On the Asymmetry of Global Spillovers: Emerging Markets Versus Advanced Economies

This paper examines growth spillovers between emerging markets and advanced economies. The empirical results, based on a two-bloc setup and cover 1991 to 2015, are twofold. First, the paper shows that the size of the spillovers running from emerging markets to advanced economies is about a fifth of those running from advanced economies to emerging markets. Second, the results point to spillovers from emerging markets to advanced economies having increased over the second half of the sample period. The paper presents suggestive evidence that the (evolving) structure of interdependencies plays an important role in explaining the existence of "asymmetrical spillovers" between these similar size blocs.


I. INTRODUCTION
Emerging markets (EMs) represent a relatively large and rising share of world GDP. Taken together, EMs now constitute the largest bloc in the world economy and contribute the most to global growth (Figures 1 and 2). From a trade standpoint, advanced economies' (AEs) imports are sourced in large part from EMs (about 50 percent) and EMs have also become important destinations for AE exports (Figure 3 and 4). 2 EMs are as a bloc the world's largest consumer of commodities, including metals and energy ( Figures 5 and 6). The EM bloc is thus expected to be a large source of (cyclical) spillovers to the rest the world. To date, however, little is known about these spillovers. Indeed, the spillover literature has focused mostly on spillovers from AEs. 3 While AEs were arguably the most important sources of spillovers for most of the second half of the 20 th century, it is high time to consider the EM bloc as not just the destination of these spillover effects but also the origin-considering the growing importance of EMs. Some recent studies have attempted to study spillovers from EMs but focusing on individual EMs, mainly China (Hsieh and Ossa, 2011). 4 The present paper attempts to fill the gap. 2 It should be noted however that trade data used in the analysis are on a gross basis and thus double or multiplecount vertical trade or flows in intermediates. Trade flows in gross basis may overestimate the extent of trade links between AEs and EMs. That said, to the extent that supply chains are at least in part regional in nature, the issue of double counting when considering trade links between AEs and EMs is less prevalent. Trade in value added allows incorporating the specificity of the new business model on which global manufacturing is based. Due to limited country coverage of value-added trade data, we are however unable to use such data in our analysis.
The main contribution of our paper is to document empirically that global spillovers are asymmetrical. Our empirical results based on a two-bloc set-up and covering the period 1991 to 2015 are twofold. First, we show that the size of the spillovers running from EMs to AEs is about a fifth of those running from AEs to EMs. Second, results point to spillovers from EMs to AEs having increased over the second half of the sample period. We present suggestive evidence that the (evolving) structure of interdependencies plays an important role in explaining the existence of asymmetrical spillovers between these similar sized blocs.
A casual look at data on trade and capital flows between AEs and EMs suggests that global spillovers may indeed be asymmetrical and changing. Figure 7 shows that the EM bloc depends relatively more on external demand, suggesting that spillovers from AEs to EMs are expected to be large. AE dependence on external demand is increasing, suggesting that spillovers from EMs maybe increasing (see Table 1). Figure 8 also shows that the EM bloc depends more on capital inflow, suggesting that spillovers from AEs maybe large. AE dependence on capital flows is increasing, suggesting spillovers from EMs are expected to increase (see Table 2). In this paper we use state of the art Vector Autogressive (VAR) techniques to systematically document the asymmetrical nature of global spillovers.
For the purpose of allowing the reader to grasp the importance of our empirical findings on the asymmetrical nature of global spillovers, it is useful to make a detour by explaining how existing theoretical frameworks used to study the international transmission of business cycles treat EMs.
There are essentially two schools of "modeling". The first school relies on the small economy assumption (see Schmitt-Grohe & Uribe, 2003;Aguiar and Gopinath, 2007;Garcia-Cicco et al. 2010). Clearly, that small open economy assumption has been justified for the most part of the second half of the 20th century but as EMs have become a large if not the largest economic bloc-with a large share of manufacturing activities in AEs having moved to EMs-it seems important to explore whether this assumption is still valid. 5 Indeed, EM growth may potentially exercise large spillovers through direct trade linkages, commodity and asset prices. The second school of modeling employs a two-country framework. Specifically, the Backus et al. (1992) framework (henceforth BKK model) is a two-bloc set-up assuming symmetry and correlated shocks. These assumptions are reasonable when considering spillovers between AEs. However, these assumptions seem less appropriate when considering spillovers between AEs and EMs that are not symmetrical in terms of the structure of their economies. The BKK also does not account for the evolving structure of interdependencies.
One challenge in documenting empirically spillovers between EMs and AEs is the choice of the identification strategy. In our benchmark VAR model, we assume that AE shocks spill over to EMs contemporaneously within a quarter but that EM shocks do not spill over to AEs within that time frame. That assumption relies on the nature of linkage between EMs and AEs. Specifically, considering that EMs rely relatively more on (fast moving) capital flows originating from AEs, it is likely that the speed of spillovers from AEs to EMs trumps the speed of spillovers running from EMs to AEs. The latter direction of the causality is likely to be channeled mostly through trade and thus may take longer to materialize. Relying on that timing restriction however imposes that a large part of co-movement between growth in EMs and AEs is attributed to AE shocks. In turn, this could lead to overestimating the magnitude of the spillovers originating from AEs onto EMs. We thus also explore different avenues, including arguably exogenous variables, to test whether our results are sensitive to the choice of decomposition. Specifically, we use fiscal shocks and damages from natural disasters to isolate growth shocks that are exogenous to other blocs. Fiscal news is constructed using the so-called narrative approach to isolate exogenous components of policy changes from endogenous policy responses. The fiscal shocks are defined as exogenous that are not driven by current and future developments on the real side of the economy. These shocks are exogenous with respect to the state of the real economy. We also exploit the arguably exogenous shock stemming from damage caused by large natural disasters. We rely on the fact that the timing of natural disasters is exogenous. Results using these identification strategies confirm the existence of asymmetrical spillovers from AEs to EMs.
The rest of the paper is organized as follows. Section II lays out the data and empirical strategy.
Section III presents our main empirical results. Section IV discusses results using an alternative identification strategy. Section V concludes.

A. Data
The sample includes 19 emerging market economies and 21 advanced economies. 6 Our classification is based on the core lists of countries that the International Monetary Fund and other organizations such as the World Bank define as AEs and EMs. The main results presented in this paper are robust to using different classifications. 7 The sample period runs from 1991Q1 to 2015Q4. The data are at the quarterly frequency. The economic growth series for the two economic blocs, namely, the EM and AE blocs, are based on purchasing power parity (PPP) weighted average of local currency real GDP growth from the IMF's World Economic Outlook live database. For the purpose of checking whether our main results are robust to different country groupings, we use smaller country groupings including the Group of Seven, representing the world's largest industrialized economies (G-7); a group composed of Brazil, the Russian Federation, India, China (BRIC); and China alone.
The data also include control variables that capture the main channels of transmission (trade, finance, and commodities). To control for trade, we use imports over total trade constructed using bilateral data from the IMF's Direction of Trade Statistics. To control for the financial channel, we construct a spread measure based on the MSCI Emerging Index over the MSCI World Index. To control for the commodity channel, we use Bloomberg's commodity price index.
Last, we also use data on fiscal expenditure and tax events in the United States and natural disasters as source of arguably exogenous variation to identify growth shocks in both AEs and EMs. U.S. spending news data are the narrative military expenditure shock obtained from Ramey (2011). U.S. tax news is the narrative tax shock obtained from Romer and Romer (2012). AE and EM natural disaster shocks are the damage stemming from large natural disasters (larger than US$1 billion) in the relevant country group over the group GDP. The data are originally from the online version of the International Disaster Database (EM-DAT). 8

B. Empirical Approach
We exploit VAR techniques to capture the interrelationships between AE and EM growth and quantify the dynamic spillovers between these blocs. Our VAR model specification is as follows: We first use a parsimonious model that simply employs a bivariate model including GDP growth of EMs and AEs. 9 Second, we augment the model with various other control (endogenous) variables such as the main financial and trade variables and commodity prices. Last we augment our bivariate model with exogenous variables to explore a different identification strategy.
We estimate the model from a Bayesian perspective. Unlike the traditional approach, we optimally choose the informativeness of our prior beliefs. The priors are treated as hyperparameters to maximize the marginal likelihood of the data. For illustrative purposes, take the case where one estimates an autoregressive process of order one, AR(1), with an unknown persistence, ρ. Instead of setting the prior associated with the persistence of the AR(1) process to follow a normal distribution as follows, ρ ~N(0.3, 1), we set the distribution to be as follows, N(0.3, s) and treat s, the "tightness of our prior" as a parameter we aim to maximize the marginal data density (equivalently the out-of-sample prediction power). The estimated value of s determines what approach we choose to pursue. If the estimated s is large, we have a loose prior and go back to using ordinary least square results. If instead s is small, we have in effect a tight prior and use that prior in our estimation. This approach is theoretically grounded and reduces the subjective choices in the setting of the prior. This approach is superior as it performs well both in out-of-sample prediction and accuracy in the estimation of impulse response functions 9 A dummy is also included to control for the period covering the global recession that is from 2007Q3 to 2009Q3.
(see Giannone et al. 2015). More details about the estimation strategy are presented in Appendix II.

III. EMPIRICAL RESULTS
We now turn to our empirical results. In the first sub-section, we present the results using our parsimonious specification. In the second sub-section, we explore how our results change across various samples. In the last sub-section, we present results using an augmented specification with the various channels of transmission.

A. Basic Results
Results obtained from our parsimonious specification taking the form of a bi-variate VAR show that the spillovers between AEs and EMs are asymmetrical. 10 The impulse response functions are reported in Figure 9. The left-hand-side panel shows the impulse responses stemming from an AE growth shock while the right-hand-side shows the impulse responses from an EM growth shock. The impulses are based on one standard deviation of GDP growth in AEs and EMs respectively. It should be noted that the standard deviation of growth of the original impulse is much higher for EMs than it is for AEs, reflecting the more volatile growth process in the former. That said, when comparing the relative importance of the spillover originating from AEs onto EMs to the opposite direction of the causality, the impulse responses clearly show that the former direction of causality yields much larger spillovers.
To further illustrate the asymmetrical spillovers between AEs and Ems, we construct an "elasticity" of the spillover as the ratio between cumulative impulse responses over a year 10 The VAR specification uses one lag. Using several lags yields qualitatively and quantitatively comparable results.
horizon. The elasticity of the spillovers running from EMs to AEs is less than a fifth of the elasticity of the spillovers in the other direction of the causality (see Table 3). These elasticities are statistically significant. These results confirm the intuitive view that AEs spillovers to EMs are much more potent. Indeed, the facts presented earlier show the relatively high reliance of EMs on external demand and capital flows from AEs.
Results also show that the spillovers from EMs to AEs have been growing. To show this, we simply split the sample into sub-periods running from 1991 to 2002 and 2003 to 2015. The impulse responses are shown in Figure 10. We normalize the shock in the block of origin to be 1 on impact for ease of comparison. Figure 10 clearly shows that the spillovers from EMs to AEs jump in the second period compared to the first. There is much less difference between the two sub-periods when considering the impulse responses capturing the spillovers running from AEs to EMs.
To illustrate further these results, we again construct the elasticity of the spillovers from EMs to AEs. The elasticity associated with that direction of the causality jumps from 0.06 to 0.37 between the two sub-periods (see Table 4). Instead, the elasticity associated with spillovers running from AEs to EMs is relatively high in both sub-samples and only increasing moderately.
The ratio of elasticity between the two directions of causality is about a third for the second subperiod while it was less than a fifth using the overall sample.
These results confirm the intuitive view that the linkage between EMs and AEs has deepened in that a growing share of exports from AEs are destined to EMs and that commodity prices are increasingly driven by growth in EMs. While the spillovers from EMs to AEs have been increasing, they remain much smaller than spillovers originating from AEs.

B. Results Using Different Country Groupings
Results using different country groupings confirm the existence of global asymmetrical spillovers. To explore that, we unpack the various country groupings for both the destination and origin of spillovers (see Table 5). While the results using G7 instead of EMs are comparable, the elasticity drops (increases) when we instead consider the United States, Japan and Germany as source (destination) of the spillover. When using BRIC instead of the EM grouping, the results are comparable for both directions of the causality, suggesting that these four countries drive the spillovers. However, when using China instead of the EM grouping, the elasticity of the spillover running from AE to China becomes much smaller and not significant. China's spillovers onto Germany and Japan appear however high, in line with China's important trade links with these two countries. These results confirm that global spillovers are asymmetrical but that the geography of trade can help explain some of the heterogeneity in the spillover effects between sub-groups. We explore the relative importance of these channels in more detail in the following sub-section.

C. Channels of Transmission
Results controlling for the various channels of transmission confirm the existence of asymmetrical spillovers between EMs and AEs. To explore the importance of controlling for the various channels, we augment our benchmark specification with variables capturing trade, financial and commodity channels. Table 6 shows that our main results are virtually unchanged when incorporating these channels.
In order to explore the relative importance of these channels in explaining the transmission of these spillovers, we conduct a decomposition exercise. Specifically, we "distribute" the spillovers to AE and EM growth between the three channels we have identified (see Table 6).
Results suggest that the structure of interdependencies matters differently for the two directions of causality of the spillovers. For AEs, the asset and commodity price channels are relatively more important in the transmission of spillovers originating from EMs. For EMs, the trade and commodity price channels matter more than the asset price channel in the transmission of spillovers originating from AEs. These results are consistent with the relatively higher reliance of EMs on external demand from AEs and the reliance of EM growth on commodities.

IV. ALTERNATIVE IDENTIFICATION SCHEMES
As mentioned earlier, the ordering of the decomposition is chosen such as AEs are ordered first, AEs are thus deemed "more exogenous" than EMs. In other words, we impose that EM growth shocks do not contemporaneously affect AE growth. One could argue that while there is a strong case for such choice decomposition, our results may be overly reliant on that identification strategy. In this sub-section, we present impulse responses using arguably exogenous U.S. fiscal news (both spending and tax) and damage from natural disasters for both EMs and AEs as a source of exogenous variation for growth shocks. To generate these impulse responses, we simply augment our benchmark bi-variate VAR with either the U.S. fiscal news or the damage from EM/AE natural disasters. Fiscal news and natural disasters are ordered first considering their arguably exogenous nature.
Results using spending and tax news in the United States confirm that the spillovers running AEs to EMs are (very) large. Figure 11 shows on the left-hand-side panel the impulse response from U.S. spending news based on the narrative military expenditure shock from Ramey (2011) on AE and EM growth. The impulse response confirms that the spillover from a U.S. spending shock is very large. The elasticity of spillovers after a year running from the United States (spending shock) to EMs (growth) is greater than one, which is larger than the elasticity from our benchmark specification. The right-hand-side panel in Figure 11 shows the impulse response from a tax shock using the narrative tax shock from Romer and Romer (2010). While Romer and Romer (2010) find that tax news shocks have a strong effect on U.S. growth, we find that the effect on AE growth overall is negative but not statistically significant. That is perhaps due to the shorter sample period used in this paper but also perhaps because when considering AEs as a whole, a tax hike in the United States may also lead to an increase in capital flows to AEs, therefore counterbalancing the negative growth effect from higher tax in the United States. That said, the impulse response from U.S. tax news shocks shows a negative and statistically significant response, suggesting that the spillovers from a U.S. tax news shock to EMs is large.
The elasticity of the spillovers from the United States to EMs is also greater than one.
Thus far, we have validated only one of our main results, that is, the spillovers from AE/US to EMs are large, using a different identification strategy. To also explore whether the other direction of the causality running from EMs to AEs is much smaller, we use arguably exogenous variation from natural disaster shocks measured as the damage from large natural disasters (greater than US$1 billion) in the relevant group over the total group GDP. Natural disasters have significant consequences. EM-DAT report that the direct economic damage from natural disasters between 1991-2015 is estimated at around $2.5 trillion and led to 1.75 million deaths. In theory, the impact of natural disasters on GDP is unclear (Strömberg, 2007).
On the one hand, the loss of productive physical and human capital may reduce GDP. On the other hand, the disaster may provide a positive contribution to measured GDP, as reconstruction efforts and humanitarian aid. The net effect of natural disasters is thus an empirical matter.
Impulse responses presented in Figure 12 confirm that a natural disaster shock in AEs has a negative and significant effect on GDP growth. The results also show that AE natural disaster shocks spill over onto EM growth. The spillover effect from AEs to EMs is greater than one. The impulse response on the right-hand-side shows the effect of an EM natural disaster on EM growth is negative and statistically significant. The spillover from an EM natural disaster on AE growth is not different from zero, suggesting that the spillovers from EMs to AEs are small.
All in all, the use of an alternative identification strategy confirms the asymmetrical nature of growth spillovers between AEs and EMs.

V. CONCLUSIONS
The paper documented the existence of asymmetrical spillovers between AEs and EMs. In particular, results showed that the size of the spillovers running from EMs to AEs is about a fifth of those running from AEs to EMs. Results also pointed to spillovers from EMs to AEs having increased over the second half of the sample. We also presented suggestive evidence that the (evolving) structure of interdependencies plays an important role in explaining the existence of asymmetrical spillovers between these similar sized blocs. Our results suggest that while quantitatively the small open economy assumption associated with EMs might still seem appropriate, more research needs to be done to model how the evolving structure of interdependencies between EMs and AEs matters for the global economy.
These results also have important implications for stabilization policy with respect to spillover effects. The asymmetric nature of spillovers between EMs and AEs implies the former are in  1991 1993 1995 1997 1999 2001 2003 2005 2007 2009 2011 2013 2015 GDP share           Notes: The cumulative responses of AE growth and EM growth to an AE shock and an EM shock are based on a bi-variate VAR with AE and EM growth. The spillover is computed as the ratio between the cumulative response in the destination and origin of the shock. "*" indicates 90 percent significance. Notes: The cumulative responses of AE growth and EM growth to an AE shock and an EM shock are based on a bi-variate VAR with AE and EM growth. The spillover is computed as the ratio between the one-year cumulative response in destination and origin of the shock. "*" indicates 90 percent significance. Notes: The cumulative responses of AE growth and EM growth to an AE shock and an EM shock are based on bi-variate VAR with AE and EM growth. For advanced economies, the growth series are either one of the following: AE, G7, Germany, Japan and US. For emerging markets, the growth series are either one of the following: EM, BRIC and China. The spillover is computed as the ratio between the one-year cumulative response of destination and origin of the shock. "*" indicates 90 percent significance. Table 6. Channels of transmission Notes: Panel A shows the spillovers of bi-variate VAR (AE and EM growth) and 5-variable VAR (AE and EM growth, export-total-trade ratio, growth of Bloomberg commodity index and the difference between total return on MSCI Emerging market index and world index). "*" indicates 90 percent significance. Panel B shows the one-year variance decomposition. Panel C contributes the variance of spillover to trade, commodity and finance channels. Between bivariate and 5-variable VAR, the ratio of growth in destination country attributed to origin country shocks indicates how much the three channel explains the spillover in total. Next, the contribution of each channel is determined by the share of contribution of variance to the variance of growth of the destination country.