TESTING THE BALASSA HYPOTHESIS IN LOW AND

MIDDLE INCOME COUNTRIES: External Version

Fentahun Baylie 1
(Professor Scott Hacker, Professor Par Sjolander, Dr. Girma Estiphanos)2
March/2016
Addis Ababa,
Ethiopia.

Abstract: This study analyses the long run relationship between economic growth and real
exchange rate for a group of 15 low and middle income countries for the period 1950-2011.
Cointegration between growth and exchange rate are established by means of an augmented
pooled mean group estimation method (which controls for heterogeneity and cross sectional
dependence). Unlike previous studies, cross sectional dependence is accounted for which
implies that the productivity effect of the Balassa term is expected to be estimated consistently
and without bias. Moreover, our results indicate that the Balassa hypothesis relies more on
level of development than on the rate of economic growth. In general, the strength of the
effect is stronger for higher level income countries in the long run. The study clearly indicates
that the Balassa hypothesis holds only for middle income countries-while productivity growth
does not lead to the appreciation of real exchange rate for low income countries. However, in
the short run, fiscal policy and exchange rate volatility rather clearly explains the variation in
real exchange rate.
Key words: Productivity growth, Real exchange rate, Balassa hypothesis.

1
2

PhD Student in JIBS based in Ethiopia, (fbaylie@yahoo.com)

Advisors from JIBS and Addis Ababa University (scott.hacker@ju.se, par.sjolander@ju.se, estifanosgirma186@yahoo.com respectively)

1. Introduction
The Balassa hypothesis tests the impact of productivity growth on real exchange rate. It states that
for a growing economy, the real exchange rate is expected to appreciate in the long run. This study
tests the hypothesis for a group of 15 growing economies. These countries are well performing
economies (in terms of rate of economic growth) in the past decade according to different reports of
the World Bank (World Bank (2014; 2015)). The sample countries represent nearly half the world
population [3.7 billion people] and one-third of the world GDP from four different continents; Asia,
Europe, Africa and Latin America. The countries grow with mean growth rate of 5% for past decade
(IMF, 2013). This group of countries are heterogeneous in many aspects but share common feature
regarding the features of their economic growth.
This study is based on a finding in Baylie (2008). It shows that real effective exchange rate is an
important policy parameter and among the most determining factors of growth in Ethiopia. Though
it recommended depreciation of the domestic currency to promote economic growth in the short run,
it discovered that it is healthier to allow appreciation in the long-run to encourage sustainable
economic growth. Therefore, it provided [an exchange rate] policy recommendation which
promotes appreciation of the domestic currency for sustainable growth in the long run.
Both depreciation and appreciation are not welcomed effortlessly by monetary authority.
Depreciation in particular, as suggested in Baylie (2008), is not favoured by the monetary authority
as it increases the burden on importing capacity for a developing country like Ethiopia. By the time
a country is recommended to appreciate, all the advantage of depreciation are assumed to be
exhausted and prospects of appreciation are awaiting. Depreciation may initially help to promote
exports and generate sufficient foreign earnings. Once this objective is met, there arises a need to
promote the import of capital goods to establish import-substituting industries, transform the
economy and sustain development in a developing economy by allowing appreciation. The only
question left unanswered in this case would be the timing to switch policy. The solution to this
dilemma is provided by Balassa hypothesis.
The time when the Balassa hypothesis holds in a particular economy, appreciation is gainful. In
short, it states that if economic growth is accompanied by appreciation of the domestic currency
[Balassa hypothesis], the monetary authority should not constrain the appreciation for just simple
reason that it may discourage exports. If economic growth by itself brings appreciation, it can be
sustained as the later further put inertia onto the former. There is a possibility for one to derive the
other in the long-run when the hypothesis holds.
1

The Balassa-Samuelson effect, first formulated by Harrod in 1934 and later by Balassa and
Samuelson in 1964 separately, describes that the distortion in purchasing power parity (PPP) is the
result of international differences in relative productivity growth between the tradable goods sector
(manufacturing and agriculture) and the non-tradable goods sector (services). Accordingly, during
the development process, productivity tends to increase more quickly in the tradable goods sector
than the non-tradable goods sector. Given that the prices of tradable goods are set by international
competition, an increase in productivity in this sector leads to an increase in wages, which is not
harmful to competitiveness but leads to a rise in relative prices in the non-tradable goods sector,
where productivity has not grown at the same pace. Given that the price index is an average of these
two sectors, there is an increase in the prices of domestic goods relative to those from abroad, which
results in an appreciation of the real exchange rate (Tica and Druzic (2006), Herberger (2003),
Coudert (2004)).
In short, the hypothesis states that the impact of growth on exchange rate is positive i.e. appreciation
of the domestic currency. The main purpose of this study is, therefore, to show whether this analysis
can be extended to a group of low and middle income countries from various continents.
While there are many evidences in favour of the hypothesis, there are also certain anti-Balassa
results in some studies. The negative results could be due to certain reasons such as lack of quality
or long period data, or the difficulty to guarantee assumptions of the model. Examples of antiBalassa studies include Chuah (2012), Genius and Tzouvelekas (2008), Canzoneri, Cumby and Diba
(1997), Funda, Lukinic and Ljubaj (2007), Gubler and Sax (2008), Wilson (2010), Drine and Rault
(2003), and Asea and Mendoza (1994) among others.
Tica and Druzic (2006) survey shows that since its discovery in 1964, the theory has been tested 58
times in 98 countries in time series or panel analyses and in 142 countries in cross-country analyses.
In these analyzed estimates, country specific Balassa Hypothesis coefficients have been estimated
164 times in total, and at least once for 65 different countries. The conclusion shows the difficulty to
ignore the significance of the Balassa-Samuelson theory.
The first empirical test of the theory was carried out by Balassa (1964). It was a cross-country
analysis of nine countries' data sets for 1955. Balassa (1964) made an OLS estimate with the ratio of
purchasing power parity and nominal exchange rate as a dependent variable and GNP per capita as
an independent variable. He found a significant and positive relationship between the two.

De Gregorio, Giovannini and Wolf (1993) made a study on OECD for the period 1970-1985 and
came to the conclusion that [in addition to demand shift to nontradables] higher productivity growth
in tradables are the main causes of inflation for the nontradables. Similarly, De Gregorio and Wolf
(1994) added the Penn effect to the Balassa hypothesis by introducing terms of trade parameter for a
sample of fourteen OECD countries. They arrived on the same conclusion. A relatively faster
productivity growth in the tradable sector [and an improvement in the terms of trade (TOT)] induce
a real appreciation.TOT, which is assumed to be constant in the hypothesis, is not only significant
but also found to have a role to play on the interaction between real exchange rate and productivity
differentials if allowed to vary.
In general, one may, therefore deduce from the population of literature on Balassa-Samualson effect
that the results of empirical studies depends on the level of development of countries in sample
studies, period of analysis and type of Econometrics methods employed. Thus, more powerful
results from the test are associated with longer period and panel data type of analysis.
2. Theoretical Framework of the model
2.1.

The Balassa-Samuelson Hypothesis3

Balassa hypothesis demonstrates the relationship between exchange rate, purchasing power parity
[PPP] and inter-country income comparisons in general. The hypothesis emanates from the PPP. It
explains why the PPP theory of exchange rate is imperfect. In the absence of all frictions, the prices
of a common basket of goods in two countries measured in the same currency should be the same at
all times for the absolute PPP to hold i.e.

/ = 1. The Balassa-Samuelson hypothesis provides

productivity growth differential between tradable and nontradable sectors and across nations as one
of the reasons why this does not hold.
The hypothesis is put in the first line in the list of modifications for the long run PPP by Kenneth
Rogoff. In contrast to the purchasing power parity theory, price levels are found to be higher in rich
countries than the poor ones when converted to a common currency. This is associated to higher
productivity growth in the tradable sector in rich countries (Rogoff 1996).
The Balassa hypothesis explains the fact that there is no reason for price differentials under perfect
competition (Prusa, 2007). Given this assumption, Balassa (1964) considers the fact that price
3

Though the idea has been mentioned by several authors [like David Ricardo (1911), Roy Harrod (1933), Jacob Viner (1937), Paul A. Samuelson (1964) and Bela Balassa (1964)], the contribution
of other authors is not as bold as Paul A. Samuelson and Bela Balassa and hence the name Balassa-Samuelson hypothesis (Tica J. and Druzic I. 2006). The name Balassa Hypothesis may be also
used in this study.

increase in the nontradable sector is the same as money wage transferred from producers to workers
in the traded sector. The PPP theory holds if and only if these productivity changes and wage
adjustments are identical and there exists neutral production and consumption effects in every
country. In the absence of which general price levels in each country follow the path of
technological improvements and wage adjustments. Samuelson (1964) also suggests that countries
should narrow their technological gap to have a relatively similar value of currency.
The Balassa hypothesis has two versions; the external and internal versions. The external version
analyzes the impact of productivity on real exchange rate. The internal version analyzes the impact
of productivity on relative price. By combining the two one can realize whether the hypothesis is
functioning through the internal version or the external version. If one finds a significant
relationship between relative price, real exchange rate and productivity, it is most likely that the
hypothesis is functioning through the internal version. The main purpose of this study is to discuss
the external version of the hypothesis.
The original Balassa hypothesis assumes a fully employed small open economy; a 2X2X2 economic
model (two-countries, two-commodities, two-factors); an inter-sector mobile labour factor [scarce]
and inter-nation mobile capital; law of one price for factors within a nation and for tradables across
nations; a constant return to scale production frontier; perfect competition in both (goods and
factors) markets; neural technical progress; and constant terms of trade (Podkaminer, 2003).
2.2.

Model Specification

The derivation of the Balassa-Samualson model may be considered as a three-stage process. The
first is to derive the relationship between productivity differential and relative price. This is referred
to as Internal Version of Balassa hypothesis. The second is to derive the relationship between
relative price differential and exchange rate. The third is to derive the relationship between
productivity differential and exchange rate. This is referred to as External Version of Balassa
hypothesis
STEP 1:
The original Balassa-Samuelson model is framed on the basis of the traditional Ricardian trade
model (Asea and Corden, 1994). It was a supply side model defined by a constant return to scale
Cobb-Douglas style production functions in two sectors as follows (Podkaminer, 2003).
=

(2.1)
(2.2)
4

Where T and N refers to the traded and nontraded sectors, and

and

represent the share of

labour in each sector respectively.

In a perfectly competitive market, factor prices must equal their respective value of marginal
products at equilibrium for both sectors.
=
(1 )

(2.3)

= (2.4)
=

(1 )

(2.5)

= (2.6)

Combing equation (2.3) and (2.4) and taking the logarithm of both sides, yields the following
results, equation (2.7).
!
!"#

=
log (
log(w) + ,-.(

"!
#

$
)

$
)

) = (1 ) log (

(1 )

) = (1 ) log(r) (1 ),-.(1 ) (1 ),-.(

) + ,-.(

log(w) = (1 ) log(r) (1 ),-.(1 ) ,-.(

,-.(

) = (1 ) log(r) (1 ),-.(1 ) + log(w) ,-.(

) (1 ),-.(

) ,-.(

) (2.7)

Similarly, Combing equation (2.5) and (2.6) and taking the logarithm of both sides, yields the
following results, equation (2.8).
,-.(

) = (1 ) log(r) (1 ),-.(1 ) + log(w) ,-.(

) (2.8)

Recall the assumption that price of tradables and interest rate [not technology] are fixed.
Differentiation of equations (2.7) and (2.8) with respect to time yields equations (2.9) and (2.10)
respectively.
12 ()
1

()

=0=

12 ()
1

()

14()
1

5()

14()
1

5()

16 ()
1

()

16 ()
1

()

(2.9)
(2.10)

Substituting the relation in equation (2.9) into (2.10) gives the final result that defines relative price
of nontradables in terms of productivity differentials.
12 ()
1

()

7 =

16 ()
1

()

16 ()
1

()

(2.11)

9 9 (2.12)

Where a hat represents growth rate of the variables.

For a foreign economy (*), equation (2.12) can be arranged as follows.
7 = :

9 9 ;....................................................................................(2.13)

Taking the difference between equation (2.12) and (2.13) helps to define price differentials across
countries as well.
7 7 = <

9 9 =:

9 9 ; (2.14)

This means price differential between sectors as well as across countries can be explained by
productivity differentials between sectors and across nations.
STEP 2:
We follow Ahn (2009) to link exchange rate and productivity differential through price index. Real
exchange rate is defined as:
>=

In log-linear form:

@ =

+ 7 7 (2.15)

Price indices may be defined as a weighted average of prices in the tradable and nontradable sectors
in both domestic and foreign.
=

In log-linear form:

7 = B7 + (1 B)7 (2.16)

In log-linear form:

C)

7 = D7 + (1 D)7 (2.17)

B and D represents the share of nontradables in consumer basket at home and foreign respectively.

Substituting equations (2.16) and (2.17) into equation (2.15) helps to define exchange rate as a
function of price differential.
@=

+ 7 7
6

@=

+ D7 + (1 D)7 7 (1 )7

+ 7 7 + E(7 7 ) (7 7 )F (2.18)

@=

In relative PPP terms, equation (2.18) may be defined as:

G@ = G + G7 G7 + E(G7 G7 ) D(G7 G7 )F (2.19)
For the PPP to hold, then G + G7 = G7 . Equation (2.19) will be:
G@ = E(G7 G7 ) D(G7 G7 )F ---------------------------------------- (2.20)
In absolute PPP terms, equation (2.20) may be defined as:
@ = E(7 7 ) D(7 7 )F ---------------------------------------- (2.21)
STEP 3:
Equation (2.21) defines exchange rate indirectly as a function of price differential between
countries. Substituting equation (2.14) into equation (2.21) helps to defined exchange rate directly
as a function of productivity differential. We assume that the share of nontradables in consumer
basket of foreign () is the same as at home (). Hence,
@J = <

9 9 =:

9 9 ; (2.22)

If home grows faster than foreign, then the domestic currency appreciates and vice versa.
In order to avoid the assumption of neutral technical progress, we introduce an intercept in the
econometric model that defines the relationship between exchange rate and productivity differential
(Kohler 1998).
The Balassa model is not complete by itself as it does not include demand side factors (De Gregorio
and Wolf1994). We also introduce demand side factors. This helps to account for other factors other
than productivity differential that may explain price differences such as the Baumol-Bowen effect4
and the Penn effect5.

Baumol and Bowen Effect argues that since the income elasticity of demand for services is greater than that of demand for goods, the share of
services in demand increases during the process of development. As a result, the relative price of services tends to increase to rebalance the supply
and demand of nontradable goods. Productivity shock may also workout through a demand channel to affect the real exchange rate. It may lead to
increase in the demand for investment and therefore a rise in the real interest rate in order to attract the necessary capital to fund it, thus leading to
an appreciation of the real exchange rate (Coudert V. 2004).
The Penn effect analyzes the role of terms of trade for the relative price of nontradable goods to determine deviations of the real exchange rate
from the PPP (Pancaro C. 2011)

Therefore, the econometric model which is used in this study and includes other factors is derived
from equation (2.22)6.
(lnQ)MN =

M +

M ln(Y/Y

MN

PM ln(G/G

MN

RM S-,(E)MN

+ UMN ------------------------- (2.23)

Q and E are real and nominal exchange rate. ln(>)MN represents log of real exchange rate for each
country measured against the US dollar. An increase in real exchange rate implies appreciation. it
refers to the ith country in period t. ln(Y/Y )MN : refers to log of real GDP per capita relative to the US
economy, the reference point. It is a proxy for Productivity growth differential in each country.
ln(G/G )MN is log of real government expenditure relative to the US economy. It is a proxy for fiscal
policy in each country. vol(E) is exchange rate volatility measured as the absolute value of
percentage change in nominal exchange rate7.
3. Econometric Methodology
I.

Cross Sectional Dependence Test

Cross sectional dependence is a problem associated with panel data that mix information from
different cross sections and results in a difficulty to interpret individual effects of each section. It
may arise from various sources such as neighbourhood or socioeconomic network effects, spatial
effects, the influence of a dominant unit or the influence of common unobserved factors. When
Cross sectional dependence is ignored, estimates may be badly biased and tests for unit roots and
cointegration may be misleading (Shin, 2014).
Factor models are used to study unobserved common factors which represent cross sectional
dependence.
dMN = eN

+ M fMN + gM hN + MN .................................................................(3.1)

where dMN is a scalar dependent variable, eN is a kzx1 vector of variables that do not differ over
units, e.g. intercept and trend, fMN is a kxx1 vector of observed regressors which differ over units, hN

is an rx1 vector of unobserved factors, which may influence each unit differently and which may be
correlated with the fMN , and MN is an unobserved disturbance with mean zero and constant variance
for each cross section i, which is independently distributed across i and (possibly) t. The covariance
between the errors

MN

= gM hN + MN is determined by the factor loadings gM (Shin, 2014).

We use the Pesaran cross sectional independence test in this study as it is the most powerful test
(Eberhardt, 2011). It is given by:

6
7

See Appendix A for derivation.

See Appendix A for variable measurement.

ij = k

lMq pqMr noMp JMp t ............................................................................(3.2)

Where ij~v(0, 1).

II. Panel Unit Root Tests

Six types of panel unit root tests are available. These are: Levin-Lin-Chu (LLC), Hariss-Tzavalis
(HT), Breitung test, Im-Pesaran-Shin (IPS), Fisher type and Hadri LM tests. The panel data of this
study are unbalanced and N is fixed and smaller relative to T. It also assumes that autoregressive
parameter, {, is panel specific. Hence, the candidate panel unit root tests that fit these criteria are
the IPS and Fisher type tests. Another advantage of these tests is they can be used to test a series
which is not serially independent across cross sections. However, the panel unit root test must be
preceded by a test of cross sectional independence.
a) ImPesaranShin test
The following is a panel unit root test as proposed by Pesaran (2007) with the presence of cross
sectional dependence. The standard augmented DickeyFuller (ADF) regressions are further
augmented with cross section averages of lagged levels and first-differences of the individual
series. Let yit be the observation on the ith cross-section unit at time t and suppose that it is
generated according to the simple dynamic linear heterogeneous panel data model:

~MN = gM hN + ;

dM,N = (1 |M )}M + |M dM,N

+ ~ -----------------------------------------(3.3)

= 1 v; = 1 o.

The initial value, yi0, has a given density function with a finite mean and variance, and the error
term, uit, has the single-factor structure.

ft is the unobserved common effect, and it is the

individual-specific (idiosyncratic) error. It is also convenient to write the model as:

where

= (1 |M )}M ,

GdM,N =
M

M dM,N

+ gM hM + ---------------------------------------------(3.4)

= (1 |M ) GdM,N = dM,N dM,N . The unit root hypothesis of

interest is expressed as:

:

M
M

= 0 h- ,, against the possibly heterogeneous alternatives

< 0, = 1, 2, . v ,

= 0, = v + 1, vP + 2, . , v

v /v, the fraction of the individual processes that are stationary, is nonzero and tends to the fixed
value such that 0 < < 1 as v . This condition is necessary for the consistency of the panel

unit root tests.

b)

Fisher-type tests

Maddala and Wu (1999) provide a fisher-type of panel unit root test which accounts for cross
sectional dependence. This test originates from Fisher (1932). Like the IPS test, the Fisher test is a
way of combining the evidence on the unit-root hypothesis from the N unit-root tests performed on
the N cross-section units. Fisher-type test make this approach more explicit.
The Fisher-type panel unit-root test combines the p-values from the panel-specific unit-root tests
using four methods. Three of the methods differ in whether they use the inverse chi-square,
inverse-normal, or inverse-logit transformation of p-values, and the fourth is a modification of the
inverse chi-square transformation. Simulation results suggest that the inverse normal Z statistic
offers the best trade-off between size and power (www.stata.com).
Let M, be a unit root test statistic for the ith group and assume that as Ti , then M, => M .
Let pi be the p-value of a unit root test for cross-section i, i.e., pi = 1 F(M, ), where F() is the
distribution function of the random variable M . The Fisher type test is given as follows (Chen M.
2013).
= 2 Mq ln 7M ........................................................(3.5)
P is distributed as 2 with 2N degrees of freedom as T for all N. When 7M closes to 0, ln 7M
closes to so that large value P will be found and then the null hypothesis of existing panel unit
root will be rejected. In contrast, when 7M closes to 1, ln 7M closes to 0 so that small value P will be
found and then the null hypothesis of existing panel unit root will not rejected.
III.

Panel Cointegration Tests

There are two possibilities to deal with nonstationary variables in a given model after the
stationarity test. One is, to test if the linear combination of nonstationary variables is stationary by
using cointegration test. If they are cointegrated, then we proceed to a long run analysis with
nonstationary variables. Otherwise, we difference the stationary variables for a short run analysis.
Granger noted (Gujarati, 2004) that a test for cointegration can be thought as a pre-test to avoid
spurious regression situations. A regression of one nonstationary variable over another
nonstationary variable may yield a stationary series and if so, it is known as cointegrating
regression and the slope parameter in such a regression is known as cointegrating parameter.
We employ a residual-based Pedroni cointegration test which is simply unit root tests applied to the
residuals obtained from a cointegrating regression. If the variables are cointegrated then the
10

residuals should be I(0). On the other hand if the variables are not cointegrated then the residuals
are not I(0) (Eviews 9 Users Guide; 2015). The test allows for heterogeneous intercepts and trend
coefficients across cross-sections. It is based on a residual obtained from a regression,
dMN =

+ BM +

M f M,N

PM fPM,N +. . . + M fM,N

M,N ............................................(3.6)

for = 1, . , o ; = 1, , ; = 1, , where f and d are assumed to be integrated of order

one, I(1). The parameters

and BM are individual and trend effects.

Pedroni has proposed seven different statistics to test panel data cointegration, namely the Panel vStatistic, Panel rho-Statistic, Panel PP-Statistic, Panel ADF-Statistic, Group rho-Statistic, Group
PP-Statistic and Group ADF-Statistic. Out of these seven statistics, the first four are based on
pooling, what is referred to as the Within dimension, and the last three are based on the
Between dimension. Both kinds of tests focus on the null hypothesis of no cointegration.
However, the distinction comes from the specification of the alternative hypothesis. For the tests
based on Within dimension, the alternative hypothesis is i = < 1 for all i, while the tests
based on Between dimension, the alternative hypothesis is i < 1 for all i (Bangake and Eggoh
2012).
IV.

Estimation Method

The choice of estimation method mainly depends on the nature of data. Analysis of nonstationary
variables calls for the use of long-panel related methods. The type of macro-econometrics method
used is also associated with assumption regarding technology parameters and factor loadings.
Therefore, we are not supposed to consider traditional estimators such as Pooled OLS, fixed effect
and first difference OLS models which are mainly used for micro data with large N and small T in
homogeneous technology parameters and factor loadings. Eberhardt et. al. (2011) and others
suggested the use of pooled mean group estimation method for the analysis of nonstationary
variables which are cointegrated in long panel setting. This method is helpful in particular when T
is large relative to N for heterogeneous technology parameters and factor loadings.
Pooled mean group (PMG) estimator involves averaging and pooling. It restricts the long-run
coefficients to be homogenous over the cross-sections, but allows for heterogeneity in intercepts,
short-run coefficients (including the speed of adjustment) and error variances. It is argued that
country heterogeneity is particularly relevant in short-run relationships, given that countries may be
affected by over lending, borrowing constraints, and financial crises in short-time horizons. On the
other hand, there are often good reasons to expect that long-run relationships between variables are

11

homogeneous across countries due to budget or solvency constraints, arbitrage conditions or

common technologies (Cavalcanti et. al., 2011).
The relationship in pooled mean group estimation may be defined by an ARDL model as follows:
G@MN =

M GfMN

+ M lDfM,N

Where @ = ,> and f = ,.

M,

@M,N t +

MN ...............................................................(3.7)

are short-run parameters, which like MP differ across countries.

Error-correction termM also differs across i, long-run parameter however is constant across the

groups. This estimator is quite appealing when studying small sets of arguably similar countries.
In I(1) panels, this estimator allows for mix of cointegration (M > 0) and noncointegration (M =
0). fM,N represent the set of explanatory variables defined in equation (2.23).

To account for cross sectional dependence which may result from any common unobserved factor
incorporated in the error term, we follow the Common Correlated Effect approach by Pesaran
(2013). The approach can handle multiple factors which can be correlated with the regressors, and
serial correlation in the errors and lagged dependent variables (Shin, 2014).
The procedure consists of approximating the linear combinations of the unobserved factors by cross
sectional averages of the dependent and explanatory variables, and then running standard panel
regressions augmented with these cross section averages. It does not require an a priori knowledge
of the number of unobserved common factors and can be applied to dynamic panels with
heterogeneous coefficients and weakly exogenous regressors (Pesaran, 2013).
The PMG estimator for a cross sectionally dependent series may be defined as follows:
G@MN =

M GfMN

+ M lDfM,N

@M,N t + gMN hN + MN ...................................................(3.8)

gM hN + MN =

MN

hN is a vector of unobserved common shocks, which captures source of error term dependencies
across countries. It may be stationary or nonstationary. The impacts of these factors on each
country are governed by the idiosyncratic loadings in gMN . The individual-specific errors, MN , are
distributed independently across i and t; they are not correlated with the unobserved common
factors or the regressors; and they have zero mean, variance greater than zero, and finite fourth
moments. The common factors, or the heterogeneous time effects, may be captured by adding cross
sectional averages of the observables to our regressions (Cavalcanti et. al., 2011).

12

The augmented Pooled mean group estimator is, therefore, defined by substituting cross sectional
averages for the unobserved common factors.
G@MN =

Where e

,N

M GfMN

+ M lDfM,N

@M,N t + q Be

,N

+ MN ...................................(3.9)

represents the set of cross sectional averages of the dependent and independent

variables and their lagged values which approximate/proxy the unobserved common factors (hN ).
The focus of this estimator is on obtaining consistent estimates of the parameters related to the
observable variables, while the estimated coefficients on the cross-section averaged variables are
not interpretable in a meaningful way: they are merely present to alter out the biasing impact of the
unobservable common factor (Eberhardt, 2012).
V.

Error Correction Mechanism (ECM)

If two/more variables are cointegrated or proved to have a long run relationship one needs to go for
an error correction mechanism. Error correction mechanism (ECM) is a method used to correct any
short run deviation of variables from their long run equilibrium i.e. it corrects for disequilibrium.
An important theorem, known as the Granger representation theorem, states that if two variables Y
and X are cointegrated, then the long term or equilibrium relationship that exists between the two
can be expressed as ECM (Gujarati, 2004). This means one shall go for the construction of an error
correction model if the two variables are cointegrated. The ECM can be given by:
@MN =

M fMN

+ D~M,N

+ MN ............................................(3.10)

denotes the first difference operator, MN is a random error term, and ~M,N

= (@MN fMN ) is one-

period lagged value of the error term from the cointegrating regression.
This ECM equation states that @MN depends on fMN and also on the equilibrium error term. If the

latter [error term] is nonzero, the model is out of equilibrium. Suppose fMN is zero and ~M,N

positive. This means @M,N

negative, the term D~M,N

is

is too high [above] to be in equilibrium. Since D is expected to be

is negative and, therefore, @MN will be negative to restore the

equilibrium. That is, if @MN is above its equilibrium value, it will start falling in the next period to
correct the equilibrium error; hence the name ECM. By the same token, if ~M,N

is negative (i.e.,

@MN is below its equilibrium value), D~M,N will be positive, which causes @MN to be positive,
leading @MN to rise in period t. Absolute value of D determines how quickly the equilibrium is

restored (Gujarati, 2004).

13

4.

Empirical Results

4.1. Data type and Collection methods

This study relies on a macro panel data. Data for all countries and variables are from the World
Bank database. The variables include exchange rate, per capita GDP and government expenditure.
The final version [Penn World Table version 8.0] of the World Bank database provides annual time
series data for the period 1950-2011. While the choice of study period for each country depends on
data availability, our sample countries are selected on the basis of their economic performance (rate
of economic growth). Basically, the sample includes 15 countries from two income groups; middle
(BRICS) and low with best performance in terms of economic growth since the beginning of the
21st century. The World Bank and reports of other international financial institution show that the
countries in our sample are among the fastest growing economies in the world (World Bank; 2015).
The countries in our sample are Angola, Brazil, China, Ethiopia, Ghana, Indonesia, India, Kenya,
Nigeria, Philippines, Russia, Rwanda, Tanzania, Uganda, South Africa plus the USA (our reference
point).
4.2. Test Results
The analysis begins by performing different econometric tests. Since not all unit roots provide the
appropriate results, cross sectional independence test was performed to decide the type of panel
unit root test to be considered.
Using the Pesaran CD test, and possibly all other tests, we reject the null of cross sectional
independence for the raw data. Hence, the series for our data are cross sectionally dependent
initially. However, after the data is augmented for cross sectional averages to eliminate the
common factors, the Pesaran CD test, and possibly three other tests, fails to reject the null of cross
sectional independence. Test results are available in table 2B, annex B.
Panel unit root tests which account for cross sectional dependence are used; the IPS and Fisher tests.
The results of the tests with different assumptions are given in table 3B, annex B. All variables are
nonstationary at levels at 1% level of significance.
The next step is to test for cointegration-whether there is a long run relation between our
nonstationary variables. The test for cointegration is residual based. We use two Pedroni type tests
(ADF and PP tests) and the IPS test. In all cases, we strongly reject the null of no cointegration for
both types of models (augmented and non-augmented). Augmented models include cross sectional
averages of dependent and independent models which account for cross sectional dependence.

14

In the final analysis of our estimation, we consider three types of augmented models; model I (with
one explanatory variable), model II (with two explanatory variables) and model III (with three
explanatory variables). In the first model, log of real exchange rate (,>) is defined as a function of
productivity (ln ( / )). In the second Model, we include government expenditure (ln (/ )). In
the third model, we add exchange rate volatility (vol (?)). Even though the model selection criterion

suggests that the model with three variables is our best model in terms of log likelihood ratio and
akaike information criteria (table 5B, annex B), we present the results of other models for
comparison purposes. However, both the log likelihood test and information criteria shows that an
ARDL(1, 1, 1, 1) model best explains/represents our relationship. The dimensions represent ,>N ,
ln ( /

), ln (/ ) and vol (?) respectively. Therefore, we focus on the results of model III [the

shaded rows in each group] in the discussion that follows.

4.3.

Findings

Since the objective in this study is to find country specific coefficients for each variable, the
assumption of homogeneous parameter is not valid. Pesaran H., Shin Y., and Smith R. (1998)
suggested the use of pooled mean group [PMG] estimation for the analysis of nonstationary
variables as T gets larger if they are cointegrated. PMG estimator involves both pooling and
averaging. It allows the intercepts, short run coefficients and error variances to differ across groups
while the long run coefficients are constrained to be the same.
Table 4.1 shows the long run results of panel cointegration estimation using augmented PMG
estimator for different groups of countries and models in the sample. We follow the tradition of
presenting the estimated coefficients of only observable variables as the cross-section averaged
variables are not directly interpretable in a meaningful way. The estimated coefficients of full
model (observable and unobservable variables) are reported in annex C.
Basically, we consider three types of groups in comparing the three types of models; ALL
COUNTRIES group (15 countries); Country groups by income-middle income countries (MICs; 5
countries) and low income countries (LICs; 10 countries); and Country groups by region (AFRICA
(9 countries) and ASIA (4 countries)). For each group, the first row shows a model with one
explanatory variable (productivity- ln( / )); the second row shows a model with two explanatory

variables (productivity and government expenditure- ln (/ ),); the third row shows a model with
three explanatory variables (productivity, government expenditure and exchange rate volatility vol (?)). The results for model III [shaded rows] are discussed for each group below.

15

Table 4.1: Panel Cointegration Estimation: Augmented PMG Estimator

Sample
[# of countries]
All countries
[15]

Type of
Model
Model I
Model II
Model III

MICs (BRICS)
[5]

Model I
Model II
Model III

LICs
[10]

Model I
Model II
Model III

Africa
[9]

Model I
Model II
Model III

Asia
[4]

Model I
Model II
Model III

Long-run coefficients

,> = 7 S ,
ln ( / )
ln (/ )
vol(?)

0.378296***
(0.108540)
0.246642***
(0.075056)
0.388355***
(0.092802)
0.109014
(0.151776)
0.382324**
(0.160930)
0.344657**
(0.149000)
0.320488*
(0.132415)
0.211173**
(0.098594)
-0.287286***
(0.086673)
0.366216**
(0.144702)
0.239056**
(0.103439)
-0.247591***
(0.071625)
0.248016
(0.258474)
0.519768**
(0.216328)
0.771434*
(0.439197)

0.170621***
(0.037414)
0.110547**
(0.045996)

-0.021531**
(0.010345)

0.220887*
(0.120288)
0.252061***
(0.092608)

0.024845**
(0.012602)

0.140663***
(0.041793)
0.010920
(0.049502)

-3.21610***
(0.419248)

0.168407***
(0.043046)
0.077565**
(0.038439)

-2.56978***
(0.285754)

-0.179409**
(0.078622)
-0.339890
(0.207397)

-7.71735***
(2.826821)

Q and E are real and nominal exchange rate, Y /Y*= real GDP of home relative to foreign (US), G and G*=real government
expenditure of home relative to foreign (US), vol(E) exchange rate volatility.
***, ** and * refer to significance level at 1%, 5% and 10%. Standard errors in parentheses.
MICs: refers to middle income countries of the BRICS group (Brazil, Russia, India, China and South Africa)
LICs: refers to low income countries (Angola, Ethiopia, Ghana, Indonesia, Kenya, Nigeria, Philippines, Rwanda, Tanzania and
Uganda)
Africa refers to African countries (Angola, Ethiopia, Ghana, Kenya, Nigeria, Rwanda, Tanzania, Uganda, and South Africa)
Asia refers to Asian countries (China, India, Indonesia, Philippines)

The results in table 4.1, in general, show that the Balassa hypothesis holds for all countries as a
group in the long-run i.e. a 1% improvement in productivity leads to an appreciation of domestic
currencies of the countries in the group by 0.388% on average. We find a different result, however,
when we categorize our sample into different groups. When categorized by level of per capita
income, the results show the Balassa hypothesis holds only for middle income countries (MICs).
The same fact holds when countries are categorized by region i.e. Balassa hypothesis holds only for
Asian countries. This may be related to the fact that most middle income countries are from Asia
and poor countries are from Africa in our sample. In both cases, a 1% increase in productivity

16

appreciates the domestic currencies of the countries in MICs and Asia groups nearly by 0.34% and
0.77% respectively (though only at 10% level of significance for the later). For LICs and Africa
groups, On the other hand, a 1% increase in productivity depreciates the domestic currencies of the
countries in the groups nearly by 0.287% and 0.247% respectively. With type II model, however,
the outcomes are uniform for all cases and are in line with the Balassa hypothesis.
The long run relationship between government expenditure and real exchange rate shows that
Expansionary fiscal policy results in appreciation of the domestic currencies in all cases except for
the LICs and Asia groups. This may not be surprising as the major countries with big economies
in both groups are almost the same (Indonesia and Philippines are members of both groups).
Exchange rate volatility has the impact of depreciating real exchange rate for all countries in all
groups except the middle income group in the long run.
Table 4.2 presents the results for short run dynamics of the same groups of countries and models in
Table 4.1. The discussions that follow focus on model III [the shaded rows].
The short-run dynamics shows that the impact of productivity on real exchange rate is also
significant though negative i.e. it has the impact of depreciating real exchange rate for all countries
in all groups in the short run. Fiscal policy does not significantly explain the variation in real
exchange rate. Exchange rate volatility significantly and negatively impacts real exchange rate in
the short run in all groups except the LICs and Africa group. In the later groups it has no impact at
all in the short run. This may be due to the reason that LICs do not use exchange rate as a policy
variable i.e. it is mainly fixed or highly managed.
An error correction term links short-run periods to the long-run period. It adjusts short-run
dynamics to achieve a long-run equilibrium. The [negative] signs and statistical significances of the
error correcting terms show that the system is stable. A stable cointegrating relationship adjusts the
short-run deviations by the extent of the error correcting term. The rate of adjustment is, however,
higher (12%) in MICs than LICs (8%). This means MICs have faster rate of adjustment and achieve
equilibrium earlier than LICs.

17

Table 4.2: Short-run dynamics of Panel Cointegration Estimation: Augmented PMG Estimator
Sample
[# of countries]

Type of
Model

Adjustment
coefficient

All countries
[15]

Model I

-0.109894***
(0.022462)
-0.142981***
(0.032498)
-0.122432***
(0.037203)
-0.167326***
(0.068251)
-0.193645***
(0.102726)
-0.12243***
(0.037203)
-0.111716***
(0.027206)
-0.141304***
(0.036735)
-0.086261***
(0.024822)
-0.098028***
(0.031963)
-0.131601***
(0.041812)
-0.107015***
(0.032358)
-0.127473***
(0.041693)
-0.165803*
(0.092356)
-0.052518***
(0.019297)

Model II
Model III
MICs (BRICS)
[5]

Model I
Model II
Model III

LICs
[10]

Model I
Model II
Model III

Africa
[9]

Model I
Model II
Model III

Asia
[4]

Model I
Model II
Model III

Short-run coefficients

,> = 7 S ,
ln ( / ) ln (/ )
vol(?)

-0.327721***
(0.081918)
-0.359401***
(0.091545)
-0.397465***
(0.082479)
-0.300195***
(0.082509)
-0.408364***
(0.133144)
-0.3974***
(0.082479)
-0.352239***
(0.118025)
-0.364848***
(0.136587)
-0.423972***
(0.099213)
-0.427619***
(0.104199)
-0.427570***
(0.122564)
-0.408247***
(0.100776)
-0.213683
(0.193959)
-0.229174
(0.192154)
-0.458148***
(0.162181)

-0.046910*
(0.027791)
-0.057736*
(0.030909)

-0.161279***
(0.038693)

-0.082710
(0.073367)
-0.057736*
(0.030909)

-0.161279***
(0.038693)

-0.042973
(0.032613)
-0.016775
(0.028716)

-0.020308
(0.031332)

-0.058400**
(0.031572)
-0.032190
(0.029437)

-0.034526
(0.046803)

0.035224
(0.067468)
-0.013347
(0.067688)

-0.044258*
(0.023795)

Q and E are real and nominal exchange rate, Y /Y*= real GDP of home relative to foreign (US), G and G*=real government
expenditure of home relative to foreign (US), vol(E) exchange rate volatility, =first difference.
***, ** and * refer to significance level at 1%, 5% and 10%. Standard errors in parentheses.
MICs: refers to middle income countries of the BRICS group (Brazil, Russia, India, China and South Africa)
LICs: refers to low income countries (Angola, Ethiopia, Ghana, Indonesia, Kenya, Nigeria, Philippines, Rwanda, Tanzania and
Uganda)
Africa refers to African countries (Angola, Ethiopia, Ghana, Kenya, Nigeria, Rwanda, Tanzania, Uganda, and South Africa)
Asia refers to Asian countries (China, India, Indonesia, Philippines)

Table 4.3 and Table 4.4 present the short run dynamics for individual countries in two income
groups (MICs and LICs). The results are presented for type-III model given our criteria.
Table 4.3 presents the short run dynamics for MICs -BRICS (Brazil, Russia, India, China and South
Africa). The short-run dynamics shows that the impact of productivity on real exchange rate is
significant and negative for all countries except Brazil and South Africa. Productivity does not have
an impact on real exchange rate in these countries. Expansionary Fiscal policy results in
depreciation of the real exchange rate in Brazil, Russia and China. The role of Exchange rate
volatility is significant in all countries. However, the effect is exceptionally positive in Russia.
18

Table 4.3: Short-run dynamics by Country: Middle Income Group (BRICS): Model III
Cases

All
countries
Brazil
China
India
Russia
South
Africa

Adjustment
coefficient
-0.12243***
(0.037203)
-0.227627***
(0.004377)
-0.064666***
(0.000587)
-0.046854***
(0.000797)
-0.562507***
(0.015174)
-0.078684***
(0.001472)

Short-run coefficients

ln ( /

-0.3974***
(0.082479)
0.228636*
(0.096920)
-0.5913***
(0.012107)
-0.4483***
(0.026819)
-0.4439***
(0.045163)
-0.194967
(0.151551)

,> = 7 S ,
ln (/ )
-0.057736*
(0.030909)
-0.088359***
(0.006094)
-0.101536***
(0.006333)
0.015571
(0.008586)
-0.451463***
(0.010053)
0.039537
(0.058783)

vol(?)

-0.161279***
(0.038693)
-0.002061***
(1.62E-05)
-0.354891***
(0.007472)
-0.301000***
(0.006721)
0.006537***
(3.27E-05)
-0.373313***
(0.009758)

Q and E are real and nominal exchange rate, Y /Y*= real GDP of home relative to foreign (US), G and G*=real
government expenditure of home relative to foreign (US), vol(E) exchange rate volatility, =first difference.
***, ** and * refer to significance level at 1%, 5% and 10%. Standard errors in parentheses.
MICs: refers to middle income countries of the BRICS group (Brazil, Russia, India, China and South Africa)

The coefficients of error correction terms are negative and statistically less than one in absolute
value for all countries. The rate of adjustment is however, highest (56.25%) in Russia followed by
Brazil (22.76%). This may be associated with the size and nature of their economies. These are the
biggest economies in the group which account for 40% and 20% of the US economy. Faster rate of
adjustment means they can achieve equilibrium earlier than other countries.
Table 4.4, on the other hand, presents the short run dynamics for LICs. The short-run dynamics
shows that the impact of productivity on real exchange rate is significant and negative for all
countries except Indonesia and Uganda. Productivity does not impact real exchange rate in the
short run in these countries. The role of Fiscal policy is significant in all countries even though the
effect is different. Increase in government expenditure results in depreciation of the real exchange
rate in all countries but in Ghana, Kenya, Indonesia and Philippines. The strongest impact of fiscal
policy is shown by Uganda (0.15%). The impact of Exchange rate volatility is significant in all
countries except Indonesia.
The coefficients of error correction terms are negative and statistically significant for all countries.
The rate of adjustment is highest (24.89%) in Kenya followed by Philippines (14.58%) and
Ethiopia (13.63%). They achieve equilibrium earlier than other countries.

19

Table 4.4: Short-run dynamics by Country: Low Income Group

Cases

All
countries
Angola
Ethiopia
Ghana
Indonesia
Kenya
Nigeria
Philippines

Rwanda
Tanzania
Uganda

Adjustment
coefficient
-0.086261***
(0.024822)
-0.001548***
(4.27E-07)
-0.136337***
(0.000650)
-0.094831***
(0.000562)
0.000108***
(1.02E-05)
-0.248868***
(0.001492)
-0.042414***
(0.000194)
-0.145762***
(0.000584)
-0.075394***
(0.000518)
-0.107271***
(0.000337)
-0.010293***
(5.76E-05)

Short-run coefficients

ln ( /

-0.423972***
(0.099213)
-0.372509***
(0.025896)
-0.816907***
(0.015748)
-0.691301***
(0.035033)
-0.006073
(0.086557)
-0.121422***
(0.012865)
-0.632917***
(0.009143)
-0.647025***
(0.022619)
-0.345743***
(0.005333)
-0.665093***
(0.013354)
0.059272
(0.032875)

,> = 7 S ,
ln (/ )
-0.016775
(0.028716)
-0.171725**
(0.003968)
-0.03830***
(0.002945)
0.03351***
(0.002317)
0.06723***
(0.010901)
0.11217***
(0.000995)
-0.02030***
(0.000807)
0.05421***
(0.003975)
-0.012847**
(0.002458)
-0.04128***
(0.002109)
-0.15042***
(0.007855)

vol(?)

-0.020308
(0.031332)
-0.005968***
(7.06E-06)
0.065656***
(0.007485)
-0.042398***
(0.002733)
-0.017159
(5.17E-05)
0.164651***
(0.003452)
-0.125337***
(0.002453)
-0.089533***
(0.002379)
-0.028748***
(0.002997)
-0.177574***
(0.003046)
0.053328***
(0.002087)

lnQ= log of real exchange rate differenced, lnY/Y*=log of real GDP relative to foreign (US) differenced, lnG/G*=log of
real government expenditure relative to foreign(US) differenced, vol(E) exchange rate volatility differenced.
***, ** and * refer to significance level at 1%, 5% and 10%. Standard errors in parentheses.
LICs: refers to low income countries (Angola, Ethiopia, Ghana, Indonesia, Kenya, Nigeria, Philippines, Rwanda, Tanzania and
Uganda)

5. Conclusions and Implications

5.1. Conclusions
In general, the results of our study confirm that the relationship between real exchange rate and
productivity does exist and is stronger for higher income countries in the long run. The level of
development [measured by real per capita income] matters more than the rate of economic growth
in explaining the effect of the Balassa term in our study. This may be related to what is stated in the
Balassa hypothesis. If a sustainable economic growth is derived by productivity growth in the
tradable sector, it will transform an economy to a higher level of per capita GDP. In this case,
productivity growth appreciates the real exchange rate. This may be the reason why the Balassa
hypothesis holds for MICs and why it does not hold in for LICs in general.
In the short run, however, we find almost uniform results. Productivity growth (mainly of
nontradables), expansionary fiscal policy and high exchange rate volatility results in real exchange
rate depreciation. More specifically,
20

=>Improvements in productivity and expansionary fiscal policies both have the impact of
depreciating real exchange rate in almost all countries both middle and low income.
=>The impact of exchange rate volatility is significant only in MICs. This may be associated with
the type of exchange rate policy/regime adopted. It is mainly fixed (unchanged) in LICs in which
case it may not be able to explain a change in real exchange rate in the short run.
5.2. Policy Recommendations
We recommend the following policy options for MICs and LICs on the basis of our findings.
=> Since the Balassa hypothesis does not hold for LICs, economic growth does not lead to real
exchange rate appreciation in these countries. Hence, countries in this group may continue to
grow by promoting productivity growth in the non-tradable sector.
=>The Balassa hypothesis does hold for the MICs. Economic growth does lead to real exchange
rate appreciation in these countries. Hence, countries in this group may promote growth by
increasing productivity in the tradable sector.
=>The role of fiscal policy may not last long in LICs but in MICs. Therefore, it should be used only
as a short run tool to alter movements in real exchange rate in LICs. The same is true for
exchange rate volatility.

21

References:
Ahn G.M. (2009); Looking for the Balassa-Samuelson Effect in Real Exchange Rate
Changes: Andong National University, Journal of Economic Research #14, pp. 219-237.
Albuquerque B., Baumann U. and Krustev G. (2014); Has US Household Deleveraging
Ended? A Model-Based Estimate of Equilibrium Debt, Working Papers, Bank of Portugal,
Portugal.
Asea P.K. and Corden W.M. (1994); The Balassa-Samuelson Model: An Overview,
Working Paper #710, USA.
Balassa B. (1964); The Purchasing Power Parity Doctrine: A Reappraisal, Journal of
Political Economy, Yale University, Vol. 72, No. 6, pp. 584-596, the University of
Chicago Press, USA.
Bangake C. and Eggoh J. (2012); Pooled Mean Group estimation on international capital
mobility in African countries, France.
Baylie F. (2008); The Impact of Real Effective Exchange Rate on the Economic Growth of
Ethiopia, master thesis, Addis Ababa University, Ethiopia.
Canzoneri M.B. Cumby R.E. and Diba B. (1997); Relative Labour Productivity and Real
Exchange Rate in the Long Run: Evidence for a Panel of OECD Countries, Georgetown
University, USA.
Cavalcanti V, Mohaddes K., and Raissi M. (2011); Commodity Price Volatility and the
Sources of Growth, IMF Working Paper, Middle East and Central Asia Department
Chen M. (2013); Panel Unit Root and Cointegration Tests, National Chung Hsing
University, USA.
Chuah K.P. (2012); How Real Exchange Rate Move in Growing Economies: Anti-Balassa
Evidence in Developing Countries, Malaysia.
Chudiky A. and Pesaran H. (2013); Common Correlated Effects Estimation of
Heterogeneous Dynamic Panel Data Models with Weakly Exogenous Regressors, USA.
Coudert V. (2004);Measuring the Balassa-Samuelson Effect for the Countries of Central
and Eastern Europe, Banque de France Bulletin Digest #122, France.
De Gregorio J. and Wolf H. (1994); Terms of Trade, Productivity and the Real Exchange
Rate, NBER Working paper No. 4407, Cambridge, Massachusetts, England.
De Gregorio J., Giovannini A. and Wolf H. (1993); International Evidence on Tradables
and Nontradables Inflation, NBER Working paper No. 4438, Cambridge, Massachusetts,
England.

22

Drine I. and Rault C. (2003); Does the Balassa-Samuelson hold for Asian Countries? An
Empirical Analysis Using Panel Data Cointegration Tests, France.
Eberhardt (2011); Panel Time-Series Modelling: New Tools for Analyzing xt Data,
University of Nottingham, UK Stata Users Group meeting, Case Business School,
England.
Eviews 9 (2015); Users Guide, HIS Global Inc.
Funda J., Lukinic G. and Ljubaj I. (2007); Assessment of the Balassa-Samuelson Effect in
Croatia, Croatian National Bank of Zagreb, Croatia.
Genius M. and Tzouvelekas V. (2008); The Balassa-Samuelson Productivity Bias
Hypothesis: Further Evidence Using Panel Data, University of Crete, Greece.
Gubler M. and Sax C. (2008); The Balassa-Samuelson Effect Reversed: New Evidence
from OECD Countries, University of Basel, Switzerland.
Gujarati (2004), Basic Econometrics, Fourth Edition, the McGrawHill Companies.
Herberger A.C. (2003); Economic growth and The Real Exchange Rate: Revising the
Balassa-Samuelson Effect, University of California, Los Angeles, USA.
IMF (2013), World Economic outlook
Maddala, G. S., and Wu, S. (1999); A comparative study of unit root tests with panel data
and a new simple test, OxfordBulletin of Economics and Statistics, 61(Special Issue), 631652.
Michael Schuman (2013), Forget the BRICS; Meet the PINEs, an article on the
Time Magazine.
Obstfeld and Rogoff (1996); Foundations of Macroeconomics, the MIT press, Cambridge,
Massachusetts, London, England.
Pesaran, M. H. (2007); A simple panel unit root test in the presence of cross-section
dependence, Journal of Applied Econometrics, 22(2), 265-312.
Pesaran H. (2013); Large Panel Data Models with Cross-Sectional Dependence: A
Survey, Cambridge, UK.
Pesaran H., Shin Y., and Smith R. (1998); Pooled Mean Group Estimation of Dynamic
Heterogeneous Panels, Trinity College, Cambridge, and University of Southern
California, USA.
Podkaminer L. (2003); Analytical Notes on the Balassa-Samuelson Effect, BNL Quarterly
Review, No. 226.
Pilbeam K. (1992), International Finance, Macmillan Publishing Enterprise, London

23

Prusa J., Nanits M., Seman V., Sierra M., Solc J. and Tanbayeva D. (2007); BalassaSamuelson Effect, Prague, Czech Republic.
Rogoff K. (1996); The Purchasing Power Puzzle; Journal of Economic Literature,
Volume 34, pp. 647-668.
Samuelson P. (1964); Theoretical Notes on Trade Problems; the Review of Economics
and Statistics, Volume 46, Number 2, pp. 145-154.
Shin Y. (2014), Dynamic Panel Data Workshop, University of York, University of
Melbourne.
Smith R. and Fuertes A. (2016); Panel Time-Series, London.
Tica J. and Druzic I. (2006); The Harrod-Balassa-Samuelson Effect: A Survey of
Empirical Evidence, University of Zagreb, Working Paper Series, Paper No. 06-7/686,
Zagreb, Croatia.
Tzilianos E. (2006); The Balassa-Samuelson Effect and Europes Southern Periphery,
Wilson E. (2010); European Real Effective Exchange Rate and Total Factor Productivity:
An Empirical Study, Victoria University of Wellington, New Zealand.
World Bank (2014); Global Economic Prospects, CIA world fact book.
World Bank (2015); The World in 2050: Will the Shift in Global Economic Power
Continue?
Zhijie Xiao and Peter C.B. Phillips (1998); An ADF coefficient test for a unit root in
ARMA models of unknown order with empirical applications to the US economy, USA.
Yongcheol Shin Y. (2014), Dynamic Panel Data Workshop, University of York,
University of Melbourne.
www.stata.com/xtunitroot -panel-data-unit-root-tests.

24

Appendices
Annex A
1A: Model Derivation (Hackers contribution)
Suppose that the growth rate of real exchange rate is defined as a function of productivity
differential between the nontradable and tradable sectors as in equation (2.22),

> = <

9 9 =:

With > 7 7 .

9 9 ; (2.22)

This is the same as equation (1A),

> = <

9 9 = 9 9 (1A)

= l 9 9 t and =
If we let

, then

+ l 9 9 t (2A)
> =
In levels form this is equivalent to:

> = (

)!

(3A)

and in log-levels it is

@ = + ,(

Where @ ln > ln
We proxy

with /

(4A)

where Y is the Home real GDP per capita and Y* is the

foreign (US) real GDP per capita, so we get equation (2.23).

@ = + (, /

25

) (5A)

Annex B
Table 1B: Descriptive Statistics (all countries)
ln(>)

Mean
Median
Maximum
Minimum
Std. Dev.
Skewness
Kurtosis
Jarque-Bera
Probability
Sum
Sum Sq. Dev.
Observations

ln ( /

-0.673730
-0.664635
0.423540
-1.626351
0.367519
0.229321
2.768579
9.159802
0.010256
-561.2175
112.3785
833

-2.701146
-2.721925
-0.434831
-4.691372
0.790716
0.058521
2.519203
8.498842
0.014272
-2250.055
520.1929
833

ln (/ )

vol(?)

0.946177
1.181342
2.860480
-2.248809
0.979527
-0.706851
2.946596
69.46539
0.000000
788.1657
798.2815
833

0.472646
0.042409
45.55217
0.000000
2.854727
11.76179
164.6432
909407.2
0.000000
386.6244
6658.114
818

Table 2B: Cross Sectional Dependence Tests

Tests
Non-augmented Model
Breusch-Pagan LM
235.4183***
Pesaran scaled LM
7.964616***
Bias-corrected scaled LM
7.837498***
Pesaran CD
11.56950***
Null hypothesis: No cross-section dependence (correlation)

Augmented Model
145.5989***
1.766490*
1.639371
-0.781383

Table 3B: Panel unit root tests (IPS and Fisher tests)
Variables
ln(>)
ln(>)
ln ( /

ln ( /

ln (/ )
ln ( /

vol(?)

Specifications

Pesaran statistics

Fisher statistics

Constant

-1.05668

32.7336

Constant and trend

Constant
Constant and trend
Constant
Constant and trend
Constant
Constant and trend
Constant
Constant and trend
Constant
Constant and trend
Constant
Constant and trend

1.61611

21.1285

-21.0684***

401.234***

-20.0749***

344.399***

-0.31331

34.8417

3.39295

21.6995

-16.2246***

296.619***

-19.4344***

327.624***

-0.92447

32.5055

Order of
integration
I(1)
I(0)
I(1)
I(0)
I(1)

0.61201

26.5756

-19.9922***

377.622***

-19.2329***

350.826***

-13.7445***

246.719***

-16.0748

230.487

I(0)
I(0)

*** indicates the rejection of the null hypothesis (unit root) at 1%.

Table 4B: Results of Cointegration tests

IPS test
NonAugmented
-19.1510***
-18.1349***
-17.3608***

Model
Model I
Model II
Model II

Augmented
-22.1827***
-22.0030***
-20.7525***

ADF test
NonAugmented
356.986***
339.992***
314.362***

PP test
NonAugmented
353.953***
351.516***
317.275***

Augmented
427.724***
426.947***
392.643***

Augmented
430.157***
499.481***
409.215***

*** indicates rejection of the null hypothesis (unit root/no cointegration) at 1%.
Table 5B: Model Selection Criteria
Model Type

LogL

AIC*

BIC

HQ

Specification

822.924667

-1.821332

-1.372511

-1.649096

ARDL(1, 1)

II

858.646483

-1.830431

-1.197479

-1.587536

ARDL(1, 1, 1)

III

920.977772

-1.980019

-1.244362

-1.697463

ARDL(1, 1, 1, 1 )

26

Annex C
Table 1C: Panel Cointegration Estimation: Pooled Mean Group Estimator (MICs)
a) Model I: 5 countries and one explanatory variable
Cases

Adjustment
coefficient

All countries

All countries
Brazil
China
India
Russia
South Africa

-0.167326***
(0.068251)
-0.094035***
(0.002577)
-0.062517***
(0.001160)
-0.088281***
(0.001965)
-0.432825***
(0.012245)
-0.158971***
(0.005060)

Long-run coefficients
ln(>) = 7 S ,

ln ( / )
ln ( / )
ln(>)
0.798599
0.109014
-0.279218
(0.392870)
(0.151776)
(0.408663)
Short-run coefficients
ln(>) = 7 S ,

ln ( / )
ln ( / )
ln(>)
-0.300195***
0.193001
1.184039*** (0.297654)
(0.082509)
(0.143309)
1.517754*** (0.092884)
0.080101
-0.134599
(0.081196)
(0.097405)
0.731379*** (0.044042)
-0.544820***
0.172248*
(0.016645)
(0.061033)
0.651760*** (0.023531)
-0.424671***
0.116570**
(0.029397)
(0.026727)
2.202588*** (0.107770)
-0.200435***
0.727058*
(0.030829)
(0.308526)
0.816714*** (0.058186)

-0.250946*
(0.106019)

0.083730
(0.075642)

b) Model II: 5 countries and two explanatory variables

Cases

Adjustment
coefficient

All
countries

All
countries
Brazil
China
India
Russia
South
Africa

-0.193645***
(0.102726)
-0.145779***
(0.004389)
-0.093285***
(0.001064)
-0.048770***
(0.001090)
-0.599758***
(0.017277)
-0.080631***
(0.001471)

ln(>)
2.572100***
(0.618219)

ln ( / )
0.382324**
(0.160930)

ln(>)
0.911459***
(0.179075)
1.251202***
(0.114768)
0.514951***
(0.041654)
0.623423***
(0.030155)
1.420242***
(0.223360)
0.747475***
(0.062188)

ln ( / )
-0.408364***
(0.133144)
0.015243
(0.092681)
-0.565473***
(0.014315)
-0.391839***
(0.032739)
-0.784496***
(0.057028)
-0.315255
(0.146280)

Long-run coefficients
ln(>) = 7 S ,

ln (/ )
ln ( / )
-2.035545** (0.784899) 0.220887*
(0.120288)
Short-run coefficients
ln(>) = 7 S ,

ln (/ )
ln ( / )
0.280367
-0.082710
(0.233861)
(0.073367)
-0.112581
-0.009625
(0.096360)
(0.005991)
0.112595
-0.13009***
(0.052182)
(0.007605)
0.117295** (0.027114)
0.002926
(0.009092)
1.200264** (0.317115)
-0.34638***
(0.012417)
0.084263
0.069632
(0.077302)
(0.042599)

ln (/ )
-1.001433***
(0.281870)

ln (/ )
0.186317***
(0.061233)
0.045260
(0.043487)
0.278712***
(0.029306)
0.105934***
(0.012866)
0.376870**
(0.101346)
0.124810**
(0.035594)

c) Model III: 5 countries and three explanatory variable)

Cases

Adjustment
coefficient

All
countries

All
countries
Brazil
China
India
Russia
South
Africa

-0.12243***
(0.037203)
-0.227627***
(0.004377)
-0.064666***
(0.000587)
-0.046854***
(0.000797)
-0.562507***
(0.015174)
-0.078684***
(0.001472)

ln(>)
3.073043***
(0.585842)

ln ( / )
0.344657**
(0.149000)

ln(>)
0.842451***
(0.104108)
1.030612***
(0.104719)
0.546403***
(0.034465)
0.58133***
(0.026750)
1.35472***
(0.171888)
0.64524***
(0.054775)

ln ( / )
-0.3974***
(0.082479)
0.228636*
(0.096920)
-0.5913***
(0.012107)
-0.4483***
(0.026819)
-0.4439***
(0.045163)
-0.194967
(0.151551)

Long-run coefficients
ln(>) = 7 S ,

ln (/ )
ln ( / )
-2.291279***
0.252061***
(0.706127)
(0.092608)
Short-run coefficients
ln(>) = 7 S ,

ln (/ )
ln ( / )
0.195265**
-0.057736*
(0.086941)
(0.030909)
-0.087487
-0.088359***
(0.086499)
(0.006094)
0.194748**
-0.101536***
(0.043771)
(0.006333)
0.144519***
0.015571
(0.022928)
(0.008586)
1.173511**
-0.451463***
(0.302887)
(0.010053)
-0.002830
0.039537
(0.067864)
(0.058783)

ln (/ )
-0.885854***
(0.253193)

vol(E)
0.024845**
(0.012602)

ln (/ )
0.064764
0.066303()
0.295247**
(0.052857)
0.224381***
(0.024206)
0.083573***
(0.011171)
0.323706**0.
(064650)
0.225998***
(0.030591)

vol(E)
-0.161279***
(0.038693)
-0.002061***
(1.62E-05)
-0.354891***
(0.007472)
-0.301000***
(0.006721)
0.006537***
(3.27E-05)
-0.373313***
(0.009758)

***, ** and * refer to significance level at 1%, 5% and 10%. Standard errors in parentheses.

27

Table 2C: Panel Cointegration Estimation: Pooled Mean Group Estimator (LICs)

a) Model I: 10 countries and one explanatory variable

Cases

Adjustment
coefficient

All countries

All countries
Angola
Ethiopia
Ghana
Indonesia
Kenya
Nigeria
Philippines
Rwanda
Tanzania
Uganda

-0.111716***
(0.027206)
-0.102129***
(0.012986)
-0.336409***
(0.003452)
-0.122792***
(0.003199)
-0.138577***
(0.004322)
-0.103606***
(0.002487)
-0.036397***
(0.000244)
-0.108977***
(0.003604)
-0.066496***
(0.000788)
-0.050526***
(0.001991)
-0.051252***
(0.000494)

ln(>)
0.316777
(0.223733)

ln(>)
0.764732*** (0.100349)
0.855895** (0.208818)
0.501835*** (0.033156)
0.311695** (0.072224)
0.809850*** (0.128066)
1.058192*** (0.039619)
0.846841*** (0.088541)
0.673336*** (0.075928)
0.850908*** (0.048690)
0.373877*** (0.057828)
1.364888*** (0.055526)

Long-run coefficients
ln(>) = 7 S ,

ln ( / )
ln ( / )
0.320488*
0.832994***
(0.132415)
(0.212990)
Short-run coefficients
ln(>) = 7 S ,

ln ( / )
ln ( / )
-0.352239***
0.287913**
(0.118025)
(0.131336)
-0.253948**
0.617786**
(0.044527)
(0.337611)
-0.971438***
0.383020***
(0.018734)
(0.041881)
-0.778050***
0.783258***
(0.043337)
(0.093796)
0.297808**
0.577364**
(0.063096)
(0.162552)
-0.460566***
0.179325** (0.045056)
(0.034020)
-0.492910***
0.859689*** (0.121913)
(0.011218)
-0.243259**
-0.198562*
(0.071328)
(0.083026)
-0.427341***
-0.007950
(0.006880)
(0.053100)
-0.292225***
-0.005919
(0.015115)
(0.065919)
0.099539*
-0.308875** (0.066104)
(0.032548)

***, ** and * refer to significance level at 1%, 5% and 10%. Standard errors in parentheses.

b) Model II: 10 countries and two explanatory variables

Cases

Adjustment
coefficient

All countries

All countries
Angola
Ethiopia
Ghana
Indonesia
Kenya
Nigeria
Philippines
Rwanda
Tanzania
Uganda

-0.141304***
(0.036735)
-0.183476***
(0.010533)
-0.411325***
(0.003066)
-0.230456***
(0.004595)
-0.157128***
(0.005046)
-0.119677***
(0.003269)
-0.027739***
(0.000211)
-0.113822***
(0.004514)
-0.100080***
(0.001786)
-0.035869***
(0.002403)
-0.033473***
(0.000409)

ln(>)
0.687212***
(0.225947)

ln(>)
0.763468***
(0.103235)
1.043815***
(0.167278)
0.447039***
(0.029137)
0.262230**
(0.069607)
0.807029***
(0.130655)
1.026971***
(0.037416)
0.796104***
(0.083229)
0.710150***
(0.072457)
0.745821***
(0.055946)
0.440817***
(0.057186)
1.354706***
(0.052564)

Long-run coefficients
ln(>) = 7 S ,

ln ( / )
ln (/ )
ln ( / )
0.211173**
0.385414
0.140663***
(0.098594)
(0.251218)
(0.041793)
Short-run coefficients
ln(>) = 7 S ,

ln ( / )
ln (/ )
ln ( / )
-0.364848***
0.265293**
-0.042973
(0.136587)
(0.135126)
(0.032613)
-0.289023***
0.620600**
-0.132596***
(0.034866)
(0.262359)
(0.004975)
-1.107956***
0.444276***
-0.139015***
(0.016134)
(0.033366)
(0.002781)
-0.885434***
0.643068***
-0.089409***
(0.051758)
(0.087026)
(0.003402)
0.386360**
0.492586*
0.045406**
(0.076477)
(0.163478)
(0.010344)
-0.404290***
0.169240**
0.074036***
(0.033373)
(0.044295)
(0.003699)
-0.507388***
0.946687***
-0.015557***
(0.010632)
(0.113823)
(0.001098)
-0.381052**
-0.142261
0.132843***
(0.071595)
(0.082116)
(0.014086)
-0.433120***
-0.064469
-0.009768*
(0.007076)
(0.054039)
(0.003485)
-0.090052**
-0.062814
-0.142793***
(0.026334)
(0.065823)
(0.005772)
0.063473
-0.393982***
-0.152872***
(0.031394)
(0.065113)
(0.006981)

ln (/ )
-0.350725***
0.095940()

ln (/ )
-0.014364
(0.109360)
-0.716392***
(0.099287)
0.112293***
(0.016147)
0.171048**
(0.042679)
-0.053054
(0.074396)
-0.153491***
(0.021064)
0.501909***
(0.043609)
-0.382858***
(0.046927)
-0.020404
(0.031347)
0.100866**
(0.030611)
0.296444***
(0.022795)

***, ** and * refer to significance level at 1%, 5% and 10%. Standard errors in parentheses.

28

c) Model III: 10 countries and three explanatory variables

Cases

Adjustment
coefficient

All countries

All countries
Angola
Ethiopia
Ghana
Indonesia
Kenya
Nigeria
Philippines
Rwanda
Tanzania
Uganda

-0.086261***
(0.024822)
-0.001548***
(4.27E-07)
-0.136337***
(0.000650)
-0.094831***
(0.000562)
0.000108***
(1.02E-05)
-0.248868***
(0.001492)
-0.042414***
(0.000194)
-0.145762***
(0.000584)
-0.075394***
(0.000518)
-0.107271***
(0.000337)
-0.010293***
(5.76E-05)

ln(>)
1.066903***
(0.212739)

ln(>)

0.563134***
(0.129416)
0.936938***
(0.125498)
0.365378***
(0.027949)
0.093121
(0.047023)
0.807090***
(0.124859)
0.261664***
(0.013310)
0.732966***
(0.061339)
0.319883***
(0.022588)
0.708706***
(0.036318)
0.066797**
(0.019542)
1.338799***
(0.054160)

ln ( / )
-0.287286***
(0.086673)
ln ( /

-0.423972***
(0.099213)
-0.372509***
(0.025896)
-0.816907***
(0.015748)
-0.691301***
(0.035033)
-0.006073
(0.086557)
-0.121422***
(0.012865)
-0.632917***
(0.009143)
-0.647025***
(0.022619)
-0.345743***
(0.005333)
-0.665093***
(0.013354)
0.059272
(0.032875)

Long-run coefficients
ln(>) = 7 S ,

ln (/ )
ln ( / )
0.272016
0.010920
(0.255193)
(0.049502)
Short-run coefficients
ln(>) = 7 S ,

ln (/ )
ln ( / )

0.189612
(0.116225)
0.422490
(0.181194)
0.34775***
(0.034343)
0.310028**
(0.062187)
0.620197**
(0.160934)
-0.038091*
(0.012417)
0.77990***
(0.085053)
0.106018**
(0.023882)
-0.161643**
(0.041584)
-0.127586**
(0.023788)
-0.362955**
(0.067288)

-0.016775
(0.028716)
-0.171725**
(0.003968)
-0.03830***
(0.002945)
0.03351***
(0.002317)
0.06723***
(0.010901)
0.11217***
(0.000995)
-0.02030***
(0.000807)
0.05421***
(0.003975)
-0.012847**
(0.002458)
-0.04128***
(0.002109)
-0.15042***
(0.007855)

ln (/ )
-0.018202
(0.128177)

ln (/ )

0.017332
(0.095085)
-0.585533**
(0.088576)
0.13374***
(0.019193)
-0.079616*
(0.025340)
0.029467
(0.068542)
-0.05375***
(0.007361)
0.59288***
(0.037783)
-0.044644*
(0.015794)
0.027850
(0.023495)
-0.12341***
(0.010226)
0.27633***
(0.027592)

***, ** and * refer to significance level at 1%, 5% and 10%. Standard errors in parentheses.

29

vol(E)
-3.216100***
(0.419248)
vol(E)

-0.020308
(0.031332)
-0.005968***
(7.06E-06)
0.065656***
(0.007485)
-0.042398***
(0.002733)
-0.017159
(5.17E-05)
0.164651***
(0.003452)
-0.125337***
(0.002453)
-0.089533***
(0.002379)
-0.028748***
(0.002997)
-0.177574***
(0.003046)
0.053328***
(0.002087)