Telechargé par Camilla Fiorina

Report DeSanto Fiorina

publicité
Spatial Patterns in US Foreign
Direct Investment
International Master in Public Finance – 2019/2020
Alessia DE SANTO, Camilla FIORINA
Table of contents
1. Introduction ......................................................................................................................................... 3
2. Theoretical background and sources of spatial interdependence. ...................................................... 4
3. Data and methodology ........................................................................................................................ 5
4. Descriptive statistics ............................................................................................................................ 8
4.1 Standard descriptive statistics ....................................................................................................... 8
4.2 Spatial descriptive statistics ......................................................................................................... 12
5. Choice of the econometric model and results: full sample ............................................................... 15
5.1 The choice of the spatial model ................................................................................................... 16
5.2 Results .......................................................................................................................................... 19
5.3 Robustness check: alternative weight matrix .............................................................................. 21
6. Choice of the econometric model and results: Europe and Central Asia .......................................... 22
7. Conclusions and limitations ............................................................................................................... 24
Appendix A ............................................................................................................................................. 26
Appendix B - Codes ............................................................................................................................ 28
Appendix C - Database: full sample ................................................................................................... 28
Appendix D – Database: ECA countries.............................................................................................. 28
References ............................................................................................................................................. 28
2
1. Introduction
Foreign Direct Investments are a very dynamic phenomenon at the lead of the globalization wave which
has characterized all markets in the last decades. In this context of international integration, the activity
of multinational enterprises (MNEs) has increased relentlessly. As reported by Dharmapala and Riedel
(2013), from 1990 to 2004, Foreign Direct Investment by MNEs grew at an average annual rate of 12,4%,
which is more than double than the 5% growth rate of economy.
As the share of international transactions represented by FDI increases, so does the research interest
for its determinants, as all countries, especially developing ones, are eager to take steps to attract
foreign funds.
While a substantial body of literature exists about factors driving FDI patterns, no consensus has been
found, meaning that there is not a universally accepted set of variables which can be considered as the
true determinants of FDI.
Moreover, most of existing empirical and theoretical findings rely on a bilateral framework which most
of the time fails to take into account the inherent complexity of FDI decisions, which are multilateral by
their own nature. For example, an investment abroad could be addressed to a geographical position
which enables the company to strategically outsource their production in different countries in order
to exploit factor cost differences. Or, a host country could serve the parent as a base to operate in
neighboring economies. Agglomeration externalities and imperfect capital mobility - limiting the funds
that MNE can invest abroad - can also create interdependent FDI decision across host countries.
Standard econometric techniques cannot take into account multilateralism of FDI dynamics, since the
latter implies that observations across countries are not independent, a violation of the most
fundamental assumption of traditional econometrics.
Spatial econometrics models help to overcome this problem allowing for spatial autocorrelation
between observations, thus capturing potential spillover effects or the impact of market potential on
the level of FDI in one country.
Last years have seen a development of the literature on spatial determinants of FDI patterns.
Coughlin and Segev (2000) were the first to apply a spatial econometric model to study the
determinants of US FDI in Chinese provinces, and found that volumes of FDI in one region were
correlated with those of neighboring regions. Subsequent studies made on different parent countries
- among others Baltagi et al. (2007) Blonigen et al. (2007) for US outward FDI, Garretsen and Peeters
(2009) for Dutch outward FDI, Ledyaeva (2009) for Russian inward FDI, and Chou et al. (2011) for
Chinese outward FDI - find spatial correlations across FDI levels in neighbor countries.
Our work fits into this latest line of research, implementing a spatial econometric analysis on the
determinants of US outbound FDI in 2013 on a sample of 85 developed and developing countries.
It is structured as follows: section 2 introduces the reader to the main existing models about FDI
patterns. It includes a literature review on the most important theoretical and empirical works on the
topic. Section 3 provides a description of the database and the weight matrix used for the analysis , and
is completed by section 4 which presents the descriptive statistics. Sections 5 and 6 include the spatial
econometric analysis performed over the full sample and over a selected subsample of countries,
respectively. Section 7 concludes and presents the limitations of the research.
3
2. Theoretical background and sources of spatial interdependence.
The first formal development of MNE theory was provided in two important theoretical works from
1984.
Markusen (1984) presents a general equilibrium model in which parent firms establish affiliates to
produce and sell in the host country, in order to gain on trade costs, which is known as “horizontal FDI”.
Since this translates in higher fixed costs, this kind of MNEs arise in countries with large markets (which
allow for economies of scale and offer many investment opportunities) and where set-up cost of
factories are low and trade costs are high.
In contrast, Helpman’s (1984) general equilibrium model assumes that affiliates are created by parent
companies to carry out some stages of the production process, in order to access cheaper input factors,
while the home market remains the main destination of the firm. This type of investment is usually
referred to by the term “vertical FDI” and is more likely to come into existence when the host country
is unskilled-labor abundant and trade costs, as well as plant set-up costs, are low.
This gravity-type framework laid the foundation for the majority of empirical studies that followed,
which generally report as the main determinants of FDI in a host country: market size, distance from
the parent country, human capital endowment, availability of infrastructure, quality of governance and
security of property rights.
However, this kind of framework fails somehow to take into account the inherent complexity of FDI
decisions, as it considers MNEs’ investment decisions as bilateral.
More recent models have tried to relax this limitation, mainly through three extensions.
First, Carr et al. (2001) and Markusen (2002) have merged the gravity-type environment into a model –
known as “knowledge-capital” model - in which the equilibrium between vertical and horizontal FDI in
MNEs is endogenously determined, depending on factor availability and barriers to trade and
investments.
“Export-platform” models (Ekholm et al. (2003), Yeaple (2003), and Bergstrand and Egger (2004))
consider the possibility that the output of foreign affiliates is rather exported than sold in the home
market. In other words, MNE invest in a particular country to use it has a “base” to reach neighboring
markets. In an empirical study from 2005, Hanson et al. consider the share of US affiliates export to
home and foreign markets and find evidence of the importance of export-platform FDI in US MNEs.
Alternatively, MNEs can fragment the vertical chain of production across multiple countries in order to
exploit the comparative advantages in costs and productivity of different markets, before the product
reaches its final destination.
Baltagi et al. (2007) named this last practice “complex-vertical” FDI and formalize the importance of
“third-country effect” in FDI patterns. They consider a model with two sectors (one homogeneous and
one differentiated), three factors of production (capital and skilled and unskilled labor) and three
countries. When more than two countries are included, the decision to invest in a host country is not
bilateral anymore, but depend also on the characteristics of the third market.
Traditional econometric techniques are not suitable for capturing a third-market effect, since the latter
implies that FDI decisions across host countries (hence FDI volumes) are not independent. Spatial
4
econometric techniques allow to overcome this problem and each of the described FDI motivations has
a different implication in term of spatial relations1.
Following Blonigen et al. (2007) in this section we fill focus on the correlation between the level of FDI
in a country and in its neighbors, and how the surrounding market potential influences the level of FDI
in a country. However, our estimation will reveal the importance of other sources of externalities.
Horizontal FDI arise with the objective to avoid high trade costs for the parent companies, which are
usually due to import barrier imposed by host countries. In pure horizontal models the only trade off
faced by companies when deciding to establish foreign affiliates is whether the trade cost avoided from
operating directly in the host market is greater than the set-up costs. In this framework, surrounding
market potential or FDI level in neighbor countries will not play any role in the final decision outcome.
For this theoretical prediction to be realistic, however, one needs to assume that all destination markets
have sufficiently high trade protection against imports from other destination markets.
When this is not the case, and the commercial protection among destination countries is lower
compared to trade costs incurred by the parent to export in those markets, it is likely that MNE will opt
for a “export-platform” type of FDI.
Establishing one affiliate in a strategic position will bring the same benefits coming from horizontal FDI
in every destination country, while saving the high set-up costs of multiple plants, since it allows the
company to rich simultaneously all the neighbor markets without incurring in important trade costs.
For this reason, it is plausible to assume that the surrounding market potential will have a positive effect
on the level of FDI in the host country of interest. At the same moment, since fixed costs coming from
setting-up multiple plants are much higher than trade costs between neighbor countries, the decision
of a parent company to establish an affiliate in a country will likely come at the cost of not establishing
in the neighbors, which translates in a negative sign of the spatial lag on the FDI.
Vertical FDI model implies that companies choose to invest in the host country which offers the lowest
production costs for the activity which needs to be relocated. Analogously to what happens for exportplatform, the decision of a pure vertical MNE to invest in a country comes at the expense of neighboring
countries. On the other hand, since in this kind of MNE the final destination of products remains the
home market, the surrounding market potential is supposed to have an insignificant effect on the
volume of FDI in the host country.
In complex-vertical MNEs the production process is “fragmented” across different countries.
Consequently, a dense supplier network in a region can decrease the costs of separating productive
activities. Moreover, there might be other agglomeration forces (e.g. location of immobile resources
as mines, IT districts) which, as long they concern the production process of parent companies, can
attract FDI. In this sense, it is plausible to expect that both spatial lag on FDI and on market potential
can have a positive impact on the FDI inflow into a host country.
3. Data and methodology
The analysis to understand the determinants and pattern of US investment abroad is performed over
85 host countries for the year 2013. The database includes 14 Asian and Pacific (AP) countries, 29
It exists a multitude of spatial relations which will be presented in the following sections. Details on the econometric model and
techniques will be provided in section 5.
1
5
countries located in Europe and Central Asia (ECA), 16 Latin American and Caribbean (LAC) countries, 8
countries located in Middle East and North Africa (MENA) and 18 Sub Saharan Africa countries2.
The complete list of countries can be found in appendix A (table A1).
Although Canada is the country attracting the highest level of US investments, it has been excluded
from the sample due to the fact that it is isolated from the rest of countries, thus it might create some
issues when performing the spatial analysis.
Following Beugelsdijk et al. (2010), the level of US Foreign Direct Investment in each country is proxied
by Sales of US MNE affiliates into third countries. Affiliates sales, contrarily to FDI stocks, represent a
reasonably unbiased proxy of the value added generated by US firms abroad.
Data for this measure are obtained from mandatory surveys of U.S. multinational enterprises
conducted by BEA3.
Outbound of FDI from the United States has been chosen because the US are one the largest foreign
investors in the world. Host countries have instead been chosen arbitrarily depending on the availability
of data.
It is important to notice that missing data in BEA surveys do not indicate absence of US FDI in a host
country, but they represent a choice of BEA to avoid disclosure of data of individual companies.
Having access to those data in further research could strengthen this kind of analysis, since the presence
of “white spots” in the map might cause problems of identification of spatial interaction effects.
To account for this issue, we tried as well to implement the analysis exclusively on European and Central
Asian countries, which form an unbroken study area.
The control variables have been chosen based on previous literature (Blonigen et al., 2007; Regelink
and Elhorst, 2015). They include traditional determinants of the standard gravity model – as GDP,
population, distance and trade costs – as well as variables stemming from the most recent literature,
as the level of instruction and institutional quality. Values of these variables for the parent country are
not taken into account. Since the parent country is always the United States it does not provide any
additional information to include the observation on these variables for the US.
Total GDP, measured in current US dollars, accounts for the market size of the economy. The sign of its
effects on FDI flow is expected to be positive, since a greater market size offers more investment and
operating opportunities. To control for the fact that FDI have the tendency to move between wealthier
markets (Blonigen et al., 2007) the total population – measured as the total number of residents – is
included among the regressors. Since, holding GDP constant, a higher population decreases GDP per
capita – which proxies the wealth of a nation – a negative impact on FDI is expected.
For completeness, GDP per capita has been included as an alternative proxy for market size.
Data for the aforementioned variables are taken from the World Bank Development Indicators4.
Trade cost represents the barriers that might affect the trade between home and parent country and
is obtained computing the inverse of the standard openness measure, calculated as the sum of exports
Countries are grouped by geographical region according to the World Bank classification. Asian countries include country from South
Asia and East Asia and Pacific regions.
3Bureau of Economic Analysis, US direct investment abroad: Operations of US parent companies and their foreign affiliates.
https://www.bea.gov/international/di1usdop
4 https://databank.worldbank.org/source/world-development-indicators
2
6
plus imports divided by GDP5. According to previous researches effects of trade cost on FDI flows are
ambiguous and depend on the prevalent form of FDI. If FDI are undertaken in order to exploit factors
cost differences by vertically outsourcing production, higher trade costs will discourage the home
country from investing in that host. On the contrary, as previously mentioned, very high trade costs
could also create an incentive for the parent firm to establish a new plant in the host country in order
to avoid high transportation cost or excise duties. This latter case will translate in a positive impact on
FDI outbound toward a country with high barriers to trade.
The distance between parent and host countries is also added and it can be considered a proxy for both
higher management costs - which increases costs of production thus decreasing benefits coming from
investing in a certain host country - and higher trade costs, whose impact is ambiguous. The data on
bilateral distances come from the CEPII database6 where they are computed using the geographic
coordinates of the capital cities using the great circle formula.
Skill endowments and investment risk are also expected to impact the inflow of US FDI in host countries.
Skill endowments is proxied by average years of schooling for people aged 25 or older. This measure is
taken from the Human Development Reports of the United Nations Development Programme, which is
able to provide a wide range of observations, gathering values from Barro and Lee (2018), the UNESCO
Institute for Statistics (2019) and ICF Macro Demographic and Health Surveys, UNICEF Multiple
Indicator Cluster Surveys and OECD (2018). As higher skills improve productivity of workers, thus
lowering production posts, their impact on FDI volumes in a host country is expected to be positive.
The degree of investments risk is instead expected to have a negative effect on US FDI inflows toward
host countries. This risk can be influenced by several factors as weak property rights, political and
economic instability and also incidence of conflicts. All these sources of risk can be grouped under the
wider concept of governance. A number of studies has proven that good governance enhances
attractiveness of host countries (see, among others, Azemar and Desbordes (2009)).
Therefore, the risk is proxied by the Quality of Government index (as in Regelink and Elhorst, 2015).
This index is constructed by Quality of Government Institute7 and is calculated as the mean value of the
variables corruption, law and order and bureaucratic quality. It ranges between 0 and 1, and increases
if the quality of government increases. Hence, it is expected to have a positive coefficient in the
econometric estimation.
These variables are taken together to form the so-called modified gravity model (Blonigen et al. 2007):
This simple specification, which throughout the paper will form our baseline result, is then modified to
take into account spatial interaction between the components of the equation. To this purpose, the
choice of an adequate spatial weight matrix is fundamental to capture the existing spatial dynamics
between countries.
Since there are white spots in our map, and world countries are quite geographically dispersed, the
most suitable matrix is the one constructed using the closest neighbors. Since the number of neighbors
varies widely from country to country, we decided to use the matrix for the 5 closest neighbors as a
baseline, and then to test different specifications using a higher number of neighbors.
Import and Export measures come, again, from the World Bank Development Indicators.
http://www.cepii.fr/CEPII/en/bdd_modele/presentation.asp?id=6
7 https://qog.pol.gu.se/data
5
6
7
The trade off in this case is to be able to take into account the underlying spatial relations without
considering very far countries as neighbors, which can invalidate the results. To this regard, the
maximum number of neighbors that has been taken into consideration is 9.
As suggested in Regelink and Elhorst (2015), when the study is performed only on European and Central
Asian countries, an inverse of distance matrix is chosen. Since countries form a more compact group,
this kind of matrix can be preferred to the closest neighbors. However, the use of a contiguity matrix is
made impossible by the presence of some islands, as Cyprus and British Isles.
4. Descriptive statistics
Before proceeding with the econometric analysis of the relationship between the determinants
described above and US FDI volumes in host countries, this section includes a descriptive overview of
the data to illustrate the database.
When performing a spatial econometric evaluation, two kinds of descriptive statistics can be
distinguished: a standard descriptive analysis, which is useful to get a general idea of the existing
dynamics between variables, and a spatial descriptive statistics, which allow to detect potential spatial
relations and patterns among observations which can be included in the econometric model.
4.1 Standard descriptive statistics
A potential problem of this analysis can be the high heterogeneity of observations. The database
includes countries from all regions of the world, without distinguishing, for example, for the level of
economic development. This inevitably translates in a very high standard deviation, especially for those
variables as GDP, population and distance which cover very wide range of values.
When observations are highly dispersed it can be difficult to detect relationships between variables and
the estimation is likely to give high standard errors.
A potential solution to this issue is the use of logarithms, which allows to smooth extreme observations.
Table 1 (Panel A and B) reports the mean and standard deviation of variables in level and in natural
logarithm. As expected, dispersion is much smaller when the logarithm form is used.
Moreover, since the database contains countries that are highly heterogeneous, having access to
regional summary statistics can be more informative about existing differences across regions. Panel C
of Table 1 reports regional averages for all variables and highlights the existence of some regional
disparities. In particular, MENA and SSA countries show consistently lower US FDI volumes with respect
to other regions.
Moreover, countries located in Europe and Central Asia rank first in the majority of indicators. In
particular they display striking differences with respect to other countries in investment risk and trade
costs, which are significantly lower than elsewhere.
This could suggest the presence of some kind of agglomeration effects, and supports the potential
existence of spatial correlation.
8
Table 1 - Mean and Standard Deviation
Panel A – Mean and standard deviation for the variables in level
Mean
Standard Deviation
62294.06
120078.08
622056333363.21
1331992566609.53
19389.26
23491.16
Population
63845369.8
201457587.09
Trade cost
1.46
0.74
Skills
8.93
3.14
Investment risk
0.55
0.23
8431.89
3594.65
Sales 2013
GDP
GDP pc
Distance
Panel B – Mean and standard deviation for the variables in level
Mean
Standard Deviation
8.65
25.64
9.06
16.58
0.26
2.1
-0.69
8.94
3.21
1.9
1.44
1.59
0.49
0.46
0.44
0.49
ln Sales 2013
ln GDP
ln GDP pc
ln Population
ln Trade cost
ln Skills
ln Investment risk
ln Distance
Panel C – Regional average for all variables in natural logarithm
Sales 2013
GDP
GDP pc
Population
Trade cost
Skills
Investment risk
Distance
AP
ECA
LAC
MENA
SSA
9.95
26.77
9.09
17.69
0.36
2.17
-0.52
9.52
9.82
26.42
10.25
16.17
-0.01
2.45
-0.36
8.87
9.6
25.44
9.07
16.37
0.5
2.14
-0.86
8.27
8.31
25.66
8.82
16.84
0.3
1.95
-0.81
9.12
5.04
23.66
7.24
16.42
0.4
1.54
-1.14
9.09
9
As regional disparities have been detected in the average volumes of US FDI, it might be useful to check
the dispersion of the dependent variable in each region. Figure 2 displays the boxplots for the
dispersions of US FDI volumes (in natural logarithm) in each of the five world’s regions outlined in
section 3.
Dispersion of data is quite different across regions. The dispersion of FDI volumes is particularly high in
ECA and SSA countries. Indeed, ECA region includes countries from western Europe, which are very
highly developed and attract huge volumes of US FDI, with countries from east Europe and Central Asia,
the majority of which are emerging economies. The SSA region also covers a very wide range of value,
since in some of the less developed countries, as Niger, Togo or Gambia, FDI volumes are close to 0
while in more dynamic economies, as South Africa, values are much higher, hitting thousands of millions
of dollars.
The majority of regions displays a right skewed distribution, with the exception of LAC and MENA
countries which display a strong left skewedness.
In conclusion, it can be noticed that the logarithm form was useful also to correct for the presence of
outliers, which instead where numerous when considering US volumes in levels.
Figure 2 - Boxplots for the dependent variable (log), by region
Before proceeding with the spatial analysis of the database, it might be helpful to have an idea of the
type of relationship existing between the dependent variable and each of the explanatory variables, as
well as the amount of correlation. In order to obtain this information, each of the predictors was plotted
against the dependent variable, and a linear regression model was used to obtain the line of best fit.
Figure 3 displays the resulting scatter plots for each of the predictors.
The strongest linear relation is the one between the amount of US FDI, as proxied by affiliates sales,
and market size, as proxied by GDP. This is in line with the previous literature, which recognizes market
potential as one of the main determinants of FDI.
10
The scatter plots also highlight a positive linear relation between FDI and GDP per capita, population,
skills (proxied by average years of schooling) and investment risk (which is proxied by the quality of
governments), even if observations are more dispersed.
The only two variables which seem to be almost uncorrelated with the level of US FDI are the distance
- which displays anyway a slightly negative relation with FDI – and trade cost – which on the contrary is
slightly positive correlated with FDI levels.
Considering the distance, this result is sharply in contrast with previous literature. However, it is
possible to hypothesize that the increasing international integration of worldwide economies and the
availability of technologies which reduce costs and time of dealing with distance is contributing in
decreasing the relevance that it covers in investment decisions of MNEs.
The low correlation between trade costs and FDI levels can instead be associated to a presence of both
vertical and horizontal FDI motives, which, as previously mentioned, react in opposite ways to the
presence of higher barriers to trade.
Figure 3 – Scatter Plots
Relationships highlighted in Figure 3 can be found also in the correlation matrix, which in addition
includes the sign and amount of relationships between all variables. This is very useful to detect for the
possible presence of multicollinearity.
In general, it does not report very high correlation across the explanatory variables, with the exception
of the proxies for skills and investment risk, which are highly correlated with GDP per capita. This could
11
rise some problems when performing the estimation for their coefficients. This potential issue will be
further investigated in Section 5.
Figure 4 – Correlation Matrix
4.2 Spatial descriptive statistics
The spatial descriptive statistics provides information about the existence and the amount of spatial
correlation among observations. In this study we are mainly interested in detecting the presence of
spatial autocorrelation in the amount of US FDI, which can signal the presence of potential spillovers
effects among countries, as well as in the proxies for market size, namely GDP and GDP per capita, to
study the surrounding market potential of countries.
A first way to check for the presence of agglomeration effects is to plot the variables of interest in a
map. This allows to check for the existence of geographical regions with higher concentration of US
capitals and of the other variables of interest.
Figure 5 plots the amount of US FDI, as proxied by affiliate sales, for all host countries included in the
sample. It is possible to detect some regions where US FDI volumes are much higher than elsewhere.
In particular, western European countries seem to attract the highest amount of US capital, followed
by East Asia and Pacific region and Latin America and the Caribbean.
In contrast, African and Middle East seem to attract very low levels of US FDI.
12
Figure 5
A Moran test performed on our observations confirms the presence of spatial autocorrelation in the
amount of US FDI worldwide. As specified in Section 3, the weight matrix of the 5 closest neighbors will
be used as a baseline. Appendix A reports the Moran test and the Moran map using the closest 9
neighbors, which is the maximum number of neighbors which will be considered.
Table 2 reports the result from the Moran Test: the p-value is close to 0, meaning that we refuse the
hypothesis of absence of spatial autocorrelation with a good amount of certainty. The Moran I statistics
is greater than 0, signaling a positive spatial autocorrelation: this means that neighboring countries tend
to display similar volumes of FDI. The result is confirmed when the test is performed using larger weight
matrixes.
Table 2 - Moran Test
Moran I test under randomization
data: ([email protected]$sales2013)
weights: PPV5.w
Moran I statistic standard deviate = 5.7692, p-value = 3.984e-09
alternative hypothesis: greater
sample estimates:
Moran I statistic
Expectation
Variance
0.311702157
-0.011904762
0.003146382
Panel A of Figure 6 reports the Moran Diagram and the Moran Map obtained using the closest 5
neighbors. Panel A reveals that the majority of observations can be found in the 3rd quadrant: there are
many countries which display below average values of US FDI and are clustered in below average
regions.
The Moran Map reported in Panel B is an illustration of the Moran diagram and allows to understand
which are the countries in the “above average” neighborhoods and in the “below average”
neighborhoods. The most clustered regions are western Europe, which confirms to attract the highest
13
volumes of US FDI, and Africa and Middle East, where many countries displaying very low volumes of
US FDI can be found in the same neighborhood.
The classification of Eastern Asia and Pacific countries as “hot points” (countries with above average
values which are clustered in above average regions) varies widely with the choice of the weight matrix.
Using a higher number of closest neighbors usually highlights the presence of a second cluster of
countries attracting above average volumes of US FDI, which includes China, Japan, India and Thailand
– as can be observed in figure A2 in the appendix.
Figure 6 - Spatial Autocorrelation
Panel A – Moran Diagram
Panel B – Moran Map
14
Appendix A includes also the Moran test and the Moran map for the logarithm of observations. The
existence of spatial correlation is confirmed and the Moran test reports higher values for the Moran I
statistics, signaling an even stronger positive correlation.
Although correlation remains positive, the majority of observations is now found in the upper right
quadrant, meaning that the majority of units has above average values for US FDI and are clustered in
neighbors with high FDI volumes. This is due to the logarithm transformation, which reduces
substantially the variability of observations; however, equilibriums across geographic regions remain
similar, with the exception that south America is now classified as an above average neighborhood.
Before proceeding with the choice of the econometric model, it is useful to check whether similar
patterns emerge also in the Moran maps for the explanatory variables, which can be find in Appendix
A.
In particular, it can be noticed that the Moran Map for the GDP displays very similar clusters to that for
US FDI: the highest levels of GDP can be found in western Europe and Asia. This is in line with the high
correlation existing between the two variables, and indicates that GDP could be a suitable proxy for
market size in this study. On the other hand, GDP per capita highlights a unique above average area,
western Europe, while the rest of the world - with the exception of Australia, New Zealand, Japan and
Saudi Arabia - is classified as cold point. This supports the idea that GDP might be a better proxy for
market size than GDP per capita.
Other explanatory variables display quite different dynamics and highlights different clusters. It can be
noticed that both patterns of average education and quality of governments display positive sptalia
correlation, and always recognize western Europe as an above average cluster. On the other hand,
observations for trade costs seem to be more negatively correlated, as it is common to find countries
with very low trade costs next to countries with above average trade costs. Two clusters can be
recognized: northern Europe, which is a below average area, and south America, whose trade costs are
on average higher than the rest of the world. However, those clusters are not the same highlighted by
the Moran map for FDI volumes, which is in line with the low correlation between those two variables.
5. Choice of the econometric model and results: full sample
In order to build a suitable model for the purpose of our analysis, we start from the simple linear
regression model which will be estimated by OLS (Ordinary Least Squares). Even if this is a non-spatial
model, it is commonly used as a diagnostic tool for model specification evaluation and as a benchmark
for comparisons with spatial models. Then we will proceed to analyze the presence of spatial
interactions among the units considered.
The simple OLS model is:
𝑦 = 𝛼 + 𝑋𝛽 + 𝜀
Where 𝑦 is the intercept coefficient, X is the matrix of the regressors and 𝛽 containes the corresponding
coefficients to be estimated; 𝜀 is the usual error term, that is a white noise term.
Including the set of regressors, we obtain the following modified gravity model:
15
The descriptive analysis revealed that a log-log specification is suitable for our data. This is supported
by previous findings as Blonigen and Davies (2004) which find that such specification leads to well
behaved residuals.
Results of this specification are reported in Table 4, where different regressions are presented in
order to show several issues that were encountered during the estimation.
The most general specification included all the regressors presented above. As it is possible to see in
column 1, this specification presents very weak coefficients and almost none of them resulted
significant. This is likely due to the presence of collinearity among GDP per capita and both investment
risk and skills proxies, as highlighted in section 4.1.
Column 2 presents the result of the estimation without the inclusion of GDP per capita. Clearly this
specification improves considerably and shows the relevance of the market size in influencing FDI flows,
GDP becomes in fact highly significant. Moreover, the level of skills endowment of a host country seems
also to matter – the coefficient is significant at 5%.
The distance between the parent and the host country results relevant and with a negative impact
which could have the straightforward interpretation that a higher distance implies higher
transportation or management costs.
The measure for investment risk remains insignificant even when removing GDP per capita. Therefore,
an additional specification is included. Column 3 shows the results of the regression excluding both GDP
and GDP per capita: only in this case investment risk resulted significant.
However, since the market size has showed to be a main determinant of FDI flows, it is important to
keep at least one variable that would allow us to see its relevance when implementing a spatial model.
Therefore only GDP per capita was excluded ant the baseline specification is the one showed in Column
2.
5.1 The choice of the spatial model
When performing a spatial analysis, one of the most crucial points is the choice of the weight matrix.
The presence of several white spots in our map influenced our choice, and made it impossible to use
the contiguity matrix. When the entire sample is taken into account, units are highly geographical
dispersed, and the most suitable matrix is those constructed using closest neighbors. Although we
performed several estimations using different number of neighbors, this section presents results using
the 5 and 9 closest neighbors, which are respectively the minimum and the maximum number of
neighbors which will be considered in this paper.
We approach the choice of the model using a bottom-up strategy: we started from a non-spatial model
and we then performed several statistical tests to detect the presence of different spatial relations. All
the following tests are applied on the baseline specification discusses above.
16
Table 3 - OLS regression results
Dependent variable:
Ln(Sales of Us Affiliates)
Complete
(1)
No GDPpc
(2)
Ln(GDP)
0.6148 (6.0603)
1.0521*** (0.2862)
Ln(GDPpc)
0.4392 (6.0794)
Ln(Population)
0.6948 (6.0890)
0.2555 (0.3028)
1.2342*** (0.1554)
Ln(Trade cost)
-0.6486 (0.4789)
-0.6467 (0.4751)
-0.3414 (0.5035)
Ln(Skills)
1.6024** (0.7369)
1.6070** (0.7295)
3.4721*** (0.5642)
Ln(Investment risk)
0.6787 (0.6867)
0.6834 (0.6792)
2.1766*** (0.5860)
-0.9532** (0.4015)
-0.9510** (0.3978)
-1.2440*** (0.4194)
-16.8278*** (4.8185)
-16.8030*** (4.7755)
-6.4097 (4.1425)
85
0.7545
0.7322
1.6613 (df = 77)
85
0.7545
0.7356
1.6506 (df = 78)
39.9571*** (df = 6;
78)
85
0.7120
0.6938
1.7766 (df = 79)
Ln(Distance)
Constant
Observations
R2
Adjusted R2
Residual Std. Error
F Statistic
33.8129*** (df = 7; 77)
Note:
No GDP no GDPpc
(3)
39.0575*** (df = 5; 79)
*p<0.1; **p<0.05; ***p<0.01
First of all, to choose between a spatial and non-spatial model, the Moran I test for the presence of
residual spatial autocorrelation is performed. The null hypothesis of this test is the absence of spatial
autocorrelation; hence, considering that the p-value is equal to 0.0001124, we will reject H0 and
confirm the hypotheses about the presence of a spatial link among our data. However, it must be
pointed out that the value of the Moran I indicate only a mild autocorrelation.
Table 4 – Moran I Test
Global Moran I for regression residuals
data:
model: lm(formula = modeleln1, data = mylndata)
weights: matrice
Moran I statistic standard deviate = 3.6894, p-value = 0.0001124
alternative hypothesis: greater
sample estimates:
Observed Moran I
Expectation
Variance
0.161251104
-0.040365644
0.002986355
17
Once the presence of spatial autocorrelation is confirmed, we proceed to perform different statistical
test in order to choose the spatial model that will fit our data in the most accurate way. There are four
possible models to be chosen: SLX, SAR, SEM and SDM.
These spatial econometrics models account for the three types of spatial interaction:
•
•
•
an endogenous interaction, when the economic decision of an agent or geographical zone will
depend on the decision of its neighbors (SAR or Dynamic model);
an exogenous interaction, when an agent’s economic decision will depend on the observable
characteristics of its neighbors (SLX model);
a spatial correlation of the effects due to the same unobserved characteristics (SEM model).
This correlation could be explained by the presence of a spatially structured latent variable or
by the presence of composition effects which marks a form of spatial heterogeneity.
Finally, the SDM model is a combination of the SAR and the SEM and by combining the two models, it
also includes the features of the SLX specification. The SDM model is the more complex one since it aim
at analyzing the presence of spatial externalities, of a spillover effect and of spatial autocorrelation in
the error terms.
To choose among these models, the Lagrange multiplier tests are performed. Below the results of four
different tests are reported: LMerr, LMlag, the RLMerr and the RLMlag.
These tests account for, respectively: error dependence, a missing spatially lagged variable, error
dependence but conditional on the presence of a missing spatially lagged variable and lastly for the
presence of a missing spatially lagged variable but in presence of error dependence.
Table 5 - Lagrange Multiplier p-value results
LMErr
LMlag
RLMerr
RLMlag
p-value = 0.0107
p-value = 0.9646
p-value = 0.00166
p-value = 0.06598
Since we cannot reject the null hypothesis of the LMErr and the RLMerr, the above results indicate that
the preferred model should be the SEM. However, before proceeding with this model, the likelihood
ratio test is performed. The latter compares the differences between the Spatial Error Model (SEM) and
the Spatial Durbin Model (SDM) and indicate the more appropriate specification. The presence of pvalue smaller than 0,05 bring to the conclusion that the SDM model is the best one.
Table 6 - Likelihood ratio test between the SDM and SEM model
Likelihood ratio for spatial linear models
data:
Likelihood ratio = 17.17, df = 6, p-value = 0.00868
sample estimates:
Log likelihood of ze.sardm
Log likelihood of ze.sem
-146.4163
-155.0012
18
Finally, the Akaike Information Criterion might be helpful with the choice, confirming or rejecting the
expected result. In this test, the model which reports the lower value should be the preferred one. In
our case the lowest value is the one assigned to the SDM model hence confirming our previous
conclusion.
Table 7: Results of the Akaike information criterion
Model
OLS
SLX
SAR
SDM
SEM
AIC
335.1172
325.2747
337.1098
322.8327
328.0023
At this point it is possible to proceed and estimate the Spatial Durbin Model with the baseline
specification discussed above. The model has the following form
𝑦 = 𝑊𝑦𝜌 + 𝑋𝛽 + 𝑊𝑋𝜃 + 𝜀
where 𝑊𝑦𝜌 account for an endogenous effect given by the lagged dependent variable; 𝑊𝑋𝜃, the
spatial lag of the explanatory, account for spatial externalities. Moreover, as mentioned, this model is
also robust to the presence of spatial autocorrelation in the error terms.
5.2 Results
Results coming from the implementation of the SDM model are showed in Table 8. The interpretation
of coefficients for a spatial Durbin model is complex since it contains both exogenous and endogenous
spatial dependencies.
Marginal effects cannot be drawn directly from the coefficients of the regression but it is instead
necessary to decompose them in 3 effects. First, the direct effect reflects the impact on the dependent
variable of a unit change in the explanatory variables. The indirect effect contains both endogenous –
spillovers – and exogenous effect, which may arise from the values of explanatory variables of one
country’s neighbors. The sum of these two effects compose the total effect. Table 8 displays the
estimates of the 3 effects for our model, while the standard regression table is reported in Appendix A
(table A6).
Starting from the direct effects, the results show that the traditional determinants of FDI remain
significant when a spatial regression is performed. Moreover, they show the expected sign.
The market size, represented by GDP, has a positive effect on US FDI flows. This comes to no surprise
since the fact that a larger market size should attract FDI flows is quite straightforward. First of all, the
GDP is an indicator of the degree of development, therefore, since investors will invest in a country
where the perceived profitability of their projects is secured, they will take into account the signal
transmitted by GDP in this regard. Moreover, a larger market size implies, usually, lower distribution
cost (if of course the production and distribution facilities are located where the majority of the
customers will be located) and a greater chance of obtaining better and more specialized inputs.
19
Skills endowment measure is significant and has a positive sign: greater skill levels are usually positively
correlated with FDI. This might signal a prevalence of horizontal FDI (both pure horizontal and exportplatform) motives, in which the firm is attracted by host countries with a higher human capital which is
believed to improve productivity.
In contrast, when pure vertical motives drive FDI decisions, skills might show a negative impact on FDI
since MNEs prefer host countries which are unskilled-labor abundant in order to benefit from low labor
costs.
Investment risk is, as expected, positively linked with the level of FDI. This measure is an indication of
the quality of the government therefore it is usually expected to attract foreign investor. The ability of
governments to maintain a favorable and stable macroeconomic environment and the strength of the
law system are in fact proven to play a positive role in influencing FDI flows in both developed and
developing countries. The presence of low inflation and stable exchange rates as well as the high
protection of the property rights increase the feasibility of economic activity and the scope for market
transactions.
The sign of the coefficients of trade costs allows to detect the prevailing motivations behind US FDI
outbounds and to understand that FDI decisions across host countries are not independent. Being a
proxy for the protectionist measures of a country, high trade costs are one of the main determinants
of horizontal or “export-platform” FDI. When strong barriers to trade are present, a parent company
might decide to directly set up a new plant in a host country or to establish one affiliates in a strategic
position and use this as a “platform” to reach the neighbors countries, a practice which is known as
tariff jumping. In both cases a positive sign on the direct effect is expected.
However, our results are in the opposite situation. This suggest presence of Vertical FDI. The negative
sign of trade costs might in fact indicate that, since MNEs are pursuing the objective of delocalize their
production in order to reduce their costs, the presence of high trade cost in a certain country will make
them opt for a different location, therefore resulting in a negative impact of FDI outbounds in that
specific country.
This result is strengthened considering the sign of the indirect effect. Trade cost is the only significant
variable among the lagged regressors and shows a positive sign.
This seems to be in line with the aforementioned reasoning in favor of vertical FDI. In fact, in presence
of vertical FDI motives, if neighboring countries have higher trade costs with respect to a certain host
country, parent companies have incentives to locate affiliates in the latter one. Therefore, an increase
in the lagged value of trade cost has a significant and positive impact on volumes of FDI toward the host
country considered.
In addition, the fact that the surrounding market potential, as proxied by GDP, shows a non-significant
coefficient seems to point to the presence of a pure vertical FDI framework, excluding the possibility of
the complex-vertical model designed by Baltagi et al. (2007). In fact, in this kind of MNE the final
destination of products is the parent market. Therefore, the surrounding market potential is supposed
to have an insignificant effect on the volume of FDI in the host country.
Regelink and Elhorst (2013) confirm that similar conditions to ours signals predominance of pure
vertical FDI.
20
Altogether, our estimations do not show a clear indication of the main type of FDI in the sample
countries considered. However, heterogeneity of regions analyzed could have created some difficulties
in this sense. This issue will be further investigated in section 6.
Table 8 - Decomposition of direct and indirect effects, SDM model (5th closest neighbors)
Impact Effects: Simulation Results
Sales of US affiliates
Mean
Direct
SIgnificance
Indirect
Mean
Significance
lngdp
1.1543
yes
-1.0808
no
lnpopulation
0.3426
no
0.4560
lntradecost
lnskills
lninvestrisk
lndistance
-1.2143
2.0120
1.3009
0.9111
yes
yes
yes
no
3.8150
-0.1971
4.0009
-2.1323
Mean
Total
Significance
no
no
0.07352
0.79858
yes
no
no
no
2.60075
1.81486
5.30184
-1.22122
no
no
yes
no
no
* Montecarlo Method, Iterations = 1:1000, Thinning interval = 1, Number of chains = 1Sample size per chain = 1000
5.3 Robustness check: alternative weight matrix
As discussed before, the choice of the weight matrix has a crucial role in a spatial analysis. Therefore,
in order to test the robustness of our results, we perform the analysis described above using the weight
matrix of the 9th closest neighbors.
The same baseline specification is considered and the same bottom-up procedure in then followed to
choose the best spatial model. The Moran Test confirms that also in this case a spatial model is to be
preferred to the standard OLS, even if the value of the Moran I (0.078288898) indicate again a mild
autocorrelation. The various statistical test on the kind of spatial correlation indicate that the most
suitable model is, also in this case, the Spatial Durbin Model.
Table 9 shows the decomposition of direct and indirect effects for the SDM model. Analyzing the
average values and the significance of the different coefficients, results are almost identical to the one
discussed in the previous section.
The market size and the skill endowment of a country stays significant and have positive signs. A larger
market size and a more skilled labor force are again proven to bring a positive effect on FDI inflows in
a country. The significance and the sign of the trade cost measure is also confirmed.
With respect to the other weight matrix, also the indirect effects don’t show any change: trade cost is
the only significant variable among the lagged regressors and it shows a positive sign.
Therefore, results previously obtained appear to be robust to a change in the weight matrix.
21
Table 9 - Decomposition of direct and indirect effects, SDM model (9th closest neighbor)
Impact Effects: Simulation Results
Mean
Direct
SIgnificance
Sales of US affiliates
Indirect
Mean
Significance
lngdp
1.0466
yes
-0.155604
no
lnpopulation
0.3878
no
-0.001142
lntradecost
lnskills
lninvestrisk
lndistance
-1.1262
2.3872
1.4315
0.7642
yes
yes
yes
no
4.415057
-1.637989
2.650763
-2.224384
Mean
Total
Significance
no
no
0.8910
0.3867
yes
no
no
no
3.2889
0.7492
4.0823
-1.4601
yes
no
no
no
no
* Montecarlo Method, Iterations = 1:1000, Thinning interval = 1, Number of chains = 1Sample size per chain = 1000
6. Choice of the econometric model and results: Europe and Central Asia
When conducting an empirical study on the location of FDI an important issue is the heterogeneity of
the sample of countries considered. Several studies find evidence of substantial differences in FDI
determinants when comparing, for example, developed and developing countries. For example,
Blonigen et al. (2007) argue that different regions of the world are likely to attract FDI for different
reasons: while OECD countries are usually linked to pure horizontal or “export platform” motives they
found that, in contrast, in non-OECD countries vertical FDI models prevail.
In addition, again having in mind the differences between developed and developing countries, the
geographical localization also plays a role in particular, of course, when the spatial dimension is taken
into account. Developed countries are in fact primary clustered in Europe (and in general they are
geographically located in the north of the equator), while developing countries are more spread in a
higher number of continents. Then, when pooling together data from both groups, important
information about spatial interaction could be missed, negatively affecting the results of the analysis.
Therefore, in this section, we split our sample and decide to focus on a restricted group of 29 countries,
all located in Europe or central Asia. Since this group of countries does not present white spots it allows
to address potential issues coming from a geographically dispersed dataset.
To proceed in the analysis of this group we chose, as the weight matrix, the inverse of distance which
is commonly preferred in the FDI literature (Regelink and Elhorst, 2015). Following previous works on
similar samples (Blonigen et al., 2007), the threshold is set equal to the distance between Brussels and
Amsterdam, that is 173 Km. The model specification is identical to the one we used for the full sample
(meaning we exclude gdp per capita) since similar results are founded when implementing the
correlation matrix.
First, the usual tests are performed to identify the more appropriate model for this sample of countries.
After having confirmed the presence of a spatial relation in our model, we implement the Robust
22
Lagrange Multiplier test and the Akaike information criterion. Both tests indicate the SAR model as the
best one. Results of the decomposition of direct and indirect effects are reported in Table 10.
The sign of direct coefficients is in line with those found in the full sample; however only GDP and trade
costs stayed significant whereas in the full sample also skills measures and investment risk indicator
proved to have an effect of FDI. Hence, in this case, not all the traditional determinants of FDI inflows
seems to be robust to the inclusion of spatial components, indicating that different dynamics might
come into play.
What is more evident when considering a smaller sample of countries and the results of the analysis
implemented on it is that countries cannot be treated as independent entities since stronger third
country effects are detected as showed in the table below.
In fact, considering indirect effects, interesting results are obtained. First of all, a positive and significant
spillover effect from GDP is found: an increase in the GDP of one country will positively impact the
dependent variable of its neighbors and this will in turn positively affect the FDI inflows in the initially
considered host country. This might suggest that the surrounding market potential plays a role in
influencing parent firm’s decisions on where to allocate its investments, thus pointing to the presence
of “complex-vertical” and “export-platform” FDI models in which a dense supplier network and the
possibility of export in neighbor markets is crucial. When countries close to a possible host are attractive
markets and offer several investment opportunities, it could positively affect the decision of the parent
company, which has now several incentives to locate its affiliates in that specific host country.
Moreover, a negative sign in the coefficient of trade costs seems to strengthen this result: an increase
in trade costs will negatively affects the FDI outbound into neighboring countries and therefore indirectly - FDI in the first country considered. More precisely, if trade costs are high in country i, this
will negatively impact the level of FDI in its neighbors, which in turn will decrease the volume of US
investments directed into i.
This effect might be due to a prevalence of “export-platform” FDI inflows: in fact, in this kind of model,
multinationals are interested in markets that could be a strategic location base for serving neighbor
countries. In this context, the presence of high barriers to trade in one or more countries in a certain
region will lower the attractiveness of the latter and will create a negative spillover effect. This
hypothesis seems to find support in the fact that export platform models are common in regions where
countries enjoy free trade agreements. This is the case, for example, of European Union member
countries which are part of the Schengen agreement on free movement of goods and citizens.
Moreover, these dynamics apply also when the so called “complex-vertical” FDI prevail. In ECA region,
very different markets – from the very rich countries of western and northern Europe to the emerging
eastern and central Asian economies – coexist, in a relatively small geographic area. In presence of low
barriers to trade, MNEs can exploit differences in productivity and factors costs across these markets,
by strategically fragment their production process across those countries.
23
Table 10 - Decomposition of direct and indirect effects, SDM model (Inverse of distance matrix)
Mean
lngdp
lnpopulation
lntradecost
lnskills
lninvestrisk
lndistance
Direct
SIgnificance
1.6122
-0.1749
-1.5525
-1.8754
1.0992
-0.1075
yes
no
yes
no
no
no
Impact Effects: Simulation Results
US affiliates sales (ECA countries)
Indirect
Total
Mean
Significance
Mean
Significance
0.76468
-0.06980
-0.72860
-0.83505
0.51414
0.04819
yes
no
yes
no
no
no
2.37685
-0.24473
-2.28105
-2.71043
1.61337
-0.05933
yes
no
yes
no
no
no
* Montecarlo Method, Iterations = 1:1000, Thinning interval = 1, Number of chains = 1Sample size per chain =
7. Conclusions and limitations
Using a sample of 85 host countries (observed in the year 2013) this work aims to investigate the
multilateral nature of US FDI decisions and the possibility of spatial patterns in US investment abroad.
In addition, results obtained in this kind of analysis allow to understand the underlying motives existing
behind investments decisions by multinational companies.
First, a spatial econometric analysis was performed on the entire sample, which includes countries from
five world regions. Results from this first step reveal, in general, a prevalence of vertical FDI motives:
US multinational companies mainly invest abroad to exploit costs differences in factors of production.
This is coherent with the composition of our sample of countries. The sample used for this study had to
exclude for technical reasons two of the world leading economies: United States and Canada. Hence,
apart from western Europe and Oceania, the remaining countries can be mainly considered emerging
economies, which typically are able to offer low production costs to foreign companies.
In this context, externalities or spillover effects play a minor role as US companies are mainly interested
in producing abroad and selling in the home market.
Then, a spatial analysis was performed exclusively over European and Central Asian countries, which
highlighted very different dynamics. Spatial lags gained significance, revealing important spillover
effects in the European market. Surrounding market potential and existing trade barriers across ECA
countries are revealed to be important determinants of US FDI patterns in this area.
This suggest that decisions of US multinational companies to invest in Europe or Central Asia are highly
supported by export-platforms motives or can be categorized as complex-vertical FDI.
Taking into account this result one may conclude that FDI decision cannot be considered as bilateral.
Instead, multilateral relationships and third country effects come into play.
However, when drawing conclusions from our results it must be taken into consideration that the
difference in the relevance of spatial dynamics in the two samples could be due to the geographical
24
dispersion of countries when the entire sample is studied. As we already mentioned, the presence of
white spots might cause problems of identification of spatial interaction effects. However, with existing
spatial econometric techniques, this is a limitation one must take into consideration when performing
this kind of analysis. In addition, when considering such a heterogeneous group of country, it results
very hard to identify a single type of FDI and it is instead more likely that different FDI motivations
coexist, and contrasting results should be expected in the estimated coefficients of explanatory
variables.
A limitation exists also regarding the second step of the analysis since the sample of selected ECA
countries is too small and risks to report some bias or not to be representative. Hence, one must be
cautious when extending related findings to larger samples or drawing policy implications.
Despite its limitations, this work brings some very interesting findings which should be further
investigated. In particular, while the majority of literature has focused on the spatial dynamics
regarding market potential, the existing barriers to trade across host countries and the related spatial
dynamics seem to be highly relevant in MNEs decisions on where to direct their investments.
25
Appendix A
Table A1 – Countries included in the sample, by region
AP
Australia, Brunei, China, Indonesia, India, Japan, Republic of Korea, Sri Lanka, Malaysia, New
Zealand, Philippines, Papua New Guinea, Thailand, Vietnam
ECA
Austria, Azerbaijan, Belgium, Bulgaria, Belarus, Switzerland, Cyprus, Czech Republic, Germany,
Denmark, Spain, Estonia, Finland, France, United Kingdom, Croatia, Hungary, Ireland, Italy,
Lithuania, Luxembourg, Latvia, Netherlands, Norway, Poland, Portugal, Russia, Sweden,
Turkey
LAC
Argentina, Bahamas, Brazil, Chile, Colombia, Dominican Republic, Ecuador, Guatemala,
Honduras, Jamaica, Mexico, Panama, Peru, Paraguay, Trinidad and Tobago, Venezuela
MENA
Egypt, Iraq, Israel, Jordan, Libya, Morocco, Saudi Arabia, Yemen
SSA
Botswana, Cote d'Ivoire, Congo (Kinshasa), Congo (Brazzaville), Gabon, Guinea, Gambia,
Kenya, Liberia, Mali, Malawi, Niger, Nigeria, Sudan, Senegal, Togo, South Africa, Zimbabwe
Table A2 - Moran test, diagram and map using PPV9.w
Moran I test under randomization
data: ([email protected]$sales2013)
weights: PPV9.w
Moran I statistic standard deviate = 7.0074,
p-value = 1.214e-12
alternative hypothesis: greater
sample estimates:
Moran I statistic
Expectation
Variance
0.279431528
-0.011904762
0.00172853
Table A3 - Moran test, diagram and map of logarithm of observations using PPV5.w
Moran I test under randomisation
data: ([email protected]$lnsales2013)
weights: PPV5.w
Moran I statistic standard deviate = 7.0178,
p-value = 1.127e-12
alternative hypothesis: greater
sample estimates:
Moran I statistic
Expectation
Variance
0.410600879
-0.011904762
0.003624615
26
Table A2 - Moran test, diagram and map of logarithm of observations using PPV9.w
Moran I test under randomisation
data: ([email protected]$lnsales2013)
weights: PPV9.w
Moran I statistic standard deviate = 9.4803,
p-value < 2.2e-16
alternative hypothesis: greater
sample estimates:
Moran I statistic
Expectation
Variance
0.411077759
-0.011904762
0.001990687
Table A3 - Moran Maps of the explantory variables
Panel A: GDP
Panel C: population
Panel E: skills
Panel B: GDP per capita
Panel D: trade costs
Panel F: quality of government
27
Table A4 – SDM regression results
Dependent variable:
lnsales2013
SDM
lngdppc
lnpopulation
lntradecost
lnskills
lninvestrisk
lndistance
lag.lngdppc
lag.lnpopulation
lag.lntradecost
lag.lnskills
lag.lninvestrisk
lag.lndistance
Constant
1.0345*** (0.2664)
1.4419*** (0.1336)
-1.1065** (0.4402)
2.3974*** (0.7057)
1.4333** (0.6298)
0.7564 (0.7425)
-0.1882 (0.9124)
-0.1120 (0.5903)
4.4072*** (1.4258)
-1.4861 (1.5912)
2.8607 (2.2315)
-2.3185** (1.0894)
-6.3954 (13.8960)
Observations
Log Likelihood
sigma2
Akaike Inf. Crit.
Wald Test
LR Test
85
-146.7181
1.8482
323.4362
0.0238 (df = 1)
0.0212 (df = 1)
Note:
*p<0.1;** p<0.05; *** p<0.01
Appendix B – Codes: full sample
Appendix C - Database: full sample
Appendix D – Codes: ECA countries
Appendix E – Database: ECA countries
References
Baltagi, B. H. & Egger, P. & Pfaffermayr, M. 2007. "Estimating models of complex FDI: Are there third-country
effects?". Journal of Econometrics, Elsevier, vol. 140(1), pages 260-281, September.
Beugelsdijk, S. & Hennart, J. F. & Slangen, A. & Smeets, R. 2010. "Why and how FDI stocks are a biased measure of
MNE affiliate activity". Journal of International Business Studies, Palgrave Macmillan; Academy of International
Business, vol. 41(9), pages 1444-1459, December.
Blonigen, B. A. & Davies, R. B. & Waddell, G. R. & Naughton, H. T., 2007. "FDI in space: Spatial autoregressive
relationships in foreign direct investment". European Economic Review, Elsevier, vol. 51(5), pages 1303-1325, July.
Paper 2
Ekholm, K. & Forslid, R. & Markusen, J. R. 2003. "Export-Platform Foreign Direct Investment". CEPR Discussion Papers
3823, C.E.P.R. Discussion Papers.
Regelink, M. & Elhorst, J. P. 2015. "The spatial econometrics of FDI and third country effects". Letters in Spatial and
Resource Sciences, Springer, vol. 8(1), pages 1-13, March.
28
Téléchargement
Explore flashcards