A project submitted to the graduate school in partial fulfillment of the requirements for the degree

The association between the change in foreign direct investment net inflows as the percentage of gross domestic product and a country’s growth of real gross domestic product is used as t

NATIONAL REALIZED ABSORPTIVE CAPACITY:

COUNTRY COMPARISONS

BY LINH THI THUY DO, B.E., M.S

A project submitted to the Graduate School

in partial fulfillment of the requirements

for the degree DOCTOR OF ECONOMIC DEVELOPMENT

Major Subject: ECONOMIC RESEARCH Minor Subject: APPLIED STATISTICS

NEW MEXICO STATE UNIVERSITY LAS CRUCES, NEW MEXICO NOVEMBER 2017

“National Realized Absorptive Capacity: Country Comparisons,” a project prepared

by Linh Thi Thuy Do in partial fulfillment of the requirements for the degree Doctor

of Economic Development, has been approved and accepted by the following:

iii

ACKNOWLEDGEMENTS

At first, I would like to emphasize my deepest gratitude to Dr Christopher A Erickson, who is my supervisor, for his valuable comments and guidance during the process of this study Without his help, I cannot complete this project

I am also grateful to my friends for the constructive discussions and useful suggestions on any inaccuracy All of them make a great contribution to the completion of this project

Finally, the sincere gratitude and appreciation go to my family for their encouragement and mental support

VITA February 24, 1987 Born in Thai Nguyen City, Vietnam

2005 Graduated from Thai Nguyen Specialized Upper Secondary

School, Vietnam

2009 Graduated from Foreign Trade University

Hanoi, Vietnam

2012 Graduated from University of Essex,

Colchester, the United Kingdom

2014-2017 Research Assistant and Teaching Assistant

Department of Economics, Applied Statistics, and International Business, College of Business

New Mexico State University

Field of Study

Major Field: Economic Development

Minor Field: Applied Statistics

v

ABSTRACT

This project investigates the association between the level of economic activity and national realized absorptive capacity Absorptive capacity is a dynamic process that has four dimensions: acquisition, assimilation, transformation, and exploitation of the external knowledge From macroeconomic perspectives, national absorptive capacity can be defined as the process by which a nation internalizes external resources for its economic growth The literature shows that many variables, including development level, drive the dynamics of national absorptive capacity We distinguish between a nation’s potential and realized absorptive capacity, then use data from the World Bank to empirically identify if there are significant differences in realized absorptive capacity among nations grouped into the four income categories defined by the World Bank The association between the change in foreign direct investment net inflows as the percentage of gross domestic product and a country’s growth of real gross domestic product is used as the indicator of the nation’s realized absorptive capacity We estimate fixed effects regressions, which show that countries are different in their national realized absorptive capacity The regression results reject the null hypothesis that countries at higher level of development have higher national realized absorptive capacity We further find evidence that investment in infrastructure has positive associations with national realized absorptive capacity

Keywords: national realized absorptive capacity, foreign direct investment, economic growth

TABLE OF CONTENTS

List of Tables vii

List of Figures viii

1 Introduction 1

2 Methodology 7

Step 1 – Estimating National Realized Absorptive Capacity 7

Step 2 – Identifying Determinants of National Realized Absorptive Capacity 11

3 Data and Estimation Techniques 13

4 Results 16

Tests of stationarity, cointegration, and Granger causality 16

Step 1- Estimates of National Realized Absorptive Capacity 20

Step 2 – Determinants of National Realized Absorptive Capacity 34

5 Conclusions and Policy Implications 41

Appendix 1 – List of countries included in the sample 45

Appendix 2 –Descriptive Statistics 46

Appendix 3: Regression results for models in Table 3 48

Appendix 4: Estimates of national realized absorptive capacity 53

Appendix 5: STATA commands 57

References 64

vii

LIST OF TABLES

Table 1: Unit root tests for the %∆FDI and the %∆GDP 17

Table 2: Results of unit root tests for other explanatory variables 18

Table 3: Regression results for models with the dependent variable %∆GDP 26

Table 4: Variance Inflation Factor of variables as in the regressions of Table 3 30

Table 5: Regression results using the averaging method 37

Table 6: Variance Inflation Factor of variables as in the regressions of Table 5 39

LIST OF FIGURES

Figure 1: National potential and realized absorptive capacity 2 Figure 2: Graph of real FDI versus real GDP for the sample of 135 countries 14 Figure 3: Visualized Regression Results of Model I from Step 1 32

1

1 Introduction

Absorptive capacity is defined as “a set of organizational routines and processes by which firms acquire, assimilate, transform and exploit knowledge to produce a dynamic organizational capability” (Zahra and George, 2002, pp 186) Absorptive capacity is a dynamic process involving the internalization of external knowledge (i.e., technology) that has four dimensions: acquisition, assimilation, transformation, and exploitation The term “absorptive capacity” can be studied at individual, group, firm, sector, and national levels By aggregating upwards from the firm level, Criscuolo and Narula (2008) define national absorptive capacity as the process of internalizing external knowledge to grow an economy

Zahra and George (2002) distinguish between a firm’s potential and realized capacity Potential capacity involves knowledge acquisition and assimilation; while realized capacity involves knowledge transformation and exploitation These four capabilities are dynamic and influence the firm’s ability to utilize the knowledge for organizational development In particular, realized capacity allows the firm to construct its competitive advantage; while potential capacity provides the firm with flexibility to respond operationally to changing market conditions Ben-Oz and Greve (2015) use the survey data of 252 decision makers from 129 Israeli early-stage high-tech organizations and find that realized absorptive capacity is affected by short-term goals and has short-term effects, while potential absorptive capacity is mainly affected by long-term goals and has long-term effects

The concept of a firm’s potential and realized capacity can be generalized into

a nation’s potential and realized absorptive capacity National potential absorptive capacity is the maximum of a nation’s capability to acquire and assimilate knowledge during a particular period of time National realized absorptive capacity is the nation’s current ability to transform and exploit the knowledge that has been acquired and assimilated up to the current time Visually, at a point in time, a national potential absorptive capacity is the external bound, while its realized absorptive capacity lies within the bound (Figure 1) The larger the green circle, the higher the efficiency ratio – the ratio of realized absorptive capacity to potential absorptive capacity (Zahra and George, 2002)

Figure 1: National potential and realized absorptive capacity National potential absorptive

capacity is the maximum of a nation’s capability to acquire and assimilate the knowledge during a particular period of time National realized absorptive capacity is the nation’s current ability to transform and exploit the knowledge that has been acquired and assimilated up to the current time The ratio of realized absorptive capacity to potential absorptive capacity is called the efficiency ratio (Zahra and George, 2002) The larger the realized absorptive capacity, the higher the efficiency ratio

Source: Authors

3

National absorptive capacity is abstract and difficult to measure, but it does exist in that two countries of similar characteristics, exposed to similar external knowledge, can experience different paces of growth Then the country that grows faster is considered to have higher absorptive capacity Whether at the individual, firm, sector or national level, exposure to external resources is a necessary condition

to exploit the new resources The sufficient condition to fully exploit the external resources is absorptive capacity as measured by the four dimensions of acquisition, assimilation, transformation, and exploitation

National absorptive capacity can be built from the bottom up in that national absorptive capacity depends significantly on the individuals’, firms’ and sectors’ absorptive capacity In particular, individuals’ absorptive capacity influences an organization’s absorptive capacity (Cohen and Levinthal, 1990), individual firms’ absorptive capacity determines a sector’s absorptive capacity (Cieślik and Hagemejer, 2014), and sectors’ absorptive capacity establishes national absorptive capacity (Criscuolo and Narula, 2008) Therefore, any factor that has impacts on absorptive capacity of individuals, firms, or sectors potentially influences national absorptive capacity And indeed, the literature shows similar factors that affect absorptive capacity at three levels – firm, sector, and nation For example, important determinants of adsorptive capacity at all levels include human capital; investment in research and development (R&D); the similarity of the partners’ knowledge; and presence and activities of multinational enterprises (MNEs) (Criscuolo and Narula, 2008; Castellacci and Natera, 2013; Cieślik and Hagemejer, 2014; Un, 2017)

We argue that difference in national absorptive capacity can help explain why some countries are successful in exploiting foreign direct investment (FDI) for economic growth, while others are not For less developed countries that lack access

to capital and technology, external resources like FDI, are especially important for economic growth But the more important thing is whether those countries have enough capacity to absorb the spillover effects of FDI Focusing on attracting more FDI without improving national absorptive capacity could lead to the waste of limited resources and could even be counterproductive to the extent that FDI crowds out domestic investment In this sense, understanding the factors influencing national absorptive capacity might help countries have proper policies to better exploit FDI

National absorptive capacity is also influenced by macroeconomic factors According to Criscuolo and Narula (2008), national absorptive capacity is determined

by international technological situation, institutions and economic actors, and the efficient use of markets In particular, Criscuolo and Narula state that both developing and developed countries increasingly rely on FDI as an explicit means to develop the nation’s technological and economic competitiveness which play key roles in improving national absorptive capacity over the stages of technological development Castellacci and Natera (2013) summarizes previous literature to find that beside human capital, the dynamics of a nation’s absorptive capacity are driven by five other dimensions, including income and development level, international trade,

1 In our study, we also add these variables in the model of the determinants of national realized absorptive capacity In particular, the countries are grouped into four development levels based on their income (GNI per capita) levels Net exports variable measures a country’s international trade Fixed telephone subscriptions indicate a country’s infrastructures Bribery incidence shows the quality of institutions and governance system GINI index presents economic inequality

countries at higher level of development have higher national realized absorptive capacity It will turn out that our hypothesis is not confirmed

This study contributes to the literature in two aspects First, this research quantifies national absorptive capacity, particularly national realized absorptive capacity, as the correlation between the external resources as measured by FDI net inflows and economic growth A set of national data is used to discover if there are significant differences in realized absorptive capacity between nations of different development levels Then, we try to determine the factors affecting national realized absorptive capacity, taking into account the distinction in development level between countries We do this using a panel data collected by the World Bank for a set of 135 countries over the period from 1990 to 2015 as the evidence of our arguments

The remainder of the research is structured as followed Section 2 is the methodology, which clarifies the theoretical background and the two steps used in our analysis Section 3 presents the data and estimation techniques used for the analyzing process of the specific data Section 4 discusses the results and limitations of analyses Section 5 concludes and suggests some appropriate policies to improve national realized absorptive capacity

7

2 Methodology

The analysis on the association between the development level and national realized absorptive capacity consists of two steps, each step using different empirical models and different econometric techniques In step 1, the absorptive capacity is estimated In particular, the estimated coefficient on the net inflows of foreign direct investment (FDI) from a regression relating gross domestic product (GDP) growth to FDI is used as a measure of the nation’s realized absorptive capacity A separate absorptive capacity is estimated for each country by using an interactive term between the country indicator variable and absorptive capacity In step 2, the coefficient on FDI estimated in Step 1 is used as a measure of national realized absorptive capacity in that the coefficients are modeled as dependent on selected socio-economic variables, including the focus variable – the development level

Step 1 – Estimating National Realized Absorptive Capacity

Economic Growth Model and Economic Growth Rate

We model the macro economy using the income identity A nation’s GDP at time t is expressed as the summation of four components as follows:

𝐺𝐷𝑃𝑖𝑡 = 𝐼𝑖𝑡 + 𝐶𝑖𝑡+ 𝐺𝑖𝑡 + 𝑁𝑋𝑖𝑡 (1) where 𝐺𝐷𝑃𝑖𝑡 is gross domestic product in time period t for country i, I is capital

investment, C is consumption, G is government spending and NX is the excess of exports over imports All variables in equation (1) as well as in the following are adjusted for purchasing power parity

National capital investment can be decomposed into domestic capital investment and foreign investment If FDI is foreign direct investment, FII is foreign indirect investment, and DI is domestic investment, then 𝐼𝑖𝑡 = 𝐹𝐷𝐼𝑖𝑡 + 𝐷𝐼𝑖𝑡+ 𝐹𝐼𝐼𝑖𝑡, and equation (1) can be rewritten as:

𝐺𝐷𝑃𝑖𝑡 = 𝐹𝐷𝐼𝑖𝑡+ 𝐷𝐼𝑖𝑡 + 𝐹𝐼𝐼𝑖𝑡+ 𝐶𝑖𝑡+ 𝐺𝑖𝑡 + 𝑁𝑋𝑖𝑡 (2)

Knowing that gross domestic product in previous time is:

𝐺𝐷𝑃𝑖,𝑡−1 = 𝐹𝐷𝐼𝑖,𝑡−1+ 𝐷𝐼𝑖,𝑡−1+ 𝐹𝐼𝐼𝑖,𝑡−1+ 𝐶𝑖,𝑡−1+ 𝐺𝑖,𝑡−1+ 𝑁𝑋𝑖,𝑡−1 (3) Subtracting (2) from (3), then dividing both sides by 𝐺𝐷𝑃𝑖,𝑡−1, results in the following equation:

Crowding Effects and Countries’ Economic Growth

FDI is a source of capital, thus, promotes economic growth directly through capital deepening But in addition to this direct influence, FDI may also influence growth through its effect on domestic capital There can be both crowding-in and crowding-out effects, by which FDI either encourages or discourages domestic

9

investment (Jan and Vladimr, 2002; Wu, Sun, and Li, 2012) For example, Gallagher and Zarsky (2007) state four ways by which foreign direct investment can crowd in domestic investment: Backward linkages—multinational enterprises create higher demand for specialized inputs produced by local firms; Forward linkages—products

of multinational enterprises enhance inputs to other domestic industries; Knowledge spillovers—the presence of multinational enterprises generates productivity and technology spillovers to the domestic firms so that they become more competitive; Multiplier effects—multinational enterprises create more jobs and increase local spending on domestically produced goods and services On the other hand, Gallagher and Zarsky (2007) argue that crowding-out effect can occur: Competition—domestic firms are out competed by more efficient foreign companies leading to the loss of local market shares; Borrowing costs—higher interest rates due to foreign investors’ borrowings in domestic financial markets discourage domestic firms from borrowing; Labor market effects – multinational enterprises start “hiring domestic entrepreneurs and skilled workers away from domestic firms (pp 25).”

Whether foreign direct investment inflows crowd in or crowd out domestic investment is an empirical question that is among the issues addressed in this project Agosin and Machado (2005) answer the question by considering the contribution of foreign direct investment to overall investment, which comprises both foreign and domestic investment If a-dollar-increase in foreign direct investment leads to one dollar increase in total investment, foreign direct investment’s influence is neutral If foreign direct investment crowds in domestic investment, then the increase in total

investment should be more than the increase in foreign direct investment If the increase in total investment is smaller than the increase in foreign direct investment, crowding-out effect occurs Another approach is carried out by Acar, Eris and Tekce (2012) who consider the direct impact of foreign direct investment on domestic investment In a regression model where domestic investment is the dependent variable and foreign direct investment is an explanatory variable, a negative coefficient shows crowding-out effect while a positive coefficient represents crowding-in effect In particular, the share of domestic investment in national GDP could be written as a function of the share of foreign direct investment in national GDP as follows:

𝐷𝐼𝑖𝑡− 𝐷𝐼𝑖,𝑡−1𝐺𝐷𝑃𝑖,𝑡−1 = 𝛾0+ 𝛾1𝐹𝐷𝐼𝑖𝑡 − 𝐹𝐷𝐼𝑖,𝑡−1

𝐺𝐷𝑃𝑖,𝑡−1 + 𝛾2𝑍𝑖𝑡+ 𝜀2𝑖𝑡 (6) where DI and FDI have the same meanings as stated in equation (2), Z is the vector of controls, and 𝜀2 is the error terms Note that 𝛾1 can be either positive or negative depending on the net effect of crowding in and crowding out

Substituting (6) into (5) leads to the new model:

%𝐺𝐷𝑃 𝑖𝑡 = 𝛼 0 + 𝛼 1

𝐹𝐷𝐼𝑖𝑡− 𝐹𝐷𝐼𝑖,𝑡−1𝐺𝐷𝑃𝑖,𝑡−1 + 𝛼 2 [𝛾 0 + 𝛾 1

𝐹𝐷𝐼𝑖𝑡− 𝐹𝐷𝐼𝑖,𝑡−1𝐺𝐷𝑃𝑖,𝑡−1 + 𝛾 2 𝑍 𝑖𝑡 + 𝜀 2𝑖𝑡 ] + 𝛼 3 𝑋 𝑖𝑡 + 𝜀 1𝑖𝑡

Step 2 – Identifying Determinants of National Realized Absorptive Capacity

Step 2 seeks to discover why national realized absorptive capacity is different across countries, i.e why foreign direct investment may have different effects on GDP growth in different countries The literature has identified some possible

2 Dummy variable trap occurs when the value of one dummy can be predicted from values of other dummies (perfect multicollinearity)

determinants According to Castellacci and Natera (2013), the dynamics of a nation’s absorptive capacity are theoretically driven by six dimensions: income and development level, international trade (or the openness of the national system), human capital, infrastructure, quality of institutions and governance system, social cohesion and economic inequality The authors use the CANA data (Castellacci and Natera, 2011) and conclude that only three factors (income per capita, infrastructure, and international trade) are statistically significant

The estimated coefficients of national realized absorptive capacity will be used as the dependent variable in the second stage of estimation process using multilevel modeling techniques (Steenbergen and Jones, 2002; Raudenbush and Bryk 2002) Lewis and Linzer (2005) state that estimated dependent variable model induces heteroscedasticity3, and ordinary least squares method with White’s heteroscedastic consistent standard errors can lead to more efficient estimates than the weighted least squares method We use the following model to run the pooled data ordinary least squares regressions:

𝑁𝑅𝐴𝐶𝑖𝑡 = 𝜃0 + 𝜃1𝐸𝑖𝑡 + 𝜃2𝐺𝑅𝑖 + 𝜖𝑖𝑡 (9)

in which NRAC is national realized absorptive capacity, that is, the coefficients 𝛽1+

𝛽1𝑖 from equation (8) in Step 1, E is the vector of explanatory variables, GR is the country group categorized based on the development level, and 𝜖 is the error terms

3 From a statistical viewpoint, heterogeneity “is the opposite of homogeneity, which refers to sameness

or similarity… While heterogeneity is a term commonly used by both ecologists and statisticians, heteroscedasticity is a purely statistical concept which concerns a particular type of heterogeneity restricted to inequality of variances” (Dutilleul and Legendre 1993, pp.152) In contrast, homoscedasticity, which indicates equality of variances, is the synonym for homogeneity of variances

13

3 Data and Estimation Techniques

This study uses the data collected from the World Development Indicators of the World Bank Using the single data source is expected to eliminate any unexpected errors due to the heterogeneity in recording systems of different countries The data source contains information on 218 countries and territories across the world from

1960, but the problem of missing data limits the study to 135 countries over the period of 26 years (1990 – 2015) The two criteria including an observation are the availability of data on foreign direct investment inflows for at least 20 years, and the availability of the purchasing power parity factor (PPP) data, which is only available since 1990

Variables are converted to international real dollars using the purchasing power parity factor FDI could be negative because “data on FDI flows are presented

on net bases (capital transactions' credits less debits between direct investors and their foreign affiliates) Net decreases in assets or net increases in liabilities are recorded as credits, while net increases in assets or net decreases in liabilities are recorded as debits Hence, FDI flows with a negative sign indicate that at least one of the components of FDI is negative and not offset by positive amounts of the remaining components” (World Bank, undated) Figure 2 shows the intuitive relationship between real FDI and real GDP, pooling the data of 135 countries It shows that high real FDI tends to be associated with high real GDP

Figure 2: Graph of real FDI versus real GDP for the sample of 135 countries

Source: World Bank, World Development Indicators

World Bank categorizes countries into four income groups – low, middle, upper-middle, and high – based on gross national income (GNI) per capita for the previous calendar year The categorization remains fixed for the entire fiscal year, but might change across years For the 2016 fiscal year, low-income countries are defined as those with GNI per capita in 2015 of $1,025 or less; lower middle-income countries have between $1,026 and $4,035 GNI per capita; upper middle-income countries have between $4,036 and $12,475 GNI per capita; high-income countries have GNI per capita of $12,476 or more Accordingly, out of 135 countries in this study, the number of economies categorized as low, lower-middle, upper-middle, and high-income is respectively 21, 43, 40, and 31 (see Appendix 1) By way of

“difference” form of explanatory variables in regression models

We use linear regressions, which take the advantage of the ordinary least squares method, for estimating the unknown parameters But ordinary least squares estimates are the best only when the assumptions are guaranteed: the errors have normal distribution, homoscedasticity4 and no serial correlation5 Violations of these assumptions might lead to inefficient estimates Therefore, it is important to check the assumptions of the regression models and fix the violations of the assumptions, if any The tests of stationarity and cointegration should be carried out before doing regressions to avoid linear regression violations The software STATA6 version 12 is used to process the data

4 Results

Tests of stationarity, cointegration, and Granger causality

There are numerous tests for stationarity of panel data Many of the tests require a strongly balanced data, which is a problem for us as our panel is not balanced Fortunately, the Fisher-type (Choi, 2001) and Im–Pesaran–Shin (2003) tests allow unbalanced panels Both tests have the similar null hypothesis that all panels contain unit roots7, but the alternative hypothesis is a little different between the two The alternative hypothesis in the Im–Pesaran–Shin test is some panels are stationary, while the alternative hypothesis in the Fisher-type test is at least one panel is stationary

Table 1 shows the results for the unit root tests of the two variables of interest, which are the change in real FDI as the percentage of real GDP in a previous year (hereafter, the %FDI) and the percentage change of real GDP (hereafter, the

%GDP) In both the Im–Pesaran–Shin and Fisher-type tests, time trend8 is included Before doing this, we also tested the stationarity of real FDI and real GDP The results from both tests show that while real FDI is stationary, real GDP is not Moreover, we are also interested in the number of lags (or past period values) to be has been inaccurately implemented and might lead to incorrect results Therefore, we use the newly suggested command pvargranger, proposed by Arbigo and Love (2016), for Granger causality tests To the best of our knowledge, however, the latest version of pvar commands is still in beta mode with no warranties, and users take their own risks in doing so

7 The presence of a unit root is a major cause of the violation of stationarity Unit root processes are those integrated of order 1 – I(1), which means differences are taken just 1 times for a process to be stationary

8 An upward or a downward trend in a time series is a common violation of stationarity In case the mean trend is deterministic, removing the trend from the data (or detrending) will make the process stationary Since we do not know the data is stationary around a time trend or not, we assume that it is,

17

included in the regression models later on Therefore, different lagged values are added into the tests It can be seen that two variables are stationary up to four lags included, and the results are robust between the two tests

Table 1: Unit root tests for the %∆FDI and the %∆GDP

Table 2 shows the results for the unit root tests of the other explanatory variables No lag is included, while time trend is included9 The tests were carried out for both the level (L) and differenced (D) of the data It is not surprising that comparable results are found using either the Im–Pesaran–Shin or Fisher-type test when the variables are in term of percentage rate For the others, there are some considerable distinctions between two tests Fisher-type tests tend not to reject the

9 For total population and fixed telephone subscriptions, time trend is not included since the test did not run properly with the inclusion of time trend

null hypothesis while Im-Pesaran-Shin tests are more likely to reject the null hypothesis that all panels contain unit roots The explanation lies in the difference between the alternative hypotheses of the two tests as mentioned above Fisher-type tests are less strict than Im–Pesaran–Shin tests

Table 2: Results of unit root tests for other explanatory variables

Cointegration is a technique to test if there is a long run association between two or more variables The two variables are said to cointegrate if there exists a linear combination of the variables that has a lower order of integration than the original

19

variables themselves10 If the two variables are cointegrated, ordinary least squares is

a valid methodology for estimating the relationship For variables that are integrated

of order one (that is, variables that are I(1)), then the variables are cointegrated if there is a linear combination of the variables with integration of order zero –that is, I(0) According to Table 1, our variables of interest are stationary, or have integration

of order zero Lütkepohl and Krätzig (2004) state that:

"occasionally it is convenient to consider systems with both I(1) and I(0) variables Thereby the concept of cointegration is extended by calling any linear combination that is I(0) a cointegration relation, although this terminology is not in the spirit of the original definition because it can happen that a linear combination of I(0)variables is called a cointegration relation (pp 89)."

Therefore, it is appropriate to run cointegration test for variables integrated of

order zero We do the test for the %∆FDI and the %∆GDP, using both Pedroni (1999,

2001) and Westerlund (Persyn and Westerlund, 2008; Westerlund, 2007) tests for cointegration in panel dataset Both test the null hypothesis of no cointegration between two variables For Pedroni’s test, the result from the dynamic ordinary least squares approach shows the t-statistic equal to 14.71 in case 2 lags and leads included

by default, and the t-statistic equal to 10.27 in case no lags and leads included Therefore, the null hypothesis is rejected This means that there is cointegration

between the %∆FDI and the %∆GDP For Westerlund test, continuous time-series are required After dropping four countries due to missing data in the %∆FDI (including

Algeria, Burundi, the Gambia, and Mauritania), the results reject the null hypothesis

at the statistical significance level of less than 1% (i.e., p-values less than 0.01); thus, results from both Pedroni’s and Westerlund’s tests indicate the cointegration between the two variables

We also run Granger causality tests (Granger, 1969) to discover the existence,

if any, of causal relationships between two variables of interest, that are the %∆FDI and the %∆GDP As mentioned in Section 3, we use the new package of programs

proposed by Arbigo and Love (2016) for Granger causality tests The result shows

that the %∆FDI Granger causes the %∆GDP (the chi-squared statistic equal to 5.160

with the p-value of 0.023), but the reverse direction does not occur (the chi-squared statistic equal to 1.746 with the p-value of 0.186) This result indicates there is one-

way causality from the %∆FDI to the %∆GDP This is an important result as it means

that %∆FDI is statistically exogenous, so we need not worry about simultaneity11

Step 1- Estimates of National Realized Absorptive Capacity

We begin with the very basic model, in which the dependent variable is the

%∆GDP and the single explanatory variable is the %∆FDI The dummies to

characterize the effects of individual countries are included We follow the typical practice in the literature, which models GDP growth as a dynamic system involving lags of both FDI and GDP growth (Alfaro, 2003; Choe, 2003; Durham, 2004; Adams,

2009) Therefore, we include the lagged values of %∆GDP and the %∆FDI into the

11 Hood, Kidd, and Morris (2008) state that employing Granger tests with panel data face the potential

of heterogeneity of individual panels in the dataset (i.e., the nature of the relationship is not constant across all countries), but panel Granger tests still have considerable benefits The tests are helpful to inform researchers on the potential relationships between variables prior to applying to multivariate

21

models Including the lags of dependent variable can reduce the probability of autocorrelation, but can have influences on the coefficients of other explanatory variables (Achen, 2001; Keele and Kelly, 2005) Since the consistency, and efficiency

of the ordinary least squares estimates will depend on the correct lag length selection (Han, Phillips, and Sul, 2017), it is important to find the optimal lag structure

There are several criteria for choosing the optimal lag length including AIC (Akaike information criterion), BIC (Schwartcz Bayesian information criterion), HQ (Hannan-Quinn criterion) In general, those criteria could be applied for a panel dataset, but there is no consensus on the selection of the optimal structure Some suggest getting an optimal lag structure for each country separately, and then use the number of lags attained the most frequently Others are more conservative and can choose the maximum lag of all countries – but this might not be the best decision in the case of large discrepancy (i.e., the criteria suggest 5 lags for a few countries and 1 lag for all the rest) For our specific panel dataset, we decided to run pilot regressions

trying all possible combinations of the lags of the %∆GDP and the %∆FDI The result

from the stationarity tests above show the two variables remain stationary up to four lags included, which is what we use The number of lags, the coefficients of the

explanatory variables of interest and the dummy variables (i.e., the %∆FDI by

individual country) remain rather consistent Furthermore, the first lag of the

dependent variable (the %∆GDP) and up to three lags of the explanatory variable (the

%∆FDI) remain statistically significant in all the pilot regressions As a result, we decided to add those lags (i.e., the lag of the %∆GDP by one period, the lags of the

%∆FDI by three continuous periods) in our final models, and the F-statistic also

shows the statistical relevance of the variables added

Using panel data requires identifying if fixed or random effects models are more appropriate A fixed effects model means a regression model where the group means are fixed, while a random effects model means a regression model where the group means are random (Ramsey and Schafer, 2002) Random effects model assumes that the unobserved variation across entities are uncorrelated with the explanatory variables in the model whereas fixed effects model allows that correlation (Greene, 2003) In random effects regression, time invariant explanatory variables (i.e., the explanatory variables that do not change across the study time) will

be included among the explanatory variables (Yaffee, 2003) But in fixed effects regression, those explanatory variables will be cancelled out (i.e., there are no estimates for those variables) and absorbed by the intercept (Torres-Reyna, 2007) Therefore, fixed effects models are often used when researchers are interested in analyzing the effects of variables that change over time within an entity (Torres-Reyna, 2007)

Fixed effects models assume that the time-invariant characteristics of individual countries are not correlated (since each country is different from others); otherwise, the inferences from the fixed effects model may not be correct (Torres-Reyna, 2007) For this reason, we conduct the Hausman test (1978) for the presence

of country-specific fixed effects to determine the appropriateness of the fixed effects models In STATA, the standard Hausman test is performed in the way recommended

23

by Wooldridge (2002) The null hypothesis of the Hausman test is the random effects model is preferred The results from the Hausman test12 show that for our specific data and model, fixed effects regression should be used

Yaffee (2003) discusses several types of fixed effects model and suggests though the variables might be homogenous across countries, there could easily be group-wise heteroscedasticity and/or autocorrelation over time We use the Wooldridge test (Wooldridge, 2002; Drukker, 2003) for autocorrelation in panel data and the modified Wald test (Greene, 2003) for group-wise heteroscedasticity in fixed effects regression model The null hypothesis in the Wooldridge test is no first order autocorrelation, and the null hypothesis in the Wald test is equal variances for all entities The results reject both null hypotheses13, which means the errors from the regression of the sample data are heteroskedastic and auto-correlated When the assumption of homoscedasticity is violated, the ordinary least squares coefficient estimates remain unbiased and consistent but they are inefficient and covariance matrix estimates are not consistent (White, 1980) Autocorrelation of the errors makes ordinary least squares estimates are not the best, but it does not make ordinary least squares coefficient estimates biased The problem with errors’ autocorrelation is the

12 For our basic model in which the dependent variable is %∆GDP and the explanatory variables include %∆FDI, the first lag of %∆GDP, three lags of %∆FDI and dummy variables for 135 countries, chi-squared statistic equals 167.60, and the p-value is 0.0000 The conclusion to use fixed effects regression remain unchanged when other control variables are added to the model

13 For our basic model, the result for the Wald test for group-wise heteroscedasticity (using the command xttest3) shows the chi-squared statistic is 23622.43 with the p-value of 0.0000; and the result for the Wooldridge test for autocorrelation (using the command xtserial) shows the F-statistic is 300.676 with the p-value of 0.0000

standard errors tend to be underestimated, leading to the overestimation of the statistics (Box and Pierce, 1970)

t-Regarding unequal variances of regression residuals, White (1980) proposed a heteroscedasticity-consistent covariance matrix estimator14, which is always appropriate even in the absence of detected heteroscedasticity (MacKinnon and White, 1985) Accordingly, inefficient but consistent ordinary least squares estimators are still used; however, an alternative “robust” standard error that allows for heteroscedasticity is calculated instead of the usual formula under homoscedasticity

In STATA, we employ the variance-covariance estimate (vce) (robust) option for every regression to obtain the “robust” variances for the ordinary least squares estimates

With respect to autocorrelation of regression residuals, Cochrane and Orcutt (1949) proposed an estimation procedure that adjusts serial correlation in the error term of a linear model15 When the error terms are correlated with first-order autoregression – AR(1), which means the error terms depend linearly on only the previous values, taking a quasi-difference can make the error terms stationary and make the statistical inferences valid In STATA, the xtregar command works for panel data based on this Cochrane-Orcutt procedure, and takes the first-order serial correlation of the error terms in the estimation process However, this can only correct autocorrelation, but does not correct heteroscedasticity

14 The approach was first proposed by Eicker (1967) and Huber (1967), then further improved by White (1980) Therefore, the heteroscedasticity-consistent standard errors are also known as Eicker- Huber-White standard errors

15 The technique is originally applied to time-series data A modification of Cochrane-Orcutt

25

What we need is a command that corrects both autocorrelation and heteroscedasticity instantaneously in a single regression Hoechle (2007) suggests the STATA command and option that produce robust standard error estimates for linear panel model in case the error terms being heteroscedastic and auto-correlated is xtreg, cluster()16 proposed by Rogers (1994) Clustered standard errors should be used if there is no restriction on the serial correlation structure of the errors (Stock and Watson, 2008) and they will correctly account for the dependence in a panel data set

to produce unbiased estimates (Williams, 2000; Petersen, 2009) For the fixed effects regressions applied to our specific data, the -robust-, -vce(robust)- or -vce(cluster)- options of the command xtreg, fe in STATA version 12 generate exactly the same results

A common practice in empirical studies is checking the robustness of the estimated coefficients, by adding or removing control variables to see how the estimates behave If the estimates remain significant and retain their sign under alternative specifications, then the estimates are considered robust We start with the basic model in which the only variables of interest are included (Model I) Model II is similar to Model I but includes three control variables, which have the most observations compared to the others Model III consists of all 12 control variables that are stationary under the Fisher-type test (See Table 2) Model IV double checks the statistical significance of three control variables, two of which are statistically significant in Model II and Model III while the other is statistically insignificant in

16 The term cluster refers to the entity (i.e., the country) Observations within a cluster are assumed to

be correlated while observations between clusters are uncorrelated (Williams, 2000)

Model III Table 3 shows a part of the regression results of these four models; the full results are presented in Appendix 3

Table 3: Regression results for models with the dependent variable %∆GDP

Dependent variable: %∆GDP Model I Model II Model III Model IV

Lag of %∆GDP (1 period) 0.1623*** 0.1491*** 0.0986 0.1502** Lag of %∆FDI (1 period) 0.3090** 0.3111** 0.0848 0.3997*** Lag of %∆FDI (2 periods) 0.1784** 0.1782* 0.0280 0.2284*** Lag of %∆FDI (3 periods) 0.1110*** 0.1115*** 0.0480 0.1266*** Change in total population 2.56e-09 1.73e-08

Change in labor force participation rate -0.0032 0.0066

Change in inflation rate (%) -8.71e-07*** -0.0002 -9.53e-07*** Change in fixed telephone subscriptions 1.26e-09

Change in real interest rate (%) -0.0020

Change in real exchange rate (%) 0.0007 -0.0000 Change in real education expenditure 2.81e-13

Change in government debt (% GDP) -0.0010

In the basic model without control variables (Model I), STATA automatically omitted Japan (i = 60) The reason for that omission is avoidance of the dummy variable trap when the value of one dummy can be predicted from values of other dummies This is relevant because Japan is a developed country that has high domestic savings and does not depend on foreign direct investment for its growth A country with positive coefficients on the interactive term is doing better than Japan in using foreign direct investment as an important external resource to develop the

27

economy A country with negative coefficients is doing worse than Japan, which means foreign direct investment has lower impacts on economic growth in that country In the four models with different control variables, Japan remains omitted

The results from Models I, II and IV are relatively consistent, which show that

in general, the %∆FDI has positive and statistically significant relationships with the

%∆GDP across the world In particular, the increase in real FDI as one percentage

change of real GDP in the previous year is associated with 2.1649 (as in Model I),

1.1738 (as in Model II) and 2.9603 (as in Model IV) increase in %∆GDP, holding everything else constant The %∆GDP in a previous year and the %∆FDI in three

previously continuous years also have positive and statistically significant

relationships with the %∆GDP For the first two models, values of R-squared show

that the included explanatory variables can explain about 30 percent of the variation

in the means of the countries’ %∆GDP

Model III consists of the most control variables; as the result, it has the highest R-squared The adjusted R-squared17 also increases, indicating that the inclusion of these control variables improves the model R2 = 0.5778 means that

57.78 percent of the variation in the %∆GDP is explained by the changes in

countries’ total population, labor force participation rate, inflation rate, infrastructure (i.e., fixed telephone subscriptions), exports of goods and services, real interest rate,

17 The adjusted R-squared is the revised version of R-squared, adjusting for the number of explanatory variables in the regression model While R-squared is defined as 1 - (SSE/SST), adjusted R-squared is derived as 1 - {[(n-1)/(n-p)][SSE/SST]} Accordingly, adjusted R-squared is always lower than R- squared, and could be negative if the explanatory variables do not help predict the response Adjusted R-squared is useful in comparing two or more models that have a different number of explanatory variables

real exchange rate, expenditure on education, business environment (i.e., the number

of start-up procedures to register a business), unemployment rate, income tax rate and government debt However, most of the estimates in Model III are not statistically significant A likely reason for this is missing data, which leads to a considerable drop

in total observations (from nearly 3,000 to 184), to a different subset of countries included in the regression, and to a different time period covered Specifically, the considerable drop in total observations reduces the degree of freedom and decreases the statistical power to detect the effect from the control variables Out of 22 countries included in Model III, there are only 12 high-income countries, seven upper-middle income countries, three lower-middle income countries, and none of low-income countries, compared to the sample total of 31, 40, 43, and 21 respectively Furthermore, 120 out of 184 observations are from high-income countries, and the number of observations per country varies from 3 to 11 The excessive weight towards the high-income countries leads to the bias, and many estimates of the dummy variables are omitted (See Appendix 3)

However, including control variables also reveals some statistically significant factors that might need some double checks In particular, Model II shows that the

change in inflation rate has a negative association with the %∆GDP (though the

practical significance is relatively weak as the estimated coefficient is very small) Model III shows that the change in a country’s exports of goods and services (percent

of GDP) has a positive relationship with the %∆GDP Therefore, in Model IV, we

include those control variables together with one statistically insignificant variable

29

(i.e., the change in real exchange rate) to see if the estimation is robust or not The result shows that the statistically significant factors remain significant, and the statistically insignificant factor remains insignificant The signs of those variables are also unchanged Furthermore, the estimates of the interested variables now become similar to Models I and II, the R-squared is relatively high at 39.83 percent and the number of observations is also considerable

One concern is multicollinearity, in which one explanatory variable is highly correlated with the other explanatory variables Multicollinearity does not decrease the reliability of the whole model, but affects the predictive power of individual explanatory variables18 It produces larger variance in the related explanatory variable; the standard error is higher; the t-value is smaller; the p-value is higher; then the explanatory variable tends to be insignificant though it might be statistically significant if there were no multicollinearity Therefore, slight changes in the data can lead to great changes in the sign and significance level of the estimated coefficient (Willis and Perlack, 1978), and out-of-sample prediction will be imprecise

A widely accepted method to detect the presence of multicollinearity is the variance inflation factor (VIF)19, which measures the inflation of the variance of an estimated coefficient as a result of multicollinearity If the VIF of an explanatory variable were 4.0, this means that the variance of the estimated coefficient of that

variable is four times greater than it would be without multicollinearity The higher the VIF is, the more serious the multicollinearity is According to Kutner, Nachtsheim, and Neter (2004), a VIF greater than 10 is considered to improperly influence the least squares estimates Checking for multicollinearity of our regression models, the results are presented in Table 4

Table 4: Variance Inflation Factor of variables as in the regressions of Table 3

Explanatory variable Model I Model II Model III Model IV

Change in total population 1.22 1.77

Change in labor force participation rate 1.05 1.45

Change in inflation rate (%) 1.00 13.72 1.00 Change in fixed telephone subscriptions 1.64

Change in real interest rate (%) 11.91

Change in real education expenditure 2.07

Change in government debt (% GDP) 1.88

Table 4 shows high multicollinearity of the %∆FDI, but relatively low

multicollinearity of the other explanatory variables The extremely high

multicollinearity of the %∆FDI is explained by the inclusion of dummy variables which are the %∆FDI by country That the variable “change in inflation rate” has low

VIF in Model I and Model IV (1.00), but high VIF in Model III (13.72) explains why the estimated coefficient is statistically significant in Model I and Model IV, but insignificant in Model III Multicollinearity makes the standard errors of the affected coefficients large; so, it may lead to a type II error (i.e., a failure to reject the null