Policy Research Working Paper 10081 The Welfare Implications of COVID-19 for Fragile and Conflict-Affected Areas Chrysostomos Tabakis Gi Khan Ten Joshua D. Merfeld David Newhouse Utz Pape Michael Weber Poverty and Equity Global Practice June 2022 Policy Research Working Paper 10081 Abstract Understanding the impacts of the COVID-19 pandemic on violence were far less likely to report receiving government households’ welfare in areas at the admin-1 level subject to assistance than those in other areas. These findings suggest fragility, conflict, and violence is important to inform pro- that the initial effects of the pandemic exacerbated pre- grams and policies in this context. Harmonized data from existing economic gaps between areas affected by fragility, high-frequency phone surveys indicate that, at the onset conflict, and violence and other areas, indicating that an of the pandemic, a higher fraction of households in areas even larger effort will be necessary in areas affected by fra- affected by fragility, conflict, and violence reported income gility, conflict, and violence to recover from COVID-19, declines and a higher fraction of respondents reported that with implications for funding needs and policy as well as they had stopped working since the beginning of the crisis. program design. Households in areas affected by fragility, conflict, and 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, upape@worldbank.org, and mweber1@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 Welfare Implications of COVID-19 for Fragile and Conflict-Affected Areas1 Chrysostomos Tabakis, Gi Khan Ten, Joshua D. Merfeld, David Newhouse, Utz Pape, and Michael Weber JEL: D31, D74, I31, O15 Keywords: Conflicts, COVID-19, household welfare 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 The ongoing COVID-19 pandemic has profoundly affected economic activity and social interactions around the globe. To combat the pandemic, governments introduced restrictions to socioeconomic activity, and individuals adopted social distancing and modified their behavior in various other ways. As past studies have shown (for example, Bundervoet, Dávalos & Garcia, 2021; Crossley, Fisher & Low, 2021; Egger et al., 2021, Kugler et al., 2021), impacts both across countries and within countries have substantial heterogeneity. In this paper, we investigate the existence of differential impacts of the COVID-19 pandemic on subnational areas based on their fragility, conflict, and violence (FCV) status. We use comprehensive data on nearly 400,000 respondents in 64 countries from the World Bank’s COVID-19 high-frequency phone surveys. A number of studies use survey data to explore the heterogeneity in the impacts of the COVID-19 pandemic both across and within countries. Some of them focus on the experiences of developed countries. For instance, Crossley, Fisher & Low (2021) look at the United Kingdom and find that the lowest income quantiles and minority ethnic groups faced the most severe labor market shocks in the first wave of the pandemic (April–May 2020). Adams-Prassl et al. (2020) exploit real-time survey data from the United Kingdom, the United States, and Germany in March and April 2020 and demonstrate that, within countries, the impacts were highly unequal, exacerbating existing inequalities to the detriment of women and less educated workers. Other authors focus on the implications of the pandemic for developing countries. In particular, Egger et al. (2021) use survey data covering nine low- and middle-income countries (LMICs) in Africa, Asia, and Latin America. They document pronounced declines in employment, income, and food security in all nine LMICs in the early period of the crisis. Khamis et al. (2021) use an early version of the high-frequency phone survey data to track overall labor market impacts and find that the onset of the pandemic had substantial adverse effects on labor markets in all areas included in their study. In a related paper, Kugler et al. (2021) focus on the distributional implications of the crisis and show that female, less educated, and younger workers were initially more severely impacted. Bundervoet, Dávalos & Garcia (2021) is probably the paper closest to ours, as it considers a variety of welfare impacts beyond employment and income. They demonstrate that the pandemic’s initial effects were both widespread and highly regressive, with the most vulnerable segments of the population being disproportionally affected. However, none 2 of the aforementioned studies focuses on FCV areas, which may be particularly vulnerable a priori to economic crises. Therefore, this study, which explores the impact of the pandemic on fragile areas, fills an important gap in the literature on the economic ramifications of the COVID-19 crisis. Our contribution to understanding the economic and social effects of the COVID-19 crisis is twofold. First, fragile and conflict-affected countries contain a significant fraction of the world’s poor. 2 In fact, according to the World Bank, up to two-thirds of the world’s extreme poor could be living in FCV settings by 2030. 3 Given that economic crises are naturally most threatening to economically vulnerable individuals or households, analyzing the impacts of the pandemic on FCV areas adds to our understanding about the channels through which crises affect the poor. Second, the empirical study of socioeconomic conditions in “hot” conflict areas using household survey data is rare since data collection there is inherently difficult (Blattman & Miguel, 2010). The nature of the phone survey data used in this study enables this analysis to make an important contribution towards better understanding the welfare impacts of crises on fragile and conflict-affected areas. Furthermore, using high(er)-frequency phone survey data to shed light on welfare impacts in FCV settings constitutes an approach that can be followed systematically in the future. We first investigate the initial economic and social impacts of the COVID-19 pandemic on FCV versus non-FCV (subnational) areas. The overall picture emerging from the analysis is that FCV areas were hit relatively harder by the COVID-19 crisis in its early months. In particular, we find that areas classified as FCV had a higher fraction of people reporting adverse changes in household income and employment. Moreover, households in FCV areas were much less likely to receive some form of government assistance during the crisis, perhaps reflecting the difficulty of distributing government assistance in FCV settings. We do not uncover, however, any significant relationship between FCV status and food insecurity. Perhaps even more surprisingly, our findings show that, in the early months of the COVID-19 crisis, having an FCV status is associated with people in the region having greater access to medical services when necessary. Further analysis reveals that the reason behind this finding is that FCV areas were less constrained by lack of transportation than non-FCV areas when it came to obtaining medical treatment. These results 2 Many studies document the negative implications of conflict for economic development, at least in the short run. For example, see Abadie & Gardeazabal (2003), Bundervoet, Verwimp & Akresh (2009), Blattman & Miguel (2010), Tandon & Vishwanath (2020), and Yamada & Yamada (2021). 3 See https://www.worldbank.org/en/topic/fragilityconflictviolence/overview#1. 3 suggest that people in FCV areas, due to the conditions typically prevailing there, might have developed effective mechanisms over the years to safeguard their level of food security and their access to essential facilities (at least in the face of transitory shocks). In other words, they have developed ways to prepare for and cope with crises. Second, we look at the recovery (or lack thereof) over time with respect to employment, food security, and access to educational and medical services in FCV versus non-FCV areas. Two main conclusions can be drawn from this portion of the analysis. First, there is no statistically significant difference between FCV and non-FCV areas in the evolution over time of any of the aforementioned welfare indicators. Second, food insecurity declined significantly over time in both FCV and non-FCV areas relative to the baseline period. This preserved the baseline (unconditional) difference in food security between FCV areas and non-FCV ones, to the detriment of the former. The remainder of the paper is structured as follows. Section 2 describes the data used in our analysis. In Section 3, we describe the methodology employed in our analysis. Section 4 presents our findings. Finally, Section 5 offers some concluding remarks. 2. Data 2.1 High-Frequency Phone Survey Since the first quarter of 2020, the World Bank has supported high-frequency phone surveys (HFPS henceforth) designed to monitor the socio-economic impacts of the COVID-19 pandemic. The topics covered include but are not limited to income and its sources, employment, food insecurity, and access to health and educational facilities. In this study, we use the August 2021 vintage that covers nearly 400,000 respondents in 64 countries. Our analysis focuses on 15 outcomes related to sources of income, employment, food insecurity, coping mechanisms, access to basic facilities, and preventive behavior. Total and earned income.—In the HFPS, income-related questions asked respondents whether a given source of income had increased, decreased, stayed the same, or had not been received since the beginning of the pandemic. The four main income categories considered in our analysis are total, wage, farm, and non-farm income. From these questions, we construct two groups of binary indicators. The first group indicates whether a given source of income has gone up since the COVID-19 outbreak. The second group indicates whether a given source of income either has decreased or has not been received at all since the onset of the pandemic crisis. 4 Unearned income.—The two sources analyzed in this paper are public assistance and remittances. The HFPS does not distinguish between the public assistance programs delivered in cash or in- kind, and thus the resulting binary variable captures whether the household has received any form of government support since the COVID-19 outbreak. In principle, the HFPS data permits checking whether, in a given household, the remittances income has gone up, stayed the same, decreased, or has not been received since the beginning of the pandemic. However, in our data set, 74% of families had not been receiving remittances before the pandemic started, and we do not have statistical power to further decompose the variable in question using the remaining sample. Therefore, we re-code the remittances variable, assigning the value of one to those households that have reported an increase/no change/decrease in the amount of remittances received since the beginning of the crisis. Zero is assigned to those families that had been receiving remittances before the pandemic started but have not received any afterward. The resulting variable is a binary indicator of whether a given family has received any flow of remittances during the period of study. Labor and food insecurity.—To access the impact of the COVID-19 pandemic on employment, we make use of the following variables in the HFPS: (i) whether the respondent has stopped working since the beginning of the pandemic, (ii) whether the respondent is currently employed, and (iii) whether in the previous week the respondent worked as usual in his/her wage job either onsite or remotely. To measure food insecurity, we use a question that asks whether, in the last 30 days, there have been adults in the household who went without eating for a whole day because of lack of money or resources. We believe this indicator objectively captures the state of food insecurity in the household, as it does not rely upon the respondent's subjective judgment of the amount of food available in the family. 4 Other outcomes.—To investigate the way households cope with the crisis, we use two variables in the HFPS. The first variable captures whether the household has used the money saved for emergencies to cover basic living expenses. The second variable reflects whether the family has sold assets during the pandemic to cope with the crisis. Both variables mentioned above are binary. 4 In the HFPS data set, there is another candidate measure of food insecurity derived from a question that asks whether there is an adult in the household who had to skip a meal because of lack of money or resources during a given reference period. However, the reference period varies across areas, even within some countries. Possible reference periods are: (i) 7 days, (ii) 30 days, and (iii) 12 months. Due to this, we abstain from using this variable in our analysis. 5 Access to basic facilities is another determinant of the family's well-being that is likely to be affected by the pandemic. To investigate the extent to which lockdown policies and social distancing have affected those aspects of families' lives, we make use of two variables available in the HFPS. The first variable is based on the following survey question: "Has the HH been able to access medical services in the past seven days when needed?". The second variable reflects whether children in the household have been engaged in any form of schooling since school closures. Finally, we also investigate the possibility that preventive behavior differs between the FCV and non-FCV areas. The first outcome is an indicator of whether the respondent has adopted social distancing/self-isolation. The second outcome variable indicates whether the respondent has been avoiding gatherings of more than ten people. It is important to acknowledge that phone surveys are not representative, and there are at least three reasons for that. First, there are several reasons for failing to reach some households by phone—e.g., a refusal to participate in an interview, an inability to pick up a call, or connection disruptions. Unless the failure to reach out to some respondents is purely random, it leads to a non- response bias. For example, Josephson, Kilic & Michler (2021) look at round-specific response rates in Ethiopia, Malawi, Nigeria, and Uganda and find that the share of completed interviews in the total number of attempted phone calls ranges between 60 and 95 percent. Even though HFPS household weights attempt to account for non-response bias, Brubaker, Kilic & Wollburg (2021) show that the interviewed respondents are more likely to be above 25, be household heads, be married, be more literate, own an enterprise, or have a wage job. Therefore, the results presented in this paper should be viewed with these caveats in mind. Next, some households were excluded from the sample due to not owning a mobile phone or being unwilling to participate in the survey. However, in many countries, especially those in Sub- Saharan Africa, HFPS respondents were selected from a sub-sample of past nationally representative surveys. In these cases, sampling weights were adjusted to correct for biases due to differences in observed characteristics, including phone ownership at the baseline. Hence, the adjusted HFPS surveys in these countries approximate the national distribution of households reasonably well. In other countries, such as those in Latin America and the Caribbean, the surveys were carried out through Random Digit Dialing. In these cases, although no adjustment was made 6 to make the data consistent with a nationally representative survey, mobile phone penetration is high. The third source of bias is due to the non-random selection of respondents within the household. As noted by Khamis et al. (2021), Bundervoet, Davalos & Garcia (2021), and Kugler et al. (2021), the phone surveys typically collect individual characteristics from only one respondent, who is in most cases the head of the household. No weighting adjustment was made to adjust for this within- household sample selection bias. This leads to biased estimates of individual level outcomes, as, for example, HFPS estimates of employment rates are substantially upwardly biased as compared with a few phone surveys that collect employment information on all household members (Kugler et al., 2021). Therefore, the analysis of household level outcomes in this analysis, such as income loss and receipt of public assistance, is more likely to be representative of the national population than the analysis of individual outcomes such as employment. However, the individual level results—concerning only employment outcomes in our case—should be interpreted with additional caution. 2.2 Conflict Deaths and Covariates Our analysis aims to investigate the difference in trajectories of welfare indicators in FCV and non-FCVs areas. For this purpose, we first collected data needed to pinpoint the FCV areas in our data. We then gathered information on the number of regional characteristics to hold fixed a list of variables that are likely to confound the relationship of our interest. This subsection describes the nature and sources of data that complement the information from the HFPS in our analysis. Conflict deaths per capita.—From the ACLED database, we obtained information on the total number of conflict deaths in 2019, or a year before the pandemic, for every region in our database. The raw data set contains the coordinates of every conflict-induced death in every region of the world. We collapse the raw ACLED data to the subnational level, thereby aggregating the total number of deaths by subnational areas in our sample. From the WorldPop project, we obtained the grid cell-based 2019 population counts, at 100m resolution. With the use of the information on subnational borders, we obtain estimates of population counts at the regional (subnational) level. Finally, we normalize the total number of deaths by population size to obtain the number of conflict-induced fatalities per capita. 7 Control variables.—The link between the overall economic performance and conflict is well- established in the literature (Collier & Hoeffler, 2004; Miguel, Satyanath & Sergenti, 2004; Blattman & Miguel, 2010). Thus, a naïve regression of a given welfare measure on the binary indicator of the FCV region might merely reveal differences driven by pre-existing poverty and weak state capacity. To address the issue in question at least partially, we collected a list of indicators intended to capture various dimensions of the pre-existing economic prosperity of subnational areas in our sample. From the HFPS data set, we obtained a set of socio-economic characteristics of areas that are plausibly determined prior to the COVID-19 outbreak. Those characteristics are the fraction of the population with a secondary degree, the pre-pandemic employment rate, average household size, and a binary indicator of an urban area. We complemented the regional figures obtained from the HFPS with two remotely sensed indicators. The first is Visible Infrared Imaging Radiometer Suite (VIIRS) nighttime lights (average radiance grids in 2019), which are a recognized proxy for economic growth (Henderson, Storeygard & Weil, 2012). We use the Annual VIIRS Nighttime Lights Version 2 product that corrects data influenced by stray lights, excludes observations affected by cloud cover, and has greater spatial coverage (see Mills, Weiss & Liang, 2013). The second indicator is the fraction of land covered by crops in 2019, intended to proxy for regional agricultural development. 5 Finally, we completed the list of controls with the country-level GDP per capita (PPP adjusted) obtained from the World Development Indicators database. After restricting the HFPS data set to those subnational areas for which we have the complete set of control variables, we ended up with 21 countries in our sample. 6 3. Methodology 3.1 Definition of the Onset of the COVID-19 Pandemic The first goal of our empirical analysis is to understand the initial socio-economic impacts of the COVID-19 pandemic on FCV areas (relative to non-FCV ones). This subsection describes two methodological challenges that are likely to complicate our analysis and our effort to address them. 5 The data on land coverage of crops was obtained from the Copernicus program’s Global Land Cover Characterization database. The variable provides information on the fraction of land covered with temporary crops followed by harvest and a bare soil period (e.g., single and multiple cropping systems). 6 Argentina, Chile, Costa Rica, the Dominican Republic, Ecuador, the Arab Republic of Egypt, Ghana, Indonesia, Madagascar, Malawi, Mexico, Mozambique, Myanmar, Paraguay, Peru, Poland, Sierra Leone, Solomon Islands, Tunisia, Uganda, and Zambia. 8 To understand the initial socio-economic consequences of the pandemic, it is crucial to define the pandemic's beginning given that (i) the timing of the pandemic's onset differs by country and (ii) countries appear in the HFPS at different points of time. Thus, in the first step, we direct our effort towards limiting our sample to the list of countries observed during the onset of the COVID-19 pandemic. Our definition of the time window that corresponds to the onset of the pandemic closely follows the work of Bundervoet, Davalos & Garcia (2021). First, for each country in our sample, we extracted time-series of the stringency index from Oxford's "Coronavirus Government Response Tracker" (Hale et al., 2021). The index in question informs about the strictness of lockdown policies implemented in a given country/point of time, with a higher value corresponding to more stringent measures. Figure 1 shows the time series of the stringency index for all countries for which data is available, by income group. As shown in the figure, on average, most countries adopted the peak stringency measures at the end of the first quarter of 2020, gradually lifting their lockdown measures starting from the second quarter of the same year. Second, for each country, we define the peak stringency period as the (first) month when the stringency index reached its maximum value. Finally, we restrict the HFPS sample to the list of those countries surveyed within three months relative to the peak stringency month. If multiple survey waves of a given country fall within the time window in question, we pick only the first cross-section. As a result, we are left with a subsample of countries observed as closely as possible to their respective period of peak stringency measures. 3.2 Initial Impacts The main threat to our analysis is the possibility of picking pre-existing economic disadvantages peculiar to the FCV areas. For this reason, we utilize a list of observable regional controls in an attempt to isolate the variation of the regional FCV status from the variation of the potential confounders. We begin by focusing on the initial impact of the crisis. As a first step, we collapse our HFPS data to the regional (state or province) level, as the FCV status varies at the subnational level. By construction, every region appears in the sample only once. Thus, we end up having a single cross- section of areas observed within three months from their country-specific peak stringency month. We then fit the following model using our final sample of areas: 9 ′ (1) + + , = is a binary indicator of where is the welfare measure observed in region r. The variable conflict fatalities exceeding 10 per 10,000 people in 2019. The term ′ is a row vector of controls. At the subnational level, we control for remotely sensed indicators (nighttime lights and crops’ coverage in 2019), the fraction of individuals with at least secondary education, the pre-pandemic employment rate, average household size, urban dummy, and survey month (linear term). At the country level, we control for the natural log of GDP per capita in 2019 (PPP adjusted) and a set of binary indicators of a given country’s region. is the error term that we allow to be arbitrarily correlated at the country level. 7 3.3 Evolution of the Crisis We also look at the recovery (or lack thereof) over time relative to the baseline period—i.e., the period over which our analysis of the pandemic’s initial impacts is conducted—of employment, food security, and access to educational and medical services in FCV versus non-FCV areas. The idea is motivated by two mutually exclusive possibilities. On the one hand, the COVID-19 pandemic may have induced persistent disruptions of the FCV areas’ economic well-being. On the other hand, a possibly steeper fall observed across a range of welfare indicators during the onset of the pandemic might imply their faster recovery. To investigate the possibilities mentioned above, we proceed as follows. We first define the baseline period as the time interval that corresponds to three months before or after the country- specific peak stringency period. Let us call this baseline period as Wave 1. We then code any cross- section of a given country’s areas observed within three months relative to its country’s peak stringency month as the one observed in Wave 1. If a given country is observed within 4 to 6 months relative to its peak stringency month, we categorize it as a cross-section of areas observed in Wave 2. Lastly, any country observed within 7 to 9 months relative to its peak stringency month is coded as an observation in Wave 3, which is the last time period. Table 1 shows the resulting distribution of the areas in our sample by FCV status and wave. 7 We do not have enough countries in our sample to apply the conventional cluster-robust inference. For this reason, following Cameron and Miller (2015), we implement the wild cluster bootstrap procedure to compute the standard errors of our regression coefficients. 10 Having constructed a region-wave panel data set, we estimate the following empirical specification: () = () + + × () + + () , (2) where is the outcome of interest observed in region , country , and wave . We control for the average stringency index, , as well as the region fixed effect, () . The term is the wave variable. Thus, the term can be interpreted as the change in the outcome of interest in the non-FCV areas relative to the baseline period, conditional on the prevailing level of stringency measures and unobserved stable characteristics of areas. The + × () term captures the same information for the FCV areas. Finally, () is the error term that we allow to be arbitrarily correlated at the country level. To investigate the sensitivity of our findings to an alternative definition of the survey wave, we re- code our monthly observations of subnational areas as follows. First, the baseline period (Wave 1) is defined as January–April 2020. Next, any region observed within May–July/August–November 2020 is coded as an observation in Wave 2/Wave 3. Finally, any region observed after November 2020 is coded as an observation in Wave 4. As we show later, our results largely hold when using this alternative definition of the survey wave. There are at least two limitations that constrain our ability to paint a detailed picture. The first limitation is the number of areas in the FCV group. As Table 1 shows, the number of FCV areas is disproportionately smaller in each wave than that of non-FCVs. The resulting variance in FCV status is unlikely to be sufficiently high to estimate the parameters of interest precisely. The second limitation is related to the number of available outcome variables. Due to the latter data limitation, our choice of welfare indicators is restricted to the following list: current employment, work as usual, food insecurity, and access to education and health facilities. 4. Results 4.1 Initial Impacts In this section, we present the results of our analysis of the initial economic and social impacts of the COVID-19 pandemic on FCV versus non-FCV areas. In the main body of the text, we report the regression results with each country being given the same weight. 11 In Table 2, we look at changes relative to the pre-pandemic period in total household income (column 1) as well as in different types thereof (columns 2–4). Our findings clearly demonstrate that, in the early months of the pandemic, having an FCV status is associated with more people in the region experiencing adverse changes in household income. In particular, it is associated with a larger proportion of individuals (14.6 percentage points) facing a drop in total household income and with a smaller share of them (5.6 percentage points) enjoying a household income increase— with the mean value of the latter variable in non-FCV areas being only 5.7%, which is a reflection of the magnitude of the economic downturn due to the pandemic in the period under consideration. With respect to different sources of household income, the estimated FCV effect on the share of individuals having an increase in household income from wage employment is −4.4 percentage points and is statistically significant at the 1% level; in the case of an increase in non-farm family business income, the estimated effect in question is −2.0 percentage points and is also significant at the 1% level—with the mean shares in non-FCV areas being 3.8% and 3.6%, respectively. In all other cases, the FCV effect is statistically insignificant. We next turn to employment and food security and examine the existence of differential impacts of the COVID-19 pandemic on them based on FCV status. According to column 1 of Table 3, the estimated FCV effect on the share of individuals stopping to work is very substantial, equaling 14.5 percentage points, and is statistically significant at the 1% level. In column 2 of the same table, we look at the proportion of individuals having worked as usual in their wage job in the previous week (including remotely). Workers in FCV areas were on average about 10 percentage points more likely to work at their usual jobs, although the difference is not precisely estimated and therefore not statistically significant. In contrast, there is almost no difference between FCV and non-FCV areas when looking at food insecurity, as measured by adults going for an entire day without eating. This finding might seem surprising at first sight given our results regarding income changes. Nevertheless, a plausible explanation for this is that people in FCV areas, due to the economic conditions typically prevailing there, have developed effective mechanisms over the years to safeguard their level of food security in the face of transitory shocks. Alternatively, this finding may simply be due to the greater provision of humanitarian aid to FCV areas, but the information available in HFPS does not permit us to further probe into this possibility. 12 In Table 4, we focus on various social safety nets and coping mechanisms that can assist vulnerable households in dealing with the adverse effects of economic shocks (such as the COVID-19 pandemic). More specifically, we look at the receipt of any income from remittances (column 1), the receipt of any form of government assistance (column 2), the use of savings (column 3), and the sale of assets (column 4). The estimated coefficient of the FCV dummy is positive in column 1 but negative in columns 2–4 (albeit statistically insignificant in column 3). The result reported in column 2 is particularly striking and is indicative of the scarcity of resources at the disposal of governments in FCV settings: the estimated FCV effect on the share of individuals receiving some government assistance in the early period of the crisis is −12.4 percentage points (significant at the 1% level), with the mean value of the variable in question in non-FCV areas being 17.3%. Last, we turn to households’ access to medical and educational services (Table 5, columns 1–2) as well as to individuals’ preventive behaviors against COVID-19 (Table 5, columns 3–4). The estimated coefficient of the FCV dummy is positive in all four columns of Table 4 but is statistically significant only when it comes to (i) households having been able to access medical services in the previous week when needed (at the 1% level) and (ii) individuals having adopted social distancing/self-isolation (at the 10% level). In other words, in the period under consideration, having an FCV status is associated with people in the region having greater access to medical services when necessary and adopting social distancing more. The former result— including the magnitude of the coefficient, which equals 0.114—seems counterintuitive at first. In Table 6, we look at different factors that prevented some households in both types of areas from accessing medical treatment. Our findings suggest that the result in question is driven by lack of transportation having constrained people less in FCV areas than in non-FCV ones when it came to obtaining medical treatment. This might stem from the fact that FCV areas are characterized by relatively (very) limited public transportation infrastructure, forcing people to access health and other essential facilities mostly via alternative means of transportation. As a consequence, with respect to accessing essential facilities, it is reasonable to expect that any disruptions in public transportation (such as the ones caused because of the COVID-19 lockdowns) represent a far bigger obstacle in non-FCV areas than in FCV ones. 4.2 Evolution of the Crisis 13 We begin with the labor market and look at two different indicators of labor market conditions. More specifically, we investigate the change over time relative to the baseline period (Wave 1) in the share of people in non-FVC and FCV areas currently employed (see Figure 2) or having worked as usual in their wage job (at their place of work or remotely) in the previous week (see Figure 3). Three main conclusions can be drawn from Figures 2–3. First, current employment in FCV areas increased about 5 percentage points more than in non-FCV areas in wave 2 and nearly 10 percentage points more in wave 3. In both waves, the increase in employment in FCV areas is statistically significant (Figure 2.A), though the difference between FCV and non-FCV areas is not (Figure 2.B). Second, with respect to people working as usual in their wage job (including remotely), the increase in FCV areas was moderately larger than in non-FCV areas by wave 3 (roughly, 6.2 percentage points). However, the estimates for FCV areas are fairly imprecise, and the difference between FCV and non-FCV areas is not statistically significant. We now turn to food insecurity, defined as the existence of hungry adults in the household having gone without eating for a whole day due to lack of money or resources in the previous 30 days. As panel A of Figure 4 shows, there was a moderate decline of about 3.4 and 3 percentage points in the second wave in FCV and non-FCV areas, respectively, while the decline strengthened in both types of areas in the third wave. The point estimates for FCV and non-FCV areas are similar to each other in both Wave 2 and Wave 3 (see Figure 4, panel B), preserving the baseline difference in food security between the two types of areas (to the detriment of FCV areas; see Figure A3 in the Appendix). Next, in Figures 5–6, we look at the change over time as compared with the baseline period in the share of respondents in FCV and non-FCV areas whose household was able to access medical services when needed in the previous week (see Figure 5) or whose children engaged in any learning/educational services after the closure of schools (see Figure 6). Figure 5 shows that both FCV and non-FCV areas broadly recovered in terms of the ability to access medical services, and the differences between them are small and not statistically significant. For children engaged in learning, however, non-FCV areas saw a stronger and more enduring recovery, as in wave 3 non- FCV areas had recovered while FCV areas saw a decline of about 10.2 percentage points relative to the baseline period. However, the estimates are not precise, and the differences are not statistically significant. 14 Finally, we assess the sensitivity of our findings to the use of the alternatively defined survey wave (see Subsection 3.3). In Panel A of Table B1, we report our baseline regression results shown earlier in Figures 2–6. Panel B of the same table shows the findings obtained after using the alternative definition of the survey wave. Broadly, the findings are consistent across the two panels of Table B1, with the exception of food insecurity in column 3. More precisely, when using the alternatively defined survey wave, we find that food security recovered more slowly in FCV areas. However, note that the estimated coefficient of the × term is not significantly different from zero in either of the two panels. 5. Conclusion This paper analyzed the differential impacts of the COVID-19 pandemic on subnational areas based on their FCV status. The analysis is based on a comprehensive phone survey data set of nearly 140,000 respondents in 21 countries. The results show that FCV areas were hit relatively harder than non-FCV areas by the pandemic in its early months. During this period, respondents in FCV areas were more likely to report adverse changes in household income and employment. Exacerbating the relative situation of FCV areas at the onset of the crisis, the estimated FCV effect on the share of individuals receiving government assistance is both negative and very substantial. On the other hand, our findings reveal—perhaps surprisingly—that having an FCV status is associated with people in the region having relatively greater access to medical services when needed in the early months of the pandemic, likely because people in FCV areas have established robust means to access medical services even in the context of crises. We also examined the evolution over time of different welfare indicators in both types of areas. The point estimates show substantial recovery in FCV areas in employment and working as usual, although differences between FCV and non-FCV areas are not statistically significant. There are signs that children in FCV areas were lagging in terms of engaging in learning activities, although the estimates are imprecise and the difference between FCV and non-FCV areas is not statistically significant. It will be important to continue monitoring these outcomes in FCV areas to see if the nascent employment recovery is sustained and whether more children engage in learning. The continued addition of new phone survey data will also allow the effects to be estimated with more precision. Taken together, the results highlight that FCV areas experienced this crisis in different ways than non-FCV areas, which should be taken into consideration when designing relief measures. 15 References Abadie, A., and Gardeazabal, J. (2003). The economic costs of conflict: A case study of the Basque Country. American Economic Review, 93(1), 113–132. Adams-Prassl, A., Boneva, T., Golin, M., & Rauh, C. (2020). Inequality in the impact of the coronavirus shock: Evidence from real time surveys. Journal of Public Economics, 189, 104245. Blattman, C., and Miguel, E. (2010). Civil war. Journal of Economic Literature, 48(1), 3–57. Brubaker, J., Kilic, T., & Wollburg, P. (2021). Representativeness of individual-level data in COVID-19 phone surveys: Findings from Sub-Saharan Africa. PLoS ONE, 16(11), e0258877. 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. Bundervoet, T., Verwimp, P., and Akresh, R. (2009). Health and civil war in rural Burundi. Journal of Human Resources, 44(2), 536–563. Cameron, A. C., and Miller, D. L. (2015). A practitioner’s guide to cluster-robust inference. Journal of Human Resources, 50(2), 317–372. Collier, P., and Hoeffler, A. (2004). Greed and grievance in civil war. Oxford Economic Papers, 56(4), 563–595. Crossley, T. F., Fisher, P., and Low, H. (2021). The heterogeneous and regressive consequences of COVID-19: Evidence from high quality panel data. Journal of Public Economics, 193, 104334. Egger, D., Miguel, E., Warren, S. S., Shenoy, A., Collins, E., Karlan, D., Parkerson, D., Mobarak, A.M., Fink, G., Udry, C., Walker, M., Haushofer, J., Larreboure, M., Athey, S., Lopez- Pena, P., Benhachmi, S., Humphreys, M., Lowe, L., Meriggi, N.F., … Vernot, C. (2021). Falling living standards during the COVID-19 crisis: Quantitative evidence from nine developing countries. Science Advances, 7(6). Hale, T., Angrist, N., Goldszmidt, R., Kira, B., Petherick, A., Phillips, T., Webster, S., Cameron- Blake, E., Hallas, L., Majumdar, S., and Tatlow, H. (2021). A global panel database of 16 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. Josephson, A., Kilic, T., and Michler, J. D. (2021). Socioeconomic impacts of COVID-19 in low- income countries. Nature Human Behaviour, 5(5), 557–565. Khamis, M., Prinz, D., Newhouse, D., Palacios-Lopez, A., Pape, U., and Weber, M. (2021). The early labor market impacts of COVID-19 in developing countries. Jobs Working Paper No. 58. World Bank, Washington, DC. 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. Miguel, E., Satyanath, S., and Sergenti, E. (2004). Economic shocks and civil conflict: An instrumental variables approach. Journal of Political Economy, 112(4), 725–753. Mills, S., Weiss, S., and 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. Tandon, S., and Vishwanath, T. (2020). The evolution of poor food access over the course of the conflict in Yemen. World Development, 130, 104922. Yamada, T., and Yamada, H. (2021). The long-term causal effect of U.S. bombing missions on economic development: Evidence from the Ho Chi Minh Trail and Xieng Khouang Province in Lao P.D.R. Journal of Development Economics, 150, 102611. 17 Figure 1 – Stringency Index 18 Figure 2 – Currently Employed NOTES: The unit of observation is region-wave. FCV is a binary indicator of conflict fatalities exceeding 10 per 10,000 people in 2019. At the subnational level, the controls comprise regional fixed effects. At the country level, we control for the Oxford Stringency Index lagged by one month. Standard errors are adjusted for clustering at the country level. Regressions are weighted by the inverse of the number of areas in a given country. 19 Figure 3 – Work as Usual NOTES: The unit of observation is region-wave. FCV is a binary indicator of conflict fatalities exceeding 10 per 10,000 people in 2019. At the subnational level, the controls comprise regional fixed effects. At the country level, we control for the Oxford Stringency Index lagged by one month. Standard errors are adjusted for clustering at the country level. Regressions are weighted by the inverse of the number of areas in a given country. 20 Figure 4 – Food Insecurity NOTES: The unit of observation is region-wave. FCV is a binary indicator of conflict fatalities exceeding 10 per 10,000 people in 2019. At the subnational level, the controls comprise regional fixed effects. At the country level, we control for the Oxford Stringency Index lagged by one month. Standard errors are adjusted for clustering at the country level. Regressions are weighted by the inverse of the number of areas in a given country. 21 Figure 5 – Access to Health Facilities NOTES: The unit of observation is region-wave. FCV is a binary indicator of conflict fatalities exceeding 10 per 10,000 people in 2019. At the subnational level, the controls comprise regional fixed effects. At the country level, we control for the Oxford Stringency Index lagged by one month. Standard errors are adjusted for clustering at the country level. Regressions are weighted by the inverse of the number of areas in a given country. 22 Figure 6 – Access to Education NOTES: The unit of observation is region-wave. FCV is a binary indicator of conflict fatalities exceeding 10 per 10,000 people in 2019. At the subnational level, the controls comprise regional fixed effects. At the country level, we control for the Oxford Stringency Index lagged by one month. Standard errors are adjusted for clustering at the country level. Regressions are weighted by the inverse of the number of areas in a given country. 23 Table 1: Distribution of the HFPS Areas by FCV Status/Wave # of adm1 locations Distance to the peak Period Non- stringency month FCV FCV -3 6 0 -2 71 3 -1 75 1 0 Wave 1 185 3 1 219 18 2 322 52 3 481 38 4 412 26 5 Wave 2 216 22 6 124 16 7 214 13 8 Wave 3 117 11 9 72 12 NOTES: This table reports the distribution of subnational areas by FCV status and the distance to the country-specific first month of peak stringency measures. Details are given in the text. 24 Table 2: Income Variables (1) (2) (3) (4) Total Farm Wage Non-farm Panel A: Income Decrease FCV 0.146* -0.069 0.025 -0.029 (0.075) (0.053) (0.058) (0.085) Baseline mean 0.614 0.654 0.488 0.780 N 283 271 274 322 Panel B: Income Increase FCV -0.056* -0.021 -0.044** -0.020** (0.027) (0.026) (0.008) (0.005) Baseline mean 0.057 0.066 0.038 0.036 N 283 271 274 322 NOTES: The unit of observation is the subnational region. Panel A reports the results for income decrease. Panel B reports the results for income decrease. FCV is a binary indicator of conflict fatalities exceeding 10 per 10,000 people in 2019. At the subnational level, the controls comprise remotely sensed indicators (nighttime lights and crops’ coverage in 2019), the fraction of individuals with at least secondary education, the pre-pandemic employment rate, average household size, an urban dummy, and the survey month (linear term). At the country level, the controls comprise the natural log of GDP per capita in 2019 (PPP adjusted) and a set of binary indicators of a given country’s region. Standard errors are adjusted for clustering at the country level. Regressions are weighted by the inverse of the number of areas in a given country. +, *, and ** denote significance at the 10%, 5%, and 1% level, respectively. 25 Table 3: Labor and Food Security (1) (2) (3) Stopped Working Usual Hours Whole Day w/o Food FCV 0.145** 0.099 0.006 (0.049) (0.085) (0.047) Baseline mean 0.318 0.778 0.155 N 374 340 307 NOTES: The unit of observation is the subnational region. FCV is a binary indicator of conflict fatalities exceeding 10 per 10,000 people in 2019. At the subnational level, the controls comprise remotely sensed indicators (nighttime lights and crops’ coverage in 2019), the fraction of individuals with at least secondary education, the pre-pandemic employment rate, average household size, an urban dummy, and the survey month (linear term). At the country level, the controls comprise the natural log of GDP per capita in 2019 (PPP adjusted) and a set of binary indicators of a given country’s region. Standard errors are adjusted for clustering at the country level. Regressions are weighted by the inverse of the number of areas in a given country. +, *, and ** denote significance at the 10%, 5%, and 1% level, respectively. 26 Table 4: Assistance Programs/Coping Mechanisms (1) (2) (3) (4) Remittances Public Assistance Savings Sold Assets FCV 0.101** -0.124** -0.040 -0.025* (0.025) (0.047) (0.069) (0.013) Baseline mean 0.863 0.173 0.205 0.059 N 247 338 249 259 NOTES: The unit of observation is the subnational region. FCV is a binary indicator of conflict fatalities exceeding 10 per 10,000 people in 2019. At the subnational level, the controls comprise remotely sensed indicators (nighttime lights and crops’ coverage in 2019), the fraction of individuals with at least secondary education, the pre-pandemic employment rate, average household size, an urban dummy, and the survey month (linear term). At the country level, the controls comprise the natural log of GDP per capita in 2019 (PPP adjusted) and a set of binary indicators of a given country’s region. Standard errors are adjusted for clustering at the country level. Regressions are weighted by the inverse of the number of areas in a given country. +, *, and ** denote significance at the 10%, 5%, and 1% level, respectively. 27 Table 5: Access to Facilities/Preventive Behavior (1) (2) (3) (4) Access to Health Access to Education Social Distancing Avoid Gatherings FCV 0.114** 0.050 0.012+ 0.005 (0.043) (0.084) (0.007) (0.018) Baseline mean 0.796 0.730 0.948 0.860 N 314 357 191 293 NOTES: The unit of observation is the subnational region. FCV is a binary indicator of conflict fatalities exceeding 10 per 10,000 people in 2019. At the subnational level, the controls comprise remotely sensed indicators (nighttime lights and crops’ coverage in 2019), the fraction of individuals with at least secondary education, the pre-pandemic employment rate, average household size, an urban dummy, and the survey month (linear term). At the country level, the controls comprise the natural log of GDP per capita in 2019 (PPP adjusted) and a set of binary indicators of a given country’s region. Standard errors are adjusted for clustering at the country level. Regressions are weighted by the inverse of the number of areas in a given country. +, *, and ** denote significance at the 10%, 5%, and 1% level, respectively. 28 Table 6: Access to Health (Reasons) (1) (2) (3) (4) (5) Lack of No Afraid of Lack of Money Personnel C19 Stringency Transportation FCV -0.023 -0.031 -0.025 -0.064 -0.062** (0.031) (0.071) (0.019) (0.059) (0.021) Baseline mean 0.147 0.266 0.188 0.111 0.078 N 251 251 203 208 179 NOTES: The unit of observation is the subnational region. FCV is a binary indicator of conflict fatalities exceeding 10 per 10,000 people in 2019. At the subnational level, the controls comprise remotely sensed indicators (nighttime lights and crops’ coverage in 2019), the fraction of individuals with at least secondary education, the pre-pandemic employment rate, average household size, an urban dummy, and the survey month (linear term). At the country level, the controls comprise the natural log of GDP per capita in 2019 (PPP adjusted) and a set of binary indicators of a given country’s region. Standard errors are adjusted for clustering at the country level. Regressions are weighted by the inverse of the number of areas in a given country. +, *, and ** denote significance at the 10%, 5%, and 1% level, respectively. 29 Appendix Figure A1 – Currently Employed, Baseline Averages 30 Figure A2 – Work as Usual, Baseline Averages 31 Figure A3 – Food Insecurity, Baseline Averages 32 Figure A4 – Access to Health Facilities, Baseline Averages 33 Figure A5 – Access to Education, Baseline Averages 34 Table B1: Results with Different Definitions of the Baseline Period (1) (2) (3) (4) (5) Children Went out Access to Current Work Engaged in w/o Eating Medical Work Usual ANY Form Whole Day Services of Learning Panel A: Baseline period: +/– 3 months relative to the peak stringency month 0.045 0.033* -0.030** 0.024 0.011 (0.031) (0.014) (0.007) (0.019) (0.037) × 0.050 0.031 -0.004 0.008 -0.062 (0.037) (0.044) (0.014) (0.041) (0.057) Baseline mean 0.586 0.827 0.158 0.810 0.684 Observations 2629 1916 1974 1837 1892 Panel B: Baseline period: January–April 2020 0.035 0.027+ -0.029** 0.007 0.035 (0.028) (0.016) (0.008) (0.028) (0.056) × 0.065 0.038 0.006 0.019 -0.035 (0.044) (0.042) (0.013) (0.053) (0.065) Baseline mean 0.561 0.603 0.233 0.795 0.516 Observations 2629 1916 1974 1837 1892 NOTES: The unit of observation is the subnational region/month of the year. Panel A reports the results with the baseline period (Wave 1) defined as the time interval that corresponds to three months before or after the country-specific peak stringency period. If a given country is observed within 4 to 6 months relative to its peak stringency month, we categorize it as a cross-section of areas observed in Wave 2. Lastly, any country observed within 7 to 9 months relative to its peak stringency month is coded as an observation in Wave 3, which is the last time period. Panel B reports the results with the baseline period (Wave 1) defined as January–April 2020. Any region observed within May–July/August–November 2020 is coded as an observation in Wave 2/Wave 3. Any region observed after November 2020 is coded as an observation in Wave 4. At the subnational level, the controls comprise regional fixed effects. At the country level, we control for the Oxford Stringency Index lagged by one month. Standard errors are adjusted for clustering at the country level. Regressions are weighted by the inverse of the number of areas in a given country. +, *, and ** denote significance at the 10%, 5%, and 1% level, respectively. 35