Dependent variables
In this study, the dependent variable, FDI inflows was denoted FDIjt. This variable was
defined as the total bilateral flows of FDI into a nation j at time t. A sample of 13 Western
Balkan and Central Eastern European countries was empirically tested in order to determine the
determinants of FDI inflows into these host nations (Slovakia, Bulgaria, Bosnia and
Herzegovina, Poland ,Macedonia, Hungary ,Montenegro, Czech Republic, Romania , Croatia ,
Albania, Slovenia, Serbia). The information or data used in this evaluation was collected
between 2000 and 2011. This paper explores the compares and contrasts the main determinants
of FDI (Foreign Direct Investment) between Western Balkan and Central and Eastern European
(CEE) countries. Prior literature reported a group of two descriptive factors: factor endowment
(human, infrastructure) and gravity (market size and distance). Other factors that have profound
effects on FDI are trade barriers, geographical closeness, incentives as well as tax policy,
regional integration as well as labour costs. Gravity factors are noted to be a good explanation
for a majority of FDI influx into these economies like Western Balkan countries with policy
environment also noted to be important for FDI. International trade is also noted as one of the
main sources of foreign direct investment in Western Balkan and Central Eastern European
countries. By means of an econometric model based on panel data analysis, this study found that
both gravity factors (GDP, population, as well as distance) and non-gravity factors, or in other
words, factors that are not related to transition factors (such as labour costs, risk, and corruption)
could be used in explaining the nature of FDI influx in the transition economies.
Independent Variables
The independent variables were carefully chosen using Altomontes (1998) OLI
framework. Altomonte (1998) noted that Gravity models were initially invented in order to help
in explaining the nature of trade flows between two nations. This is analogous to the law of
gravity, in that the attractions between nations are weakened by the extent of geographical
disparity(distance ) between them. This can be simplified as the cost of transport between them.
The independent variables in this case were as follows:
a) Size of the host nations market (denoted by its GDP). This variable is most likely to have a
positive correlation with its coefficient also having a positive sign. High level of GDP translates
to higher levels of FDI. This is due to the logic that larger economies seems to attract more
capital as a result of the gravity approach
b)
variable (DISTANCE).This represents the actual distance in kilometres from the capital city of
the one country to the capital city of the regional headquarter (This is represented by distance
from Brussels). This variable haves a negative coefficient and hence a negative relationship with
the nature of FDI movements.
Additionally, we looked at several other variables that we deemed to be significant to the
determination of FDI. These included :
1.
The changes in labour costs in the host nation and these were incorporated into the
econometric model with as a (WAGE) variable. This represented the overall change in
labour costs in the host nation. It is worth noting that this is part of an efficiency
driven consideration. This was defined by Altomonte (1998) as the host nations
comparative advantage over the source nations (in terms of wage differentials) . The
variable is significant since measures the relative changes in the nations business
environment in entirety. This is as suggested by Janiniski and Wunnava (2004) on the
expectation for comparative wages to have a negative correlation to FDI flows.
2.
The other factor that was considered is trade liberation, and this was represented by
the variable TRADE. An infrastructure reform was also considered as a variable
(INFRASTRUCTURE).This was following the work of Beer and Cory (1996).
3.
Host nations investment climate was incorporated into the model using a special variable
named RISK. This was done in accordance with the work of Carstensen and Toubal (2004).
Model selection
The second details the model used in the study as well as the application of the CLRM
assumptions.
Models (A and B) have data from a 12-year consecutive observation of multiple determinant of
FDI that were appropriately regressed against the annual FDI confidence index for the 12-year
period.
Model A
The model has a 12-year consecutive observation of multiple determinants of FDI that were
appropriately regressed against the annual FDI confidence index for the 12-year period.
The Model is indicated below;
Ln(FDI jt ) 0 1 Ln(Distance jt ) 2 Ln(GDPjt ) 3 Ln(Population jt )
4 Ln(Wages jt ) 5 Ln(Risk jt ) u jt
(1)
In which:
j = Slovakia, Bulgaria, Bosnia and Herzegovina, Poland ,Macedonia, Hungary ,Montenegro,
Czech Republic, Romania , Croatia , Albania, Slovenia, Serbia
FDI = Foreign Direct Investment
Distance = Distance from a significant capital (Brussels) in kilometers
GDP = Gross Domestic Product of a given country
Population =Population of a given country
Wage = the rate of wages in a given economy
Risk = the level of risk according to the country risk index
= intercept
= coefficients of the various determinants of FDI
t = 2000 2011
Model B
(2)
Assumptions
The best model must satisfy the conditions set aside by the CLRM assumptions. The final
acceptable model must fulfill a series of criteria as set aside by Brooks (2008, p.192):
term;
The model must have stable parameter estimates that do not decay or flatter over the
patterns;
The model must be proficient for elucidating the outcomes of all available models
If all the assumptions are true, then the model estimators alpha ( ) and beta () are duly are
arrived at via the Ordinary Least Square (OLS) method and must have the required properties 1.
The ideal feature for the Ordinary Least Square (OLS) estimator is that they must possess a high
degree of consistency, fairness (unbiased) and efficiency.
In the process of validating the five assumptions, model estimation as well as economic
interpretations is provided.
The models A and B are not linear but were nevertheless estimated using OLS
Model A:
Ln(FDI jt ) 0 1 Ln(Distance jt ) 2 Ln(GDPjt ) 3 Ln(Population jt )
4 Ln(Wages jt ) 5 Ln(Risk jt ) u jt
LnDistanc
e
LnGDP
LnPopulati
on
LnWages
LnRisk
Constant
The
model
Coefficie
nts
0.054433
1
0.72737
83
0.38013
31
0.14789
85
-0.1045
477
2.127787
The Pvalues
Significance
levels at < 0.05
Significance
at < 0.10
Not significant
Not significant
0.000
Not significant
Not significant
0.036
Not significant
Not significant
0.342
Not significant
Not significant
0.614
Not significant
Not significant
0.185
Not significant
Not significant
0.655
Model B:
Ln(FDI jt ) 0 1 Ln(Distance jt ) 2 Ln(GDPjt ) 3 Ln(Population jt )
u jt
The
model
Coefficie
nts
The Pvalues
Significance
levels at < 0.05
Significance
at < 0.10
LnDistanc
e
LnGDP
LnPopulati
on
Constant
0.0863145
0.7918045
0.397
0.000
Not significant
Not significant
Not significant
Not significant
0.2827625
-2.052628
0.008
0.114
Not significant
Not significant
Not significant
Not significant
The value of R square can result in a negative value. This means that the models average
(y) accurately provides a good explanation for most of the difference in y than it does
The first assumption is not violated because it includes a constant term. In the model, the
constant term is -2.127787
The magnitude of the intercept is negative. This is justifiable due to the fact that some of
the FDI inflows were transformed by the logarithm function/transformation
The intercept in our model is negative. This indicates that the Balkan states and the CEE
states reveal less FDI that they should.
Model A
The visual observations outcomes were confirmed using the STATA command line hettest
residuals.
Fig. 3
The Stata results indicated that the constant variance of the obtained error terms (Ho) could
never be rejected. Therefore, the outcome indicates the existence of equivalent variance of the
residuals in line with the model line, which was predicted much earlier.
Model B
These visual observations outcomes were also confirmed using the STATA command line
hettest residuals.
In addition, visual observations were run using Stata. The figures below clearly indicate the
scatter plots of all residuals against each of the independent variable. It is amazing to see that all
the graphs obtained, indicated a constant variance of the models error term.
Again, it would be wise to ask ourselves what would happen if the regression models never had a
constant variance of the error term. The outcome would be heteroscedasticity that makes the
standard errors to be completely biased. The situation can only be solved by employing elaborate
standard errors in Stata. The given function would then dismiss either one or all of the originally
crafted OLS assumptions; these assumptions are that (A) that the errors are independent of each
other and (B) the errors are distributed identically.
The use of standard errors is noted to be more reliable. This however, never changes the original
coefficient estimates but rather, adjusts obtained standard errors. A process that in the end;
induces accurate values for the p-values.
The 3rd assumption: cov (ui , uj ) = 0
The third assumption states that there is no autocorrelation. All that exists is total lack of what is
commonly known as serial correlation between the error terms. This means that there is totally
no correlation between the models error terms.
The STATA visual plot below confirmed this.
Model A
Stata function having formal statistical test accurately confirmed the visual observation;
Model B
The Durbin-Watson d-statistic that was obtained was slightly close to one (unity) while the
correlograms indicated a total lack of autocorrelation in the models residuals of ten (10) lags.
Additionally, for model A, the plot of residuals versus lag residuals over time never indicated any
peculiar patterns. Signs of autocorrelation were also absent. The correlograms for model A and B
had similar values
What of a case whereby the regression models has autocorrelation in its residuals? The result
would be the failing by the model to correct for any cases of autocorrelation. However, when it is
present it would be similar to that of case heteroscedasticity is ignored. This is the case even if
the coefficients remain unbiased. In this case, they would not be described as Best Linear
Unbiased Estimators. As a result, estimates of standard would be wrong in some cases. If this is
the case, then R square is inflated with a corresponding increase of type one (1) error as well as
rejection of the null hypothesis in case it is correct.
The 4th assumption : the xts are all non-stochastic
The presented OLS model must assume that all the dependent variables are deterministic. This
implies that if a given regression were to be duplicated with a very new set of data, then the
independent variables would be determinate and remain purely the same. The dependent
variables however, would be different due to the new data stream. This would therefore generate
new values for the models error terms.
The 5th assumption: the disturbances are all normally distributed
This assumption suggests that single as well as joint hypothesis tests regarding the econometric
model parameters, demands a normality assumption i.e (ut~N(0, 2)). It is quite clear that the
proposed model Fulfilled all the CLRM assumptions. Further tests were necessary in order to
assess or evaluate the most suitable model for determining the FDI in Balkan and CEE countries.
Models A and B Fulfilled all the five CLRM assumptions. It was then necessary for us to
conduct additional tests aimed at assessing the most appropriate model for gauging the
determinants of FDI inflow into Balkan and CEE countries.
Omitted variable bias
An omitted variable bias is said to occur when two conditions are present: (A) the variable,
which is omitted, is directly correlated with the included regressor; and (B) when the variable,
which is omitted, has an obvious influence over the models dependent variable.
The omitted variable test that was performed in STATA for models A and B never indicated the
presence of an omitted variable bias.
Model A
Model B
What of case where the regression models A and B have omitted variable bias? In such a case,
The estimator of the models slope in its regression test on determinants of FDI (Distance, GDP,
population, wages, and risk) could be wrong. As a result, the beta coefficients magnitude could
be overstated or understated; a case that would not perfectly represent the correct determinants of
FDI.
Multicollinearity
The CLRM assumptions categorically states that the regressors are never perfectly
multicollinear.It also states that the explanatory variables never correlated with each other.
Model A indicated a high level of interrelationships between its independent variables.
Multicollinearity is therefore most likely to exist between variables that posses a correlation
value of more than 0.80, such as the one between lnGDP and lnWAGE.
It is imperative to note that high level of relationship existing between the models explanatory
variables is not always an indicator of a high level of multicollinearity. In certain instances,
multicollinearity is removed or completely ignored in cases where the model has either high
magnitude, or suitable sign coefficients. This is while it satisfies the extant economic theory.
Model A
A regression test on model A indicated that it lacks any significant coefficients both at five and
ten percent significance level.
LnDistanc
e
LnGDP
LnPopulati
on
LnWages
LnRisk
Constant
The
model
Coefficie
nts
0.054433
1
0.72737
83
0.38013
31
0.14789
85
-0.1045
477
2.127787
The Pvalues
Significance
levels at < 0.05
Significance
at < 0.10
Not significant
Not significant
0.000
Not significant
Not significant
0.036
Not significant
Not significant
0.342
Not significant
Not significant
0.614
Not significant
Not significant
0.185
Not significant
Not significant
0.655
Large outliers
In economic modelling such as FDI inflow, the concepts of large outliers are normally
present and may cause a total rejection of presumed normality assumption. The variables
in our model A and B were appropriately transformed by means of natural logarithms.
Large outliers were detected during the early years of EU accession (see figure below).
Summary
CLRM Assumptions
Model A
Model B
Fulfilled
fulfilled
1: E(ut ) equals 0
2: cov (ui , uj ) equals 0
3: var (ut ) equals 2 <
4: the disturbances exhibit normal distribution
5: the xt values are non-stochastic
Further Tests:
Fulfilled
Fulfilled
Fulfilled
models coefficients
Multicollinearity test
Fulfilled
Failed
Larges Outliers
Failed
Failed
Economic Theory
Fulfilled
Fulfilled
variable (DISTANCE).This represents the actual distance in kilometres from the capital
city of the one country to the capital city of the regional headquarter (This is represented
by distance from Brussels). This variable haves a negative coefficient and hence a
negative relationship with the nature of FDI movements.
Additionally, we looked at several other variables that we deemed to be significant to the
determination of FDI. These included :
1.
The changes in labour costs in the host nation and these were incorporated
into the econometric model with as a (WAGE) variable. This represented the
overall change in labour costs in the host nation. It is worth noting that this is
part of an efficiency driven consideration. This was defined by Altomonte
(1998) as the host nations comparative advantage over the source nations (in
terms of wage differentials) . The variable is significant since measures the
2.
The other factor that was considered is trade liberation, and this was
represented by the variable TRADE. An infrastructure reform was also
considered as a variable (INFRASTRUCTURE).This was following the work
of Beer and Cory (1996).
3.
Host nations investment climate was incorporated into the model using a special
variable named RISK. This was done in accordance with the work of Carstensen and
Toubal (2004).
In a nutshell, we see that the pattern of foreign direct investment (FDI) inflow into
the Western Balkan as well as Visegrd economies (Central Eastern European (CEE)
countries) is mainly determined by similar gravity factors like population, distance and
gross domestic product. Privatisation process is however quite different between the two
groups of countries with the Central Eastern European (CEE) countries or Visegrd
economies having a speedy privatisation process via their private sector. The process is
rather slower for the Balkan economies. This means that prioritisation-based FDI is lower
in the Balkan economies.
The CLRM Assumptions proved to be very important for this analysis. According
to the CLRM assumptions, the best model must satisfy the conditions set aside by the
CLRM assumptions. The final acceptable model must fulfill a series of criteria as set
aside by Brooks (2008, p.192). The model must be logically reasonable and be consistent
with fundamental financial theory. That includes satisfying any of the applicable
parameter restrictions. It must have regressors too that are uncorrelated with the existing
error term. The model must also have stable parameter estimates that do not decay or
flatter over the entire data sample. It also must have residual values that are very random
and should exhibit no patterns. And finally, the model must be proficient for elucidating
the outcomes of all available models .The objective is to come up a model which is
statistically satisfactory and capable of answering the laid down CLRM assumptions,
must be is parsimonious, and have an accurate economical interpretation and theoretical
backing. The aim is to come up with a model that can practically mimic the economic
realities within the various countries of interest.
This test has therefore revealed that the main determinants of FDI (Foreign Direct
Investment) between Western Balkan and Central Eastern European (CEE) nations are
numerous. As noted in prior literature, a special a pair of descriptive factors; i.e. factor
endowment (human, infrastructure) as well as gravity (market size and distance) are
important in this regard. The other factors that have great effect on FDI are indicated to
be barriers to trade, geographical closeness, tax policy as well as incentives and regional
integration and labour costs. Our model was chosen in order to represent both factor
endowment and gravity parameters. Future econometric models must aim at showing the
exact nature of multicollinearity existing between these determinants.
References
Altomonte, C. (1998). FDI in the CEECs and the theory of real options: an empirical
assessment. Discussion Paper No.7698, LICOS - Centre for Institutions and Economic
Performance, K.U.Leuven, Belgium.
Carstensen, K., & Toubal, F. (2004). Foreign direct investment in Central and Eastern
European countries: a dynamic panel analysis. Journal of Comparative Economics, 32(1),
322.
Janicki, H., & Wunnava, P. (2004). Determinants of foreign direct investment: empirical
evidence from EU accession candidates. Applied Economics, 36(5), 505509.
Variable: Lndistance
Variable: Lngdp
Variable:lnpopulation
Variable:lnwage
Variable:lnrisk