International Trade and the Transmission of Temperature Shocks

Schenker, Oliver; Osberghaus, Daniel Article — Published Version International Trade and the Transmission of Temperature Shocks Environmental and Resource Economics Provided in Cooperation with: Springer Nature Suggested Citation: Schenker, Oliver; Osberghaus, Daniel (2025) : International Trade and the Transmission of Temperature Shocks, Environmental and Resource Economics, ISSN 1573-1502, Springer Netherlands, Dordrecht, Vol. 88, Iss. 4, pp. 965-1007, https://doi.org/10.1007/s10640-025-00957-3 This Version is available at: https://hdl.handle.net/10419/323358 Standard-Nutzungsbedingungen: Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Zwecken und zum Privatgebrauch gespeichert und kopiert werden. Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich machen, vertreiben oder anderweitig nutzen. Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, gelten abweichend von diesen Nutzungsbedingungen die in der dort genannten Lizenz gewährten Nutzungsrechte. Terms of use: Documents in EconStor may be saved and copied for your personal and scholarly purposes. You are not to copy documents for public or commercial purposes, to exhibit the documents publicly, to make them publicly available on the internet, or to distribute or otherwise use the documents in public. If the documents have been made available under an Open Content Licence (especially Creative Commons Licences), you may exercise further usage rights as specified in the indicated licence. http://creativecommons.org/licenses/by/4.0/ Accepted: 6 January 2025 / Published online: 30 January 2025 © The Author(s) 2025 Oliver Schenker [email protected] Daniel Osberghaus [email protected] 1 ZEW - Leibniz Centre for European Economic Research, L7,1, DE-68161 Mannheim, Germany International Trade and the Transmission of Temperature Shocks OliverSchenker1· DanielOsberghaus1 Environmental and Resource Economics (2025) 88:965–1007 https://doi.org/10.1007/s10640-025-00957-3 Abstract We examine how the adverse impacts of weather shocks are distributed through the trade network. Exploiting a rich, theoretically derived, fixed effects structure, we find significant negative short-run effects of high temperature on exports. A month with an average temperature above 30 ◦C implies export losses of around three percent. These effects are increasing in the labour-intensity of exports. Using our structural Gravity model, we assess the general equilibrium incidence of these temperature shocks. We find that equilibrium adjustments reduce the economic costs by around 20 percent, but significant costs arise also for countries not directly exposed to high temperatures. Keywords International trade · Temperature · Extreme weather · Structural Gravity JEL Classification F14 · F18 · Q54 · Q56 1 Introduction Climate change is a global problem. The emission of greenhouse gases impacts ecosystems and economies around the globe, independently of the location of the emitter. In order to infer the sensitivity of economic activities to climate change a growing literature estimates how weather in general, and temperature in particular, affects aggregate economic outcomes. Most studies in this field analyse local outcomes of local weather variation (Dell et al. 2012; Burke et al. 2015). But economies are not isolated from each other. As international trade links the fortune of economies, the economic impacts of local weather events may disseminate through the global trade network, creating a spatial disentangling of the occurrence of weather events and their full economic consequences. Furthermore, most weather 1 3 O. Schenker, D. Osberghaus events occur only for a few days or weeks, which may make it challenging to identify their impact in annual economic data. Thus, in order to comprehensively assess the economic costs of weather events, we need to understand the propagation of the costs of these events across space using data with high temporal resolution. Our study is—at least to our knowledge—the first that provides robust ex-post empirical evidence of the international transmission of the costs of extreme weather events using monthly trade data. Based on a refined structural Gravity model we show that high temperature events in one country cause economically and statistically significant costs also in countries not directly exposed to the event. Recently, new economic geography models started to include spatial heterogeneity of climate change impacts, showing that the margins of spatial economic adjustments to climate impacts by either migrating or by shifting specialisation patterns are important mechanisms to adapt to climate change (Conte et al. 2021; Cruz and Rossi-Hansberg 2024). Understanding the spatial spillovers of local weather events has also important consequences for cost-assessments of climate change and thus the decision-making of policy makers. As the frequency and severity of extreme weather events is likely increasing with further global warming the propagation of the costs of extreme weather events through the global trade network is relevant for a comprehensive account of the social cost of carbon (SCC), a key figure used in cost-benefit analysis that puts a monetary value on the impacts of climate change caused by one ton of carbon (Wagner et al. 2021). In the U.S., the former Trump administration revised the official, formerly global, SCC figures, considering only domestic climate impacts in continental United States. This has been reversed by the Biden administration, mainly based on ethical and fairness arguments. But if costs of climate change are propagated through international trade, this provides an argument to take into account global climate impacts for the SCC calculation even from a pure selfish perspective.1 It is therefore important to measure the extent of climate impacts on trade flows and their propagation through the global trade system. This paper therefore aims at answering three research questions: (i) Do extreme weather events affect exports? (ii) If so, through which channels do these events affect bilateral trade and which characteristics govern the effects? (iii) And, finally, what is the spatial incidence of the costs of these events and how much of these costs arise in not directly exposed countries? In order to identify both partial and general equilibrium effects of weather events on international trade, we build a structural Gravity model, where weather events affect monthly output and, consequently, importers face supply losses from affected exporters. The model explicitly describes demand and price shifts, as importers respond with substitution from other sources. This model is then estimated using five decades of monthly observations of bilateral trade and weather data, including more than 20,000 country pairs and about 4.5 million observations. Estimating a structural Gravity model provides a well-suited framework to identify the impact of weather on aggregate economic outcomes as the identification builds on established trade theory and a robust empirical relationship. While we perform the empirical analyses for various types of extreme weather events, we focus on episodes of high ambient temperature in the main part of the analysis. Consistent with our derived general equilibrium model, we are able to exploit a rich and theory-derived structure of exporter (importer) × 1 See Kotchen (2018) for a theoretical analysis of this argument. 1 3 966 International Trade and the Transmission of Temperature Shocks year, country pair × year, and exporter(importer) × calendar month fixed effects that allows to tightly estimate the temperature impacts on monthly bilateral exports. We find highly significant negative non-linear effects of high absolute temperature and extreme temperature deviations in the exporting country on the value of contemporaneous gross exports. In a month when the average temperature is at least 30 ◦C , exports decrease by 3.4 percent relative to a month with an average temperature below this threshold. Using an alternative specification of extreme deviations from country-specific mean temperatures, we find that a top-percentile temperature shock in the export country reduces the export value by 2.1 percent. We then examine if specific exporter characteristics govern the effect size and find that the impact rises with the labour-intensity of the exports. This is consistent with the notion that short-run temperature impacts on exports might be governed by labour productivity or labour supply effects in the exporting country due to of physiological heat stress, as suggested by substantial micro-empirical evidence. Equipped with these estimates, which inform our structural Gravity model, we compute the counterfactual global trade equilibria in absence of high temperature events. This allows us to calculate the cost-incidence of such an event for the country directly exposed to the temperature shock as well for countries only indirectly exposed through trade links with the affected location. Measuring costs as losses in trade relative to the counterfactual scenario without temperature shock, we find that the mean high temperature shock has statistically significant global costs of 360 million USD. About two-thirds of these costs appear in countries not directly exposed to the temperature event, suggesting that substantial parts of the costs of these shocks are transmitted and propagated through the trade network. In a final step, we analyse the magnitude of these spillovers under climate change projections based on twenty-year monthly temperature averages from global climate models. We find that under a middle-of-the-road climate projection of the period 2020–2039, annual global trade is reduced by about 735 million USD due to additional high temperature events relative to 2015. 1.1 Literature Review Growing microand plant-level evidence suggests that high ambient temperature has detrimental impacts on labour productivity and supply. Using survey data on time allocation of individuals, Zivin and Neidell (2014) find evidence for a substantial reduction of labour supply in climate-exposed industries such as agriculture, construction and manufacturing in non-climate-controlled facilities on days with maximum temperature above 85 ◦ F (29.4 ◦ C). Given the evidence on the temperature-productivity relationship of individuals, one might suggest that such temperature effects prevail also on plant-level. Looking at the nearuniverse of Chinese manufacturing plants from 1998 to 2007, Zhang et al. (2018) find an inverted U-shape relationship between temperature and total factor productivity (TFP). Their estimates show that for the average plant on a day with maximum temperature above 90 ◦ F (32.24 ◦ C), TFP decreases by 0.56 percent relative to a day with 50–60 ◦ F (10–15 ◦ C), translating into an estimated output loss of 0.45 percent for the average plant. Similar findings are documented using firm-level data from India. Somanathan et al. (2021) provide evidence that annual plant output falls by about 2 percent if every day would warm by 1 ◦ 1 3 967 O. Schenker, D. Osberghaus C. This loss appears to be driven by a reduction in the output elasticity of labour due to an increasing rate of absenteeism and a decrease in labour productivity. This research provides the micro-economic foundation of the macro-level impact of high ambient temperatures on economies. Measuring aggregate impacts of temperature changes on economic growth rates has been the aim of a number of influential studies such as Dell et al. (2012) and Burke et al. (2015). Assuming a log-linear relationship of temperature and economic activities, Dell et al. (2012) find a substantial negative effect of temperature changes on GDP growth, but only in poor countries: a 1 ◦C rise in annual average temperature reduces economic growth by about 1.3 percentage points. Burke et al. (2015) argue that the aggregate impact of temperature on economic outcomes is non-linear, suggesting a concave function with economic productivity peaking at 13 ◦C . Based on sub-national data, Kalkuhl and Wenz (2020) support the evidence that temperature variation affects aggregate outcomes in a non-linear fashion. However, this literature focuses on local effects of local events. Given the economic relevance of international trade, focusing on local temperature might provide an incomplete picture of the impacts of weather events on the economy. As Jones and Olken (2010) say: "international trade links the fortunes of countries providing important conduits for geographically limited climatic impacts to have global economic effects." Using reduced-form regressions with product-level export panel data they find that export growth is reduced by 2.0–5.7 percentage points in poor countries if annual temperature increases by 1 ◦ C. Product-level analyses show that this is driven in particular by adverse effects on agriculture and light manufacturing. This finding has been confirmed by Dallmann (2019), who additionally controls for temperature (and precipitation) impacts at the importer location. Also using a linear temperature specification in an annual time-scale, she finds that each 1 ◦ C warming in the exporter country reduces bilateral exports by 3.1 percent, but does not find significant effects of the importer’s temperature. Our paper differs in five important aspects from these previous studies. First, our estimation is derived from a general equilibrium trade model. The derived estimated Gravity equation has been proved to be empirically robust in many applications in international economics. Jointly with our tight, theoretically derived, high-dimensional fixed effects structure, this should minimize omitted variable bias, improve identification and generate robust estimates of the weather variation effects on an aggregate economic outcome such as exports. Second, we use data with a monthly temporal resolution while most of the previous macro-level studies rely on annual data. Heat waves typically last only for a few weeks. But as aggregate statistics such as GDP are only available on a quarterly or even annual basis, it may be difficult to identify the effects of those events in aggregate data with sufficient accuracy. The few existing studies using monthly trade data do either not analyze global data (such as Karlsson (2021), focusing on U.S. exports), or do not study temperature effects (Tembata and Takeuchi 2019; Felbermayr et al. 2020). The monthly temporal resolution additionally enables an analysis of lagged impacts of weather events. We find that exports are mainly affected during and directly after the event but do not find evidence for a substantial compensation for the lost exports in post-event months. Third, previous studies show that the functional relationship between aggregate economic outcomes and temperature remains heavily debated. Newell et al. (2021) test several hundred functional forms and find that non-linear temperature specifications dominate the 1 3 968 International Trade and the Transmission of Temperature Shocks model set in terms of predictive ability. Based on our large data set, we estimate the effects of single ◦C bins, thereby allowing for a high degree of flexibility in the functional form. Fourth, while we, similar to Jones and Olken (2010), find evidence for larger effects of high temperature events on exports of manufactured goods, we use a more direct approach to identify labour-productivity effects as a key channel. Using input-output data we find that high labour-intensity of exports correlates with a stronger negative impact of high temperature events on exports. Fifth, and of particular relevance, by exploiting our estimated structural model, we simulate the equilibrium adjustments caused by a high temperature episode. This enables us to estimate the cost-incidence of such events also for countries only indirectly affected, taking into account their opportunities to adjust imports in response to a shock abroad. Besides contributing to literature on weather effects on international trade, our paper adds to a growing literature that studies the spatial transmission of natural shocks such as disasters or weather extremes more explicitly. This is particularly relevant as impacts of climate change are and will be unevenly spread across countries and, as Costinot et al. (2016) point out: "[i]n a globalized world, the impact of micro-level shocks depends not only on their average but also on their dispersion over space." Using a spatially highlyresolved crop field model, they study general equilibrium adjustments of crop productivity shocks from climate change. They show that adjusting planted crop types in response to changes in comparative advantage is an important force to reduce costs of climate change in agriculture. Relative to this, international trade plays only a minor role in alleviating the consequences of climate change. Similarly, Conte et al. (2021) develop a calibrated spatial economic model to explore changes in specialization as an adaptation mechanism. In contrast, we do not address this margin explicitly as in our model each country produces a specific, non-homogeneous good. Thereby, we implicitly abstract from adjustments in the domestic production processes. Desmet et al. (2021) study the changing spatial distribution of the economic activity due to climate change induced sea level rise in a calibrated model of the world economy at a 1◦ by 1◦ resolution and show as well that the dynamic spatial equilibrium adjustment is an important adaptation mechanism. Lower resolved country-level computable general equilibrium (CGE) models informed by climate impact projections, such as Schenker (2013) and Knittel et al. (2020) also point out that international trade is an important redistribution mechanism of climate impact costs such that for some regions these imported impacts can be responsible for a substantial part of the total cost of climate change. Different to these studies, which rely on carefully calibrated models of future climate conditions, we exploit past weather and trade data to estimate the dispersion of these effects across space. At the firm level, Barrot and Sauvagnat (2016) find that when one of their suppliers is hit by a large natural disaster, firms experience an average drop of 2–3 percentage points in sales growth. This is supported by findings of Pankratz and Schiller (2024) who show that firm-performance is negatively affected if large suppliers of these firms have been exposed to extreme weather events but also that the downstream firms respond and adjust their supply chains to less exposed suppliers. While these papers study firm-level responses to exogenous shocks, we focus on aggregate impacts on the macro-level. The remainder of the paper develops in Sect. 2 the analytical general equilibrium Gravity model from which we derive our estimation equation. Section 3 discusses the empirical approach and introduces the estimation framework. Section 4 presents the data and the 1 3 969 O. Schenker, D. Osberghaus construction of the variables. Section 5 shows the estimated temperature effects on bilateral trade. These estimates lay the groundwork for the counterfactual simulations, presented in Sect. 6, including ex-ante simulations based on future climate projections. Finally, Sect. 7 concludes the analysis. 2 Model We build on a simple general equilibrium trade model where each country produces a specific variety which is traded with the rest of the world—i.e. goods are differentiated by origin as in Armington (1969). Consumers have constant elasticity of substitution (CES) preferences for these country-specific goods. The CES-Armington general equilibrium model, whose theoretical underpinning goes back to Anderson (1979), is the workhorse model in structural Gravity research (Head and Mayer 2014). We extend this model in two important dimensions: First, we model how weather shocks can affect the production of output. Second, we take into account intra-annual variation in production and consumption. 2.1 Consumption Each point in time can be characterized by the set union of year index t∈{1, .., T} and calendar month m∈{1, .., 12} . Each country j∈{1, .., N} is populated by a representative agent with CES utility. As we explain below, the model controls for important known determinants of weather shocks such as the geography and intra-annual climate variation. Thus, the realisation of a weather shock is ex-ante unknown and economic agents are myopic with respect to the occurrence of these idiosyncratic weather shocks. Hence, we assume that the representative agent maximizes her utility at each point in time independently of past or future expectations or decisions. Thus, utility of the representative agent in country j in year t and month m is described by U jtm = ( N ∑ i=1 λ 1−σ σ iC σ−1 σ ijtm )σ σ−1 , (1) where Cijtm denotes country j’s consumption of the specific variety imported from country i at point in time {t, m} . This consumption expression can be decomposed in two components: First, there is aggregate annual consumption of good i in j in year t, Cijt . Second, holiday seasons, accounting or exogenous inventory management motivations, as well as other factors shape the intra-annual demand variation, captured by the exogenous consumption shifter ϕjm , which is normalized such that ∑12 m=1 ϕ jm =1 . Hence, Cijtm =Cijt ϕjm . λi>0 describes an exogenous preference parameter for goods from country i and σ>1 is the elasticity of substitution among varieties of different origins. While trade balances are exogenously fixed across years, we assume that intra-annually countries can run trade balance surpluses or deficits. Thus, consumers in j maximize equation (1) subject to the annual budget constraint ∑i τ ijt p it C ijt = E jt , where Ejt is total 1 3 970 International Trade and the Transmission of Temperature Shocks annual expenditure for consumption, pit denotes factory-gate prices of good i and τijt are iceberg trade costs. Important determinants of these trade costs are time-invariant characteristics such as the geographical distance between exporter and importer and common cultural attributes. In addition, trade costs may also have a time-varying component as new trade agreements come into force or improvements in infrastructure reduce transport costs. Note that we assume that bilateral trade costs change only annually rather than monthly. While in reality, trade agreements or infrastructure improvements come into force in a particular month, this assumption simplifies our identification strategy.2 We also assume that factorygate prices pit are sticky and change only annually as menu costs impede price adjustments and contractual agreements fix prices for certain periods.3 Solving the representative agents optimization problem yields annual demand C ijt = ( λiτijt pit Pjt )1−σ Ejt , (2) where the associated consumer price index in country j is given by P jt = ( N ∑ i=1 (λiτijt pit)1−σ ) 1 1−σ . (3) For given prices, we can thus define the propensity of country j to spend on imports of good i at date {t, m} with θijtm =pit τijt Cijt ϕjm . 2.2 Output and Weather Shocks But j’s import spending propensity for good i cannot be satisfied in any case. First, also the production in i is exposed to exogenous, intra-annual shifts and subject to seasonal variation. This is most obvious in the case of agricultural production which depends on harvesting cycles. Other drivers of these country-specific intra-annual cycles are holiday seasons or annual cyclical weather impacts which also affect production and transport infrastructure such as tropical cyclones. This is captured by the weight parameter φim , which describes the exogenous monthly country-specific variation of output, normalized such that ∑12 m=1 φ im =1 . Second, there are potential weather shocks affecting output beyond these cyclical patterns.4 Let us assume that Witm =exp(ρ1itm(Ditm)) describes the potential weather shock affecting output in country i in year t and calendar month m. If the indicator 1itm(Ditm) is equal to one a weather shock materialises and production of good i is exposed to the extreme 2 Most trade costs changes are long-term and their temporal implementation probably uncorrelated to monthly weather variations. This assumption should therefore not lead to biased estimates of weather impacts. 3 As seen below, presumably temperature-induced labour-productivity effects in the manufacturing sector are a key determinant of the measured aggregate impact on trade. These goods are often relatively specific and are not traded on global spot markets but have rather sticky prices. For instance, Apel et al. (2005) find that the median Swedish firm in their sample adjusts prices just once a year. 4 Motivated by the above discussed micro- (e.g., Somanathan et al. (2021)) and macro-level (e.g., Burke et al. (2015)) studies, weather shocks affect output in our model. However, one could also model weather shocks on preferences or trade costs. 1 3 971 O. Schenker, D. Osberghaus weather event Ditm . In general, this could be a month of extreme high or low temperature, heavy or poor rainfall, or high wind speed. Otherwise, 1itm(Ditm)=0 . The parameter ρ measures then the weather shock’s impact on output. Identifying ρ is one of the key aims of our empirical exercise. Let us denote Xitm as the total free on board (f.o.b) value of exports of i. Then, X itm =∑jX ijtm, where Xijtm is the value of actual bilateral exports net of intra-annual supply shifts and weather shocks from country i to j. 2.3 International Trade Hence, for given prices, actual exports of country i to j at time {t, m} can be expressed as Xijtm =θjitm φim Witm . Solving for θjitm and plugging this into the demand equation (2) leads to X ijtm = ( λiτijt pit Pjt )1−σ Ejt ϕjm φim Witm . (4) Annual total exports of country i in year t are thus X it = N ∑ j=1 12 ∑ m=1 φim Witm ( λiτijt pit Pjt )1 −σ Ejt ϕjm . After rearranging this expression and dividing by ( λ i p it)1−σ , following Anderson and Van Wincoop (2003), we define the term on the right hand side as Π 1−σ it = N ∑ j=1 12 ∑ m=1 ( τijt Pjt )1 −σ Ejt ϕjmφim Witm , (5) the so-called outward multilateral resistance term. Since ( λ i p it)1−σ= X it / Π1−σ it , we plug this into equation (4) and get the Gravity equation that describes the monthly exports of country i to j: X ijtm =Xit φim Witm Ejt ϕjm ( τijt ΠitPjt )1−σ . (6) Equation (6) is the equation we are going to estimate. It is this equation that transmits the weather shock from the exporters location to the importer, affecting the availability of goods from i. Assuming that ρ<1 , the occurrence of a weather shock in i reduces bilateral import of j by exp(ρ) at that particular point in time. Ceteris paribus, households in j face a potential import loss. But ignoring the general equilibrium response from substitution and price adjustments may be misleading. With a positive σ , consumers in j are able to substitute goods of different origin, so a weather-caused shortage of supply from one country can, at least partially, be compensated by imports from other locations. 1 3 972 International Trade and the Transmission of Temperature Shocks abnormally high temperature shocks, identified by observations in the highest percentile of the country-specific temperature distributions. This specification, presented in column (2) in Table 1, confirms a contemporaneous short-run non-linear temperature effect on exports. Comparing columns (1) and (2) in Table 1, the question arises whether the two specifications describe the same economic processes, or whether the effects of absolutely and abnormally high temperatures are independent impacts. In the latter case, the underlying impact channels may differ, indicated by significant marginal effects in an estimation including both temperature specifications. Therefore, in column (3), we include both absolutely and abnormally high temperature, and in column (4) we additionally study the interaction between them. For exporters, both temperature variables ( Dt30 itm and Dp99 itm ) remain individually significant. Moreover, the effects are independent from each other as there is no statistically significant interaction effect. Hence, both effects can be observed in the short term: a substantial effect on exports from episodes of absolute hot temperature, and a somewhat smaller effect of months of abnormally high temperatures, given the country-specific temperature distribution. In the Appendix, we also test a temperature impact model with a linear, as well as a quadratic specification, as often assumed in the literature — see, for example, Missirian and Schlenker (2017) or Burke et al. (2015). The results, presented in columns (1) and (2) in Table 4 in the Appendix, confirm a non-linear effect of temperature on exports. Moreover, we assess the question whether the effects of absolute and abnormally hot months affect the extensive or intensive margin or both. Therefore, we estimate linear probability models of the binary variable whether there is trade between countries (columns 3 and 4), and restrict the sample to positive export flows (columns 5 and 6). The results (see Table 4 in the Appendix) suggest that the effects stem from changes at the intensive margin: Temperature events have no significant impact on the decision of whether exports occur or not, but reduce the value of exports in the subsample of positive export flows. Equipped with this series of estimation results for contemporaneous effects of temperature, we obtain one robust finding: Months with high temperatures, either measured in absolute or relative terms, lead to a statistically and economically significant reduction of exports (1) (2) (3) (4) Abs. Temp. Abn. Temp. Both Interaction Dt30 itm −0.0336*** −0.0253** −0.0166 (0.0100) (0.0124) (0.0140) Dt30 jtm −0.0124** −0.0092 −0.0095 (0.0056) (0.0058) (0.0059) Dp99 itm −0.0210*** −0.0206*** −0.0203*** (0.0076) (0.0076) (0.0077) Dp99 jtm −0.0149** −0.0148** −0.0149** (0.0064) (0.0064) (0.0063) D t 30 itm × D p99 itm −0.0236 (0.0190) D t 30 jtm × D p99 jtm 0.0019 (0.0089) Observations 3821155 3821155 3821155 3821155 Table 1 Contemporaneous temperature effects ρ on exports The dependent variable is bilateral exports in current USD. Index i (j) indicates the exporting (importing) country. Standard errors in parentheses 1 3 979 O. Schenker, D. Osberghaus relative to months facing lower (or less extreme) temperature levels. This effect is well captured by the parsimonious models focusing on Dt30 itm or Dp99 itm (columns (1) and (2) in Table 1). Reviewing the prior literature on temperature effects on macro-economic outcomes (see literature review in Sect. 1), reveals a clear focus on absolute temperature specifications, as there is a sound and robust micro-economic and physiological empirical foundation of these effects, while there is less theoretical and micro-econometric support for economic impacts from abnormally high temperatures. Further analysis reveals that the observed effect is primarily driven by episodes of extremely high absolute temperatures in warmer countries, rather than by episodes of relatively high temperatures in colder countries, which are moderate in absolute terms (see Fig. 13 in the Appendix). Therefore, in the remainder of the analysis, we are going to focus on impacts of high absolute temperature ( Dt30 itm ). However, we replicate all analyses with specifications based on abnormally high temperatures, report the results in the Appendix, and highlight potential qualitative differences. 5.2 Lagged Temperature Effects We have shown that months of high temperatures have a detrimental effect on exports in the month of their occurrence. However, it is important to understand the duration of these impacts. Are they only short-lived or do they have longer term consequences? For assessing this question, we estimate a finite distributed lags model with four months before and twelve after the temperature event. To reduce computational complexities, we rely on the parsimonious model with a dummy for months with average temperature greater or equal 30 ◦C . Figure 3 depicts the estimated coefficients, relative to a month where temperature has been below 30 ◦C . The estimation confirms the contemporaneous effect of a temperature shock on exports in the month of the event (−6.7 percent, p=0.031 ). We also find a lagged negative effect three months after the event (−5.5 percent, p<0.001 ) and a positive effect after seven months (+4.2 percent, p=0.009 ). The estimated effects prior to the temperature event, serving as placebos, are statistically non-significant. The cumulative effect over the period of one year after the event remains negative, albeit at non-significant levels (see Fig. 15 in the Appendix). We conclude that temperature shocks on exports manifest mainly in the short term during and directly after the event. The effect is neither substantially aggravating over time nor is there any evidence of a substantial catching up or compensation for lost exports in post-event months, such that the cumulative effect after one year is in the same order of magnitude as after a few months. For abnormally high temperatures, the analysis of lagged effects yields similar results: The effect proves to be short-lived and only existent for exporters (see Figs. 16 and 17 in the Appendix). However, the initially negative effect of abnormally hot months is followed by some (statistically non-significant) positive lagged effects, such that the cumulative effect after one year of the event is not statistically different from zero. In summary, it can be stated that across all model specifications the temperature shock in the exporting country is more relevant than in the importing country, supporting similar findings based on reduced-form, parsimonious estimates with annual trade data (Dallmann 2019). These identified temperature impacts are relatively short-lived. Consequentially, we focus in the remainder of the analysis on the contemporaneous temperature effects in the 1 3 980 International Trade and the Transmission of Temperature Shocks exporting country. However, we keep the temperature in the importing country as a covariate in all estimations. 5.3 Heterogeneous Effects of Temperature So far, we estimated average effects of high temperature events on exports across the full sample. But the effect magnitude might be conditional on country or trade flow characteristics. Understanding these differences allows us to infer more precisely about impact channels and economic mechanisms translating the temperature shock in economic outcomes. One important impact channel identified by prior microand plant-level evidence suggests that high absolute ambient temperature reduces productivity and supply of labour (Somanathan et al. 2021; Zhang et al. 2018; Zivin and Neidell 2014). Hence, we incorporate the labour intensity of yearand importer-specific exports ( labourintijt ). If labour productivity (or supply) is a major determinant of the magnitude of the adverse temperature effect on exports as suggested, the estimated effect should vary with the labour intensity of bilateral exports.9 We augment the basic estimation equation (10) with interaction terms of the heterogeneity variable with Dt30 itm . Direct effects of labour intensity and other heterogeneity variables are not estimated as they are perfectly collinear with fixed effects at country pair-year level. 9 While we focus here on labour intensity given the micro-economic evidence, we similarly test for other potential sources of effect heterogeneity in the Appendix. Inspired by Dell et al. (2012), we interact the temperature effect with the annual income in the exporting country ( gdpit ) and the annual share of agricultural production in the exporter’s GDP (including forestry, fisheries and hunting, agriit ). We further hypothesize that the exporter’s resilience towards climate change (measured by the annual ND-GAIN index, ndgainit ) may govern the response to a temperature shock. Finally, effects may vary with the product composition of total exports (Jones and Olken 2010). Therefore, we interact the temperature shock with product-specific shares of total exports in the preceding four years (e.g., prF oodijt for food and live animal products). Fig. 3 Lagged impact on exports of an average monthly temperature of at least 30 ◦C . Estimated effects ρl of a hot month on exports, including 95-percent confidence intervals. The effects are relative to a month with a temperature below 30 ◦C . Lagged impacts of hot months at the importer location are depicted in Fig. 14 in the Appendix 1 3 981 O. Schenker, D. Osberghaus We estimate marginal effects of Dt30 itm on exports over different levels of the heterogeneity variable. Figure 4 depicts the results for labour intensity, and Fig. 19 in the Appendix summarizes similar plots for the other potential heterogeneity variables. The results suggest that the contemporaneous effects of absolutely hot months is indeed governed by the labour intensity of exports. The interaction effect shown in Figure 4 is highly significant ( p=0.003 ). A 10 percentage points increase in labour intensity of exports is associated with an increase of the adverse impact by approximately 5 percentage points. On the contrary, there is no significant interaction effect with relative temperature extremes ( Dp99 itm , see Figure 18 in the Appendix). This is broadly in line with prior micro-economic and physiological literature (see e.g., Dunne et al. (2013)), which postulates that absolute temperature levels are crucial for potential declines of work capacity.10 Hence, an unusually warm summer in a cold environment would be treated as an extreme temperature event in Dp99 itm , but has no expected adverse impact on labour productivity. In a similar vein, Jones and Olken (2010) detect a negative effect of high temperatures on exports of light manufacturing goods and speculate that productivity effects of workers might be responsible. Our results provide further evidence that labour productivity or supply is key for understanding the mechanism behind the identified effects of temperature shocks on aggregated exports. Note that our data set contains only trade in goods but not in services. Since services are typically more labor-intensive than goods, our estimate of the full impact of temperature shocks on exports therefore likely represents a conservative assessment and a lower bound of the full effect. 10 While other factors such as wind speed and humidity are important, substantial losses of work capacity are generally only observed at absolute temperature levels of higher than 25 ◦C . Fig. 4 Estimated effect of an average monthly temperature of at least 30 ◦C on exports for different levels of labour intensity. Estimated contemporaneous effects ρ of Dt30 itm on exports for given levels of labour intensity, including 95-percent confidence intervals. The effects are relative to a month with temperature below 30 ◦C . Labelled values at the x-axis are the 5th , 50th , and 95th percentile of labourintijt 1 3 982 International Trade and the Transmission of Temperature Shocks Most of the other analyzed variables show no significant interaction with the temperature effect (Fig. 19 in the Appendix). In our structural Gravity model, the temperature effects on exports do neither significantly vary with economic development, the share of agricultural goods in the exporting country’s production, or the exporter’s assessed resilience to climate change, nor over time (Fig. 21 in the Appendix). Similarly, most of the product category shares do not interact with temperature impacts — with the exceptions of exports characterized by high shares of Crude Materials (more adversely affected) and Mineral Fuels (less affected). These results, however, are compatible with the interpretation that labour productivity is the underlying channel of temperature effects — as labour intensity is relatively high for the former and low for the latter. 5.4 Extensions: Impacts of Precipitation and Storms While we focus on the effects of high temperature, the data and the employed empirical methodology generally allow for an equivalent analysis of the impacts of other weather phenomena. We are particularly interested in weather events that may be affected by climate change, and therefore additionally assess impacts of extreme precipitation and storms. The employed data and obtained results are summarized in Appendix 8.4.. Regarding precipitation, we do not find any contemporaneous effects on exports or imports (see Table 6). Considering the potential impact channels of hydrological events on production processes, this non-effect may be plausible: precipitation is most relevant for the production of agricultural goods. But these goods have certain growing periods, and their processing also needs time, such that a contemporaneous effect of a lack of precipitation on the production and hence on exports is unlikely. But perhaps it is the combination of high temperature with low precipitation that leads to droughts and negative impacts on exports. If high temperature events are highly correlated with episodes of low precipitation, our estimated temperature effects might mask drought effects. We find a positive but low correlation of the appearance of low precipitation events with high temperature events (Spearman’s ρ=0.07 ). This aligns with the findings of an estimation model that incorporates low-precipitation events as controls, yielding temperature effects nearly identical to those of the baseline model (see Table 6 in the Appendix). Extreme high precipitation, in contrast, may have adverse effects on trade, e.g. by flooded transport infrastructure or production facilities. The fact that we do not find such an effect may be due to an inaccurate measure of flood intensity. As a monthly mean value, our precipitation measure may not properly indicate short-term extreme events which last only for one or two days. Furthermore, for being harmful to the economy, the occurrence of intense precipitation events must coincide with the location of vulnerable assets or infrastructure — issues that are not sufficiently detectable by country-month averages. Similarly, we find only limited evidence for the existence of storm impacts on exports. Linear and quadratic specifications of monthly maximum wind speed values yield insignificant effects. Extreme wind speed events, defined as being in the country-specific top percentile of wind speeds, are without effect as well. However, we find non-linear effects of very intense storms in absolute terms (of 140 knots maximum wind speed and higher) when analyzing a flexible model using wind speed bins. In these months, exports decrease substantially (up to 7 percent), which may hint to impaired transport infrastructure or production facilities. 1 3 983 O. Schenker, D. Osberghaus 6 Counterfactual Simulations Our previous analysis revealed that exports are negatively affected by extreme temperature events. A high temperature episode in the exporting country reduces trade in the month of the event. But the global costs of such an event remain unclear since importers are able to either source goods from somewhere else or compensate for lost imports by purchasing more in later periods from the same source. We studied possible compensation across time using distributed lag models but did not find robust evidence for this (see Sect. 5.2). But importers could also seek for compensation across space. Although in our model imports of different origin are imperfect substitutes only, buyers can adjust the source of their imports and, at least partially, recoup losses on one trade link by additional imports from other sources. This demand adjustment has, of course, repercussions on relative prices and available income, changing the world trade equilibrium. Therefore, we compute the new global trade equilibrium resulting from a temperature shock, the global costs, and their distribution on directly and indirectly affected countries. Finally, we approach the question how these global costs will evolve under climate change. 6.1 Methodology As discussed above, and as shown by Fally (2015), we are able to retrieve from our PPML fixed effects estimation of (10) the underlying structural Gravity model, described in equations (5)– (9). See Appendix 8.1. for the derivation of the model from the fixed effect estimation. However, there are two parameters which we do not observe. One is the preference parameter λi , which informs the pricing equation (8). But we can rewrite equation (8) such that it provides the price change relatively to the estimated baseline which is independent of this structural parameter. ∆ pit = ( Xit ˆ Xt ˆ XitXt) 1 1−σ ˆ Πit Πit , where the hat denotes estimated parameters, i.e. the predicted exports and the estimated outward multilateral resistance terms. The second not retrievable key parameter is the elasticity of substitution between varieties of different origin, σ . We take this from the literature and inform our model by the elasticity estimate of Simonovska and Waugh (2014), who estimate it using disaggregate price and trade-flow data. Their estimate yields an elasticity of roughly four. However, we are going to conduct extensive sensitivity analyses and test the robustness of our findings with respect to σ . Obviously, the ease of substitution from one good to another substantially influences the magnitude of the effects. Equipped with our estimated model and the additional parameter, we run counterfactual simulations, computing the hypothetical trade equilibrium in absence of a high absolute temperature event. This allows to assess the full international trade costs of a temperature shock, accounting for equilibrium adjustments in the trade network. But it needs to be pointed out that we only observe international trade flows, omitting domestic impacts on production and income as well as domestic adjustment effects. For instance, a temperature 1 3 984 International Trade and the Transmission of Temperature Shocks shock might induce a reallocation of resources within the country and subsequently also affect imports and exports. Given the rather short term nature of the temperature shock, substantial equilibrium adjustments of domestic resources seems rather improbable. However, without observing the full domestic impact in the affected country, we are also not able to derive comprehensive welfare effects of temperature shocks from our simulations. Rather, our calculations provide the costs of temperature shocks in terms of lost trade.11 To administer the counterfactual simulations and compute the new equilibria we need to reduce the model dimensionality in terms of spatial and temporal coverage. First, we reduce the number of countries incorporated in the analysis. This more parsimonious model covers the bilateral monthly trade of 43 countries (listed in Table 8 in the Appendix). The trade within this sub-sample makes up roughly 90 percent of the global trade volume. With this reduced data set we bootstrap 100 times the estimation of the structural Gravity equation (11) with specific exporter × year, importer × year, and country-pair × year fixed effects. The mean coefficient of the weather shock Dt30 itm on exports estimated with this smaller data set spanning 43 countries only is almost identical to the point estimator based on the full data set in our preferred specification reported in column (1) of Table 1. In the next step, we select the year 2015 — the year where we have the best data coverage of monthly bilateral trade — for our assessment.12 The bootstrapped fixed effects for 2015 plus the estimated temperature shock coefficient calibrate our CES-Armington monthly-trade model described in section 2. As these estimated coefficients, jointly with our assumed trade elasticity, fully and consistently specify our trade model, we are ready to conduct counterfactual analyses. For each set of bootstrapped estimates, we randomly draw an observed high temperature event in 2015, and compute the counterfactual global trade equilibrium assuming this event would not have happened. We report the mean loss of imports as well as the 95 percent-confidence interval for the country experiencing the temperature shock (directly affected) and all other countries (indirectly affected). 6.2 Global Trade-loss Incidence of Local Temperature Shocks Equipped with our estimated structural Gravity model, we aim at answering how a single high temperature event changes the world trade equilibrium and impacts directly and indirectly affected countries. As a first approximation for computing the global international trade losses of the average temperature shock, an analyst could just sum up the reduction of imports from the directly affected exporter across all importers, ignoring substitution and income effects. This is shown in the left bar in the left panel of Figure 5. This "naïve" approach leads to aggregated international trade losses of 2015-USD 454 million for the average high temperature event. However, as the right bar in the same panel of Figure 5 shows, when equilibrium adjustments are taken into account, importers are able to partially 11 Note that in welfare terms — ignoring any frictions and rigidities — the maximum costs at stake for the importing country are the total welfare gains from trade with the exporting country exposed to the weather event. 12 The global climate in 2015 has been characterized by an El Niño situation — a phase of the El NiñoSouthern Oscillation (ENSO) climate phenomena that influences sea temperatures and weather in large parts of the globe (Blunden and Arndt 2016). As a consequence, 2015 saw heatwaves in France, high temperatures and drought in South America, in particular in Argentina, Brazil, and Colombia, as well as severe drought in South Africa. 1 3 985 O. Schenker, D. Osberghaus substitute their purchases from the directly exposed country to other sources, and global trade losses decrease to 360 million USD, a reduction of about 20 percent relative to an assessment that ignores equilibrium effects. This is a clear indication of international trade’s potential to distribute and reduce costs of such shocks through equilibrium adjustments and substitution. This total amount of import losses appears in directly and indirectly affected countries. Indirectly affected importers can only partially substitute their import losses via alternative sources. In this sense, international trade transmits a share of the costs of the temperature event. A cost assessment of these events not including these cross-border effects is therefore incomprehensive. The right panel of Figure 5 shows this. Of the 360 million USD total international trade losses appear 136 million USD in the directly affected exporting country due to losses in purchasing power caused by the temperature-induced export reduction. This means that the group of countries that have not been directly experienced the temperature event bear import losses of 224 million USD from the average single temperature shock. This are average costs of about 5.3 million USD per temperature shock for an indirectly affected country via these trade spillovers. While this effect is not huge, our bootstrapped confidence interval suggests that these costs are statistically significant different from zero given a five percent significance level. But these indirect costs are distributed unevenly across countries. As Gravity theory tells us, absolute indirect costs of temperature shocks are governed by the total value of imports and therefore depend on the size of the importing country, as well as the trade costs of shipping a good to the importing country. Thus, larger importing countries and countries with lower costs of trade with the directly affected country have to burden larger absolute costs. Figure 6 ranks the average indirect costs in a decreasing order. While the country at the 25th percentile faces costs of about 8.6 million USD, the country at the 75th percentile faces costs of 2.3 million USD. 6.3 Sensitivity Analysis The magnitude of the general equilibrium effects is fully specified by the estimates of the underlying Gravity model with one degree of freedom: The estimation of the Gravity system Fig. 5 General equilibrium trade losses of average high temperature event. The left panel shows the aggregated global trade loss (summing up directly and indirectly affected countries) of the average high temperature event (monthly mean temperature at least 30 ◦C ). The left bar depicts the mean estimate of lost imports before equilibrium adjustments (naïve). The right bar shows estimated lost imports after the equilibrium-adjustment (equil.). The right panel splits the aggregated losses into average losses for the directly affected country and average aggregated losses for the indirectly affected countries. Bootstrapped 95 percent confidence interval 1 3 986 International Trade and the Transmission of Temperature Shocks does not identify the elasticity of substitution σ which governs the ease to adjust demand and source goods from somewhere else. As discussed above, in our central estimates we set σ=4 , following Simonovska and Waugh (2014). However, it is key to understand how sensitive our results are depending on the choice of σ . Figure 7 plots the effect of the mean high temperature event for values of σ ranging from 2.5 to 5.5. The right panel shows the aggregate import losses for only indirectly affected countries. Assuming an elasticity of substitution of 2.5 leads on average to an import loss of about 520 million USD. These losses decrease with a substitution elasticity of 5.5 to about 124 million USD for all importing countries. It is also important to note that the point estimator of the mean temperature event on the average indirectly affected country, although being small (2.9 million USD at σ=5.5 ), is statistically different from zero with 95 percent confidence over the whole range of the tested elasticities. As theory suggests, the costs are Fig. 7 Sensitivity analysis–elasticity of substitution σ . Sensitivity analysis of general equilibrium effects conditional on elasticity of substitution σ ranging from 2.5 to 5.5. The left panel shows the trade losses of the mean >30 ◦C event on directly affected countries. The right panel shows the losses on indirectly affected countries. Confidence intervals are bootstrapped Fig. 6 Distribution of indirect costs. Distribution of average costs of the top-15 indirectly affected countries in decreasing order. Bootstrapped 95 percent confidence interval 1 3 987 O. Schenker, D. Osberghaus lower for higher elasticities as costs can be more easily mitigated by adjusting the sourcing of goods. The left panel of Fig. 7 shows the effects for directly affected countries. If cost-sharing is limited under a low elasticity of substitution of 2.5, the average directly affected country faces mean losses 220 million USD. Assuming an elasticity of substitution of 5.5, these losses decrease to about 100 million USD. 6.4 Costs Under Future Climate Projections So far, our estimated structural Gravity model has been applied to compute ex-post general equilibrium effects of high temperature events. However, the model allows also to compute the impacts on international trade — of both, directly and indirectly affected countries— under future climate projections. We thus compute the counterfactual impact a projection of the future climate would have on our estimated world economy of 2015 and compare it with the historic climate means. There is therefore an important caveat: Our simulations assume that the future climate happens in the world economy of 2015, ignoring any socioeconomic adjustments and changes that will happen. But trade costs, total expenditures but also the vulnerability to high temperature might change substantially in future.13 We therefore believe that this is a sensible approach, and are fully aware of its limits. We use data provided by World Meteorological Organization in the KNMI Climate Change Atlas. This database provides modelled mean temperatures at the country-month level, both for the past and for projections up to the year 2100. We rely on the median output of the multi-model ensemble CMIP5 (Coupled Model Inter-comparison Project Phase 5 used in the 5th IPCC Assessment Report), and focus on projections based on the Representative Concentration Pathway (RCP) 4.5. RCP4.5 is a rather "optimistic" emission scenario which assumes that global emissions peek before 2050, and radiative forcing stabilizes by 2080–2100. However, due to the inertia of the climate system, projections for the near future do not vary substantially across RCPs. In our analysis, we examine the consequences of the average climate of 2020–2039 due to the reasons discussed above. We compare this period to the latest available historic 20-year period in the data, which is 1980–1999. Thus, the differences to the historic climate are not substantial. For both 20-year periods, we compute for each country and calendarmonth the probability that the monthly mean temperature exceeds 30◦C . We then compute the difference in annual import expenditures, comparing a scenario with the historical distribution of high temperature events to a scenario with the expected distribution of heat events given the climate projection for 2020–2039. We express these changes in percent of the historical baseline. While the differences depicted in Fig. 8 are small, we observe additional trade losses in all countries that are part of the simulated world trade equilibrium. While these additional losses are small in Europe, they become substantially higher in Asia (in particular India) and Oceania, as well as in South Africa. If we sum this up across countries, we find that annual global trade is reduced by 735 million USD due to additional months with high temperatures in a projection of the average 13 An important factor is future access to air conditioning. Davis and Gertler (2015) show that with rising income, demand for cooling will increase substantially in many of today’s middle income and developing countries. 1 3 988 International Trade and the Transmission of Temperature Shocks that the effect is on the intensive margin, as there is no significant effect on the binary variable indicating positive trade (Models 3 and 4), but a substantial effect on the trade value in the sub-sample of positive exports (Models 5 and 6). Note that the number of observations in the Models 3 to 6 is lower than in the Models 1 and 2, since we can only safely identify zero trade flows since the year 2000. For assessing the effects of high temperatures in differently exposed countries, we divide the sample into ’cold’ and ’warm’ exporters, depending on their maximum temperature throughout the analysis period. ’Warm’ countries exhibit maximum temperatures above the median of the maximum temperature distribution. Figure 13 presents the results in the form of 5-degree temperature bins. Similar to the proceeding analyses, these results suggest that Fig. 12 Contemporaneous effects of temperature on exports in 5 ◦C bins. Estimated coefficients for temperature bins in the exporting country (left panel) and the importing country (right panel) and their 95 percent confidence intervals, using 15–20 ◦C as the baseline category. Based on a PPML regression with fixed effects at the country-pair-year-level, exporter-calendar month-level, and importer-calendar monthlevel (N=3,821,155) Fig. 11 Contemporaneous effects of temperature on exports in 1◦C steps, unrestricted plot. Estimated coefficients for each ◦C in the exporting country (left panel) and the importing country (right panel) and their 95 percent confidence intervals, using 20 ◦C as the baseline category. Based on a PPML regression with fixed effects at the country-pair-year-level, exporter-calendar month-level, and importer-calendar month-level (N=3,821,155) 1 3 995 O. Schenker, D. Osberghaus very high temperatures are the main driver of the estimated negative effect. Moderately high temperatures in absolute terms, even if abnormally high for the exposed locations, do not show a similar effect on exports - neither in ’warm’ nor in ’cold’ countries. Figure 14 illustrates the lagged impacts of a temperature shock at the importer’s location, estimated as covariates in the context of assessing lagged impacts on exports (presented in Fig. 3). The estimated coefficients are not statistically different from zero throughout the first year after the temperature shock. Table 4 Linear and quadratic temperature effects on exports, estimates of the extensive and intensive margin (1) (2) (3) (4) (5) (6) Model 1 Model 2 Model 3 Model 4 Model 5 Model 6 tempitm 0.0013 0.0022* (0.0010) (0.0011) temp2 itm −0.0001** (0.0000) Dt30 itm 0.0017 −0.0398*** (0.0035) (0.0125) Dp99 itm 0.0009 −0.0235*** (0.0013) (0.0090) tempjtm −0.0005 −0.0002 (0.0008) (0.0009) temp2 jtm −0.0000 (0.0000) Dt30 jtm 0.0024 −0.0239*** (0.0016) (0.0073) Dp99 jtm −0.0020** −0.0194** (0.0009) (0.0086) Observations 3821155 3821155 2698697 2698697 1725813 1725813 Model 1 and 2: PPML estimations of bilateral exports in current USD, including zero exports. Model 3 and 4: Linear probability estimations of a binary variable indicating the existence of positive exports (extensive margin). Model 5 and 6: PPML estimations of bilateral exports in current USD, excluding zero exports (intensive margin). All regressions are with fixed effects at the country-pair-year-level, exporter-calendar month-level, and importer-calendar month-level. * p<.1 ** p<.05 *** p<.01 (1) (2) (3) Clustering 1 Clustering 2 Clustering 3 Dt30 itm −0.0336*** −0.0337*** −0.0337*** (0.0100) (0.0129) (0.0100) Dt30 jtm −0.0124** −0.0125** −0.0125** (0.0056) (0.0062) (0.0055) Observations 3821155 3776991 3776991 Column 1: Clustered by importer, exporter, and time step (replication of baseline specification). Column 2: Clustered by spatial distance between importer and exporter. Column 3: Clustered by all four variables. * p<.1 ** p<.05 *** p<.01 . The number of observations is slightly lower in column 2 and 3 due to missing data for distance Table 3 Contemporaneous effects of temperature with differently clustered error structures 1 3 996 International Trade and the Transmission of Temperature Shocks In Fig. 15, we depict the cumulative effects of Dt30 itm during twelve months after the temperature shock. The estimation confirms the negative effect on exports in the first months after the temperature shock, and shows that these export values are not recovered in subsequent periods. There is no significant impact on imports. Regarding relatively warm temperature events ( Dp99 itm ), we depict the lagged and cumulative effects on exports in Figs. 16 and 17, respectively. As for the case of absolute hot temperatures, the effect is concentrated on exporters, and short-lived. However, in contrast to the Dt30 itm -specification, the effect is not aggravating in the first few months but there is a slight tendency towards catching-up, such that the cumulative effect within one year after the event is statistically equal to zero. In Fig. 18, we replicate the interaction analysis for the potential labour intensity channel for temperature events in the top percentile ( Dp99 itm ). In this specification, the effect is not sigFig. 14 Lagged impact on imports of an average monthly temperature of at least 30 ◦C . Estimated effects on imports of a hot month on exports, including 95-percent confidence intervals. The effects are relative to a month with a temperature below 30 ◦C . The model is estimated with PPML and include temperature at the exporter location, country-pair-year, exporter-calendar month, and importer-calendar month fixed effects. Standard errors are multi-way-clustered at exporter, importer and time step-level Fig. 13 Contemporaneous effects of temperature on exports in 5 ◦C bins in ’cold’ and ’warm’ countries. Estimated coefficients for temperature bins in ’cold’ exporting countries (left panel) and ’warm’ exporting countries (right panel) and their 95 percent confidence intervals, using 15-20 ◦C as the baseline category. Based on a PPML estimation with fixed effects at the country-pair-year-level, exporter-calendar monthlevel, and importer-calendar month-level (N=1,956,682 for ’cold’ subsample and 1,864,473 for ’warm’ subsample 1 3 997 O. Schenker, D. Osberghaus nificant which may be rationalized by the insights of prior studies that absolute temperature levels are more important for effects on labour capacity. Figures 19 and 20 summarize the estimates of various interaction effects. Most of the analyzed potential heterogeneity variables show no significant impact on the estimated effects of Dt30 itm or Dp99 itm . Figure 21 analyses heterogeneous effects over time. We interact Dt30 itm with time bins of five year length in order to check if the effect of temperature shocks changes over time. We find statistically significant effects of high temperature shocks on exports in early years of the sample but have to interpret them with a grain of salt as the sample is small and biased towards a small number of developed countries. In later periods, when the sample becomes Fig. 16 Lagged effects of an average monthly temperature in the top percentile. Estimated effects of a hot month on exports (left panel) and imports (right panel), including 95-percent confidence intervals. The effects are relative to a month with non-extreme temperature. The model is estimated using PPML and include country-pair-year, exporter-calendar month, and importer-calendar month fixed effects. Standard errors are multi-way-clustered at exporter, importer and time step-level Fig. 15 Cumulative effects of an average monthly temperature of at least 30 ◦C . Estimated cumulative effects of a hot month on exports (left panel) and imports (right panel), including 95-percent confidence intervals. The effects are relative to a month with a temperature below 30 ◦C . The model is estimated with PPML and include country-pair-year, exporter-calendar month, and importer-calendar month fixed effects 1 3 998 International Trade and the Transmission of Temperature Shocks more comprehensive, we find mostly negative effects of temperature shocks on exports but not all of them are statistically significant. Effects of Precipitation and Storms As for the case of temperature, our main source of historical precipitation data is CCKP (World Bank 2022a). As for the case of temperature, the values are monthly means aggregated at the country level. In Table 5, we present the descriptive statistics of the additional weather variables used in this and the subsequent section. For effects of storms, we use data on the maximum wind speed at the country-month level from the ifo GAME data set (Felbermayr and Gröschl 2014), which is based on two primary data sources: First, it uses the International Best Track Archive for Climate StewFig. 18 Estimated effect of an average monthly temperature in the top percentile on exports for different levels of labour intensity. Estimated contemporaneous effects of Dp99 itm on exports for given levels of labour intensity, including 95-percent confidence intervals. The effects are relative to a month with non-extreme temperature. The model is estimated with PPML and includes country-pair-year, exporterand importer-calendar month fixed effects. Labelled values at the x-axis are the 5th, 50th, and 95th percentile of labourintijt Fig. 17 Cumulative effects of an average monthly temperature in the top percentile. Estimated cumulative effects of a hot month on exports (left panel) and imports (right panel), including 95-percent confidence intervals. The effects are relative to a month with non-extreme temperature. The model is estimated with PPML and include country-pair-year, exporter-calendar month, and importer-calendar month fixed effects. Standard errors are multi-way-clustered at exporter, importer and time step-level 1 3 999 O. Schenker, D. Osberghaus ardship (IBTrACS) which is provided by the National Climatic Data Center of the National Oceanic and Atmospheric Administration (NOAA) and contains data of individual hurricane events. Second, in order to capture tornadoes, summer and winter storms not included in IBTrACS, the hurricane data is matched to daily data of the Global Surface Summary of Day (GSOD) data (version 7) on maximum wind speed and wind gust from over 9,000 weather stations worldwide. Wind data from ifo GAME is available at the country-month level for the period 1979-2010. In Table 6 we study the effect of different models estimating the precipitation impacts on exports. In more detail, we examine the linear and quadratic effects of absolute precipitation (Models 1 and 2, respectively), and effects of extremely dry and wet months (Models 3 and 4). We do not find significant effects of precipitation in any of the specifications. In Model 5, we add extremely dry months as a control variable to our baseline estimates. The estimated effect of high temperature on exports remains almost identical but episodes of low precipitation have no significant effect. Model 6 shows that also the interaction of extremely dry months with very hot months does not have a significant effect on exports. Table 7 shows the results of similar estimations of wind speed impacts on exports. There are neither effects in the linear and quadratic specifications, nor do months with countryspecific high wind speeds show an effect on exports. Fig. 19 Estimated effects of an average monthly temperature of at least 30 ◦C for various heterogeneity variables. Estimated contemporaneous effects of Dt30 itm on exports for given levels of various heterogeneity variables, including 95-percent confidence intervals. The effects are relative to a month with a temperature below 30 ◦C . The model is estimated with PPML and includes country-pair-year, exporterand importer-calendar month fixed effects. Standard errors are multi-way-clustered at exporter, importer and time step-level. Labelled values at the x-axes are the 5th , 50th , and 95th percentiles of the heterogeneity variables. The reported p-values refer to the significance of the interaction term (n.s.: p ≥ 0.1) 1 3 1000 International Trade and the Transmission of Temperature Shocks However, using country-specific distributions may not be the appropriate strategy for identifying non-linear effects of storms, since the occurrence of harmful wind speeds is not equally distributed on countries (as the variable combiP99), but is mainly an issue in tropical cyclone-exposed locations. Therefore, we use a similar strategy as for temperature bins (Fig. 12), and estimate effects of absolute wind speed bins in Fig. 22, using the bin with the largest number of observations (40-45 knots) as the baseline category. The results suggest that for wind speeds above 140 knots, exports decrease substantially in the month of the weather event, while imports remain largely unaffected. These potential short-term and nonlinear effects of intense storms on exports are beyond the scope of this analysis, and may be a promising avenue for further research. Additional Information on Counterfactual Simulations Table 8 presents the list of countries included in the simulation analysis. Fig. 20 Estimated effects of an average monthly temperature in the top percentile for various heterogeneity variables. Estimated contemporaneous effects of Dp99 itm on exports for given levels of various heterogeneity variables, including 95-percent confidence intervals. The effects are relative to a month with non-extreme temperature. The model is estimated with PPML and includes country-pair-year, exporterand importer-calendar month fixed effects. Standard errors are three-way-clustered at exporter, importer and time step-level. Labelled values at the x-axes are the 5th , 50th , and 95th percentiles of the heterogeneity variables. The reported p-values refer to the significance of the interaction term (n.s.: p ≥ 0.1) 1 3 1001 O. Schenker, D. Osberghaus Table 5 Descriptive statistics of additional weather variables Variable Description Mean Std. dev. Minimum Maximum Obs. precitm Mean precipitation in mm, average of country area 94.67 93.57 0.00 1,063.69 49,597 precP 1itm Mean precipitation in lowest countryspecific percentile 0.015 0.121 0.00 1.00 49,597 precP 99itm Mean precipitation in country-specific top percentile 0.010 0.099 0.00 1.00 49,597 winditm Maximum wind speed in knots 48.58 19.12 0.00 165.00 26,559 windP 99itm Maximum wind speed in countryspecific top percentile 0.007 0.082 0.00 1.00 26,559 Descriptive statistics are calculated at the country-month-level. For reasons of brevity, country-specific statistics are only reported for exporters Fig. 21 Estimated effects of average monthly temperature of at least 30 ◦C over time. Estimated effects on imports of a hot month on exports for given time periods, including 95-percent confidence intervals. Based on an interaction of the temperature variable with 5-year-dummy variables. The effects are relative to a month with a temperature below 30 ◦C . The model is estimated with PPML and include temperature at the importer location, country-pair-year, exporter-calendar month, and importer-calendar month fixed effects. Standard errors are three-way-clustered at exporter, importer and time step-level 1 3 1002 International Trade and the Transmission of Temperature Shocks Table 6 Contemporaneous precipitation effects on exports (1) (2) (3) (4) (5) (6) Model 1 Model 2 Model 3 Model 4 Model 5 Model 6 Dt30 itm −0.0335*** −0.0331*** (0.0100) (0.0102) Dt30 itm ×precP 1itm −0.0719 (0.0678) precitm −0.0000 −0.0000 (0.0000) (0.0001) prec2 itm −0.0000 (0.0000) precP 1itm 0.0026 0.0026 0.0026 (0.0190) (0.0190) (0.0190) precP 99itm 0.0019 (0.0080) Observations 3821155 3821155 3821155 3821155 3821155 3821155 All models are estimated using PPML and include country-pair-year, exporter-calendar month, and importer-calendar month fixed effects. In all estimates we control for weather effects in the importing country (1) (2) (3) Model 1 Model 2 Model 3 windi −0.0001 0.0003 (0.0001) (0.0005) wind2 i -0.0000 (0.0000) windP 99i 0.0003 (0.0085) Observations 2155697 2155697 2155697 Table 7 Contemporaneous wind speed effects on exports All models are estimated using PPML and include country-pairyear, exporter-calendar month, and importer-calendar month fixed effects. In all estimates we control for weather effects in the importing country 1 3 1003 O. Schenker, D. Osberghaus Argentina Australia Austria Belgium Bangladesh Brazil Canada Switzerland China Colombia Czechia Germany Denmark Algeria Spain France Great Britain Hungary Indonesia India Ireland Iraq Italy Japan South Korea Mexico Malaysia Netherlands Norway New Zealand Oman Philippines Poland Russia Saudi Arabia Singapore Sweden Thailand Turkey United States Venezuela Viet Nam South Africa Table 8 List of countries in used in simulations List of countries used in counterfactual simulations. Trade between these countries covers 90 percent of the trade volume over the sample period Fig. 22 Contemporaneous effects of maximum wind speed on exports in 5 knots bins. Estimated coefficients for wind speed bins in the exporting country (left panel) and the importing country (right panel) and their 95 percent confidence intervals, using 40-45 knots as the baseline category. Based on a PPML regression with fixed effects at the country-pair-year-level, exporter-calendar month-level, and importer-calendar month-level (N=2,153,305). Standard errors are clustered at exporter, importer and time step-level 1 3 1004