Upload
others
View
4
Download
0
Embed Size (px)
Citation preview
The Predictive Capacity of the Gravity Model of Trade
on Foreign Direct Investment
Nationalekonomiska Institutionen Shen Gao Uppsala Universitet Handledare: Christian Nilsson HT-2008 2009-01-06
Index 1 Introduction……………………………………………………………………………3 2 Theoretical background ……………………………………………………………....5
2.1 Reasoning behind FDI………………………………………………………5
2.2 The Gravity Model ………………………………………………………….8 3 Empirical approach………………………………………………………………......10 4 Basic Model Specifications…………………………………………………………...12
4.1 The OLS model………………………………………………………..........12
4.2 OLS estimations……….……………………………………………………14 5 The Fixed-effect Model……………………………………………………………….19
5.1 Fixed-effect model specification……..……………………………………..19
5.2 Fixed-effect estimations……….……………………………………………21 6 Conclusions……………………………………………………………………………25 Appendices………………………………………………………………………………27
1 Introduction
The link between foreign direct investments (FDI) and trade is firmly established in
economic literature. Casson (1990) has for example suggested that FDI is a “logical
intersection” of the theory of international capital markets, the theory of the firm and
trade theory. Singh & Jun (1995) and Tanaka (2006) mention that firms might conduct
FDI for the specific purpose of “tariff hopping” and avoiding trade costs, suggesting that
trade issues have significant sway when firms make investment decisions. Yet despite the
vast amount of literature on this subject, very few have tried to look at FDI through the
lens of trade theory, choosing rather to approach the subject on either a macroeconomic-
level or on firm-level. The purpose and scope of this paper is not to extend and build
upon the ideas from such studies, but rather to explore FDI through the lens of trade-
theory.
The gravity model has been widely used in trade-theory to predict the level of trade
between different countries based on their economic size and distance from each other,
and it has been recognized for its empirical success and consistently high statistical
explanatory power (Bergstrand 1985). However, the vast majority of FDI studies have
chosen to incorporate trade theory and certain components of the gravity model of trade
into the macroeconomic-level of study. It is the intention of this study to refrain from
doing so, but rather to use the gravity model exclusively in modeling FDI values.
Questions that will be asked are whether the gravity model of trade can serve as a reliable
model for FDI value as well? Are there certain variables in the gravity model that are
distinctively powerful determinants of FDI? The idea is of course to see how strong the
link between FDI and trade actually is; whether the gravity model can obtain as
consistently strong results for FDI as it does for trade.
For practical purposes, FDI in this paper will be defined as when an investor from one
country obtains controlling interest in a (new or existing) firm in another country, and
then operates that firm as a part of the multinational business of the investor. FDI may be
financed through parent company transfer of funds to the new affiliate, borrowing from
3
home-country lenders, borrowing in the host country by the parent company, or any
combination of these strategies. A foreign investor is considered to have control over a
firm if they have 10% of shares or voting power in an enterprise (or the equivalent in an
unincorporated firm). FDI also pertains to investments in infrastructure, equipment and/or
organizations that allow the foreign investor to influence the management of the firm.
The data for FDI values come from the OECD.stat statistical database and the data used
in this paper span from the years 1990 to 2005. This time-period was chosen specifically
due to the post-Cold War market and trade liberalization initiatives that were prevalent.1
Countries included in the study are OECD countries chosen for geographic dispersion
and relevancy (Australia, Belgium, Canada, Czech Republic, France, Germany, Italy,
Japan, Korean Republic, Mexico, Netherlands, Spain, Sweden, Switzerland, United
Kingdom and United States), and five major transitioning countries (Brazil, China, India,
Russia and South Africa). One issue that Ceglowski (2006) mentions in her study, which
also applies to this study, is that the OECD data were supplied by national statistical
offices. Consequently, the investment partner detail in the data varies by country. For
some country pairs, only a single source of foreign direct investment is available. In such
instances, these are the data used in the analysis. In other cases, both partners report
incoming and outgoing direct investments. These two reported values are usually not
identical, and the size of the difference varies considerably. The discrepancies reflect the
methods, scope, and quality of the data collection used by the national statistical offices
that supplied the data. In these cases, the values from the country with the most thorough
and consistent reporting was used.
Since the FDI data is only available as calculated using current US dollars, the GDP and
GDP per-capita data are also in current US dollars. The GDP data however is taken from
the World Bank World Development Index due to an easier format to use and the
availability and accuracy of the data. One concern with using current US dollars is of
course the fluctuations of the exchange rates, and whether converted current values can
1 Data for the Russian Federation start from 1991, and they start from 1993 for the Czech Republic since these countries did not gain their independence prior to these dates.
4
accurately capture the effect these fluctuations had on investments in the past. The
problems associated with fluctuating exchange rates should of course be taken into
consideration when making economic inferences and drawing conclusions from the
models.
The next section serves as a theoretical background and review for FDI theory and the
Gravity Model of Trade, as well as their applications in this particular study. It is
followed by an overview of the empirical approach of this study. A specification as well
as analysis of the two models used in this study (OLS and fixed-effect) follows, and a
summary concludes.
2 Theoretical background
Traditionally studies on FDI have approached the problem either on an economic-level or
on firm-level. Firm-level approach to FDI is influenced by conventional investment
theory and microeconomics, whereas the economic-level approach is based on
international macroeconomics. This study will seek to use the firm-level approach to map
out the theoretical groundwork of the paper, and then use the economic-level approach to
empirically test the hypotheses proposed. As will become evident, this study mostly
concerns the so-called horizontal direct investment theory due to both data-constraint as
well as the nature of the study.
2.1 Theories on the reasoning behind FDI decisions
A central hypothesis of this paper will be the complementary relationship between FDI
and trade. One way to explain the FDI decision process and how it relates to trade is
through Helpman, Melitz & Yeaple’s (2004) theory of proximity-concentration trade-
offs.2 In this theory, firms engage in foreign markets because markets are imperfect, and
in weighing cost-benefits firms can decide to either; a) pull out of the foreign market, b) 2 For a detailed walkthrough of the theoretical model, refer to Appendix II.
5
export to the foreign market exclusively, or c) invest in a foreign production facility to
serve that specific market (FDI). The decision of interest to my study is whether a firm
exports or invests in a foreign market. According to Helpman, Melitz & Yeaple (2004) it
is the different relative costs of the modes of access that determine whether a firm
engages in exports only or whether they decide to do FDI. Exporting involves lower fixed
costs whereas investing means lower variable costs. The choice by the firm in this case is
driven by “the proximity-concentration trade-off: relative to exports, FDI saves transport
costs, but duplicates production facilities and therefore requires higher fixed costs. In
equilibrium, no firm engages in both activities for the same foreign market.” Essentially,
Helpman, Melitz & Yeaple find that exports are more profitable than FDI for low-
productivity firms and less profitable for high-productivity firms.
The detail of importance to this paper is that all firms identify potential profits in a
foreign market; the only difference is that their mode of access varies depending on
productivity. One firm investing a production plant in a foreign market does not preclude
another firm from exporting a comparable product to that same market. In fact, a market
that is identified to be profitable by one firm is highly likely to receive the same
assessment by a competing firm. Hence it is highly possible that trade and FDI are
complementary unless there is considerable information asymmetry or barriers to entry. It
is this connection between FDI and exports/trade that is the theoretical basis for the
empirical models of this study.
Worth noting about the proximity-concentration trade-off theory is that it is only
applicable to horizontal investments3; which fits in well with my study. As Markusen,
Venables, Konan and Zhang (1996) note in their paper, horizontal direct investment is
more relevant to developed countries, whereas vertical direct investment is more relevant
to investment in developing countries. In their study they find that “horizontal
multinationals dominate when the countries are similar in both size and in relative
3 In Horizontal Direct Investment, firms usually choose to produce roughly the same product in different locations, essentially substituting international production for trade. In Vertical Direct Investment firms spread the production of a single product in several locations, taking advantage of differences in factor prices.
6
endowments, and when trade costs are moderate to high”. Since a considerable majority
of the countries in my sample are developed countries, horizontal investments will be of
greater concern. However, a number of countries are transitioning economies (China,
India, Brazil etc.), in which case a blend of horizontal and vertical FDI might exist.
The issue of interest in horizontal investment is that these investments are high when
trade-costs are moderate to high. If one were to apply the proximity-concentration trade-
off theory, it would be obvious that the variable costs associated with exports are higher
in these cases and firms that would’ve otherwise engaged in exports will either pull out of
the market or simply invest (FDI). Another strand of theories worth noting is the Export-
Platform FDI that Ekholm, Forslid and Markusen (2005) have developed. In their study
they identify conditions under which large countries use small countries as an export
platform to server other high income countries. The reasoning behind this type of
horizontal investment is to avoid trade barriers as well as draw benefits from potential
free trade agreements (FTAs). The implications of their theory will be discussed further
in the empirical analysis section of this paper. The issue however is that this type of
horizontal investment seems to contradict the gravity theory of my study. According to
the export-platform theory firms inside the EU for example wouldn’t have to worry about
trade barriers or other costs, therefore their need to invest in other EU countries should be
very small. However, in this paper I would argue that investments between countries in
close proximity is still greater than distant ones due to two factors: 1) since similar
countries trade more with each other than dissimilar ones, then there should be more
investments between neighboring countries or countries that are part of the same FTA
due to less costs and risks; and 2) conventional FDI theory only deals with Multinationals,
whereas I would argue that Small and Medium-sized Enterprises (SMEs) constitute a
significant portion of FDI. One explanation for this theory would be that Multinational
firms nowadays have become Global firms and their investment costs are minimal
irrespective of region in the world. Investment costs for SMEs however, are still
considerable, and therefore these investments tend to end up in closer countries both in
geographic and cultural terms.
7
Vertical investments are only of peripheral interest in this study. As Zhang and Markusen
(1997) explain in their study, vertical multinationals exploit factor-price differences in the
world economy and allocate their investments and production accordingly, essentially
relocating labor-intensive but low-skill production to low-wage countries. This type of
investment might be of interest in data from countries such as China and India who have
traditionally been recipients of such vertical investments. However, with these economies
in transition and their accumulation of capital, a “reverse” vertical investment could
potentially be of interest. This is to say that firms from these traditionally low-wage
countries could potentially invest in more developed countries in order to quickly gain
access to a high-skill labor-pool. If this is the case, then FDI in both directions would be
of interest to observe, rather than just one direction as might be suggested in conventional
vertical investment theory. This thought process might explain potential discrepancies
between expected FDI based on distance and market size, and actual observations.
To summarize, the important theoretical components of this paper are that firms from one
country identify similar foreign markets as being potentially profitable, and the means by
which they serve these foreign markets is determined by the productivity of an individual
firm. While there are several factors that affect whether a market is deemed profitable by
firms, the central theory of this paper holds that distance (both physical and cultural) and
market size are the two most prominent factors due to perceptions of risk and the
involvement of all firms (and not just MNEs as is conventional in FDI theory) in FDI
activity. Finally, growing markets are not only targets for investment, but have large
firms capable of investing abroad themselves. This changing market outlook could
potentially make international investment more reciprocal than before, and relationships
that were previously identified as being “vertical” could potentially be “horizontal” in
nature, which would make these new markets highly interesting for this study.
2.2 The Gravity Model
The core idea behind the gravity model of trade is the notion that trade is determined by
the economic size of the countries involved as well as the physical distance between them.
8
In trade-theory, the gravity equation in its most basic and frequently used form is
specified as:
(1) 0 1 2 3 4ln ln ln ln lnij i j ij ij ijX Y Y D Fβ β β β β= + + + + + μ
where is the amount of trade between country i (host) and country j (home), is the
nominal GDP of each country, is the distance between the two countries, and
represents any other factors that might affect the amount of trade conducted between
country i and j. In conjunction with the economic size of a country is its market size,
meaning larger countries have greater potential markets which would attract more firms
to export to that country. To account for this possibility, some theories have suggested an
extension of the gravity model to include the population size of each country into the
equation.
ijX Y
ijD ijF
iN
4
(2) 0 1 2 3 4 5 6ln ln ln ln ln ln lnij i j i j ij ij ijX Y Y N N D Fβ β β β β β β= + + + + + + + μ
In this extended model, the economic size coupled with the actual size of the countries is
supposed to account for the market potential of a country that serves to predict trade
value.
A technical detail worth noting is that in practical application of the gravity equation, the
miscellaneous factors are frequently represented by dummy variables. This is because
more often than not, these factors tend to remain constant for each individual country.
Examples of such factors that can affect trade value are common language, common
borders, if they are members of the same RTA or FTA
ijF
5, common historical background
etc.6
4 See Cheng and Wall (2005) 5 Regional Trade Agreement and Free-Trade Agreement 6 For a common example of the use of multiple dummy variables in gravity modeling, see Stijn (2003).
9
Another noteworthy detail is that in most applications of the gravity equation, the data
tends to be cross-sectional, meaning the time-component t is held constant for all
observations. As Mátyás (1997) mentions in his paper, by holding the time-component
constant one is actually creating a restriction on the effectiveness of the model since an
assumption is being made that trade value does not change with the passing of time. If
this assumption proves to be false, incorrect interpretations about the independent
variables can be made and improper economic inference would most likely occur. To
overcome this problem, Mátyás (1997) propagates the use of panel-data to incorporate
e time-component. I will also use panel-data in this study to avoid making unnecessary
fixed-effect model to
etermine whether these concerns are warranted for my study as well. A thorough fixed-
ffect model specification can be found in section 5 of this paper.
ity in
th
restrictions on the model.
In addition to the general OLS gravity model, this study will also employ the use of a
fixed-effect gravity model in predicting FDI. The Fixed-effect gravity model is primarily
used because recent literature have identified some concerns regarding the viability of an
OLS model due to faulty model specifications that could lead to inaccurate parameter
estimations. Their solution has been to essentially equate the miscellaneous factors ijF of
the OLS model with unobserved heterogeneity between each country-pair of the fixed-
effect model. In my study, I will use both an OLS model and a
d
e
3 Empirical approach
This study will use the macroeconomic approach to FDI as a framework for the empirical
study. The reason for choosing a macroeconomic approach rather than the
microeconomic firm-level approach is two-fold. Firstly, the gravity model of trade is a
model that measures macroeconomic data and to use it in modeling microeconomic data
would be problematic to say the least. Secondly, almost all empirical studies on FDI have
been done using the macroeconomic approach due to availability of data and simplic
10
creating a feasible model. Essentially, the macroeconomic approach is more suited to
observe how FDI flows in certain patterns, which is also the purpose of this paper.
The idea behind the macroeconomic approach to FDI is that it emphasizes the
determinants of why net investment among pairs or groups of nations tends to flow in
certain patterns (Grosse, R. & Trevino, L.J. 1996). It attempts to explain FDI behavior
with macroeconomic variables such as inflation, national income and exchange rate etc.
(Trevino, L.J & Mixon Jr., F.G. 2004). As Grosse and Trevino (1996) demonstrate in
their study of foreign FDI in the U.S, most macroeconomic studies on FDI focus on three
separate groups of independent variables: 1) economic, such as GDP, per capita income,
exchange rate, interest rates etc.; 2) political risk; and 3) distance in both absolute terms
and cultural terms. Since this paper approaches FDI using the gravity model of trade,
only those economic factors included in the gravity model such as GDP and the distance
variable will be included explicitly. Political risk is an interesting variable since it not
only pertains to societal and governmental affairs, but also includes operations costs as
part of the overall risks involved in an investment. These operations costs can sometimes
be attributed to cultural factors, which make political risk a highly interesting variable for
this study. However, as Singh and Jun (1995) note in their paper, most empirical studies
on political risk have not been statistically significant, rendering it a rather unreliable
variable to include in the model. Thus, rather than including a separate risk variable, this
udy will use per capita income to serve as a rough proxy for political risk since in
y market size indicates the number of firms that have the capacity to invest
nd operate abroad. Thus both host country and home country GDP are expected to be
st
sweeping terms, well-off countries tend to have more stable institutions than poorer
countries.
Thus the most distinguishing variables used in this study will be home and host country
GDP and GDP per-capita, as well as the distance between each country-pair. The
reasoning behind including host country as well as home country GDP in the model is
that host country market size is an indicator of potential returns on an investment and
home countr
a
11
positively correlated with FDI, meaning the larger the market size the more FDI will be
conducted.
Per-capita income being included (aside from aforementioned proxy of risk) is built upon
the concept that countries with similar markets tend to trade with each other more than
dissimilar ones. 7 As Helpman, Melitz & Yeaple (2004) mention in their proximity-
concentration trade-off theory, higher productivity firms are more likely to engage in FDI
activity and well-off countries are most likely to have higher productivity firms.
Therefore it is reasonable to expect that home country per-capita GDP should have a
positive effect on FDI. In addition Grosse and Trevino (1996) mention in their paper that
demand patterns as well as firm behavior explain the trade between similar countries.
This concept is especially interesting to this study since the data encompasses OECD
member countries as well as the five countries nearest OECD accession. Roughly
eaking, these are the “richest” countries in the world and one could expect that their
onal costs and the perception of risk that can influence FDI
ecisions. In the case of geographic distance FDI is expected to decrease with increasing
istance, and for cultural distance, FDI is expected to increase if the countries share a
Basic Model Specifications
sp
markets tend to be rather similar. Thus the per-capita income of the host country is also
expected to have a positive impact on FDI.
Distance is central to the gravity model of trade and I have included two variables in
order to test its importance to FDI. Distance in this paper pertains to geographic distance
and the language dummy serves as an indicator of cultural distance. The rationale behind
including geographic distance to explain FDI is the greater cost of obtaining relevant
information as well as the difficulties in managing affiliates in distant regions. Cultural
distance may affect operati
d
d
common official language.
4
7 This ties in with theories concerning horizontal FDI, which will be discussed further on in this section.
12
4.1 The OLS Model
he basic gravity equation that will be used in this study is as follows:
(3)
T
0 1 2 3ijt it jt it 4 5 6 7ln ln ln ln ln lnjt ij ij ij ijtFDI GDP GDP PC PC DIS lang bordβ β β β β β β β= + + + + + + + + μ
j (home) to country i (host)
at time t, expressed in current US dollars.
: The GDP of country j at time t in current US dollars.
: Per-capita GDP of country i at time t in current US dollars.
: Per-capita GDP of country j at
between country i and j as measured by the distance
between each country’s capital.
my variable that takes the value 1 if country i and j share a common official
nguage, 0 otherwise.
o
ijtFDI : Total value of foreign direct investments from country
itGDP : The GDP of country i at time t in current US dollars.
jtGDP
itPC
PC jt time t in current US dollars.
DISij : The distance in kilometers
langij : A dum
la
ijbord : A dummy variable that takes the value 1 if country i and j share a comm n border,
0 otherwise.
As can be seen, per-capita GDP ( PC ) has taken the place of population size ( N ) in this
model. This is in part to account for the fact that similar countries (in terms of economic
development) have been observed to trade more with each other than dissimilar ones.8
Also, as Ceglowski (2006) mentions in her paper, some studies on trade have included
per-capita income in the gravity equation in order to capture elements of economic size
that are not fully contained in the income terms themselves. An example of this would be
8 Econometrically however, per-capita GDP (Y/N) and population (N) are essentially equivalent. As an explanatory variable, per-capita GDP only contributes through the variations in N. Variations in Y are redundant in the per-capita GDP case since these changes are already captured by the GDP variable.
13
the productivity of firms, which in the proximity-concentration trade-off would in turn
determine their means of serving foreign markets and affect FDI values. In addition, as
as been mentioned earlier in this paper, per-capita GDP can be seen as a crude measure
ry, and in this case would serve as a proxy measurement for
2 OLS estimations
ble 1: OLS regressions
udy.
nally, the fourth column lists the coefficients for an OLS regression with a year dummy.
h
of the development of a count
political stability.
4. Ta
*All the above values are robust and have been corrected for heteroscedasticity.
Four different OLS regressions were done in this study. The first column lists the
coefficients of a simple OLS regression with all the countries from the dataset. The
second column includes coefficients from a regression where the five prominent
developing countries, the so-called BRICS 9 , have been excluded. The third column
represents an OLS regression of only the European OECD countries included in the st
Simple OLS Simple OLS w/o BRICS
Simple OLS only Europe
OLS with year dummy (BRICS)
lngdpi .7576971 (0.000)
.7345006 (0.000)
.5106558 (0.000)
.7359769 (0.000)
lngdpj .9246042 (0.000)
.6833374 (0.000)
.8785434 (0.000)
.6848451 (0.000)
lnpci .3575239 (0.000)
.5958057 (0.000)
1.069413 (0.000)
.5854841 (0.000)
lnpcj 1.348072 (0.000)
2.913072 (0.000)
3.446074 (0.000)
2.904034 (0.000)
lndis -.8023343 (0.000)
-.7934467 (0.000)
-.9541462 (0.000)
-.794027 (0.000)
lang 1.297567 (0.000)
1.078454 (0.000)
-.6577832 (0.000)
1.083427 (0.000)
bord -.0417723 (0.583)
.1277888 (0.113)
.3696065 (0.000)
.1269802 (0.115)
yrdummy .0731315 (0.162)
Fi
9 Brazil, Russia, India, China and South Africa
14
The year dummy assumes the value 0 for all observations prior to 1999, and 1 otherwise.
The reasoning behind including this shift-dummy can be observed in Figure 1 below.
Figure 1: Absolute FDI according to year 0
1.00
0e+1
12.
000e
+11
3.00
0e+1
1FD
I
1990 1995 2000 2005Year
It is evident that FDI displays a growing trend during the time-period pertinent to this
study. However, the period after 1999 is of special interest since the IT boom took place
during this time. The variable yrdummy is thus an attempt to see whether the IT boom
had an impact on FDI, causing the sample to form a spline function. The regression that
my is based on the previous model that excludes the BRICS
includes this year dum
countries since their share of FDI has increased after the millennium shift making it hard
to determine the IT boom component of their FDI10.
Looking at the results of the simple OLS regression, we can observe that all of the
variables are significant except for the common-border variable, which is only significant
10 When doing the same year dummy regression WITH the BRICS countries, a significant value for yrdummy was attained.
15
in the sample containing strictly European countries. It is difficult to ascertain why
common borders would not be a significant factor in determining FDI levels. One way of
explaining this result might be that even though FDI and trade have a proven link, the
actual determinants still vary in certain key areas. Common borders is an important factor
in determining trade values because of transport costs, yet since investments do not suffer
from these transport costs common borders would probably be less of an issue. Of note
for this study is that the common borders variable is only significant in the regression
with European countries. This could be attributed to the theory previously put forth in
is paper that SMEs do in fact play a role in FDI. SMEs are more reliant on outside
an indication there is a causal
lationship between exports and FDI, several factors point towards these two variables
th
capital for their investments and well-off countries are more likely to have excess capital,
also SMEs are probably more likely to invest in close proximity to their home-market due
to higher costs of investment.
In conjunction with the SME theory as well as the weak results found with the common
borders variable, it would be interesting to see how the relationship between exports and
FDI holds up in the data sample pertinent to this study. In appendix III I have compiled
the results from regressing FDI using exports (lnexp) as an explanatory variable in order
to determine whether there is a causal relationship. Whereas the results weren’t very
strong, two issues merit some consideration. Firstly, the correlation coefficient between
FDI and exports in this data sample is 0.5602, suggesting there is a strong positive
correlation between the two variables. Secondly, adding an export variable significantly
increases the standard errors of all the GDP-related variables, signifying strong chances
that multicollinearity is present. Most likely the GDP-related variables are also
explanatory variables for exports. Thus while there isn’t
re
being highly correlated. This high correlation could point towards the complementary
relationship between trade and FDI, and the existence of both high-productivity firms and
low-productivity firms serving a similar foreign market.
As can be seen in table 1, the GDP per-capita of the home country (lnpcj) is in all four
cases the most pronounced positive determinant of the amount of FDI between country-
16
pairs. This result might tie in with Grosse & Trevino’s (1996) ideas of home market size
being a proxy for firm capacity to invest. However, instead of market size, it is per-capita
income that is of importance. One explanation for this higher capacity to invest ties in
ith Helpman, Melitz & Yeaple’s (2004) ideas of higher productivity firms being more
an issue when
e model is restricted to these countries. Yet another interpretation could be that lnpcj
l inferences.
variable (lang) makes quite drastic shifts depending on the sample being used. One
reason why the lang variable is negative for the European dataset could be that
w
likely to engage in FDI and well-off countries are prone to have more productive firms
able to invest abroad. Additionally, these countries have large pools of capital that enable
firms to make investments that they could otherwise not afford.
In conjunction with home-country per-capita GDP is the trend observed in the OLS
regressions that the more “concentrated” the sample, the more importance is attributed to
both home and host country per-capita GDP. Per-capita GDP has a higher coefficient for
strictly OECD countries than when the BRICS are included, and it becomes higher still
when the sample size is reduced to European OECD countries. This trend may follow the
observations made in several studies that countries with similar relative factor
endowments engage in more FDI activity11 (Tanaka 2006). Another way to interpret this
result would be that European economies tend to be smaller when compared with the
overall OECD dataset, and thus firm capacity to invest becomes more of
th
and lndis aren’t strictly uncorrelated, however neither a correlation matrix12 nor standard
errors suggest autocorrelation being an issue. The uncertainty surrounding the trend in
per-capita GDP makes it difficult to put forth any meaningfu
An additional observation of note is that the year dummy had no significant effect on FDI.
So there is nothing in this particular sample that would suggest that the IT boom had any
significant effect on FDI values amongst OECD countries.
One final observation from the set of OLS regressions is that the common language
11 In Appendix V I have tried to clear up the questions surrounding per-capita GDP by using dummy variables for richer countries. 12 Correlation between lngdppcj and lndis never goes above 0.08
17
Switzerland has a common language with three different countries; France, Germany and
Italy. This example illustrates the difficulties when using dummy variables to model
specific effects13. One way to attempt to correct this problem is to account for what
Cheng & Wall (2005) refer to as “heterogeneous effects” by using fixed-effect models.
Aside from the ambiguity of the lang and bord variable, taking a closer look at residual
plots also reveals problems with the OLS regressions. As can be seen in Figure 2 below,
there is a slight trend of more residuals being negative than positive. In addition, if one
looks at Figure 3 one can see that the range of the residuals becomes greater the larger
lndis is. As was mentioned earlier, these trends make inference-making from the OLS
e problem is to use fixed-
effect models, which will be the topic of the following section. Figure 2: Plot of the residuals from fitted OLS values
regressions somewhat unreliable. One possible solution to th
-10
-50
510
Res
idua
ls
10 15 20 25 30Fitted values
13 In Appendix IV I have excluded certain fixed dummy variables in favor of distance squared to determine whether the results change.
18
Figure 3: Regression residuals in relation to lndis
-10
-50
510
Res
idua
ls
5 6 7 8 9 10lnDIS
5 The Fixed-effect Model
5.1 Fixed-effect model specifications
The main issue with conventional OLS-methods of having a basic model and adding
dummy variables for each additional factor is that they pose several problems when
estimating results. As Cheng and Wall (2005) mention in their paper, the conventional
OLS gravity model yields biased results because they do not control for heterogeneous
effects for each country-pair. These heterogeneous effects are numerous and hard to
quantify. One example of this is if one considers the distance variable. In the OLS-
regression above I have chosen to define distance as being the distance in kilometers
between each country’s capitals. This might be an accurate measure of economic distance
if one were to compare smaller countries like Sweden and Switzerland. For larger
19
countries however, the distance between capital-cities might not necessarily be a good
indicator of economic distance since a large country can have several economic centers,
each with distinctive characteristics. In addition, physical distance doesn’t always
quantify the economic distance that is of interest. An example of this would be Moscow,
London and New York. The physical distance between London and New York in
kilometers is farther than that of London and Moscow, yet the economic links between
New York and London are much greater. Even if one were to compare New York and
Washington D.C with London, one would most likely find that the economic activity
between New York and London is greater simply because of both cities being large
commercial and financial centers. These heterogeneous effects are very hard to capture
ith conventional OLS methods.
s is a blunt instrument
r modeling the variables that affect FDI between country-pairs.
w
The other dummy variables included in the OLS regressions also cause some concern.
Common language, as was mentioned in the Switzerland case earlier, is a difficult
variable to capture. It is exceedingly difficult to designate a specific point at which two
countries can be said to share a common language or not. Sweden for example does not
officially have the same language as Norway or Finland; however there is a considerable
chance that one could make oneself understood speaking Swedish in these two countries.
In a broader perspective, the Asian countries might display certain language-related
positive effects simply because their languages are more closely related, and the same
could be said for the countries in northwestern Europe. The difficulty in quantifying
language is also applicable for the border variable. Island countries for example do not
have any neighboring countries, yet they might still experience proximity effects similar
to that of common borders. Countries can also have links that provide common border
effects without actually having any real borders (the channel tunnel between Great
Britain and France, the bridge between Sweden and Denmark). With these issues in mind,
it is clear therefore that an OLS regression with dummy variable
fo
20
Because of the difficulty in capturing these heterogeneous effects, some studies have
ested that a fixed-effect model approach would be most appropriate.14 sugg
(4) 0FDIijt t ij ijtijtZα α α β μ= + + + +
In the fixed-effect model above, FDIijt is trade value (or in the case of this paper, value of
FDI) from country j to country i in year t and Zijt = [zit, z ] is the vector of gravity jt …
variables (gross dom lation or per-capita GDP). The intercept
has three parts: one common to ears and country pairs,
estic product [GDP], popu
all y 0α ; one specific to year t and
common to all pairs, tα ; and one specific to the country pairs and common to all years,
ijα. In my case, since I am dealing with pane sted in the bilateral
eneous effects in FDI, the intercept of interest would b
l data and only intere
e heterog ijα. There is some debate
whether the country-pair effect ijα should be symm rding to the direction of etric acco
trade or FDI, meaning whether ij jiα α= or whether these fixed effects are indeed even
unique considering the direction of trade or FDI ij jiα α≠ . All of the fixed effect
regressions done in this study are in accordance with the two-way theory,
meaning ij jiα α≠ . It is altogether viable that a symmetric fixed-effect approach would be
equally feasible, or perhaps even better. However, due to time-constraints and concerns
generating symmetric pairs with STATA, I chose not to do regressions of symmetric
2 Fixed-effect estimations
in
fixed-effects.
5.
14 See Cheng and Wall (2005), Glick and Rose (2001), and Mátyás (1997).
21
Table 2: Fixed-effect regressions
Fixed Effect
Fixed Effect w/o
BRICS
Fixed Effect Europe
Fixed Effect Year
Dummy lngdpi .5540553
(0.215) -2.571137 (0.000)
-4.042991 (0.004)
-2.524095 (0.000)
lngdpj 5.30848 (0.000)
4.141524 (0.000)
12.05305 (0.000)
4.135882 (0.000)
lnpci 1.840277 (0.000)
5.559181 (0.000)
6.346213 (0.000)
5.718852 (0.000)
lnpcj -3.203955 (0.000)
-.9550657 (0.400)
-8.601446 (0.001)
-.7691635 (0.505)
yrdummy -.0869341 (0.061)
The same four datasets are used in the fixed effect regressions as in the previous OLS
models. The difference here is that distance, common language and common borders
have been taken out of the actual equation and instead contribute to the intercept ijα. A
noticeable result at first glance is that host country GDP (lngdpi) is not significant in the
first regression. This is a striking result since lngdpi was significant and robust in the
OLS model, and the result can in essence be interpreted as host country GDP not having
a significant impact on FDI value, which is almost counterintuitive. It is hard to draw any
concrete conclusions from these results, but one could assume that certain factors that the
OLS regressions attributed to lngdpi were in fact country-pair specific, meaning these
effects can be attributed to specific relations between country-pairs. Another perplexing
result for host country GDP is that it, when significant, has a negative impact on FDI.
Meaning the bigger a country’s economy, the less FDI it receives. This goes against both
what the gravity model as well as conventional FDI theory predicts. One should be
careful when making inferences about these results, but one hypothesis would possibly be
that since the sample is restricted to OECD countries, certain countries dwarf others in its
capacity for investment. For example, the U.S receives a considerable amount of FDI,
however due to its sheer size and abundance of capital; their outward investments dwarf
those of lesser countries thereby “skewing” FDI numbers.
Another very noteworthy result from the fixed effect models is that home country per-
capita GDP has a negative effect on FDI as well. This is fairly surprising since lnpcj was
22
found to be the most important positive contributor to FDI in the OLS regressions. This is
somewhat alarming since the results could also undermine the proximity-concentration
trade-off hypothesis and firm productivity effects on FDI. However, lnpcj is only
significant in the first and third model, making the results very ambiguous and it
wouldn’t be a stretch to assume that home-country per-capita GDP simply has no
significant effect on FDI. If one were to hypothesize about how home-country per-capita
GDP could negatively impact FDI, one suggestion could be that well-off countries have
sufficiently large internal markets as to make firms deem FDI to be unnecessary.
In contrast with the OLS regressions, the fixed effect models all point to home country
GDP (lngdpj) to be significant and having a highly important effect on FDI. This result
might in fact support Grosse & Trevino’s (1996) original theory that home country GDP
can be a proxy for firm ability to invest, rather than lnpcj as the OLS models seemed to
suggest. Also, the per-capita GDP of the destination country (lnpci) is found to have a
considerable effect on FDI in the fixed effect models. One explanation could be that
firms invest in well-off countries since these countries have more consumers willing to
spend, a nod to the Keynesian economy where firms cater to the market so to speak.
If one were to speculate as to what factors might explain the wildly different results of the
fixed-effect regressions, one aspect would be the existence of Free-Trade Agreements
(FTA) and Regional Trade Agreements (RTA). As Ekholm, Forslid and Markusen (2005)
mention in their Export-Platform theory, firms sometimes set up a local affiliate to serve
an entire region in order to circumvent certain trade barriers and take advantage of the
free-trade agreement. Thus, an American firm might only invest in plants in one EU
country, when its intention is actually to serve the EU in its entirety thereby affecting FDI
figures. Indeed, some of these fixed effects could also be rules and regulations that make
direct investments prohibitively costly, thereby forcing even highly productive firms to
resort to exports rather than investing. In conjunction with the trade bloc issue is the
problem with exchanges and currencies. All values in this study are designated using
current U.S dollars; inflation rate as well as foreign exchange fluctuations might affect
23
FDI values. Also, since the Euro wasn’t in public circulation until 2000, its role in the
sample time period chosen for this study is still undetermined.
Another issue that relates to fixed effects to an extent is that different countries attract
certain types of FDI. Some firms engage in resource seeking FDI, other firms do market
seeking FDI. FDI can also be done for efficiency purposes or simply out of strategic
(market share) reasons. Investing in a copper mine entails entirely different commitments
and risks than setting up an assembly plant for beverage cans. Larger markets are also
more likely to attract FDI for the sole purpose of having a strategic presence in that
market. The differing characteristics enjoyed by the countries in this sample could thus
attract specific types of FDI, which in turn could affect the total FDI value in that country.
The ambiguous results from the fixed-effect models warrant the question whether it is
useful in determining FDI values. Indeed, one major flaw with the fixed-effect models
used in this study is that the vast majority of countries included have had fairly stagnant
population growth during the time-period studied. This almost “fixed” nature of the
population could have had unobserved effects on per-capita GDP, which is one of the
major explanatory variables of this study. To test whether the population is indeed
“fixed”, I redid the fixed-effect regression using population rather than per-capita GDP as
an explanatory variable (see appendix VI for results). The results were inconclusive but
overall the results from the population fixed-effect models were still weaker than in the
OLS regressions. This suggests that the population variable does in fact cause some
concerns for the fixed-effect models. Indeed, there seems to be certain underlying issues
that puts into doubt the stability and viability of the fixed-effect model. The causes for
these problems are beyond the scope of this paper; however it would be an interesting
topic for future studies.
24
6 Conclusions
This study estimates FDI value using two modified versions of the gravity model of trade.
The findings when using OLS regressions are that the components of the gravity model
of trade are indeed key determinants of FDI value, and the two most significant positive
determinants were home country GDP (lngdpj) as well as home country per-capita GDP
(lnpcj). These results tie in with FDI theory where countries that are well-off are more
likely to have high-productivity firms whose profit levels allow for riskier FDI endeavors.
Larger well-off economies could also have better developed capital markets that enable
firms to make more investments. Distance (lndis), which is a key component of the
gravity model of trade, is also found to have a significant negative effect on FDI as was
expected. Several variables were found to have no significant effect on FDI value in the
fixed-effect model however, and only home country GDP (lngdpj) and host country per-
capita GDP (lnpci) were consistent positive determinants of FDI. The differing results
from the OLS models and the fixed-effect models cause inference-making to be
somewhat unreliable. There could indeed be country-pair specific factors that cannot be
modeled using simple dummy variables, yet the fixed-effect model might not be stable
enough or properly specified to accurately predict FDI.
As for the tie-in of results with the original hypothesis, the proximity-concentration trade
off seems to be accurate in the OLS model as was mentioned earlier. Distance and GDP
are also significant components, suggesting that “gravity” is indeed at work even for FDI
activity. The theory of an increasingly reciprocal international investment climate
however is hard to determine. The greater importance of the GDP variable for the dataset
including the BRICS countries, coupled with lesser importance for per-capita variables
suggest that the nature of these investments might still be “vertical”.
Finally, while the results in this study are inconclusive, there are several ways in which to
improve upon the study. One flaw is of course that OECD data is gathered from several
different sources and the accuracy of the data isn’t always completely reliable. Further
studies could also be conducted using the original population variable rather than per-
capita GDP as was used in this study. Another approach could be to test the gravity
25
model regionally according to continent or economic region. Since the regressions using
only European countries exhibited different results than when using all OECD countries
in this study, it could be possible that other regional FDI values display the same
characteristics as the European regression. One final improvement would be to use
symmetric as well as two-way fixed effect models as opposed to only two-way models as
was done in this study. It is altogether possible that home and host country characteristics
do not matter when determining FDI, and that it is rather the specific combination of
home and host country that is the key.
26
Appendix I
Cities Used to Calculate Distances
Country: City: Country: City:
Australia Canberra Mexico Mexico City
Belgium Brussels Netherlands Amsterdam
Brazil Brasilia Russian Federation Moscow
Canada Ottawa South Africa Johannesburg
China Beijing Spain Madrid
Czech Republic Prague Sweden Stockholm
France Paris Switzerland Bern
Germany Berlin United Kingdom London
India New Delhi United States Washington D.C
Italy Rome
Japan Tokyo
Korean Republic Seoul
27
Appendix II
Below is a simplified version of Helpman, Melitz & Yeagle’s model for exports and FDI.
Df : Fixed Overhead Labor Costs in domestic market
Xf : Fixed Cost of Entering Foreign Market (Exports)
If : Fixed Cost of Entering Foreign Market (FDI)
a : Total demand for the good (market size)
π : Profit functions
Df is simply the costs of starting a firm and continuing its operations in the domestic market. We think about Xf as the costs of forming a distribution and servicing network in a foreign country (similar costs for the home market are included in Df ). The fixed costs If include these distribution and servicing network costs, as well as the costs of forming a subsidiary in a foreign country and the duplicate overhead production costs embodied in Df . Simply: I X Df f f> > . Country i and j are assumed to be fairly similar and therefore the demand patterns as well as fixed and operating costs are assumed to be comparable as well. The reasoning behind why Xπ has a different slope than the other two profit functions has to do with the variable trade costs involved in exports. We are assuming that the two
28
markets are relatively similar and therefore the fixed costs involved in starting and operating a firm doesn’t differ that much. Iπ is further out due to the duplicated overhead costs of maintaining two separate plants in both country i and j.
Da is the break-even point for a firm in the domestic market. Above this point firms will continue to produce and if profits are below this point the firm will close down. is the break even point for exporting to a foreign market. If a firm has profits above this point then they will engage in exporting their good to the foreign market. is the point at which a firm switches from exporting to a foreign market to investing and creating a foreign affiliate to serve that specific market.
Xa
Ia
As can be seen, in this model the decision between exporting and engaging in FDI is decided by profits. The more profitable firms will seek to invest in foreign affiliates while less profitable firms settle with exporting. These differences in profits can be caused by several factors, but in Helpman, Melitz & Yeagle’s model they attribute the differences to firm productivity. The more productive a firm is the less cost it incurs from its operations, and therefore it is more likely to engage in costlier and riskier alternatives in supplying a market.
29
Appendix III
Table A.1: OLS regressions with exports as independent variable |
Simple OLS Simple OLS w/o BRICS
Simple OLS only Europe
lngdpi .2198556 (0.000)
.1368717 (0.000)
-.1596564 (0.050)
lngdpj .0798136 (0.050)
.0987361 (0.002)
.2023012 (0.022)
lnpci .2924683 (0.000)
.4391195 (0.030)
.9210574 (0.000)
lnpcj 2.614567 (0.000)
2.615896 (0.000)
3.359847 (0.000)
lndis -.2334437 (0.000)
-.2310847 (0.000)
-.4944987 (0.000)
lang .890915 (0.000)
1.111937 (0.000)
-.4655214 (0.000)
bord -.3732731 (0.000)
-.4915177 (0.000)
-.2229595 (0.009)
lnexp .7167134 (0.000)
.7737565 (0.000)
.840576 (0.000)
Table A.2: Fixed-effect regressions with exports as independent variable Fixed
Effect Fixed
Effect w/o BRICS
Fixed Effect Europe
lngdpi -.8716026 (0.082)
-2.396176 (0.000)
-3.229703 (0.015)
lngdpj 3.684273 (0.000)
4.206138 (0.000)
11.58231 (0.000)
lnpci 2.724313 (0.000)
5.227924 (0.000)
4.708106 (0.001)
lnpcj -.4380808 (0.674)
-1.992399 (0.073)
-9.122407 (0.000)
lnexp .217473 (0.000)
.3323291 (0.000)
.5817689 (0.000)
The variable lnexp above refers to exports from country j to country i during time t.
30
Appendix IV Table A.3: OLS regressions with distance squared Simple OLS Simple OLS
w/o BRICS Simple OLS only Europe
lngdpi .7614893 (0.000)
.7253432 (0.000)
.5765249 (0.000)
lngdpj .9294527 (0.000)
.6790035 (0.000)
.9242915 (0.000)
lnpci .3689059 (0.000)
.7295531 (0.000)
.9251114 (0.000)
lnpcj 1.367651 (0.000)
3.022054 (0.000)
3.339807 (0.000)
lndis -2.629262 (0.000)
-2.462159 (0.000)
-4.738608 (0.000)
lndis2 .1166761 (0.000)
.1048452 (0.000)
.286567 (0.000)
Squaring lndis and adding it as one of the independent variables in the regression doesn’t change the results very much. The only noticeable difference is that lndis has a higher coefficient in all three samples when compared with the OLS regressions done with the bord and lang variables. The larger coefficients could be due to the model attributing the common language and common border effects on FDI onto the distance variable. Appendix V The regressions below were done using a “rich country” dummy variable in order to see whether richer countries indeed trade more with each other as the horizontal direct investment theory suggests.
The basic idea behind the dummy is ( *ln ln )a richdummy pci pci+ . In essence, if the host country’s per-capita GDP (lnpci) is above $15,000 during the entire time-period of the sample (1990-2005), then that country is considered a “rich country” and the richdummy will be 1. If a country’s per-capita GDP is below $15,000 then it will have a richdummy value of 0. When doing the regressions, I found that if only the host country was classified as a “rich country” then the richdummy had very little effect on fdi (usually the coefficient was below 0.07). Also, including a richdummy in addition to the original per-capita GDP variables caused all of the per-capita variables to be less significant by a large margin.
31
However, if the richdummy accounted for both the host and the home country being “rich countries” the results became more favorable. The explanatory capacity of the actual richdummy is mostly inconsequential though. The chart below is a regression done using the richdummy with the European OECD country sample as well as the non-BRICS OECD sample. Linear regression Number of obs = 1028 F( 8, 1019) = 367.14 Prob > F = 0.0000 R-squared = 0.8301 Root MSE = .97981 ------------------------------------------------------------------------------ | Robust lnfdi | Coef. Std. Err. t P>|t| [95% Conf. Interval] -------------+---------------------------------------------------------------- lngdpi | .5179119 .0371958 13.92 0.000 .4449227 .5909012 lngdpj | .8552277 .0350776 24.38 0.000 .786395 .9240603 lngdppci | 1.082321 .0879805 12.30 0.000 .9096772 1.254965 lngdppcj | 3.660155 .1151881 31.78 0.000 3.434123 3.886188 lndis | -.9186476 .0563186 -16.31 0.000 -1.029161 -.8081339 lang | -.6762215 .0826395 -8.18 0.000 -.8383846 -.5140585 bord | .2954176 .0627482 4.71 0.000 .1722871 .4185481 richjlnpci | -.0005561 .000158 -3.52 0.000 -.0008662 -.000246 _cons | -55.36585 1.929752 -28.69 0.000 -59.15259 -51.57911 ------------------------------------------------------------------------------
Linear regression Number of obs = 2828 F( 8, 2819) = 860.28 Prob > F = 0.0000 R-squared = 0.7575 Root MSE = 1.3601 ------------------------------------------------------------------------------ | Robust lnfdi | Coef. Std. Err. t P>|t| [95% Conf. Interval] -------------+---------------------------------------------------------------- lngdpi | .7459619 .0286185 26.07 0.000 .6898465 .8020773 lngdpj | .6927902 .0247652 27.97 0.000 .6442304 .74135 lngdppci | .3228597 .0939411 3.44 0.001 .1386595 .5070599 lngdppcj | 2.696669 .0857324 31.45 0.000 2.528565 2.864774 lndis | -.7899063 .0272503 -28.99 0.000 -.8433389 -.7364737 lang | 1.014429 .0765585 13.25 0.000 .8643126 1.164545 bord | .1124194 .0806779 1.39 0.164 -.0457744 .2706132 richjlnpci | .0399937 .0089627 4.46 0.000 .0224196 .0575677 _cons | -42.13194 1.366784 -30.83 0.000 -44.81194 -39.45194
32
Appendix VI
Table A.4: Fixed-effect regressions using population
Fixed Effect Fixed Effect w/o BRICS
Fixed Effect Europe
lngdpi 2.378301 (0.000)
2.613812 (0.000)
2.04444 (0.000)
lngdpj 2.233647 (0.000)
3.213383 (0.000)
4.319855 (0.000)
lnpopi -1.614999 (0.000)
-4.196301 (0.260)
-7.21433 (0.000)
lnpopj 2.428034 (0.000)
1.06274 (0.000)
5.293708 (0.018)
The population data is taken from the United Nations UNData online database. The population figures are done using 5 year intervals, meaning accurate data are only available for 1990, 1995, 2000 and 2005. The 1990 figures are used for the time period 1990 – 1994. 1995 figures are used for the time period 1995 – 1999. 2000 figures are used for the time period 2000 – 2004. 2005 figures are used for the year 2005 only.
33
Data Sample References GDP and GDP Per-Capita numbers are taken from the World Bank World Development Indicators Online 2008 (WDI 2008). Export figures taken from SourceOECD ITCS International Trade by Commodities Statistics: Total Trade in Values Vol 2007 release 01.
FDI figures taken from SourceOECD International Direct Investment Statistics: International direct investment by country Vol 2008 release 01.
Population data are taken from the United Nations online statistical source (UNData).
Literary References
Bergstrand, Jeffrey H. “The Gravity Equation in International Trade: Some Microeconomic Foundations and Empirical Evidence”. The Review of Economics and Statistics, Vol. 67, No. 3, (Aug., 1985), pp. 474-481. Bergstrand, Jeffrey H. “The Generalized Gravity Equation, Monopolistic Competition, and the Factor-Proportions Theory in International Trade”. The Review of Economics and Statistics, Vol. 71, No. 1, (Feb., 1989), pp. 143-153. Casson, M. 1990. "The Theory of Foreign Direct Investment." In P. Buckley, ed., InternationalInvestment. Aldershot, England: Edward Elgar Publishing Ltd., pp. 244-73. Ceglowski, Janet. “Does Gravity Matter in a Service Economy?” Review of World Economics, July 2006, Volume 142, Issue 2, pp. 307-329. Cheng, I-Hui and Wall, Howard J. “Controlling for Heterogeneity in Gravity Models of Trade and Integration”. Federal Reserve Bank of St. Louis Review, January/February 2005, 87(1), pp. 49-63. Dunning, John H. “The Eclectic Paradigm of International Production: A Restatement and Some Possible Extensions”. Journal of International Business Studies, March 1988, Volume 19, Number 1, pp. 1-31. Ekholm, Karolina., Forslid, Rikard., & Markusen, James R. ”Export-Platform Foreign Direct Investment”. NBER Working Paper No. 9517, March 2003. Glick, Reuven and Rosen, Andrew K. “Does a Currency Union Affect Trade? The Time Series Evidence”. NBER Working Paper No. 8396, Jul. 2001.
34
35
Grosse, Robert., and Trevino, Len J. “Foreign Direct Investment in the United States: An Analysis by Country of Origin”. Journal of International Business Studies, March 1996, Volume 27, Number 1, pp. 139-155. Helpman, Elhanan., Melitz, Marc J., & Yeaple, Stephen R. ” Export Versus FDI with Heterogeneous Firms”. American Economic Review, March 2004, Volume 94, Number 1, pp. 300-316. Markusen, James R., Venables, Anthony J., Eby Kohan, Denise., & Zhang, Kevin H. “A Unified Treatment of Horizontal Direct Investment, Vertical Direct Investment, and the Pattern of Trade in Goods and Services”. NBER Working Paper No. 5696, Aug. 1996. Mátyás, Lászlό. “The Gravity Model: Some Econometric Considerations”. The World Economy, May 1998, Volume 21, Number 3, pp. 397-401(5). Mátyás, Lászlό. “Proper Econometric Specification of the Gravity Model”. The World Economy, May 1997, Volume 20, Number 3, pp. 363-368. Singh, Harinder., Jun, Kwang W. ”Some New Evidence on Determinants of Foreign Direct Investment in Developing Countries”. The World Bank Policy Research Working Paper No. 1531, Nov. 1995.
Stijns, Jean-Philippe. “An Empirical Test of the Dutch Disease Hypothesis using a Gravity Model of Trade”. EconWPA International Trade Series No. 0305001, May 2003. (For presentation at the 2003 Congress of the EEA, Stockholm, August 20 to August 24)
Tanaka, Kiyoyasu. “The Relative Importance of Horizontal Versus Vertical FDI: Evidence from Japanese and US Multinational Firms”. University of Hawaii Press, Oct. 2006. (Online Source) Zhang, Kevin H. and Markusen, James R. “Vertical Multinationals and Host-Country Characteristics”. NBER Working Paper No. 6203, Sep. 1997.