Financial shocks transmission to unemployment in developing countries: A GVAR analysis

Transcription

1 Financial shocks transmission to unemployment in developing countries: A GVAR analysis Hany Abdel-Latif a * Swansea University, Department of Economics, UK 213 This revision: May 213 Abstract Two possible transmission channels of financial shocks are empirically examined in this paper, namely foreign direct investment and export volume. These channels are assumed to proxy for financial and real transmission channels, respectively. The paper imports into effect a relatively new econometric technique to labour and development literature, Global Vector Autoregressive model (GVAR). The proposed GVAR modelling approach considers transparently interdependencies between countries through: a) conditional individual country models; b) allowing for non-zero pairwise correlations between countries and equations to capture dependency in idiosyncratic shocks; and c) explicitly incorporates common global factors. Accordingly, using quarterly data for 197Q1-212Q4, the current paper builds on a GVAR model of 21 developing countries and 8 developed countries. The GVAR approach allows examining individual courtiers as well as grouped countries based on regions and on their development achievements. To this end, the variables of interest show heterogeneous responses to the shock. The main findings show that while financial channel represented by FDI has an immediate negative impact on unemployment in developing countries, the FDI flows pull through but perhaps seeking different destinations afterwards the shock. As well, there is empirical evidence of the shock lagged impact on unemployment through export volumes from developing countries. However, once the shock adverse effect has been realized, it persists over a longer period of time. This can be built upon to conclude some policy implications. * Corresponding Address: Hany Abdel-Latif, School of Business, Economics and Law, F4 Richard Price Building, Swansea University, Singleton Park, Swansea, SA2 8PP, UK. Address: @swansea.ac.uk Telephone number: +44 () The author is grateful to Vanessa Smith (University of Cambridge), Mathew Greenwood (University of Melbourne) and Timo Bettendorf (University of Kent) for their useful discussion and helpful comments. The correspondent author is a PhD student and gratefully acknowledges financial support from the University of Swansea 1

2 1. Introduction The recent financial crisis has revealed that current world residents are connected through more channels than ever was in the bygone century. More particularly, developing countries became more exposed to the economical turbulences in the developed economies. In this context, there are number of channels through which financial crises could be transferred into domestic developing economies. These channels could be subtitled under the following wide head titles. (1) Trade channels (trade volume and conditions); and 2) financial channels (declines in foreign investments, drops in remittances inflows, unfulfilled development assistance inflows and limited access to world credit markets), see (Perry and Lederman, 1998, Meyn and Kennan, 29, Robinson and Willenbockel, 29). Historical evidence in Reinhart and Rogoff, (29) show that sharp recession is associated to financial crises. They show that financial crisis-based recession is more likely to last approximately two years after. Other scholars such as Choudhry et. al., 21 record empirical evidences of substantial increases in unemployment levels in time of financial crises. Additionally, there is a general consensus among these studies regarding a slower improvement in unemployment rates after the shock in spite of any recovery the GDP may have witnessed (Thangavelu et al., 29). As well, the hysteresis effects of financial crises have been widely accepted in labour market theoretical models (Blanchard and Wolfers, 2, Nickel et al., 25, Bassanini and Duval, 26, Choudhry et. 21). These models assert that the higher than pre-crisis levels of unemployment tend to persist over time, up to some years after the shock. On the other hand, understanding how external shocks transmit into domestic economies is indispensably important to policy makers. Presumably, the lack of international dynamicity and inadequate interdependency is the key problem encounters most of macroeconomic models used by policy makers. Generally speaking, there are few known macroeconomic models that stand for the precedent problem. Therefore, this paper analyses international shocks transmission in a global context. More particularly, utilizing a Global Vector Autoregression (GVAR) model which is developed by (Pesaran et al., 24, Pesaran and Smith, 26), the current paper investigates transmission channels of financial shocks originated in developed countries and its propagation mechanism to developing countries, namely their labour markets. To proxy for trade channels and financial channels, the paper uses export volumes and FDI inflows, respectively. As potential shocks adjusting mechanisms, the labour market outcome of interest is unemployment rate. In this occasion, 2

3 the GVAR is not only perceived to be a rich and flexible modelling approach but also it keeps dimensionality curse of macroeconomic models manageable (Eickmeier and Ng, 211). First, the GVAR modelling approach directly incorporates many observable macroeconomic and financial market variables which allow policy makers to craft scenarios that influence some specific variables and their response to different types of shocks. Second, the interdependencies between countries are transparently considered in three ways; a) combining individual country models that include domestic and foreign variables, b) capturing dependency in idiosyncratic shocks by allowing for non-zero pair-wise correlations in residuals between countries and equations, and c) by explicitly incorporating common global shocks such as oil prices. Thus, based on a GVAR model, this paper empirically investigates the trade and financial transmission channels of financial crisis to real economy in developing countries. The export volumes and FDI inflows are used to proxy for these channels. In particular, the outcome variable is unemployment rate in developing countries. The paper contributes to the present literature in several dimensions. First, it imports a relatively new econometrics technique, namely GVAR model, to labour and development economics literature. On contrast to existed literature, the GVAR modelling approach allows to investigate the transmission mechanism of financial crises to developing country labour markets considering the global context. Second, the paper stands among the GVAR literature by plugging more developing countries to the model as well as considering the most influential countries in the world economy. Third, given that developing countries do not compose a homogenous group, the paper looks into the relative importance of the FDI and export volumes as two possible sock transmission channels by opting three distinctive model specifications; individual countries, regional and development success grouped countries. Finally, the paper builds upon the empirical findings in order to stitch some policy implications. In this context, the main findings of the current paper show that while financial channel represented by FDI has an immediate negative impact on unemployment in developing countries, the FDI flows pull through but perhaps at different destinations afterwards the shock. As well, there is empirical evidence of the shock lagged impact through export volumes from developing countries. However, once the shock adverse effect has been realized, it persists over a longer period of time. 3

4 In addition to the proceeding introduction in section (1), this paper is stitched as follows. Section (2) summarizes the literature review. The GVAR modelling approach and the dataset are discussed in sections (3) and (4), respectively. Section (5) explains the model specification. The dynamic analysis is summarized in section (6). Finally, a summary of the main findings and conclusions could be found in section (7). 2. Literature review 2.1 Financial shock international propagation Within the literature of financial economics, the transference of financial crises among countries and/or sectors is known by contagion effect. According to (Baur, 211), there are three brands of contagion as follows: (1) Contagion of the financial sector across countries (i.e. spillover); (2) contagion of the financial sector and the real economy within a country; and (3) contagion of the financial sector and the real economy across countries. Thus, this study empirically investigates how a financial crisis could pass through the real economy, more particularly to the labour markets, in developing countries which makes it ideally fits into the third category of contagion effect literature. On other hand, given the lack of connectedness between financial markets in both developing and developed economies, it is expected that financial shocks effect on poor countries to take place through shrinking international trade and drops on other forms of capital flows such as FDI, ODA and remittances. Apart from the long lasting debate of how FDI inflows underwrite development process, it is commonly consented it has a direct impact on injected capital, modern technology, marketing and management skills into domestic economies which may potentially spillover and create positive externalities to local firms, see (Kawai, 1994, Caves, 1999, De Mello, 1999, Djankov and Hoekman, 2, Xu, 2, Nair Reichert and Weinhold, 21, Buckley et al., 22). Moreover, it is widely accepted that, compared to the short-term capital flows, the FDI is normally less footloose. However, the experience from past financial crisis suggests that FDI inflows significantly affected by the financial crises (Thangavelu et al., 29). 2.2 GVAR modelling approach literature Since the seminal work of (Sims, 1986), the Vector Autoregressive models (VAR) are in a continuously progression to contribute in macroeconomic modelling literature. The prevailing property of VAR models is that they are atheoretical models that exploit information from 4

5 data without the need to impose any theoretical or statistically motivated restriction a priori. The VAR model imports its origin from the prevailing structural simultaneous equations models paradigm. Hence, in a situation of simultaneity, (Sims, 1986) argues that there is unwarranted distinction between endogenous and exogenous variables. Therefore, the VARs treat all variables as endogenous. The foundation of GVAR formwork imports its origin from (Pesaran et al., 2) where they propose a general structural VAR model with exogenous I(1) variables; which becomes known as VARX * model. Their argument plots upon a small open economy that is trivial to affect the world economy which allows incorporating foreign variables as weakly exogenous. Within GVAR approach, a country s domestic variables are linked to foreign variables depending on its bilateral trade or financial exposure to other countries. This innovative way is used to construct country-specific foreign aggregates which are allowed to directly influence domestic variables in the model. The GVAR approach implicates estimating country-specific VARs individually each of which takes in a set of co-integrating relations derived from a New-Keynesian small open economy model 1. Based on certain weights, these separately estimated models are combined into a global model. In this model, the world variables and country specific variables are jointly determined. In this occasion, (Rebucci and Spatafora, 26) state that GVAR may be interpreted as an empirical counterpart to a simplified, global, dynamic general equilibrium model as well as an approximation to a global common factor model. (Pesaran and Smith, 26) show that the VARX * models can be derived as solutions of dynamic stochastic general equilibrium (DSGE) models. An approximation of the GVAR model to a common factor model has been derived by (Dees et al., 27). Henceforth, the GVAR model embraces VARs of each country linked together via weighted averages of all other countries variables and/or global variables such as oil prices. In effect, applying weighted-averaging to foreign variables cures dimensionality curse associated to empirical research on international transmission of financial shocks (Eickmeier and Ng, 211) 2. 1 For each country, these restrictions are first tested using an unrestricted model; if not rejected; they are then imposed on the data. 2 Having potentially many more variables and parameters than observation requires heavily imposing restrictions on the parameter space. One possible way to reduce dimension is applying weighted averaging to the variables. 5

6 Studies use a GVAR to model international financial shock transmission include (Galesi and Sgherri, 29, Chen et al., 21, Beaton and Desroches, 211, Bussière et al., 211, Chudik and Fratzscher, 211, Eickmeier and Ng, 211, Xu, 212). (Rebucci and Spatafora, 26) construct a GVAR model to investigate the role of oil prices in a global context. Their model incorporates, beside United States and China, three large bulks of countries, namely advanced, developing and oil exporter countries. Their findings support a negative effect of oil price shocks on the external balance (current account) of non-oil exporter countries including the US. Based on a panel of 21emerging and advanced economies, (Bussière et al., 29) utilize GVAR to model global trade imbalances. Distinctly, they jointly model exports-imports response to three possible shocks; a shock to US output, a shock to the US real exchange rate and a shock to German output. (Galesi and Lombardi, 29) apply GVAR to address the impact of oil and food price shocks on inflation. (Chen et al., 21) apply GVAR to examine international transmission of bank and corporate distress. (Eickmeier and Ng, 211) use sign restrictions on short-run impulse responses to identify credit supply shocks. They criticised other GVAR studies (such as (Xu, 212, Beaton and Desroches, 211)) that use generalized impulse responses to shocks that have not been orthogonalised, which makes economic interpretation of the shocks difficult. Other studies which examine financial shocks tend to use simple recursive schemes (e.g., (Bagliano and Morana, 212, Rigobon, 23). Pesaran and Smith (26) discuss the use of Beveridge- Nelson decompositions to extract long-run equilibrium solutions. Cesa-Bianchi et al. (212) use time-varying weights to capture changing trade patterns china and Latin America. Finally, lacking the focus on shocks transmission to developing countries is one common characteristic among GVAR literature. Also, GVAR modelling approach is relatively new in studying developing country labour market shock responses. The current paper stands to these gaps in the literature. 3. The GVAR model: There has been an increase in popularity of VAR models in empirical macroeconomics since the influential work of (Sims, 198). By combining country-specific models, the GVAR evades dimensionality curse and allows for analyzing the interactions between countries/regions into a global framework. The GVAR allows unrestricted coefficients for the 6

7 domestic variables and carefully construct country-specific foreign variables which are treated weakly exogenous when estimating the individual models. To capture the long-run relationships as well as the short-run dynamics, the individual VAR models are estimated individually in the form of country-specific vector error-correcting models VECMs. Posteriorly, these VECMs are stacked up to simultaneously generate impulse response functions. Hence, the GVAR allows for the interdependencies between countries and regions explicitly (Galesi and Sgherri, 29). GVAR modelling involves greater use of prior information in choosing observed weights (Eickmeier and Ng, 211). Most of GVAR literature relies basically on bilateral international trade to compute these weights, e.g. (Pesaran et al., 24, Dees et al., 27). However, few studies opt different weight mechanisms; for instance, Galesi and Sgherri (29) uses financial weights, Vansteenkiste (27) uses geographical distances based weights, Vansteenkiste and Hiebert (27) adopt weights based on sectoral input-output tables across industries. Country-specific models: Given countries in the global economy, indexed by, where and country serves as a reference country 3, each country, on its own, is considered as a small open economy thereby cannot affect the whole system. However, the world economy affects individual countries through country specific foreign ties. For each country a model, where and are the lags orders of the domestic and foreign variables respectively, can be presented as follows: Where is the vector of domestic variables, is the vector of country-specific foreign variables, denotes the matrix of observed global factors (here, oil prices), and are the coefficients of the deterministic components, and is the vector of idiosyncratic, serially uncorrelated, country specific shocks, where (1) and have a zero mean with a covariance matrix ( ), for.further, there is,,, where is the lag operator and and are the lag order of the domestic and foreign variables for the ith country. 3 Given its sizable effect on the world economy, the US has been extensively referred to as the reference country in GVAR literature and this paper show no exception. 7

8 Country-specific VARX * models are endogenous variables conditioned on country-specific foreign variables which are constructed based on the trade weights,, that capture the importance of country for country s economy such that, where and,,. These weights are fixed and computed based on countries average bilateral trade (see appendix B). For notation ease, the corresponding VARX * (2, 2) in the error correcting representation is as follows [ ], (2) where, ( ), is a matrix of rank, is a matrix of rank (the number of cointegration relationships in the system). By partitioning as conformable to, the error correction terms defined above can be written as, which allows for the possibility of cointegration within the endogenous variables and between endogenous and foreign variables in a given country model, as well as endogenous variables across country models. For estimation, the country-specific foreign variables are treated as long-run forcing weakly exogenous with the respecting to the parameters of the conditional model. The GVAR indispensable initiation to circumvent the dimensionality predicament is via conducting the estimating on a country by country basis. However, the model is solved for the system as a whole considering all the variables are endogenous to the whole model. Recall that and consider the individual models written as,, where [ ], 8

9 The vector can be written as,, (4) Where a link matrix of dimension constructed based on country specific weights. Substituting results in the following,, (5) Then, the vector of endogenous variables of the global economy, stacking the country specific models as, can now be obtained by, (6) where ( ), ( ),. The model above can be solved recursively and used for generalized impulse response analysis and forecast. 4. Dataset Countries included in the dataset are chosen on the basis of several considerations. The first is to give the taste of developing countries by pondering more developing countries and, in the meantime, keeping the most influential world developed countries in the model. This has been missing in GVAR literature as the majority, if not all, restrict their focus to developed countries. The second consideration is to select a group of developing countries for which data of reasonable quality could be assembled. The third consideration is to include developing economies representing different geographical areas and wide range of development experiences. Model specification includes regional and income level classification as shown below. The dataset composes Quarterly data covers 29 countries for the period of (1979Q1 212Q4). Table (1) classifies the dataset countries according to both country income level and regional grouping. Data source is introduced in appendix (A). 9

10 Table 1: Dataset country classification according to income level and regions Income level * Region * Low income Lower middle income Upper middle income MENA NA ** Egypt Algeria Sub-Saharan Africa Tanzania Nigeria South Africa South Asia Bangladesh India NA *** East Asia and Pacific Myanmar Indonesia-Philippines Latin America and Caribbean Haiti **** Peru China, Malaysia, Thailand Argentina- Brazil- Chile-Mexico- Venezuela Others Turkey High income countries G7 Canada- France- Germany- Italy- Japan- UK- US Others Singapore * Both income and regional classification are based on the World Bank. The country income level is based on the 212 classification. ** Based on the World Bank country classification, the MENA region includes 13 developing economies none of which is a low income country. *** Out of 8 south Asian developing economies, only Maldives falls within the upper middle income category for which no reliable data could be assembled. **** For the Latin America and Caribbean region, all countries are categorized as middle income countries except Haiti which is considered as a low income country. 5. Model Specification The variables for the GVAR model are unemployment rate, real GDP, inflation rate, FDI flows, export volumes and equity price index. Additionally, foreign unemployment rate, foreign real GDP, foreign FDI flows, foreign export volume are included as weakly exogenous variables into the system constructed based on a constant weights based average bilateral trade (26-28). 4 As well, the model contains unweighted oil prices (poil) as a global variable which is weakly exogenous to all countries in the system except the US in which poil enters as endogenous variable. 5 Subsequently, the resulting GVAR model is no more than a tremendous VAR (P) model containing global endogenous variables. Table (2) defines the GVAR model variables corresponding to each group of countries. As shown in table (2), the FDI flows and export volumes do not enter the model as exogenous variables in the US model since it plays as the reference country in the model. Also, as stated before, the oil prices variable enters all individual models as exogenous except for the US 4 The foreign variables are constructed as the sum weighted contributions of the N-1 other countries. 5 GVAR literature usually justifies the US model selection to contain oil prices as endogenous variable by the high demand of the US to oil. This is actually understandable and may be justifiable for the same reason in this chapter whereby there is no considerable exporter of oil in the sample. However, this justification is very weak and critical in most of the other GVAR literature as they more likely to consider more sizable oil exporters such as Saudi Arabia. In other words, it is still not understandable why oil price enters to the Saudi Arabia as exogenous in the vast majority of the GVAR literature. 1

11 model in which it is endogenous. On other front, the paper considers three focal models; individual countries, region based, and income level based models. These are shown in table (1) earlier. Variables Table (2): model specifications US Other developed Developing countries Unemployment Rate Real GDP FDI Export volume Equity Prices Oil Prices Estimation of the country-specific models: To seize the possibly unobserved common factors, GVAR incorporates the cross section averages of the endogenous variables which are assumed to be weakly exogenous. Also, being individually integrated is a presumption should one aspire making use of long-run information as well as short-run dynamics. Such allegations, among others, have to be formally tested and for which this subsection intends to stand. Unit root test: To ensure series have univariate integration properties, Augmented Dickey-Fuller (ADF) proposed in (Dickey and Fuller, 1981) and weighted symmetric ADF (hereafter ADF-WS) introduced by Park and Fuller (1995) are considered. Different unit-root tests are set for the purpose of evasion the low power featuring such tests as some test superiorly perform when compared to others, see for example (Perron, 1989) (Leybourne et al., 25), (Pantula et al., 1994) (Elliott et al., 1996). These tests mull over the null hypothesis of unit root against the alternative of no unit root. Thus, ADF and ADF-WS tests are conducted with a time trend and with no trend for level as well as first and second differences. Table (3) presents ADF- WS statistics of all the country-specific variables whilst table (4) summarizes the test results for the country-specific foreign variables. 6 The results in tables (3) and (4), overall, support the treatment of the variables as being I(1). 6 Note that the Akaike Information Criterion is employed for lag selection of these tests. 11

12 Domestic Variables Table (3): Unit Root Test Statistics for Domestic Variables Country a Argentina Brazil China Turkey Algeria Egypt Tanzania S. Africa Peru Haiti Note: Based on univariate autoregressive specifications, the ADF-WS statistics for the level, first differences, and second differences of the variables are all computed on the same sample period, namely, 1979Q1-212Q4. The ADF-WS statistics for all level variables are based on regressions including a linear trend. The 95% critical value of the ADF-WS test for a regression with a linear trend is -3.24, and for a regression with an intercept only is a The unit root test statistics for all the countries are given in a Supplement that can be obtained from the authors on request. 12

13 Foreign Variables Table (4): Unit Root Test Statistics for Foreign Variables Country a Argentina Brazil China Turkey Algeria Egypt Tanzania S.Africa Peru Haiti Note: Based on univariate autoregressive specifications, the ADF-WS statistics for the level, first differences, and second differences of the variables are all computed on the same sample period, namely, 1979Q1-212Q4. The ADF-WS statistics for all level variables are based on regressions including a linear trend. The 95% critical value of the ADF-WS test for a regression with a linear trend is -3.24, and for a regression with an intercept only is a The unit root test statistics for all the countries are given in a Supplement that can be obtained from the authors on request. Cointegration relations: Given model specifications in table (2) and in light of the AIC lag selection criterion, the corresponding models are estimated and the rank of their cointegrating space determined as given by equation (2). These are computed based on Johansen s trace and maximal eigenvalue statistics as set out in (Pesaran et al., 2). Moreover, to avoid having a quadratic trend in level, the model utilizes case IV defined by (Pesaran et al., 2) wherein the intercept coefficient is not redistricted but trend coefficient is so. Table (5) presents the cointegration rank statistics for a number of developing countries. The order of the VARX * models as well as the number of cointegration relationships are presented in table (6). 13

14 Table (5): Cointegration Rank Statistics for a number of developing countries Country a H H 1 Argentina Brazil China Turkey Algeria Egypt Tanzania S.Africa Peru Haiti Maximum Eigenvalue Statistics r = r = r 1 r = r 2 r = r 3 r = r 4 r = Trace Statistics r = r > r 1 r > r 2 r > r 3 r > r 4 r > r 5 r > a Test results for the remaining countries are provided in a Supplement that can be obtained from the authors on request. Table (6): VARX * Order and Number of Cointegration Relationships in the Country-Specific Models VARX * Country a p i q i No. Of Cointegrating Relationships Argentina Brazil China Turkey Algeria Egypt Tanzania S.Africa Peru Haiti a Test results for the remaining countries are provided in a Supplement that can be obtained from the authors on request. Results in table (5) are based on the Maximum Eigenvalue and Trace statistics at the 5% significance level along with the cointegrating relationships for the individual VARX models. As shown in table (6), the number of cointegrating relationships varies across individual VARX models. However, most of the models have two or three cointegrating relationships (21 countries out of 29 countries) and Japan has the highest number of cointegrating relationships (i.e., 5 cointegrating relationships). 14

15 Testing for residual serial correlation: The F statistic for the residual serial correlation test is the F-version of the familiar Lagrange Multiplier (LM) statistic (see (Godfrey, 1978a, Godfrey, 1978b)), also known as modified LM statistic. The F-statistics for the serial correlation of the VECMX models along with the corresponding critical values at 5% significance level are reported in table (7). Table (7): F Statistics for Tests of Residual Serial Correlation for Country-Specific VARX * models Country a F Critical at 5% Domestic Variables Un rg cp fd Eq Ev Argentina F(4,16) * * * Brazil F(4,114) * China F(4,11) * 4.95 * * Turkey F(4,11) * 3.53 * Algeria F(4,11) * 4.33 * 2.51 * 7.1 * * Egypt F(4,19) * 8.77 * * * Tanzania F(4,112) * * S.Africa F(4,19) * Peru F(4,19) * * * Haiti F(4,12) * *.57 Note: * denotes statistical significance at the 5% level or less a Test results for the remaining countries are provided in a Supplement that can be obtained from the authors on request Testing for weak exogeneity: The weak exogeneity of the foreign variables with respect to the long run parameters of the conditional model is one crucial assumption that GVAR approach builds upon. In the context of cointegration, this assumption implies no long run feedback from endogenous to exogenous variables without the need of ruling out lagged short run feedback between them, see (Johansen, 1992) and (Granger and Lin, 1995). Thus, the country-specific endogenous variables are said to be long run forcing for the corresponding foreign variables. As described in (Johansen, 1992) and (Harbo et al., 1998), this involves a test of the joint significance of the estimated error correction terms in auxiliary equations for the countryspecific foreign variables,. In particular, for each element of the following regression is carried out 15

16 ,, are the estimated error correction terms corresponding to the cointegrating relations found for the country model, and and are the lag orders of the lagged changes for the domestic and foreign variables, respectively. The test for weak exogeneity is an F test of the joint null hypothesis that for in the above regression. The results from the exogeneity test are reported in table (8). These results show that the exogenous variables pass the test. Results of weak exogeneity test show that the null hypothesis could not be rejected for the majority of the variables being considered. Table (8): F Statistics for Testing the Weak Exogeneity of the Country-specific Foreign Variablesselected countries Country F test Fcrit_.5 un * rg * cp * fd * eq * ev * poil * ARGENTINA F(3,97) BRAZIL F(2,112) CHINA F(1,13) TURKEY F(3,94) ALGERIA F(1,13) EGYPT F(2,12) SAFRICA F(2,12) TANZANIA F(1,17) PERU F(2,12) HAITI F(2,95) a Test results for the remaining countries are provided in a Supplement that can be obtained from the authors on request. Contemporaneous effects of foreign variables on their domestic counterparts The contemporaneous effects can be interpreted as impact elasticities between domestic and foreign variables and they are given by the estimated coefficients on the contemporaneous foreign specific variables. In particular, they reveal the international ties between the domestic and foreign variables. The stronger the co-movements between domestic and foreign variables, the higher the estimated coefficients are expected to be. The contemporaneous effects coefficients are estimated together with t-ratios computed on standard, as well as White and Newey-West adjusted variance matrix. For briefness, the contemporaneous effects estimated coefficients along with only their t-ratio based on the Newey-West adjusted standard errors are shown in table (9). Focusing on developing countries, the estimated coefficients show mixed (sign +/-) findings regarding the contemporaneous effect of foreign unemployment rate on a country specific unemployment rate. For example, results show that while a one percentage increase in foreign unemployment rate leads to a decreased unemployment rate in Argentina by.29 percentage points, it increases the unemployment rate in Mexico by.17 percentage points. Concisely, countries that their unemployment rates show positive co-movements along with 1% increase in the 16

17 foreign unemployment rate are Chile (.67%), India (.3%), Indonesia (.16%), Mexico (.17%), Philippines (1.1%), Thailand (.35%), Turkey (.64%), Venezuela (.6%), Nigeria (.19%), Peru (.16%) and Haiti (.3%). Table (9): Contemporaneous Effects of Foreign Variables on their Domestic Counterparts Un Rg cp fd Ev ARGENTINA (-.896) (1.26) (-1.566) (.862) (2.178) BRAZIL (-1.863) (1.594) (.987) (-1.493) (2.13) CHINA (-1.984) (3.328) (-1.44) (-.567) (-1.363) TURKEY (2.138) (1.882) (-.141) (-.844) (-1.252) ALGERIA (-.191) (.193) (-.173) (1.459) (-.47) EGYPT (-1.57) (2.635) (.114) (2.41) (-1.86) SAFRICA (-2.4) (1.17) (1.655) (-1.456) (1.4) TANZANIA (1.22) (.418) (2.94) (1.162) PERU (1.242) (3.234) (.92) (2.142) (.343) HAITI (.386) (.922) (-.637) (1.49) (-.748) Note: Standard errors are in parentheses. a Test results for the remaining countries are provided in a Supplement that can be obtained from the authors on request. Average pairwise cross-section correlations The idiosyncratic shocks of the individual country models are assumed to be crosssectionally weak correlated. This implies that the covariance between the foreign variables and the error terms tend to be zero as the number of cross section units goes to infinity. As a result, the weak exogeneity of the foreign variables is ensured 7. Given that the weakly exogenous foreign variables upon which individual VARX models are conditioned to proxy for the common unobserved global factors, one should expect that the degree of correlation of the remaining shocks across countries/ regions to be inconsequential. 8 Therefore, a simple diagnostic of the extent to which the country specific variables have been effective in reducing the cross-section correlation of the variables in the GVAR model is provided by the average pairwise cross-section correlations. 9 7 The weak exogeneity test indirectly support the view that the idiosyncratic shocks could only be weakly correlated. However, by using the average pair-wise cross section correlation, GVAR modelling approach provides direct evidence on the extent to which this is likely to be true. 8 These residual interdependences could reflect, for example, policy and trade spillover effects. 9 They are computed for the level and first differences of the endogenous variables of the model, as well as those of the associated residuals over the selected estimation period. For every country for each given variable, 17

18 Table (1) reports the average pairwise cross section correlations for selected countries for unemployment rate. 1 Although the results vary widely across variables and countries, the average cross section correlations are generally high for the level of the endogenous variables and remarkably fall once the first differences are considered. For example, in the case of unemployment rate, sharing common trend, for Argentina show cross section correlations of 8.6% and 3.4% for level and first different, respectively. In all cases, there is evidence of cross country correlations for the domestic variables in the GVAR model. However, the extent of this correlation depends on the variable and country. On other hand, the cross section correlations from the VECMX models are shown to be very small and, contrast to the endogenous variables only, they do not depend on the choice of the variable or country. For example, once adding the foreign variable to the model for Argentina, the average cross section correlation dropped drastically from 8.9% (in level) and 3.4% (first different) to hit.9%. The latter findings show that the model has successfully captured the unobserved global common effects. In other words, results show that once country specific models are conditionally constructed on foreign variables, the remaining degree of correlations across the shocks form different countries are modest. Table (1): Average Pairwise Cross-Section Correlations for unemployment rate Country a First VECMX Levels Differences Residuals ARGENTINA BRAZIL CHINA TURKEY ALGERIA EGYPT SAFRICA TANZANIA PERU HAITI a Test results for the remaining countries are provided in a Supplement that can be obtained from the authors on request. Structural Stability tests: Likewise to other macroeconometric modelling approaches, GVAR is not immune to the structural breaks problem. Unfortunately, this shed a number of insurmountable difficulties the pairwise correlation of that country with each of the remaining countries is computed, and averaged across countries. 1 Obviously, these results do not introduce a formal statistical test of how important the inclusion of country specific foreign variables is. However, the average pairwise correlations may be taken as an indication of the usefulness of the foreign variables in the model. 18

19 especially on modelling future breaks for forecasting and policy analysis purposes. 11 To some extent, relying on conditional country-specific model is expected, one way or another, to lessen this problem. This expectation conforms to the concept of co-breaking introduced by (Hendry, 1996) and examined further in (Hendry and Mizon, 1998). Accordingly, (Dees et al., 27) argue that the country VARX models contained in GVAR might be more robust compared to country VAR models. The current paper utilizes the Wald for of Quandt s (196) likelihood ratio statistic (QLR) to test for structural breaks. The critical values of the test are computed under the null of parameter stability, using the bootstrap samples obtained from the solution of the GVAR (p) model Dynamic Analysis: Given the space constraint and the large size of the model, this section presents a selection of results of interest of the GVAR model dynamic analysis. In particular, the model has been used to simulate the effects that shocks to selected variables in the system may have on the other variables over time. More specifically, this section investigates the effects of the US equity price shocks on unemployment rate, FDI inflows and export volumes in developing countries. To this end, the current paper makes use of the Generalized Impulse Function (GIRF) to cope with the difficulty in identifying the structural shocks in the GVAR context. What makes this advantage is that the GIRF is invariant to the ordering of the equations in the system which implies that the shocks are not orthogonalised. It shows what is most likely to happen after a shock to the l th equation in country i. Therefore, although no direct policy implication could be driven based on these responses, the GIRF is still informative in terms of demonstrating the most likely effect of certain shocks. Accordingly, this section shall be divided into a number of subsections. Firstly, it will define the GIRF in the GVAR context. Secondly, the shock effects of US equity prices will be discussed regarding the variables of interest. These will be based on the median estimates resulting for bootstrap the model (5 replications) considering 9% confidence intervals of the error bands. 6.1 Generalized impulse response functions in GVAR model: The time profile of the effects of variable-specific shocks or theoretically identified shocks on the future states of a dynamical system is known as the impulse response. Imposing restrictions may result in more economically interpretable shocks. For example, to estimate 11 There is little known about how best to model these breaks. Although Bayesian procedure may be use to identify in-sample breaks, allowing for future breaks is still a serious challenge though. 12 Results of the QLR structural break statistics are not shown here, but are available upon request. 19

20 the impulse responses in a structural VAR model, (Sims, 198, Sims, 1986) suggest imposing k(k-1)/2 restrictions in the form of a Cholesky decomposition on the variance covariance matrix. As well, while (Blanchard and Quah, 1989) impose long-run restrictions, (Uhlig, 25) directly restrict the impulse response. The GVAR model considers the GIRF introduced in Koop, Pesaran and Potter (1996) and adapted to VAR models in Pesaran and Shin (1998). In the same vein, (Pesaran et al., 24) show that the full identification of such a system by imposing k(k-1)/2 is difficult. Alternatively, they show that applying GIRF (Pesaran and Shin, 1998) is more preferable in order to examine the effects of different shocks. 6.2 A negative shock to US equity price This subsection summarizes the effects of one standard deviation negative shock to the US equity prices index. The responses of three variables of interest are investigated. These variables are unemployment rate as main labour market outcome- and FDI inflows and export volumes as proxies of financial and real transmission channels of developed countries financial shocks. Following a 1SD shock to the US equity prices, the US real GDP immediately falls by 8% and the shock negative effect reached its peak after one year by 26 percentage points. Whereas it has no immediate effect on the US unemployment rate, the shock negative effect starts at the first quarter and gets its maximum of 17 percentage points in the fifth quarter. Also, the shock instantaneously hits the US FDI outflows as it suffers immediate drop by 4 percentage points. Even though, it seems that the effect dies out in the sixth quarter and the US FDI outflows show recovery in quarter seven. On other hand, the equity prices shock confirms a negative and persistent effect on the US export volumes albeit it has not shown a pressing effect. The shock harmed the US export volume by a 2 percentage points drop in quarters five and six whereas it has not vanished away from 1 percentage points for a long time after the shock till 1 years after the incipient shock. Notably, the shock has similar effects on all the G7 countries and Singapore. For brevity, the results discussed here are narrowed down to developing countries only. Noticeably, GVAR treats all variables in the whole system as endogenous and, hence, restricting the discussion to developing countries should not cause any concerns. To this aim, the analysis shall cover the three model specifications that are defined in section (5) earlier. Precisely, individual developing countries (21 countries), five groups-based region (MENA, 2

21 MENA Sub Africa South Asia East Asia Latin America Low income Lower Middle Upper Middle Argentina Brazil Chile China India Indonesia Malaysia Mexico Philippines Thailand Turkey Venezuela Algeria Egypt South Africa Tanzania Nigeria Myanmar Bangladesh Peru Haiti median Sub-Saharan Africa- South Asia, East Asia and Latin America) and three income level groups (Low, lower middle and upper middle income countries). Appendix (C) presents the GIRF graphs of for all developing countries in the system. At this instant, the analysis is more engrossed on a certain comparison between different countries and groups of countries A negative shock to US equity price on unemployment rate: Grippingly, given the sign of the GIRF, whilst countries show different instant responses to the shock, most of them have experienced increasing in their unemployment rate starting from one quarter after the shock. Those countries with 2 quarters lagged response are Brazil and Italy. All countries in the system are negatively affected by the shock except for Algeria and Nigeria. Figure (1) shows the one year effect of the shock on unemployment rate in all countries along with grouped countries in the model. Figures (2) and (3) show the GRIFs of countries grouped by region and income level, respectively. The individual GIRFs graphs are presented in appendix (C). Figure (1) US Equity Price negative shock on unemployment rate Effect after one year Note: Tanzania unemployment rate is missing Region/Country Figure (1) shows the one-year effect of the shock is most sever on Venezuela in which the unemployment rate has increased by 26 percentage points as a result of the shock. It persists at 35 percentage point after 1 years forward. Argentina comes next by 21 percentage points but it persists around 13% for the same period. The one year effect in Chile is at 13 percentage points. However, it seems that the shock effect dies out after five years. Mexico one year effect is at 7% and persists around 4%. Brazil effects 4% and persists at 5% (no 21

22 immediate effect, the effect started from the second quarter). The remaining Latin American countries do not show to be affected by the shock. Surprisingly enough, china and India unemployment rates seem to be immune from the US equity shock. For Turkey the same shock results in 23% change to the unemployment rate. Regarding the regional responses, on average percentage points, they are as follows: Latin America (8%), Sub-Saharan Africa (3%), East Asia (2%), South Asia (1%) and MENA ( 2%). On other hand, these responses are 5% for upper middle income countries, 2% for both lower and low middle income countries. Figure (2) Impact responses of a negative unit shock to equity prices on regional unemployment rates MENA South Asia Latin America Sub-Saharan East Asia Turkey Figure (3) Responses of a negative unit shock to equity prices on unemployment rates Upper middle Lower middle Low income

23 MENA Sub Africa South Asia East Asia Latin America Low Income Lower Middle Upper Middle Argentina Brazil Chile China India Indonesia Malaysia Mexico Philippines Thailand Turkey Venezuela Algeria Egypt South Africa Tanzania Nigeria Myanmar Bangladesh Peru Haiti median A negative shock to US equity price on FDI inflows to developing countries: Following a one SD negative shock to the US equity prices, most developing countries show instantaneous drops in their FDI inflows. However, some countries show positive FDI inflows even after one year of the shock. Surprisingly enough, Brazil is one of these countries. The effect on China was temperate and vanished out after four quarters time. Figure (4) plots the one year effect of the shock. Figures (5) and (6) show the GIRFs for developing countries classified based on regions and income level, respectively. The GIRFs plots for all developing countries are shown in appendix (c). As shown by figure (4), the shock year effect hits Tanzania worst where the FDI inflows dropped sharply by 22 percentage point. Venezuela, Egypt, South Africa and Philippines come next in terms of the shock adverse effect on FDI inflows that have declined by 16, 12, 1 and 1 percentage points, respectively. On other hand, the graph shows that the most beneficial countries in terms increased FDI inflows were Algeria (5%), Nigeria (4%) and Peru (1%). South Asia, MENA and Sub-Saharan witnessed 6 percentage point fall in FDI inflows. East Asia and Latin America as least harmed by FDI inflows drops 2% and 3%, respectively. Figure (4).2 US Equity Price negative shock on FDI inflows Region/Country 23