COVID-19 and Food Security in Ethiopia: Do Social Protection Programs Protect?

We assess the ability of Ethiopia’s flagship social protection program, the Productive Safety Net Program (PSNP), to mitigate the adverse impacts of the COVID-19 pandemic on food and nutrition security of households, mothers, and children. We use both prepandemic in-person household survey data and a postpandemic phone survey. Employing a household fixed effects difference-in-differences approach, we find that household food insecurity increased by 11.7 percentage points and the size of the food gap increased by 0.47 months in the aftermath of the onset of the pandemic. Participation in the PSNP offsets virtually almost all of this adverse change; the likelihood of becoming food insecure increased by only 2.4 percentage points for PSNP households, and the food gap increased by only 0.13 months. The protective role of the PSNP was greater for poorer households and those living in remote areas. Results are robust to definitions of PSNP participation, different estimators, and how we account for the nonrandomness of mobile phone ownership. Furthermore, PSNP households were less likely to reduce expenditures on health and education by 7.7 percentage points and were less likely to reduce expenditures on agricultural inputs by 13 percentage points.


I. Introduction
The COVID-19 pandemic is testing global food and social protection systems at an unprecedented scale. The spread of the pandemic is disrupting food systems and undermining food and nutrition security of households (Amjath-Babu et al. 2020;Barrett 2020;Béné 2020;Devereux, Béné, and Hoddinott 2020;GAIN 2020;Reardon et al. 2020;Swinnen 2020;Torero 2020). There is concern that developing countries with poor health care and more limited social protection systems may be especially badly affected. Recent projections show that globally the pandemic is likely to push 110-150 million people into extreme poverty by 2021 (World Bank 2020), a third of these being from sub-Saharan Africa. Globally, the pandemic is projected to double the number of people facing acute food insecurity by the end of 2020 (about 135 million people before the crisis; WFP 2020). This study was supported by funding from the CGIAR Research Program on Policies, Institutions, and Markets, which is led by the International Food Policy Research Institute, and the Research Support Budget from the Development Research Group of World Bank. Berhane and Hoddinott thank the Bill and Melinda Gates Foundation for their funding for work on the 2019 survey (award OPP1162182). Abay acknowledges funding from the Partnership for Economic Policy, which is financed by the Department for International Development of the United Kingdom (UK Aid) and the International Development Research Centre of Canada. The findings, interpretations, and conclusions expressed in this paper are entirely ours. 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. We are grateful to Abraha Weldegerima, Teame Tesfay, and Mehari Abay, who helped in collecting and cleaning the data. Data are available through Dataverse (https://doi.org/10.7910/DVN/YUVCNI and https://doi.org/10.7910/DVN/JODCEJ). Contact the corresponding author, John Hoddinott,at jfh246@cornell.edu. There are at least three ways in which the pandemic may affect household food security. First, individuals in some households may contract the virus, and this will have both direct economic impacts, such as loss of earnings, and indirect impacts due to increases in medical expenses. Even in the absence of direct contraction, fear of contracting the virus could reduce income-generating activities. Second, government restrictions on movement and gatherings aimed at slowing the spread of COVID-19 have disrupted livelihood activities, thereby reducing household incomes (Abay, Tafere, and Woldemichael 2020;Amare et al. 2020;Arndt et al. 2020;World Bank 2020). Third, access to food has been affected by disruptions to markets and food value chains (e.g., Aggarwal et al. 2020;Mahajan and Tomar 2020). However, empirical evidence on the magnitude of the impact of the pandemic on household food security remains scant, partly because the pandemic is still unfolding and detailed household survey data are not available yet.
Within this context, this paper makes three contributions. First, we add to the small but growing literature on the impact of the pandemic on household food security in sub-Saharan Africa, specifically rural Ethiopia. Second and most importantly, we assess the effectiveness of a social protection intervention in mitigating these malign impacts. Gentilini et al. (2020) note that since the outbreak of the pandemic, more than 150 countries and territories have implemented or announced plans to implement social protection measures, yet little is known about the effectiveness of these interventions. Similarly, in the context of the COVID-19 pandemic, the mechanisms through which social protection and safety nets benefit recipients is ambiguous (Banerjee et al. 2020). 1 Third, we assess the impact of the pandemic on the diets of women and preschool children. As is well understood, the impact of shocks on households can have unequal effects on individual household members (Alderman et al. 1995;Hoddinott 2006), and our data allow us to assess some of the intrahousehold impacts of COVID-19.
In March and August 2019, we conducted face-to-face surveys with mothers of children under the age of 24 months to assess how access to Ethiopia's flagship social protection program, the Productive Safety Net Program (PSNP), had affected their food security and nutritional status. In June 2020, we reinterviewed these mothers by phone. Our data cover drought-prone areas of Ethiopia, spanning the four main highland regions of the country. Having access to prepandemic data allows us to assess the extent to which household food security and diets of individual members have changed following the start of the pandemic in Ethiopia. The longitudinal nature of our data, together with a sample that included both PSNP and non-PSNP households, allows us to combine a difference-in-differences approach with a household fixed effects estimator, allowing us to control for a wide range of confounding factors.
We find that the percentage of households reporting a food gap, a widely used measure of food insecurity in Ethiopia, increased by 11.7 percentage points among program nonbeneficiaries, and the size of the food gap increased by 0.47 months. Participation in the PSNP offsets virtually almost all of this adverse change; the likelihood of becoming food insecure increased by only 2.4 percentage points for PSNP households, and the duration of the food gap increased by only 0.13 months. The protective role of the PSNP in food security is higher for poorer households and those living in remote areas. Results are robust to alternative definitions of PSNP participation, different estimators, and how we account for the nonrandomness of mobile phone ownership. PSNP households were less likely to reduce expenditures on health and education by 7.7 percentage points and were less likely to reduce expenditures on agricultural inputs by 13 percentage points. By contrast, mothers' and children's diets changed little despite some changes in the composition of diets, as consumption of animal source foods declined significantly. Our findings highlight the value of having a well-functioning social protection program in place before the pandemic to protect the food security of poor households.
The paper is organized as follows. Section II describes the PSNP in Ethiopia and details our survey design and implementation. Section III provides contextual information on the pandemic in Ethiopia, how it affected the households in our sample, and how we define PSNP participation. Section IV describes our estimation strategy. Section V presents our results, and we end with concluding remarks in section VI.

II. The PSNP and Survey Design
A. The PSNP The PSNP is Ethiopia's flagship rural food security program benefiting approximately 8 million people. It was established in 2005 by the Ethiopian government and its development partners in response to the recurrent food insecurity in rural areas, which used to be addressed through ad hoc appeals for emergency food aid. While the emergency food aid had been instrumental in averting starvation, it had failed to fundamentally improve the vulnerability of rural households to food shortages. The PSNP aims to address food insecurity and vulnerability to shocks while also enabling households to build assets, which may enable them to escape from poverty. The PSNP aims to protect and promote food insecure households by facilitating consumption smoothing and asset accumulation (Gilligan et al. 2009;Hoddinott et al. 2012;Berhane et al. 2014). It provides predictable transfers to qualifying households in food program areas.
The PSNP is targeted both geographically-operating in woredas (districts) that are considered chronically food insecure-and at the household level, with households selected for inclusion according to a series of criteria of which household food insecurity is particularly important. Households targeted for PSNP participation are historically food insecure and have low household asset holdings (e.g., land, oxen) and limited income from alternative sources of employment. Historical data on food aid allocations are used to identify drought-prone woredas as well as determine the number of beneficiaries within these woredas. Community-based targeting is then used to select eligible households. The beneficiary list is updated annually to reflect historical and contemporaneous food security status of households in the community (Berhane et al. 2014(Berhane et al. , 2020. Eligible households that have able-bodied workers are employed in laborintensive public works for about 6 months per year. A small fraction of beneficiaries (approximately 15%) whose main income earners are elderly or disabled are provided unconditional payments. Payments are largely made in cash, but in-kind (food) payments are also used in localities that have poor access to markets. Between January and August 2019, the average payment per household was ETB 3,648 (approximately US$128). About 73% of these payments were cash, and the rest were in-kind transfers (Berhane et al. 2020). The program has evolved through several phases, with the most recent phase (PSNP-4) including important nutrition-sensitive packages and activities to address existing maternal and child malnutrition (Berhane et al. 2020). Beginning in late 2018 and in the highland regions of Ethiopia, additional activities were introduced into the PSNP that included the provision of information on improved maternal and child nutrition practices with the aim of making the program more nutrition sensitive.
Previous evaluations of the PSNP show that it improved food security (Gilligan et al. 2009;Berhane at al. 2014). Whether these impacts realized during "normal times" extend to extraordinary times such as the current COVID-19 crisis remains an important empirical question. On the one hand, the overwhelming economic disruptions associated with the pandemic and the covariate nature of the pandemic may limit the effectiveness of relatively "small" safety net transfers. On the other hand, given that the PSNP targets very poor households, who heavily rely on these transfers, the protective role may be even higher during the pandemic when other sources of income are limited.
Safety nets can protect households against COVID-19-type aggregate shocks through several mechanisms. First, immediate and recent transfers can help cushion consumption shortfalls caused by COVID-19-induced disruptions in economic activities and drops in incomes. This is especially the case with inkind PSNP transfers, which may support households even under COVID-19-induced supply chain disruptions. Second, continuous transfers may help households build resilience against aggregate shocks either by encouraging livelihood investments or by facilitating savings that can help households smooth consumption during income fluctuations. Third, continuous PSNP transfers may shape households' coping strategies against aggregate shocks that can affect longer-term livelihood outcomes.
The PSNP operations remained largely the same during the pandemic, with little change to payment modalities, payment levels, or inclusion in the program. The two exceptions were that (1) participants were exempted from public works to avoid physical contact and gatherings during public works and (2) participants were guaranteed to receive their entitlements that covered the public work months. A one-time lump-sum payment of transfers for subsequent months was also planned to minimize physical contact due to repeated interactions. Rapid assessments made in June 2020 indicated that these changes were not implemented in many places, and in fact, payment delays were observed in some locations (as has been the case in prepandemic periods). Overall, mobility restrictions, limits on large gatherings, and slowdowns in local markets had some impacts in the PSNP areas. Nonetheless, Ethiopia did not implement a full lockdown measure, and most restrictions were eventually abandoned to avoid further disruptions.

B. Survey Method and Implementation
In March and August 2019, we conducted face-to-face surveys in 88 woredas where the nutrition-sensitive PSNP was supposed to operate. Three rural kebeles (subdistricts) were randomly selected from each woreda, and within these, one enumeration area was randomly chosen. 2 For each enumeration area, a list of households was constructed. Inclusion in this list was based on the following criteria that were implemented sequentially. First, a household was eligible if it had a child younger than 24 months of age. Conditional on having a child in this age range, households were eligible for survey inclusion if (a) they had been included in the PSNP or (b) they were not included in the PSNP but were considered poor. 3 From this list, we randomly selected five PSNP and five 2 Kebele is the smallest administrative unit in Ethiopia. 3 Poverty status was determined using a subjective poverty measure in which households were asked to rank themselves on a seven-rung poverty ladder. The first rung represented the very poorest households in the village, and the seventh rung represented the very richest households in the village.
Our previous work in Ethiopia shows that this poverty ladder is well correlated with other (more non-PSNP households. Our March 2019 sample included 2,626 households with a young child aged 0-24 months, of which 2,551 were successfully reinterviewed in August 2019 (Berhane et al. 2020). 4 In June 2020 we conducted a phone survey of these households. As part of the 2019 surveys, we collected information on households' ownership of mobile phones, their phone numbers, and permission to contact them by phone if needed. About 54% (1,387/2,551) of the August 2019 sample had access to a phone. In addition, members of our 2019 survey teams were able to help us locate additional households who had obtained mobile phones after August 2019. Using this information, we interviewed 1,497 of the 2,551 households who took part in the August 2019 survey, which is 59% of our original sample, an attrition rate that is comparable to several phone surveys in Africa. 5 As was the case in the 2019 surveys, the primary respondent was the mother of the preschool child. Several methods were deployed by the survey team to minimize nonresponse, including using built-in out-of-network reminders, allowing extended appointments, and making several call attempts at different hours of the day. On average, each interview required about two call attempts to succeed, about 53% were reached on the first attempt, and about 3% of the interviews needed a minimum of seven (and a maximum of 17) call attempts to succeed. 6 An obvious concern with this approach is that the phone sample will differ in systematic ways from the original 2019 sample, not least because ownership of mobile phones is correlated with wealth in this population (see table 1). We account for this in the following way. Our 2019 survey contained a rich set of observable characteristics that we can use to predict the probability of response to our phone survey (see table 1) using a logit model (see table A1; tables A1-A8 are available online). We then construct sampling weights as the inverse of the predicted probability of response in the phone survey. These weights were applied in all of our analyses and estimations. As table 1 shows, use of these objective) welfare measures: durable asset levels, livestock holdings, and self-reported food security. Non-PSNP households were chosen from the bottom four rungs of the ranking. See Berhane et al. (2020) for detailed discussion about this selection process. 4 This is an attrition rate of 2.9% largely related to two reasons: households being absent in the rainy season and some areas being inaccessible because of ongoing civil unrest. The March 2019 sample considered children 0-24 months old. In June 2020 our the sample included children 13 months older. In addition, a new child had been born to 15% of the mothers in our sample. Given the interest in younger children, whenever possible, we have replaced the old index children with the young ones. Thus, our current child sample is composed of children 0-36 months old. 5 The attrition rates reported by recent World Bank (Living Standards Measurement Study-Integrated Surveys on Agriculture) phone surveys are comparable to this. See https://www.worldbank.org/en /programs/lsms/brief/lsms-launches-high-frequency-phone-surveys-on-covid-19. 6 The median interview time was 32 minutes. weights markedly reduces the unweighted differences in the observable characteristics in the full sample and phone sample. Furthermore, the weighted distribution for some observables in our phone survey is almost indistinguishable from those of the full sample (see fig. 1). In our analysis, we deploy these sampling weights to recover appropriate and representative statistics under the assumption that this list of observable factors can account for systematic nonresponses in the phone survey (Korinek, Mistiaen, and Ravallion 2007;Wooldridge 2007).

III. Descriptive Results and Study Context
Here we provide some descriptive information of our sample, our outcome variables, the spread of COVID-19 in Ethiopia, and our measures of PSNP participation.

A. Descriptive Results
About 90% of our sample households are headed by males who are, on average, 39 years old and have 3 years of schooling (see table 1). Households have an average family size of six, operate just under a hectare of land, and own livestock equivalent to three tropical livestock units (TLUs). 7 More than half of mothers had never been to school, and those who had completed, on average, only two grades. In evaluations of the PSNP, household food security is measured using a selfreported indicator called the food gap, the number of months the household was not able to satisfy its food needs (Berhane et al. 2014(Berhane et al. , 2020. In both the August 2019 and the June 2020 surveys, households were asked to report their food gap over the previous 6 months. This outcome ranges from 0 (no food gap) to 6 months (acute food gap). We use both this count outcome as well as a binary food insecurity indicator constructed from these responses. This binary food insecurity indicator assumes a value of 1 for households who experienced problems in satisfying their food needs, those reporting a food gap above 0, and a value of 0 for those reporting no difficulty satisfying their food needs (those households reporting no food gap). Note, however, that the difference in the timing of the two survey rounds mechanically contributes to difference in food gap. The 6 months before the August baseline cover February to July, and the 6 months before the June follow-up cover December to May. The two survey rounds have a 4-month overlap (February to May), but June and July are generally leaner months than December and January. As a result, reported food insecurity (food gap) is likely to be higher for the August 2019 survey compared with the June 2020 follow-up. This suggests that our estimated changes in food insecurity are likely underestimates.
More than 51% of households reported to have experienced food insecurity in the prepandemic period, with the figure higher for PSNP beneficiaries ( fig. 2). This incidence of food insecurity increased to 59% in the aftermath of the onset of the pandemic, largely driven by the sharp increase in food insecurity among non-PSNP households. The average household reported a food gap of 1.3 months in the August 2019 survey, but this grew to 1.6 months in June 2020. Non-PSNP households reported a 0.5-month increase in the food gap, while the corresponding increase for PSNP households was 0.1 month. Compared with August 2019, the food gap in June 2020 increased in three of the four surveyed regions. The largest increases were in Southern Nations, Nationalities and People (SNNP; from 1.8 to 2.6 months) and Tigray (from 0.5 to 0.9 months). While the mean food gap was highest in Oromia, it changed little between the two rounds. PSNP beneficiaries report smaller increases in the mean food gap (0.2 vs. 0.5 months).
We asked respondents about their perception of their food security status in the past 3 months preceding the June 2020 survey compared with similar months the previous year. The food security question is framed in terms of households' ability to satisfy their food needs. As shown in figure 3, almost half of the respondents (49%) reported that their ability to satisfy their food needs worsened, and the rest reported that it remained about the same (48%) or improved (3%). The worsening of the problem of satisfying food needs was greater in areas that are more severely affected by the pandemic. In addition to the household survey, we collected zone-level (subregional) data on spread of COVID-19 until June 2020. The first case of COVID-19 in Ethiopia was recorded on March 13, and the total number of cases reached 5,846 by June 30 (EPHI 2020). We construct an indicator variable of zone-level status of the spread of COVID-19 using the number of cases reported as of the end of June 2020. The indicator variable takes a value of 1 for zones in the top tercile of COVID-19 cases and a value of 0 for zones in the bottom two terciles. Households reporting that their food security situation worsened (34% vs. 15%) or remained about the same (35% vs. 13%) are concentrated in zones with a high number of COVID-19 cases. We obtained comparable data on maternal and child diets in both the 2019 face-to-face surveys and the 2020 phone surveys. Mothers were asked about their and the index child's food consumption in the 24 hours before the interview using a listing of food items. Mothers' food items are grouped into 10 food categories: all starchy staple foods; beans and peas; nuts and seeds; dairy; flesh foods; eggs; vitamin A-rich, dark green, leafy vegetables; other vitamin A-rich vegetables and fruits; other vegetables, and other fruits (FAO and FHI 360 2016). Thus, mothers' dietary diversity score ranges from 0 to 10 food items. Children's food items are grouped into seven categories: grains, roots, and tubers; legumes and nuts; dairy products; flesh foods; eggs; vitamin A-rich fruits and vegetables; and other fruits and vegetables. Children's dietary diversity score ranges from 0 to 7.
Reductions in incomes and restrictions in access to food markets may lead to a decline in the diversity of diet as well as shifts away from high-quality but more expensive items to cheaper substitutes. Figure 4 shows the percentage of mothers reporting that they or their children have consumed a specific food category in the past 24 hours. Starting with mothers, their diets are dominated by starchy staples, vegetables, and beans and peas, with almost all (97%) consuming starchy staples ( fig. 4a). There was a small increase in the number of mothers reporting consumption of these items in June 2020 (from 93% to 97% for starchy staples and from 61% to 63% for other vegetables) but no change in the consumption of beans and peas (60%). We observe a larger increase in consumption of vitamin A-rich, dark green, leafy vegetables (from 20% to 32%). There was a decline in the consumption of eggs (from 5% to 2%), dairy products (from 20% to 13%), and flesh foods (from 7% to 2%).
Changes in the food children consumed follow a similar pattern, except for eggs, though the size of the change is much larger for children ( fig. 4b). There is a sharp increase in the consumption of grains, roots, and tubers (from 84% to 96%); vitamin A-rich fruits and vegetables (from 8% to 31%); other fruits and vegetables (from 39% to 62%); and legumes and nuts (from 35% to 51%). The consumption of dairy products decreased from 16% to 11%, and consumption of poultry, fish, and meat decreased from 3% to just 1%, whereas consumption of eggs increased from 6% to 10%. Overall, both mothers and their children increased consumption of vitamin A-rich fruits and vegetables, while consumption of animal source foods declined.

B. Context around COVID-19
The first COVID-19 case in Ethiopia was confirmed on March 13, 2020, in Addis Ababa. It has since spread to all 11 regions of the country. As of the end of 2020, the number of confirmed cases exceeded 120,000, with more than 1,900 confirmed deaths. About 56% of the confirmed cases were in Addis, with Oromia (17%), Tigray (5%), Amhara (5%), and SNNP (4%) being the other most affected regions (EPHI 2020). To slow the spread of the pandemic, the Ethiopian federal government put in place several measures, including restrictions on movement of people, school closures, and shelter in place orders. These culminated in the declaration of a state of emergency on April 8, 2020. Before that, a range of restrictions had been put in place, including bans on overcrowded public transport and large gatherings, closure of primary and secondary schools, and closure of universities and colleges on March 16, 2020. Domestic travel restrictions were put in place on March 21, 2020, and a halt on the movement of people along all borders was imposed on March 22, 2020. A mandatory quarantine for international travelers was imposed on March 23, 2020, followed by stay at home/shelter in place orders on March 23, 2020 (EPHI 2020).
As of June 2020, all respondents (99.8%) had heard of COVID-19. The initial source of information was primarily mass media. Just more than half of respondents indicated that they first learned about COVID-19 either on radio (51%) or on television (8%). Neighbors (18%) and family members (4%) were also noteworthy sources of initial information. On average, respondents could describe 1.8 symptoms of COVID-19, and 93% could identify at least one symptom. Slightly more than one-third (36%) reported that they did not leave their homesteads during the previous week. On average, respondents reported taking 3.2 actions to reduce the likelihood that they or someone in their household would contract COVID-19, and only 5% reported taking no actions at all. The most common actions were washing hands for 20 seconds or more (82%) and, conditional on going out, avoiding shaking hands or kissing when greeting others (77%) or avoiding large gatherings or queues (66%).
We asked respondents how worried or stressed they were because of the coronavirus pandemic, using a scale from 1 (not worried/stressed at all) to 10 (extremely worried/stressed). Two-thirds of respondents (68%) reported that they were extremely worried, 10% said that they were not worried at all, and the rest reported stress levels scattered between these. We then asked which aspect of the coronavirus crisis had had the greatest impact on the respondent and the household. (Enumerators read these out and asked the respondent to indicate which item had the greatest effect.) Aggregate responses are shown in figure 5. Market closure, fear of getting infected by the virus, high food prices, and loss of income were the most important effects of the pandemic on livelihoods. More than 24% of respondents reported market closure as the greatest impact of the pandemic. Responses disaggregated by self-reported stress levels (not stressed, responding from 1 to 3 on the 10-point scale; moderately stressed, responding from 4 to 8; extremely stressed, responding either 9 or 10) are reported in table A2.
Not surprisingly, about 30% of respondents who are found in the notstressed category indicated that they had not been affected in any way. Extreme self-reported stress is strongly associated with fear of illness or death. A loss of employment or income is cited as the most important impact by more than 8% of all respondents, just under a third (30%) report that disruptions in food access were the most important impact, and another 9% stated that higher food prices were the most important impact. Figure 5 lists the single most important impact, not the only impact. To further assess the consequences of the pandemic, we asked respondents to provide a qualitative assessment of changes in household income compared with incomes usually received at this time of the year. Across all respondents, two-thirds stated that incomes were much less (26%) or somewhat less (41%; table A3). Only 27% reported that incomes were unchanged, and few, just 6%, stated that incomes had increased. A potentially confounding factor, however, is that parts of southern and eastern Ethiopia were affected by locust swarms from February 2020 onward. We asked respondents whether in the past month their crops or livestock had been adversely affected by locusts on a scale from 1 (not at all) to 5 (totally lost). Most respondents (84%) reported that they were not affected, 8% said that they had been affected "a little bit," and the remaining 8% said they that were more severely affected. In table A3, we disaggregate income losses by households affected by the locust invasion and those that were not. Households affected by the COVID-19 pandemic and the locust swarm reported higher rate of income losses. However, even excluding those households that were affected by desert locusts, we still see a large fraction (64.5%) of households reporting that their incomes were much less or less than usual.
We next explore how households cope with potential impacts of the pandemic and associated income losses. Specifically, we asked whether in the previous 30 days anyone in the household had undertaken certain actions because of a lack of food or a lack of money to buy food or meet other basic needs. These coping strategies are reported in figure 6. These include reductions in food consumption, expenditures on nonfood, expenses on agricultural inputs, and other approaches to smooth consumption. Several results emerge from figure 6. First, households reporting income losses were most likely to report the use of all of these coping strategies. Second, households were more likely to report undertaking reductions in food consumption or expenditures on nonfood items, such as health, education, and clothing, or to reduce purchases of agricultural or livestock inputs than actions that reduced asset holdings or increased indebtedness. Third, borrowing money to buy food was used by more than 44% of households reporting that income was less than usual. Fourth, few households reported selling consumer durables, possibly because these are relatively illiquid. A larger proportion (20%) reported selling productive assets, but note that this category includes livestock, which also serves as a store of value.

C. Alternative Definitions of PSNP Participation
We define PSNP participation in three different ways. First, we define access to the PSNP according to self-reported receipt of any payments in the August 2019 survey. According to this definition, about 40% of the households in our baseline sample are PSNP beneficiaries (see table 1). PSNP participation could be endogenous to the pandemic if, for example, the government expanded the program in response to the spread of the coronavirus; using pre-COVID-19 access to PSNP transfers to define participation in the program addresses this concern. This approach assumes that program participation remains stable during our study period. This assumption appears to be correct; of those households receiving PSNP transfers in June 2020, 94% were also PSNP beneficiaries in August 2019.
Second, we use information on actual PSNP transfers made in the 6 months before the August 2019 survey and generate an indicator variable assuming a value of 1 for households who received transfers exceeding ETB 100 (about US$3.5) and a value of 0 for households receiving less than ETB 100 or none at all. This definition reduces potential measurement errors due to misreporting of program participation. Similarly, very small transfers may not have meaningful impacts, and distinguishing these from relatively larger transfers is important. Among those households who reported to be PSNP beneficiaries, 1% had not received income greater than ETB 100, and 6% of them received transfers less than ETB 1,000 in the 6 months before the 2019 survey. 8 Third, we generate aggregate district-level access to PSNP transfers. Among the 88 woredas that were supposed to be included in the PSNP, there were six woredas that did not make any payments because of delays in implementation and related logistical problems. That is, among the 88 woredas that were meant to be PSNP beneficiaries, 9% of these (translating to about 9% of the households in our baseline sample) were not actually providing PSNP transfers. This provides a relatively more exogenous variation in households' access to PSNP transfers.
We note that there exists significant discordance between these alternative definitions and measures of PSNP participation (see table A4).

IV. Estimation Strategy
To assess the impact of PSNP participation on household food security as well as maternal and child diets under COVID-19, we compare the temporal evolutions of food security and nutrition outcomes between PSNP beneficiary and nonbeneficiary households. We employ the following difference-in-differences specification: where Y ht stands for food security and related maternal and child diet measures for household h and round t; a h stands for household fixed effects, which capture all time-invariant differences between PSNP beneficiaries and nonbeneficiaries; PSNP h represents households' access to PSNP transfers; and Post t is a dummy variable, assuming a value of 1 for the phone survey conducted during the COVID-19 pandemic (henceforth, "COVID-19 round") and a value of 0 for the pre-COVID-19 round. The parameter associated with this time dummy captures aggregate trends in food security across rounds, including those driven by seasonality in our sample, changes in coping strategies, or other factors unrelated to the COVID-19 pandemic. The time trend may also capture potential differences in our outcomes of interest driven by differences in survey methods (face-to-face or phone survey). Given that we have employed face-to-face surveys in the pre-COVID-19 round and phone surveys in the COVID-19 round, such differences can be substantial. This time trend can effectively capture these differences, especially if they remain similar across PSNP beneficiaries and nonbeneficiaries.
The key parameter of interest in equation (1) is b 1 . This parameter identifies potential differences in the temporal evolution of food security outcomes of PSNP beneficiaries and nonbeneficiaries. In the absence of unobservable time-varying variables with differential impact on PSNP beneficiaries, this parameter can be interpreted as the protective impact of the PSNP against food insecurity caused by the COVID-19 pandemic. This estimation strategy requires that in the absence of the pandemic, the food security status of PSNP beneficiaries and nonbeneficiaries would follow similar trends. We assess this assumption by running the same specification using our two pre-COVID-19 surveys (March 2019 and August 2019). As shown in table A5, the parallel trend assumption holds in most of our empirical specifications and for most food security indicators. Before the COVID-19 pandemic, PSNP beneficiaries and nonbeneficiaries followed similar trends in the food gap and the dietary diversity of mothers and children but not in the children's minimum dietary diversity score. 9 In all of our specifications, we control for observable potential time-varying confounders, such as the locust invasion that hit parts of the country. We include a categorical variable that measures whether households were affected by locust swarms. The use of alternative definitions of access to the PSNP helps us to probe the robustness of our findings to potential time-varying unobserved factors. Our outcome variables are measured as binary and count outcomes. Hence, we estimate both linear and nonlinear specifications. More specifically, we estimate a standard linear fixed effects model as well as fixed effects logit and fixed effects Poisson regressions. While the linear fixed effects model extracts time-invariant differences without functional form assumptions, the nonlinear approaches employ a fully parametric approach to control for time-invariant differences across households. Moreover, the nonlinear models discard observations with time-invariant dependent variables, leading to significant differences in the sample sizes reported in our results. These differences reflect the robustness of our results to some parametric assumptions and model specifications and choices.
To account for potential systematic nonresponse in the phone survey, we use the weights described in section II. By using these sampling weights, we can generate representative statistics under the assumption that these data are "missing at random," conditional on the observables used in the construction of the weights (Korinek, Mistiaen, and Ravallion 2007;Wooldridge 2007). Access to the PSNP and related unobserved factors are potentially correlated among households living in the same sampling unit, the kebele. To account for this, we cluster standard errors at the kebele level.
The impacts of the pandemic and the role of the PSNP in mitigating them are likely to vary across households with different socioeconomic characteristics. To uncover such potential differential impact of the PSNP across various groups of households, we run our preferred specification in equation (1) for several dimensions of heterogeneity, including wealth quintiles and remoteness of location of residence.

V. Results
We first discuss our results on the impact of the PSNP in protecting household food security before considering impacts on maternal and child diets. Table 2 presents difference-in-differences estimates comparing the temporal evolution of food security of PSNP beneficiaries and nonbeneficiaries under alternative definitions of PSNP participation and controlling for time-invariant and observable time-varying household characteristics. We report results from linear fixed effects regressions in the text and present results from nonlinear models (fixed effects logit and fixed effects Poisson) in table A6. We present results with and without sampling weights, although weighted regressions are our preferred specifications, to probe the robustness of our results. Columns 1-3 of table 2 report results for the food insecurity dummy variable, and columns 4-6 show results for the continuous food gap measure. Columns 1, 3, 4, and 6 are based on weighted fixed effects specifications, while columns 2 and 5 are based on unweighted fixed effects models. Panel A of table 2 provides results using self-reported access to PSNP benefits, panel B reports results using information on the amount of transfers, and panel C provides estimates based on an aggregate indicator of access to the PSNP at the woreda level. Three important findings emerge from the results in table 2. First, the share of food insecure households increased by 11.7 percentage points in the 6 months before the June 2020 round compared with a similar period in 2019 (panel A, col. 1), and the size of the food gap increased by 0.47 months (col. 4). Second, inclusion in the PSNP offsets virtually all of this adverse change. The magnitude of the coefficient b 1 is 20.093 in panel A, column 1, indicating that PSNP participation reduced the likelihood that the household was food insecure by 9.3 percentage points. Adding coefficients b 0 and b 1 together shows that the likelihood of becoming food insecure increased by 11.7 percentage points for non-PSNP households and by 2.4 (11:7 2 9:3) percentage points for PSNP households. Looking at the results found in panel A, column 4, adding coefficients b 0 and b 1 together shows that the duration of the food gap increased by 0.47 months for non-PSNP households and by 0.13 (0:474 2 0:341) months for PSNP households. Third, our findings that the PSNP offset the impact of the pandemic are robust to the definition of access to the PSNP (see panels B, C) and whether we weight or do not weight our data (see cols. 2, 5). While the interaction terms in panels A and B capture actual protective impact (average treatment on the treated) of the PSNP, the corresponding parameter in panel C represents the role of living in a PSNP-covered district (intention to treat). 10 Thus, both impacts are not directly comparable.

A. The Role of the PSNP in Protecting Household Food Security
We note four additional features of our results. First, our identification hinges on the assumption that in the absence of the pandemic, food security outcomes for PSNP beneficiaries and nonbeneficiaries follow parallel trends. We probe this using pre-COVID-19 rounds and trends of food security. As discussed above, the food security trend of PSNP beneficiaries and nonbeneficiaries for the March 2019 and August 2019 surveys were similar for most food security indicators. Second, alternative estimators, fixed effects logit and Poisson regressions, give comparable results (see table A6). We note that these nonlinear estimation approaches discard observations with constant dependent variable values across rounds, and hence the number of observations reported in table A6 is much smaller than that reported in table 2. Despite these significant differences in sample size, estimation methods, and associated parametric assumptions, our main results remain robust.
Third, we control for a second shock that occurred at this time-the locust invasion. Dropping this variable does not affect our findings (see table A7). Fourth, to probe the role of other shocks, including conflict and civil unrest, we reestimate our main specification by dropping regions that have been affected by recent conflicts and civil unrest. The Oromia region experienced conflicts and demonstrations the most in the months before our June 2020 survey. Excluding Oromia from our sample does not affect our results (see table A8).
The role of the PSNP in protecting food security and smoothing consumption is expected to be higher for poorer households, for whom the share of PSNP transfers in total consumption expenditure is higher. The impact of the pandemic is also likely to be higher for poorer households who face greater adverse exposure, potentially leading to further deterioration in food security due to COVID-19-related income losses. Similarly, PSNP transfers, particularly in-kind transfers, may be more impactful for households with limited access to markets. PSNP cash transfers may also prove useful given that rural transactions take different forms, including reciprocities. Conversely, COVID-19-related disruptions in access to food may have limited impact on households who rely less on markets-often the poor and those residing in remote locations. The PSNP would bridge the gap in own production that otherwise would need to be filled through purchases from food markets, an option unavailable for nonbeneficiaries.
We test these hypotheses by disaggregating the data along several dimensions. In table 3 we split the sample by wealth quintiles: combining the bottom three quintiles into one group (poorest households) and the top two quintiles into another group (less poor households). The results in table 3 suggest that PSNP transfers are more protective for poorer households' food security. Poorer PSNP nonbeneficiaries report higher increase in food insecurity experience. In table 4 we split the sample into remotely located and accessible households, using distance to urban centers. Using median distance to urban centers with population of 20,000 or higher, households located in areas with distances above the median are classified as remote. The results in table 4 show that households in remote areas and not receiving any PSNP transfers are more likely to experience a significant deterioration in food security.
Finally, the protective role of the PSNP can be higher for those households experiencing income or job losses because of the pandemic. In table 5 we test for differential impacts by self-reported experience of income loss. We find that the protective role of the PSNP is significantly higher for those households reporting loss of income. Column 4 of table 5 shows that the PSNP has marginal but statistically insignificant effect for those households who reported that their income has remained the same or increased in the latest round.
Social protection and safety nets can also help households adopt effective coping strategies when they face income losses and related shocks. In our phone survey we elicited households' coping strategies in the past 1 month. Households were read a long list of non-mutually-exclusive coping strategies and were asked to pick ones that applied to them. The most dominant strategies include (i) spent savings, (ii) reduced food consumption, (iii) borrowed money from others, (iv) reduced health and education expenditures, (v) reduced expenditures on agricultural inputs, and (vi) sold assets. Some of these  Note. Estimates are from linear regressions controlling for household fixed effects. The dependent variable in cols. 1 and 2 is an indicator variable equaling 1 for households unable to satisfy their food needs. The dependent variable in cols. 3 and 4 is the food gap measured in months. Odd columns provide results for households above the median distance to urban centers of 20,000 or more people, and even columns report results for households below the median distance to urban centers. PSNP participation is defined as reported receiving PSNP payments in 2019. Standard errors, clustered at kebele level, are in parentheses. ** p < .05. *** p < .01. Note. Estimates are from linear regressions controlling for household fixed effects. The dependent variable in cols. 1 and 2 is an indicator variable equaling 1 for households unable to satisfy their food needs. The dependent variable in cols. 3 and 4 is the food gap measured in months. Odd columns provide results for households who reported loss in income because of the pandemic, and even columns report results for households who reported that their incomes remained the same or increased in the latest round. PSNP participation is defined as reported receiving PSNP payments in 2019. Standard errors, clustered at kebele level, are in parentheses. ** p < .05. *** p < .01.
coping strategies are likely to have lasting adverse impact on the livelihoods of rural households. For instance, decrease in education-and health-related investments will adversely affect human capital accumulation and reduce future earnings. Similarly, reduction in agricultural investments may limit households' production potential. Table 6 presents single-difference estimates, as data on coping strategies were collected only in June 2020 and thus should be interpreted cautiously. It shows that households who received PSNP transfers are less likely to reduce expenditures on health and education by 7.7 percentage points and were less likely to reduce expenditures on agricultural inputs (fertilizer, seeds, and livestock) by 13 percentage points. These associations are intuitive and consistent with our fixed effects results and provide suggestive evidence that PSNP transfers can help households avoid detrimental coping strategies to deal with income losses associated with COVID-19.
B. The Role of the PSNP in Protecting Mothers' and Children's Diets COVID-19 could affect maternal and child diets through reductions in income or reduced market access. Households might alter their consumption as incomes drop because of limited on-and off-farm income-generating opportunities and a fall in remittances and other private transfers. The effect on diet could manifest both in the number of items households consume and in the quality of diets. To measure the effect of COVID-19 on these dimensions of diet, we use diet diversity indexes for mothers and children as well as changes in consumption of animal source foods and vitamin A-rich fruits and vegetables. Table 7 shows how diet diversity has changed in response to the COVID-19 crisis and the implication of the PSNP in protecting mothers' and children's diets. We present two sets of results using a continuous diet diversity index for mothers and children as well as a dummy variable that takes a value of 1 if minimum diet diversity is met and a value of 0 otherwise. The minimum diet diversity is defined at five food categories for mothers and four food categories for children (WHO 2010). There has been some increase in diet diversity in the COVID-19 period, with the effect particularly large and statistically significant for children. Similar patterns are observed for the likelihood of meeting the minimum diet diversity requirement. Children were about 16 percentage points more likely to meet the minimum diet diversity compared with the pre-COVID-19 period. Disruptions in supply value chains, particularly those of perishable foods that rural households may supply (e.g., fruits, vegetables, eggs), may have resulted in these foods being consumed at home-indirect evidence of which Hirvonen, de Brauw, and Abate (2020) report for urban food markets in Ethiopia. This limited role that the PSNP plays in protecting mothers' and children's diets could be explained by several factors. First, as part of the transfers is given in cash, with limited market access, households cannot use cash transfers to buy food. Second, in-kind transfers are likely given in consumption items that households already consume (e.g., wheat, cooking oil) and contribute mainly to the intensive margin of how much of the food items they consume but not much to the extensive margin of whether they consume foods from specific food categories. Third, dietary diversity of mothers and children was already at a very low level and thus unlikely to decline further despite the COVID-19 pandemic. Fourth, the PSNP transfers may not be sufficiently nutrition sensitive, pointing to the need for further refinements to the program (Berhane et al. 2020). Although the PSNP had little differential impact on the diet diversity of mothers and their children, it could conceivably improve the quality of their Note. The outcome variables in cols. 1 and 3 are the number of food categories consumed in the past 24 hours, taking values of 0-10 for mothers and 0-7 for children and estimated using ordinary least squares. The outcome variables in cols. 2 and 4 are dummy variables taking a value of 1 if mothers (children) meet minimum diet diversity defined at five categories or more (four categories or more), estimated as a linear probability model. PSNP participation is defined as reported receiving PSNP payments in 2019. Child estimates are based on children between 6 and 24 months old. Standard errors, clustered at the kebele level, are in parentheses. *** p < .01.
diet. The cash and in-kind transfers from the PSNP may allow participants to substitute lower-quality foods (e.g., cereals and beans) with higher-quality foods, such as meat, fish, dairy, and eggs. To gauge whether this was indeed the case, we group dairy, flesh foods, and eggs together into "animal source food" and group vitamin A-rich fruits and vegetables into another group. Table 8 presents regression estimates for mothers and children separately. We find that there has been a sharp decline in animal source food consumed by mothers, whereas children's intake of animal products has changed little in the COVID-19 period. On the contrary, consumption of vegetables and fruits increased for both mothers and children, though it is statistically insignificant for mothers. The increase in consumption of vegetables and fruits was particularly large for children. We do not find any evidence of a protective role of the PSNP on the diets of mothers and children. This is not surprising given that the composition of the in-kind transfers is unlikely to have changed in the COVID-19 period, and restrictions in access to markets may mean cash transfers are unlikely to make meaningful difference in the range of foods available to households. Note. The outcome variables in cols. 1 and 3 are dummy variables that a take value of 1 if the mother (child) consumed dairy products, flesh foods, or eggs in the past 24 hours. The outcome variables in cols. 2 and 4 are dummy variables taking a value of 1 if the mother (child) consumed vitamin A-rich vegetables and fruits in the past 24 hours. PSNP participation is defined as reported receiving PSNP payments in 2019. Child estimates are based on children between 6 and 24 months old. Standard errors, clustered at the kebele level, are in parentheses. *** p < .01.

VI. Summary and Concluding Remarks
We combine data from a prepandemic face-to-face survey with a phone survey conducted in the aftermath of the onset of the COVID-19 pandemic in Ethiopia, finding that household food security deteriorated significantly. About half of the households surveyed reported that food security had worsened compared with the same period before the pandemic. Market closures, food price increases, and loss of income appear to be the most important aspects in which the pandemic impacted livelihoods. We do not find significant changes in mothers' and children's diets despite some changes in the composition of diets. Consumption of animal source foods declined significantly, perhaps due to the closure of markets associated with the pandemic. Social protection, specifically Ethiopia's PSNP, mitigated the impacts of the pandemic on food security. We find that PSNP beneficiaries report relatively less deterioration in food security compared with nonbeneficiaries. The protective role of the PSNP was slightly higher for poorer households and those living in remote areas. The PSNP also reduced the likelihood of households adopting detrimental coping strategies, such as reducing expenditures on education, health, and agricultural inputs. However, we do not find evidence that the PSNP protects mothers' and children's diets.
Our findings highlight the value of having a well-functioning social protection program in place before the pandemic to protect the food security of poor households. This lends empirical support to the argument for expanding social safety nets (Devereux, Béné, and Hoddinott 2020;Gentilini et al. 2020;Gilligan 2020). This is a particularly important finding in the context of Ethiopia, where the pandemic is still unfolding and the government is weighing alternative measures to support poor and vulnerable households.