Economic Systems 37 (2013) 155–170

Contents lists available at SciVerse ScienceDirect

Economic Systems

journal homepage: www.elsevier.com/locate/ecosys

Capital mobility in the Caucasus

Rustam Jamilov *

Department of Research, Central Bank of the Republic of Azerbaijan, 32, R. Behbudov Str., Baku, Azerbaijan

A R T I C L E I N F O

Article history:

Received 24 April 2012

Received in revised form 3 December 2012

Accepted 12 December 2012

JEL classification:

F3

F31

C23

O53

Keywords:

Capital mobility

Saving-investment relationship

Panel cointegration

A B S T R A C T

This paper examines the degree of capital mobility in the countries

of the Caucasus. I estimate a simple model developed in the seminal

paper by Feldstein and Horioka (1980). I construct a panel of 6

countries of the Caucasus – Armenia, Azerbaijan, Georgia, Kazakh-

stan, Russia, and Turkey – and employ a panel cointegration

approach. To that end, I make use of the Dynamic OLS (DOLS), Fully

Modified OLS (FMOLS), and Pooled Mean Group (PMG) techniques

for heterogeneous panels. Preliminary cross-dependency tests

reject the presence of cross-sectional dependence. Panel unit root

and cointegration tests confirm that investment and saving are non-

stationary and cointegrated. The estimated long-run saving

retention ratios using DOLS, FMOLS, and PMG are 0.90, 0.73, and

0.83, respectively. These results suggest that capital mobility in the

Caucasus is very low. I put these findings in an international context

and confirm that the Caucasus is considerably financially restrained

compared to other regions. I also look at the country ratings of the

Index of Economic Freedom (IEF) and find that my results work well

in predicting the IEF rank. Finally, I discuss some implications for

the region’s policy-relevant issues such as financial integration,

human capital mobility, cross-border trading, fiscal and monetary

policy, solvency management, responsive consumption smoothing,

and recession resistance.

� 2013 Elsevier B.V. All rights reserved.

1. Introduction

Capital mobility, especially in the case of developing countries, is a valuable area of research –particularly from the practical point of view. The degree of capital mobility sheds light on the historical‘‘policy trilemma’’ – the choice between monetary independence, capital mobility, and exchange rate

* Tel.: +994 505218070; fax: +994 124938951.

E-mail addresses: rustam_jamilov@cbar.az, jamilovrustam@gmail.com.

0939-3625/$ – see front matter � 2013 Elsevier B.V. All rights reserved.

http://dx.doi.org/10.1016/j.ecosys.2012.12.004

R. Jamilov / Economic Systems 37 (2013) 155–170156

stability. With concrete practical relevance, capital mobility is important for determining the optimalfiscal and monetary policy (Mundell, 1968), managing the exchange rate (Levich, 1985), setting the taxinduced by inflation (Easterly et al., 1995), and for numerous other purposes as well.

There are multiple ways in which capital mobility can be estimated. Caprio and David (1984),Penati and Dooley (1984), Obstfeld (1986), and Calvo et al. (1992) have suggested a simple method ofanalyzing the aggregate values of capital flows. Others, such as Haque and Montiel (1991), Faruqee(1991), and Reisen and Yeches (1991), claimed that a good methodology for capital mobilityestimation is the uncovered interest parity formula. The chief proposition of this method is thatexpected returns from domestic and foreign assets are equalized via the inter-border arbitrage.

Another approach, and the one which I will adopt in this paper, is the Feldstein and Horioka (1980)assertion that a good measurement for capital mobility can be the interplay between domesticinvestment and saving. The logic behind the Feldstein–Horioka (FH) model is that for a small openeconomy which is very strongly integrated with the international financial markets, the correlationbetween domestic investment and saving rates will be minimal, if not completely zero. This occursbecause regardless of how much the population decides to save, investment rates are determined onthe global markets and are ‘‘imported’’ home.

More broadly speaking, an economy for which there exists a perfect co-variation betweenaggregate investment and saving does not exhibit signs of capital mobility because the country seemsto be investing only the capital which has been accumulated within its own borders. If there is asubstantial gap between domestic investment and saving, then the country is allowing capital to flowin from abroad, permitting the accounting mismatch. Of course, in a very long run, it is fiscallyimpossible to continue to maintain that gap, since nations must meet the basic solvency constraint.However, over specific periods of time, it is interesting to observe whether certain countries or regionsshow any significant tendencies toward capital mobility or financial closeness.

In general, the discussion on saving and investment is important in its own right, even if I discard thecapital mobility prism. There is little doubt in the conventional belief that domestic saving is animportant factor of economic growth. Even the most basic of all economic growth models show thatdecreasing saving rates have a negative effect on the gross domestic product. The theoretical relationshipbetween saving and growth builds on the long-run congruence of the investment and saving rates. Risingdomestic saving improves the investment climate, which in turn spurs the financial dynamic andultimately leads to higher income growth. A natural response, therefore, is to try to estimate therelationship between saving and investment. Not only for the sole purposes of description andobservation, but in order to derive important policy-relevant recommendations and implications.

This paper will compute the saving-investment relationship, as a measurement of capital mobility,for six countries of the Caucasus. These are Azerbaijan, Armenia, Georgia, Russia, Kazakhstan, andTurkey. Although Turkey is sometimes geographically excluded from the Caucasus, its economic tiesto the region are substantial enough for me to decide to include it in this analysis. To some extent, thesame logic applies to the inclusion of Russia and Kazakhstan. While Azerbaijan, Armenia, and Georgiaconstitute the Caucasian ‘‘core’’, often referred to as the ‘‘Southern Caucasus’’, I wish to extend thediscussion by bringing Russia, Kazakhstan, and Turkey (all with significant economic ties with theregion) into the picture.

The Feldstein–Horioka model has been estimated very extensively, using a plethora of variouseconometric methods, with datasets ranging from OECD countries to the sub-Saharan states. Themodel itself is quite straightforward and takes the form of the following regression equation:

I

Y

� �t

¼ b0 þ b1

S

Y

� �t

þ et (1)

in which I/Y is the ratio of aggregate domestic investment over GDP, S/Y is ratio of aggregate domesticsaving over GDP, and et is the error term.

The purpose of the FH regression is to estimate the saving retention ratio �b1. In an extreme case ofcomplete financial autarky, the coefficient is equal to unity. This scenario implies that all investmentthat occurs in the country of discussion comes from domestically accumulated savings, as no capital atall arrives from abroad. In an alternative extreme case of perfect capital mobility, the coefficient would

R. Jamilov / Economic Systems 37 (2013) 155–170 157

be zero. This case suggests, rather intuitively, that the country is financing its domestic investmentoperations only with foreign funds. Both extremes are rare, as theory would predict most economies todemonstrate a saving coefficient of between 0 and 1.

Feldstein and Horioka estimated Eq. (1) using long-run averages of cross-sectional regressions ofinvestment over saving for the 16 countries of the OECD. They achieved rather puzzling results – veryhigh coefficients for the investment–saving relationship, which effectively signaled low capitalmobility in the OECD. These results, with additions from studies on developing countries, gave birth tothe notion of the so-called ‘‘Feldstein–Horioka’’ puzzle – high saving retention ratios for developednations (indicating low capital mobility) and small coefficients for the developing states (suggestingfinancial openness and free flow of capital). Economic logic and conventional theory would predictexactly opposite empirical conclusions. This inconsistency was eventually called the ‘‘mother of allpuzzles’’ (Obstfeld and Rogoff, 2000).

The reason the FH model has been so popular is that it is pretty simple to estimate. The economiclogic of the model is straightforward and elegant, with numerous ways to interpret the results forpolicy purposes. While the interplay of investment and saving opens the door to vivid intellectualdebates, there is a natural limit to what the model can offer: in its original set-up, the equation has justone independent variable. There is not much one can do with the original model without extending itto a general equilibrium framework or adding additional regressors. The only reasonable innovationsthat one can hope for in this stream are either to study the countries and regions that have not beenexamined before or to experiment with new and more sophisticated econometric methods. It appearsto me that this is indeed the narrative of the FH model: as economies in transition continue to breakinto the international economic and financial arena, we gain empirical access to broader markets andare therefore able to re-apply the seminal model to newer blocks of countries. Also, as the econometricliterature progresses, newer techniques allow us to learn more and more about the nuances andshortfalls of the original model and to overcome the limitations with state-of-the-art modeling. TheFH research stream has really taken a wild empirical journey from simple OLS regressions to complexequilibrium models and panel cointegration techniques. However, there seems to be a technical aswell as a geographical ceiling on how long this stream may continue.

This paper effectively follows the described narrative: I attempt to apply the most recent advancesin econometrics while simultaneously extending the analysis to a new region. While the methodsemployed in this paper are indeed recent and proven, there have already been quite a few researcherswho experimented with panel cointegration. The true distinction of this study is that, to the best of myknowledge, no previous work has been done on the analysis of capital mobility for the Caucasus as adistinct group. Thus, this paper is important first and foremost as a precedent-setter for capitalmobility analysis in the Caucasus, and secondarily as another application of advanced panelcointegration estimation.

While Montiel (1994) has estimated the saving ratio for Turkey to be 0.47, the saving retentioncoefficients for Armenia, Azerbaijan, Georgia, Kazakhstan and perhaps even for Russia will be obtainedfor the very first time. In general, developing countries should demonstrate some forms ofinfrastructural imperfections and thus low capital mobility. I therefore base my expectations on thisassumption and expect the coefficients of mobility to be closer to unity than to zero.

From a bigger picture, this paper aspires to put quantified results on the tables of regionalpolicymakers as countries of the Caucasus continue with their development strategies. It is extremelydesirable that this study will spark a new stream of research on this region, as very little empiricalwork has been done on the nations of the Caucasus until now. In addition, there have not been awfullymany applications of panel data econometrics toward this topic, which is flooded with cross-sectionand time-series papers. Therefore, our application of panel cointegration will also provide someadditional variety in the estimation methodologies that this sphere requires.

2. Literature review

Some cross-country studies of the FH model include Feldstein (1983), Feldstein and Bachetta(1991), Tesar (1991), Artis and Bayoumi (1992), and Coakley et al. (1996). There exists a strikingheterogeneity in the cross-section numbers, with the saving ratios ranging from as low as 0.42 in

R. Jamilov / Economic Systems 37 (2013) 155–170158

Obstfeld (1986) to as high as 0.85 in Murphy (1984). Most of the cross-sectional empirical work,especially in the 1980s, was performed over samples of developed economies, such as the OECD states.Results seem to confirm the existence of the Feldstein-Horioka puzzle.

There are quite a few notable time-series estimations of the FH model. These studies include papersby Miller (1988), Alexakis and Apergis (1992), Tesar (1993), and Afxentiou and Serlitis (1993). Thepeculiarity of the time-series method is in the ability to estimate the FH estimates for individualcountries directly. However, the time-series approach suffers from the low power of times-series unitroot tests and from variable endogeneity.

It is also possible to evaluate the performance of economic regions, or groups of countries, with theimplementation of a panel setting. A panel data approach toward FH estimation is a relatively newdevelopment in this research stream, although there are several notable studies. These includeAmirkhalkhali and Dar (1993), Coakley et al. (1999), Jansen (2000), Kim (2001), Coakley et al. (2004),and Cyrille (2010). The relationship between standard cross-section, panel-data, and time-seriesestimates are described well in Pesaran and Smith (1995).

The most recent advances in econometrics enable us both to account for short-term variation inaggregate saving and to provide a long-run equilibrating equation of the investment-savingrelationship. Like the majority of studies that deal with the panel-data approach, this paper workswith panel sets suffering from heterogeneity of the country components. In this paper I have adoptedthree distinct but similar panel cointegration methods that will solve the econometric hindrances andprovide us with robust saving retention ratios.

The key assumption behind the panel cointegration approach is that the standard OLS applicationproduces biased estimates of the saving retention ratio, because both investment and saving sufferfrom the spurious regression problem. Elimination of this bias is possible by allowing for fixed effectsamong different panel members to be heterogeneous, while simultaneously permitting the short-rundynamics. Panel cointegration methods eliminate regressor endogeneity and serial correlation in theresiduals, the problem present in most long-run cointegration estimations. The estimates computedthis way also tend to be asymptotically unbiased. While there are quite a few tweaked approachestoward panel cointegration, I highlight the three most fundamental methods that I have decided toapply: Dynamic OLS (DOLS), Fully Modified OLS (FMOLS), and Pooled Mean Group (PMG) estimations.

Bahmani-Oskooee and Chakrabarti (2005) have applied the FMOLS technique in their study of alarge heterogeneous panel of 126 countries. They provide useful decompositions of their global panelinto groups of economies which vary by degree of openness, income status, oil production, etc.Bangake and Eggoh (2011) employed the DOLS, FMOLS, and PMG methods in their analysis of Africaneconomies. Murthy (2009) performed a panel cointegration analysis for Latin American and Caribbeancountries. Kim et al. (2005) apply the FMOLS panel cointegration technique for 11 Asian economies.For excellent literature reviews on capital mobility and the FH model, consult Coakley et al. (1998) andApergis and Tsoumas (2009).

On a slightly digressing note, it really is remarkable that despite the abundance of studies there isstill no definitive conclusion on what the FH coefficient actually should be. Research has producedextremely varying results, and there is no unifying structural approach as to what the optimal degreeof capital mobility is in either developing or developed economies. There is a substantial gap betweenempirical findings and predictions of the theory: highly developed countries sometimes empiricallydemonstrate surprisingly high FH coefficients, suggesting low capital mobility levels, while emergingand transition economies are often supposedly very open to capital and financial flows. I think thatsince theory has basically failed at consistently predicting empirical outcomes of capital mobilityestimations, this is a fundamentally empirical and contextual issue, highly vulnerable to the choice ofeconometric method, the region of analysis, and the time frame of focus. Changing one or more ofthese three parameters can produce strikingly varying results. I believe that capital mobility literaturewill probably never engineer a final, concluding note, but will instead diverge into clusters of studieswhich differ by method and geography.

The remainder of the paper is structured as follows. Section 3 describes the data and explains theeconometric methods applied in this study. Section 4 reports the panel cointegration results. Section 5compares our outcomes with the official rankings of market openness, puts our results in aninternational context, and discusses various policy-relevant issues. Finally, Section 6 concludes.

R. Jamilov / Economic Systems 37 (2013) 155–170 159

3. Econometric methods and data

3.1. Cross-dependency tests

Before testing for unit roots in the variables, I perform quick preliminary tests for cross-sectionaldependence. It is widely regarded that cross-sectional dependence would distort the conventionalpanel unit root and cointegration tests, deeming them irrelevant. If cross-dependence is indeedproven, we need to apply cross-dependence augmented tests like the Pesaran CIPS test and theWesterlund error-correction based cointegration test (Pesaran, 2006; Westerlund, 2007). In order totest for cross-sectional dependence I first perform the Breusch-Pagan test (Breusch and Pagan, 1980).It is based on a Lagrange Multiplier (LM) statistic for testing the null of zero cross equation errorcorrelations. The test statistic could be represented as follows:

CDlm ¼ TXN¼1

i¼1

XN

j¼iþ1

pi j (2)

where pi j is the sample estimate of the pair-wise correlation of residuals.In addition, and more relevantly to our case, I perform the Pesaran Cross-Dependence test (CD),

which is more appropriate to small samples when the T dimension is small (Pesaran, 2004). Unlike theBreusch and Pagan LM test statistic, the Pesaran statistic has exactly mean zero for fixed values of T andN under a wide array of panel data models, including heterogeneous dynamic models. The applicationof this test to our case of a heterogeneous panel is particularly relevant. The Pesaran CD test can beformulated as follows1:

CD ¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi

2T

N N � 1ð Þ

s XN¼1

i¼1

XN

j¼iþ1

pi j

0@

1A (3)

3.2. Panel unit root and cointegration tests

Having tested for cross-sectional dependence, and assuming that we fail to reject the nullhypothesis of cross-sectional independence, I proceed with unit root and cointegration testing. For thepanel cointegration method employed in this study, the core assumption of the process is thatinvestment and saving both possess a unit root. In order to determine that this is indeed the case, Iemploy the Im-Pesaran-Shin (IPS) panel unit root test as proposed in Im et al. (2003). The IPS test willbe constructed on the base of a group mean of the individual member unit root tests. I must estimatethe following equations to perform the test:

DI

Y

� �it

¼ m1i þ l1iI

Y

� �i;t�1

þX

k

u1i;kDI

Y

� �i;t�k

þ d1it þ e1it (4a)

� � � � X � �

D

S

Y it

¼ m2i þ l2iS

Y i;t�1

þk

u2i;kDS

Y i;t�k

þ d2it þ e2it (4b)

where mi are the individual fixed effects; li are corrected for inter-temporal serial correlation.The IPS test statistic is computed based on the average of the individual ADF statistics over cross-

sections (Maddala and Wu, 1999). The null hypothesis of the test is the presence of a unit root in theseries. The following equation is used to estimate the IPS statistic:

t j ¼1

N

Xi

l ji

sl ji

(5)

1 For a more thorough technical discussion of both the Breusch–Pagan and the Pesaran CD tests, consult Pesaran (2004).

R. Jamilov / Economic Systems 37 (2013) 155–170160

In addition to the IPS test, I will perform alternative panel unit root tests as described in Maddalaand Wu (1999). The Maddala–Wu proposition is a special case of the Fisher test (Fisher, 1932) and isbased on the combination of the p-values of the test-statistic for a unit root in each cross-sectionalentry. The MW test statistic is distributed as Chi-square with 2N degrees of freedom under theassumption of cross-sectional independence.2 This test does not depend on different lag lengthselections in the individual ADF regressions and is superior to the IPS test. The MW statistic is given bythe following3:

�2XN

i¼1

logðpiÞ ! x22N (6)

In the next step, assuming cross-sectional independence and non-stationarity of the variables, Idetermine if investment and saving are cointegrated. For a panel data set like ours, I perform acointegration test for heterogeneous panels as described in Pedroni (2004). The test takes into accountthe heterogeneity which is most certainly present in the model, asserting that the vectors ofcointegration are not identical. In order to perform the test, I must first estimate the following long-run cointegration relationship:

I

Y

� �it

¼ ai þ dit þ b1i

S

Y

� �1;it

þ b2i

S

Y

� �2;it

þ � � � þ bMi

S

Y

� �xM;it þ eit (7)

for i=1,. . ., N; t=1,. . ., T; m=1,. . ., M; and where ai – idiosyncratic fixed effect, different acrosscountries; bi – effect, common to all countries; ei – error term. The residuals that are estimated in (7)will therefore have the following nature:

eit ¼ rieit�1 þ uit (8)

Pedroni’s method presents seven different statistics that test panel data cointegration. Four of themare based on pooling, and are thus named the ‘‘within’’ dimension, while the last three are based on the‘‘between’’ dimension. For both types of tests, the null hypothesis is that the series are notcointegrated in the long run.

3.3. The FMOLS and DOLS estimators

After establishing the presence of an underlying long-run cointegrating equation between the twovariables in focus, let’s examine how we can estimate that equation using the method of panelcointegration. It is a known fact that the standard pooled OLS will suffer from the spurious regressionproblem if the variables are not stationary in level forms (Chen et al., 1999). Thus, the ordinary OLSapproach is basically irrelevant. For the estimation of panel cointegration, I have employed threestrategies which are considered to be superior to traditional OLS (Kao and Chiang, 2000). First, I discussthe usage of the Fully Modified OLS (FMOLS) and Dynamic OLS (DOLS) techniques and then the PooledMean Group (PMG) method.

The FMOLS approach takes care of the spurious regression problem which arises in standard OLSmodels in addition to correcting for serial correlation and the endogeneity of regressors (Pedroni,1999, 2000). In order to estimate the long-run cointegrating equation, we must first consider thefollowing panel regression:

I

Y

� �it

¼ ai þS

Y

� �it

b þ uit (9a)

for i=1,. . .N, t=1,. . .T, and where yit is a matrix (1,1), b is a vector of slopes, ai is the individual fixedeffect, and uit is the stationary disturbance. The vector xit is an integrated process of order one for all i,

2 This is precisely where the issue of cross-sectional independence becomes crucial, since it is the key assumption behind our

unit root testing. This validates our use of preliminary Breusch–Pagan and Pesaran cross-dependence tests.3 For more information on testing for unit roots in the time-series context, consult Dickey et al. (1986).

R. Jamilov / Economic Systems 37 (2013) 155–170 161

and is defined as follows:

I

Y

� �it

¼ S

Y

� �it�1

þ eit (9b)

Under the conditions outlined above, Eq. (8) depicts a system of cointegrating regressions. In otherwords, (I/Y)it is cointegrated with (S/Y)it. The FMOLS estimator itself is constructed in a way that thetwo chief problems associated with standard OLS estimation (endogeneity and serial correlation) areproperly corrected for. The estimator is defined as:

bFMOLS ¼ ðb�NT � bÞ ¼

XN

i¼1

L�2

22i

XT

i¼1

ðxit � xiÞ2 !�1XN

i¼1

L�1

11iL�1

22i

XT

t¼1

xit � xið Þm�it � Tg i

!(10)

where g i is the serial correlation correction term, m�it is the endogeneity correction, L11i and L22i are thelower triangular matrices of the scalar long-run variance of the residual mit and m�m long-runcovariance among the eit, respectively.

From the FMOLS panel estimator I can derive t-statistics which are asymptotically normallydistributed (Pedroni, 2004). The null hypothesis for such a t-test is H0: bi =b0 for all i versus thealternate hypothesis Ha: bi =ba 6¼b0; where b0 is some value for b under the null hypothesis, and ba

some alternative value for b which is homogeneous across all the members of the panel. The t-statisticcan be constructed using the following equation:

tb�NT¼ ðb

�NT � bÞ

XN

i¼1

L�2

22i

XT

i¼1

ðxit � xiÞ2 !�1=2

(11)

The Dynamic OLS (OLS) estimator is equally effective at eliminating the endogeneity and serialcorrelation issues. DOLS is deemed to be more powerful for small samples and tends to outperformFMOLS (Kao and Chiang, 2000). The DOLS estimator engineers a parametric adjustment to the errorterms by including the past and future values of the differenced regressors. This allows us to getunbiased long-run parameters. The following equation is needed to obtain the DOLS estimator:

yit ¼ ai þ xitb þXj¼q2

j¼�q1

ci jDxi;tþ j þ vit (12)

where cij is the coefficient of a lead or lag of first differenced explanatory variables. The DOLS estimatoritself is given by the following formulation:

bDOLS ¼XN

i¼1

XT

t¼1

zitz0it

!�1 XT

t¼1

zit yþit

!(13)

where zit ¼ ½xit � xi; Dxi;t�q; :::; Dxi;tþq� is the 2(q+1)�1 vector of regressors.

3.4. The PMG estimator

The third and final method that I employ to estimate the long-run cointegrating relationshipbetween saving and investment is the Pooled Mean Group (PMG) estimator described in Pesaran et al.(1999). The PMG is effectively an intermediate estimator between averaging and pooling theparameters in dynamic panel data models. The former permits the parameters to be freelyindependent across groups but does not allow for homogeneity between groups. The latter methodincludes the conventional random effects, fixed effects and GMM models, and restricts the parametersto be equal across groups, while the intercept may be different between groups.4 The PMG, being themiddle ground between averaging and pooling, has an advantage over both the FMOLS and DOLS

R. Jamilov / Economic Systems 37 (2013) 155–170162

methods in its capacity to allow short-run dynamic behavior to be heterogeneous across countrieswhile constraining the long-run coefficients to be the same.

Following Pesaran et al. (1999), the PMG estimator can be obtained by estimating the followingauto-regressive distributed lag model (ARDL):

yit ¼Xp

j¼1

li jyi;t� j þXq

j¼0

d0i jxi;t� j þ mi þ eit (14)

where xit (k�1) is the vector of explanatory variables for group i, mi is the fixed effect, lij are the scalarrepresentation of the coefficients of the lagged dependent variables, and dij are k�1 coefficientvectors. Eq. (14) can then be re-written in the following form:

Dyit ¼ fiyi;t�1 þ b0ixit þXp�1

j¼1

l�i jDyi;t� j þXq�1

j¼0

d0i j�Dxi;t� j þ mi þ eit (15)

where i=1, 2,. . ., N, and t=1, 2,. . ., T. fi ¼ �ð1 �P p

j¼1 li jÞ, bi ¼Pq

j¼0 di j, l�i j ¼ �P p

m¼ jþ1 lim for j=1,2,. . ., p�1, and d�i j ¼ �

Pqm¼ jþ1 dim for j=1, 2,. . ., q�1.

The model operates on the assumptions that disturbances eit are independently distributed across i

and t, that the roots ofPp

j¼1 li jzj, i=1, 2,. . ., N, lie outside the unit circle, and that the long-run

coefficients of the regressors, defined by ui =�bi/fi, are identical across the groups (the so-called long-run homogeneity assumption).

The PMG estimator refers to the estimation of the long-run coefficients ui. The method with whichwe estimate the long-run coefficients is the Maximum Likelihood (ML) technique. The ML estimationof ui pools the long-run coefficients by applying the homogeneity restriction and averages the meansof the estimated short-run parameters of the model. It should thus be clear why the method has beennamed ‘‘pooled mean group’’. The ML parametric form of Eq. (15) is maximized with respect to v,where v=(u0, f0, s0), f=(f1, f2, ..., fN)0, s ¼ ðs2

1; s22; :::; s2

NÞ0.

3.5. Parameter homogeneity test

Finally, after estimating the saving retention ratio for the Caucasus using the panel cointegrationtechniques described above, I will look at the nature of the obtained coefficients. Specifically, I want totest if the parameters are hetereogeneous even after accounting for individual fixed effects whichdiffer across countries. In other words, the question is whether the convergence of individual countryparameters is consistent with a single saving-investment function. Failing to reject the hypothesis ofheterogeneity would imply that the saving-investment relationship behaves heterogeneously as afunction, and each country in our analysis has internally different investment and saving dynamics. Totest for parameter heterogeneity I follow Pedroni (2007) and employ an F-test for heterogeneouscoefficients. The test produces a Wald statistic which compares the sum of squared errors for therestricted case when b1 =b for all i against the case with unrestricted heterogeneous bi values. The nullhypothesis is parameter homogeneity, H0: b1 =b for all i.

3.6. Data description

In this study I use annual data for gross capital formation and gross domestic saving, both as apercentage of GDP. The former is referred to as ‘‘investment’’ and the latter as ‘‘saving’’. The numberswere obtained from the World Bank. I use gross and not net saving flows because it is precisely thegross saving which responds to the investment and saving yield differentials. The time period in thestudy spans from 1996 to 2010. The countries in the sample are the conventionally accepted membersof the Caucasus – Armenia, Azerbaijan, Georgia, Kazakhstan, Russia, and Turkey.

4 Consult Pesaran et al. (1999) for a more in-depth discussion of dynamic panel data models in general and the pooled method

in particular.

Table 1Cross-dependency tests.a

Statistic p-Value

CDlm 1.207281 0.227324

CD 1.170357 0.241857a CDlm refers to the Breusch–Pagan LM test and CD to the Pesaran test for cross-dependency. H0: cross-sectional independence.

Table 3Panel cointegration test.

Test statistic Statistic p-Value

Within dimension

Panel v-statistic 1.110013 0.1335

Panel rho-statistic �0.274966 0.3917

Panel PP-statistic �0.450769 0.3261

Panel ADF-statistic �3.390061 0.0003**

Between dimension

Group rho-statistic 0.363219 0.6418

Group PP-statistic �0.941604 0.1732

Group ADF-statistic �2.731586 0.0032**

** Rejects the null hypothesis of no cointegration at the 1% significance level.

Table 2Panel unit root tests.

Variable IPS ADF PP

Statistic p-Value Statistic p-Value Statistic p-Value

Investment �1.65252 0.0492* 19.9404 0.0682** 13.7003 0.3203**

D (investment) �2.97461 0.0015^ 30.8711 0.0021^ 40.1900 0.0001^

Saving �0.62890 0.2647** 14.0696 0.2963** 8.19991** 0.7693**

D (saving) �3.55525 0.0002^ 35.0200 0.0005^ 49.9657 0.0000^

^ Rejects the null hypothesis of unit root process in first-differences at the 1% significance level.* Cannot reject the null hypothesis of unit root process in stationary form at the 1% significance level.** Cannot reject the null hypothesis of unit root process in stationary form at the 5% significance level.

R. Jamilov / Economic Systems 37 (2013) 155–170 163

4. Results

I begin to report the results by first demonstrating the outcome of the Cross-Dependency (CD) tests.Table 1 presents the test statistics and p-values of the Breusch–Pagan and Pesaran tests for cross-sectional dependence. Both tests fail to reject the null hypothesis of cross-sectional independence atthe accepted levels of statistical significance. We can thus proceed with the conventional tests for unitroots and cointegration.

Table 2 reports the outcomes from the panel unit root tests. As described in Section 3 oneconometric methods, I have performed the IPS and the ADF-Fisher and PP-Fisher unit root tests. Allthree methods uniformly do not reject the null hypothesis of unit root in levels, but reject it in firstdifferences. This shows that investment and saving are non-stationary. Regressions with suchvariables produce a spuriousness problem, and we need to account for this. Our panel data estimationswill take care of this problem.

Having established the non-stationary nature of investment and saving, I try to determine whetherour two variables are cointegrated, i.e. form a single long-run equilibrating relationship. Table 3 hasthe results from the Pedroni panel cointegration test. All seven statistics are presented. Small samplesizes, like in our case, usually make the rejection of the null of cointegration more difficult (Montiel,1994). Based on the results of the inferentially more powerful ADF statistics of both the within- and

Table 4Cointegration estimates for individual countries.

FMOLS DOLS PMG

Estimate t-stat Estimate t-stat Estimate t-stat

Armenia 1.25 0.86 1.10 0.44 0.20 1.78*

Azerbaijan �0.19 �2.16* 0.84 �0.21 0.89 4.29**

Georgia 1.17 0.62 1.14 0.48 0.29 4.26**

Kazakhstan 1.61 2.04* 1.97 5.68** 1.22 19.51**

Russia 0.21 �1.79* �0.02 �4.04** 1.49 22.39**

Turkey 0.32 �2.60** 0.40 �1.52 0.87 23.71**

Caucasus 0.73 �0.83** 0.90 0.34** 0.83 18.76**

* Rejects the null hypothesis of no cointegration at the 5% significance level.** Rejects the null hypothesis of no cointegration at the 1% significance level.

R. Jamilov / Economic Systems 37 (2013) 155–170164

the between-dimension groups, which reject the null of no cointegration at the 1% confidence interval,I can now estimate the long-run cointegrating equation.5

Table 4 presents the results from our three panel cointegration estimation methods. This tablereports just the individual country figures. One lag has been chosen in the individual country set-upaccording to the Schwarz Bayesian criterion. First, I want to highlight the heterogeneity of the resultsacross the estimation methods. The numbers are not consistently significant, and the PMG methodtends to produce more robust results. I will therefore try to emphasize the results from the PMG forpolicy purposes, since drawing implications and conclusions from statistically non-significant figureswould not be enduring. Based on the statistically more powerful outcomes of the PMG method, weobserve signs of relatively mobile capital in Armenia and Georgia, with the rest of the regiondisplaying high saving retention ratios and thus a considerably low degree of capital mobility.

I also highlight that on several occasions, for example the PMG estimates of Kazakhstan and Russia,the saving coefficient is higher than unity, i.e. the response of investment overshoots the impulse fromsaving. Although theory would predict that the FH coefficient must lie strictly on the 0–1 spectrum ofpossible numbers, empirical studies have shown that this is not always the case. Notably, Coakley et al.(1994, 1995) have calculated the coefficient to be 1.019, 1.043, and 1.182 for Germany, Japan, and theCzech Republic, respectively. In addition, Obstfeld (1986) computed an estimate of 1.422 for theperiod of 1970–1979 for the original Feldstein-Horioka grouping of 16 OECD countries. It really is notclear what, if anything, is the commonality for the overshooting in the OECD countries, the Caucasus,Germany, and the Czech Republic. The targets are so geographically and economically heterogeneousthat rationalizing overshooting with a small economy, or resource-rich state, or advanced countryargument yields no systematic conclusion.

One possible solution to this puzzle would be to examine the same set of countries over multipleperiods of time. It is possible that a country or region is using a certain period t to return to its long-runsolvency condition by compensating the lower-than-accepted saving-investment coefficient in t-1

with a higher-than-unity coefficient today. In other words, it is only expected that economies remainsolvent in the long run, but the dynamic toward that long-run condition may vary from region toregion and country to country. It is probable that the higher-than-unity saving retention ratios aretemporary in the bigger picture, the long-run point of reference. This could be proven if some futurestudies were to examine the FH model for the Caucasus over, say, the 2010–2020 decade, and find thatthe coefficients drop substantially. In that case, the 1996–2010 period discussed in this study would beviewed merely as a sub-period on the path toward the long-run solvency requirement.

Table 5 delivers the estimates of the cointegrating equation for groups of countries. Similarly tothe individual country case, one lag is selected. Unlike our findings for individual member states ofthe Caucasus, the region as a whole exhibits quite homogeneous results across all three estimationmethods. I find that the long-run saving retention ratio for the Caucasus based on the FMOLS, DOLS,

5 I also performed the Kao Residual Cointegration Test (McCoskey and Kao, 1998), which firmly rejects the null of no

cointegration at the 1% significance level. The results of this test are omitted for brevity.

Table 5Cointegration estimates for panels.

FMOLS DOLS PMG

Estimate t-stat Estimate t-stat Estimate t-stat

South Caucasus (SC) 0.76 �0.16** 1.03 0.41** 0.46 5.63**

SC+Kazakhstan 0.99 0.90** 1.26 3.20** 0.65 10.28**

SC+Russia 0.66 �0.87** 0.76 �1.66** 0.72 11.29**

SC+Turkey 0.66 �1.31** 0.87 �0.40** 0.56 9.07**

Caucasus 0.73 �0.83** 0.90 0.34** 0.83 18.76**

** Rejects the null hypothesis of no cointegration at the 1% significance level.

Table 6Parameter heterogeneity test.

Statistic p-Value

F-test for heterogeneity 24 0.0002*

* Rejects the null hypothesis of homogeneous parameters at the 1% significance level.

R. Jamilov / Economic Systems 37 (2013) 155–170 165

and PMG methods is 0.73, 0.90, and 0.83, respectively. I also decompose the global panel into severalsub-panels. The saving retention ratio for the South Caucasus (SC), which includes only Armenia,Azerbaijan, and Georgia, is substantially different from the global panel in the PMG case (0.46 vs.0.73), while for the FMOLS and DOLS cases the estimates are quite similar (0.76 vs. 0.73 and 1.03 vs.0.90). Then, I selectively add Kazakhstan, Russia, and Turkey into a panel with the SC to allow for apolicy-friendly representation of how the region performs on the capital mobility criteriondepending on the changing composition.

I briefly highlight the observation that the addition of Turkey to the SC panel systematically lowersthe saving retention ratio across all three methods. This suggests that, compared to Kazakhstan andRussia, Turkey is a more benevolent factor for the financial openness and capital flexibility of theCaucasus. I do not wish to use this finding to speculate on the arguments of regional integration in theCaucasus, but this observation needs to be taken into account by the region’s policymakers.

Overall, for individual countries, we have seen some substantially heterogeneous results acrossdifferent estimation methods. For the case of the Caucasus as a group, however, the results are quitesimilar (0.73, 0.90, and 0.83). I do not find evidence for the Feldstein–Horioka puzzle (low saving-investment correlations for developing countries), since for a region which is still very much indevelopment and emergence the saving retention ratio is very high. Most of the country estimates forDOLS and FMOLS are not significant, and thus any policy conclusions for specific countries must bedrawn from the individual country PMG estimates. For the Caucasus as a whole all three methodsproduce homogenously high results, thus rendering a chance to uniformly claim capital immobility inthe Caucasus.

The presentation of results is concluded with a test for parameter heterogeneity. Table 6 providesthe estimates of the Wald statistic and the p-value of the F-test for heterogeneous parameters. The nullhypothesis of homogeneous parameters is rejected at the 1% significance level, suggesting that savingretention ratios are heterogeneous even after allowing for individual fixed effects to differ acrosscountries. This shows that the imposition of idiosyncratic fixed effects is not enough to establish acommon long-run saving-investment relationship for the Caucasus, implying that the interplaybetween saving and investment varies across the internal financial structures of the individualCaucasian member states.

I would like to comment briefly on the apparent heterogeneity in our country-specific estimates asnoted in Table 4.6 It is important to remember that the fundamental research question of the paper iswhether capital mobility exists in the Caucasus as a distinct economic region, not in any individualcountry in particular. This question is answered in a rather conclusive manner, since our panel

Table 7Comparison of capital mobility estimates.

Region Estimate Method Source

Caucasus 0.73 FMOLS Jamilov (2013)a

OECD 0.69 MLE Coakley et al. (1994), Kim (2001)

LDCs 0.44 Time-Series Coakley et al. (1999)

Asia 0.39 FMOLS Kim et al. (2005)

Africa 0.38 FMOLS Bangake and Eggoh (2011)

Low-Income 0.58 FMOLS Bahmani-Oskooee and Chakrabarti (2005)

Middle-Income 0.54 FMOLS Bahmani-Oskooee and Chakrabarti (2005)a For ease of comparison with other studies, only the FMOLS estimate from this paper is included in the table. Results of the DOLS and

PMG methods are 0.90 and 0.83, respectively.

R. Jamilov / Economic Systems 37 (2013) 155–170166

estimates are all homogeneously high, in fact larger than 0.7, and create an easy case for capitalimmobility. Heterogeneity in the single-country estimates is not very severe if we consider only thestatistically significant estimates, and the heterogeneity is present only in selected countries of thewhole group. Besides, we advise to focus single-country analysis on the PMG method since it fits ourpurposes by econometric design. Finally, as discussed in greater detail in the following section, wehighlight our application of the economic openness ranking from the Index of Economic Freedom. Ourestimates are actually perfect at predicting the IEF rank, which is a clear sign of passing the real-liferobustness check for validity and relevance.

5. Discussion

In the previous section, we have concluded that the saving retention ratio is large and close to unityin the case of the Caucasus. However, in order to draw policy implications, we need to compare theregion with other notable geographic and economic groupings. Table 7 puts the results of this paper inan international perspective. A quick examination of the table makes it clear that the Caucasus has aconsiderably low degree of capital mobility in an international context. It is particularly puzzling toobserve that the Caucasus ranks substantially lower than either Africa or Asia. Of course, our resultswould change if we altered the definition of the ‘‘Caucasus’’, i.e. by adding or subtracting a country fromthe global panel. However, for the given definition of this region, the Caucasus is indeed financiallyrestrained. As stated in the introductory section, I expected the saving retention ratio to be high for theCaucasus, reflecting the nature of its emerging and developing status. However, it is surprising to seethe Caucasus rank lowest among all the regions and groups that it is compared to in Table 7.

I now compare my results with the market openness category of the Heritage Foundation’s Index ofEconomic Freedom (IEF).7 Table 8 provides the country rating for the IEF’s category of marketopenness and the rank based on the saving-retention coefficients obtained in this paper. I find thatthere is a great degree of correlation between my capital mobility estimates and the IEF ranking.Indeed, the correlation is 100% perfect because if we rank our six member states of the Caucasus basedon their individual PMG capital mobility coefficients, then the ranking we get is exactly equal to thatproposed by the IEF. Assuming that the IEF is in itself a good proxy for market openness and financialflexibility, we can thus claim that the IEF method serves as a good predictor of individual country FHcoefficients. On the other hand, assuming that my own results are robust, panel cointegration (or atleast the PMG technique) can be a good predictor of overall financial and market openness. The keytake-away point from this discussion can be that the IEF is a good robustness check for capital mobilityestimation.

Based on the results of this study I can now draw some interesting policy-relevant inferences. First,it is rather obvious that our substantially high saving-investment correlation reflects a low degree offinancial openness in the Caucasus and its poor integration with the world’s capital markets (Sarnoand Taylor, 1998). This, in fact, is the original cornerstone logic of the whole FH research stream.

6 I am thankful to the anonymous referee for pointing this out.7 The Index is readily available in HTML and book formats from the IEF’s website: http://www.heritage.org/index/.

Table 8Index of economic freedom.

Country Trade

freedom

Investment

freedom

Financial

freedom

Average score

and openness rank

Capital

mobility rank*

Armenia 85.4 75 70 76.8 (1) 1

Azerbaijan 77.2 55 40 57.4 (4) 4

Georgia 89.2 70 60 73.0 (2) 2

Russia 68.2 25 40 44.4 (6) 6

Kazakhstan 79.6 30 50 53.2 (5) 5

Turkey 85.4 70 60 71.8 (3) 3* Higher rank indicates a smaller PMG coefficient obtained in this paper.

R. Jamilov / Economic Systems 37 (2013) 155–170 167

Second, our results shed some empirical light on the age-old discussion of the impossible policytriangle. Recent research has shown that most emerging market economies are showing signs ofconvergence toward a ‘‘middle ground’’ of a semi-fixed exchange rate, somewhat restricted capitalflows, and strong although incomplete monetary independence (Aizenman and Ito, 2012). Low capitalmobility in the Caucasus, as evidenced by our high saving-investment coefficient, suggests thatgovernments in this region have their hands untied with respect to both monetary independence andexchange rate management. Without capital flexibility, policy-makers in the Caucasus are believed tobe following a fixed or at least a semi-fixed exchange rate regime. It also follows that Caucasian statesenjoy the privileges of considerably independent monetary policy-making.

Third, Barro et al. (1995) have suggested that the relationship between aggregate domestic savingand investment does not necessarily relate to the financial or even physical capital. Instead, theyargue, the saving retention ratio demonstrates the degree of human capital mobility. Using modifiedRamsey growth models, they show that a high correlation between saving and investment suggeststhat human capital is inflexible and its cross-border penetration is jeopardized. Effectively this meansthat, assuming Barro’s analysis is correct, human capital mobility is very low in the Caucasus.

Fourth, the high saving-investment correlation can reflect a home-country bias in portfolioselection (Feldstein, 1994). There is a variety of potential risks that could prevent domestic investorsfrom purchasing financial assets from abroad. The most prominent of these are political and currencyrisks. Due to these and any other preventative obstacles, investors hold a disproportionally high shareof domestic securities, thus forgoing international portfolio diversification in favor of the apparentlyless risky domestic holdings. When applied to the case of the Caucasus, the home-bias argumentsuggests that this region presents quite substantial political and/or currency risks for purchasingfinancial assets overseas.

Fifth, national governments typically have two substantial policy tools at their disposal – monetaryand fiscal. A policy design targeted specifically at either investment or saving could affect the interplaybetween the two aggregates in an unexpected way. In other words, the relationship would suffer froma peculiar form of omitted variable bias, i.e. government policy, which would distort an otherwisemarket-established equilibrium (Summers, 1988). Thus the endogeneity in domestic savings andinvestment; the regressors are correlated with the policy response variable sitting in the error term. Itis therefore possible that with the right policy mix, the saving retention ratio would stay highregardless of the dynamics in the financial environment (Bayounmi, 1990). In the case of the Caucasus,given the high FH coefficient, I suggest the possibility that national governments design specificpolicies that target their respective domestic current accounts. Current account targeting, in turn,leads to a disproportionally high saving retention ratio.

Sixth, the saving-investment correlation may be an even better indicator for solvency manage-ment. For any economy, exposure to international capital markets presumably implies some form oflong-run solvency constraint. The degree to which countries respect this constraint in the short rundepends on their policy design, deficit buffers and financial infrastructure, cross-border trading risks,portfolio preferences, etc. A high saving-investment relationship could therefore be a signal for astrong solvency imposition (Coakley et al., 1996). I believe that solvency pressures along with currentaccount targeting are the best explanations for the high saving retention ratio that is observed in theCaucasus.

R. Jamilov / Economic Systems 37 (2013) 155–170168

Seventh, the two-country general equilibrium model of Tesar (1993) with utility maximizingagents, two types of goods (traded and non-traded), and two types of exogenous shocks to non-tradedgoods (investment and productivity), shows that economic agents change their preference for tradedand non-traded goods in response to a sudden exogenous shock. High saving-investment correlations,like in the case of the Caucasus, can reflect a poor degree of consumption smoothing in response toexogenous productivity shocks. Considering the prevalence of the non-traded sector of the energyexporting economies of Azerbaijan, Russia, and Kazakhstan, the question of consumption smoothingmight prove to be worth further research and exploration.

Finally, Kaplan and Kalyoncu (2011) claim that capital mobility and the saving-investmentrelationship can reflect the country’s crisis stabilization mechanisms. High levels of capital mobilitycan signal large exposure to foreign incoming capital, which stabilizes the economy during economicslumps. Low mobility, on the other hand, mirrors a restriction on the free movement of capital that hasallowed the country to withstand the negative foreign shocks and ‘‘isolate’’ itself from the crisisoccurring elsewhere. For the Caucasus with its very large saving retention ratio I can conclude thatfinancial closeness and capital inflexibility have not been too detrimental for the region with respectto surviving through the 2008 Global Financial Crisis.

6. Conclusion

This paper has applied three recently developed techniques in heterogeneous panel cointegrationto the analysis of saving and investment dynamics for the countries of the Caucasus. All threetechniques have computed a considerably high, positive relationship between savings andinvestment. These results provide strong empirical evidence for the lack of capital mobility in theCaucasus. Putting the numbers in an international context gives even more support to our propositionas the Caucasus falls noticeably behind Asia, Africa, and middle-income countries. I perform arobustness check by comparing this paper’s estimates with the Index of Economic Freedom and find aperfect correlation between the two sets of results. My findings imply that the Caucasus is integratedpoorly with the global financial markets; policymakers possess monetary independence and conductfixed or semi-fixed monetary policy; mobility of human capital is quite low; there exists a substantialhome-country portfolio formation bias due to high political and currency risks of holding overseasfinancial assets; national governments design current account targeting policies that indirectly affectthe saving-investment interplay; financial markets impose a strong long-run solvency constraint; theconsumption smoothing reaction to exogenous production shocks is poor; capital immobility hashelped the region go through the recent recession relatively undamaged.

Acknowledgements

I would like to thank the editor, Richard Frensch, and two anonymous referees for valuablecomments. I gratefully acknowledge technical advice from Peter Pedroni. I am also thankful to HasinulSiddique, Cem Tintin and all the participants of the First Istanbul International Conference on Businessand Economics for fruitful discussions. The opinions presented in this paper are solely my own and donot necessarily reflect the views of the Central Bank of the Republic of Azerbaijan.

References

Afxentiou, P.C., Serlitis, A., 1993. International capital mobility and the long run investment and saving in Canada. EconomiaInternazionale 46, 147–167.

Alexakis, A., Apergis, S., 1992. The Feldstein–Horioka puzzle and exchange rate regimes: Evidence from cointegration tests.Journal of Policy Modeling 16, 459–472.

Amirkhalkhali, S., Dar, A.A., 1993. Testing for capital mobility: a random coefficients approach. Empirical Economics 18,523–541.

Aizenman, J., Ito, H., 2012. Trilemma policy convergence patterns and output volatility. NBER Working Paper 17806, Cambridge,MA.

Apergis, N., Tsoumas, C., 2009. A survey on the Feldstein–Horioka puzzle: what has been done and where we stand. Research inEconomics 63, 64–76.

R. Jamilov / Economic Systems 37 (2013) 155–170 169

Artis, M.J., Bayoumi, T., 1992. Global capital market integration and the current account. In: Taylor, M.P. (Ed.), Money andFinancial Markets. Blackwell, Cambridge, MA and Oxford, pp. 297–307.

Bangake, C., Eggoh, J., 2011. The Feldstein–Horioka puzzle in African countries: a panel cointegration analysis. EconomicModelling 28, 939–947.

Bahmani-Oskooee, M., Chakrabarti, A., 2005. Openness, size, and the saving-investment relationship. Economic Systems 29,283–293.

Barro, R.J., Mankiw, N.G., Sala-I-Martin, X., 1995. Capital mobility in neoclassical models of growth. American Economic Review85, 103–115.

Bayounmi, T., 1990. Saving-investment correlations: immobile capital, government policy, or endogenous behavior? IMF StaffPapers 37, 350–387.

Breusch, T.S., Pagan, A.R., 1980. The Lagrange multiplier test and its applications to model specification in econometrics. Reviewof Economic Studies 47, 239–253.

Caprio, G., Jr., David, H.H., 1984. Domestic Saving, Current Accounts, and International Capital Mobility. International FinanceDiscussion Paper 244, Washington, DC.

Calvo, G., Leiderman, L., Reinhart, C., 1992. Capital inflows and real exchange rate appreciation in Latin America: the role ofexternal factors. IMF Working Paper 92/62, Washington, DC.

Chen, B., McCoskey, S., Kao, C., 1999. Estimation and inference of a cointegrated regression in panel data: a Monte Carlo study.American Journal of Mathematical and Management Sciences 19, 75–114.

Coakley, J., Kulasi, F., Smith, R., 1994. The saving-investment association. Birkbeck College Working Paper 10:94, London.Coakley, J., Kulasi, F., Smith, R., 1995. The Relationship Between Cross Section, Time Series and Other Estimators of the FH

Regression. Mimeo, Birkbeck College, London.Coakley, J., Hasan, F., Smith, R., 1996. Current account solvency and the Feldstein-Horioka puzzle. Economic Journal 106,

620–627.Coakley, J., Kulasi, F., Smith, R., 1998. The Feldstein–Horioka puzzle and capital mobility: a review. International Journal of

Finance and Economics 3, 169–188.Coakley, J., Hasan, F., Smith, R., 1999. Saving, investment, and capital mobility in LDCs. Review of International Economics 7,

632–640.Coakley, J., Fouertes, A., Spagnolo, F., 2004. Is the Feldstein–Horioka puzzle history? Manchester School 72, 569–590.Cyrille, S.M., 2010. Saving-investment correlation and capital mobility in sub-Saharan African countries: a reappraisal through

inward and outward capital flows’ correlation. International Journal of Economics and Finance 2, 97–106.Dickey, D.A., Bell, W.R., Miller, R.B., 1986. Unit roots in time series models: tests and implications. American Statistician 40,

12–26.Easterly, W.R., Mauro, P., Schmidt-Hebbel, K., 1995. Money demand and seigniorage maximizing inflation, Journal of Money.

Credit, and Banking 27, 583–603.Faruqee, H., 1991. Dynamic capital mobility in Pacific Basin developing countries: Estimation and policy implications. IMF

Working Paper 91/115, Washington, DC.Feldstein, M., 1983. Domestic saving and international capital movements in the long run and the short run. European Economic

Review 21, 129–151.Feldstein, M., Horioka, C., 1980. Domestic saving and international capital flows. Economic Journal 90, 314–329.Feldstein, M., Bachetta, P., 1991. National saving and international investment. In: Bernheim, D., Shoven, J. (Eds.), National

Saving and Economic Performance. University of Chicago Press, Chicago, IL, pp. 201–226.Feldstein, M., 1994. Tax policy and international capital flows. Review of World Economics 130, 675–697.Fisher, R.A., 1932. Statistical Methods for Research Workers, fourth ed. Oliver & Boyd, Edinburgh.Haque, N., Montiel, P., 1991. Capital mobility in developing countries: some empirical tests. World Development 19,

1391–1398.Im, K.S., Pesaran, M.H., Shin, Y., 2003. Testing for unit roots in heterogeneous panels. Journal of Econometrics 115, 53–74.Jansen, W.J., 2000. International capital mobility: evidence from panel data. Journal of International Money Finance 19,

507–511.Kao, C., Chiang, M.H., 2000. On the estimation and inference of a cointegrated regression in panel data. Advances Econometrics

15, 179–222.Kim, H., 2001. The saving-investment correlation puzzle is still a puzzle. Journal of International Money and Finance 20,

1017–1034.Kim, H., Oh, K.-Y., Jeong, C.W., 2005. Panel cointegration results on international capital mobility in Asian Economies. Journal of

International Money and Finance 24, 71–82.Kaplan, M., Kalyoncu, H., 2011. Measuring the level of international capital mobility for MENA countries. Journal of Econonomic

and Social Studies 1, 25–32.Levich, R.M., 1985. Empirical studies of exchange rates: price behavior, rate determination and market efficiency. In: Jones,

R.W., Kenen, P.B. (Eds.), Handbook of International Economics, vol. II. Elsevier, Amsterdam, pp. 979–1040.Maddala, G.S., Wu, S., 1999. A comparative study of unit root tests with panel data and a new simple test. Oxford Bulletin of

Economics and Statistics 61, 631–652.McCoskey, S., Kao, C., 1998. A residual-based test of the null of cointegration in panel data. Econometric Reviews 17,

57–84.Miller, S.M., 1988. Are saving and investment cointegrated? Economic Letters 27, 31–34.Montiel, J.P., 1994. Capital mobility in developing countries: some measurement issues and empirical estimates. World Bank

Economic Review 8, 327–342.Mundell, R.A., 1968. International Economics. MacMillan, New York.Murphy, R.G., 1984. Capital mobility and the relationship between saving and investment rates. Journal of International Money

and Finance 3, 327–342.Murthy, N.R.V., 2009. The Feldstein–Horioka puzzle in Latin American and Caribbean countries: A panel cointegration analysis.

Journal of Economics and Finance 33, 176–188.

R. Jamilov / Economic Systems 37 (2013) 155–170170

Obstfeld, M., 1986. Capital mobility in the World Economy: theory and measurement. In: Carnegie-Rochester Conference Serieson Public Policy, 24. pp. 55–104.

Obstfeld, M., Rogoff, K., 2000. The six major puzzles in international macroeconomics: Is there a common cause? NBERMacroeconomics Annual 15, 340–390.

Penati, A., Dooley, M., 1984. Current account imbalances and capital pormation in industrial countries, 1949–81. IMF StaffPapers 31, 1–24.

Pedroni, P., 1999. Critical values for cointegration tests in heterogeneous panels with multiple regressors. Oxford Bulletin ofEconomics and Statistics 61, 653–670.

Pedroni, P., 2000. Fully modified OLS for heterogeneous cointegrated panels. Advances in Econometrics 15, 93–130.Pedroni, P., 2004. Panel cointegration: asymptotic and finite sample properties of pooled time series tests with an application to

the PPP hypothesis. Econometric Theory 20, 597–625.Pedroni, P., 2007. Social capital, barriers to production and capital shares: implications for the importance of parameter

heterogeneity from a nonstationary panel approach. Journal of Applied Econometrics 22, 429–451, 2007.Pesaran, M.H., Smith, R.P., 1995. Estimating long run relationships from dynamic heterogeneous panels. Journal of Econometrics

68, 79–113.Pesaran, H., Shin, Y., Smith, R.P., 1999. Pooled mean group estimation of dynamic heterogeneous panels. Journal of the American

Statistical Association 94, 621–634.Pesaran, M.H., 2004. General diagnostic tests for cross section dependence in panels. In: Cambridge Working Papers in

Economics 0435, Cambridge.Pesaran, M.H., 2006. A simple panel unit root test in the presence of cross-section dependence. In: Cambridge Working Papers in

Economics 0346, Cambridge.Reisen, H., Yeches, H., 1991. Time-varying estimates on the openness of the capital account in Korea and Taiwan. In: OECD

Development Centre Technical Paper 42, Paris.Sarno, L., Taylor, M.P., 1998. Saving-investment correlations: transitory versus permanent. Manchester School 66, 17–38.Summers, L.H., 1988. Tax policy and international competitiveness. In: Frenkel, J. (Ed.), International Aspects of Fiscal Policies.

NBER Conference Report. University of Chicago Press, Chicago, IL, pp. 349–375.Tesar, L.L., 1991. Saving, investment and international capital flows. Journal of International Economics 31, 55–78.Tesar, L.L., 1993. International risk-sharing and non-traded goods. Journal of International Economics 35, 69–89.Westerlund, J., 2007. Testing for error correction in panel data. Oxford Bulletin of Economics and Statistics 69, 709–748.