}} \) and in \({\mathbf{Z}}_{{jit}}^{\varvec{k}}\) (except for the potentially endogenous variables, CAB, GDPG and GDPP, and the three core variables).¹³ The focus in this second application is to determine whether country specific institutions (such as democracy, rule of law, corruption, bureaucracy, ethnic and international conflict and type of political regime),¹⁴ cultural factors (languages) or geographical locations (number of boundaries, coastal location, abundance of natural resources, proximity to particular regions) can influence FDI. Many geographical and political/institutional factors have been conclusively linked to economic growth (e.g., Durlauf et al. 2005) and remain active areas of research. The results of our second EBA application are reported in Table 6.

Ten of the 28 geopolitical variables are considered as robust determinants of FDI as both of their CDFs are at least 0.90. These include the dummies for: countries in the South Asia region (SA), countries in the East Asia and Pacific region (EAP), countries with more than 3 boundaries (GTBUN), countries that are not land-locked (LANDUNLOCKED), Spanish (SPN)- and Arabic (ARB)-speaking countries as well as nations with greater democratic accountability (DEMO). These seven determinants are all generally positively correlated with FDI inflows. The other three robust geopolitical variables are dummies for countries experiencing low international and internal conflict (CONFLICTINT) and economies with an abundance of the natural resources: oil (OILDUMMY) and gas (GASDUMMY).¹⁵ These three determinants are all generally negatively correlated with FDI inflows.

DEMO is a robust determinant with a generally positive coefficient sign, which is expected as democracy increases transparency and reduces arbitrariness and red tape (Jensen 2008 and Li 2009).¹⁶ The internal and external conflict variable (CONFLICTINT) is robust with a generally negative coefficient sign. This is consistent with a priori expectations as less conflict reduces incertitude amongst potential investors, which raises FDI. Hence, our results support the notion that an increase in institutional quality (as indicated by greater democracy and lower conflict) would strengthen incoming FDI; these results are consistent with previous analyses (e.g. Globerman and Shapiro 2002; Sachs 2003). It is noteworthy that various other variables whose relevance has been highlighted in the theory of growth, such as rule of law, corruption, bureaucracy, ethnic conflict or type of political regime, are fragile determinants of FDI.

Our results also suggest that language is an important factor in attracting FDI. We found that countries where Arabic and Spanish are the main language have higher incoming FDI ceteris paribus. This result may be driven by the higher incoming FDI into countries such as the Middle East and Latin America, as opposed to others in Africa and elsewhere in which English or French are the official languages; dummies representing countries where English and French are the main language are found to be fragile. International languages such as English and to some extent French may play a role in attracting FDI, but they are often spoken by much of the population in countries where they are not the main language and this may help explain why countries where English and French are the main languages do not receive any significant increases in incoming FDI. We also find that coastal countries tend to attract more FDI: the dummy “LANDUNLOCKED”, indicating countries that are not landlocked, is a robust determinant with a generally positive sign. This is consistent with findings in growth empirics (Easterly and Levine 2003). Furthermore, countries with more than three boundaries attract more FDI than those with fewer boundaries given the robust and generally positive coefficient. This is also in the spirit of the previous finding (the “landunlocked” feature): a country with more neighbours has more freedom to trade and, hence, better prospects for incoming FDI. While “landlockedness” has been emphasised in the past as a factor affecting growth and FDI, the finding that the numbers of borders affects FDI is, we believe, novel.

Natural resource abundance in the form of oil and gas (OILDUMMY and GASDUMMY, respectively) are both found to be robust determinants of FDI with generally negative coefficient signs. This is a kind of “Dutch disease”, akin to that highlighted by Sachs and Warner (1995) in relation to growth; see also Tietenberg and Lewis (2006). This reasoning will of course not apply to specifically resource-seeking firms, which would naturally be attracted by resource abundance; this would explain the inflows of FDI into the Arab Gulf and African countries. All other geopolitical variables exert only a fragile influence on FDI.

From Table 6 we see that eight non-core economic variables are robust determinants of FDI: CAB, GDPG, GDPP, CGD, FDIO, INTERNET, RATIOT and TEL. These findings for economic variables are similar to those in Table 5 in that FDIO, RATIOT, CAB, GDPG and GDPP are found to be robust in both of our EBA applications—broadly confirming the robustness of these results. For economic variables, the average coefficient signs are the same in Tables 6 and 5 except for RATIOT which has a generally negative coefficient sign in Table 6; this change in coefficient sign between the two EBA applications may be due to RATIOS being a core variable in the second application and not the first. Table 6 suggests three additional robust economic variables, which are central government debt (CGD), internet use (INTERNET) and telephone mainline use (TEL). CGD appears as robust with a generally negative coefficient: this is expected, as debt may have a number of adverse consequences, such as inducing higher interest rates and raising default risk. The latter two capture communication facilities. As expected an increase in internet and telephone use increases FDI inflows, as indicated by the generally positive coefficient signs for these variables. All of the other economic variables in our second EBA application are fragile.