Policy Research Working Paper 10080 How Well Can Real-Time Indicators Track the Economic Impacts of a Crisis Like COVID-19? Gi Khan Ten Josh Merfeld David Newhouse Utz Pape Poverty and Equity Global Practice June 2022 Policy Research Working Paper 10080 Abstract This paper presents evidence on the extent to which a set of not others. Data on nighttime lights show no clear drop in real-time indicators tracked changes in gross domestic prod- March outside East Asia. Linear models selected using the uct across 142 countries in 2020. The real-time indicators Least Absolute Shrinkage and Selection Operator explain include Google mobility, Google search trends, food price about a third of the variation in annual gross domestic prod- information, nitrogen dioxide, and nighttime lights. Google uct growth rates across 72 countries. In a smaller subset of mobility and staple food prices both declined sharply in higher income countries, real-time indicators explain about March and April, followed by a rapid recovery that returned 40 percent of the variation in quarterly gross domestic prod- to baseline levels by July and August. Mobility and staple uct growth. Overall, mobility and food price data, as well food prices fell less in low-income countries. Nitrogen as pollution data in more developed countries, appeared to dioxide levels show a similar pattern, with a steep fall and be best at capturing the widespread economic disruption rapid recovery in high-income and upper-middle-income experienced during the summer of 2020. The results indi- countries but not in low-income and lower-middle-income cate that these real-time indicators can track a substantial countries. In April and May, Google search terms reflecting percentage of both annual and quarterly changes in gross economic distress and religiosity spiked in some regions but domestic product. This paper is a product of the Poverty and Equity Global Practice. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://www.worldbank.org/prwp. The authors may be contacted at dnewhouse@worldbank.org and upape@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team How Well Can Real-Time Indicators Track the Economic Impacts of a Crisis Like COVID-19? 1 Gi Khan Ten, Josh Merfeld, David Newhouse, Utz Pape JEL: C82, E01, I31 Keywords: COVID-19, GDP impact estimation, big data 1 The team would like to thank the Korea Trust Fund for Economic and Peace-Building Transitions (KTF) for providing funding. The KTF, supported by the Ministry of Economy and Finance, Republic of Korea, is a global fund administered by the World Bank to finance critical development operations and analysis in situations of fragility, conflict, and violence. 1. Introduction SARS-CoV-2, the virus that causes COVID-19, has reached nearly every corner of the world, resulting in millions of deaths (Dong et al., 2020). To decrease the spread of the virus, governments around the world have implemented lockdowns and mobility restrictions. These restrictions are apparent in Figure 1, which visualizes the Oxford Stringency Index, broken down by regions. Governments introduced peak stringency immediately following the onset of the COVID-19 pandemic in the first quarter of 2020, after which there was a mild decline in stringency. These restrictions led to massive economic disruption, as the COVID-19 pandemic has resulted in a reversal of a decade-long declining global poverty trend, pushing upwards of 100 million people into extreme poverty in 2020 (World Bank, 2020). In response to the growing demand for accurate statistics of the economic cost of the pandemic and resulting policies, scholars have adopted three general approaches designed to shed light on the impact of measures like mobility restrictions and lockdowns. One strand of literature used past data to forecast the state of the world's economic well-being during the pandemic. In April 2020, during the peak stringency adopted in most countries, the IMF projected a 3 percent contraction of world economic output, much larger than during the 2008 global financial crisis (IMF, 2020a). The 2021 growth rate was projected to be 6.5 percentage points below the pre-pandemic forecast made in January 2020 (IMF, 2020b). Following the onset of the COVID-19 crisis, other work has used real-time forecasting. Combining pre-pandemic household surveys with the growth projections predicted that nearly 49 million people would fall below the $1.90 poverty threshold in 2020 real terms (Mahler et al., 2020a). An updated set of estimates published in June 2020 painted a gloomier picture, predicting that the number of newly poor individuals would range from 71 million to 100 million by the end of 2020 (Mahler et al., 2020b). The second strand of literature uses interviews of representative parts of the population to understand socio-economic impacts, also as input to parameterize modeling and forecasting efforts. In the context of the pandemic, most interviews were conducted by phone 2 and – to a smaller extent – by using online surveys. Khamis et al. (2021), for example, use data from the high-frequency phone survey (HFPS) to estimate the early impact of COVID-19 on the labor markets of 39 countries. Their findings show that the pandemic has negatively affected labor market outcomes in these countries with respect to job and income losses, lack of payment and job changes. In a related paper, Kugler et al. (2021) focus on the distributional implications of the crisis, showing that female, less educated, and younger workers were initially more severely impacted. Bundervoet et al. (2021) demonstrate that the pandemic’s initial effects were both widespread and highly regressive, with the most vulnerable segments of the population being disproportionally affected. The third strand of studies makes use of remote sensing indicators to monitor economic and other human activity during the COVID-19 pandemic. For example, the literature has long recognized that satellite images can be used to monitor the state of economic development in places where this information is difficult to obtain by other means (e.g., Henderson, Storeygard and Weil, 2012; Donaldson and Storeygard, 2016; Beyer et al., 2018; Kim, 2022). Studies have documented 2 For example, the World Bank conducted phone surveys in dozens of countries and published a harmonized dashboard available at https://www.worldbank.org/en/topic/poverty/brief/high-frequency-monitoring-surveys. 2 declines in nighttime lights in China (Elvidge et al., 2020), India (Beyer, Franco-Bedoya and Galdo, 2021), and Morocco (Roberts, 2021). By the same token, some scholars have taken advantage of real-time information on air pollution. Improvements in air quality during the onset of the COVID-19 pandemic were particularly noticeable in China and the Republic of Korea, two countries with generally poor air quality (Liu et al., 2020; Koo et al., 2020; Seo et al., 2020; Zhang et al., 2020; Ju, Oh, and Choi, 2021), but were also apparent in many cities in Sub-Saharan Africa (Masaki et al, 2021). Finally, several studies investigate the ability of Google data and food prices to track the COVID- 19 crisis. For example, Woloszko (2020) uses Google Trends data to track economic activity in 46 OECD and G20 countries. Abay et al. (2020) utilize Google search terms to investigate changes in demand for selected services across 182 countries during the COVID-19 pandemic. Sampi and Jooste (2020) employ Google mobility data to nowcast monthly industrial production dynamics in Latin America and the Caribbean in roughly the same period. Dietrich et al. (2021) document increases in food prices following the introduction of more stringent social distancing policies. In this paper, we contribute to the latter strand of studies, seeking to answer the following question: how well did real-time indicators track the COVID-19 crisis across a broad range of countries? To answer this question, we collected information on nitrogen dioxide (NO2) emissions and nighttime light intensity at the subnational level for 142 countries with a population of more than 500,000 people. Our findings reveal substantial heterogeneity in the ability of remote sensing indicators to track the COVID-19 crisis, across regions of the world and country income groups. Following the beginning of the crisis in the first quarter of 2020, NO2 emissions showed the steepest drops in the Middle East and North Africa (when grouped by regions) and upper-middle-income countries (when grouped by income level). Nighttime lights showed the steepest decline in East Asia and Pacific (when grouped by regions) and lower-middle-income countries (when grouped by income level). Although not every indicator seems to capture the beginning of the COVID-19 crisis in all corners of the world, both NO2 and nighttime lights show recovery starting from the second quarter of 2020. In addition to these satellite-derived indicators, we compiled a set of indicators that have the potential to reflect the COVID-19 crisis, namely Google mobility data, Google search data, and data on food prices. Of these additional real-time indicators, Google mobility, religious Google search terms, and food prices appear to reflect the onset of the crisis best. Following a sharp change in Google search for religious terms, food prices, and mobility trends at the end of the first quarter of 2020, all these indicators show recovery starting from the second quarter of 2020. Finally, we also examine whether these real-time indicators predict deviations from projected GDP in a subset of countries, most from the OECD, in which quarterly GDP estimates are available. Selected features from grouped lasso explain around 88 percent of the total variation from predicted GDP in the subsample. Around half of this is explained by real-time indicators, with the remaining half due to quarter and country fixed effects. Two indicators – a price indicator and a mobility indicator – alone explain more than 30 percent of the total r-squared, or just short of 30 percent of the total variation in the dependent variable. The results suggest that real-time indicators can complement existing tools to monitor the economic impacts of a crisis like COVID-19, in which mobility was sharply restricted. 3 The remainder of the paper proceeds as follows. Section 2 describes our data. Section 3 presents our descriptive graphical findings, while Section 4 presents our empirical results. Section 5 concludes. 2. Data and Methodology 2.1 Data We consider a variety of real-time indicators, including data on nitrogen dioxide (NO2), night- time lights, Google mobility indicators, Google search data, and food price data. NO2 is a pollutant produced by everyday human activity, such as burning fossil fuels. As such, it is attractive as a potential measure of the changes brought about by the COVID-19 crisis, as changes in patterns of human behavior can directly affect NO2 emissions. We use the TROPONOMI Level 2 Nitrogen Dioxide product obtained initially by the Sentinel-5 Precursor, an Earth observation satellite launched as a part of the Copernicus Programme in October 2017. We obtained data from January 2019 to December 2020, with the resulting series containing monthly concentrations of NO2 observed at the subnational level. We also collected information on monthly nighttime lights intensity from the Visible Infrared Imaging Radiometer Suite (VIIRS) sensor onboard the Soumi-NPP Earth observation satellite processed by Elvidge et al. (2017). Nightlights have been found to correlate well with some measures of economic growth and development (see, e.g., Henderson et al., 2012) and may therefore provide a window into possible changes in economic activity brought about by the COVID-19 crisis. We use the Annual VNL Version 1 product that corrects data influenced by stray lights, excludes observations impacted by cloud cover, and has greater spatial coverage (especially in the northern part of the globe; for additional information, see Mills et al., 2013). As in the case of the NO2 series, the resulting data set contains monthly time series on nighttime light intensity observed at the subnational level from January 2019 to December 2020. To account for seasonal fluctuations in NO2 and nighttime lights series, we construct year-over-year deviations in 2020 relative to the same month in 2019, expressed in percentage changes. A potential problem with these satellite-derived variables is the possibility of capturing noise, and its consequences are exacerbated given that we work with large-area zonal statistics. For example, Mills et al. (2013) point out that the Annual VNL Version 1 product has not been filtered out from aurora lights, fires, boats, biomass burning, and glass fares. Also, a failure to account for cloud pressure may artificially elevate the air pollution level observed in the data. To remove the influence of extreme outliers—whose values are more likely to be driven by noise—we look at the median values of nighttime lights/NO2 concentration in a given group of subnational regions when analyzing the data. For example, when aggregating data by region/income group, we look at the median subnational region in a given month. Likewise, we pick the median subnational region in a given country when constructing the country-level data. In response to the crisis, Google has published trends in the movement of people across different categories of places, like workplaces, transit stations, and residences, using data derived from cell 4 phones. 3 The data presents changes in visits to these different places relative to some baseline value for a given geographic location, and it is available at the subnational level in some countries. These data are only obtained from mobile phone users with location services turned on but nonetheless can reflect useful signals in the data. We extracted Google mobility and search trends at the subnational level. The raw GPS data on mobility were first grouped by categories of places with similar characteristics, namely: grocery and pharmacy, parks, transit stations, retail and recreation, residential areas, and workplaces. 4 The resulting data shows how visits to categorized places change relative to the baseline period, which is the region-specific median value picked from the 5-week period between January 3 and February 6, 2020, just before the onset of the global pandemic. The Google mobility information is available only from February 2020, which naturally limits its usage—when tracking changes in economic activity—to the COVID-19 period. Google search data—that is also available at the subnational level in some countries—provide information regarding the search intensity of a given combination of keywords for the target location and time period. Each data point can be interpreted as a share of a given request in the total searches in a given location and time period. However, as Abay et al. (2020) point out, the raw Google search series are not comparable across terms and geographic areas, which in our case are subnational regions. The primary reason is the following: when requesting Google search data for a given search term, it is possible to obtain the information only for a limited number of geographic locations per request. Thus, to obtain the information for the desired number of geographic locations, one has to file multiple requests in succession. The resulting data points are comparable across terms and geographic locations only within one request. To achieve the desired comparability of Google search data points across different requests – and following Abay et al. (2020) – we constructed our data set as follows. First, we extracted the Google search data for every subnational region in our sample and California jointly for a given search term and query period. In this case, the data on search terms from a given subnational region and California define one request. At the second stage, within each request, we normalized the time series from the target subnational region by the maximum value of the same search term index observed in California during the query period. Having repeated the above steps with every search term, we obtained a data set containing a list of search term indices comparable across different subnational regions. We extracted the information on the following search terms: food bank, assistance, online education, prayer, Bible, and Quran. Finally, we extracted information on food prices from FAO’s Daily Food Price Monitor. The raw data set contains information on percentage changes in food prices in 180 countries relative to the baseline period (February 14, 2020). We have information on prices for the following food staples: apple, banana, bread, cheese, eggs, lettuce, meat, chicken, onions, oranges, potato, milk, and rice. When aggregating data by regions or income levels, we weight each observation by the population in a given subnational region, normalized by the total country-level population in 2018. Thus, every country in our sample is given the same weight, whereas higher weights are given to more 3 https://www.google.com/covid19/mobility/ 4 This preprocessing takes place at Google. 5 populous subnational regions within a given country. We obtained subnational estimates of the 2018 population counts from WorldPop. 5 2.2 Methodology First, we look at broad trends in the real-time indicators. Using our current knowledge of the evolution of the crisis throughout 2020, we identify key patterns in trends using purely descriptive graphical representations of those indicators that we believe to be the most reflective of the COVID-19 crisis in most parts of the world. We start with using both satellite-derived variables– nighttime lights and NO2–to track economic activity. We continue with the Google mobility indices and the Google Search data set. Finally, we use food prices restricted to the following six food items commonly consumed in many countries: bread, chicken, eggs, milk, onions, and rice. Second, we estimate simple cross-section regressions with three separate dependent variables: GDP growth, deviation of actual GDP growth from the pre-pandemic expected growth, and the stringency index presented in Hale et al. (2021), which we refer to as the “Oxford Stringency Index.” We obtained country-level GDP – as well as values for predicted GDP from prior to the crisis – for a large group of countries for the entire year of 2020 from the IMF. We then constructed deviations of actual GDP from predicted GDP to analyze how well the real-time indicators predict these deviations. There are 62 countries for which we have information on the 2020 growth rates plus the complete list of real-time indicators described above. Table A1 lists the countries included in this analysis. We first present regressions including all candidate predictors. Then, to create a more parsimonious model, we also employ lasso to select features. We use the cross-validation (CV) method to select the value of that minimizes the mean squared error (MSE) in the validation sample and present the resulting estimates. We include information on 2019 GDP per capita, total population, and primary school enrollment rate as candidate covariates taken from the World Development Indicators database. Third, we obtained more detailed quarterly GDP data for a subset of countries that mainly consists of the OECD members. 6 In addition, since the quarterly data form a panel of countries, we employ group lasso to select features, grouping quarter fixed effects and country fixed effects. 7 The quarterly GDP growth data primarily come from high-income states, whereas the cross-sectional data set consists of countries with varying income statuses. One might expect that the predictive power of lasso-selected covariates is different across the samples, so we focus on these differences throughout the results below. There are 39 countries with non-missing data for quarterly GDP and the candidate features; a list is available in Table A2 of the appendix. Some candidate features are likely to be strongly collinear. We thus use factor analysis to group the data as follows: 1. Two factors for food prices: apples, bananas, bread, cheese, eggs, lettuce, meat, chicken, onions, oranges, potatoes, milk, and rice (all relative to January of 2020) 5 https://www.worldpop.org/ 6 These GDP statistics come from the OECD (2021). 7 In other words, lasso will either select all quarterly dummies or none of them (and the same for country dummies). We implement this using gglasso in R. 6 2. Two factors for Google search terms related to labor: money, salary, LinkedIn, unemployment, employment 3. Two factors for all Google mobility data 4. One factor for food and assistance search terms: assistance, money, bread, milk, wheat 5. One factor for education and sport search terms: school, student, exercise 6. One factor for religion search terms: prayer, bible A key question in the regression results is what share of the total r-squared is attributable to different combinations of variables. To shed some light on this issue, following Shorrocks (1982), we investigate the contribution of each group of real-time indicators to the overall goodness of fit of each regression, using Shorrocks–Shapley decomposition of r-squared. 8 The resulting decomposed values of r-squared allow checking the fraction of the goodness-of-fit contributed separately by each group of real-time indicators and fixed effects, with the latter especially pertinent in the quarterly results where we include both quarter and country fixed effects. 3. Findings 3.1 Satellite indicators a. Nitrogen dioxide Figure 2 presents trends in median NO2 emissions across seven different geographic regions: East Asia and the Pacific, Europe and Central Asia, Latin America and the Caribbean, the Middle East and North Africa, South Asia, Sub-Saharan Africa, and North America. At first glance, there is a noticeable decrease in NO2 emissions in all regions – Latin America being a possible exception – with declines in Europe/Central Asia and Sub-Saharan Africa starting in February. Interestingly, the largest decrease is in Europe and Central Asia, and it began in February, even though the large changes in human behavior began in March. The decreases in the Middle East/North Africa, South Asia, and East Asia and the Pacific took place in March and generally did not recover until much later in the year. Trends in North America and East Asia and the Pacific are very comparable in the first quarter. However, North America faced a steeper increase in NO2 since April 2020, followed by a steady decline during the next three months and another rebound in the last quarter of the year. Figure 3 presents results disaggregated by country income levels instead of geographic regions. We see a noticeable decrease across all income categories except low-income countries. The largest decrease is in high-income countries, but this decrease again starts in February. The upward trend is more consistent in this figure than in the previous one; after April, there is a noticeable uptick across all four income categories, with numbers returning to 2019 levels in the last quarter of the year. b. Nightlights 8 We implement this using shapley in Stata (Kolenikov, 2000). 7 Figure 4 presents data for 2020 across the same regions as in Figure 1. All regions are increasing from January to February, and most at similar rates. However, in March, all regions show a marked change, with four of the six decreasing coincident with the start of the crisis. South Asia shows a dip in May and June, before rebounding in July and August. Four regions continue decreasing over the few months following the onset of the crisis, while Sub-Saharan Africa seems to recover and even increase rather substantially in the last several months of the year. Finally, consistent with the NO2 results shown in Figure 2, North America shows a steep increase in the second quarter of the year followed by a steady decrease in the next three months. A rebound is observed only in the last quarter of the year. . When we disaggregate by country income in Figure 5, we see a similar dip in March in three of the four categories, but these dips are relatively small. The only category that appears to have a sustained decrease is lower middle income; nightlights decrease in March, April, and May, before showing a brief increase in the following months. None of the categories shows consistent decreases from the previous year. In fact, high-income countries show the lowest increases but are right at zero for the entire year. The other three categories are above zero in almost all months and sometimes substantially above zero. Overall, NO2 appears to correlate with the onset of the crisis in upper-middle-income and high- income countries, but nightlights do not seem to exhibit a consistent pattern. 3.2 Google data a. Mobility Figure 6 presents the Google mobility data for all the countries in our sample. The baseline for most areas is February, which is why we see all of the trends clustered around zero (i.e. no change from the baseline) in that month. Visits to all locations outside residences decreased markedly in March, bottoming out in April, with the largest relative decreases in transit stations and retail/recreation locations. On the other hand, time in residences increased markedly in April before slowly decreasing but remaining above pre-crisis levels. Moreover, while visits to grocery and pharmacy locations rebounded to February levels by around October, the other four non- residence locations – with the exception of parks, which see a bit of a jump during the northern hemisphere summer – remain well below pre-pandemic levels throughout the year. In January of 2021, all these locations were still more than 20 percent below their February levels. We generally see the same trends across regions. Figure 7 presents trends for four separate locations – workplaces, grocery/pharmacy, residences, and retail/recreation – for all seven regions. Beginning with workplaces, we see the same marked decrease in visits in March and April, with visits again bottoming out in the latter month. There is quite a bit of heterogeneity across regions, however. For example, Latin American and the Caribbean countries see a decrease of almost 50 percent in April, while East Asia and the Pacific and Sub-Saharan Africa see declines of just half that amount. There are similar amounts of heterogeneity in the other three categories, though the patterns sometimes differ. For example, while grocery/pharmacy shows a decrease in most regions, there is a noted increase in visits to these locations in all regions except Latin America and the Caribbean 8 following the April nadir. There is an uptick in visits to residences in all regions and a drop in visits to retail/recreation locations, again with the biggest changes in April. Interestingly, Latin America and the Caribbean show the largest changes in all four categories, with the biggest drops in visits to workplace, grocery/pharmacy, and retail/recreation locations and the biggest increase in visits to residential locations. The indicators may even track variability in the severity of the crisis, not just the initial drop. For example, there is an increase in visits to retail and recreation in Europe and Central Asia in July and August, and it coincides with summer in these areas, as well as with a lull in new coronavirus cases. Across all regions, the pattern is clear: there is a sudden, large decrease in visits to non-residential locations in March and April. Moreover, visits to these locations – especially workplaces and retail/recreation – remain far below their February levels throughout the entirety of 2020. Figure 8 presents the final set of mobility data, which disaggregates the data by country income. While we do see decreases in visits to workplace, grocery/pharmacy, and retail/recreation locations and an increase in visits to residential locations, one clear pattern is that visits to workplaces decreased the least in low-income countries. Whether this is due to differences in the underlying data – e.g. differences in the people who are contributing to the data – or actual differences in behavior is an open question. One possible explanation is that the classification of residences and workplaces is more difficult in developing economies, where a large proportion of the population is self-employed. In Sub-Saharan Africa as a whole, for example, more than 80 percent of workers may be self-employed (Cho et al., 2016), which means that “residence” and “workplace” may be one and the same. In all the mobility data, April shows the biggest change from baseline. While visits to many locations increased somewhat rapidly after April, overall visits remained significantly below the baseline month well into the year. This finding appears especially true for both workplaces and retail/recreation. b. Search data We present trends for six different search terms – food bank, assistance, online education, prayer, bible, and quran – in Figure 9 and Figure 10. We begin with trends by region in Figure 9. The top- left panel presents results for how often people searched for the term “food bank.” There is a large increase in just one region: North America. This could be due to the availability of food banks in that region relative to the others, rather than differences in actual needs. The top-middle panel presents results for “assistance.” Here, we see the most pronounced spike in two regions: South Asia and East Asia and the Pacific. The top-right panel shows trends in searches for “online education.” East Asia and the Pacific and Europe and Central Asia show marked increases in searches for the term in April. Sub-Saharan Africa also shows an increase, though the magnitude of the change is much smaller than for the other two regions. 9 The bottom three panels present results for searches related to religion. The bottom-left panel is a generic “prayer” search, while the other two are specific to different religions: “bible” for Christianity and “quran” for Islam. At first glance, there appears to be a jump in searches for “prayer” and “quran” in April/May of 2020. However, there is an important caveat: we see the same increase in 2019. This may be driven by the fact that Ramadan took place around late April/early May to late May/early June. In 2019, for example, Eid al-Fitr was June 3rd to 4th, while in 2020 the dates were May 23rd and 24th. On the other hand, searches for “bible” do not show a bump in 2019, yet we see a large bump in April 2020, similar to the mobility trends. Again, the jump is biggest in Sub-Saharan Africa, but there is also a sizeable bump in Latin America and the Caribbean. Figure 10 presents the results disaggregated by country income. In the top panel, all three search trends – food bank, assistance, and online education – show large increases in searches in April and May. Interestingly, we also see a small increase in searches for online education in low-income countries. We see similar patterns – if not levels – across all income categories for the bottom three searches. However, it is again worth noting that only for “bible” is the trend different from 2019; both “prayer” and “quran” show similar patterns across the two years, although the 2020 bump happens to line up with the start of the pandemic, though the jump in prayer seems to be larger in 2020. 3.3 Food prices Figure 11 breaks down the average prices by region. There is a single pattern tying all of these different food items together: there is a marked drop in average prices between March and May/June. The largest drop differs by food – for example, the largest drop for eggs is in South Asia in June through August, while the largest drop for chicken is in East Asia and the Pacific as well as the Middle East and North Africa. Moreover, while exact patterns differ across regions – bread and egg prices seem relatively flat in the Middle East and North Africa, for example – there appear to be clear similarities. It seems unlikely that this is driven by, for example, simple seasonal patterns for a couple of reasons. First, the same patterns are seen throughout all regions, with relatively few exceptions. Since these regions are quite heterogeneous – for example, the agricultural season is at different times of the year in many of the countries – it seems unlikely that seasonal patterns could drive the shared variation. Second, Figure A1 in the appendix presents average global prices for rice, taken from the Federal Reserve Bank of St. Louis and the IMF. It does not appear that yearly prices, at least for rice, tend to be lower in May and June than in January. If anything, it is usually the opposite. The final set of results breaks down food prices by country income category. We present these results in Figure 12. The results are consistent with the patterns seen in the previous figure, although this breakdown adds a bit more information. Specifically, the drop was much lower in low-income countries for several foods – eggs, milk, rice, and, especially, chicken and onions. For 10 example, the price of onions increased by around 10 percent in March relative to January before dropping back to January levels in May/June. Likewise, the price of chicken is consistently higher – relative to January – in low-income countries. In fact, chicken does not appear to drop below January prices until the end of the year. 4. Which real-time indicators are the best predictors of the 2020 growth rate and the severity of stringency measures? Our next aim is to understand which real-time indicators best predict 2020 growth rates across countries and the severity of stringency measures. We begin with regressions of growth rates and the Oxford Stringency Index on the groups of variables described above, as well as our set of macro variables – log of GDP per capita (PPP, 2019), log of 2019 total population, and 2019 primary school enrollment rate. We present results from these country-level regressions in Table 1. The first column uses the actual growth rate as the outcome, while the second column reports the results for deviation from the predicted annual growth rate. The last column of Table 1 uses the Oxford Stringency Index on the left-hand side. In the first two columns of Table 1, the coefficients’ signs are consistent across the columns. Both price factors have negative signs; Google mobility data is positively associated with growth and its deviation from the 2019 projections. Satellite-derived indicators are also positively correlated with economic activity. The last column of Table 1 paints a slightly different picture. Both price factors have opposite signs. Higher mobility is associated with the lower value of the stringency index. Of the two remote sensing indicators, only NO2 significantly predicts the stringency index. It is worth noting that when a variable is significant in both columns two and three, it is always of the opposite sign across the two columns. Column two of Table 2 shows the coefficients of features selected for the growth rate’s deviation from its 2019 projection. Besides three features selected in the previous column–the second price factor and two mobility indices–lasso selects the second labor search index. Together with the macro variables, these lasso-selected features explain 49.5% of the variation in the outcome. Finally, except for the education and sports search index, lasso selects all the indices for the stringency regression. The lasso-selected variables explain 68% of cross-country variation in stringency measures and a higher value of adjusted r-squared indicates a better fit compared to the unrestricted specification reported in column three of Table 1. To assess the individual predictive power of every lasso-selected feature, we check the contribution of every variable to the r-squared in each model, expressed in percent. Table 3 reports the results of this decomposition. According to our findings, Google mobility and food prices have the largest explanatory power. The two groups of indices contribute to 72.5%, 87.2%, and 58.4% of the overall r-squared from growth rates, their deviations from projected values, and stringency index regressions, respectively. The search terms also contribute to a reasonably large share of r- squared in every regression reported in Table 2. Finally, satellite-derived data and macroeconomic variables account for the minor shares of the goodness-of-fit measures. 11 It is worth noting that a sizeable proportion of people in some countries do not have internet access, which might limit the ability of some predictors—e.g., Google search trends—to track the crisis. We address this concern in two ways. First, we link the absolute values of the OLS residuals to the data on the internet users per capita to see whether the accuracy of our prediction depends upon the internet penetration. 9 As shown in Figure 13, our regressions with LASSO-selected features perform better in high-income countries with high internet penetration—as manifested by the negative relationship between the absolute values of the OLS residuals and the internet users per capita. However, we cannot say that LASSO-selected features have a greater capability to track the Stringency Index in countries with higher shares of internet users—if anything, the actual values of the index in question have smaller deviations from their predicted counterparts in countries with lower internet penetration. Second, we look at the within-country dynamics of the economic performance in a panel of states with comparable levels of income that mainly consists of OECD members. To get a sense of the temporal heterogeneity of economic growth across quarters in 2020, Figure 14 plots quarter-over- quarter, seasonally adjusted GDP growth rates for the 41 countries for which we have data. There are two striking patterns. First, every country recorded a lower growth rate in the second quarter than in the first, consistent with the peak of stringency measures shown earlier in Figure 1. Second, every country recorded a higher growth rate in the third quarter than in the second, though the overall pattern is negative growth across the second half of the year. Table 4 presents the regression results for the quarterly-level data. Column one presents results with all variables and quarter dummies, but without country fixed effects. Column two includes country dummies to ascertain whether these features can predict changes within countries during the crisis. The first mobility factor seems to be the only consistent predictor of GDP growth according to the first two columns of Table 4. The inclusion of country dummies seems to improve the precision of some coefficients and increases r-squared and adjusted r-squared quite substantially. Column three presents the group lasso-selected features and their coefficients. The group lasso selects both sets of fixed effects and all other candidate features except for the religion index and the first labor index. The overall predictive power as measured by r-squared decreases only slightly, and the adjusted r-squared goes up slightly. At this stage, it might be useful to compare the relative strength of the real-time indicators’ predictive power across the two samples of countries used in our analysis. The first is a sample of 61 countries with different income status. The sample in question includes 34 high-income, 18 upper-middle-income, eight lower-middle-income and one low-income countries. The second sample consists of 39 countries with relatively more high-income states (29 high-income countries). 9 We have obtained the information on the internet users per capita (most recent year available) from the World Telecommunication/ICT Indicators Database. See https://www.itu.int/en/ITU- D/Statistics/Pages/publications/wtid.aspx.f 12 We note that in Table 4, the value of r-squared from the regression with lasso-selected features – 0.882 – is above the one obtained from the identical exercise reported in column one of Table 3, which is 0.486. However, this includes both quarter and country fixed effects, so the comparison is not apples-to-apples. According to the last column of Table 4, the second quarter dummy explains the most considerable fraction of r-squared from the GDP growth regression. This finding is not surprising, given the dip in 2020Q2 GDP growth shown in Figure 13. Consistent with the r- squared decomposition results from the cross-section of countries – reported earlier in Table 4 – among all the real-time indicators, Google mobility and food prices are the first and second largest contributors to the r-squared from the GDP growth model, respectively. Altogether, the lasso- selected real-time indicators account for 45.7% of the r-squared obtained after fitting the model reported in column three of Table 4. In other words, the features in question explain 0.457 × 0.882 = 40.3% of the variation in growth rates in the panel of high-income countries. It might be useful to contrast this figure with that obtained from the cross-section of countries that includes states with lower income. According to column one of Table 3, lasso-selected features contribute to (1 − 0.0574) = 94.3% of the overall r-squared reported in column one of Table 2. The r-squared in question is equal to 0.486. Therefore, the lasso-selected features explain 0.486 × (1 − 0.0574) = 45.8% of the variation in GDP growth in the cross-section of 61 countries. This explanatory power of real-time indicators is comparable to that observed in the sample of high-income countries. From this observation, we conclude that, among all the real-time indicators considered in our analysis, Google mobility data and staple food prices are the most robust predictors of GDP growth and that their explanatory power is highly comparable across the two samples of countries available to us. The latter conclusion speaks to the promise of the ability of the two sets of real-time indicators in question to track different crises irrespectively of the country’s income status. 5. Conclusion In this paper, we examined the capability of a set of real-time indicators to track the economic impacts of the COVID-19 crisis around the world. Of five different categories of real-time indicators, Google mobility, Google search trends, and food prices seem to have the strongest potential to track the onset of the pandemic. Google mobility data showed a steep reduction in mobility trends across all six categories of places, with a steady recovery that followed a gradual relief of stringency measures worldwide. The ability of Google search terms to track the COVID- 19 crisis appeared to be income- and region-specific. Except for high-income countries, religious search terms lined up with the onset of the pandemic, returning to the pre-crisis level in the second half of 2020. However, the generalizability of these terms to other crises remains uncertain. Furthermore, searches for food banks, assistance programs, and online education are highly dependent on the location and level of income. Food prices also painted an informative picture of the COVID-19 crisis state throughout 2020. The data showed that prices for basic food categories dropped sharply in most countries, rising to their pre-pandemic values only at the end of 2020. While NO2 and nighttime lights do not appear to follow the beginning of the COVID-19 crisis everywhere, both indicators show recovery starting from the third quarter of 2020. A more rigorous investigation would require longer time series of real-time indicators and a higher geographic resolution of the analysis. For example, utilizing an extended temporal coverage, future 13 work could investigate the ability of real-time indicators observed in the past to produce ex- ante predictions of the economic losses implied by the actual values of predictors obtained during the COVID-19 pandemic. The accuracy of such predictions can be validated by comparing them against the actual GDP figures that usually come with a one-year lag. Also, the results point out potential caveats associated with the use of the “digital” real-time indicators, as a non-trivial proportion of people in low-income countries face limited internet access. With the availability of data at a higher geographic resolution, it would be possible to assess the models’ performance in settings with comparable rates of internet penetration—e.g., at the subnational level in a given country—and examine how the prediction accuracy depends upon the internet coverage in a more detailed way. Despite these potential extensions, the analysis above can help shed light on how best to monitor the economic impact of a crisis like COVID-19 when survey data is scarce. Real-time indicators have several advantages in monitoring the impact of crises. They are often readily available, low cost, and do not involve face-to-face interactions, which is particularly important in a pandemic. Some indicators, particularly those relating to mobility, may be particularly valuable during the COVID-19 crisis when mobility was restricted. Nonetheless, the analysis suggests that Google search terms and food prices are promising indicators for tracking the trajectory of economic impacts during crises, particularly when combined in a parsimonious model that predicts post- crisis outcomes. References Abay, K. A., Tafere, K., & Woldemichael, A. (2020). Winners and losers from COVID-19: Global evidence from Google Search. Policy Research Working Paper No. 9268. World Bank, Washington, DC. Beyer, R. C., Chhabra, E., Galdo, V., and Rama, M. (2018). Measuring districts' monthly economic activity from outer space. The World Bank. Beyer, R. C., Franco-Bedoya, S., and Galdo, V. (2021). Examining the economic impact of COVID-19 in India through daily electricity consumption and nighttime light intensity. World Development, 140, 105287. Bundervoet, T., Davalos, M. E., and Garcia, N. (2021). The short-term impacts of COVID-19 on households in developing countries: An overview based on a harmonized data set of high- frequency surveys. Policy Research Working Paper No. 9582. World Bank, Washington, DC. Cho, Y., Robalino, D., and Watson, S. (2016). Supporting self-employment and small-scale entrepreneurship: potential programs to improve livelihoods for vulnerable workers. IZA Journal of Labor Policy, 5(1), 1-26. Dietrich, S., Giuffrida, V., Martorano, B., & Schmerzeck, G. (2021). COVID‐19 policy responses, mobility, and food prices. American Journal of Agricultural Economics. Donaldson, D., and Storeygard, A. (2016). The view from above: Applications of satellite data in economics. Journal of Economic Perspectives, 30(4), 171-98. Dong, E., Du, H., & Gardner, L. (2020). An interactive web-based dashboard to track COVID-19 in real time. The Lancet Infectious Diseases, 20(5), 533-534. Elvidge, C. D., Baugh, K., Zhizhin, M., Hsu, F. C., & Ghosh, T. (2017). VIIRS night-time lights. International Journal of Remote Sensing, 38(21), 5860-5879. 14 Hale, T., Angrist, N., Goldszmidt, R., Kira, B., Petherick, A., Phillips, T., Cameron-Blake E., Hallas, L. M., and Tatlow, H. (2021). A global panel database of pandemic policies (Oxford COVID-19 Government Response Tracker). Nature Human Behaviour, 5(4), 529-538. Henderson, J. V., Storeygard, A., and Weil, D. N. (2012). Measuring economic growth from outer space. American Economic Review, 102(2), 994-1028. IMF (2020a). World Economic Outlook: The Great Lockdown, April 2020. Washington, D.C.: International Monetary Fund. IMF (2020b). World Economic Outlook: A Crisis Like No Other. An Uncertain Recovery. June 2020. Washington, D.C.: International Monetary Fund. Ju, M. J., Oh, J., and Choi, Y. H. (2021). Changes in air pollution levels after COVID-19 outbreak in Korea. Science of the Total Environment, 750, 141521. Khamis, M., Prinz, D., Newhouse, D., Palacios-Lopez, A., Pape, U., & Weber, M. (2021). The Early Labor Market Impacts of COVID-19 in Developing Countries. Jobs Working Paper No. 58. World Bank, Washington, DC. Kim, K. (2022). The North Korean economy seen by satellite: Estimates of national performance, regional gaps based on nighttime light. Journal of Asian Economics, 78, 101405. Kolenikov, S. (2000). SHAPLEY: Stata module to perform additive decomposition of sample statistic. Statistical Software Components S411401, Boston College Department of Economics. Koo, J. H., Kim, J., Lee, Y. G., Park, S. S., Lee, S., Chong, H., Cho, Y., Kim, J., Choi, K., and Lee, T. (2020). The implication of the air quality pattern in South Korea after the COVID- 19 outbreak. Scientific reports, 10(1), 1-11. Kugler, M.D., Viollaz, M., Vasconcellos, A., Gaddis, I., Newhouse, D.L., Palacios Lopez, A., and Weber, M. (2021). How did the COVID-19 crisis affect different types of workers in the developing world? Jobs Working Paper No. 60. World Bank, Washington, DC. Liu, Q., Sha, D., Liu, W., Houser, P., Zhang, L., Hou, R., Lan, H., Flynn, C., Lu, M., Hu, T., and Yang, C. (2020). Spatiotemporal Patterns of COVID-19 Impact on Human Activities and Environment in China Using Nighttime Light and Air Quality Data. arXiv preprint arXiv:2005.02808. Mahler, D.G., Lakner, C. Aguilar, R.A.C., and Wu, H. (2020a). The Impact of COVID-19 (Coronavirus) on global poverty: Why Sub-Saharan Africa might be the region hardest hit. World Bank Blogs: Data Blogs, April 20, 2020. Mahler, D.G., Lakner, C. Aguilar, R.A.C., and Wu, H. (2020b). Updated Estimates of the Impact of COVID-19 on Global Poverty. World Bank Blogs: Data Blogs, June 8, 2020. Masaki, T., Nakamura, S., & Newhouse, D. (2020). How is the COVID-19 Crisis Affecting Nitrogen Dioxide Emissions in Sub-Saharan Africa?. Mills, S., Weiss, S., & Liang, C. (2013). VIIRS day/night band (DNB) stray light characterization and correction. In Earth Observing Systems XVIII (Vol. 8866, p. 88661P). International Society for Optics and Photonics. OECD (2021), OECD Economic Outlook, Interim Report September 2021: Keeping the Recovery on Track, OECD Publishing, Paris. Roberts, M. (2021). Tracking Economic Activity in Response to the COVID-19 Crisis Using Nighttime Lights. Sampi, J.,and Jooste, C. (2020). Nowcasting Economic Activity in Times of COVID-19 : An Approximation from the Google Community Mobility Report. Policy Research Working Paper No. 9247. World Bank, Washington, DC 15 Seo, J. H., Jeon, H. W., Sung, U. J., & Sohn, J. R. (2020). Impact of the COVID-19 outbreak on air quality in Korea. Atmosphere, 11(10), 1137. Shorrocks, A. F. (1982). Inequality Decomposition by Factor Components. Econometrica, 50(1), 193– 211. Stephens-Davidowitz, S. (2014). The cost of racial animus on a black candidate: Evidence using Google search data. Journal of Public Economics, 118, 26-40. Woloszko, N. (2020), "Tracking activity in real time with Google Trends", OECD Economics Department Working Papers, No. 1634. World Bank. (2020). Poverty and Shared Prosperity 2020: Reversals of Fortune. The World Bank. Zhang, R., Zhang, Y., Lin, H., Feng, X., Fu, T. M., & Wang, Y. (2020). NOx emission reduction and recovery during COVID-19 in East China. Atmosphere, 11(4), 433. Zou, H., and T. Hastie. 2005. Regularization and variable selection via the elastic net. Journal of the Royal Statistical Society, Series B 67: 301–320. Figure 1 – Oxford Stringency Index, by region. 16 NOTES: Average values of Oxford Stringency index, by region. Higher values indicate stronger stringency measures in a given region/month. For more details about Oxford Stringency Index, see Hale et al. (2021). 17 Figure 2 – NO2 year-over-year change (percent), by region NOTES: Each line shows a year-over-year change (percent) in the median subnational region in a given world region. 18 Figure 3 - NO2 year-over-year change (percent), by income group NOTES: Each line shows a year-over-year change (percent) in the median subnational region in a given income group. 19 Figure 4 - Nightlights year-over-year change (percent), by region NOTES: Each line shows a year-over-year change (percent) in the median subnational region in a given world region. 20 Figure 5 - Nightlights year-over-year change (percent), by income group NOTES: Each line shows a year-over-year change (percent) in the median subnational region in a given income group. 21 Figure 6 - Google mobility trends NOTES: Average values of Google Mobility index, by category. 22 Figure 7 - Google mobility indices, by region NOTES: Average values of different categories of Google Mobility index, by region. 23 Figure 8 - Google mobility indices, by income group NOTES: Average values of different categories of Google Mobility index, by income group. 24 Figure 9 - Google search trends, by region NOTES: Average values of Google search terms, by region. 25 Figure 10 - Google search trends, by income group NOTES: Average values of Google search terms, by income group. 26 Figure 11 - Selected food prices, by region NOTES: Percentage changes in food prices relative to the baseline period (February 14, 2020), by region. 27 Figure 12 - Selected food prices, by income group NOTES: Percentage changes in food prices relative to the baseline period (February 14, 2020), by income group. 28 Figure 13 – OLS residuals from regressions with LASSO-selected features and internet penetration NOTES: Country-specific absolute values of OLS residuals—obtained from fitting the models with LASSO-selected features reported in Table 2—plotted against the number of internet users per capita. 29 Figure 14: GDP Growth Rates by Quarter of 2020 NOTES: Quarterly GDP growth rates in 2020 for a list of countries available in OECD (2021). 30 Table 1: Country-Level Regressions (1) (2) (3) Growth rate Growth rate (deviation from Stringency Index (actual) predicted) b/se b/se b/se Price factors Factor 1 -0.284 -0.270 1.915* (0.622) (0.336) (1.125) Factor 2 -1.806*** -1.522*** -0.451 (0.547) (0.496) (1.014) Google mobility Factor 1 1.839*** 1.831*** -5.318*** (0.484) (0.477) (0.920) Factor 2 1.757*** 1.612*** -2.644* (0.626) (0.466) (1.451) Satellite-derived data NO2 0.675 0.529 2.277** (0.686) (0.603) (0.886) Nightlights 2.846 0.380 -7.891 (1.735) (2.159) (5.725) Search terms factors Labor 1 0.037 0.135 -1.242 (0.499) (0.408) (0.816) Labor 2 0.134 0.494 -2.032 (0.810) (0.831) (1.404) Food and assistance -1.682** -1.185* 2.860** (0.752) (0.626) (1.258) Education and sports 0.870 0.178 -0.830 (0.922) (0.746) (1.545) Religion -0.738 -1.207 2.373 (2.825) (1.943) (9.067) Macro variables Log (GDP per capita 2019, ppp) 1.123 1.812*** -0.231 (0.706) (0.636) (1.601) Log (population 2019) -0.157 -0.012 2.211*** (0.354) (0.337) (0.500) % Primary school 0.060 0.046 0.225*** (0.052) (0.045) (0.081) r-squared 0.495 0.506 0.682 31 adjusted r-squared 0.345 0.359 0.585 Observations 62 62 61 NOTES: OLS regression results with full sets of controls. Robust standard errors are in parentheses. * 0.1 ** 0.05 *** 0.01. 32 Table 2: Country-Level Regressions, LASSO-Selected Grouped Features (1) (2) (3) Growth rate Growth rate (deviation from Stringency Index (actual) predicted) b/se b/se b/se Price factors Factor 1 1.918* (1.113) Factor 2 -1.693*** -1.392*** -0.338 (0.462) (0.384) (1.006) Google mobility factors Factor 1 1.818*** 1.916*** -5.445*** (0.437) (0.384) (0.892) Factor 2 1.725*** 1.559*** -2.608* (0.549) (0.375) (1.435) Satellite-derived data Nightlights 2.916* -8.222 (1.617) (5.607) NO2 2.240** (0.883) Search terms factors Labor 1 -1.507** (0.719) Labor 2 0.492 -2.197 (0.755) (1.448) Food and assistance -1.632** -1.001** 2.637** (0.648) (0.401) (1.218) Education and sports 0.880 (0.791) Religion 1.934 (8.837) Macro variables Log (GDP per capita 2019, ppp) 1.236* 1.783*** -0.089 (0.690) (0.468) (1.562) Log (population 2019) -0.091 0.052 2.172*** (0.316) (0.298) (0.495) % Primary school 0.056 0.040 0.234*** (0.047) (0.042) (0.077) 33 r-squared 0.486 0.495 0.680 adjusted r-squared 0.397 0.418 0.592 Observations 62 62 61 NOTES: OLS regression results with LASSO-selected controls. Robust standard errors are in parentheses. * 0.1 ** 0.05 *** 0.01. 34 Table 3: R2 decomposition of LASSO-Selected Features (% of the overall r-squared) (1) (2) (3) Growth rate Growth rate (actual) (deviation from Stringency Index predicted) Price factors 20.08 23.83 22.71 Google mobility factors 52.49 63.37 55.65 Satellite-derived data 2.84 3.86 Search terms factors 18.85 10.1 13.77 Macro variables 5.74 2.7 4.01 NOTES: The table shows the contribution of each group of real-time indicators to the overall goodness of fit of each regression using Shorrocks–Shapley decomposition of r-squared. 35 Table 4: Country Panel Regressions (1) (2) (3) (4) R2 Growth rate (actual) decomposition b/se % Price factors Factor 1 -0.001 -0.001 0.002 10.1 (0.005) (0.006) (0.006) Factor 2 0.001 0.005* 0.006** 0.67 (0.004) (0.003) (0.003) Google mobility factors Factor 1 0.024*** 0.013** 0.012** 22.88 (0.004) (0.005) (0.005) Factor 2 -0.006 -0.005 -0.005 7.11 (0.005) (0.006) (0.006) Satellite-derived data NO2 0.001 0.001 0.001 0.04 (0.001) (0.001) (0.001) Nightlights 0.003 -0.000 (0.013) (0.008) Search terms factors Food and assistance -0.006 -0.046** -0.039** 2.28 (0.004) (0.018) (0.018) Labor 1 0.001 0.020 (0.008) (0.027) Labor 2 -0.005 -0.021 -0.009 0.63 (0.006) (0.016) (0.009) Education and Sports 0.005 -0.019 -0.009 1.95 (0.006) (0.017) (0.013) Religion 0.057 0.169 (0.040) (0.181) Quarter dummies Quarter 2 -0.073*** -0.077*** -0.072*** 25.12 (0.015) (0.015) (0.014) Quarter 3 -0.037*** -0.035*** -0.032*** 2.74 (0.012) (0.011) (0.011) Quarter 4 -0.014** -0.021*** -0.023*** 1.55 (0.006) (0.007) (0.007) Country FE NO YES YES Lasso-selected features NO NO YES 36 r-squared 0.671 0.885 0.882 adjusted r-squared 0.639 0.827 0.828 Observations 156 156 156 NOTES: Columns (1)–(3) show the OLS regression results with LASSO-selected controls. Robust standard errors are in parentheses. * 0.1 ** 0.05 *** 0.01. Column (4) shows the contribution of each group of real- time indicators to the overall goodness of fit of each regression, using Shorrocks–Shapley decomposition of r-squared. 37 Appendix A Figure A1 – Global price of rice, 2015-2019 38 Table A1 – Countries included in the cross-section analysis A. High income countries (34 countries) ARG, AUS, AUT, BEL, CAN, CHE, CHL, CZE, DEU, DNK, ESP, EST, FRA, GBR, GRC, HRV, HUN, IRL, ISR, ITA, JPN, KOR, LTU, LVA, MLT, NOR, NZL, POL, PRT, SAU, SVK, SVN, SWE, URY B. Upper middle income countries (18 countries) ARG, BGR, BLZ, BRA, COL, DOM, ECU, GEO, GTM, KAZ, MEX, MYS, PER, RUS, SRB, THA, TUR, ZAF C. Lower middle income countries (8 countries) GHA, IND, KHM, LAO, LKA, MAR, PHL, VNM C. Low income countries (1 countries) AFG 39 Table A2 – Countries included in the panel analysis A. High income countries (29 countries) AUS, AUT, BEL, CAN, CHE, CHL, CZE, DNK, ESP, EST, FRA, GBR, GRC, HUN, IRL, ISR, ITA, JPN, KOR, LTU, LVA, NOR, NZL, POL, PRT, SAU, SVK, SVN, SWE B. Upper middle income countries (9 countries) ARG, BGR, BRA, COL, IDN, MEX, RUS, TUR, ZAF C. Lower middle income countries (1 countries) IND C. Low income countries (0 countries) 40