You are on page 1of 19

Empirical Studies in Social Sciences

6th International Student Conference, Izmir University of Economics, Izmir Turkey

TESTING THE BALASSA-SAMUELSON HYPOTHESIS: EVIDENCE FROM 10 OECD COUNTRIES
Cem Tintin Department of Economics, Lund University P.O. Box 7082, S-220 07, Lund, Sweden E-mail: mas08ct1@student.lu.se
This study tests the Balassa-Samuelson (BS) hypothesis for 10 OECD countries between 1975 and 2007 using the Johansen cointegration approach. The cointegration analysis confirms the existence of cointegration among the real effective exchange rate, relative productivity and terms of trade in 10 OECD countries. The study¶s major findings suggest that the BS hypothesis still works well in OECD countries while explaining the long-run movements of the real effective exchange rates that it holds for 8 out of 10 countries in the unrestricted model.

Key words: The Balassa-Samuelson Hypothesis, Real Effective Exchange Rate, Cointegration JEL Classification: C12, C13, F31

1. Introduction
The investigation of the differences in exchange rates across countries has always been an interesting and tough job for economists. The history of attempts that trying to explain the differences in exchange rates across countries might go back till 16th century, however, the modern theories of exchange rates start with the Purchasing Power Parity (PPP) theorem. In 1918, Gustav Cassel, a famous Swedish economist, proposed the Purchasing Power Parity (PPP) theorem which based on law of one price, and states that in the long-run exchange rates should be identical across countries. In other words, ³the PPP theory predicts that, in the long-run, relative prices determine the exchange rate (i.e., e=P/ P*); and any deviation of relative prices from the equilibrium exchange rate will be transient and ultimately mean-reverting in the long-run´ (Chowdhury, 2007, p.4). Nonetheless, empirical studies reject this version of (absolute) PPP theorem. 1 Apart from reasons such as transaction costs, transportation costs and inefficient markets, the most convincing explanation came from Balassa (1964) and Samuelson (1964) -known as the BalassaSamuelson hypothesis- which states that the productivity differences in tradables and nontradables sectors across countries lead to differentiation of wages, price levels and, hence the real exchange rates. In particular, the BS hypothesis explains the two effects: (i) The price level differences across countries (The Penn Effect): According to the BS hypothesis, when the productivity level of tradables sector in home increases relative to foreign country¶s tradables sector, home experiences higher price level due to increase in
1

Antweiler (2008) and OECD (2005) give clear definitions of PPP. And see (Pilbeam, 2006, p.135-139) for empirical studies which reject the PPP theorem.

1

All cointegration tests verify the long-run relation among the real effective exchange rate. In particular. Third. we concentrate on and test the second effect of the BS hypothesis regarding the real exchange rate differences across countries. and assumptions are explained. First. the paper tests the BS hypothesis not only for a single country but also for 10 OECD countries for the robustness of the analysis. the ³terms of trade´ is added as the second explanatory variable. which can be seen as a test of ³the extended (unrestricted) BS model´. Section 1. relative productivity and terms of trade. which is an explanation of the high price levels in rich (high per capita income) countries. (ii) The real exchange rate differences across countries (The BS effect): The BS hypothesis claims that in a country where the productivity of tradables sector is higher than the other country. since it was (re)discovered in 1964. the OECD average of labor productivity is used as the benchmark to calculate the relative productivities in the analysis. The next section briefly explains the importance of the BS hypothesis. Throughout the study. The study differs from others in three respects. The findings of the study mostly confirm the validity of the BS hypothesis. 2 . the popularity of the BS hypothesis has increased over time.3 The main conclusion of the study is that the BS hypothesis still keeps its importance in explaining the real exchange rate movements across OECD countries. in addition to the relative productivity explanatory variable. This increase in the real exchange rate index -which is known as real appreciation. 1. According to a survey conducted by Tica and Druzic (2006). Second.P*). the countryspecific estimations indicate that the BS hypothesis is valid for 7 countries out of 10 when the terms of trade variable excluded (the original BS model). even though the results are country and model specific. in which they explain the real exchange rate differentials across countries with productivity differences. This study aims to test the validity of the BS hypothesis for 10 OECD countries between 1975 and 2007 using the recent data sets and econometric methods.1 Why does the Balassa-Samuelson Hypothesis Get a Great Deal of Attention? After the seminal papers of Balassa and Samuelson in the same year (1964). and estimation results. Section 3 reviews the main findings of the study and concludes. Section 2 revisits the BS model formally in which the deterministic and empirical BS models are derived. The organization of the study is as follows. 2 This is known as the Penn Effect. See Alexius and Nilsson (2000) for a use of similar name. When the terms of trade included (the extended BS model). then the real exchange rate index (R) of this country will be higher.2 gives a literature review in comparative perspective paired with a summary literature results table. the BS hypothesis is rejected only for 2 countries out of 10. ³In total.stems from the definition of the real exchange rate in which price level of home (P) stands in the numerator and price level of foreign country (P*) and nominal exchange rate (e) stand in the denominator (R= P/e. the theory has been tested 58 times in 98 countries in time series or panel analyses and in 142 2 3 Assume that the productivity levels of nontradables sectors are identical in two countries. cointegration tests.general wage level. Section 2 also describes the data sources and transformation moreover it presents the unit root tests.

2 Literature Review In here. they develop a simple model of a small open economy producing exportable and nontradable goods and consuming importable and nontradable goods. (Tica and Druzic. They use the Johansen cointegration approach 3 .51 with a positive intercept term. 1. p. oil prices and openness. In their influential study.´ This result implies that 1% increase in per capita income levels lead to 0. we present the summary findings of some selected studies in Table 1. 1994. especially the sectoral productivity databases enable scholars to test the BS hypothesis without assuming all sectors have the same productivity level within a country. In his pioneering study. we discuss some selected empirical studies regarding the BS hypothesis in which authors used similar methods and variables with us. country-specific BS coefficients have been estimated 164 times in total.countries in cross-country analyses.51% increase in real exchange rate levels. Balassa (1964) employed OLS analysis in order to estimate the equation in which real exchange rates were used as the dependent variable and per capita income levels as the independent ones.4) Some of the main reasons can be counted as follows why the BS hypothesis preserves its importance and popularity in empirical economics: (a) The continuing desire of researchers to explain why law of one price fails across countries. (e) The theoretical contributions to the theorem by adding additional variables such as terms of trade. p. but also affect the long-run economic growth thereby the development of the country. Froot and Rogoff (1994) note that ³Balassa (1964) reports a regression for a cross-section of twelve industrial countries for the year 1960 in which the estimated BS effect was 0.´ (De Gregorio & Wolf. (f) The availability of new data sets throughout the time. (b) The examination of the determinants of real exchange rates is economically very important for economists. they conclude that ³The evidence from OECD countries broadly supports the predictions of the model. De Gregorio and Wolf (1994) successfully integrated the ³terms of trade´ into the BS model. This enables researchers to test the BS hypothesis by using various kinds of models and techniques. Since the real exchange rates not only affect the net trade and current account balance of a country. and present empirical evidence for a sample of fourteen OECD countries. (d) The invention of new econometric techniques and easy implementation of the techniques via new econometric software programs. Alexius and Nilsson (2000) use the terms of trade and relative real GDP (as a proxy of productivity) to explain the real exchange rate movements in 15 OECD countries from 1960 to 1996. In these estimates. and at least once for 65 different countries´. Then. namely that faster productivity growth in the tradable relative to the nontradable sector and an improvement in the terms of trade induces a real appreciation. (c) The desire of researchers to explain the high price levels in developed countries. 2006. Clearly.i) In a benchmark article for our study.

And they report the presence of cointegration among three series in all countries and estimate the correct sign (positive) for the BS effect in two thirds of the cases. tradables sector produces the goods which can be traded within country and internationally. 4 .9 and 1. openness.in search of cointegration relation among variables and estimate the BS model by using FMOLS method. In this framework. 2... (1) Simply. Assume that labor (L) is immobile internationally whilst perfectly mobile within the country. Non-tradables sector produces the goods in a country which cannot be traded internationally. terms of trade Per capita GDP. Recall the real exchange rate definition.2 between 0.6 and 1. And capital (K) is perfectly mobile internationally. In contrast. After Balassa (1964) and Samuelson (1964) and with theoretical contribution of De Gregorio and Wolf (1994) regarding the ³terms of trade´.5 between 0.3 and 2. Price levels VAR-ECM.51 between -1 and 1. And * (asterisks) denotes the foreign country.9 between -0.. Summary of Some Selected Studies Regarding the BS Hypothesis Authors Balassa (1964) Alexius and Nilsson (2000) Egert (2002) De Broeck and Slok (2001) Choudhri and Khan (2005) Drine and Rault (2005) Dependent Variables RER RER Independent Variables Per capita income Domestic and foreign GDP.. terms of trade Labor productivity differences. the BS hypothesis predicts that PRO and TOT variables are assumed to have a positive effect on the real exchange rate. Pedroni cointegration 2.4 between 0 and 3. Johansen cointegration Pooled mean estimation DOLS Estimated BS effect 0. which necessarily implies e =1 and . the BS model is presented formally and an empirical BS model is derived to use in our estimations.1 The Basic Framework of the Balassa-Samuelson Model Let there be 2-country and 2-sector world in which T stands for tradables and N stands for non-tradables sector. international is normalized trade assumed to be equalize prices of traded goods in two countries that to 1. today widely accepted the BS model can be formalized to explain the real exchange rates as follows: RER= f (PRO. terms of trade Labor productivity growth Productivity. Table 1. The Balassa-Samuelson Model In this section.. productivity differences Method OLS Johansen cointegration and FMOLS VAR.5 REER REER RER REER. TOT) «««««««««««««««««««««««««.

(4) (e = Nominal exchange rate. nominal wages ) are equalized across sectors within the country. plugging (8) and (9) into equation (5) and taking the natural logarithms yield the deterministic BS model equation: «.. 5 . 2007... whose productivity growth rate is higher than the foreign one in tradables. in home country an increase in the productivity of tradables relative to foreign country should associate with an increase in the real exchange rate level of home country. Put differently home country.. ««««««««««««««.e.208-212) explicitly show the derivation of this type of equation. p. p. (11) is known as the BS effect and theoretically In equation (11)..««««««««««««««««««««««  Now we turn to the production side of the economy.. in the long-run as an effect of ³perfect labor mobility´ within the country. (9)  Under the assumption of ³the shares of labor income in non-tradables and tradables sectors are equal´6 .«« (8) .) If we divide (3) with (4) and apply the assumption = 1..5 That is to say.. W= profit maximization context....1745-1747).. (3) In where.. However.. Hence. defined as the price of the domestic currency in terms of the foreign one.885-886) and (Lothian & Taylor.1405). p... 6 in the If we do not make this assumption there should be an additional ratio in front of square brackets.... And by using the first order conditions of equations (6) and (7) in (i. they also use the equation (10) in the sense that empirics showed that this ratio is almost one... 5 For an extensive exposition of the derivations see (Dumitru & Jianu.. and the marginal product of labor equals to the nominal wage (factor price of L) in two sectors... one may show that the marginal product of capital equals to the world interest rate (factor price of K)....... the relative price level (the real exchange rate) between the countries depends solely on the relative price of non-tradables´ (Muscatelli et al. (general price level equation in foreign country). (2) (general price level equation in home country)... which represents the shares of the labor income in non-tradables and tradables sectors..«««« (10) With the assumption that = ...«. (Obstfeld & Rogoff.. p...«.. 2009.... the real wage levels are identical in T and N sectors ).... then non-tradables sector vanishes from equation (10) and it becomes..««««. ««««. ³in a two-good world.. firstly experiences an increase in wage levels in both sectors due to factor price equalization across sectors within 4 In other words. Thus the BS hypothesis suggests that...... 1996.«««««««««««««««««««««««««... equation (2) can be rewritten as 4: RER= (5) «... 2008.. we can write... the coefficient assumed to be positive.

e. we get the basic empirical BS model 7: ««««««««««««««««.. many studies such as Balassa (1964)... Before converting it into an empirical model. Consequently.i) If we denote the relative productivity as ³PRO´. it is crucial to define the variables and coefficients of the equations (12) and (13).. Then the general price level. we need to convert it into an empirical one.11) in our estimations. ³the real effective exchange rate is computed as the weighted geometric average of the price of the domestic country relative to the prices of its trade partners´. 2. In this respect.. The definitions of the variables and expectations about the coefficients are as follows: : According to IMF. simply we can write.g. there is no consensus among scholars today. non-tradables sectorµs productivity growth rate can be negligible. (12) And as we mentioned. (13. goes up in home due to high wages level. in empirical studies it has been widely employed such as in Egert (2002) and Lothian and Taylor (2008). relatively high price level in home leads to an increase in RER. and Jaunky (2007). we add the terms of trade to equation (12) and get equation (13). In fact. «««««««««. Moreover. the derived empirical BS model (eq.ii).. which use relative GDP as a proxy of relative productivity of countries implicitly make this assumption..2 The Empirical Balassa-Samuelson Model In order to use the deterministic BS model (eq.12) is used as the restricted BS model. it is beneficial to explain the assumption that we make in equation (11) that ³there is no productivity growth in non-tradables sector over time´ ( = ). And CPI-based REER can be expressed as 8 : 7 Note that we use REER in eq. Alexius and Nilsson (2000). there is no long time series data of sectoral productivity for cross-section analysis. this assumption implies that relative to tradables sector. The rationale behind this assumption is two-fold for our study. about the classification of some sectors whether they are tradables or not. since we use REER data in model estimations..13. for the robustness of the BS hypothesis test of 10 countries.the country... And if we write equation (11) as a stochastic model with an intercept term in natural logarithmic form and eliminate the T (sectoral) subscript. Thus. is used as the benchmark equation (unrestricted model) and the model without the terms of trade (eq.ii) In our estimations.. Although it seems as a strong assumption. ««««««««« (13.2). CPI.12 instead of RER. First. especially after the study of De Gregorio and Wolf (1994) the terms of trade has been used widely as a second independent variable in the BS model such as in Choudhri and Khan (2005). Second.. the assumption we make seems plausible in point of view of theory. Before proceeding to data description section.. from the definition of the real exchange rate (eq. which is known as real appreciation. which includes the terms of trade. 6 .

and coefficients can be interpreted as the elasticity of the PRO and TOT with The respect to REER. 7 . whereas a decrease implies a real depreciation. In this type of REER definition. see IMF (2008).. Ri : nominal exchange rate of home country in US dollars. ³Labor productivity (output per employed person) in manufacturing sector´ is chosen as the proxy of productivity in tradables sector.««««««««««««««««««««««««.Samuelson (BS) effect. : Labor productivity in manufacturing sector (tradables) in country i. UK. the dataset is constructed with 10 cross-section units and 33-year sample. Nonetheless. i : Canada. the selected 10 countries have similar properties in the sense that all of them are OECD countries and developed countries in terms of per capita income by the end of 2007. The real effective exchange rate data are extracted from World Development Indicators. Simply. (14) In where . 2. : The Balassa. an (%) increase in ³the relative productivity´ and an (%) increase in ³the terms of trade´ in country i. : Unweighted mean of ³Labor productivity in manufacturing sector (tradables) of 14 OECD countries´. Thus. Zanello and Dominique (1997). Sweden.5 is to find out the sign and size of these coefficients for 10 countries to test the validity of the BS hypothesis. Denmark. Rj : nominal exchange rate of country j¶s currency in US dollars.. simply ³average productivity of OECD´. Italy. In other words. t : 1975. since the independent and dependent variables are in natural logarithmic forms. the expected positive signs of the coefficients point out that. Netherlands. More precisely. Pi : home country¶s CPI index. Japan. Wij: country j¶s weight for home. : Relative productivity in where. USA. Pj : price index of country j (foreign). WDI uses the IMF-IFS statistics database as the main source to extract the CPI-based real effective exchange rate data. theoretical expected sign is positive.3 Data Sources and Description of Data To test the BS hypothesis. Norway. the main goal of the estimations in section 2. CBRT (2008). should associate with an (%) increase in the real effective exchange rate. our CPI-based REER data set is taken from the IMF-IFS via WDI database which uses 2000 as the base year (2000=100). 1976«2007. Cross section units (countries) and data length are selected according to data availability. : The terms of trade (TOT) effect. theoretical expected sign is positive. an increase in the index denotes a real appreciation of the home currency. Although ³labor 8 For a detailed exposition of calculations and weights. : Terms of trade (relative prices of country i¶s export to import). 10 OECD countries are selected for the 1975-2007 period. Germany. It is worth mentioning that there is a long lasting debate about the best productivity proxy issue in the literature.

Department of Labor. And to construct the relative labor productivity data the ³unweighted mean of labor productivity of 14 OECD countries´ is used as the benchmark. Denmark. R-square and Durbin-Watson (d. In time-series econometrics. to calculate the relative productivity for Canada in 1980. we follow a two-step approach to test the BS hypothesis for 10 OECD countries: y First. and according to estimation results. Sweden. 2006. although t values of the coefficients are significant. p. it is known as integrated of order 1 or I (1). Netherlands. More importantly. we multiply it by 100 in order to make the scale of the TOT index 100. Belgium.5. Then. we test whether or not. and USA. we estimate the coefficients of the BS model (equations 12 and 13). The terms of trade is defined as the relative prices of a country's export to import. some authors use GDP as a proxy and some argue that total factor productivity (TFP) is a better proxy than the labor productivity. as mentioned in section 2. Bureau of Labor Statistics´ in March 2009 which uses 1996 as a base year (1996=100). In a spurious regression. we get the raw terms of trade data. there is cointegration relation among REER. By taking the ratio of two prices.11) These 14 countries are Australia. we employ the country-specific unit root tests to identify the stationarity of the series. Before the calculation of the relative productivity data. As usual before starting the cointegration tests and estimations. And if a series is becoming stationary after taking the first difference. Canada. the estimated 9 See the discussion on this issue in (Tica & Druzic. The second explanatory variable of the model ³terms of trade´ data are calculated from the IMF-IFS data base. in section 2. Spain.S. Norway. tests to detect the existence of stationarity (or nonstationarity) in the series are known as unit root tests. export and import unit prices of 10 countries are extracted from the IMF-IFS data base in which 2000 is used as the base year (2000=100). France. y Second. Germany.productivity´ has been used in empirical studies widely. we decide whether the BS hypothesis holds or not. 10 8 . moreover comparable long time series data are not available for our study. Japan. first we calculate the ³unweighted mean of labor productivity in manufacturing sector of 14 OECD countries´ in 1980. However.10 For example. Italy.w) values are low. as the REER and relative productivity indices.4 Methods As mentioned in introduction. it is always mentioned that TFP is difficult to calculate for different type of sectors and countries. Then Canada¶s labor productivity index in 1980 is divided by this mean value to get the ³PRO´ (the relative productivity) data for Canada. we converted the base year 1996 into 2000 for consistency with two other variables¶ base years. PRO and TOT variables. Then. one of the distinguishing features of the study is the use of Johansen approach to test the BS hypothesis for 10 OECD countries. 9 The labor productivity data in manufacturing sector are taken from the publication of ³U. In the econometrics literature. UK. non-stationary series (unit root) phenomenon is crucial in the sense that linear combinations of these series may generate spurious or nonsense regressions. 2. In this regard.

the Engle±Granger (EG) two-step approach is the usual one while testing the existence of cointegration. The ADF unit root test results of country-specific series are documented in Table 2. Thus.coefficients are biased.53 is not significant at 5%. -1. It is worth mentioning that in both approaches the necessary condition to search for cointegration is that ³integration order of series should be the same´. all series should be I (d). Before cointegration tests. We employ the Johansen approach to test the existence of cointegration in country-specific series since we have three series. our first task is to conduct the unit root tests in order to understand whether the series are stationary or not. or equilibrium.´ (Gujarati. Since. a linear combination of two or more time series can be stationary´.57 which is significant at 5%. if there are only two series. In this respect. moreover it has some distinct advantages on the EG approach such as it is not a two-step approach. The ADF test uses the null hypothesis that ³series has a unit root´. two variables will be cointegrated if they have a long-term. For the level (the original values of series) we accept the null. 9 . One way to escape from spurious regression in existence of non-stationary series is looking for a cointegration relation among variables. 2003. In case of more than two series. H1: REER (Canada) does not have a unit root. we conclude that REER of Canada is I (1). the Johansen approach is the most applicable one. Simply. Cointegration defined as: ³Cointegration means that despite being individually nonstationary. relationship between them. and we reject the null. the unit root test for Canada¶s REER series is conducted as follows: H0 : REER (Canada) has a unit root. ³Economically speaking. p. For example. However if we take the first difference t-value becomes -5. detecting the existence of unit root in time-series becomes vital in order to get rid of spurious regression risk. We mainly use the Augmented Dickey±Fuller (ADF) unit root test for individual series.822 and 830) In the cointegration literature.

For consistency.0001) -2.0021) -3.3368) (0.194638 -3. except series with (*).6472) ( 0.141961 -4.9917) ( 0.043640 ( 0.921323 ( 0.0037) -2.875087 -4.389405 -4.780036* ( 0.680097 ( 0.0033) -1. the Phillips-Perron unit root test showed that the series is I (1). 10 .4484) ( 0.0056) -2.5040) ( 0.3666) ( 0.1229) ( 0. However.5366) ( 0.785665 ( 0. when we conduct the unit root tests for all country-specific series. (3) Bold numbers denote that they are significant at 5% level.216767 -4.0007) NORWAY -2.0032) -2.1106) ( 0.4485) ( 0. In addition.248376 -4.1195) ( 0.9035) ( 0.0001) 0.0760) ( 0.0048) -2.248193 -5.161792 -4.0004) DENMARK -2.0235) -2. (5) Tests are conducted for series with (*) by including ³intercept´.759665 ( 0.025691 -4.511697 ( 0.3777) ( 0.0432) -2.412536 -5.152128 -4.891747 -4.0024) JAPAN -1.397041 -5.739137* ( 0.383519* -3.8351) ( 0. we reported the Phillips-Perron test result for the TOT series of Netherlands that we marked the values with (**) in Table 2.688790 -4.0058) -1.1544) ( 0.559828** ( 0.0027) NETHERLANDS -2.036180 ( 0.080568 -4.762738 ( 0.141317 -6. (7) Automatic lag length selection (Schwarz) is used with maximum 8 lags.570857 ( 0.508787* -5.0175) -2.9448) ( 0.0015) -0.694513 ( 0. ADF Unit Root Test Results of Country-Specific Series REER Level First difference CANADA -1.094322* ( 0. (2) Probabilities are in parentheses.0374) -1.8729) (0.864695 ( 0.631978 ( 0.2474) ( 0.487358 ( 0.356682 -4.0072) DECISION REER is I(1) for all series VARIABLES PRO Level First difference -3.086615 ( 0.2992) (0.0045) -2.3740) ( 0.119237* -4.239519 ( 0.423442 ( 0.0016) TOT is I(1) for all series Notes: (1) Values without parentheses are t-statistics.528129 ( 0. 11 That is to say.421048** -4.184688 (0.117222 -5.0062) GERMANY -3. (6) Values with (**) are Phillips-Perron t-statistics.735821 ( 0.167158 -4. the Elliot-RothenbergStock unit root test also confirmed that this series is I (1).739853* -4.0006) -1.476182 -5. In a similar fashion.160361 ( 0.499314 -4.0052) -0. we can search cointegration relation among the series.479582 ( 0.4763) ( 0.6959) ( 0.341529 ( 0. (4) All tests are conducted by including ³intercept and trend´.1833) ( 0.0039) USA -3.535167 -5.7950) ( 0.0019) UK -2.287677 ( 0.Table 2.0002) ITALY -2. we conclude that all series are I (1).0005) SWEDEN -2.5 Country-Specific Cointegration and Estimation Results 11 Only TOT series of Netherlands was found I (2) by the ADF test. 2.9911) ( 0.4648) ( 0.287855 -3.8088) ( 0.963406 ( 0.868569 -4.0118) PRO is I(1) for all series TOT Level First difference -1.0099) -1.966596* -4.4908) ( 0.1417) ( 0.287496* (0.618982 ( 0.

(ii) Decision about whether or not to add a trend (deterministic or quadratic) component to equation (16) and trend assumption in level data.. p. PRO. and TOT variables for each country using the Johansen approach. p. The specification of equation (16) regarding (i) and (ii) significantly affects the coefficients in matrix.«« (16) ³ contains information about the long-run relationships among the variables included in vector Z. in which optimal lags are selected by three different information criterions.Since all country-specific series are found I (1). (Harris & Sollis. that we only reported three information criterion results. Put differently. 12 Stock and Watson (1993) show that Johansen¶s cointegration test is sensitive to the lag lengths used in the VAR models.195) Johansen (1988) estimates equation (16) by using maximum likelihood method and extracts trace and maximum eigenvalue ( -max) statistics to determine the number of cointegrating vectors. Johansen (1988) demonstrates that could be decomposed into the product of two p×r matrices. PRO and TOT variables as the BS hypothesis suggests. By using this approach we constructed Table 3. this approach is used in other studies such as in Alexius and Nilsson (2000). 14 For instance. 2005. This is not surprising.. in the last column we stated the ³preferred lags´ as a conclusion of our search process. These are: (i) Optimal lag length of the VAR system (K in eq. As seen in Table 3. there is a long-run relation among REER. After the influential papers of Johansen (1988) and Johansen and Juselius (1990). In addition. Ahking (2002) and Turner (2007) find out that the Johansen approach is sensitive to the specification of the deterministic terms (trends). we can seek cointegration relation among REER. The aim of the cointegration test is to understand whether or not.12 In order to decide the optimal lag length in our VAR model.««. 13 For example. 11 . the Johansen approach is sensitive to the specification of VARequation. Johansen (1988) expresses the following VAR system: ««««««««. Johansen (1988) also reports the matrix which contains the normalized and coefficients. ³Preferred lags´ are the lags which give the best results in terms of ³rank of cointegration´ and ³significance of estimated coefficients´ in matrix. there are two important things to decide before running the equation (16) for our model. Number of cointegrating vectors r is determined by the rank of .. we employ the ³information criterion (IC)´ approach.. the Johansen approach has become popular in the cointegration literature. In the literature. 14 In addition. However. = ' where the elements of matrix form the long-run cointegrating coefficients and the elements of matrix form the adjustment parameters. there is no consensus among scholars which information criterion (IC) is the best one.´ (Bahmani-Oskooee & Economidou. Khim and Liew (2004) report the AIC as the best information criterion in determining the optimal lag length. i. 13 In this regard.16). sometimes the preferred lags are not determined by any three information criterions.117) suggest the HQC. 2009.e. we did not report FPE (Final prediction error) criterion.

) There is no a systematic approach to decide whether or not to include a trend component into VAR models such as the IC approach for optimal lag length.VAR lag selection command.Table 3. (Model 1 and Model 2 are found irrelevant for our analysis. we follow two steps and test the following hypotheses: Step 1: H0 : r = 0 (no cointegration). H1: r ” 1 (at most one cointegration relation) 12 .1-under Johansen cointegration command-. and HQC: Hannan-Quinn criterion. as we did in ³lag length selection´. Table 4 presents the trace statistics and max-eigenvalue results that we use to determine the ³rank of cointegration´ among three variables. For example. AIC: Akaike criterion. y Consistency between the ³rank of cointegration´ test results of trace statistics and max-eigenvalue. determining the trend component and its type in VAR models is more problematic. Some authors. we follow a similar ³search process´. We tried all five specifications with ³preferred lags´ in countryspecific VAR models and reported the best results in the last column of Table 3 according to: y Significance of estimated coefficients in matrix and trend component. and linear trend in VAR) yielded the best results and in others at least a linear trend added to the model to capture the dynamic structure of the series. the model 5 (quadratic trends in level data. Thus. Model 4: Level data and VAR have linear trends. we run the VAR models for each country. Using the preferred lags and chosen model specifications. and linear trend in VAR. BIC: Schwartz Bayesian criterion. But in here. Model 3: Linear trends in level data. Lag Length and Model Selection CANADA DENMARK GERMANY ITALY JAPAN NETHERLANDS NORWAY SWEDEN UK USA AIC 4 4 5 5 5 1 1 1 2 5 BIC 1 1 1 1 1 1 1 1 1 4 HQC 4 2 5 5 5 1 1 1 1 5 Preferred lags 5 4 4 5 4 5 5 5 2 5 Preferred Model Model 5 Model 3 Model 4 Model 4 Model 5 Model 3 Model 3 Model 4 Model 4 Model 5 Notes: Lag length selection is conducted by using Gretl . not in VAR. In E-views there are five different trend specification options under the Johansen cointegration approach. to find the rank of cointegration among three series of Canada. ³Model numbers´ are the same as in E-views 5. In most cases. Model 5: Quadratic trends in level data. such as Ahking (2002) and Cheong (2005) prefer reporting VAR model estimation results with trend and without trend to tackle with this problem. with k-max=5.

for each country at least one cointegration relation has been found.76 is not significant at 5%.96371 (0.699328 (0.12569* ( 0.7477) Rank 1 1 2 3 2 2 2 3 1 2 Notes: (1) Values in parentheses are probabilities.00705* 3 29.0215) 22.0001) ( 0. We reach the same conclusion.0015) ( 0.41837* 20.40175* ( 0.219398 ( 0.96144* 36.0001) 22.547679 1 26.626* 1.24862* 23.0010) 98. 80.1409) ( 0. PRO.0000) ( 0.66074* ( 0.0000) 113. That is to say.17209* 26.44064* 20.1560) 32. Thus we accept the null hypothesis.00705* (0.547679 ( 0.0300) ( 0.0285) 63.0414) 6.08409 2 46.0235) ( 0.0322) 55.99207* 3 49. if we apply the same steps by using maximum eigenvalues.0107) ( 0.3660) ( 0. Step 2: H0: r ” 1 (at most one cointegration relation).0141) 17. 86. we accept the alternative hypothesis.0390) 18.0000) 73.2988) ( 0.0000) ( 0. according to maximum eigenvalue rank test the rank of cointegration among Canada¶s three series is 1.0000) ( 0.0006) 64. In other words.1738) (0. H1 : r ” 2 (at most two cointegration relations) Since the probability of maximum eigenvalue is 0.64898* 16.0414) ( 0. r ” 2.25 is significant that its probability is 0.95751* 2. there is cointegration (long-run relation) among REER.0000) 62.39815* ( 0. that implies 4.99207* ( 0.981138 1.78245* 11.0000) ( 0.2978) ( 0.87 that implies 5.0112) (r 2) 1. we see that the rank of cointegration among three series ranges between 1 and 3.0161) ( 0. Hence. hence we accept the alternative hypothesis.66304 0.0671) 26.85273* 21. Test Results For Cointegrating Rank Unrestricted Cointegration Trace Rank Test Unrestricted Cointegration Maximum Eigenvalue Rank Test CANADA DENMARK GERMANY ITALY JAPAN NETHERLANDS NORWAY SWEDEN UK USA tr (r = 0) 86.91.817101 ( 0.0059) (0.74183 2 65.62224* 3.0235) ( 0.0000) ( 0.04062* ( 0. So far.15384 ( 0.74183 ( 0.9193) 11.0016) tr (r 1) tr (r 2) Rank (r = 0) (r 1) 5.168101 ( 0.55111* ( 0.2695) 0.7477) ( 0.1332* (0. thus we accept the null. according to trace rank test the rank of cointegration among Canada¶s three series is 1.Using the trace statistic¶s critical value at 5%.0032) 33.23475* ( 0.699328 1 30.000) 42. If we repeat the same steps for 10 countries.219398 1 80.0001) 2.761740 ( 0.0002) 44.23 is too significant that its probability is 0.1409) 1.0000) 50.4010) 27.70152 6.18001* ( 0. 13 .08409 ( 0. H1: r ” 1 (at most one cointegration relation) At 5%. H1 : r ” 2 (at most two cointegration relations) Since the probability of trace statistic is 0.2978) 0.2695) ( 0.4030) (0.78941* ( 0.1335* (0.0052) ( 0.13274* 13.0671) ( 0.0464) ( 0. So far.0183) ( 0.63518* ( 0. Although the results of the rank tests are countryspecific.0042) ( 0.07275* 12.3941) 3.3941) ( 0.98 is not significant at 5%.3660) 13. Step 2: H0: r ” 1 (at most one cointegration relation). and TOT variables in each country.25361* 4.27498* (0.98727* 10.8703) ( 0. (4) There is full consistency between rank test results of trace and maximum eigenvalue.54191* ( 0. (2) The null hypotheses are r = 0.91435* 21. r ” 1.87086* ( 0.21885* 0. Step 1: H0 : r = 0 (no cointegration).77427* ( 0.168101 2 27.961498 2 28. (3) * denotes the rejection of the null at 5% level. Table 4. we reject the null.961498 ( 0.817101 2 50.4030) 11.65031* (0.0000) ( 0.

00 1.00 1 (PRO) 4.00 1.500234 (8. If we take the REER as the dependent variable.584104 (3. in which estimation results of the unrestricted (equation 13) and restricted models (equation 12) are reported. we decide whether the BS hypothesis holds or not. size and sign of the normalized coefficients. Canada has unrestricted models due to positive and significant the biggest coefficient in both the restricted and unrestricted models among 10 countries with values of 4.994009 (9. 14 .5%.45) 1.690840 (3. These results imply that in Canada.00 1.2) -2.25) 0. Put another way. In addition is significant and equals to 1 for Canada. Table 5.019265 (2.84) RESTRICTED MODEL (Restriction 2 =0) (eq. The interpretations of country-specific estimation results are as follows: Canada is a country in which the BS hypothesis holds in both the restricted and coefficients.12) -1.75.00 1.00 1. respectively.012) -1.78) 0. (2) All values are significant at 5% level.13) REER CANADA DENMARK GERMANY ITALY JAPAN NETHERLANDS NORWAY SWEDEN UK USA 1. the estimated normalized coefficients show the size and sign of the relation among REER.55) -0.042141 (5.13) 0.674) 4.740930 (-7.55 and 4.318104 (8. by excluding the TOT variable from the unrestricted model.00 1.02619** (-0.561349 (3.00 1.104034 (-19.515238 (10.74) 0.568617 (3.115919 (2. Moreover.83) 0. To this end.25) 0.814363* (1.296982 (7.09) 2 (TOT) 0 0 0 0 0 0 0 0 0 0 Notes: (1) Values in parentheses are t-values.363578 (3.95) 1.27) -0.00 1.833648 (-2. (3) Values with (*) are significant at 10% level.25) 2 (TOT) 1.07) 0.757827 (7.747938 (6.542179 (13.00 1.00 1.00 1.00 1 (PRO) 4.12) -0.00 1.010976** (-0.99) 1.22972 (-1.240538 (4.882087* ( 1. Country-Specific Estimation Results UNRESTRICTED MODEL (eq. 1% increase in PRO leads to an increase in REER around 4.00 1. We do this by reporting the elements of matrix.037656 (-7.812291 (-4. Moreover.553164 (19. the biggest effect of relative productivity on REER takes place in Canada among 10 countries.25) 1.00 1. which is consistent with theory.00 1.97) -1.73) -0.01) 0.71) 0.Now.2) 0. PRO and TOT.5) 1. the remaining task is to find the size and sign of this long-run relation among three variables.00 1. we present Table 5.33) 1.35) 3.00 1.23) -2. By interpreting the significance.775681 (19.00 1.12) REER 1.83088** (-0. except the values with (*) and (**). (4) Values with (**) are insignificant. we can see whether it is important for the BS model.

the biggest estimated coefficient belongs to Netherlands among 10 countries. As seen in Table 5. in the restricted model models due to negative signs of coefficient in the coefficient is found insignificant. 15 15 . effect Denmark and Italy which is inconsistent with the prediction of the BS hypothesis. 16 Drine and Rault (2005) also rejected the BS hypothesis for USA due to the negative sign of the BS effect. USA is the other country in which the BS hypothesis fails in both the restricted and coefficients. it is fair to conclude that the BS hypothesis performs well for these countries. Thus.99 in Germany and 1. That is to say.56 in particular. Germany and Italy are three countries in which the BS hypothesis holds in both the restricted and unrestricted models due to positive and significant coefficients. 0. and total consumption may generate a less-biased coefficient.11 in Sweden and -2. In the unrestricted model. in the unrestricted model coefficients are 3.3).8 in UK and all estimated are significant.51 in Denmark. In coefficients are 0. 2007. USA Such a big BS effect seems to be over-estimated for Canada. In addition.16 Although the estimated BS effect of USA is negative.74 in Netherlands. 0.83) unrestricted model implies that 1% increase in the relative productivity leads to -1. A similar BS effect is estimated by Chowdhury (2007) for Australia. If we exclude the TOT variable.88 in Netherlands. Similarly. The negative and significant (-1. Put merges with the BS effect which leads to over-estimation of another way.58 in Netherlands. the estimated coefficients are 4.83% decrease in the REER of Japan. government consumption. In the restricted model the estimated these countries are as follows. one should use the TOT variable as an explanatory variable which may help her to estimate a less-biased BS effect. the estimated coefficients are 0. the estimated coefficients of Italy in the unrestricted model. In the restricted model. and 1. This might stem from the relative importance of TOT for Netherlands.36 in UK. which implies the TOT variable has no effect on REER of Italy. It is worth noting that exclusion of TOT led to a significant increase in the BS effect of Netherlands. for the BS model of Canada inclusion of some relevant explanatory variables such as openness.6 and he suspects that ³the elasticity coefficient is over-estimated due to the exclusion of relevant explanatory variables´ (Chowdhury. to explain the long-run REER movements in Netherlands. Moreover. 0.5 in Sweden and 1.31 (consistent with prediction) in Germany. which is against the prediction of the BS hypothesis.This implies that 1% increase in the ³terms of trade´ associates with 1% increase in the real effective exchange rate of Canada.77 in Sweden and 1 in UK. Hence. Furthermore. has been found insignificant for Italy. In both models. the BS effect unrestricted models due to negative signs of works in an opposite way that the relative productivity growth leads depreciation in the ³REER of USA´. He estimated the BS effect 5.69 in Germany. p. 0. Sweden and UK are other three countries in which the BS hypothesis holds in both the restricted and unrestricted models due to positive and significant coefficients. Japan is a country in which the BS hypothesis fails in both the restricted and unrestricted coefficients.03) in the unrestricted model which is inconsistent with the prediction of the BS hypothesis.56 in Italy. is found negative (-1. 1. Although the TOT has negative sign in is 1. Netherlands. perhaps the TOT effect in the restricted model. thus it has to be used with a caution.54 in Denmark. the estimated coefficients are significant at 5%. as in Japan. 0. 15 Denmark.

the inclusion of terms of trade). This result shows that TOT is a crucial variable in explaining the ³REER of Norway´.13) 8 out of 10 2 out of 10 1.29) and significant in the unrestricted model that the BS hypothesis holds. In the study.34.1.g. Generally speaking. The findings of this study are mostly in favor of the BS hypothesis. BS hypothesis for Norway in the restricted model. Since 1964. In particular. when we exclude the TOT turns out negative (-0. In this respect. Norway is the only country that the BS hypothesis holds in the unrestricted model and fails in the restricted one. The theoretical contributions (e. which explains the real exchange rate movements with productivity differences between countries. In Netherlands. the BS hypothesis is valid 8 out of 10 countries when the unrestricted model is used and valid 7 out of 10 when the restricted model is employed. And the TOT effect has been found as predicted (positive and significant) in 6 out of 10 countries.34 RESTRICTED MODEL (eq. cointegration analysis) motivate scholars to test the validity of the BS hypothesis. And the BS effect has been found positive and significant in 7 out of 10 countries in the restricted model. it is found out that the TOT variable is relatively more important for Netherlands and Norway. 16 . the cointegration analysis confirms the existence of long-run relation among REER. country-specific coefficient estimations indicate that the BS effect is positive and significant in 8 out of 10 countries. the exclusion of the TOT variable from the BS model led over-estimation of the BS effect and in Norway it led to the rejection of the BS hypothesis. In the restricted model.has a positive and significant TOT effect which is 0.6 and the average TOT effect of 10 OECD countries is 0. many published studies have tested the hypothesis by using different methods.02) and insignificant which implies the rejection of the variable. the average BS effect of 10 OECD countries is 1. and improvements in econometric techniques (e. the average BS effect of 10 OECD countries is 1. Summary of Country-Specific Estimation Results BS HOLDS BS FAILS AVERAGE BS EFFECT AVERAGE TOT EFFECT UNRESTRICTED MODEL (eq. the estimated is positive (0. which is consistent with theory and previous studies¶ results (see Alexius & Nilsson.1 3. Conclusion The BS hypothesis. Nonetheless. In the unrestricted model. According to Table 6.g. 2002). PRO and TOT variables in all countries. In the unrestricted model. In particular.6 0. Table 6 summarizes the country-specific estimation results regarding the validity of the BS hypothesis. the unrestricted BS model performed partially well for USA since the TOT effect has a correct (positive) sign. TOT is a critical variable in the BS hypothesis that the TOT coefficient is estimated positive (as predicted) and significant in 6 out of 10 countries. still keeps its importance in empirical economics. 2000 and Egert. Table 6.12) 7 out of 10 3 out of 10 1.81.

W.ubc. Are Malaysian Exports and Imports Cointegrated? A Comment.gov. Cassel. Antweiler. 72. T. Purchasing Power Parity. B. 413-415. Real Exchange Rates and Fundamentals: Evidence from 15 OECD Countries. Growth Led Exports: LDCs Experience. Sunway Academic Journal. (2002). 1918). Model Mis-specification and Johansen¶s Co-integration Analysis: an Application to the US Money Demand. 24. government consumption. Open Economies Review. Economics Working Paper Series. the inclusion of some relevant explanatory variables such as total consumption.T. especially for USA in where non-tradables (services) sector productivity growth is more evident. University of British Columbia. still plays an important role in explaining why the PPP fails and the real exchange rates differ across countries. M. W. 584±596. (2007). Economic Journal. Chowdhury. Export Led Growth vs. A. (2005). (1964). and M. 11. The Journal of Political Economy. 101-107.sauder. References Ahking. 383±397. K.tcmb.ca/PPP. Economidou (2009).Among 10 OECD countries. E. 17 . Nilsson (2000). accessed on May 3. Alexius. The Journal of Developing Areas. Abnormal Deviations in International Exchanges. Journal of Macroeconomics. 51±66. (1918). relaxing the assumption that we made about non-tradables sector would also affect the estimation results for these countries. WP 07-11. 2.tr/yeni/evds/yayin/reel_efktf/MethodologicalExplanation. 2009: http://fx. In this respect. Cheong. F. Japan and USA are two specific countries. 387-409. University of Wollongong. which cannot be rejected widely in empirical studies (as in this study). Volume 42. With the improvements in econometric methods and availability of new sectoral productivity data. in which the BS hypothesis fails in both restricted and unrestricted models probably due to country-specific reasons. and openness would help us to get better results for Japan and USA regarding the BS effect.pdf. Number 2.html. G. http://www. the BS hypothesis seems to keep its importance also in the future. Balassa-Samuelson Effect Approaching Fifty Years: Is it Retiring Early in Australia?. and C. Finally. IMF Staff Papers. Balassa. Choudhri. 3. Bahmani-Oskooee. Real Exchange Rates in Developing Countries: Are Balassa-Samuelson Effects Present. CBRT (2008) (Central Bank of the Republic of Turkey). 52. today the BS hypothesis. (2008). 179209. Khan (2005). and J. In addition. The Purchasing Power Parity Doctrine: a Reappraisal. (December.

J. 3. C. and I. Structural Change and Real Exchange Rates: Evidence from Historical Data. A. Khim. England.org. Johansen. International Financial Statistics (IFS) Database: http://www. R. NBER Working Paper Series. M. 3-4. (2007). 873±887. 15:4. Productivity. 4952. Real Exchange Rates over the Past Two Centuries: How Important is the Harrod-Balassa-Samuelson Effect?. Department of Economics. 18 . Jaunky. Juselius (1990). and C. Rault (2005). B. 313. Journal of International Money and Finance. International Financial Statistics Country Notes. (1988).De Broeck. and M. and T. Some Structural Hypotheses in a Multivariate Cointegration Analysis of the Purchasing Power Parity and the Uncovered Interest Parity for UK. and K. Basic Econometrics. and S. Economics Bulletin.Volume LXI. Economic Systems. (2002). Applied Financial Economics. Trecroci (2007). 1±16. K. S. Wolf (1994). IMF. (2003). Muscatelli. Estimating the Impact of the Balassa±Samuelson Effect on Inflation and the Real Exchange Rate During the Transition. Slok (2001). 26.imfstatistics. Harris. Froot. R. Johansen. Liew (2004). D. No. John Wiley & Sons Ltd. Dumitru. J. 4807. Statistical Analysis of Cointegration Vectors. and the Real Exchange Rate. The Economic Journal. Perspectives on PPP and Long-Run Real Exchange Rates. C.318. V. 15:8. 1742±1763. Gujarati. V. F. Terms of Trade. Drine. Applied Economics Letters. 33.. De Gregorio. P. 118 (October). The Balassa±Samuelson Effect in Romania ± The role of Regulated Prices.Sollis (2005). European Journal of Operational Research. and H. No. and R. 231±254. 1í9. IMF (2008). 12. Interpreting Real Exchange Rate Movements in Transition Countries. I. Working Paper No.. Working Paper No. S. I. V. Vol. Journal of Economic Dynamics and Control. and. University of Copenhagen. Can the Balassa-Samuelson Theory Explain Long-run Real Exchange Rate Movements in OECD countries?. IMF Working Paper.90-05. Applied Time Series Modeling and Forecasting. 519530. McGraw-Hill. Macroeconomic Shocks. Discussion Papers. NBER Working Paper Series. 26. Taylor (2008). 194. Which Lag Length Selection Criteria Should We Employ?. and K. WP/01/56.C. Spinelli. Rogoff (1994). The Real Exchange Rate and Productivity Differentials: a Panel Cointegration Approach. Egert. 1403-1423. Jianu (2009). Lothian.

and D. WP 2007±12. 783±820. Department of Economics. Review of Economics and Statistics. Cambridge.Obstfeld. M. Database: Tica. IMF Working Paper. Rogoff (1996). J. and K. Working Paper Series. Samuelson. 335±346. (2007). International Finance. MIT Press. 19 . 1950-2007. The World Bank. International Comparisons of Productivity and Unit Labor Cost Trends: 17 countries or areas. (2006). 06-7/686). Dominique (1997). OECD (2005).org/data/onlinedatabases/onlinedatabases.S. and I. Foundation of International Macroeconomics. 61. University of Zagreb. P. Bureau of Labor Statistics (2009). The Harrod-Balassa-Samuelson Effect: A Survey of Empirical Evidence. Division of International Labor Comparisons. Faculty of Economics and Business. (1964). Stock. and M. H. Theoretical Notes on Trade Problems.W. http://www. Croatia. 06-07 (No. A Primer on the IMF's Information Notice System. Department of Labor. Discussion Paper Series. Loughborough University. Zanello. Econometrica. Paper No. K.org/dataoecd/59/41/37985056. 46.pdf.worldbank. P. J. World Development Indicators (WDI) http://www. Watson (1993). Pilbeam. A Simple Estimator of Cointegrating Vectors in Higher-Order Integrated Systems. MA. Turner. . WP/97/71.html. Druzic (2006). Testing for Cointegration Using the Johansen Approach: Are We Using the Correct Critical Values?.oecd. Palgrave Macmillan. U. 135-139. PPP Methodological Manual. A. A.