The trade impact of the COVID‐19 pandemic (2024)

3.1. Data

We use China's monthly export data at the HS 8‐digit level over January 2019‐December 2020, obtained from China Customs Statistics. Our dataset covers all of the 242 destination countries or regions that China exported to in 2019 and 2020. Starting in 2020, China Customs report the combined January and February data, rather than for each individually. Thus, we also combine January and February's data of 2019 and consider it as one month. We aggregate our data to month–destination–product (HS6) level and control for China's specific factors over a year with month fixed effects. We carry out the analysis at the HS 6‐digit level, instead of doing it at the 8‐digit level, for two reasons. First, product classifications at higher than 6‐digit levels are not internationally comparable. Second, some variables that we use in our analysis are also defined at the HS 6‐digit level. At this level of aggregation, the number of observations in our analysis is already close to 2million. We describe below our data sources and variable definitions. For easier reference, we also summarise in the Appendix, Table ​TableA1,A1, the definitions and data sources of all variables used in the paper.

The Oxford COVID‐19 Government Response Tracker (OxCGRT), compiled by Hale et al. (2021), systematically collects publicly available information on many COVID incidences (cumulative cases, death cases and tests) and policy indicators.⁹ There are nine indicators, which record information on containment and closure policies, such as school and workplace closures and restrictions in internal and international movement.¹⁰ Based on these indicators, a stringency index is constructed to measure the strictness of ‘lockdown style’ policies that primarily restrict individuals’ behaviour. The original stringency index ranges from 0 to 100. We rescale it by dividing the original index by 100 to help the interpretation of estimated coefficients in the regressions. The higher the lockdown stringency index (Stringency) is, the more restrictions to individuals and to economic activities the country has.¹¹

Among all COVID‐19 incidence measures collected by OxCGRT, the death‐related measures seem the most reliable and internationally comparable both across countries and within countries over time. Other measures, such as number of tests and number of positive cases, are highly dependent on a country's testing capacity and reporting methods. More importantly, such capacity changed significantly within countries during 2020. Accordingly, we use the measures of new deaths per thousand people as our proxy for COVID‐19 incidence (CovidD). The original COVID‐19 data from OxCGRT are available at daily frequency. The number of new deaths is smoothed over the last seven days to fill gaps when data are missing for a day. We aggregate the data to the monthly level.¹²

As explained in the previous section, we also consider the pandemic variables of a country's trading partners. Specifically, country i's Stringency_ROW measure is the average stringency of the rest of the world in month t, weighted by country i's import share of product p in 2018 from all countries except China.¹³ That is,

where Mijp,2018 denotes the value of imports of product p by country i from country j in 2018. The import weights make lockdowns in the country's main trading partners more important. We use an analogous procedure to construct CovidD_ROW. Both variables are defined to vary at the HS6‐country–month level. The trade value data of 2018 used as weights are from BACI‐CEPII. Because our regressions cover years 2019 and 2020, using the pre‐determined trade data in 2018 as weights avoid potential endogeneity problems of bilateral trade.

In our heterogeneous analysis, we use several additional variables defined at the product or the country level. We explain below how each of them is defined and their data sources.

Since we study a health shock, we give special attention to medical goods (MG) in our analysis. The list of COVID‐related medical goods is provided by the World Customs Organization (WCO), together with the World Health Organization (WHO). We use its 3.01 Edition.¹⁴

As Bems et al. (2013) show in their review, during the ‘trade collapse’ of 2008–2009, the largest group of goods affected were durable consumption goods. Accordingly, we investigate whether a similar conclusion applies to the current situation as well. We define consumer durable and non‐durable goods based on the UN Broad Economic Category (BEC) classification (5th revision).

Also because of the nature of the shock, activities that can be performed from home are affected very differently than those that require physical presence. To make that distinction, we use the work‐from‐home shares from Dingel and Neiman (2020) and Bonadio et al. (2021). Dingel and Neiman (2020) calculate the work‐from‐home share at the occupational level. Based on it, Bonadio et al. (2021) compute the sectoral work‐from‐home intensity measure from the average of Dingel and Neiman (2020)'s, weighted using sectoral level expenditure shares on each occupation. We use the concordance between ISIC Rev.4 and ISIC Rev.3 and the concordance between HS 2017 and ISIC Rev. 3 to calculate the HS6level work‐from‐home intensity following Bonadio et al. (2021).¹⁵

Another product‐level feature that may lead to heterogeneous effects of COVID‐19 on trade is the level of contract intensity, which determines how much relationship‐specific capital is required to establish a trading relationship. We use the measure of contract intensity constructed by Nunn (2007), which corresponds to the share of intermediate inputs that require relationship‐specific investment. We convert Nunn (2007)'s original data at the 3‐digit level of ISIC Rev. 2 to the HS 2017 at the 6‐digit level.

The trade modes in which products are exported could also yield different responses from the pandemic, because the level of relationships may vary depending on the trade regime. Our data from China Customs Statistics provide information on trade regimes, mainly processing trade and normal trade.¹⁶ Thus, we create the share of processing trade among processing and normal trade at the HS6 product level to investigate whether products exported in that way are affected differently. This share varies at the country–product–month level.

The effects of the pandemic may also vary depending on the position of products along global value chains, because the pandemic is likely to affect firms and families differently. To investigate that possibility, we use the UN BEC 5th Revision data to distinguish goods between capital goods, intermediate goods and final goods for consumption, evaluating them separately.¹⁷

In addition to the COVID‐related deaths and the stringency index, we also take advantage of data collected by OxCGRT on worldwide government responses to the pandemic in the form of economic support policies. The economic support index records measures such as income support to those who lose jobs or cannot work, as well as debt relief.¹⁸ The original economic support index is at a daily frequency. We take a simple average to calculate the index at the monthly level, which ranges from zero to about 100 in our final dataset. We rescale it to between 0 and 1 by dividing the original index by 100, similar to the rescaling of the stringency index.

Some countries have implemented trade policies in response to the pandemic. We use the measures compiled by the WTO on temporary COVID‐19 related trade policies to investigate whether they affect our estimates.¹⁹ Since our analysis is on the pandemic effects on countries’ imports, we only consider import measures, most of which aim to promote imports of selected medical products or materials. For a few cases when a policy applied to all medical goods, we use the WCO‐WHO definition of medical goods discussed earlier. We consider the temporary nature of these policies based on their initiation and revocation dates. The trade policy indicators vary by country–product–month.

Finally, we include the exchange rate in one specification to further control for possible unobserved heterogeneity at the monthly level. We obtain the monthly exchange rate data from the CEIC database. It measures the amount of a currency per USD. A higher exchange rate means a weaker currency relative to the USD.

Table ​TableA2A2 in the Appendix provides summary statistics for the variables used in our analysis. The data used in our baseline regressions cover a large number of countries/regions (174) and products at the HS6level (4636) over 12months of 2019 and 2020 (January and February combined).²⁰ Naturally, the COVID‐related variables reflect only the information for 2020, since such data did not exist before 2020. The table shows that there is substantial variation in both our dependent and independent variables, including those that we use for the heterogeneity analysis.

Table ​TableA3A3 in the Appendix shows the pairwise correlations among the four main independent variables. The correlation coefficients among them are all positive, as expected, but are below 0.5 for most of them, except for Stringency and Stringency_ROW, which have a correlation coefficient of 0.6. This indicates that multicollinearity should not be a major concern in our analysis.

3.2. Econometric specification

We estimate how deaths from COVID‐19 and the stringency of lockdowns affect countries’ imports from China using monthly trade data at the HS 6‐digit product level. We consider a country's own COVID‐19 variables and those in its trading partners except China (ROW). A natural empirical specification for such an analysis is the standard log‐linear gravity regression, as follows:

The dependent variable is the log of the volume of imports of product p by country i from China in period t, referring to a specific month between January 2019 and December 2020. The main explanatory variables are the number of COVID‐related deaths per thousand people (CovidD) and a stringency index capturing the overall strictness of a country's policies to control the spread of the virus (Stringency). The ROW measures of Stringency and CovidD are the average stringency and COVID‐related deaths in ROW, weighted by each country's imports of a product from all countries except China in 2018. In some regressions, we also include interaction terms between COVID‐19 variables and other variables. Naturally, all COVID‐related variables are available only for the year 2020; for years before 2020, they are set to zero.

It is clear that COVID deaths and lockdown stringency are endogenous to each other—more deaths tend to induce harsher lockdowns, which in turn tend to reduce deaths. Estimates would therefore be biased if one included just one of them in the analysis. By including both, we capture the independent effect of each variable. It is also possible that other unobservables affect both monthly imports from China and lockdown stringency. However, the nature of those unobservables is not very clear, and perhaps for that that reason the literature on the topic has been silent about it (and about potential instrumental variables).

Instead, to control for other contemporaneous shocks and characteristics, we rely on a large set of fixed effects. We consider a particularly demanding set of fixed effects at the country–product–year level (a_ipy), which represent any factor that affects imports from China of a particular country–product pair in the same way over the months of a year. These effects capture differences in imports from China due to specific characteristics of the importing country, such as its size, and due to specific characteristics of the product, such as those that make it more or less appealing. They also capture similar effects at the country–product level—for example, factors that make a country have a particular large or small demand for imports from China of a specific product. Furthermore, they are allowed to vary by year.²¹ In turn, a_t refers to time (year‐month) fixed effects, which control for monthly level unobserved heterogeneity, including worldwide and Chinese‐specific macro and health factors, as well as seasonal elements. With this wide set of fixed effects, the variation that our coefficients capture comes only from within‐country or within country–product pairs over time.

Instead of estimating (2), we take the year‐over‐year (yoy) difference in trade and COVID‐related variables. This has the practical advantage of eliminating all time‐invariant idiosyncratic country–product effects. Country–product effects that vary at the year level are still present; however, and in the yoy specification, we denote them simply as α_ip. This gives our baseline specification:

The dependent variable, Δimportsipt, is the log difference between country i's imports of product p from China in month t of 2020 and its imports in the same month of 2019, that is Δimportsipt=100×[log(imports2020)ipt‐logimports2019)ipt. We multiply it by 100 to help with the visualisation of the estimated coefficients, so it measures the yoy change in trade value in percentage terms. The explanatory variables are exactly the same as in equation (2), since they all take value zero during the whole 2019. Parameter m_t corresponds to a month fixed effect. Thus, this chained log difference yoy estimation at the monthly level also removes seasonality and avoids potential issues arising from combining the trade data for January and February. Considering that the trade of a product may be correlated across countries and over time (such as trade in medical goods) or across country–product over time, we cluster the standard errors at the country–product level.²²

Compared with the traditional log‐linear gravity regressions, our method has three advantages. First, we can avoid estimating equation (2) with a very large number of fixed effects corresponding to the two sets of country–product fixed effects (one for each year).²³ Second, the yoy difference allows an apples‐to‐apples comparison of trade values, instead of comparing them month‐over‐month sequentially, where there may be large seasonal changes, especially since we have to combine January and February trade values. Even with the full sets of fixed effects, the traditional log‐linear gravity regression, based on demeaning the data along various dimensions, cannot fully address the country–product‐specific seasonality (e.g. the sudden drop in trade for most products during Chinese New Year except for holiday‐related goods). Third, with the yoy measure, results are formatted as percentages, making it straightforward to interpret the economic significance of the results.

4.1. Baseline results

In Table ​Table1,1, we consider our baseline specification, assessing the impact of our main variables on the monthly log difference in imports from China, defining goods at the 6‐digit level. In columns 1 and 2, we have only the Stringency and CovidD variables. In the first column, to have an initial picture of how they are correlated with the dependent variable, we add only month dummies. In the second column, we then add country dummies and product fixed effects (absorbed). Naturally, this increases the fit of the regression significantly. In both cases, the two variables display negative and statistically significant effects, indicating that pandemic's negative demand effect prevails over the negative supply effect. The absolute value of the coefficients drops considerably from column 1 to column 2, highlighting the importance of controlling for the additional fixed effects.

Now, just like in any other gravity estimation, it is important to control for changes in economic conditions in third countries. In our specific case, we are directly interested in understanding how the state of the pandemic in third countries affects trade flows between a country/region and China, as this can clarify the strength of diversion and complementary forces between China and other countries. To do so, we consider in our analysis COVID‐19 incidence and lockdown policies in the rest of the world (excluding China), with greater weights given by those variables in the country's main trading partners, as defined in equation (1). The results from adding those variables to our regression are reported in columns 3 and 4. In column 3, we have the same set of fixed effects as in column 2, to make it clear the sole effect of adding the third‐country variables. In terms of sign, the estimated coefficients on CovidD and on Stringency do not change with the introduction of third‐country variables, remaining negative and estimated very precisely. In column 4, we have instead country–product fixed effects, which we keep in all subsequent specifications. As the results reveal, moving to that stricter specification increases the absolute magnitudes of all estimated coefficients, indicating that some trends at the country–product level mask the effects of the pandemic‐related variables.

We find that lockdown stringency in a country's main trading partner does not have a significant effect on import growth, suggesting that either trade diversion and complementarity effects offset each other or neither is relevant. On the other hand, more COVID‐19 deaths in a country's main trading partners increases that country's imports from China. That effect is statistically significant and large, indicating strong trade diversion when economic conditions in a country's main trade partners are affected by the pandemic.

The three coefficients that are statistically significant are precisely estimated and are economically meaningful. For example, using the estimates from our baseline specification in column 4 of Table ​Table1,1, we find that moving from a situation of no restriction to economic activities (as in 2019) to the highest level of stringency observed in 2020 (for Honduras in April and May, and for Philippines in April) would yield a reduction in imports of 17.6%.²⁴ For COVID‐19 deaths, going from zero (as in 2019) to the highest level of COVID‐19 deaths per capita excluding micro‐states (0.63 per thousand people in Slovenia in December) would induce a reduction in imports from China of 12%.²⁵ The effect stemming from COVID‐19 incidence in third countries goes in the opposite direction. It implies an increase in imports from China of 16.4% if the country's main trading partners experience an increase in COVID‐19 deaths from 0 to 0.628 per thousand people, the highest value in our sample (for Croatia in December).

In Table ​TableA4,A4, in the Appendix, we show the individual impact of each of the three statistically significant coefficients following a one standard deviation increase, a move from zero to the sample mean and from zero to sample maximum. In the three thought experiments, the positive effect stemming from COVID‐19 deaths in third countries prevails over the negative effect due to domestic COVID‐19 deaths. However, for non‐extreme values, the negative effect of a country's own lockdown on imports from China is the dominating force. Combining all of them, the net effect when each variable goes from zero to the sample mean is a decrease in imports from China of nearly 10 per cent.

4.2. Heterogeneity

Now, countries differ in how well they can absorb the consequences of the health crisis, and the effects likely differ by type of product as well. We explore various sources of heterogeneity, starting in Table ​Table2,2, where we look more closely at two types of goods.

The most obvious distinction in our setting is whether the good is helpful in containing the virus during the pandemic. Thus, we split the sample between medical goods (MG) and non‐MG. This group of products includes Personal Protective Equipment products, as well as many other goods, such as ventilators, test kits and syringes. As is well known, demand for some types of MG products skyrocketed at the onset of the pandemic and remains at historically high levels. However, while trade in MG goods is likely to display an idiosyncratic pattern, they correspond to only about 4% of the observations in our sample. Indeed, when we estimate our main specification without MG goods (column 1 of Table ​Table2),2), we obtain results that are very similar to our baseline regression. Now, if we do the opposite and restrict the sample to MG goods (column 2), we observe that the pattern is indeed very different. Neither lockdowns nor the number of COVID deaths has a significant effect on the imports of MG products from China, indicating that demand and supply effects offset each other. The impact of lockdown stringency in the main trading partners remains statistically not significant. Meanwhile, more COVID‐related deaths in the main trading partners increase imports from China, as it does with other goods, except that the magnitude of the effect is much larger for MG goods. Specifically, imports of MG from China would increase by 5.1% if the ROW experiences an increase in COVID‐19 deaths from zero to 0.051 (the sample mean) per thousand people.

We carry out a similar analysis for durable consumption goods. We are motivated by the previous findings of, for example, Bems et al. (2010) and Eaton et al. (2016), who have showed that the decrease in the demand for durable goods was responsible for a large part of the sharp decline in international trade flows in 2008–2009. Thus, in column 3 of Table ​Table2,2, we exclude durable goods from the analysis. The results are qualitatively similar to the baseline, but the magnitudes of the coefficients are between 12 and 33 per cent lower in absolute value. Unsurprisingly, then, we observe the opposite pattern when we restrict the analysis to durable goods. All coefficients that are statistically significant in the baseline remain so and maintain the same signs. However, they are between 35 and 92 per cent higher than in the baseline. Furthermore, more stringent lockdowns in the main trading partners cause a statistically significant reduction in imports of durable goods from China. These results show that, despite many differences in the nature of the current crisis and the one after the 2008 financial crisis, both show a particularly strong negative impact on imports of durable consumption goods.

In Table ​Table3,3, we consider other aspects of product‐level heterogeneity, related to how they are produced and traded. An important element mediating the impact of the pandemic in an economy is whether activities can be performed remotely. In particular, in countries where a large share of the population can work from home, both the direct effect and the government‐mandated indirect effect of the pandemic can be absorbed more smoothly than in countries where most activities require workers to leave home to carry out their jobs. Since both the negative demand and the negative supply effects are mitigated when working from home is feasible, the net result is a priori ambiguous. We assess whether that economic characteristic is quantitatively relevant for a country's imports by interacting a product's work‐from‐home share (wfh_sh) with our main variables of interest.²⁶

Table ​Table3,3, column 1, shows that the negative coefficients of both lockdown stringency and COVID‐related deaths increase in absolute value but are partially offset by products’ work‐from‐home share. Using the point estimates from column 1, they indicate that the effect of Stringency becomes positive for wfh>0.74, while the effect of CovidD becomes positive for wfh_sh>0.79. In our data, only 0.5% of the HS6 products have a value of wfh_sh above 0.74 (same for the cutoff value 0.79), and the median of our sample has wfh_sh=0.33. Still, the results reveal that having a substantial share of workers that can work from home can significantly dampen the negative impact of the pandemic on imports. Note that the estimated coefficients on the ROW variables hardly change when the interactions with wfh_sh are introduced.

In column 2, we turn to the contractibility of products. Products with high levels of contract intensity tend to be more heterogeneous, depend on long‐term arrangements and on relationship‐specific investments and be match‐specific. This makes it more difficult to switch suppliers, at least in the short run. We assess whether that economic characteristic is quantitatively relevant by interacting contract intensity at the product level with our main variables of interest.²⁷ The estimates for Stringency and for CovidD remain negative but increase in absolute terms relative to the baseline, especially the CovidD estimate. On the contrary, their interactions with contract intensity are positive and significant. This reveals that, for products with high contract intensity, for which it is costly to sever relationships, imports are more resilient to the impact of COVID‐19 deaths and to lockdowns, which have a temporary nature. This result resembles a key finding of Bas et al. (2022). Note that the estimated coefficients on the ROW variables remain virtually unchanged when the interactions with contract intensity are introduced.

Finally, in column 3, we consider potential differences in trade flows that are classified as ‘processing’. The estimates for the effects of our main variables remain very similar to the baseline results. The coefficient of the share of processing trade, which is defined at the country–product–month level, is positive, indicating that this type of trade has become more important compared with ‘normal’ (non‐processing) trade. Furthermore, its interaction with Stringency and CovidD is positive and statistically significant. This shows that, for products crossing borders with a greater share of processing trade, imports are more resistant to the negative impact of COVID‐19 deaths and of lockdowns. This could reflect the fact that processing trade usually involves closer relationships between domestic processing firms in China and their foreign partners.

Product heterogeneity may also depend on the position of a good in the production process. In Table ​Table4,4, we carry out that investigation by splitting products among consumption, intermediate and capital goods. Of the three categories, intermediate products behave more closely to the average product, which is partly explained by the fact that this group encompasses about half of the original sample. Stringency has a similar negative effect on the imports of each type of good, although higher for consumption and lower for capital goods. The sensitivity to own CovidD is substantially more diverse among the three groups: Imports of consumption goods are heavily affected by CovidD, whereas capital goods are virtually unaffected. This difference may be explained by the temporary nature of the negative income shock, which has a larger effect on family consumption, but little impact on firms’ long‐run investment plans. The effect of lockdowns in the main trading partners on each group of products is even more diverse. It is mute for intermediate products (like for the average product), negative and significant for consumption goods and positive and significant for capital goods. This suggests that, if it is difficult to import capital goods from the main trading partners due to lockdown policies there, firms turn to China to keep their investment plans.

Now, the level of development, and of the wealth, of a country is also likely to be an important mediator of how a country reacts to the pandemic. For example, we have seen that, on average, both CovidD and Stringency have a strong negative effect on countries’ imports from China. This is likely caused by the negative income effect of the pandemic, which tends to lower consumption, especially in nations where individuals are more credit constrained and have lower savings/wealth that could be used to smooth consumption. In contrast, in rich countries, the temporary loss of income, which is likely to have a small effect on lifetime income, may have a smaller impact due to better access to credit and to personal savings. To verify that possibility, in Table ​Table5,5, we split the sample between OECD countries (column 1) and non‐OECD countries (column 2). The sample without OECD countries yields results that are qualitatively similar to the baseline results in Table ​Table1.1. This is to be expected, since they comprise 69 per cent of the sample. In contrast, column 1 reveals that rich countries have reacted very differently to lockdowns with respect to their imports from China. Specifically, the effect of stricter lockdowns is positive for them, indicating that the negative demand effect is dwarfed by the negative supply effect, which induces local consumers and firms to turn to China to replace domestic goods when these cannot be produced locally due to restrictions on economic activities. The other coefficients are, however, relatively similar to those in the group of non‐OECD countries.

One possible reason for the differential effects of a country's own lockdown on its imports for developed and developing economies may be their fiscal redistributive policies during the pandemic. Generally, they have been significantly more generous in rich countries than in poor ones, not only in absolute terms but also as a percentage of GDP.²⁸ We consider that explicitly in column 3, where we introduce a measure of the value of a country's economic support to the population due to the pandemic. The coefficient on that variable is positive, as expected, and is statistically significant at the 10% level. Nevertheless, the estimates of the four main variables are hardly affected by the inclusion of that variable in the regression, indicating that the pandemic‐motivated fiscal policies increased demand for imports, but did so in a way that was largely unrelated to the specific effects of the pandemic on imports from China.

In column 4 of Table ​Table5,5, we do something different: eliminate the United States from the regression. A concern is that, because of the ongoing trade war between the United States and China, and because the United States is the largest importer of China, there could be confounding effects due to the extra tariffs on Chinese exports. We do not expect the trade war to play a central role in our estimations, since it took off in 2018 and between 2019 and 2020 there were relatively few policy changes (despite an ambitious but unenforced bilateral agreement to manage bilateral trade flows). The results in column 4 confirm our prior: The estimates are remarkably similar to those in the baseline estimation, indicating that including or not the United States in the estimations makes little difference, and therefore, the United States–China trade war is not a driver of our results.

4.3. Path‐dependence, extensive margin and robustness

In Table ​Table6,6, we consider five additional issues: path‐dependence, pandemic‐related trade policies, time‐varying sector level heterogeneity, control for exchange rates and the extensive margin of trade. Both the Stringency and the CovidD variables are serially autocorrelated. Thus, it is plausible that the trade effects of the pandemic in a month may be affected by the level of those variables in previous periods. In column 1 of Table ​Table6,6, we evaluate that possibility, adding the two variables measuring the cumulative sum of Stringency and COVID‐related deaths in the previous months of 2020. The results show that the estimates for the own variables remain similar to the baseline after the introduction of their previous values, as do the estimates for the ROW variables, with relatively small changes in the magnitudes of the estimated coefficients. In contrast, both previous Stringency and previous CovidD have a positive and statistically significant sign. This indicates that the pandemic introduces a component of intertemporal substitution in countries’ imports. A nation that restricts its imports in a month when it is badly hit will at least partially compensate for that reduction in the future. We note further that, relative to their contemporaneous effects, the magnitude of this intertemporal compensatory effect is greater for CovidD than for Stringency effects.²⁹

A few countries have implemented temporary trade policies in 2020 to cope with COVID‐19. Most of these measures sought to promote imports or restrict exports of medical goods or materials to meet the increasing domestic demand, with a few cases of import restriction measures. Omitting these policies may lead to biased estimates. Since we study countries’ imports, we consider only the import measures, as reported by the WTO. Specifically, in column 2 of Table ​Table6,6, we drop the observations affected by these policies (about 1% of the sample). The estimated coefficients are very similar to what we obtain from the baseline regression.³⁰

Recall that our empirical strategy, based on yoy changes and country–product fixed effects, eliminates country–product‐specific heterogeneity that is either fixed over time or varies annually. Now, there may also exist product‐specific short‐term shocks at the monthly level. To control for that, in column 3 of Table ​Table6,6, we add HS2‐month fixed effects along with the HS6‐country fixed effects. The signs and significance of the coefficients are the same as in our baseline regression. Except for the estimate for the effect of CovidD_ROW, which falls by about 30%, the magnitudes of the estimated coefficients hardly change either, indicating that those shocks are not central for our estimation.

There is also a concern about potential omitted variable bias in our regressions. To address this issue, in addition to the large set of fixed effects described above, we also consider cross‐country factors. Although most of the available data are reported at an annual frequency, one important trade‐related variable that is available for most countries in 2020 at a higher frequency is the exchange rate. We define it as the amount of a currency per USD. Since the Chinese Yuan is closely linked to the USD through a basket of key currencies, a higher exchage rate implies a stronger Yuan, which is likely to hurt imports from China. Given the multiple ways COVID affects the economy, the exchange rate may be related to the country's COVID variables, although it is not clear ex‐ante what the main channels would be. In column 4 of Table ​Table6,6, we report the results from adding the exchange rate variable to the benchmark specification. The coefficient of the exchange rate variable is negative and highly significant, as expected. However, its magnitude is very small. More importantly, the coefficients of the main variables bear the same signs and similar magnitudes as in our benchmark specification, except that the coefficient of Stringency in ROW is now statistically significant. Note that this regression is not directly comparable to the benchmark, because they use different samples: nearly a third of million observations are dropped due to missing monthly exchange rate data.

Now, since our dependent variable is the log difference in imports from China at the country–product level, we are effectively carrying out the analysis at the intensive margin; any country–product pair not observed in a month does not contribute to the estimations. We can also look at the extensive margin. A simple way to do that is to consider how the number of products imported from China by a country has changed from a given month in 2019 to the same month in 2020. More precisely, we compare the change in the number of HS6 product lines at the country–month level between the two years, so the dependent variable is also a yoy measure at the country–month level. The results are in column 5 of Table ​Table6.6. Interestingly, the effects are almost all mute.³¹ Although the sample is substantially smaller in this case, it still has almost 1800 observations. The absence of effects suggests that the impacts are mostly at the intensive rather than at the extensive margin. This is consistent with the findings about the nature of the great trade collapse following the financial crisis of 2008, which was also largely driven by the intensive margin of trade (Bems et al., 2013).

Finally, in Table ​Table7,7, we evaluate whether our results may be driven by outliers. The histogram of our dependent variable (Figure ​(FigureA1)A1) shows that its distribution is fairly ‘well‐behaved,’ indicating the presence of few outliers. However, Table ​TableA1A1 shows that the maximum and minimum of the distribution are indeed very large in absolute values. To check whether they are key to our findings, we run three different specifications. In column 1, we drop all observations where the absolute value of the dependent variable is above 500 per cent. This amounts to dropping approximately 2.6% of the observations. In column 2, we drop all the observations for micro‐states, defined as countries with population smaller than a half million in 2018. They correspond to 18 of the 174 countries in the baseline regression. Since those tiny countries may behave differently and be responsible for outliers both in the dependent variable and in some of the key independent variables—as in the case of San Marino, which has the highest value for CovidD in the sample—we want to investigate whether they are disproportionately affecting the results. Finally, in column 3, we run a median regression instead of an OLS. The results in column 1 are very similar to our baseline. The results in column 2 are almost identical to the baseline, showing that micro‐states are largely irrelevant for the estimates. Similarly, the median regression also shows results that are strikingly similar to our baseline, indicating that the latter is not highly affected by outliers.

FIGURE A1

Histogram of the dependent variable. Notes: This diagram shows a histogram of the dependent variable, defined as 100 times the log difference between countries’ imports from china in 2020 and that in 2019 at HS6 product level, monthly, i.e., Δimportipt=100∗log(import2020)ipt‐log(import2019)ipt

TABLE A1

Variable list

TABLE A2

Summary statistics

TABLE A3

Pairwise correlation among key COVID‐related variables

TABLE A4

Economic significance of the estimates

¹Source: https://hbs.unctad.org/total‐merchandise‐trade/.

²We also find that the effects are concentrated on the intensive margin. This pattern is similar to what has been found for the ‘great trade collapse’ that followed the financial crisis of 2008. For example, using Belgium firm‐level data, Behrens et al. (2013) find that nearly all of the fall in trade in that episode happened at the intensive margin. Bricongne et al. (2012) obtain similar results for French exporters. Bems et al. (2013) provide a review of the literature on the great trade collapse. This is not to say that the two shocks had similar global effects, however. As Le Moigne and Ossa (2021) point out, in aggregate, world trade has displayed much greater resilience in 2020 than during the trade collapse of 2008–2009.

³See Table ​TableA4A4 in https://www.wto.org/english/res_e/statis_e/wts2020_e/wts2020_e.pdf.

⁴Friedt and Zhang (2020) use China's exports data until May 2020 at the province–country–product (HS2) level and find that China's exports are very sensitive to foreign countries’ new cases and domestic new cases. Che et al. (2020) use China's export data until May 2020 at the country–product (HS6) level and find that China's exports decline when foreign cases increase. Pei et al. (2021) use data up to April 2020 for China's exports at the city‐country level and find that Chinese cities under lockdown experience a 34 percentage‐point reduction in year‐over‐year export growth rate.

⁵Like us, Bas et al. (2022) use both COVID‐19 deaths and lockdown stringency throughout their analysis. They use monthly import data for the United States, Japan, Germany and France until July 2020, and conclude that the negative trade impact of the pandemic on these four countries stems mainly from inputs whose supply relies on China and that require a high degree of customization.

⁶We know of two exceptions. Berthou and Stumpner (2022) use trade data from 31 reporting countries with the rest of the world until November 2020. They concentrate on the influence of lockdown policies, and in a robustness specification construct a similar measure for third‐country stringency, although not for COVID‐19 incidence variables. Espitia et al. (2022) also consider a third‐country effect, but in a very different way, based on changes in third countries’ industrial output, and without using COVID‐19 variables.

⁷We do not isolate the role of global value chains (GVCs) in shaping the impact of the epidemic on trade, but several other authors have done precisely that. Bonadio et al. (2021) model the pandemic‐induced lockdown as a labour supply shock, studying quantitatively its trade and welfare impacts through input‐output linkages that transmit the shock across countries through GVCs. Based on a quantitative Ricardian model including input‐output features, Eppinger et al. (2021) analyse the influence of GVCs in mediating countries’ exposure to foreign shocks. Sforza and Steininger (2020) calibrate a Ricardian model with production networks with data from the first quarter of 2020, showing that the transmission of the COVID‐19shock through production networks magnifies the impact of local supply disruptions. There are also authors who investigate the role of GVCs in the COVID‐19shock with reduced‐form approaches (Hayakawa and Mukunoki, 2021; Kejzar and Velic, 2020).

⁸An additional reason for the reduction in import demand is precaution. Given the uncertainty created by the pandemic, consumers may want to postpone consumption and firms may decide to postpone production and investment. This precautionary effect adds to the negative effect on imports.

⁹Sources: https://covid.ourworldindata.org/data/owid‐covid‐data.xlsx and https://github.com/OxCGRT/covid‐policy‐tracker

¹⁰Source: https://github.com/OxCGRT/covid‐policy‐tracker/blob/master/documentation/codebook.md

¹¹The lockdown policy indicators are updated on a bi‐weekly basis. The original stringency index is constructed based on nine sub‐indexes at daily frequency. The first eight relate to containment and closure: school closing, workplace closing, cancellation of public events, restrictions on gathering size, closure of public transport, stay‐at‐home requirements, restriction on internal movement and restrictions on international travel. The ninth relates to health systems, which is the public information campaign. For each subcategory, based on the judgement of the policy, a certain number of points is given, and then aggregated to generate the overall stringency index, which ranges from 0 to 100.

¹²Occasionally, a country may record a negative number for new COVID‐19 deaths to adjust for previous over‐reporting. Such negative values become very rare after we aggregate the data to a monthly frequency: they are present for only three country–month pairs among nearly two thousand pairs. We drop those three observations from our regressions.

¹³We exclude Hong Kong, SAR and Macau, SAR, because they are entrepots handling mainly the exports of mainland China.

¹⁴Source: http://www.wcoomd.org/‐/media/wco/public/es/pdf/topics/nomenclature/covid_19/hs‐classification‐reference_edition‐3_es.pdf?la=en.

¹⁵Concordance between ISIC Rev. 4 and ISIC Rev. 3 is from the WIOD SEA Source and Methods, https://www.rug.nl/ggdc/valuechain/wiod/wiod‐2016‐release, ‘WIOD Socio‐Economic Accounts 2016’, pp. 26–27. Concordance between HS 2007 and ISIC Rev. 3 are obtained from the WITS, https://wits.worldbank.org/product_concordance.html.

¹⁶Processing trade refers to the business activity of importing all or part of the raw materials, parts and components, packaging materials from abroad in bond (i.e. duty‐free), and re‐exporting the finished products after processing or assembly by firms within China. Besides processing trade and normal trade, there are more than a dozen other minor trade regimes, but they only account for about 13% of total exports in China during 2019–2020, with normal and processing trade accounting for 59% and 28%, respectively. We exclude these minor categories when calculating the share of processing trade.

¹⁷Source: https://unstats.un.org/unsd/trade/classifications/bec.asp.

¹⁸Codebooks: https://github.com/OxCGRT/covid‐policy‐tracker/blob/master/documentation/codebook.md#economic‐policies and https://github.com/OxCGRT/covid‐policy‐tracker/blob/master/documentation/index_methodology.md.

¹⁹Source: https://www.wto.org/english/tratop_e/covid19_e/trade_related_goods_measure_e.htm (as of March 26, 2021).

²⁰Some countries are covered by our trade data, but not by the COVID‐related data (mostly constrained by the availability of the stringency index). Nevertheless, our sample coverage of importing countries is among the largest among similar studies.

²¹Observe that we allow the country–product effect to vary by year, but not by month of the year, which is the periodicity of our sample—otherwise, it would absorb all of the variation in the dependent variable.

²²In the few specifications where we have separate country and product fixed effects, we cluster the standard errors at the HS6 product level.

²³With 174 countries/regions and 4636 products at the HS6level in our baseline analysis, this amounts to 2*174*4636 =1,613,328 fixed effects. Using yoy measures, we still have one set of fixed effects, but they do not cause similar computational difficulties.

²⁴1‐exp‐19.371/100=17.6%, where −19.371 is the estimated coefficient of Stringency. We divide it by 100 because the dependent variable is in percentage terms.

²⁵1‐exp‐20.861×0.63/100=12%.

²⁶We do not add the work‐from‐home variable by itself in the regressions because it is fully absorbed by the country–product fixed effect.

²⁷We do not add the contract intensity variable by itself in the regressions because it is fully absorbed by the country–product fixed effect.

²⁸See United Nations (2021).

²⁹Here, we compare the coefficient of the CovidD previous cumulative value relative to the absolute value of the coefficient of its contemporaneous effect (i.e. 4.02/20.26=0.2) with the analogous ratio for Stringency (2.44/26.54=0.09).

³⁰Alternatively, we add to our baseline regression two indicator variables for observations covered by a temporary import liberalization or import restriction policy. The result, available upon request, shows that the coefficient of the import liberalization dummy variable is positive, as expected, but is not statistically significant. The coefficient of the import restriction measure is negative and statistically significant, indicating that these measures, although relatively rare, did reduce countries’ imports. The other coefficients of the regression change very little.

³¹In contrast to us, Berthou and Stumpner (2022) find a negative and significant effect of Stringency on the extensive margin of trade, defined just as we define here. An important difference between the two analyses is that Berthou and Stumpner include only Stringency as the explanatory variable. If we drop from our regression CovidD, Stringency_ROW and CoividD_ROW, the coefficient on Stringency turns negative and significant also in our dataset.

• Anderson, J., & Wincoop, E. (2003). Gravity with gravitas: A solution to the border puzzle. American Economic Review, 93(1), 170–192. 10.1257/000282803321455214 [CrossRef] [Google Scholar]
• Antràs, P., Redding, S. J., & Hansberg, E. R. (2020). Globalization and pandemics. Covid Economics, 49, 1–84. [Google Scholar]
• Baldwin, R. & Tomiura, E. (2020). Thinking ahead about the trade impact of covid‐19. In Baldwin R., Weder di Mauro B. (Eds.), Economics in the time of COVID‐19 (pp. 59‐71). CEPR Press. [Google Scholar]
• Bas, M., Fernandes, A., & Paunov, C. (2022). How resilient was trade to COVID‐19? No 9975, Policy Research Working Paper Series. The World Bank. [Google Scholar]
• Behrens, K., Corcos, G., & Mion, G. (2013). Trade crisis? What trade crisis?The Review of Economics and Statistics, 95(2), 702–709. 10.1162/REST_a_00287 [CrossRef] [Google Scholar]
• Bems, R., Johnson, R. C., & Yi, K.‐M. (2010). Demand spillovers and the collapse of trade in the global recession. IMF Economic Review, 58(2), 295–326. 10.1057/imfer.2010.15 [CrossRef] [Google Scholar]
• Bems, R., Johnson, R. C., & Yi, K.‐M. (2013). The great trade collapse. Annual Review of Economics, 5(1), 375–400. 10.1146/annurev-economics-082912-110201 [CrossRef] [Google Scholar]
• Berthou, A., & Stumpner, S. (2022). Trade under lockdown. Banque de France Working Paper No. 867.
• Bonadio, B., Huo, Z., Levchenko, A. A., & Pandalai‐Nayar, N. (2021). Global supply chains in the pandemic. Journal of International Economics, 133, 103534. 10.1016/j.jinteco.2021.103534 [PMC free article] [PubMed] [CrossRef] [Google Scholar]
• Bricongne, J.‐C., Carluccio, J., Fontagné, L., Gaulier, G., & Stumpner, S. (2021). From macro to micro: Heterogeneous exporters in the pandemic. Mimeo. [Google Scholar]
• Bricongne, J.‐C., Fontagné, L., Gaulier, G., Taglioni, D., & Vicard, V. (2012). Firms and the global crisis: French exports in the turmoil. Journal of International Economics, 87(1), 134–146. 10.1016/j.jinteco.2011.07.002 [CrossRef] [Google Scholar]
• Che, Y. I., Liu, W., Zhang, Y., & Zhao, L. (2020). China's exports during the global COVID‐19 pandemic. Frontiers of Economies in China, 15(4), 541–574. [Google Scholar]
• Dingel, J. I., & Neiman, B. (2020). How many jobs can be done at home?Journal of Public Economics, 189, 104235. [PMC free article] [PubMed] [Google Scholar]
• Eaton, J., Kortum, S., Neiman, B., & Romalis, J. (2016). Trade and the global recession. American Economic Review, 106(11), 3401–3438. 10.1257/aer.20101557 [CrossRef] [Google Scholar]
• Eppinger, P., Felbermayr, G., Krebs, O., & Kukharskyy, B. (2021). Decoupling global value chains. Mimeo. 10.2139/ssrn.3848341 [CrossRef] [Google Scholar]
• Espitia, A., Mattoo, A., Rocha, N., Ruta, M., & Winkler, D. (2022). Pandemic trade: Covid‐19, remote work and global value chains. The World Economy, 45(2), 561–589. 10.1111/twec.13117 [PMC free article] [PubMed] [CrossRef] [Google Scholar]
• Friedt, F. L., & Zhong, K. (2020). The triple effect of Covid‐19 on Chinese exports: First evidence of the export supply, import demand and GVC contagion effects. Covid Economics, 53, 72–109. [Google Scholar]
• Hale, T., Angrist, N., Goldszmidt, R., Kira, B., Petherick, A., Phillips, T., Webster, S., Cameron‐Blake, E., Hallas, L., Majumdar, S., & Tatlow, H. (2021). A global panel database of pandemic policies (Oxford COVID‐19 Government Response Tracker). Nature Human Behaviour, 5, 529–538. 10.1038/s41562-021-01079-8 [PubMed] [CrossRef] [Google Scholar]
• Hayakwa, K., & Mukunoki, H. (2021). Impacts of COVID‐19 on global value chains. The Developing Economies, 59(2), 154–177. [PMC free article] [PubMed] [Google Scholar]
• Kejzar, K. Z., Veli, & A. C. (2020). Covid‐19, trade collapse and GVC linkages: European experience. Covid Economics, 61, 219–240. [PMC free article] [PubMed] [Google Scholar]
• Le Moigne, M., & Ossa, R. (2021). Crumbling Economy, Booming Trade: The Surprising Resilience of World Trade in 2020. Working Paper 01‐21 Kühne Center Impact Series.
• Nunn, N. (2007). Relationship‐specificity, incomplete contracts, and the pattern of trade. Quarterly Journal of Economics, 122(2), 569–600. 10.1162/qjec.122.2.569 [CrossRef] [Google Scholar]
• Pei, J., de Vries, G., & Zhang, M. (2021). International trade and Covid‐19: City‐level evidence from China's lockdown policy. Journal of Regional Economics, Forthcoming, 1–26. 10.1111/jors.12559 [PMC free article] [PubMed] [CrossRef] [Google Scholar]
• Sforza, A. & Steininger, M. (2020). Globalization in the Time of Covid‐19. CESifo working paper 8184.
• Socrates, M. K. (2020). The effect of lockdown policies on international trade flows from developing countries: Event study evidence from Kenya. Working Paper. University of Nairobi. [Google Scholar]
• United Nations(2021). World Economic Situation and Prospects: February 2021 Briefing, No. 146.