How Did GVC-Trade Respond to Previous Health Shocks? Implications for COVID-19

Using diﬀerence-in-diﬀerence analysis in a gravity model, we examine the response of GVC-trade to two previous health shocks, SARS and MERS. Our baseline estimates suggest a decline in GVC-trade, both gross and value-added, from SARS, emanating from supply and demand shocks, though a similar eﬀect is not observed for MERS. There is some evidence for “reshoring” and “near-shoring” in the stylized facts on SARS while empirical analysis also suggests geographical diversiﬁcation of value chains due to MERS and their non-resilience to SARS in particular. The ﬁndings are observed at both the intensive (value) and extensive (number of products) margins and for both intermediate and ﬁnal goods. The SARS eﬀects are driven by non-OECD countries that were also more integrated and downstream in GVCs, and by products that were less capital- and technology-intensive. We expect similar disruptions to GVC-trade from COVID-19.


Introduction
COVID-19 is proving to be an unprecedented health and economic crisis with long-term implications for countries across the world. It has already affected over hundred million lives since its outbreak in Wuhan, China in December 2019. A burgeoning literature has developed around analysing the epidemiological and macroeconomic impact of this pandemic (Baldwin and di Mauro, 2020;Baqaee and Farhi, 2020a,b;Djankov and Panizza, 2020;Eichenbaum et al. 2020;Fornaro and Wolf, 2020;Guerrieri et al. 2020;McKibbin and Fernando 2020). The crisis also has significant implications for trade and investment due to disruption of global value chains (GVCs). These disruptions emanate from the demand shock that lockdowns and stalled economic activity have caused as well as the supply shock resulting from temporary or permanent breaks in supply networks. Additionally, there is the GVC contagion effect (Baldwin and Freeman, 2020;Friedt and Zhang, 2020) -the pandemic has affected many locations simultaneously and the high level of interconnectedness of the global economy has amplified the impact, especially on the global hubs (World Bank, 2020).
While these disruptions have reignited the discussion on geographical/supplier diversification, nearshoring and repatriation of value-chains, it may be both too soon and premature to examine the extent, nature and impact of these disruptions using data currently available. We thus aim to inform the discussion by studying GVC responses to two past health epidemics, SARS and MERS, both to understand how value-chains may have responded to those crises and to draw implications for the COVID-19 outbreak. We focus on SARS and MERS for reasons that are common to the current pandemic -both outbreaks originated at an epicentre but spread around quickly; the diseases are characterized by flu-like symptoms; and manufacturing value-chains were likely disrupted by both episodes. 1 Infectious disease outbreaks have a profound impact on GVCs, simultaneously affecting multiple countries and industries, with the fear of contagion resulting in unanticipated changes in demand and supply of products (Sheffi, 2015). This fear can lead to under-reporting of an outbreak, especially if the country fears an ex-post application of trade sanctions against it by non-outbreak countries (Brahmbhatt and Dutta, 2008). It is believed that epidemic outbreaks are a unique type of supply-chain risk characterized by long-term disruption in demand, supply and logistics as well as unpredictable ripple effects. The location of supply bases in severely affected regions creates disruptions in supply networks; suppliers may close their plants or may be unable to deliver their products (Ivanov, 2020;Miroudot, 2020). For example, a supply-side contagion in East Asia's (China, Japan, Korea and Taiwan) manufacturing sectors may hurt manufacturing sectors of other countries as well due to supply linkages, especially in automobile, textiles and ICT goods sectors (Baldwin and Tomiura, 2020). Similarly, the decrease in domestic output in Thailand due to COVID-19 is attributed to increasing trade costs and under-utilization of capital, especially in the ICT goods industry that has the highest level of fragmentation of production in that country (Maliszewska et al. 2020). Moreover, the scope and timing of disruptions play a vital role in determining the impact of an epidemic outbreak on supply chains; the asynchronous opening and closing of facilities creates uncertainty at the firm-level, necessitating a guided framework for better decision-making (Ivanov, 2020).
Building resilience during a pandemic is thus the topmost priority for firms integrated in supply chains. Brandon Jones et al. (2014) and Miroudot (2020) distinguish between building robustness and resilience in supply chains -the ability to recover in the post-crisis period is resilience, while the ability to continue firm operations during a crisis is robustness. Extant literature proposes two opposing solutions to build resilience. One, insurance against a disruption by diversifying the supplier base, albeit at an additional cost, to reduce excess dependence on one country and compensate loss from a few supplier breakdowns (Henriet et al. 2012;Baldwin and Tomiura, 2020); and two, isolation from any disruption through reshoring manufacturing firms back home (Henriet et al. 2012;di Mauro, 2020).
Exclusive reliance on suppliers from one or a few countries can be detrimental by exposing importers to localized risks from health crises or natural disasters. Hence diversifying to alternative suppliers or locations of production during a crisis is more of a robustness strategy compared to reshoring manufacturing back home to a localized setting (Miroudot, 2020). However, long-term firm-to-firm relationship with a single supplier can assist in an easy bounce-back in the post-crisis period (Antràs, 2020), besides avoiding sunk costs from diversification at the eleventh-hour. Hence, there is an apparent downside to diversification vis-à-vis recovery, as supplier diversification is associated with slower recovery from interruptions (Jain et al. 2016). Strange (2020) recommends diversification over reshoring citing increased firm costs, reduced competitiveness and foreign sale of goods due to reshoring of firms closer home. The quantitative analyses in Bonadio et al. (2020) and Eppinger et al. (2020) also find reshoring to be sub-optimal from an economic welfare perspective. The negative sentiment around reshoring is also corroborated by firms: 32% of executives interviewed in an UNCTAD survey associated reshoring of manufacturing functions with a significant decline in global FDI (UNCTAD, 2015). Similarly, Hassan et al. (2020) show that discussions about diversifying supply chains in firm-level conference calls on corporate sector resilience during epidemics peaked during Q1 of 2003, clashing with the SARS outbreak.
Existing literature has studied the macroeconomic consequences of natural disasters, including health crises (Toya and Skidmore, 2007;Noy, 2009;Raddatz, 2009). Previous research suggests that economic development and institutional quality may provide implicit insurance against natural disasters (Kahn, 2005). Recent work provides both historical evidence (Ceylan et al. 2020) and empirical analysis (Fernandes and Tang, 2020;Friedt and Zhang, 2020), including on the the role of GVCs in the propagation of the COVID-19-induced shock (Bonadio et al. 2020;Eppinger et al. 2020;Sforza and Steininger, 2020).
We aim to contribute to this literature by studying GVC responses to SARS and MERS as observed in actual trade data. In doing so, we also explore the following hypotheses: one, the disease outbreaks were associated with a rise in domestic production at the expense of total imports of GVC-based products ("reshoring"); two, there was a tendency to reduce reliance on disease epicentres during the epidemics towards alternative suppliers ("geographical diversification of value-chains"); three, the disease outbreaks were associated with a decline in the concentration of the partner distribution ("GVC-widening"); four, there was a tendency to import more from suppliers in the geographical neighbourhood at the expense of the disease epicentres ("near-shoring"); and five, disruptions to GVC-trade coincided with the time period of the virus outbreaks but dissipated over time ("GVC-resilience"). 2 Our empirical strategy involves using difference-in-difference analysis to examine the effect of supply and demand shocks induced by SARS and MERS on bilateral trade in GVC-based products and value-added trade in a gravity setting. We measure the shocks using binary dummy variables that take the value one for the SARS-worst-affected countries (China, Hong Kong, Canada, Singapore and Vietnam) in 2003, and for the MERS-worst-affected countries (Saudi Arabia, South Korea and the UAE) 3 over 2013-17, 2015 and 2014, respectively. 4 Our 2 We take these terms from the international business literature, where 'reshoring' and 'near-shoring' are defined more on the basis of the location of production of multinational enterprises (MNEs) and whether they create foreign or domestic affiliates (or affiliates in neighbouring countries for near-shoring). In that sense, we use these terms differently; for instance, we use reshoring to mean 'renationalization' (Bonadio et al. 2020) or 'repatriation' (Eppinger et al. 2020) of value-chains.
3 These countries reported the largest number of cases and amongst the highest case fatality rates according to data from the WHO (https://www.who.int/csr/sars/country/table2004_04_21/en/; https://www.who.int/emergencies/mers-cov/en/). For instance, the number of SARS cases in China (the disease epicentre) and Hong Kong during January-June 2003 were 5327 and 1755, respectively, while the case fatality rate was 17% in Hong Kong and Canada (251 cases), 14% in Singapore (238 cases) and 7% in China. Vietnam reported 63 SARS cases and a case fatality rate of 8%. Similarly, Saudi Arabia, where MERS originated, had 158, 662, 454, 249 and 233 cases during each year of 2013-2017, followed by 185 cases in South Korea in 2015 and 86 cases in the UAE in 2014. While Saudi Arabia again witnessed a spike in MERS cases in 2019, there was a distinct break in trend in 2018, which we exploit in our identification strategy. Significantly, this break is also consistent with the stylized facts observed in Section 3 (see Figure  3 for details), where the decline in imports of GVC-based intermediates from Saudi Arabia in some sectors seems to have been arrested in 2018. 4 The time periods span severe incidence of the virus outbreaks in the worst-affected countries.
identification strategy thus exploits differences in the time period of severe incidence of each disease and in the samples covering exporting (supply-shock) and importing (demand-shock) countries that were more adversely affected than others.
We use bilateral trade data from BACI (Gaulier and Zignago, 2010), disaggregated at the HS 6-digit level over the 2000-2018 period, for over 200 countries (see Annex Table 1). The HS 6-digit products are classified as intermediate and final products in GVCs in the apparel, automobiles, electronics, footwear, pharmaceuticals and textiles sectors, following Sturgeon and Memedovic (2010) and the World Bank WITS classification 5 . In complementary analysis, we also use data from the UNCTAD-EORA GVC database (Casella et al. 2019) to examine how value-added trade was affected during these outbreaks.
Our baseline estimates suggest that the supply-shock from SARS reduced bilateral exports by 1.5% and 2.4% for GVC-based intermediate and final products, respectively; the respective decline from the SARS-induced demand shock was 1.2% and 2.5%. These results are consistent with the findings on the impact of SARS on China's exports and imports from firm-level analysis in Fernandes and Tang (2020). However, similar adverse effects are not observed at the intensive margin in the case of the MERS outbreak. The results are found to be robust to using alternative estimation strategies and to matrix completion analysis (Athey et al. 2017;Xu, 2017). Qualitatively similar findings are observed for value-added trade, underlining SARS' adverse impact on both gross and value-added trade.
We also find some evidence for reshoring and near-shoring in the stylized facts -the SARS epidemic, for instance, may have been associated with an increase in US imports from Mexico and EU15 imports from Turkey at the expense of Chinese exports in some sectors. Differencein-difference estimates also provide evidence for geographical diversification of value-chains -import shares of GVC-based products from China declined during SARS, while MERS was accompanied by a rise in the number of trading partners for GVC-based intermediate products and by a decline in partner concentration for GVC-based final goods from the outbreak-induced supply-shock. The value-chains also seem to not have been resilient to the SARS outbreak -the adverse effects on both gross and value-added trade were accentuated significantly over time, underlining the medium to long-term impact of that epidemic. These results are consistent with SARS's medium term impact observed on Chinese firm-level trade in Fernandes and Tang (2020); they also emphasize the need for countries to adopt strategies to minimize any longer-term disruption to GVCs from the current pandemic.
We also exploit the heterogeneity of our dataset along different dimensions to explore the likely drivers of GVC-disruptions due to SARS using both aggregate and disaggregated 5 The classification is available at https://wits.worldbank.org/referencedata.html.

5
This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=3798581 P r e p r i n t n o t p e e r r e v i e w e d data. Consistent with conceptual aspects of GVCs and their participation (Antràs, 2020), the SARS effects are found to be driven by non-OECD countries that were also integrated and downstream in GVCs; and by products that were less capital-and technology-intensive. We expect similar disruptions to GVC-trade from COVID-19.
Firms participating in GVCs exchange highly customized inputs on a repeated basis (Antràs, 2020); our results can thus be explained using the theoretical framework in Acemoglu and Tahbaz-Salehi (2020). The virus outbreaks result in negative shocks to the economy (observed as lower productivity or higher fixed costs of operation for some firms, sectors, or in aggregate) that alter the distribution of surplus throughout the production network, causing customized firms to fail either due to the direct negative shock to their production technology or indirectly from a break in their supplier network; reduced demand from their customers; or other losses along their network. This firm failure also leads to a decline in trade, which is more likely to be pronounced for firms and clients located in countries more severly affected by the disease outbreaks, a fact that we exploit in our empirical strategy that focuses on both the supply and demand shocks emanating from these health crises.
Our paper adds to the existing literature on the impact of health epidemics; it also provides stylized facts on and an empirical analysis of hypotheses underlying GVC-responses to external shocks. The paper is also related to different strands of the empirical literature examining the determinants and effects of (i) the 2008/09 global financial crisis (Baldwin, 2009;Bems et al. 2010;Levchenko et al. 2010;Ahn et al. 2011;Crowley and Luo, 2011;Chor and Manova, 2012) and (ii) natural disasters, especially the 2011 earthquake in Japan (Todo et al. 2015; Barrot and Sauvagnat, 2016;Carvalho et al. 2016;Zhu et al. 2016;Boehm et al. 2019;Freund et al. 2020).
The rest of the paper is organized as follows. Section 2 provides a brief background on the two virus outbreaks and studies examining their impact. Section 3 provides stylized facts on changes in GVC-trade patterns in the aftermath of the two epidemics. Section 4 discusses the empirical methodology used to examine the impact of SARS and MERS on GVC-trade. Section 5 discusses results from estimation while Section 6 provides additional analysis on SARS. Section 7 concludes with some implications for the COVID-19 pandemic. P r e p r i n t n o t p e e r r e v i e w e d was declared a global threat with 8,437 cumulative cases, of which, 7,452 reported recoveries (WHO, 2003), putting the mortality rate at 9.6%. Majority of the cases were concentrated in China (63.1%) and Hong Kong (20.8%) and these also accounted for 79.5% of the total SARS-reported deaths (WHO, 2003). As the virus spread through contact with the infected individual, controlling measures consisted of an early warning system, isolation of suspected cases, and contact tracing. SARS had costs beyond immediate health concerns; it created widespread panic, halted tourist activity in the region as well as greatly impacted trade and the overall far-eastern economy with losses worth US$ 30 billion by May 2003 (Demmler and Ligon, 2003). The disruption to international travel also impacted business meetings, leading to cancellation of factory orders and adding to the medium-term impact of the disease (Fernandes and Tang, 2020).
Several studies have examined the economic cost of SARS (Hai et al. 2004;Hanna and Huang, 2004;Lee and McKibbin, 2004;Smith, et al. 2019). The overall impact was felt across sectors, as diverse as seafood to microchips (ADB, 2003;NIC, 2003;IMF, 2004). SARS deterred global FDI in industrial production in China (Fan, 2003;Hanna and Huang, 2004) and in Hong Kong and Japan (Keogh-Brown and Smith, 2008). The threat to manufacturing sectors in China was to the extent that new orders were placed on hold and investors halted expansion plans for the year. Lee and McKibbin (2004) show that Hong Kong and China experienced the largest shocks to their GDPs from the SARS outbreak compared to Taiwan and Singapore, primarily due to their greater reliance on trade. In fact, Taiwan may have faced a wave of delayed shocks to its trade and investment due to linkages with mainland China (Chou et al. 2004

Stylized facts
In this section, in the spirit of Yeats (2001), we use disaggregated HS-6 digit-level data to look at the pattern of import shares of GVC-based intermediates from the SARS-and MERS-worst-affected countries, before, during and after the incidence of these outbreaks to explore the hypotheses outlined in the introduction. We also see if these episodes were associated with a fall in the number of intermediate products exported by the worst-affected countries or with a fall in the number of their export destinations. Our analysis covers GVCbased intermediates in the apparel, automobiles, electronics, footwear and pharmaceuticals sectors; note that a few HS6-digit products are classified as intermediates common to the automobiles and electronics sectors.
We begin by exploring the reshoring hypothesis by looking at the trend of mean total imports and domestic production of GVC-based intermediate and final products over 2000-2017, with the period covering the two virus outbreaks. 7 Figure 1 shows that mean imports of GVC-based intermediate and final products may not have declined during the SARS outbreak suggesting an absence of reshoring in the wake of that epidemic. In contrast, these imports seem to have witnessed a clear decline during the MERS outbreak in automotives and electronics (intermediates) and in the electronics and textiles sectors (final goods). This decline in imports seemed to have been accompanied by a rise in domestic output in the auto and electronics sector for intermediates and in electronics for final products, which is suggestive of reshoring in these sectors during the MERS outbreak.
<Insert Figure 1 here> 7 Disaggregated data on domestic output for GVC-based products included in our analysis are only available from UNIDO's INDSTAT database according to the ISIC Rev.3 and Rev.4 classification. Since the four-digit ISIC classification is more aggregate than the HS6 classification, the HS6 products in the data were aggregated to the four-digit ISIC Rev.3 level using concordance tables in United Nations (2002) for the purpose of this analysis. A decline in the share of intermediate imports by value across sectors (barring footwear and pharma) was also observed for South Korea and the UAE during the MERS epidemic (see Figure 3 top panels); for Saudi Arabia, a consistent decline was observed in auto and electronics over 2014-17. These countries also seem to have witnessed a decline in the number of their trading partners (middle panels) and in the number of GVC-based intermediates (bottom panels) exported across sectors in the wake of this outbreak.
Thus, there is suggestive evidence for GVC-disruptions in the data along both extensive and intensive margins and it may well have been the case that importing countries were bringing the value-chains closer home. To explore this "near-shoring" hypothesis, Figure 4 plots the ratio of US (EU15) imports of GVC-based intermediate products from Mexico (Turkey) to those from China over time in the top (bottom) panel to examine if these countries may have imported more GVC-based intermediates from Mexico and Turkey, respectively, at the expense of China in the wake of the SARS epidemic. To explore if the value-chains were indeed brought closer home and not simply moved to other low-cost suppliers, the figure also traces the evolution of the ratio of US and EU15 imports from two geographically-distant but competitive comparator countries relative to China -Thailand and Vietnam.

9
This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=3798581 Figure 4 suggests that the US may have switched imports of auto, electronics and pharmaceutical intermediates to Mexico 8 ; and the EU15 may have preferred Turkey for intermediate imports of apparel and auto & electronics; all at the expense of Chinese exports. This can be seen from the spikes in the respective sectoral ratios for the US and EU15 in 2004 and 2005, one and two years after the SARS outbreak. The US spikes in these sectors are much bigger in the case of Mexico compared to Thailand and Vietnam, suggesting that the associated value-chains may have been brought closer home following SARS, though the US may have also increased its reliance on Thailand for intermediate imports of apparel products in 2004. Meanwhile, the EU15 also seems to have enhanced its preference for Thailand in particular, for sourcing intermediates across all six sectors in 2004-2005 and to some extent, for Vietnam in 2005, for importing apparel intermediates, thereby providing strong preliminary evidence for the geographical diversification of all associated value-chains.
<Insert Figure 4 here> While these stylized facts are suggestive of a reconfiguration of GVCs in response to these disease outbreaks, they do not provide conclusive evidence of the "impact" of these health crises on GVC-trade. The identification of these effects requires more rigorous causal inference, which is the subject of the following section.

Empirical strategy
Our empirical strategy employs difference-in-difference (DiD) analysis in a gravity model setting. Following Anderson and van Wincoop (2004), the value of exports from country i to country j of product p at time t can be written as follows: where X ijt denotes the value of exports from country i to j at time t, E j is the expenditure in the destination country j, Y i denotes the total sales of exporter i towards all destinations, Y is the total world output, φ ij are the bilateral trade costs and σ is the elasticity of substitution.
P it and Π jt are the outward and inward Multilateral Resistance Terms (MRTs) as defined in the literature.
Trade costs in φ ijt can arise from different sources such as import tariffs (τ ijt ); geographical distance between trading partners [ln(DIST ij )]; cultural distance proxied by dummy variables identifying whether the trading partners share a common border (CN T G ij ), had a colonial relationship (CLN Y ij ) and share a common language (LAN G ij ); and membership of preferential trade agreements (P T A ijt ). Recent advancements in the estimation of structural gravity advocate the use of three-way fixed effects to mitigate endogeneity-induced biases in estimation (for instance see Baier and Bergstrand, 2007;Piermartini and Yotov, 2016). The dyadic trade cost variables (lnDIST ij , CN T G ij , CLN Y ij and LAN G ij ) are thus subsumed in bilateral pair-wise fixed effects.

Intensive margin analysis
The baseline DiD equations take the following forms: where X I/F ijt is the value of country i's exports of GVC-based intermediate/final (I/F ) products in destination country j at time t; P T A ijt denotes membership of preferential trade agreements; µ it , γ jt and δ ij are time-varying exporter and importer and dyadic fixed effects; and ijt is the error term. The shocks are measured on the supply-and demand-side by binary dummy variables that take the value one for the SARS-worst-affected countries and for the MERS-worst-affected countries amongst the sample of exporting and importing countries, respectively. Thus, in equation (2), SARS it and SARS jt take the value one when exporting and importing countries include {China, Hong Kong, Canada, Singapore and Vietnam} and the year is 2003; in equation (3), M ERS it and M ERS jt take the value one for the following exporting and importing countries: Saudi Arabia over 2013-2017, UAE in 2014 Korea in the year 2015. Under the parallel trends assumption, the estimates of χ 1 , ψ 1 and χ 2 , ψ 2 identify the average treatment effects of the supply-and demand-shocks emanating from the virus outbreaks on the treated.
The equations are estimated separately over the time periods 2001-2006 and 2011-2018, respectively, to examine the effects of the SARS and MERS outbreaks. Given the collinearity of the binary dummy variables with the time-varying exporter and importer fixed effects, the coefficients on these variables are retrieved in a two-stage estimation following Head and Mayer (2014) and Anderson and Yotov (2016). In the first stage, ln(X ijt ) is regressed on P T A ijt , and all three fixed effects. Estimates of multilateral resistance from this firststage regression are then used as outcome variables in auxiliary regressions to retrieve the coefficients of SARS it , M ERS it and SARS jt , M ERS jt , respectively. A priori, we expect the estimates of χ 1 , ψ 1 and χ 2 , ψ 2 to be negative. 9

Extensive margin analysis
Variants of equations (2) and (3) were also used to examine empirically if the epidemics were associated with a decline in the number of HS-6 products exported (P rod HS6 ijt ). Negative values of the estimated coefficients would provide evidence for the adverse effects of SARS and MERS at the extensive margin in each case.

Geographical diversification of value-chains
To examine if the supply-shocks emanating from the disease outbreaks were associated with a widening of value-chains, we depart from the gravity framework and estimate the baseline and augmented 10 versions of the following equations at the disaggregated HS6-digit level for GVC-based products for SARS and MERS separately: where HHI ipt is the Hirschmann-Herfindahl index of partner concentration 11 for country i at the HS6-digit level (p) at time t; µ pt and γ ip are the HS6-year and exporter-HS6 fixed 9 In alternative regressions, we estimated the coefficients of SARS it , M ERS it (and SARS jt , M ERS jt ), in a single step using the fixed effects selectively i.e. exporter (importer), importer-time (exporter-time) and bilateral fixed effects along with additional exporter-time (importer-time) varying controls comprising population and GDP. The adverse (non-adverse) effects of SARS(MERS)-induced supply and demand-shocks were confirmed in these results, available upon request. In further sensitivity analysis, we also used matrix completion analysis (Athey et al. 2017;Xu, 2017) that nests synthetic control and unconfoundedness in a generalized approach and was implemented in R using the Gsynth package; these estimates were also found to be qualitatively similar and are available upon request.
10 Augmented versions of equations (4)- (7) include Z zit which is a vector of exporter-time varying control variables, comprising the log of population (lnP op it ) and the log of nominal GDP (lnGDP it ). In specifications using disaggregated data, the control vector also included the log of tariffs [ln(1 + T ar ijpt )].
11 The value of the HHI lies between 0 (fully diversified) to 1 (unitary partner).

12
This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=3798581 P r e p r i n t n o t p e e r r e v i e w e d effects; and ipt is the error term. Estimated ϕ 1 , ϕ 2 < 0 would provide evidence for widening of value-chains from the supply-shocks emanating from the two outbreaks.
Finally, to examine if the supply-shocks emanating from the disease outbreaks were associated with a change in the number of destination markets for the exported GVC-products, we estimate the baseline and augmented versions of the following equations for SARS and MERS separately: where P AR it is the number of destination markets for the exported GVC-products for country i at time t; µ t and γ i are the year and exporter fixed effects; and it is the error term. Estimated α 1 , α 2 < 0 would provide evidence for a decline in the number of destination markets for the exported GVC-products from the outbreaks-induced supply-shocks.
All equations in this section are estimated using OLS. Summary statistics are reported in Annex Table 2 Finally, note that an essential pre-condition for DiD analysis is the existence of parallel trends between treated and control groups in the pre-treatment period (see Meyer, 1995;Angrist and Pischke, 2009). To examine this pre-condition, we implemented the 'common pre-dynamics test' proposed by Mora andReggio (2012, 2015). 12 Results from this analysis, available upon request, suggested that the assumption of parallel trends between treated and control groups in the pre-treatment period may have been met in all cases.

Intensive margin
The results from estimating equations (2) and (3) at the intensive margin are reported in Table 1, columns (1) and (2)

13
This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=3798581 P r e p r i n t n o t p e e r r e v i e w e d top and bottom panel for SARS and MERS, respectively. All coefficients are retrieved using the two-stage estimation procedure outlined in Section 4.1, with standard errors clustered by dyad-year in stage one and by exporter-year, importer-year in the auxilary regressions.
In these results, the supply-shock from SARS is found to reduce bilateral exports by 1.5% and 2.4% for GVC-based intermediate and final products, respectively; the respective decline from the SARS-induced demand shock is found to be 1.2% and 2.5%. The combined adverse effects of SARS-induced supply and demand shocks at the intensive margin thus range from 2.7% for intermediate goods to 4.8% for final goods. The inability of firms producing GVCbased final products to substitute customized inputs used in production with generic inputs (Acemoglu and Tahbaz-Salehi, 2020) likely explains the more adverse trade effects observed for GVC-based final products in these results. The findings are also consistent with those in Fernandes and Tang (2020) wherein Chinese firms located in SARS-affected regions observed 11 and 6 percentage-point lower export and import growth relative to unaffected firms and the pre-SARS period. <Insert Table 1 here>

Extensive margin
The intensive margin effects of SARS are partly corroborated at the extensive margin on the demand side in the results reported in columns (3) and (4)  This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=3798581 P r e p r i n t n o t p e e r r e v i e w e d ( 3) and (4)

Widening of value-chains
There is some evidence for geographical diversification in the form of widening of value-chains in response to the supply-shock emanating from the MERS outbreak for final products; the coefficient estimates reported in columns (7) and (8) of Table 2 in the bottom panel suggest that the Hirschmann-Herfindahl index of partner concentration may have declined by 1.0% to 1.2%: While a similar diversification in distribution is not observed for GVCbased intermediate products -the coefficient estimates in columns (3) and (4) of Table 2 in the bottom panel lack statistical significance, the number of trading partners may have increased in response to the MERS-induced supply-shock (coefficient estimates reported in columns 1 and 2 in the bottom panel are positive).
<Insert Table 2 here> While the corresponding HHI results for SARS are also statistically indifferent from zero for GVC-based final products, the coefficient estimates in columns (3) and (4) of Table 2 in the top panel are positive, suggesting that the distribution of partners for GVC-based intermediate products may have become more concentrated as a result of the SARS-induced supply-shock, though there may have been no effect on the number of trading partners (coefficient estimates reported in columns 1 and 2 in the top panel are statistically insignificant).

GVC-resilience
Since we would need data for 2019 and beyond to examine GVC-resilience to the MERS outbreak, we can only examine resilience of value-chains to the SARS epidemic. The results reported in Table 3 suggest that the adverse effects seemed to have accentuated significantly over time for both supply and demand-shocks emanating from the SARS outbreak, which is consistent with the epidemic's medium term impact observed on Chinese firm-level exports and imports in Fernandes and Tang (2020 P r e p r i n t n o t p e e r r e v i e w e d <Insert Table 3 here> On the whole, the results in this section suggest that there may have been a disruption to GVCs from the SARS outbreak at both intensive and extensive margins and that the adverse effects also seem to have persisted and intensified over time. This has implications for the current pandemic which we shall discuss in Section 7. In the following section, we dig deeper into the results for SARS to examine their likely drivers but before that we apply the Goodman-Bacon (2019) decomposition to the average treatment effect of the MERS outbreak as there is variation in treatment timing in this case.

Goodman-Bacon (2019) decomposition of the average treatment effect of the MERS outbreak
Goodman-Bacon (2019) shows that when there is variation in treatment timing, the 2FE DiD estimator is a weighted average of all possible two-group/two-period DiD estimators in the data. His decomposition cautions "against summarizing time-varying effects with a single-coefficient" showing that "when already-treated units act as controls, changes in their treatment effects over time get subtracted from the DD estimate". Figure 5 shows the results from the decomposition for the impact of MERS on GVC-trade. 13 The DiD estimates were found to be 0.54 and 0.17 emanating from the supply and demand-shock, respectively, which confirms that the MERS outbreak may not have had an adverse impact at the intensive margin. It also follows from the decomposition that the information contained in the different timing groups as well as the within estimates did not play a major role in the overall coefficient estimates; most of the weight (0.99) in the averages was given to the estimates of the difference in outcome between units that were never treated, versus those that were treated at different times.

Value-added trade
The results so far have examined the effects of SARS and MERS on gross trade in GVC-based intermediate and final products. In this sub-section, we use data from the UNCTAD-EORA 13 The Goodman-Bacon decomposition was implemented in STATA using the bacondecomp package. The bacondecomp implementation requires that the panel be strongly balanced, which meant that 24 countries, primarily small island states, were dropped from analysis over the period 2011-2018.

16
This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=3798581 P r e p r i n t n o t p e e r r e v i e w e d GVC database (Casella et al. 2019) and estimate equations (2) and (3) to examine how value-added trade was affected by these outbreaks incorporating both supply and demandside shocks in analysis. The results from these estimations are reported in Table 1, column (5) and provide qualitatively similar findings for the adverse effects of SARS on value-added trade, together with the absence of a statistically significant effect of the MERS outbreak. Table 1, column (5) in the top panel suggest that SARSinduced supply and demand-shocks may have reduced value-added trade by 0.6% and 5.4%, respectively; the combined adverse effect of SARS on value-added trade in these results is 5.9%. PTA-membership is also found to have a positive statistically significant effect on value-added trade in the results reported in the top panel.

Coefficient estimates reported in
6 Additional analysis on SARS

What drives the SARS effects?
In this sub-section, we exploit the heterogeneity of our dataset along different dimensions to examine the likely drivers of GVC-disruptions caused by the SARS epidemic using both aggregate and disaggregated data on gross trade in GVC-based products.

Aggregate analysis
To explore the likely drivers of the SARS results from aggregate analysis, we estimate equation (2) with additional interaction terms -SARS it * V ar i and SARS jt * V ar j -where V ar i/j = {N onOECD i/j , HGV C i/j , U i/j }. These time-invariant variables denote, respectively, the exporting/importing country not belonging to the group of OECD countries; dummy variables that are unity when measures of GVC participation 14 are greater than the respective median values for the sample of exporters/importers in the year 2000; and being upstream 15 from final demand. All coefficients are retrieved using the two-stage estimation procedure described in Section 4.1.
14 GVC participation is defined as the sum of backward and forward participation; these terms were constructed using EORA MRIO data for the year 2000 as the share of foreign value added (FVA) and indirect value added (DVX) in gross exports (GX), respectively (for instance see Aslam et al. 2017). FVA and DVX are vertical specialization measures from import and export perspectives, respectively (Hummels et al. 2001). In subsequent work, Johnson and Noguera (2012) and Koopman et al. (2014) have generalized the Hummels et al. (2001) vertical specialization measures to complex production chains.
15 Following Antràs and Chor (2018), upstreamness was measured by the share of a country's output sold directly to final consumers in the year 2000; the smaller the share, the more upstream is the country's position in GVCs.

17
This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=3798581 P r e p r i n t n o t p e e r r e v i e w e d <Insert Table 4 here> The incidence of weak contract enforcement on GVC-trade is disproportinately large (Antràs, 2020). Since institutional quality is correlated with economic development, high income countries are more likely to manage crises better by, inter alia, providing contractual security to minimize shock-induced GVC-disruptions effectively. This inference seems to be corroborated by the results reported in columns (1) and (2) of Table 4 for both GVC-based intermediate and final goods -the estimates of SARS it and SARS jt , that reference OECD countries, are positive; in contrast, the sums of the estimates of SARS it , SARS jt and their respective interaction terms are strongly negative. The coefficient estimates translate into average treatment effects for non-OECD countries ranging from -4.0% to -3.3% on intermediate and final goods on the supply side and to -1.9% and -12.5% on the demand side.
Since our analysis covers trade in GVC-based products, the extent of disruption to GVCs is likely to be positively correlated with the extent of countries' GVC-participation. This is borne by the results reported in columns (3) and (4) of Table 4, wherein the sums of the estimates of SARS it , SARS jt and their respective interaction terms are strongly negative. The coefficient estimates suggest that the outbreak-induced supply shock may have reduced GVC-based intermediate and final products imports of above-median GVC-integrated countries by 6.0% and 7.1%, respectively; the average treatment effects on the demand side range from -2.8% to -13.8%. These findings are also consistent with the theoretical literature on the role of supply-chains in propagating shocks through the economy (Acemoglu et al. 2012;Acemoglu and Tahbaz-Salehi, 2020). In contrast, the coefficient estimates of SARS it , SARS jt , that reference countries exhibiting low integration in GVCs, are positive for both GVC-based intermediate and final goods.
Finally, results reported in columns (5) and (6) of Table 4 suggest that countries more upstream from final demand may have not been adversely affected by SARS on both the supply (especially for GVC-based intermediate goods) and demand side, which is consistent with the findings in Fernandes and Tang (2020). The sum of the estimate of SARS it and its interaction term is positive for intermediate goods and less negative than SARS it for final goods, while the sum of estimated SARS jt and its interaction term is positive for both. In contrast, the SARS it , SARS jt estimates, which reference countries whose position is more downstream in GVCs, are strongly negative. The coefficient estimates translate into average treatment effects for countries downstream from final demand ranging from -6.5% to -5.5% for intermediate and final goods on the supply side and from -3.7% to -7.4% on the demand side. Trade costs have a higher incidence on downstream stages (Antràs, 2020) and these trade costs get compounded by shocks, which likely explains these results.

18
This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=3798581 P r e p r i n t n o t p e e r r e v i e w e d

Disaggregated analysis
To examine the likely drivers of the SARS effects using disaggregated product-level data, we estimate equation (2) at the product-level (a la Dai et al. 2014) with additional interaction terms -SARS it * V ar ip and SARS jt * V ar jp -where V ar ip/jp = {HRCA ip/jp , HKL ip/jp , M ED − T ECH p , HIGH − T ECH p }. These time-invariant variables denote, respectively, dummy variables that are unity when value of the normalized revealed comparative advantage index 16 and capital-to-labour ratio are greater than the respective median values for the sample of exporters/importers in the year 2000; and belonging to the technology class classified as medium-and high-tech. All coefficients are again retrieved using the two-stage estimation procedure described in Section 4.1.
<Insert Table 5 here> Exporting countries more competitive in producing certain goods are also likely to be relatively insulated from supply-side macroeconomic shocks affecting such products. This is observed in the results reported in Table 5 columns (1) and (2) for GVC-based intermediate and final goods. The analysis makes use of the revealed comparative advantage (RCA) indicator as a product-level proxy of competitiveness. The SARS it estimate and its sum with its interaction term is strongly positive for both intermediate and final goods.
More capital-intensive 17 and high-technology 18 products are also likely to be more sophisticated and less easily substitutable, and hence, prone to fewer disruptions (Fernandes and Tang, 2020). Results reported in columns (3)-(6) of Table 5 confirm these hypotheses in the context of SARS-induced disruptions to GVC-trade. The supply-and demand-shocks emanating from SARS are found to reduce exports of low capital-and technology-intensive intermediate and final products -the estimated coefficients of SARS it and SARS jt are negative in columns (3), (4) and (5), (6), respectively. In contrast, the SARS it estimate and its sum with its interaction term is strongly positive in columns (3)-(4) and in column (5) for high-tech products, while the SARS jt estimate and its sum with its interaction term is strongly positive for both high-tech intermediate and final goods in columns (5) and (6). Meanwhile, low-tech final products seemed to have been insulated from the adverse effects of SARS-induced supply-shocks in the results reported in column (6), though this may not have been the case for high-tech final products.
17 This is measured by the ratio of GFKF to the number of employees in data sourced from UNIDO's INDSTAT database at the ISIC Rev 3. four-digit level.
18 The technology classification of products is taken from UNCTAD.

19
This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=3798581 P r e p r i n t n o t p e e r r e v i e w e d In sum, the analysis in this subsection suggests that shock-induced GVC-disruptions may be more pronounced for non-OECD countries that are also more integrated and downstream in GVCs; and are less capital-and technology-intensive in intermediate goods production. The results from this subsection also show that our overall findings on the adverse effects of SARS are robust to using different sub-samples.

Alternative estimation strategy
In this sub-section, we use an alternative estimation strategy to estimate the trade effects of SARS, using the following specification: (8) The dependent variable in equation (8) now also includes data on intra-national trade flows, which not only makes the model theory-consistent (Fally, 2015) but also enables us to quantify the effect of SARS-induced supply-and demand-shocks (which are otherwise collinear with the time-varying exporter and importer fixed effects), using interaction terms between SARS it , SARS jt and a binary dummy (IN T L ij ) that takes the value one for international trade flows and the value zero for intra-national trade flows (see Piermartini and Yotov, 2016 for similar applications). As an additional robustness check, equation (8) is estimated using the Poisson Pseudo-Maximum Likelihood (PPML; Silva and Tenreyro, 2006), which also accounts for heteroskedasticity-related concerns in estimation, using data on cross-border and intra-national trade flows in manufacturing sectors 19 from the EORA26 MRIO database (Lenzen et al. 2012(Lenzen et al. , 2013.
The results from this estimation are reported in Table 6 and confirm both the adverse effects of SARS-induced supply-shocks on trade as well as their intensification over time.
The coefficient estimates translate into average treatment effects of -4.9% and -6.8% for intermediate and final goods exports over 2001-05, increasing significantly to -15.5% and -23.4%, respectively, over 2001-08. However, the effect of the demand shock emanating from SARS could not be estimated in these results as that interaction term was omitted due to collinearity. Meanwhile, PTA membership is found to have a positive impact on both intermediate and final goods exports in all these results.
Using difference-in-difference analysis in a gravity framework, we examine the response of GVC-trade to two previous health shocks. Our baseline estimates suggest a decline in GVCtrade, both gross and value-added, from SARS, emanating from supply and demand shocks, though a similar effect is not observed for MERS. Empirical analysis also suggests geographical diversification of value chains as well as their non-resilience to SARS in particular; even a relatively localized epidemic like MERS was associated with widening of value-chains from the outbreak-induced supply-shock.
It is tempting to compare SARS and COVID-19 given that both originated in China, but one must be mindful of the evolution of China's share in global GDP and trade over time, the inter-connectedness of the world economy and the severity of the current pandemic. During SARS, China accounted for 4% of global output; today, that number has quadrupled. Thus, any slowdown in China today will likely impact the world much more severely than in 2003, coupled with the effect of GVC-contagion as demonstrated by Friedt and Zhang (2020).
Moreover, the unique spatial dimension of the COVID-19 pandemic will not only yield a large global impact emanating from China, but also in smaller localized disruptions creating regional contagion effects. These effects will exhibit heterogeneity depending on a region's level of integration in global trade and the response of governments to the resultant economic crisis (OECD, 2020). For example, the EU is highly integrated in GVCs and is also a large producer and consumer of manufacturing goods; national governments in the region have responded with complete lockdowns, intermittently, during the pandemic. These shutdowns are likely to create strong ripple-effects in the exports of other countries located in Asia, Africa and the Americas that are dependent on European supply chains through backward and forward linkages (Solleder and Velasques, 2020).
The overall impact of COVID-19 is also likely to be worse than SARS because of three additional reasons. One, the current scale of the pandemic is much larger than that of SARS both in terms of incidence of cases and geographical spread (less than 10,000 lives were affected by the relatively more localized-SARS versus over 110 million confirmed cases of COVID-19 as of 28 February 2021 all over the world). Two, the state-mandated lockdowns in March 2020 resulted in immediate supply and demand shocks with lingering adverse effects on economies. Three, services trade is being more severely impacted this time as three of the four modes of services delivery require physical proximity betweeen buyers and sellers 21 This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=3798581 P r e p r i n t n o t p e e r r e v i e w e d and the latter is the first casualty of social-distancing practices in the wake of this pandemic (Shingal, 2020;WTO, 2020).
At the same time, the impact of COVID-19 will also depend on the type of products being traded through supply networks. Fernandes and Tang (2020) found capital and skill-intensive products to have been more resilient to the export disruption caused by SARS. These results are also consistent with our findings and with Taiji et al. (2018) who found that sourcing of differentiated inputs is less vulnerable to trade-shocks. The Chinese economy now specializes in a variety of products that are tech-and skill-intensive, which likely explains its faster recovery or small disruption given the inability of countries to find alternative sources for intermediates that are harder to substitute. This said, given that we find the adverse effects of SARS on GVC-trade to have intensified significantly over time, countries need to adopt strategies, both unilaterally and collectively, to mitigate likely medium and longer-term disruptions to GVCs from the current pandemic.
While any prior epidemic experience may be significantly associated with a less negative sentiment towards COVID-19 (Hassan et al. 2020), one lesson from the findings of our study for managing the current and future crises more effectively would be to diversify the portfolio of suppliers geographically. Recent anecdotal evidence suggests that this may already be happening 20 , as part of a larger China+1 strategy that began in 2019 following the US-China trade war with the aim of diversifying away from China towards other low-cost Asian countries. However, a second-order effect of the pandemic is expected to hit companies willing to relocate to ASEAN countries as these countries continue to be dependent on China for imports of intermediate inputs. Hence any relocation of production to these countries is not really diversification away from China.
While Southeast Asian countries seem to be the most preferred destinations after China, Mexico is also emerging as a close favourite especially for American and Japanese firms. Trade data are beginning to show that US companies are opting for suppliers closer home, 20 For example, Apple, Google and Samsung have already begun diversifying away from China since February 2020.
Apple will manufacture some mobile phones in Vietnam, India, Taiwan and Mexico and has already planned for an expansion into these new markets (https://www.bloomberg.com/news/articles/2020-03-27/iphone-makers-look-beyond-china-in-supply-chainrethink; https://asia.nikkei.com/Spotlight/Coronavirus/Google-Microsoft-shift-production-from-Chinafaster-due-to-virus). Google smartphone unit is set to move to Northern Vietnam, while it has already chosen Thailand for its smart-home product unit. Microsoft is also expected to start manufacturing in Vietnam soon. A similar trend is being observed in the textiles and clothing industry. During this pandemic, the Indonesian textiles industry has witnessed a 10% rise in the number of orders, primarily from global brands looking to substitute trade with China (https://www.fibre2fashion.com/industry-article/8679/indonesiantextiles-industry-likely-to-pull-through-global-pandemic). Another favored destination for relocation of textiles manufacturing has been Vietnam (for instance, the Japanese megabrand UNIQLO has moved sourcing from China to Vietnam).

22
This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=3798581 P r e p r i n t n o t p e e r r e v i e w e d chiefly local suppliers and those based in Mexico, a trend that resonates with the stylized facts on SARS presented in this paper. It seems that the US has also used the pandemic as an excuse to move pharmaceutical production back home from China. Moreover, US companies have already begun relying on locally-sourced electronic parts rather than sourcing them from China. This has led to an increase in orders to both local and Mexican firms.
Many non-US firms are also seeking new markets to shock-proof their supply chains. There is talk of nearshoring by companies in the Eurozone, for instance, to Hungary, Czech Republic, Slovakia, Romania and Poland (for instance see Javorcik, 2020) as a new manufacturing hub to move out of China following the pandemic. EU members have also urged automobile and pharmaceutical companies to strengthen local value chains as a way to reduce dependence on China. Japan is another country that has expressed concerns over reliance on China for imported inputs. As part of its COVID-19 stimulus package, the government set aside USD two billion to incentivize shifting production back home for high-tech manufacturing and in other sectors to South-east Asian economies or India.
In sum, while value-chains may have exhibited selective resilience to previous health shocks despite disruptions, there may be more permanent changes this time around, including a conscious move towards geographical and supplier diversification. At the same time, the pandemic has spurred e-commerce and is also likely to accelerate the fourth industrial revolution through adoption of automation, 3D printing and extreme customization. It would be interesting to study these changes and their ramifications on GVCs in future research on this subject.

23
This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=3798581 P r e p r i n t n o t p e e r r e v i e w e d  Note: All columns report estimates from the two-stage regression described in the text. Robust standard errors, clustered by dyad-year in stage one and by exporter-year, importer-year in stage two, reported in parentheses. Significance levels: *10%, **5%, ***1%. Note: All results based on the two-stage regression approach described in the text. Coefficients of PTA ijt , NonOECD i/j , HGVC i/j and U i/j nor reported. Robust standard errors, clustered by dyad-year in stage one and by exporter-year, importer-year in stage two, reported in parentheses. Levels of significance: *10%, **5%, ***1%. Legend: INT = Intermediate products; FNL = Final products. Note: All results based on the two-step regression approach described in the text.The product is not defined at HS6 in columns (3) and (4) where it is defined at ISIC Rev.3, which also accounts for the lower number of observations in those columns. Coefficients of ln(1+TAR ijpt ), HRCA ip/jp , HKL ip/jp , MED_TECH p and HIGH_TECH p are not reported. Robust standard errors, clustered by dyad-product-year in stage-one and by exporter-product-year, importerproduct-year in stage two, reported in parentheses. Levels of significance: *10%, **5%, ***1%. Note: Data on X ijt also include intra-national trade flows. INTL ij is a binary dummy that takes the value one for international flows and zero otherwise. SARS it , SARS jt and SARS jt .INTL ij omitted due to collinearity. Robust standard errors, clustered by dyad-year, reported in parentheses. Levels of significance: *10%, **5%, ***1%. This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=3798581 P r e p r i n t n o t p e e r r e v i e w e d Annex