Pull Factors of FDI: A Cross-Country Analysis of Advanced and Developing Countries

Abstract

In a cross-country bilateral FDI flow setup, we examine macroeconomic determinants of FDI flow to advanced economies (AE) from AE, to AE from developing economies (DE), to DE from AE and to DE from DE. It is observed that determinants vary significantly across these broad groups. Further, we construct composite index based on these macroeconomic determinants and rank the countries within these broad groups of FDI flow to understand macroeconomic enabler in the host country which attract FDI. We also propose a new methodology to circumvent multicollinearity issue which arises as selected determinants of FDI are found to be interrelated.

An earlier version of this paper was presented at XI Annual Conference of Knowledge Forum held at the IIT Madras, Chennai, 3–5 December 2016. Authors are grateful to participants at the conference for their helpful comments. Views expressed in this paper are those of the authors’ alone and not of the institution to which they belong.

Appendices

Appendix 1

Variables and Their Definitions

• i. d_RD: difference of Researchers in R&D (per million people) at host and home

• ii. d_Trademark: difference of Trademark applications: total at host and home

• iii. home_ict_imp: ICT goods imports (% total goods imports) by home country

• iv. home_ict_exp: ICT goods exports (% of total goods exports) by home country

• v. ict_exportimport: sum of ICT goods exports (% of total goods exports) by host and ICT goods imports (% total goods imports) by home country

• vi. ict_importexport: sum of ICT goods imports (% total goods imports) by Host and ICT goods exports (% of total goods exports) by Home and

• vii. d_edu_exp: difference of Adjusted savings: education expenditure (% of GNI) as technology and human capital factors of FDI inflow (host–home) at host and home

• viii. d_Co2: CO₂ emission (metric tons per capita) difference between host and home

• ix. d_mfg_imp: difference of Manufactures imports (% of merchandise imports) of host and home.

• x. d_elec: difference of Electric power consumption (kWh per capita) at host and home

• xi. d_Air: difference of Air transport, passengers carried/population at host and home

• xii. d_ATM: difference of Automated teller machines (ATMs) (per 100,000 adults) of host and home

• xiii. yoy_exch_rate: appreciation/depreciation of bilateral level of the exchange rate; Exch rate (annual freq) = 1 unit of host currency = ‘x’ unit of home currency; yoy exch rate = (x _t /x _t−1) * 100 − 100

• xiv. exch_rate_cv: coefficient of variation of exchange rates exch = std dev (x _t …x _t−12)/average (x _t …x _t−12)

• xv. d_popuabove65: difference (Host–home) of Population ages below 65 (% of total)

• xvi. d_real_int: difference of Real interest rate (%) at host and home

• xvii. d_sal: difference of compensation of employees (% of expense) at host and home

• xviii. d_GDP: difference of GDP growth (annual %) at host and home

• xix. d_infl: difference of Inflation in term of GDP deflator (annual %) at host and home

• xx. d_sp: difference o fS&P Global Equity Indices (annual % change) at host and home

• xxi. host_gdp_avg2010: log of host GDP

• xxii. d_distance: geographical distance between capital of two country

• xxiii. d_custom: difference of Customs and other import duties (% of tax revenue) at host and home

• xxiv. d_imp_and_exp: sum of Cost to import (US$ per container) at host and Cost to exports (US$ per container) by home country

• xxv. host_home_pol_stability: difference (host–home) of Political Stability and Absence of Violence/Terrorism: Estimate,

• xxvi. d_cont_corruption: difference (host–home) of control of corruption estimate.

• xxvii. d_law: difference (host–home) of rule of law

• xxviii. d_regul: difference (host–home) of regulatory quality

• xxix. d_fisc_def: difference (host–home) of Central government debt, total (% of GDP)

• xxx. d_cur_act: difference (host–home) of Current account balance (% of GDP)

Appendix 2

Multicollinearity

Multicollinearity refers to high inter-dependence among the explanatory variables in a regression set up. Particularly when the economic indicators are considered, multicollinearity is natural outcome of economic interaction. Such high correlation among the regressors creates the problem of near singularity of the residual variance-covariance matrix and thereby inflates the parameter estimates leading to unrealistic interpretation. For instance, the parameter estimates $\widehat{\beta }$ in the following regression set up is given by

$$\beta = \left( {X^{\prime } X} \right)^{ - 1} X^{\prime } \underbrace {Y}_{{}}$$

where

$$\begin{aligned} & \left( 1 \right)\;Y_{t} = \beta_{0} + \beta_{1} X_{1t} + \beta_{2} X_{2t} + \cdots + \beta_{n} X_{nt} + \in_{t} \\ & \Rightarrow \left( 2 \right)\;Y_{t} = \beta X_{t} + \in_{t} \\ \end{aligned}$$

Thus the presence of strong dependence among variables is not desirable for robust estimation of model parameters. Now considering (1), we get

$$\begin{aligned} & \left( 1 \right)\;Y_{t} = \beta_{0} + \beta_{1} X_{1t} + \beta_{2} X_{2t} + \cdots + \beta_{n} X_{nt} + \in_{t} \\ & \Rightarrow {\text{Var}}\left( {\widehat{\beta }_{i} } \right) =\underbrace{ \left[ {\frac{{\sigma^{2} }}{{\mathop \sum \nolimits_{t = 1}^{T} \left( {X_{it} - \overline{X}_{i} } \right)^{2} }}} \right]} \times \underbrace {{\left[ {\frac{1}{{1 - R_{i}^{2} }}} \right]}}_{{}} \\ \end{aligned}$$

Here the variance of OLS estimate of $\beta_{i}$ is proportional to $R_{i}^{2}$ where $R_{i}^{2}$ is the R-square value of $x_{i}$ when $v$ is regressed over other regressor. The first part of the variance expression depends upon error variance and sample variance of $X_{i}$ and the second part is proportional to $R_{i}^{2}$. Thus higher $R_{i}^{2}$ is likely to inflate the variance of $\widehat{\beta }_{i}$. This concept has been extended to define VIF for ith variable which can be used for detecting multicollinearity. Higher value of VIF indicates presence of multicollinearity among the regressor. VIF is defined as

$${\text{VIF}}_{i} = \frac{1}{{\left( {1 - R_{i}^{2} } \right)}}$$

In case the variable $x_{k}$ for some $k \in \left[ {1,n} \right]$ is highly correlated with other regressor, then $R_{k}^{2}$ is likely to be higher and so will be $\frac{1}{{\left( {1 - R_{k}^{2} } \right)}}$. Generally, for a variable VIF value exceeding 4 indicates high multicollinearity and requires further investigation.

Appendix 3

Principal Component Analysis (PCA)

PCA is a powerful data analysing tool often used to identify patterns in data, and highlight their similarities and differences, especially, in data of high dimension where graphical representation of data is not available. Once patterns in the data are found, data can be compressed by reducing the number of dimensions, without loss of much information. PCA transform the variance-covariance matrix in term of eigenvector and eigenvector with the highest eigenvalue is the principle component of the data set. PCA give us the original data solely in terms of the eigenvectors. Once eigenvectors are found from the covariance matrix, the next step is to order them by eigenvalue, highest to lowest so that one can decide to ignore the components of lesser significance and the final data set will have fewer dimensions than the original.

The space spanned by the observed variables can be widely sparse in nature which can be measured by the variance-covariance matrix of the observed data. Due to inter-dependence among the variables, the covariance terms are expected to be non-zero. PCA tries to transform the original variables into a new set of orthogonal variables such the variability of the observed data can be explained by the transformed data. In vector space, this is equivalent to transform the observed data (X) into new set of variables (Y) such that

$$\left( 1 \right)\;P{:}\; R^{N} \to R^{N} { \ni }Y = P.X$$

Geometrically P is a rotation of X into space of Y in such a manner that Y is orthogonal, i.e. $Y_{i}^{\prime } *Y_{j} = 0$ for $i \ne j$. Thus given orthogonality of P, the row vectors of P (i.e. $P_{i}$, for i = 1(1)N) is the basis of space spanned by X. Now

$$\begin{aligned} S_{Y} & = \left( {\frac{1}{N - 1}} \right)YY^{T} = \left( {\frac{1}{N - 1}} \right)\left( {PX} \right)\left( {PX} \right)^{T} \\ & = \left( {\frac{1}{N - 1}} \right)PXX^{T} P^{T} = \left( {\frac{1}{N - 1}} \right)PAP^{T} = \left( {\frac{1}{N - 1}} \right)PEDE^{T} P^{T} \\ & = \left( {\frac{1}{N - 1}} \right)\left[ {PE} \right]D\left[ {PE} \right]^{T} = \left( {\frac{1}{N - 1}} \right)ZDZ^{T} \\ \end{aligned}$$

We select P in such a manner that $P_{i}$, is the eigenvector of $XX^{T}$. Now

$$S_{Y} = \left( {\frac{1}{N - 1}} \right)\left( {PP^{ - 1} } \right)D\left( {PP^{ - 1} } \right)^{ - 1} = \left( {\frac{1}{N - 1}} \right)D$$

Thus the eigenvectors are principal components and the proportion of variability explained by each principal component is determined by the eigenvalue share of corresponding eigenvector. The application of PCA can be many folds from dimension reduction to orthogonal transformation.