Access to health services, food, and water during an active conflict: Evidence from Ethiopia

Civil conflict began in Ethiopia in November 2020 and has reportedly caused major disruptions in access to health services, food, and related critical services, in addition to the direct impacts of the conflict on health and well-being. However, the population-level impacts of the conflict have not yet been systematically quantified. We analyze high frequency phone surveys conducted by the World Bank, which included measures of access to basic services, to estimate the impact of the first phase of the war (November 2020 to May 2021) on households in Tigray. After controlling for sample selection, a difference-in-differences approach is used to estimate causal effects of the conflict on population access to health services, food, and water and sanitation. Inverse probability weighting is used to adjust for sample attrition. The conflict has increased the share of respondents who report that they were unable to access needed health services by 35 percentage points (95% CI: 14–55 pp) and medicine by 8 pp (95% CI:2–15 pp). It has also increased the share of households unable to purchase staple foods by 26 pp (95% CI:7–45 pp). The share of households unable to access water did not increase, although the percentage able to purchase soap declined by 17 pp (95% CI: 1–32 pp). We document significant heterogeneity across population groups, with disproportionate effects on the poor, on rural populations, on households with undernourished children, and those living in communities without health facilities. These significant disruptions in access to basic services likely underestimate the true burden of conflict in the affected population, given that the conflict has continued beyond the survey period, and that worse-affected households may have higher rates of non-response. Documented spatial and household-level heterogeneity in the impact of the conflict may help guide rapid post-conflict responses.

Introduction and background crop-cutting and harvesting activities but also impeding the functioning of local and interregional markets, especially in conflict hotspot areas. Because different staples are produced in different areas of Ethiopia, local markets may experience commodity-specific impacts of disruptions to trade. On the other hand, the war also curtailed households' livelihood activities, contributing to likely reductions in income and the ability to afford food purchases.
Households' access to health services. In the phone surveys, households were first asked if they needed any medical services (treatment or consultation) in the past 4 weeks. We define an indicator variable for health services demand which equals 1 if any member of the household needed any medical services, 0 otherwise. Households who needed health services were next asked if they were able to access the required medical services in the past 4 weeks. Therefore, conditional on the demand for health services, we define another indicator variable that takes a value of 1 if the household is unable to access the health services needed, 0 otherwise. In addition, we generate a third indicator variable that takes a value of 1 if the household is not able to buy enough medicine, 0 otherwise.
Households' access to WASH. Households' access to WASH is measured using three indicator variables, each taking value of 1 if household has sufficient access to water for drinking, sufficient access to water for washing, and sufficient access to soap for washing, 0 otherwise.
Estimation strategy. We employ two empirical strategies to identify the impact of the conflict on households' access to food, health and WASH services. First, we employ a Difference-in-Differences strategy to identify the overall impacts of the war by comparing trends in households' access to these services before and after the outbreak of the war across affected and unaffected households. Second, we use the ACLED database to construct household-level and time-varying exposure to violent conflict and estimate two-way fixed effects model. In both cases we use linear probability models (LPM); alternatively, we also present logistic regression models in the S1 Appendix. Our first approach implements the following DID specification: A hrt represents access to food, health and WASH services associated with household h living in region r observed in time t. α h captures household fixed effects. Wartime t is a binary indicator of the conflict period, taking a value of 1 for periods affected by war, and 0 otherwise. The war started on November 04, 2020, but the November 2020 round interviews were completed before that date and hence appear as a pre-war round. α 1 captures aggregate changes in households' access to food, health and WASH services including in the absence of the war. Tigray r represents an indicator variable for households living in Tigray region. Thus, in the first phase of the conflict, Tigray is the region affected by the conflict while the other regions in Ethiopia serve as control group. α 2 captures the interaction of residence within the Tigray region and the conflict period. Under the assumptions of this DID model, this interaction term identifies the impact of the war on households' access to food and health services. � hrt captures other unobserved factors that may affect households' access to these services.
The specification in Eq (1) assumes that all survey households from Tigray have been affected by the war and Tigray is considered as the conflict-exposed region and the rest of the country as a comparison group. This assumption is plausible, given the depth and breadth of the conflict (see Fig 1). However, some households in Tigray may not have been affected by the conflict, suggesting that the estimates in Eq (1) can only measure a parameter analogous to intention to treat (ITT) impacts; that is, the effect of living in a conflict-exposed region, rather than the effect of direct conflict exposure itself. If the share of households in Tigray who were unaffected by the armed conflict is significant, these estimates would be considerably smaller than the average treatment effect (ATE) of the war. To quantify estimates that are closer to the ATE of the war, we construct a more granular measure of exposure to violent conflict using ACLED's battle events recorded during the survey period (between August 2019 and May 2021). We use the cumulative number of battles that took place within 20 and 30 km radius of households. This distance-based measure of exposure to battle events reduces potential misclassification of households (into affected and unaffected), while also capturing impact of conflicts realized throughout the country, including in regions bordering Tigray. While the estimates generated using these distance-based measures are still technically ITT, they are likely to be similar to the ATE of the war. Thus, we estimate the following two-way fixed effects specifications: where Battles hrt is the cumulative number of battles experienced within 20 km or 30 km radius from households' residence. α h and α m stand for household and survey round fixed effects, respectively. φ 1 is the main coefficient of interest, and measures the impact of an additional battle event on the outcome variables. To account for potential spatial correlation in battle experiences and the outcome variables, standard errors are clustered at district (woreda) level.
To account for systematic non-responses in the phone surveys, we constructed inverse probability sampling weights (used in all analyses). The sample consists of households appearing in the pre-war and at least once in the wartime phone survey rounds, implying that the weights need to be constructed considering attrition and non-responses in both phases. Households which were not observed in both pre-war and wartime rounds were considered as non-responses in the construction of sample weights. We use household and location PLOS GLOBAL PUBLIC HEALTH characteristics from the 2019 face-to-face baseline survey to predict the joint probability of response in pre-war and wartime rounds (see Table F in S1 Appendix). We then construct sampling weights as the inverse of the predicted probability of responses in both pre-war and post-war-onset phone surveys. We note that applying the weights markedly reduces the differences between the unweighted means in the observable characteristics of baseline (in person) and phone surveys (Table G in S1 Appendix). In particular, Table G in S1 Appendix shows that applying the sampling weights renders observable characteristics of the phone survey sample largely comparable to the full sample. The statistical software used for the analysis in this paper is Stata 17.0.

Ethics
The Institutional Review Board (IRB) of the Harvard T.H. Chan School of Public Health determined that this study did not qualify as human subjects research (Protocol number IRB22-0341), as the study uses only publicly available, anonymized secondary data.  Table 1 shows pooled summary statistics of the study outcomes. 24 percent of households report being unable to purchase some types of food. When disaggregated by item, about 11 percent of households were not able to purchase teff (panel (a)), the most important staple food in Ethiopia. About 30 percent of households had some need for health services in the week prior to the survey, of which 11 and 6 percent of households were unable to access the health services and not able to purchase medicine, respectively (panel (b)). About 24 percent of households lacked sufficient drinking water, while 7 percent lacked washing water and 11 percent lacked access to soap (panel (c)).

Difference-in-differences results
Fig 2 shows trends in households' ability to purchase staple foods from the market. Panel (a) shows trends in access to food items in the four major highland regions of Ethiopia. Before the war, the share of households reporting inability to buy enough food was lowest in Tigray. This figure increased from 5 percent in October 2020 (just prior to the start of the war) to 29 percent in May 2021. The temporal trends in households' inability to buy teff, wheat and maize is shown in panels (b)-(d). In all three panels, access to food markets remained relatively stable over time for Amhara, Oromia and SNNP regions, while in Tigray, the share of households that reported not being able to purchase teff increased from 5 to 23 percent between the last pre-conflict survey round (October 2020) and the final survey round (May 2021). The corresponding value for wheat increased from 2 to 21 percent. Table 2 provides estimation results for regression models of the factors associated with food access outcomes. The interaction term between the indicator variable for Tigray and wartime indicator variable captures differential trends in households' access to food markets between Tigray and the rest of Ethiopia; this is the quantity of interest in our model. The first column provides results associated with households' ability to buy sufficient foods of any type, while the remaining columns show results for specific foods. Households reporting inability to buy enough staple foods increased by 5 percentage points in those regions outside Tigray in the wartime period but by an additional 26 percentage points in Tigray region. The disaggregated Not able to buy teff Not able to buy maize results indicate larger disruptions to households' ability to access teff, an inter-regionally traded staple, and wheat. By contrast, impact estimates are smaller for maize and statistically insignificant for oil. Fig 3 provides temporally and spatially disaggregated trends in households' access to health services. Panel (a) shows that the patterns of demand for health services in Tigray did not change appreciably after the start of the war in November 2020. However, the share of households who were unable to access health services (panels (b)) or buy medicines (panel (c)) had sharply increased.
Columns 1-3 of Table 3 provide estimates of the impacts of the conflict on households' access to health services using the DID regression framework. The first column shows the impact of the conflict on demand for health services, while the remaining two columns show the impacts on access to health services and medicine among households who reported need for health services. Among this group, the war reduced access by 35 percentage points, while households' ability to buy medicine dropped by 8 percentage points.
Columns 4-6 of Table 3 show regression estimates of impacts of the conflict on households' access to WASH services: access to enough drinking water, access to enough washing water, and access to enough washing soap. The results in columns 4-5 indicate that war caused no significant change in access to drinking or washing water. However, households' ability to access enough soap decreased by 16.6 percentage points.

Heterogeneity analyses
In Table 4, the sample is divided into rural and urban areas, poor and non-poor households and households with and without undernourished children under age 5, and the empirical specification in Eq (1) is re-estimated. Poor households and households with undernourished children are identified using the (baseline) face-to-face LSMS survey in August 2019. Households in the bottom three quintiles of the welfare distribution (consumption expenditure) are defined as poor. Households with undernourished children are those with any child under the age of 5 whose height or length and body weight qualified as wasted, stunted, or underweight (2 Households including children with pre-war nutritional deficits also suffered more because of the conflict relative to households without undernourished children. We also estimate the impact of the conflict on households' access to health services, by splitting the sample into communities with and without access to health services before the outbreak of the conflict (Table 5, Panel A). Communities without a health center experienced a 60 percentage point reduction in access to health services, almost three times the effect on those communities with health services pre-war (Table 5, Panel A).
Finally, we explore potential heterogeneities in impacts on households' access to WASH services by rural and urban areas. The results shown in Table 5 (Panel B) indicate that the impact of the war on WASH services was higher in urban areas. While we find no effect on access to drinking water, households living in urban areas experienced a 23.5 percentage points reduction in access to washing water and 13.8 percentage point reduction in access to soap.

Fixed effect results: Proximity to large scale conflict events (battles)
As a counterpart to the regional DID models, regression models of the impacts of proximity to battles show qualitatively similar results. Table B in S1 Appendix shows the estimated impacts of the number of battles within 20 and 30km on food access outcomes. For each additional battle that takes place within 20 km in the months preceding the survey, the likelihood of being unable to buy enough food increased by 0.6 percentage points. These disruptions were most pronounced for teff and wheat, with smaller impacts on maize and oil. In Table C in S1 Appendix, we present results from similar models in which the dependent variables are measures of access to health services. Each battle within 20km is associated with a 0.8 percentage point increase in the likelihood of not being able to access health services and a 0.3 percentage point increase in the likelihood of being unable to access medicine. Because our outcome variables are binary in nature and as a sensitivity analysis, we also run all our estimations using standard and fixed effects logit regressions. These results show qualitatively similar findings (Tables D and E in S1 Appendix).

Discussion
This study examines the impacts of an ongoing large-scale conflict in the Tigray region of Ethiopia on households' access to health services, food, and WASH services. Using a high frequency phone survey data conducted on a panel of 2,677 households and a difference-indifferences methodology, the central findings are that conflict exposure has large, statistically significant negative impacts on access to health services and food. These impacts were more pronounced on the poorer than non-poor, on households with undernourished children, and on rural compared to urban households. Although not unexpected, the magnitude of the disruptions we document in terms of access to food, basic health services and sanitation items are strikingly high given these were measured immediately after the first phase of the conflict, suggesting the likelihood of deeper subsequent impacts with the continued blockade of access to these services for months afterwards. These results echo previous evidence on the dramatic negative consequences of armed conflicts on public health services and public health delivery [16]. While there is a large literature documenting the multiple harmful impacts of war, the findings of this study document the magnitude of the harms ex durante (i.e. while the conflict is still ongoing), their localized nature, and the additional harms suffered by already disadvantaged groups. The large impacts recorded-including the dramatic decrease in access to health care among those in need of treatment-likely reflects both restrictions on mobility and harm to livelihoods, as well as the destruction of public health services and facilities documented recently [8,9,17,18]. Similarly, the magnitude of the disruption to food systems is striking. The ability to acquire staple foods is dramatically affected, with disruptions occurring through both price effects (likely reflecting diminished supply, due to supply chain disruptions) and reductions in household's ability to afford food (because of reductions in income). These findings are consistent with previous studies showing increased malnutrition in Tigray after the onset of the conflict [9].
The pronounced heterogeneity in impacts across geography, income, nutritional status, and baseline access to services is also consistent with previous literature. The fact that rural households have experienced larger disruptions in access to food and health services reflects the more tenuous nature of rural markets and services, relative to those in urban areas. Given the reported road blockages and disruptions in public transport, it is not surprising that rural markets and public health services are relatively more affected. In addition, because rural households in Ethiopia are poorer on average than urban households, food price increases are likely more problematic in these areas. These varying impacts of the conflict across different value chains is important and consistent with other recent studies [19]. In contrast, however, our finding that urban households have suffered greater disruptions to sanitation services accords with their greater reliance on piped water infrastructure, which has been extensively damaged in the conflict. Recent reviews have highlighted major data gaps with respect to health service availability, utilization, and health outcomes in conflict settings [20,21]. These reviews have called for increased use of various data collection strategies to fill these gaps, including the Health Resources Availability Mapping System (HeRAMS) [22], surveys in camps for displaced persons, improved administrative data, and improved coordination of existing and new data tools [23], to quantify gaps in access to health services. This paper demonstrates the potential for mobile phone surveys to also play an important role in filling gaps, especially when a pre-conflict baseline survey has been conducted.
As with many studies relying on phone surveys, our analysis is constrained by limitations associated with non-random selection and non-response. While the pre-conflict LSMS-ISA sample in Ethiopia is nationally representative, the follow-up phone surveys faced attrition, in part due to the war's effects. While we employ the inverse probability sampling weights, our weighting may not fully capture systematic differences between responding and non-responding households. Under the seemingly plausible assumption that more conflict-affected households are less likely to be able to respond to phone surveys, these results represent a lower bound of the actual impacts of the conflict on the outcomes that we study. Furthermore, although the difference-in-difference methodology adjusts for time-invariant and pre-exposure differences in the outcome of interest between exposed and non-exposed communities, residual time-varying confounding factors, such as variables correlated with conflict exposure and residence in Tigray region, remain possible threats to a causal interpretation of these results. Another limitation is that household exposure to conflict could be misclassified due to population movements. However, this misclassification of exposure would typically result in attenuation bias, suggesting that the reported findings may underestimate the harms of the ongoing conflict. In addition, while our analysis captures service disruptions from the perspective of system users; future analysis could incorporate information on disruptions to health facility functioning. A further limitation is that the LSMS-ISA HFPS only continued until May 2021. As a result, we can only examine the impact of the first phase of the conflict. Further research to understand the ongoing dynamics of the conflict after this initial phase are warranted. The results emphasize that policy actions to mitigate the harmful effects on populations are urgently needed.
Supporting information S1 Appendix. Tables A-G.  Table B: Local conflict events and access to food markets. Table C: Local conflict events and access to health services. Table D: The impact of violent conflict on households' access to food markets; logit and fixed effects logit models. Table E: The impact of violent conflict on access to health and WASH services: logit models. Table F: Modeling the probability of response in both pre-and post-war-onset phone surveys.