Policy Research Working Paper 10660 The Impacts of Disasters on African Agriculture New Evidence from Micro-Data Philip Wollburg Yannick Markhof Thomas Bentze Giulia Ponzini Development Economics Development Data Group January 2024 Policy Research Working Paper 10660 Abstract Disasters affect millions of people each year and cause adverse climatic events. In the countries and time period economic losses worth many billions of dollars globally. analyzed, these losses reduced total crop production by an Reporting on disaster impacts in research, policy, and average of 29 percent. Importantly, many of these losses are news primarily relies on macro statistics based on disaster underreported or undetected in key disaster inventories and inventories. The macro statistics suggest that a relatively therefore elude macro statistics. In the case of droughts and small share of disaster damages accrues in Africa. This floods, the economic losses recorded in the micro-data are paper, instead, uses detailed survey micro-data from six $5.1 billion higher than in the macro statistics, affecting African countries to quantify disaster damages in one key 145 million to 170 million people, more than four times as sector: crop agriculture. The micro-data reveals much many as the macro statistics suggest. The difference stems higher damages and more people affected than the macro mostly from smaller and less severe but frequent adverse statistics would indicate. On average, 36 percent of the events that are not recorded in disaster inventories. agricultural plots in the sample suffer crop losses due to This paper is a product of the Development Data Group, Development Economics. 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 pwollburg@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 The Impacts of Disasters on African Agriculture: New Evidence from Micro- Data Philip Wollburg1,2, Yannick Markhof1,3, Thomas Bentze1, Giulia Ponzini1 JEL codes: Q54, N57, O13, Q15, C81, C83 Keywords: agriculture, climate change, disaster risk, survey data, Africa, loss and damage 1 Development Economics Data Group, World Bank; 2 Wageningen University & Research; 3 United Nations University, UNU-MERIT. This paper received funding support from the 50x2030 Initiative to Close the Agricultural Data Gap and the World Bank Research Support Budget grant “On the Measurement of Agricultural Productivity Trends in Africa”. The authors are grateful to Douglas Gollin, Erwin Bulte, Gero Carletto, Ruth Hill, Stephane Hallegatte, Talip Kilic, Travis Lybbert and participants of the CSAE 2023 Conference, the ICAS IX Conference, the EAAE 2023 Congress, and at the World Bank for their comments and feedback. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Introduction In 2022, natural disasters led to over $220 billion in economic losses, affecting 185 million people.1 Losses in 2023 are on track to exceed the previous year’s2 and large-scale disasters, such as record extreme heatwaves, the violent monsoon in India, and a prolonged severe drought in the Horn of Africa have received widespread media and public attention.3–6 The frequency and intensity of disasters and their impacts has increased over the last decades, a trend that is set to continue, and likely accelerate, due to climate change and global warming.7–12 Reporting on disaster impacts relies predominantly on macro statistics. A key data source is the Emergency Events Database (EM-DAT), which is a publicly available global inventory of disaster impacts that is widely used in media,13 research,14 and policy reports, including recently the World Bank’s 2023 ‘Atlas of the Sustainable Development Goals’ and the Food and Agriculture Organization’s (FAO) 2021 report on ‘The impact of disasters and crises on agriculture and food security.’15,16 Here, we offer a different approach to studying disaster impacts, based on survey micro-data. We quantify the value of crop production losses due to adverse climatic events on more than 120,000 fields across six African countries and study the impacts of these events on African agriculture, rural populations, and the national economies. Agriculture is a key sector, on which many households in the region depend for their livelihoods, especially the poor and rural households.17 Agriculture dependent households are thought to be particularly at risk of suffering the impacts of climate change and adverse shocks. Climate change and natural disasters are expected to be especially severe in rural areas in this region,18,19 while smallholder agricultural production remains predominantly rainfed and the adoption of drought or heat resistant seeds or other such climate-smart technologies is limited.20 We document that crop losses due to adverse shocks are common and costly both to individual farm households and to the economy at large, and that farmers often suffer multiple shocks in the same season. Taken together, production losses have a substantial aggregate impact. Importantly, these events and their impacts are underreported or undetected in common macro data sources for disaster reporting, such as EM-DAT. Our analysis of micro-data offers an important complementary perspective to analyses based on aggregate statistics derived from disaster inventories. Aggregate statistics are critical to the study of disaster impacts, providing annual data at a global scale. They are less well-suited to capture the differential impacts of disasters on different population groups, especially poor and vulnerable people.21– 23 This is because they account primarily for damages to assets and losses in agricultural production whose value is greater and better documented among richer households and in richer countries. For instance, according to the most recent estimates of EM-DAT, about 70% of economic losses due to disasters occurred in the Americas, compared to just under 4% in Africa.24 A recent study using the same data source concluded that disaster impacts do not affect poor people as much as the general population.25 In contrast, evidence from survey micro-data suggests that poorer households and individuals are more exposed and less resilient to adverse climatic and environmental shocks and suffer disproportionately greater well-being losses than better-off households.18,22,26 Our analysis suggests that production losses due to adverse climatic events are meaningful not only for the well-being of low-income households individually but, because of how many households are affected, they are significant also for the whole economies of our study countries and on a global scale. 2 Results Crop losses are widespread and significant The data used in this analysis is from the Living Standards Measurement Study-Integrated Survey on Agriculture (LSMS-ISA) in Ethiopia, Malawi, Mali, Niger, Nigeria, and Tanzania. These data were harmonized across countries and cover close to 120,000 fields on around 30,000 farms. The data show that crop losses due to disasters and adverse climatic events are widespread and significant in African smallholder agriculture. Farmers report crop losses on between 11% (Nigeria 2018/19) and 90% of plots (Niger 2011), depending on country and year (Figure 1, Panel A and Table 1). Overall, 36% of plots in our sample report a crop loss. Farmers reported losing, on average, 53% of their harvest on plots affected by crop shocks (Panel B and Table 1). Losses vary across countries and years, ranging from 48% of harvest (Ethiopia 2018/19) to 71% of harvest (Niger 2011). Disaster losses have also become more common over time (Table 3). In the 11 years from 2008 to 2019 that our dataset spans, the estimated likelihood of a plot incurring a disaster loss increased by close to 10 percentage points. This is seemingly driven by a higher prevalence of small shocks as the estimated share of harvest lost on plots with any loss decreased by on average 1.1 percentage points with every year studied. In aggregate, crop losses due to adverse climatic events reduce the total national crop production by between 3% in Nigeria 2018-19 and 81% in Niger in 2011. A total of 29% of potential harvest value is lost across the countries and agricultural seasons observed in our dataset (Figure 1, Panel C and Table 4]). 3 Figure 1. Panel A displays the prevalence of crop shocks on plots across country-waves. Panel B displays the mean percent of potential harvest value lost on plot, by country-wave, as well as the fraction of aggregate potential harvest lost (valued with current prices), per country-wave A B Crop production is impacted by multiple shocks Farmers face a diversity of adverse climatic shocks. Multiple shocks are recorded to affect agricultural production in each year and across all countries (Table 1). There are also some instances of multiple shocks 4 affecting the same farm in a given agricultural season (Table 5). This ranges from 1.5% of farms (Tanzania 2014) to 21% of farms (Ethiopia 2018-2019). Overall, drought is the most common shock, with 22% of plots in our sample recording a crop loss due to drought (Table 1). One in ten plots records losses due to irregular rains, meaning erratic rainfall at unusual times in the agricultural season. Pests are also widespread across our sample, affecting 6.3% of all plots. Still, there is substantial variation across countries and years. The severity of the damages caused varies between different events (Table 2). Floods in particular cause more damage than other shocks, reducing crop production per plot on average by 62%. Losses from pests and irregular rains tend to be smaller. However, there is again some variation between different countries and farming environments (Table 6). Which shocks are the most prevalent varies also within countries. Figure 2 illustrates this for selected countries and years, showing the most reported events by subnational administrative divisions. There is some geographical clustering, but we commonly see different events accounting for most of the impacted plots in different areas of the same country in the same year. This is true even in years with exceptionally severe disasters such as the droughts in Niger in 2011 and Ethiopia in 2015-16 where the vast majority but not all areas of the country recorded drought as the primary loss reason. Figure 2. Most common disaster events by administrative unit, selected countries and years 5 Crop losses differ locally and between farmers Not all farmers and plots are equally affected. Some are less likely to experience a loss even in the face of an adverse climatic event. Here we show that shock exposure and impacts can differ even between neighboring plots in the same area. We limit this analysis to droughts. Given the nature of droughts, all plots in the same small geographic cluster should be faced with the same drought shock – but the impacts of that drought can differ. Indeed, in 41% of the geographical clusters in our sample, some but not all plots report being affected by a drought (Table 12). This finding holds also for plots growing the same crops. In 31% of clusters, some but not all maize plots suffer drought losses (32% for sorghum plots and millet plots). The result extends to plots with the same crops cultivated by the same households (Table 13): for farms that record a drought shock on one of their maize plots, close to two-thirds of other maize plots on the same farm also record drought-related crop losses. These findings suggest that disaster impacts are highly localized, consistent with the high spatial concentration that meteorological events can have.28 Further, idiosyncratic factors, such as land characteristics and management practices, and happenstance play a role in determining whether and how much production is affected. We find that plot elevation is negatively associated with the likelihood of experiencing losses and the size of the losses incurred (an effect almost twice as strong for floods compared to other disasters), while smaller plots are less likely to suffer losses but record higher losses when they are affected (Tables 7 and 8). Losses on intercropped plots are 7.5 percentage points lower than on mono-cropped plots, though intercropped plots are more likely to experience a loss in the first place (+3.6 percentage points). Plots farmed in more input and technology intensive ways appear more resilient to crop losses due to adverse events. Disaster exposure and impact also vary according to who manages the plot. Plots managed by women are more often affected by disaster losses (+2.2 percentage points) than plots managed by men and these losses are also larger on average (+4.4 percentage points; Table 10). These differences are likely because plots managed by women are endowed and farmed differently than plots managed by men, which in turn may follow from differential access to inputs and land between women and men.27 Aggregate data sources underestimate impacts of extreme events on crop production How do disaster impacts as captured in the survey data compare to estimates from other commonly used data sources? Here, we contrast the results from the survey microdata with publicly available estimates of disaster impacts from the Emergency Events Database (EM-DAT). EM-DAT aggregates reports from UN agencies, governments, insurance companies, research institutes and the media into a global inventory of disaster impacts.29 EM-DAT is the preeminent and only publicly available data source of this kind, used widely in disaster reporting and research.30 We focus on two disaster types, droughts and floods, and compare two estimates: the number of people affected and the total economic damages caused in the years which the survey micro-data covers. We create aggregate figures from the micro-data using population sampling weights. On both metrics, and for both drought and flood impacts, the micro-data estimates on average exceed the EM-DAT estimates, that is, for years in which there is information from both sources, the survey micro- data find more people affected and higher damages from droughts and floods (Figure 3 and Tables 14 and 6 16). Moreover, there are many instances in which the EM-DAT records no disaster impacts at all. This is true especially for droughts, where the microdata suggests that droughts are prevalent to some degree across every country-year combination covered, while EM-DAT records droughts affecting the population in only a third of cases. Estimates of the economic value of disaster impacts are mostly missing in the EM- DAT data for the study countries, even in years when drought and flood events were recorded to affect the population in the study countries (Tables 15 and 17). Large, salient drought and flood episodes have better coverage in the EM-DAT, such as the severe droughts in Niger in 201131,32 and Ethiopia in 2015-16,33,34 or the droughts and floods Malawi in 2015- 16,35,36 which were widely covered in international media at the time. The events that go unreported in EM-DAT are smaller, on average, in terms of the population affected and the damages caused. However, we show that such smaller, under-covered events have substantial overall impacts. More than a fifth of the population suffered production and income losses in the droughts in Malawi in 2009-2010 and in Mali in 2074, according to our micro-data estimates, while there is no coverage of these events in the EM-DAT for the same years. Overall, we estimate the total number of people affected by droughts or floods in all instances covered by the microdata is between 145 million and 170 million, more than 4 times higher than what is reported in the EM-DAT for the same periods and the same shocks. The micro-data analysis suggests that the aggregate value of the disaster impacts on crop production is substantial. For the drought in Ethiopia in 2015-2016, the micro-data crop loss estimates are much larger than the total economic damage reported in the EM-DAT. For the 2014 floods in Niger, the estimated value of crop losses exceeds the total damage reported in the EM-DAT data by almost USD 78 million (in 2022 USD values). In the other years there is no damage estimate in EM-DAT, but our survey micro-data documents even some large disaster impacts, such as in Niger in 2011 and Ethiopia in 2018-2019 with estimated losses of USD 1.6 billion and USD 1.4 billion, respectively. Taken together, we estimate that across the countries and years captured in the microdata, there were USD 5.1 billion in drought and flood damages unaccounted for in the EM-DAT data (Table 18). What explains these discrepancies? Disaster inventories such as EM-DAT and survey microdata differ in a number of meaningful ways. Most importantly, disaster inventories do not measure shock impacts themselves but instead aggregate data from government sources, humanitarian organizations, the media, and others. They therefore rely on the comprehensiveness and accuracy with which shocks due to natural hazards are covered by one or more of these sources.30,37 Less salient events as well as those affecting marginalized population groups are less likely to be reported on and less likely to have detailed information on the affected population or economic and welfare impacts.37–40 This is particularly acute in LMICs where the density of information for disaster repositories to draw on is much lower and a large share of damages is uninsured.30,37 Shocks in LMICs in general and smaller events (in terms of intensity or the population affected) in particular are more likely to have incomplete or inaccurate information in disaster repositories or are not covered at all.30,37,41,42 Microdata such as the LSMS-ISA measure shock impacts on smallholder farmers where they occur by asking farmers directly. They therefore do not suffer from the same limitations regarding the recording of smaller, less salient, or more localized shocks and their impacts as disaster repositories. Smaller shocks or adverse climatic events may not be considered disasters as disasters suggest a minimum level of severity. For an event to be recorded in the inventory, the EM-DAT requires a minimum of 100 people to be affected (injured, homeless, in need of immediate assistance) or an official declaration of 7 emergency or appeal for international assistance – arguably a sensible set of criteria for a disaster inventory. Not all events recorded in the micro-data meet these requirements. Importantly, the events recorded in the micro-data have substantial impacts on the livelihoods of farmers and the economies of the study countries. At the same time, micro-data has drawbacks and limitations. First, it is rare that microdata in low- and middle-income countries are available annually, with surveys typically implemented every few years. Shock coverage and detail depend on the survey design, which typically differs from country to country. Finally, microdata does not provide the same cross-country coverage as disaster repositories. With these limitations, the microdata naturally also provides an incomplete picture (see discussion in Appendix A). 8 Figure 3. Comparison of shock prevalence and impact between EM-DAT and LSMS-ISA data. A B Percent affected in EM-DAT Percent affected in LSMS Total damages in LSMS (with 95% CI) Total damages in EM-DAT (with 95% CI) C D Percent affected in EM-DAT Percent affected in LSMS Total damages in LSMS (with 95% CI) Total damages in EM-DAT (with 95% CI) Note: Panel A displays a comparison of the total estimated individuals affected by droughts between EMDAT (in blue) and LSMS-ISA data (in orange), while panel B shows a comparison of the estimated damages (in millions of 2022 dollars), in years where damages could be estimated in the LSMS-ISA surveys. Panel C displays a similar comparison for floods, in years where floods are listed as a potential shock in the LSMS-ISA data, while panel D shows a comparison of estimated damages from floods. Confidence intervals for panels B and D were calculated before log- transformation, and are hence asymmetrically situated around log-scaled point estimates. Discussion We explore the crop production impacts of adverse climatic events on 120,000 fields on 30,000 smallholder farms in Sub-Saharan Africa. Smallholder agriculture is of special interest for achieving SDGs 1 and 2 as it remains the primary means of livelihood for many of the world’s poor.43 Our findings generate new insights and advance our understanding of the disaster risks and losses that smallholder farmers face. Other studies have investigated the vulnerability of smallholder farmers to 9 disasters and environmental shocks.44–47 These studies have mostly focused on single geographies and stopped short of quantifying the value of disaster related losses in smallholder agriculture. Other studies have relied on macro-data from disaster inventories to assess the impact of disasters on agriculture.14 Here we offer a cross-country perspective using harmonized survey micro-data from six African countries. We value crop production losses to assess the economic importance of disaster impacts on rural households and African agriculture and the regional economies more broadly. The analysis shows that disaster related crop production losses among African smallholder farmers are widespread and significantly reduce the production of affected farms. In any given year, between 18% and 94% of crop- farming households suffer such crop losses and on average 53% of plots’ harvest potential is lost when they are hit by adverse shocks. In aggregate, disaster impacts reduce the national crop production by 29% every year on average and by up to 81% in years with large-scale disasters. These results show that crop production losses due to adverse climatic events are significant both for individual farms and for the entire sector and the study countries’ economies. We show that the EM-DAT disaster inventory misses out on a meaningful share of disaster impacts in the agricultural sector in Africa when compared to the micro-data analysis. The micro-data captures many smaller, more localized disaster events, which are less salient and therefore less likely to be reported on and register in the disaster inventory.37,48 At the same time, disaster impacts are more likely to be missing entirely in lower-income settings.30 Our analysis shows that less salient and underreported disaster events still have significant economic impacts. The findings have implications for policies and interventions aimed at disaster risk reduction and resilience building. For such policies to be effective, it is important to recognize the risks from less salient and under- reported adverse events and offer ways to insure households’ livelihoods against their impacts. The findings also have implications for research and measurement. Research and analysis using aggregate global data such as EM-DAT are likely missing some disaster impacts in poorer countries and among poorer people. This concerns also tracking progress towards the SDGs, in particular SDG target 1.5, which seeks to reduce the vulnerability of the poor to extreme weather events, and SDG target 13.1, which seeks to strengthen resilience and adaptive capacity to climate-related hazards and natural disasters in all countries.15 Survey micro-data such as the LSMS-ISA have more limited country and temporal coverage. Inventory and survey data offer complementary perspectives on disaster impacts and combining both sources will likely yield a more complete and nuanced understanding of the issue that will promote more effective policy design. In particular, information on disaster impacts from survey micro-data could be incorporated systematically into disaster inventories. Improving micro-data systems is key to systematically utilizing micro-data for monitoring and reporting of emergency events. More flexible and higher frequency data collection is needed to provide better temporal coverage and account for disaster impacts when they occur. Phone surveys, which have only recently become more widely adopted in low- income settings, may provide this function, for instance as part of mixed-mode survey systems that combine traditional in-person surveys with data collection over the phone.49 Integration of geospatial data can improve spatial coverage and facilitate better identification of natural hazards and disaster occurrence.50,51 Sampling protocols of household surveys can be optimized to better capture disaster impacts and greater harmonization of survey methods and measurement instruments in line with best practices can benefit data quality. 10 Our study faces several limitations. The valuation of losses relies on human reporting which is subject to human error, respondents’ incentives, misreporting, and misperceptions. The data allows for a detailed analysis of disaster impacts on crop production, but other aspects of disaster impacts on agriculture are absent. This concerns, for example, damages to agricultural assets, storage losses, or impacts on livestock. Further, the data is representative of the household sector in study countries but misses commercial and larger farms. This implies that we are likely underestimating the full extent of disaster crop losses. As the analysis is focused on agriculture, we do not discuss damages incurred in other sectors. Methods A. LSMS-ISA microdata We use plot-level survey data from the Living Standards Measurement Study – Integrated Surveys on Agriculture (LSMS-ISA) in Ethiopia, Malawi, Mali, Niger, Nigeria, and Tanzania. The LSMS-ISA comprise a series of harmonized, national, multi-topic household panel surveys with a distinct focus on agriculture. We harmonized data on agricultural outputs, inputs, and plot characteristics for close to 30,000 households over the six countries for a total of 18 survey waves collected between 2008 and 2019. The combined dataset contains over 120,000 plot observations. More specifically, the dataset includes data from the Ethiopian Social Survey (waves 1 to 4), Malawi’s Integrated Household Panel Survey (waves 1 to 4), Mali’s Enquête Agricole de Conjoncture Intégrée (waves 1 and 2), Niger’s Enquête National sur les Conditions de Vie des Ménages et Agriculture (waves 1 and 2), Nigeria’s General Household Survey (wave 4) and Tanzania’s National Panel Survey (waves 1 to 5). Households are selected to be representative of the population at the national and sub-national level using a two-stage stratified sampling design with census enumeration areas (EAs) as primary sampling units and households as secondary sampling units. Households are then tracked through time, except for Mali which only tracks enumeration areas (EAs). Each survey wave covers an agricultural production cycle or season. B. EM-DAT aggregate data We compare survey estimates to country-level data from is the Emergency Event Database (EM-DAT). EM- DAT is the preeminent public database taking stock of shocks on a global scale and widely used for research and to inform policy52. Both natural (e.g., geophysical, meteorological) and technological (e.g., industrial accidents) events are recorded, along with information on disaster damages valued in 2022 USD. The EM-DAT compiles information from a broad range of sources including insurance companies, international organizations, press agencies and governmental agencies. Disasters are recorded if they provoke 10 or more deaths, affect 100 or more people (injured/ homeless/ in need of immediate assistance) or are accompanied by an official declaration of emergency or appeal for international assistance.29 C. Variable construction Valuation of crop production Plot-level crop production was calculated using survey variables which were collected after the harvest in each season and country. Farmers report harvest quantities for each seasonal crop grown on each plot. 11 The harvest quantity on each plot was then valued using a set of constant crop prices. Specifically, each price corresponds to the median crop sale price calculated in one survey round in each country. Applying constant prices for each crop-country combination eliminates the effect of relative price fluctuations over time. This allows us to isolate the impact of disasters on harvest quantity rather than measuring their impact on quantity and prices. An alternative approach was used to calculate and report aggregate losses, such as in Table 4 and Figure 3. In these cases, country and wave specific prices were calculated, by estimating median farmer sale prices for different crop types. This was done to reflect losses as they were perceived by farmers in the year of the shock. Whether current or constant, harvest values are initially calculated in local currency units and then converted and adjusted to 2020 USD using exchange rates and a CPI drawn from a library of the World Bank World Development Indicators. 53 Yields, defined as the value of production per hectare, were obtained by dividing the total harvest value on the plot by GPS-measured plot area. Finally, yields are winsorized at the 1st and 99th percentiles (while allowing full losses to be equal to 0). Identification of disaster events and loss size Identification of disaster events is based on farmers reporting crop production losses before harvest for each crop on each of their plots. Specifically, for each crop cultivated on each plot, farmers are asked whether the area harvested was less than the area planted, i.e. if some of their crop has been lost, along with the cause of the loss in harvested area. In Ethiopia, farmers are further asked whether the crops they harvested had any damage on them and what the cause of damage was. We define disaster crop losses as any loss in crop area or any damage on the crops harvested due to climatological (drought, irregular rain, hail, wildfire), hydrological (flood) or biological (insect infestation, disease) reasons. This does not include losses due conflict, unavailability of inputs or other household or socio-economic events. We denote a plot with a disaster loss as any plot for which at least one crop on the plot had a disaster-induced loss. To calculate the size of disaster losses, we use farmers’ reports of the share of the planted area lost due to disasters and, where available, the percent of damage on crops that were harvested. To determine the potential harvest that could have been achieved in the absence of disaster losses, we employ the methodology proposed by the Food and Agriculture Organization (FAO).54,55 This means determining the realized harvest for each crop on the plot, as described in the preceding section, and scaling it up in proportion to the planted area lost and degree of damage on the crop that is attributed to disasters. The quantity lost equates to the difference between the potential harvest without disaster losses and the realized harvest. To aggregate losses across crops, we value loss quantity of each crop on each plot using a constant set of prices, which we then convert to 2020 US dollars. In order to correct for outliers in the self-reported data, we winsorize these loss values at the 99th percentile. Equation (1) formalizes the construction of the plot-level loss aggregate. 12 , = � × �,, − ,, � =1 1 1 =� × �� ,, × × � − ,, � =1 1 − ,, 1 − ,, 1 =� × � ,, × � − 1�� (1) =1 (1 − ,, )(1 − ,, ) Where the value of harvest loss on plot i in agricultural season s is equal to the difference between the potential harvest in the absence of disasters, and the realized harvest reported by the farmer . The potential harvest is calculated by scaling up realized harvest in proportion to the share of the planted area lost to disasters, ,, , and the percent of damage on the crop, ,, . The quantity lost is then valued in terms of the median selling price for crop j in the country p and aggregated at the plot level across all crops on the plot. Imputation of full losses In case the harvest for a crop on a plot is fully lost (i.e., ,, = 1 or ,, = 1), Equation (1) is not defined. Instead, we estimate the quantity lost in these cases by imputing potential harvest values using a Gaussian normal regression imputation method.56 To this end, we define a model where potential harvest is the outcome variable, regressed on the set of explanatory variables, along with country and crop fixed effects. The explanatory variables used in the imputation are the following: (i) agricultural input variables, specifically, plot area, non-hired labor days spent working on the plot (e.g., family labor), as well as hired labor value, inorganic fertilizer value and seed value. In similar fashion to the production values described above, a constant set of prices was computed within each country, based on median purchase prices. These input variables are all expressed in per hectare terms, winsorized and logged; (ii) an agricultural asset index was computed using a principal component analysis based on an inventory of household assets; (iii) plot-level dummy variables were included to indicate if a plot is irrigated, pesticides are used, organic fertilizers are applied, the plot is intercropped, and if the plot is owned by the household; (iv) gender of the primary decisionmakers on each plot; (v) household-level variables household size and dummies for livestock ownership, electricity access and urban/rural residence; (vi) finally, a set of geophysical variables consisting of plot elevation, a topographic wetness index, and the distance of the household from the closest population center and closest road. Our final imputed value is obtained by calculating the mean of 100 imputations. Table 19 contains an overview of the available data and variables for each country and agricultural season covered. D. Estimation i. Disaster crop losses in Sub-Saharan Africa Our main descriptive analysis of disaster prevalence and intensity is conducted at the plot-level and involves the estimation of means, proportions, and frequencies at the national level as well as pooled across countries. These estimates, as well as any household-level estimates of disaster exposure formed 13 by aggregating across plots belonging to the same farm, are weighted using the probability weights described in a separate section below (Sampling weights). ii. Disaster type and frequency Similarly, our estimates of the prevalence of different shock types is conducted at the plot-level and involves the estimation of frequencies at the national level and pooled across countries using the household sampling weights. Our estimates of the most common shock type within enumeration areas are based on simple, unweighted frequencies. iii. Heterogeneity in disaster impacts across farms Our multivariate analysis focuses on two main outcome variables: a binary variable indicating any disaster crop loss and a continuous variable denoting the percentage share of the total potential harvest that was lost to disasters. We estimate all models with the binary crop loss indicator as outcome variable via maximum likelihood using logistic regression. Models with the percent share of harvest lost as outcome variable are estimated via ordinary least squares regression. Our independent variables for these multivariate regressions are comprised of plot characteristics (plot size, elevation, a topographical wetness index, an indicator for ownership of the plot, and main crop fixed effects), as well as plot management (hired labor and fertilizer input use, irrigation, intercropping), plot manager (age, gender, and education), and household characteristics (urban/rural residence, an indicator for livestock farming, and electricity access). Models pooling the sample across countries further include country fixed effects. We also conduct multivariate analysis using a binary variable capturing the gender of the plot manager as outcome variable and plot- and plot-management characteristics as independent variables. As before, all multivariate regressions are weighted using the sampling weights. iv. Different disaster impacts on neighboring plots Our analysis of differences in drought impact within the same enumeration areas first determines whether some but not all plots belonging to the same enumeration area recorded a drought loss and then estimates the simple, unweighted proportion of enumeration areas for which this is the case. Our analysis of within-household differences in shock impacts first limits the sample to households with multiple maize plots and where at least one of the household’s maize plots recorded a drought shock. We then calculate the share of remaining maize plots belonging to the same household that also record a drought loss and report a simple, unweighted average of this share across households for each country and year. v. Comparison of survey micro data with aggregate sources Our estimates of the number of people affected by disasters and aggregate economic losses are totals at the national level and employ the household-level sampling weights. 14 In order to the compare drought and flood impacts using LSMS-ISA data with those using EM-DAT data, we use two metrics: the share of individuals “affected” by the shock and the estimated total value of damages. Since the LSMS-ISA surveys run every two to three years, we only retain events in the EM-DAT database for which the start or end date is within a year containing LSMS-ISA data. The comparison for flood shocks is possible in fewer countries and years due to limitations in the microdata questionnaire’s scope in some cases (Table 18). In order to compute the total number of individuals impacted by a shock within a specified period in the EM-DAT database, we aggregate the total number of people “affected” by the shock in the macro data. Affected persons are those that are reportedly injured, homeless, or otherwise in need of “immediate assistance.”29 To estimate the total number of individuals affected by a shock in the LSMS-ISA microdata, we construct population weights by multiplying household weights by household size. These weights are then used as expansion factors, which we multiply by a dummy variable equal to one in the household report a shock on any of its cultivated plots. We then add up this product to calculate an expansion estimator.57,58 To obtain shares of the total population, the numbers of individuals affected in both EM-DAT and the LSMS-ISA data are divided by total yearly population estimates drawn from a library of the World Bank World Development Indicators.53 We then compute the estimated damages from both droughts and floods in the periods and years covered by LSMS-ISA data. We first aggregate the “total damages” estimated in the EM-DAT database, defined as the values of total losses “directly or indirectly related to the disaster”, in 2022 USD values. Using LSMS- ISA microdata, we aggregate the estimated value of crop losses for each household. In this case, the value of losses is calculated by multiplying the potential value of total output using current prices by the estimated percentage of output lost at the plot level. Contrary to the rest of our analysis, we do not apply a common set of prices to value output and losses, but rather median crop prices faced by farmers at the time of their crop production. The loss values are then converted to 2022 USD values, in order to allow comparison with EM-DAT data. As above, we use population weights as expansion factors, that we multiply with our loss value estimates.57,58 Sampling weights In the survey data, household sampling weights are used to compute estimates that are representative of the national or subnational population. These reflect the inverse probability of selection into the sample, are adjusted to account for non-response and survey design choices, and are post-stratified to ensure that they sum to known household population totals.59 Moreover, a further adjustment was made to ensure that the weights in the study sample sum to the total population of households, because only a subset of LSMS-ISA households report cultivating seasonal crops. More formally, we can define a set of households indexed by ℎ (where ℎ = 1, … , 1 ) within a country-wave, where each household is associated with a sample weight ℎ . After restricting the dataset to households that cultivate crops, the total number of households in the country-wave drops from 1 to 2 , and weights are adjusted such that: 15 1 ℎ ∑ ℎ=1 ℎ ℎ = × 2 ℎ ∑ℎ=1 ℎ Where ℎ denotes the number of plots in household ℎ on which seasonal crops are grown, and ℎ is the final adjusted weight. The resulting weights were used to compute estimates in our analysis. Wherever population means are estimated, the standard errors provided with the estimate takes into account the clustered and stratified sampling design.57 All regression models with continuous variables as outcomes are estimated using ordinary least squares regression, while regressions with binary outcomes are estimated via maximum likelihood using a logistic regression model. References 1. CRED. 2022 Disasters in numbers. https://cred.be/sites/default/files/2022_EMDAT_report.pdf (2023). 2. Lörinc, M. & Hotový, O. Global Catastrophe Recap. First Half of 2023. https://www.aon.com/getmedia/760ea02e-ce76-4348-800a-cb8d99ca2a8f/20230720-1h-2023- global-cat-recap.pdf (2023). 3. Paddison, L. Catastrophic drought that’s pushed millions into crisis made 100 times more likely by climate change, analysis finds. CNN (2023). 4. Zachariah, M. et al. Extreme heat in North America, Europe and China in July 2023 made much more likely by climate change. http://spiral.imperial.ac.uk/handle/10044/1/105549 (2023) doi:10.25561/105549. 5. UN OCHA. Horn of Africa Drought Regional Humanitarian Overview & Call to Action. https://reliefweb.int/report/ethiopia/horn-africa-drought-regional-humanitarian-overview-call- action-revised-26-may-2023 (2023). 6. Visualising India’s record-breaking rainfall. The Economist (2023). 7. United Nations Office For Disaster Risk. Global Assessment Report On Disaster Risk Reduction 2022 - Our World At Risk: Transforming Governance For A Resilient Future. (UN, 2022). 16 8. Agnolucci, P. et al. Impacts of rising temperatures and farm management practices on global yields of 18 crops. Nat. Food 1, 562–571 (2020). 9. Deryng, D., Conway, D., Ramankutty, N., Price, J. & Warren, R. Global crop yield response to extreme heat stress under multiple climate change futures. Environ. Res. Lett. 9, 034011 (2014). 10. Gourdji, S. M., Sibley, A. M. & Lobell, D. B. Global crop exposure to critical high temperatures in the reproductive period: historical trends and future projections. Environ. Res. Lett. 8, 024041 (2013). 11. Barlow, K. M., Christy, B. P., O’leary, G. J., Riffkin, P. A. & Nuttall, J. G. Simulating the impact of extreme heat and frost events on wheat crop production: A review. Field Crops Res. 171, 109–119 (2015). 12. Teixeira, E. I., Fischer, G., Van Velthuizen, H., Walter, C. & Ewert, F. Global hot-spots of heat stress on agricultural crops due to climate change. Agric. For. Meteorol. 170, 206–215 (2013). 13. Ritchie, H. Opinion | 3 charts show how better buildings save lives in earthquakes. Washington Post (2023). 14. Lesk, C., Rowhani, P. & Ramankutty, N. Influence of extreme weather disasters on global crop production. Nature 529, 84–87 (2016). 15. Pirlea, A. F., Serajuddin, U., Wadhwa, D., Walsh, M. & eds. Atlas of Sustainable Development Goals 2023. (World Bank, 2023). 16. FAO. The impact of disasters and crises on agriculture and food security: 2021. (FAO, 2021). doi:10.4060/cb3673en. 17. World Bank. Poverty and Shared Prosperity 2020: Reversals of Fortune. (World Bank, 2020). doi:10.1596/978-1-4648-1602-4. 18. Park, J., Bangalore, M., Hallegatte, S. & Sandhoefner, E. Households and heat stress: estimating the distributional consequences of climate change. Environ. Dev. Econ. 23, 349–368 (2018). 19. Tol, R. S. J. The Economic Impacts of Climate Change. Rev. Environ. Econ. Policy 12, 4–25 (2018). 17 20. Arslan, A., Floress, K., Lamanna, C., Lipper, L. & Rosenstock, T. S. A meta-analysis of the adoption of agricultural technology in Sub-Saharan Africa. PLOS Sustain. Transform. 1, e0000018 (2022). 21. Hallegatte, S. & Rozenberg, J. Climate change through a poverty lens. Nat. Clim. Change 7, 250–256 (2017). 22. Hallegatte, S., Vogt-Schilb, A., Rozenberg, J., Bangalore, M. & Beaudet, C. From Poverty to Disaster and Back: a Review of the Literature. Econ. Disasters Clim. Change 4, 223–247 (2020). 23. Markhvida, M., Walsh, B., Hallegatte, S. & Baker, J. Quantification of disaster impacts through household well-being losses. Nat. Sustain. 3, 538–547 (2020). 24. Centre for Research on the Epidemiology of Disasters, United States Agency for International Development & Université Catholique de Louvain. 2021 Disasters in numbers - World | ReliefWeb. https://reliefweb.int/report/world/2021-disasters-numbers (2022). 25. Song, Z., Hochman, G. & Timilsina, G. R. Natural Disaster, Infrastructure, and Income Distribution : Empirical Evidence from Global Data. https://documents.worldbank.org/en/publication/documents- reports/documentdetail/099312406272326097/IDU018bdbef90c42104d730a46700d5b8c4c1c83 (2023). 26. Hallegatte, S., Vogt-Schilb, A., Bangalore, M. & Rozenberg, J. Unbreakable: building the resilience of the poor in the face of natural disasters. (World Bank Publications, 2016). 27. FAO. The status of women in agrifood systems. (Food and Agriculture Organization of the United Nations, 2023). doi:10.4060/cc5343en. 28. Philippon, N., Mougin, E., Jarlan, L. & Frison, P.-L. Analysis of the linkages between rainfall and land surface conditions in the West African monsoon through CMAP, ERS-WSC, and NOAA-AVHRR data. J. Geophys. Res. Atmospheres 110, (2005). 29. EM-DAT Guidelines. https://public.emdat.be/about. 18 30. Jones, R. L., Guha-Sapir, D. & Tubeuf, S. Human and economic impacts of natural disasters: can we trust the global data? Sci. Data 9, 572 (2022). 31. Rice, X. Severe drought causes hunger for 10 million in west Africa. The Guardian (2010). 32. World Bank. Niger Takes Action to Reverse 50 Years of Food Insecurity. World Bank News https://www.worldbank.org/en/news/feature/2011/09/01/niger-takes-action-to-reverse-50-years- of-food-insecurity (2011). 33. Munro, T. & Wild, L. As drought hits Ethiopia again, food aid risks breaking resilience. The Guardian (2016). 34. Jeffrey, J. Ethiopia drought: How can we let this happen again? Al Jazeera (2016). 35. Chonghaile, C. N. Malawi floods devastation far worse than first thought. The Guardian (2015). 36. Africa drought fears grip Malawi and Mozambique. BBC News (2016). 37. Gall, M., Borden, K. A. & Cutter, S. L. When Do Losses Count?: Six Fallacies of Natural Hazards Loss Data. Bull. Am. Meteorol. Soc. 90, 799–810 (2009). 38. Moriyama, K., Sasaki, D., Ono, Y., & International Research Institute of Disaster Science (IRIDeS), Tohoku University 468-1-S302 Aoba, Aramaki, Aoba-ku, Sendai, Miyagi 980-0845, Japan. Comparison of Global Databases for Disaster Loss and Damage Data. J. Disaster Res. 13, 1007–1014 (2018). 39. Panwar, V. & Sen, S. Disaster Damage Records of EM-DAT and DesInventar: A Systematic Comparison. Econ. Disasters Clim. Change 4, 295–317 (2020). 40. Wirtz, A., Kron, W., Löw, P. & Steuer, M. The need for data: natural disasters and the challenges of database management. Nat. Hazards 70, 135–157 (2014). 41. Osuteye, E., Johnson, C. & Brown, D. The data gap: An analysis of data availability on disaster losses in sub-Saharan African cities. Int. J. Disaster Risk Reduct. 26, 24–33 (2017). 19 42. Harrington, L. J. & Otto, F. E. L. Reconciling theory with the reality of African heatwaves. Nat. Clim. Change 10, 796–798 (2020). 43. Editorial. Ending hunger: science must stop neglecting smallholder farmers. Nature 586, 336–336 (2020). 44. Derbile, E. K., Chirawurah, D. & Naab, F. X. Vulnerability of smallholder agriculture to environmental change in North-Western Ghana and implications for development planning. Clim. Dev. 14, 39–51 (2022). 45. Chandra, A., McNamara, K. E., Dargusch, P., Caspe, A. M. & Dalabajan, D. Gendered vulnerabilities of smallholder farmers to climate change in conflict-prone areas: A case study from Mindanao, Philippines. J. Rural Stud. 50, 45–59 (2017). 46. Harvey, C. A. et al. Extreme vulnerability of smallholder farmers to agricultural risks and climate change in Madagascar. Philos. Trans. R. Soc. B Biol. Sci. 369, 20130089 (2014). 47. Zeleke, T., Beyene, F., Deressa, T., Yousuf, J. & Kebede, T. Vulnerability of smallholder farmers to climate change-induced shocks in East Hararghe Zone, Ethiopia. Sustainability 13, 2162 (2021). 48. Marulanda, M. C., Cardona, O. D. & Barbat, A. H. Revealing the socioeconomic impact of small disasters in Colombia using the DesInventar database. Disasters 34, 552–570 (2010). 49. Gourlay, S., Kilic, T., Martuscelli, A., Wollburg, P. & Zezza, A. Viewpoint: High-frequency phone surveys on COVID-19: Good practices, open questions. Food Policy 105, 102153 (2021). 50. Carletto, G. & Banerjee, R. Strengthening Disaster Resilience: A Microdata Perspective. in Capitalism, Global Change and Sustainable Development (ed. Paganetto, L.) 87–96 (Springer International Publishing, 2020). doi:10.1007/978-3-030-46143-0_6. 51. Carletto, G., Chen, H., Kılıç, T. & Perucci, F. Positioning household surveys for the next decade. Stat. J. IAOS 38, 1–24 (2022). 20 52. Jones, R. L., Guha-Sapir, D. & Tubeuf, S. Human and economic impacts of natural disasters: can we trust the global data? Sci. Data 9, 572 (2022). 53. World Bank. World Development Indicators (database). (2022). 54. Conforti, P., Markova, M. & Tochkov, D. FAO’s methodology for damage and loss assessment in agriculture. http://www.fao.org/documents/card/en/c/ca6990en (2020) doi:10.4060/ca6990en. 55. Markhof, Y. V., Ponzini, G. & Wollburg, P. R. Measuring Disaster Crop Production Losses using Survey Microdata: Evidence from Sub-Saharan Africa. http://elibrary.worldbank.org/doi/book/10.1596/1813-9450-9968 (2022) doi:10.1596/1813-9450- 9968. 56. Gelman, A. Bayesian data analysis. (CRC Press, 2014). 57. Heeringa, S., West, B. T. & Berglund, P. A. Applied survey data analysis. (Chapman & Hall/CRC, 2010). 58. Haziza, D. & Beaumont, J.-F. Construction of Weights in Surveys: A Review. Stat. Sci. 32, (2017). 59. Himelein, K. Weight Calculations for Panel Surveys with Subsampling and Split-off Tracking. Stat. Public Policy 1, 40–45 (2014). 21 Appendix Table 1: Frequency of plots recording a shock, by country-year Frequency of plots Frequency of plots recording Frequency of plots Frequency of plots recording an Frequency of plots Country Year recording a disaster a drought shock recording a flood shock irregular rain shock recording a pest shock shock Disaster Disaster All plots All plots All plots Disaster plots All plots Disaster plots plots plots 2011 - 2012 31.19 % 16.96 % 55.16 % missing missing 11.39 % 37.41 % 4.89 % 16.06 % 2013 - 2014 28.32 % 11.14 % 40.14 % missing missing 9.94 % 36.60 % 7.35 % 30.52 % Ethiopia 2015 - 2016 53.83 % 41.89 % 78.29 % missing missing 17.55 % 34.88 % 3.48 % 7.97 % 2018 - 2019 39.76 % 19.81 % 50.63 % missing missing 14.15 % 37.17 % 6.95 % 21.12 % 2009 - 2010 40.40 % 24.34 % 60.26 % missing missing missing missing 1.05 % 2.61 % 2012 - 2013 36.82 % 8.18 % 22.22 % 8.64 % 23.47 % 17.80 % 48.33 % 2.85 % 7.75 % Malawi 2015 - 2016 64.54 % 30.77 % 47.68 % 1.06 % 1.56 % 40.61 % 62.92 % 2.21 % 3.42 % 2018 - 2019 47.18 % 5.60 % 11.86 % 21.65 % 45.81 % 19.77 % 41.91 % 7.97 % 16.76 % 2014 18.59 % 14.44 % 77.71 % missing missing 1.75 % 9.41 % 1.05 % 5.63 % Mali 2017 24.07 % 22.78 % 94.64 % missing missing 0.45 % 1.89 % 3.08 % 1.97 % 2011 89.88 % 65.81 % 73.22 % 0.64 % 0.71 % missing missing 27.69 % 30.81 % Niger 2014 40.11 % 23.43 % 58.41 % 2.75 % 6.86 % missing missing 11.08 % 27.63 % Nigeria 2019 11.39 % 2.59 % 26.06 % 4.88 % 50.72 % missing missing 5.49 % 48.85 % 2008 32.21 % 14.52 % 47.11 % missing missing missing missing 13.99 % 45.45 % 2010 36.70 % 21.52 % 60.74 % missing missing missing missing 10.00 % 28.29 % Tanzania 2012 32.78 % 19.91 % 62.79 % missing missing missing missing 9.86 % 31.05 % 2014 23.96 % 11.38 % 49.72 % missing missing missing missing 8.05 % 35.00 % 2019 40.89 % 25.80 % 65.03 % missing missing missing missing 12.39 % 31.21 % All 2008 – 2019 36.34 % 22.17 % 61.66 % 6.80 % 14.44 % 10.49 % 31.31 % 6.30 % 16.83 % countries Note: some questionnaires (the IHPS, for example) allow respondents to report multiple shock types on a single plot. 22 Table 2: Mean fraction of potential harvest lost at the plot level, by country-year Mean fraction Mean fraction Mean fraction Mean fraction of potential of potential of potential of potential Mean fraction of harvest lost, harvest lost, harvest lost, harvest lost, Country Year potential harvest on plots on plots on plots on plots lost, all plots affected by affected by affected by affected by drought floods irregular rains pests 2013 - 2014 48.66% 53.46 % missing 49.36 % 45.22 % Ethiopia 2015 - 2016 54.84% 59.71 % missing 55.90 % 47.35 % 2018 - 2019 47.63% 46.91 % missing 47.65 % 42.36 % Malawi 2018 - 2019 55.00% 57.17 % 62.65 % 49.79 % 49.26 % 2014 48.08% 48.37 % missing 38.11 % 40.73 % Mali 2017 47.96% 47.50 % missing 35.75 % 66.73 % 2011 71.09% 71.79 % 77.53 % missing 66.96 % Niger 2014 49.33% 52.90 % 56.79 % missing 40.97 % Nigeria 2019 56.38% 36.38 % 61.87 % missing 30.71 % All countries 2011 – 2019 53.49% 56.75 % 62.28 % 51.52 % 46.52 % Note: only point estimates are reported. Plots with no losses are excluded. Sample weights are used to calculate estimates. 23 Table 3: Linear trend in shock exposure and size (1) (2) (3) Any loss Percent lost Percent lost Full sample With any loss Full sample Year 0.00829*** -1.137** 0.0986 (0.00231) (0.519) (0.368) Constant 2,344** -178.4 (1,047) (741.5) Observations 121,983 31,746 91,280 Country FE YES YES YES Note: Marginal effects from logit regression (Column 1) and results from OLS regressions (Columns 2 and 3). Columns 1 and 3 are unconditional, column 2 is conditional on any loss on the plot. The estimates are weighted to be nationally representative. Standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1 24 Table 4: Total fraction of aggregate potential harvest lost, by country-year Fraction of total potential harvest Country Year lost in shocks 2013 - 2014 15.14 % Ethiopia 2015 - 2016 34.63 % 2018 - 2019 31.52 % Malawi 2018 - 2019 46.53 % 2014 7.59 % Mali 2017 8.22 % 2011 81.12 % Niger 2014 29.24 % Nigeria 2019 3.19 % Pooled 2011-2019 28.57 % Note: Sample weights are used to calculate estimates. Current prices are used to value losses and attainable harvest 25 Table 5: Share of households affected by multiple disaster shocks on their plots Percent of households that Country Year report disaster losses from multiple sources Ethiopia 2011 – 2012 8.99 % Ethiopia 2013 - 2014 12.58 % Ethiopia 2015 - 2016 20.92 % Ethiopia 2018 - 2019 21.02 % Malawi 2012 - 2013 4.50 % Malawi 2015 - 2016 15.03 % Malawi 2018 - 2019 15.05 % Tanzania 2008 5.09 % Tanzania 2010 4.25 % Tanzania 2012 3.04 % Tanzania 2014 1.46 % Tanzania 2019 4.52 % Note: Sample weights are used to calculate estimates. Only waves which allow the reporting of multiple shock types on each listed crop are retained. 26 Table 6: Loss size by shock type (1) (2) (3) (4) (5) Dependent variable: Percent of harvest lost Ethiopia Malawi Mali Niger Nigeria Drought 6.570*** 7.767*** 2.536 3.183 -4.939 (1.853) (2.668) (3.966) (2.235) (4.152) Pests 5.955*** -4.283 11.48* -6.561** -10.28** (1.731) (2.867) (6.384) (2.710) (4.711) Flood 15.11*** -1.935 9.263* (2.995) (5.656) (4.984) Constant 34.08*** 47.68*** 45.31*** 64.71*** 31.80*** (1.145) (1.944) (3.875) (2.292) (4.936) Observations 10,478 2,180 7,389 6,537 449 Crop FE YES YES YES YES YES Survey Wave FE YES YES YES YES YES Note: Results from OLS regressions with the share of potential harvest lost as outcome variable. Base category for shock dummies is other shock. The estimates are weighted to be nationally representative. Standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1 27 Table 7: Heterogeneity in disaster exposure by plot characteristics (1) (2) (3) (4) (5) (6) (7) Dependent variable: Any disaster loss on plot Pooled Ethiopia Malawi Mali Niger Nigeria Tanzania Any hired labor on plot -0.0151 -0.00717 -0.0513*** -0.000535 -0.0291 -0.00315 -0.0130 (0.0111) (0.0242) (0.0167) (0.0142) (0.0205) (0.0217) (0.0151) Any inorganic fertilizer used -0.0445*** -0.0482** 0.0236 -0.0621*** -0.0527* -0.0357** -0.0885*** (0.0109) (0.0187) (0.0161) (0.0171) (0.0289) (0.0154) (0.0290) Any organic fertilizer used 0.0285*** 0.0666*** 0.0382** -0.00299 -0.00833 -0.0106 -0.0126 (0.0105) (0.0195) (0.0158) (0.0160) (0.0255) (0.0188) (0.0266) Plot is irrigated -0.00339 0.000150 -0.0647 -0.201*** -0.169*** 0.00297 0.0370 (0.0211) (0.0334) (0.0592) (0.0461) (0.0535) (0.0305) (0.0502) Plot is intercropped 0.0491*** 0.0353 0.146*** -0.0335 -0.0271 -0.0221 0.0419** (0.0117) (0.0242) (0.0181) (0.0292) (0.0364) (0.0164) (0.0207) Plot is owned -0.000302 0.0200 -0.00111 -0.0542** -0.0529 -0.0296 -0.00966 (0.0121) (0.0199) (0.0215) (0.0246) (0.0329) (0.0184) (0.0234) Log plot area (ha) 0.00875*** 0.00983* 0.0219*** 0.000514 -0.0272*** -0.0122* 0.0339*** (0.00333) (0.00504) (0.00568) (0.00562) (0.00835) (0.00677) (0.00803) Plot topographic wetness index 0.00731*** 0.00315 0.00503 0.00210 0.00349 0.00134 0.0121*** (0.00165) (0.00407) (0.00345) (0.00244) (0.00267) (0.00161) (0.00225) Plot elevation (m) -9.16e-05*** -8.41e-05*** -0.000371*** -0.000299** -0.000172 -2.12e-05 -6.58e-05*** (1.60e-05) (2.64e-05) (3.41e-05) (0.000131) (0.000186) (4.61e-05) (2.37e-05) Urban household -0.0211 0.00330 -0.115*** -0.0356 0.0446 -0.0212 0.0273 (0.0206) (0.0529) (0.0410) (0.0484) (0.0472) (0.0249) (0.0291) Household engaged in livestock farming 0.00545 0.0210 -0.00556 0.0271 -0.145*** -0.0153 0.0398** (0.0101) (0.0211) (0.0133) (0.0203) (0.0466) (0.0182) (0.0173) Household has access to electricity 0.0711*** 0.120*** 0.0173 -0.0240 -0.0302 0.00143 0.0195 (0.0161) (0.0241) (0.0283) (0.0167) (0.0493) (0.0197) (0.0358) Observations 116,773 39,801 17,900 31,257 9,004 6,623 12,187 Crop FE YES YES YES YES YES YES YES Country FE YES NO NO NO NO NO NO Note: Average marginal effects from multivariate logit regressions. Base category for crop fixed effects is 'other crop'. The estimates are weighted to be nationally representative. Standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1 28 Table 8: Heterogeneity in loss size by plot characteristics (1) (2) (3) (4) (5) (6) Dependent variable: Percent of harvest lost Pooled Ethiopia Malawi Mali Niger Nigeria Any hired labor on plot -5.120*** -3.632 -3.179 -0.707 -4.162*** -9.870* (1.762) (2.236) (3.237) (1.977) (1.454) (5.065) Any inorganic fertilizer used -8.073*** -7.744*** -2.133 -15.47*** -4.415** -17.28*** (1.512) (1.885) (1.885) (2.563) (1.955) (4.320) Any organic fertilizer used 0.707 3.476* -3.111** -4.282** -6.589*** -8.291** (1.315) (1.842) (1.522) (2.106) (1.183) (4.103) Plot is irrigated 3.450 1.126 9.046 -20.15*** 4.865 14.80* (3.063) (3.636) (14.07) (6.356) (4.255) (8.736) Plot is intercropped -7.513*** -3.781 -1.550 -6.306* -6.367*** -36.05*** (1.754) (2.329) (2.454) (3.550) (1.773) (4.865) Plot is owned -1.945 -2.300 6.057* -2.874 -2.586 -3.114 (1.556) (2.017) (3.122) (3.702) (1.645) (4.379) Log plot area (ha) -2.565*** -2.782*** -4.610*** -0.255 -0.599 0.394 (0.410) (0.497) (1.109) (0.686) (0.427) (1.383) Plot topographic wetness index 0.263 0.576* 0.0928 0.993*** 0.0383 0.0686 (0.206) (0.301) (0.434) (0.344) (0.249) (0.451) Plot elevation (m) -0.00727*** -0.00650*** -0.0250*** -0.0179 -0.00721 0.00584 (0.00214) (0.00228) (0.00404) (0.0187) (0.0108) (0.00892) Urban household 4.463 12.26*** -3.747 5.014 1.584 -5.863 (2.772) (2.829) (3.387) (6.587) (2.659) (5.539) Household engaged in livestock farming -2.395 1.289 0.157 3.039 -9.035*** -2.391 (1.733) (2.353) (1.820) (3.219) (2.203) (4.460) Household has access to electricity 2.388 1.600 3.479 -1.266 -2.625 5.226 (1.946) (2.219) (4.478) (2.336) (2.222) (4.404) Constant 64.21*** 53.17*** 61.56*** 55.52*** 84.30*** 104.4*** (6.177) (7.224) (9.642) (10.15) (6.901) (10.39) Observations 30,845 14,896 2,059 7,130 6,071 689 Crop FE YES YES YES YES YES YES Country FE YES NO NO NO NO NO Note: Results from OLS regressions with the share of potential harvest lost as outcome variable. Base category for crop fixed effects is 'other crop'. The estimates are weighted to be nationally representative. Standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1 29 Table 9: Heterogeneity in disaster exposure by plot manager characteristics Pooled Ethiopia Malawi Mali Niger Nigeria Tanzania (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14) Dependent variable: Any disaster loss on plot No control Control No control Control No control Control No control Control No control Control No control Control No control Control Female plot manager 0.0222** 0.0267*** 0.00350 0.0126 0.0529*** 0.0272** 0.0843*** 0.0614*** 0.0503* 0.0475* 0.0386* 0.00467 0.0138 0.0355* (0.00939) (0.00904) (0.0171) (0.0169) (0.0137) (0.0123) (0.0237) (0.0215) (0.0293) (0.0280) (0.0207) (0.0165) (0.0201) (0.0186) Age of plot manager (decades) 0.00727*** 0.00708*** 0.00829* 0.00750* 0.00112 0.00407 0.0124*** 0.0124*** -0.0137*** -0.00896* 0.00774 0.00564 0.0119** 0.00878* (0.00249) (0.00237) (0.00457) (0.00418) (0.00446) (0.00416) (0.00443) (0.00429) (0.00525) (0.00517) (0.00542) (0.00533) (0.00523) (0.00481) Plot manager has primary educ -0.0140 -0.0138 -0.0176 -0.0258 -0.0467*** -0.00906 -0.0386* -0.0178 -0.0609 -0.0594 0.0186 -0.00118 -0.0216 -0.00847 (0.0125) (0.0124) (0.0300) (0.0285) (0.0145) (0.0140) (0.0224) (0.0216) (0.0480) (0.0473) (0.0155) (0.0179) (0.0314) (0.0336) Observations 115,041 115,041 39,289 39,289 17,123 17,123 31,019 31,019 8,911 8,911 6,591 6,591 12,107 12,107 Controls NO YES NO YES NO YES NO YES NO YES NO YES NO YES Crop FE YES YES YES YES YES YES YES YES YES YES YES YES YES YES Country FE YES YES NO NO NO NO NO NO NO NO NO NO NO NO Note: Average marginal effects from multivariate logit regressions. Controls include dummy variables for input use (any hired labor, any inorganic fertilizer, any organic fertilizer), plot characteristics (plot area, irrigration, intercropping, plot ownership, elevation, and a topographic wetness index), household characteristics (urban/rural residence, livestock farming, and electricity access) as well as main crop and country fixed effects. The estimates are weighted to be nationally representative. Standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1 30 Table 10: Heterogeneity in loss size by plot manager characteristics Pooled Ethiopia Malawi Mali Niger Nigeria (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) Dependent variable: Percent of harvest lost on plot No controls Controls No controls Controls No controls Controls No controls Controls No controls Controls No controls Controls Female plot manager 4.405*** 2.303** 3.345** 1.968 3.952 2.363 5.167 0.988 5.025*** 3.789** 10.76** -0.0518 (1.158) (1.136) (1.448) (1.385) (2.383) (2.128) (3.452) (2.895) (1.742) (1.754) (5.463) (6.238) Age of plot manager (decades) 0.231 0.285 0.373 0.407 -0.995 0.0282 -0.537 -0.385 -0.382 -0.0241 0.974 -0.303 (0.320) (0.315) (0.384) (0.377) (0.871) (0.797) (0.663) (0.544) (0.481) (0.474) (1.837) (1.328) Plot manager has primary educ -0.0273 -1.641 -1.508 -3.215 0.497 1.915 -4.344 -1.184 -5.307* -3.501 5.493 -1.332 (1.831) (1.686) (2.362) (2.192) (2.893) (2.226) (2.950) (2.775) (2.759) (2.510) (4.983) (5.349) Constant 49.31*** 62.68*** 48.94*** 51.42*** 55.91*** 60.52*** 49.98*** 57.01*** 65.99*** 83.15*** 46.19*** 106.0*** (1.923) (6.404) (2.152) (7.456) (4.869) (9.858) (3.617) (10.82) (2.491) (7.111) (10.01) (15.73) Observations 30,230 30,230 14,680 14,680 1,739 1,739 7,100 7,100 6,027 6,027 684 684 Crop FE YES YES YES YES YES YES YES YES YES YES YES YES Country FE YES YES NO NO NO NO NO NO NO NO NO NO Note: Results from OLS regressions with the share of potential harvest lost as outcome variable. Controls include dummy variables for input use (any hired labor, any inorganic fertilizer, any organic fertilizer), plot characteristics (plot area, irrigration, intercropping, plot ownership, elevation, and a topographic wetness index), household characteristics (urban/rural residence, livestock farming, and electricity access) as well as main crop and country fixed effects. The estimates are weighted to be nationally representative. Standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1 31 Table 11: Heterogeneity in plot endowment and management by gender of the plot manager (1) (2) (3) (4) (5) (6) (7) Dependent variable: Female plot manager (dummy) Pooled Ethiopia Malawi Mali Niger Nigeria Tanzania Any hired labor on plot 0.0342*** 0.0336** 0.000582 -0.0254** 0.0644*** 0.0473** 0.0560*** (0.00878) (0.0145) (0.0148) (0.0104) (0.0148) (0.0216) (0.0201) Any inorganic fertilizer used -0.00474 0.0159 -0.0149 -0.0206 -0.00489 -0.0373** -0.00251 (0.00879) (0.0125) (0.0154) (0.0138) (0.0349) (0.0176) (0.0275) Any organic fertilizer used -0.0128 0.0127 0.0234 -0.0381*** -0.0375** -0.0367* -0.0167 (0.00924) (0.0114) (0.0186) (0.00935) (0.0150) (0.0211) (0.0309) Plot is irrigated -0.0332 0.0234 -0.0671 -0.0956*** -0.244** -0.0418 -0.0718 (0.0246) (0.0339) (0.0703) (0.0274) (0.100) (0.0610) (0.0489) Plot is intercropped 0.0184** -0.0168 0.0698*** -0.0224 -0.0544*** 0.0287** 0.0430* (0.00844) (0.0149) (0.0129) (0.0225) (0.0206) (0.0144) (0.0225) Plot is owned 0.0268*** 0.0907*** 0.0897*** -0.00327 -0.0335** -0.00298 -0.00916 (0.00988) (0.0173) (0.0239) (0.0189) (0.0148) (0.0181) (0.0195) Log plot area (ha) -0.0423*** -0.0245*** -0.0524*** -0.0229*** -0.0461*** -0.0498*** -0.0619*** (0.00258) (0.00297) (0.00706) (0.00363) (0.00667) (0.00567) (0.00795) Plot topographic wetness index 0.00105 -0.00154 0.00110 0.00154 -0.00162 -0.000648 0.00310 (0.00141) (0.00271) (0.00386) (0.00135) (0.00156) (0.00176) (0.00323) Plot elevation (m) -2.96e-05** 2.91e-07 -9.40e-05*** -0.000297*** 0.000227** -0.000188* -5.01e-06 (1.17e-05) (1.59e-05) (3.27e-05) (5.32e-05) (0.000110) (9.64e-05) (2.13e-05) Urban household 0.00322 0.0744** 0.0166 -0.0568** -0.0634* -0.0347 0.0214 (0.0153) (0.0356) (0.0350) (0.0274) (0.0378) (0.0281) (0.0268) Household engaged in livestock farming -0.0888*** -0.142*** -0.103*** 0.00674 -0.00526 -0.00275 -0.0921*** (0.00932) (0.0152) (0.0147) (0.0124) (0.0200) (0.0202) (0.0195) Household has access to electricity -0.00878 -0.0308** -0.0621* 0.00794 -0.0857** 0.0115 -0.0408 (0.0115) (0.0144) (0.0345) (0.00813) (0.0384) (0.0205) (0.0333) Observations 118,311 40,243 17,932 31,138 8,803 7,773 12,307 Crop FE YES YES YES YES YES YES YES Country FE YES NO NO NO NO NO NO Note: Average marginal effects from multivariate logit regressions. Base category for crop fixed effects is 'other crop'. The estimates are weighted to be nationally representative. Standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1 32 Table 12: Heterogeneity in drought reports within enumeration areas, by crop type Drought All crops Maize Sorghum Millet 2011-2012 39.4 16.1 18.6 13.3 2013-2014 43.8 26.3 24.9 12.7 Ethiopia 2015-2016 55.8 34.3 29.0 23.3 2018-2019 48.6 27.7 26.9 10.5 2009-2010 62.6 58.4 76.0 30.0 2012-2013 25.7 28.0 36.2 30.0 Malawi 2015-2016 33.0 36.3 43.2 40.0 2018-2019 17.5 11.7 20.5 16.7 2014 29.7 13.9 20.6 14.4 Mali 2017 44.5 28.1 32.2 32.2 2011 78.6 41.7 68.1 82.4 Niger 2014 66.4 28.6 62.6 63.6 Nigeria 2019 20.7 8.9 7.6 6.4 2008 55.9 51.9 0.0 50.0 2010 59.3 57.6 58.1 55.6 Tanzania 2012 46.9 44.4 37.9 26.3 2014 32.1 29.7 66.7 50.0 2019 42.1 44.6 87.5 33.3 All countries All years 41.1 31.1 31.5 32.1 Note: Share of enumeration areas with heterogenous drought reports, pooled and by crop type. Heterogeneous reports are defined as enumeration areas where at least one plot recorded a drought loss and at least one plot did not. Columns 2, 3, and 4 compare only plots that grew the same crop. Enumeration areas with only a single plot with the crop are excluded. 33 Table 13: Mean frequency of plots with drought shocks within household, conditional on at least one drought shock being recorded Mean drought shock Mean drought shock frequency of plots with frequency of plots with the same crop within the same crop within Country Survey household household Restricted to plots with Restricted to plots with maize maize (1) (2) Ethiopia 2011 – 2012 68.75 % n≤10 Ethiopia 2013 – 2014 62.52 % 70.76 % Ethiopia 2015 – 2016 82.33 % 91.07% Ethiopia 2018 – 2019 59.68 % 72.81 % Malawi 2010 88.08 % 93.52 % Malawi 2013 65.77 % 100.00 % Malawi 2016 71.77 % 88.89 % Malawi 2019 38.89 % 40.00 % Mali 2014 65.83 % 69.91 % Mali 2017 62.97 % 62.72 % Niger 2011 n≤10 90.29 % Niger 2014 n≤10 69.01 % Nigeria 2019 24.40 % 57.14 % Tanzania 2008 – 2009 43.22 % n≤10 Tanzania 2010 – 2011 52.65 % 67.39 % Tanzania 2012 – 2013 56.57 % n≤10 Tanzania 2014 – 2015 44.25 % n≤10 Tanzania 2019 – 2020 51.19 % n≤10 All countries 66.56 % 79.55 % Note: These frequencies are conditional on at least one plot with the same crop (maize in col (1) or sorghum in col(2)) experiencing a drought shock, the household having more than one crop plot with the same specific crop (maize in col (1) or sorghum in col(2)), and a sample size of over 10 observations to calculate shares. 34 Table 14: Share of the national population affected by droughts according to EM-DAT and LSMS-ISA data Percent of the Percent of the population affected population affected Country Year by drought using by drought using EM- LSMS-ISA data DAT data 7.88 % 2011 - 2012 6.20 % [ 5.73 % ; 10.04 % ] 11.45 % 2013 - 2014 missing Ethiopia [ 8.83 % ; 14.07 % ] 32.50 % 2015 - 2016 9.68 % [ 27.33 % ; 37.67 % ] 14.30 % 2018 - 2019 missing [ 10.88 % ; 17.72 % ] 23.82 % 2009 - 2010 missing [ 20.44 % ; 27.20 % ] 8.98 % 2012 - 2013 missing Malawi [ 6.98 % ; 10.98 % ] 25.51 % 2015 - 2016 missing [ 19.25 % ; 31.78 % ] 7.66 % 2018 - 2019 missing [ 5.06 % ; 10.26 % ] 15.89 % 2014 missing [ 13.35 % ; 18.43 % ] Mali 24.53 % 2017 missing [ 21.54 % ; 27.53 % ] 56.14 % 2011 17.03 % [ 48.77 % ; 63.50 % ] Niger 20.78 % 2014 missing [ 15.78 % ; 25.78 % ] 2.39 % Nigeria 2019 missing [ 1.49 % ; 3.29 % ] 12.38 % 2008 missing [ 10.55 % ; 14.21 % ] 18.98 % 2010 missing [ 16.54 % ; 21.42 % ] Tanzania 15.38 % 2012 2.09 % [ 13.08 % ; 17.68 % ] 10.27 % 2014 missing [ 4.08 % ; 16.45 % ] 16.50 % 2019 missing [ 10.07 % ; 22.93 % ] Note: missing entries correspond to cases where either no event was reported or information on the population affected was missing in the EM-DAT. 95% confidence intervals for estimated totals are calculated with LSMS-ISA microdata 35 Table 15: Total estimated value of crop production lost (LSMS-ISA) and total aggregate economic losses (EM-DAT) due to droughts Total aggregate losses Total aggregate losses Country Year due to drought (in 2022 due to drought (in 2022 dollars) in LSMS-ISA dollars) in EM-DAT 589 million 2013 - 2014 missing [ 281 ; 896 ] 2,980 million Ethiopia 2015 - 2016 2,134 million [ 2000 ; 3960 ] 1,354 million 2018 - 2019 missing [ 229 ; 2479 ] 41 million Malawi 2018 - 2019 missing [ 23 ; 60 ] 70 million 2014 missing [ 55 ; 85 ] Mali 165 million 2017 missing [ 136 ; 194 ] 1.583 million 2011 missing [ 1294 ; 1871 ] Niger 185 million 2014 missing [ 129 ; 241 ] 80 million Nigeria 2019 missing [ 29 ; 131 ] Note: missing entries correspond to cases where either no event was reported or information on the population affected was missing in the EM-DAT. 95% confidence intervals for estimated totals are calculated with LSMS-ISA microdata 36 Table 16: Share of the national population affected by floods according to EM-DAT and LSMS-ISA data Percent of the Percent of the population affected population affected Country Year by floods according by floods according to LSMS-ISA data to EM-DAT data 8.26 % 2012 - 2013 0.95 % [ 5.73 % ; 10.78 % ] Malawi 1.18 % 2015 - 2016 3.74 % [ 0.49 % ; 1.87 % ] 22.75 % 2018 - 2019 5.38 % [ 16.37 % ; 29.12 % ] 1.72 % 2011 0.24 % [ 0.30 % ; 3.13 % ] Niger 3.76 % 2014 0.85 % [ 1.58 % ; 5.95 % ] 4.62 % Nigeria 2019 0.03 % [ 3.17 % ; 6.08 % ] Note: missing entries correspond to cases where either no event was reported or information on the population affected was missing in the EM-DAT. 95% confidence intervals for estimated totals are calculated with LSMS-ISA microdata 37 Table 17: Total estimated value of crop production lost (LSMS-ISA) and total aggregate economic losses (EM-DAT) due to floods Total aggregate Total aggregate flood loss (2022 flood loss (2022 Country Year dollars) in LSMS- dollars) in EM- ISA DAT 346 million Malawi 2018 - 2019 missing [95 ; 598] 9 million 2011 missing [ 3.00 ; 15 ] Niger 81 million 2014 3 million [ 0 ; 205 ]* 352 million Nigeria 2019 missing [ 208 ; 496 ] Note: missing entries correspond to cases where either no event was reported or information on the population affected was missing in the EM-DAT. 95% confidence intervals for estimated totals are calculated with LSMS-ISA microdata *Negative lower bounds were cut at 0. 38 Table 18. Estimated combined impacts of drought or flood events captured in LSMS-ISA Share of Share of Number of Value of Number of Value of total total individuals damages individuals damages population population affected by from affected by from Country Year affected by affected by droughts droughts droughts droughts droughts droughts and/or and/or and/or and/or and/or and/or floods floods floods floods floods floods LSMS-ISA LSMS-ISA LSMS-ISA EM-DAT EM-DAT EM-DAT 7.9 % 2011 - 7.4 million [ 5.7 % ; 10.0 5.846 million 6.30 % [ 5.4 ; 9.5 ] 2012 %] 11.4 % 589.6 million 2013 - 11.4 million [ 8.8 % ; 14.1 [ 280.9 ; 898.2 52 thousand 0.10 % 3.5 million [ 8.8 ; 14.0 ] 2014 %] ] 2,981.3 Ethiopia 34.2 million 32.5 % million 10.904 2,134.4 2015 - [ 28.8 ; 39.6 ] [ 27.4 % ; [ 1,999.0 ; million 10.40 % million 37.6 % ] 2016 3,963.7 ] 1,355.7 16.3 million 14.3 % million 200 2018 - [ 13.1 ; 19.5 [ 11.5 % ; [ 227.9 ; thousand 0.20 % ] 17.1 % ] 2019 2,483.4 ] 23.8 % 2009 - 3.5 million [ 20.5 % ; 38 thousand 0.30 % 2010 [ 3.0 ; 4.0 ] 27.1 % ] 16.5 % 2012 - 2.6 million [ 13.5 % ; 2.05 million 13.00 % 2013 [ 2.2 ; 3.1 ] 19.5 % ] Malawi 26.4 % 2015 - 4.6 million [ 20.1 % ; 7.341 million 42.40 % 594.6 million [ 3.5 ; 5.7 ] 2016 32.8 % ] 28.5 % 379.9 million 2018 - 5.4 million [ 21.5 % ; [ 123.0 ; 636.7 1.001 million 5.40 % [ 4.1 ; 6.7 ] 2019 35.4 % ] ] 15.9 % 2.8 million 70.0 million [ 13.4 % ; No entry 2014 [ 2.4 ; 3.2 ] 18.4 % ] [ 54.7 ; 85.3 ] Mali 24.5 % 164.7 million 4.7 million [ 21.7 % ; [ 135.4 ; No entry [ 4.2 ; 5.3 ] 2017 27.4 % ] 194.0 ] 1,589.9 56.7 % 9.8 million million [ 49.1 % ; 3.041 million 17.40 % [ 8.5 ; 11.1 ] [ 1,298.4 ; 64.4 % ] Niger 2011 1,881.3 ] 24.1 % 265.6 million 4.7 million 166 [ 18.4 % ; [ 129.5 ; 401.7 0.90 % 3.1 million [ 3.6 ; 5.8 ] thousand 2014 29.8 % ] ] 13.6 million 6.8 % 427.5 million Nigeria [ 10.5 ; 16.7 [ 5.3 % ; 8.4 [ 274.4 ; 71 thousand 0.00 % 0.0 million 2019 ] %] 580.6 ] 12.4 % 5.3 million [ 10.6 % ; 10 thousand 0.00 % 2008 [ 4.5 ; 6.1 ] 14.2 % ] 19.0 % 8.6 million Tanzania [ 7.5 ; 9.7 ] [ 16.5 % ; 50 thousand 0.10 % 2010 21.4 % ] 15.4 % 7.3 million [ 13.1 % ; 1.0 million 2.10 % 2012 [ 6.3 ; 8.4 ] 17.7 % ] 39 10.3 % 5.2 million [ 4.1 % ; 40 thousand 0.10 % 3.1 million 2014 [ 2.1 ; 8.4 ] 16.5 % ] 16.5 % 9.9 million [ 10.4 % ; 5 thousand 0.00 % 2019 [ 6.2 ; 13.6 ] 22.6 % ] 7,824.2 All 2008 - 157.0 million million 31.814 2,738.6 [ 144.6 ; countries 2019 170.3 ] [ 6,168.6 ; million million 9,479.8 ] Note: both point estimates and 95% confidence intervals are reported 40 Table 19. Overview of disaster loss information availability in the LSMS-ISA microdata Country Year Ability to quantify List of disaster shocks partial losses options provided in the dataset 2011 - 2012 No Drought; Rains: Fire; 2013 - 2014 Yes Ethiopia Insects; Crop disease; 2015 - 2016 Yes Locusts; Hail 2018 - 2019 Yes 2009 - 2010 No Drought; Irregular rains; 2012 - 2013 No Malawi Floods (not in wave 1); Fire; 2015 - 2016 No Insects; Disease 2018 - 2019 Yes 2014 Yes Drought; Rains; Fire; Mali 2017 Yes Insects; Disease Yes Insects and bird attacks; Niger 2011 2014 Yes Plant illness; Drought; Flood Nigeria 2019 Yes Drought; Flood; Pest; 2008 No 2010 No Tanzania 2012 No Drought; Rain; Fire; Insects 2014 No 2019 No 41 Appendix A: Comparison of data from disaster inventories and microdata Data from disaster inventories such as EM-DAT and from microdata such as the LSMS-ISA differ in a number of meaningful ways. This includes their coverage of different shock types and sizes, the affected population, and time frame, as well as their detail, accuracy, and comparability across countries. These dimensions determine the respective strengths and weaknesses of each data source and can give rise to differences in estimates of disaster incidence and severity. Disaster inventories collate reports from a variety of data sources on a per-event basis and are subject to the coverage and detail of information provided in these underlying sources.37 Typically, this data comes aggregated, most often at the country-level. On the other hand, microdata, in our case, is based on household surveys that interview a sample from the target population and population-level estimates are formed based on this sample. Shock coverage Coverage of shocks in disaster inventories varies depending on the underlying source data for each event. While large, salient shocks are likely reported across one or more sources that disaster inventories rely on (such as news reports or data bases from international organizations) and are therefore likely covered in disaster inventories, this is somewhat less likely for small or idiosyncratic shocks.30,37 Such small shocks may therefore go unrecorded. For similar reasons, some shock types are more likely to be covered in disaster inventories than others.37,42 Conversely, microdata is based on reports directly elicited from those affected by the shocks in question. They are therefore more granular and able to cover small and localized shocks. At the same time, shock recording in microdata sources is subject to the design of the survey questionnaire such as the list of shocks covered, and the number of shocks recorded. Population coverage Disaster inventories are not limited to a specific population of interest and can, in principle, cover any event affecting any population group so long as this is reported in a source the disaster inventory can draw on. The lack of an explicit focus and reliance on reports from underlying sources, in turn, means that coverage for poor and marginalized population groups is more likely to be incomplete. Microdata from household surveys focuses on (stratified) samples from a well-defined target population. In the case of the LSMS-ISA surveys, the samples are typically drawn to be nationally representative of the general population at the national and sub-national levels, and of urban and rural areas. This means that even poor and marginalized populations can be explicitly covered. At the same time, coverage is limited to the sample and affected by gaps in coverage wherever the sample is not fully representative of the population affected by a shock. Temporal coverage 42 A major strength of disaster inventories is their ability to record shocks at any time they occur so long as they get reported. Over the long run, this means that coverage of shocks improves as shock recording in underlying sources gets more exact and comprehensive.37 In contrast, temporal coverage in microdata is intermittent: limited to the years in which the survey is implemented and the recall period of the survey. Detail of available information The breadth of coverage in disaster inventories across time and space often comes at the cost of limited detail and missing information.30,37,48 This particularly concerns the economic and welfare impact of shocks which are often difficult to quantify, especially so for uninsured damages.37,40 Detail and completeness of the available information in disaster repositories is thus related to the size and salience of a shock. Further, the data provided in disaster inventories typically offers little opportunity for disaggregation, be it by affected population groups, geographies, or other key variables. As a consequence, disaster inventories often lack sufficient information to quantify disaster impacts on (asset- ) poor but highly vulnerable population groups.21 As microdata sources elicit shock reports directly from those affected by them, they offer a high level of detail and completeness of information even for small or idiosyncratic events, for shocks that affect poor or marginalized population groups, and for their impact on uninsured damages and other, less salient domains of disaster impact. Depending on the design of the survey, shock and impact recording is often highly granular and allows for disaggregation and analysis along a broad range of ancillary information collected in the survey. Accuracy of data Data recorded in disaster inventories typically relies on reports from one or few sources. Especially where events are small and therefore only reported in a single data source, this exposes disaster inventories to possible inaccuracies in the source from which information is drawn.37 Even though reports of the number of people affected and the economic impact of any disaster are almost always estimates, disaster inventories rarely provide information that quantifies the uncertainty associated with any individual or collection of estimates. Estimates from microdata, on the other hand, are based on a large number of data points collected from the population of interest. This also allows to explicitly quantify (part of) the uncertainty inherent in estimates of disaster impact. While the large samples underlying estimates from microdata attenuate the effect of inaccuracies in individual data points, microdata, as any data source, is still subject to possible systematic measurement error. Comparability of source data Disaster inventories, and EM-DAT in particular, stand out for their broad coverage of events across geographies and time. To achieve this, disaster inventories need to rely on a wealth of different sources that vary according to their reliability, reporting standards, and definitions. Such idiosyncrasies are seldom 43 explicit in collated records as part of disaster inventories but limit the comparability of data across different events and contexts.37 Within the same survey, information collected in microdata is typically harmonized and comparable across events covered. While microdata are subject to differences in survey and questionnaire design across contexts, such differences are usually explicit. Analysts can therefore take these differences into account when making comparisons and there is scope for harmonization across contexts as in the case of the LSMS-ISA suite of surveys. Geographical differences Many of the previously discussed sources of inaccuracies in estimates of disaster incidence and severity are more pronounced in low- and middle-income countries (LMICs).21,30,37,42 This goes in particular for data from disaster inventories that depend on the comprehensiveness of coverage in the underlying data sources they draw on. For example, the density of information on shocks and natural disasters is lower in LMICs meaning that shocks are more likely to go unreported, information is more likely to be incomplete or inaccurate, poor and marginalized population groups are more likely not to be covered, and the impact on them is more likely to be underestimated due to a higher share of uninsured damages and a lower value of affected assets. On the side of microdata, surveys in high-income countries may provide more accurate data due to the availability of more sophisticated measurement approaches and higher levels of education among respondents which may positively affect the accuracy of self-reported information. Dimension Disaster Inventories Microdata Coverage • Subject to reporting in aggregate data • Based on “grassroots” reports elicited sources and reports from those affected by shocks Shock coverage • Salient shocks likely covered but data • Granular and able to cover small and (Threshold and hazard less sensitive to idiosyncratic or small localized shocks bias) shocks • Shock recording subject to questionnaire • Coverage varies by shock type design (list of shocks, number of shocks recorded) • Coverage depending on • Limited to (stratified) sample of target comprehensiveness of coverage in population underlying data sources (e.g. news • Potential for gaps in coverage wherever Population coverage reports) but not limited to a specific shock impacts not well represented by (Population coverage population of interest sample biases) • under-coverage of poor and • Poor and marginalized population groups marginalized population groups within explicitly covered countries likely Temporal coverage • Continuous coverage but subject to • Intermittent coverage limited to years in (Temporal coverage improvements in quality of reports in the which survey was conducted and/or biases) long run recall period of survey questions Detail and accuracy Detail of available • Dependent on information reported in • High level of detail and completeness of information sources, limited detail and frequently data collected, even for small and (Missing data biases) missing information in some dimensions 44 of shock impacts (e.g. economic and idiosyncratic shocks as well as poor or welfare impacts) vulnerable population groups • Detail and completeness of information • Information collected at highly related to size and salience of shock disaggregated level but can be • No or limited ability to disaggregate aggregated information • Available information subject to survey • Incomplete recording of uninsured design damages • Lack of information to quantify impact on (asset-)poor but highly vulnerable population groups • Regularly relying on one or few sources • Estimation based on (large) sample from per event and exposed to measurement population of interest Accuracy of data error therein • Explicit quantification of uncertainty in (Accounting biases) • Usually based on approximations estimates without ability to quantify uncertainty or • Subject to systematic (non-classical) accuracy measurement error • Broad coverage across geographies and • Increased scope for harmonized data time collection across contexts Comparability of • Exposed to idiosyncrasies in reporting • Idiosyncrasies in data collection between source data protocols between different data sources different microdata sources are explicit without them typically being explicit Geographical differences • Lower information density in LMICs • Potentially greater accuracy of and lower ability to draw on ancillary microdata in HICs due to use of more data sources for shock recording sophisticated measurement approaches • Systematic under-coverage of events and and higher levels of education among marginalized population groups in respondents LMICs Geographical • Lower accuracy of data due to lower differences density of (independent) sources of information • Underestimation of impacts in LMICs due to higher share of uninsured damages and low value of affected goods and assets 45 Appendix B: Estimated impact for floods or droughts, and for all shocks, in the LSMS-ISA Table B1. Estimated impact of all adverse shocks captured in the LSMS-ISA Estimate share of Estimate number of total population Estimate damages Country Year individuals affected affected by all from all shocks by all shocks shocks 14 million 15.00 % 2011 - 2012 [ 12 ; 17 ] [ 12.40 % ; 17.60 % ] 34 million 34.20 % 34 million 2013 - 2014 [ 30 ; 39 ] [ 29.80 % ; 38.60 % ] [ 30 ; 39 ] Ethiopia 52 million 49.30 % 52 million 2015 - 2016 [ 46 ; 57 ] [ 43.90 % ; 54.60 % ] [ 46 ; 57 ] 31 million 27.60 % 31 million 2018 - 2019 [ 27 ; 36 ] [ 23.60 % ; 31.60 % ] [ 27 ; 36 ] 6 million 39.00 % 2009 - 2010 [5;6] [ 34.20 % ; 43.90 % ] 6 million 36.40 % 2012 - 2013 [5;7] [ 32.00 % ; 40.80 % ] Malawi 9 million 52.30 % 2015 - 2016 [ 7 ; 11 ] [ 41.70 % ; 62.80 % ] 9 million 49.30 % 9 million 2018 - 2019 [ 7 ; 11 ] [ 39.00 % ; 59.50 % ] [ 7 ; 11 ] 4 million 21.90 % 4 million 2014 [3;4] [ 18.90 % ; 25.00 % ] [3;4] Mali 5 million 26.20 % 5 million 2017 [4;6] [ 23.10 % ; 29.30 % ] [4;6] 12 million 70.80 % 12 million 2011 [ 11 ; 13 ] [ 63.70 % ; 77.90 % ] [ 11 ; 13 ] Niger 7 million 36.70 % 7 million 2014 [6;8] [ 30.00 % ; 43.40 % ] [6;8] 18 million 9.20 % 18 million Nigeria 2019 [ 15 ; 22 ] [ 7.50 % ; 11.00 % ] [ 15 ; 22 ] 11 million 25.20 % 2008 [ 10 ; 12 ] [ 22.50 % ; 28.00 % ] 14 million 30.80 % 2010 [ 13 ; 15 ] [ 27.80 % ; 33.90 % ] 11 million 23.90 % Tanzania 2012 [ 10 ; 13 ] [ 21.00 % ; 26.90 % ] 9 million 18.30 % 2014 [ 5 ; 14 ] [ 9.90 % ; 26.60 % ] 14 million 23.70 % 2019 [ 10 ; 19 ] [16.20 % ; 31.20 % ] 268 million 10,611 million All countries 2008 - 2019 [ 251 ; 285 ] [ 8,892 ; 12,330 ] Note: both point estimates and 95% confidence intervals are reported 46