Policy Research Working Paper 11167 Unpacking the Disaster-FCV Nexus Household Economic Impacts of Conflict and Floods in Nigeria Karima Ben Bih Bramka Jafino Chloe Desjonqueres Solene Masson Urban, Disaster Risk Management, Resilience and Land Global Department July 2025 Policy Research Working Paper 11167 Abstract This paper investigates the impact of conflicts on flood-af- down to flood-affected households. The results indicate fected households in Nigeria, utilizing a balanced panel that conflict-affected households experience lower con- dataset derived from the Living Standards Measure- sumption expenditure compared to non-conflict-affected ment Survey data collected in 2012, 2015, and 2018, households, with the adverse effects being significantly and geo-spatial conflict data from the Armed Conflict more pronounced for those also affected by floods. The Location and Event Data Project. The analysis employs study also investigates these effects on households’ income, difference-in-difference regressions to examine whether albeit with a smaller sample. Similar findings, although less conflicts have a measurable effect on households and robust, were noted when analyzing income trends. The find- whether this effect is intensified when considering flood ings underscore the compounded vulnerabilities faced by exposure. The study focuses on households’ consumption households in conflict and flood-prone areas, highlighting expenditure outcomes, comparing conflict-affected and the need for integrated policy interventions to address the non-conflict-affected households, and further narrows compounded impacts of these shocks. This paper is a product of the Urban, Disaster Risk Management, Resilience and Land Global Department. 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 kbenbih@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 Unpacking the Disaster-FCV Nexus: Household Economic Impacts of Conflict and Floods in Nigeria Karima Ben Bih; World Bank Bramka Jafino; World Bank Chloe Desjonqueres; World Bank Solene Masson; World Bank Keywords: Conflict Impact; Flood-Affected Households; Nigeria; Difference in Difference; Consumption Expenditure; Geo-spatial Data, LSMS, Microeconomics JEL: D10, D12, D74, D81, F35, F51, R15 The authors are grateful to Arden Finn, Alvina Erman, Edmundo Murrugarra, and Mook Bangalore for their thoughtful comments, suggestions, and guidance. The paper has benefitted from support from Esesua Ikpefan, Katie Peters, and Hiromi Akiyama. Introduction This paper investigates the impact of conflicts on flood-affected households in Nigeria. Using observational household survey data (the Living Standards Measurement Survey – “LSMS”) and geo- spatial conflict data (from ACLED), the analysis relies on a balanced panel dataset derived from LSMS data collected in 2012, 2015 and 2018, and restricts the sample to conflicts occurring only after 2015. This paper uses difference-in-difference regressions to investigate the measurable effects of conflict on household consumption expenditure, and whether these effects are intensified by flood exposure. We analyze outcomes for both conflict-affected and non-conflict-affected households, considering their flood exposure, and specifically assess the impact of conflict on flood-affected households. Results show that conflict-affected households experience lower consumption expenditure than households not affected by conflict. This difference is accentuated for conflict-affected households that are also affected by a flood. We further corroborate these findings with a smaller scale analysis of income utilizing the same methodology. The findings of the analysis of impact on income can be found in Annex A. Data and methodology Data The data used in this analysis is built upon household survey and geospatial data for conflict and flooding. The survey data is a balanced panel dataset derived from Living Standards Measurement Survey (LSMS) data collected across Nigeria in 2012, 1 2015, 2 and 2018, 3 to estimate the impact of conflict on households’ consumption expenditure (in Nigerian naira, ₦). 4 Although LSMS surveys are constructed to facilitate longitudinal studies, the LSMS data of Nigeria is characterized by a lack of harmonization over time, marked by high attrition rates that may have been caused by cross-state displacement, and by significant questionnaire differences introduced in the 2018 survey. The full dataset includes 4,770 households from the 2012 survey, 4,582 households from the 2015 survey, and 4,980 from the 2018 survey. To longitudinally track the same households over time for each variable, we limited the sample to a strict balanced panel dataset, which resulted in a total of 3,405 observations (1,135 households per survey year), and one viable outcome 1 World Bank. (2013). Nigeria - General Household Survey, Panel 2012-2013, Wave 2. Retrieved from https://worldbankgroup.sharepoint.com/sites/afr-wsafr- nisegknowledgeplatform/SitePages/PublishingPages/Country-Survey-Data-11032021-081638.aspx 2 World Bank. (2016). Nigeria - General Household Survey, Panel 2015-2016, Wave 3. Retrieved from https://worldbankgroup.sharepoint.com/sites/afr-wsafr- nisegknowledgeplatform/SitePages/PublishingPages/Country-Survey-Data-11032021-081638.aspx 3 World Bank. (2019). Nigeria - General Household Survey, Panel 2018-2019, Wave 4. Retrieved from https://worldbankgroup.sharepoint.com/sites/afr-wsafr- nisegknowledgeplatform/SitePages/PublishingPages/Country-Survey-Data-11032021-081638.aspx 4 LSMS data is available for 2010 in Nigeria, and using it would have increased the statistical inference and robustness of our model. However, given the need to use a balanced panel dataset, relying on 2010 implied losing too many observations, significantly impeding statistical inference. We thus elected to use three instead of four survey years. 2 variable: consumption expenditure. To assess the validity of our findings, we have also run the regressions on income, presented in Annex A. 5 Consumption expenditure Consumption expenditure is proxied by the utilities’ expenditure variable utl_exp in 2012 and 2015 data, located in the ‘household SSA harmonized’ section. In 2018, such expenses are captured by the non-food expenditures variable nfdtexp, located in the ‘poverty SSA harmonized’ section. Despite being located in different datasets and having different names, both variables capture household-level expenses for water, electricity, and gas consumption, 6 to make up our consumption expenditure outcome variable. The data is summarized in Table 1. Table 1. Summary statistics for annual consumption expenditure (₦) for each wave and overall. Year Mean Std. Dev. Numb. Obs. Numb. Of 0 2012 232,574.6 428,249.5 1,027 0 (0%) 2015 278,941.4 359,591.3 1,027 0 (0%) 2018 65,901.4 98,385.6 1,027 100 (9%) Total 192,472.5 340,238.8 3,081 (total) 100 (3%) ACLED conflict data This study’s primary objective is to estimate the impact of conflicts on populations that are also affected by a disaster (flooding). To define what constitutes conflict in Nigeria this paper follows the World Bank definition, which primarily encompasses violent conflicts. This definition is closely aligned with the ACLED dataset, which documents geolocated conflicts, their intensity, and their type. Accordingly, this paper focuses on violent conflict and includes events such as battles, violence against civilians, and explosions or remote violence. To maintain temporal coherence and separate conflict from non-conflict households, this analysis focuses specifically on conflicts that occurred from 2015 onwards, thus separating Treatment households affected by conflict from 2016 to 2018 (the post-period) from Control households that are considered not to have been exposed to conflict for the entire period (2011 to 2023 7). Household geocoordinates Geolocated conflict data from ACLED is integrated into 2012 LSMS data, which is the only survey year used in this analysis that contains household GPS coordinates (Figure 1). This linkage facilitates a comprehensive understanding of the geographic distribution and impact of conflicts on surveyed households. LSMS survey households are randomly geolocated within a 2km buffer zone in urban areas and within 5km to 10km buffer zones in rural and extremely remote areas. Given this feature and taking the spatial distribution of conflicts in the country into account (Figure 2), we apply a buffer zone of 10km around households GPS points. This buffer zone allows for 5 Income is presented in the annex due to data quality limitations. To compare income through time without accounting for inflation, yearly income is divided by the percentage annual Consumer Price Index provided here https://data.worldbank.org/indicator/FP.CPI.TOTL.ZG?locations=NG. 6 LSMS data does not provide a breakdown of spendings on each of the utilities. 7 The wide time range of ACLED conflict data available enables the inclusion of more years to ensure the low likelihood of a household being exposed to conflict at any point in time before, during, and after LSMS surveys. This not only guarantees low likelihood of exposure, but also helps avoid any reverse causality (e.g. a significant flood event influencing the incidence of conflict). 3 enough observations across the two groups: households affected by conflict, and households located outside conflict-affected areas. Figure 1. Household GPS points from Nigeria’s 2012 LSMS survey 4 Figure 2. ACLED conflict data from January 2011 to December 2023 (black dots) and 2012 LSMS household geocoordinates with a 10km buffer (red dots) Flood exposure While the exogeneous shock presented in this paper is violent conflicts, the analysis is also concerned with flood-affected households. The paper does not explore flood as a shock per se, it investigates how flood-affected households are affected by conflict. This paper uses two ways to identify flood-affected households: survey-based, using LSMS self-reported data, and satellite- based, overlaying 2012 household geocoordinates with 2012 flood footprint maps. 8 LSMS data contains a self-reported variable on whether the household has been affected by a climate or disaster shock in any of the 3 survey years named ‘shock_cd’. Based on that variable, the first approach this paper applies is to restrict the sample to floods using the ‘shock_desc’ string variable, which describes the type of shock the households claim to have experienced. There is debate in literature over which of these two measures is most accurate. Survey data is sometimes considered more subjective, with a potential bias introduced by idiosyncrasies of household respondents (Guiteras, Jina and Mobarak, 2015) (Bangalore, 2023). Satellite imagery could be perceived as being more objective than self-reported data, however, the indirect impacts of a flood event may also be significant outside satellite-observed zones, making it difficult to identify 8 Humanitarian Data Exchange. (2012). Nigeria Flood Extents - November 2012. Retrieved from https://data.humdata.org/dataset/nigeria-flood-extents-nov-2012-fod 5 impacted households. Bangalore (2023) mentions that in Nigeria, flood impacts vary greatly even at very local scales, as geographical determinants and water dynamics can inundate certain areas but leave neighboring homes unaffected, a fact further supported in other countries (Patankar, 2015; Hallegatte, Bangalore, et al., 2016; Winsemius et al., 2018). As both measures have their own limitations, we chose to apply both methods to this analysis. On the one hand, using LSMS self-reported data for a single survey year leads to a significant drop in the number of observations. To address this limitation, the first approach identifies flood-affected households as those reporting flood impact in any of the three survey years. On the other hand, relying solely on satellite imagery includes too few surveyed households. The second approach creates a balanced dataset by applying a 5km buffer around surveyed households. This restriction results in a considerable loss of observations, from about 1,000 observations a year to 200 observations a year. Despite this drop in the number of observations, the satellite imagery-based results presented below support the evidence brought about by the survey-based specification, which includes all three survey years. Methodology Difference-in-difference Difference-in-difference (DiD) regressions, a quasi-experimental method commonly known as an ex-post impact evaluation, are commonly used to analyze contrasting outcomes between two groups – a treatment and a control group – over time, accounting for their distinct characteristics. The difference-in-difference research design is well suited for event study, and the quantification of the impact of an unexpected shock on economic outcomes. This method has been widely used in the literature (Card & Krueger, 2000; Galiani et al., 2005). A critical assumption underlying this approach is the existence of a common trend between the groups over time, barring exogenous shocks. The parallel trends assumption in DiD regressions states that, in the absence of treatment, the treated and control groups would have experienced the same trajectory in outcomes over time. This allows any observed post-treatment difference between the groups to be attributed to the treatment effect, as opposed to pre-existing differences in trends. In the absence of this parallel trend, it would not be possible to estimate causal evidence of impact. The main research aim of this paper is to investigate whether conflicts have a measurable effect on households (1), and whether this effect changes when factoring in flood exposure (2). Conflicts thus serve as the exogenous shock, and household consumption expenditure as the outcome variable. The treatment group includes households affected by conflicts between 2016 and 2018 (not in 2012), while the control group encompasses all households not affected by a conflict before and after 2015. The first difference-in-difference specification is as follows: Yi,t = Treated β1 + Post Period β2 + , β3 + ε, (1) whereby Yi,t stands for the logarithm of the consumption expenditure of household i at time t, where t takes a value of 2012, 2015 or 2018; i is a dummy variable that takes a value 1 when household i is affected by conflict after 2015, and a value of 0 when household i is not affected by conflict during the entire period. Post Periodt is a dummy variable that takes a value of 1 for 2018, the time period after the exogeneous shock (i.e. conflicts that have occurred after 2015); The variable, the variable of interest in a difference-in-difference specification, is the interaction between the Treated and Post Period variables. 6 The second difference-in-difference specification, which accounts for flood exposure, is as follows: Yi,t = Treated β1 + Post Period β2 + β3 + ε (2) The second specification takes two approaches to restrict the sample to flood-affected households: the first is survey-based, relying on self-reported LSMS data, and the second is satellite-based, restricting the dataset to households affected by the 2012 flood event using satellite imagery. In both specifications, errors are clustered at the household level to control for serial correlation and the average values of the 5th-95th percentile are assigned to outliers. Results First specification: Conflict- versus non-conflict-affected households The first specification investigates the relationship between conflicts and households’ consumption expenditure. Results are presented in Table 2, for a sample of 422 observations in the Treatment group and 605 observations in the Control group, for a total of 1,027 observations per survey year. The distribution of consumption expenditure allows for the normalization of the distribution using a logarithmic transformation, as only a few of observations are registered as zero in 2018. As such, results can be interpreted as percentages. Conflicts are associated with an average reduction of up to 16.7% households’ consumption expenditure compared to unaffected households (Table 2). These preliminary results demonstrate high significance, which indicates the substantial impact of conflict. Table 2. DiD for the log(Consumption expenditure) in 2012-2018 (first specification) VARIABLES log(Consumption expenditure) Treated -0.425*** (0.0484) Post Period -1.630*** (0.0500) -0.167** (0.0838) Observations 3,081 Number of HH 1,027 Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 Figure 3 and Table 3 show the common trend and robustness check for the pre-time period (2012 to 2015), with 2015 acting as the Post Period variable. The common trend of the consumption expenditure variable of the full dataset is visually validated (Figure 3), further supported by the Treatment variable in Table 3, which is not significant for the pre-time period. The relatively large 7 dataset (more than 1,000 observations) reinforces confidence around the causality of the first specification’s estimation. Figure 3. Changes in log(consumption expenditure) due to the occurrence of conflict (first specification) Note: The exogeneous shock is based on conflicts that have occurred from 1st January of 2016. This means 2015 is the baseline, whereby neither group (control and treatment) is conflict-affected. It does not mean that there was no conflict in Nigeria before 2016. The dataset is restricted to allow for a difference in difference regression analysis. Table 3. DiD for the log(Consumption expenditure) in 2012-2015 VARIABLES log(Consumption expenditure) Treated -0.459*** (0.0532) Post Period 0.203*** (0.0254) 0.0671 (0.0408) Observations 2,054 Number of HH 1,027 Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 8 Second specification: Impact of conflict on flood-affected households After confirming the impact of conflicts on Nigerian households’ consumption expenditure, this analysis aims at estimating the impact of conflict on households exposed to flood risk as identified via (a) survey-based (self-reported) and (b) satellite imagery-based (2012 flood map) methods. Restricting the sample to flood-affected households significantly reduces the number of observations, potentially compromising statistical inference and introducing some bias. The results presented below are thus rather interpreted as correlations than causations. a. Survey-based identification The survey-based sample of flood-affected households contains 122 observations per survey year, including 52 observations in the Treatment group and 70 observations in the Control group. As few observations are reported as 0 in 2018, the distribution of consumption expenditure is normalized using a log transformation and results can be interpreted as percentage. However, due to the low number of observations, and given the low variation in consumption expenditure in this sample’s statistical distribution, the outliers assignment applied in the first specification is not replicated in the second one. Applying the 5th-95th percentile assumption on outliers would reduce variation and increase potential estimation bias. Compared to the first specification, which included the entire balanced panel dataset, households reporting flood-related loss during the 2012-2018 experience a larger negative impact on their consumption expenses when also affected by conflicts, as opposed to households not affected by conflicts. Households affected by conflict (exogenous shock) and by floods experienced a reduction in consumption expenditure of an average of 152%, compared to flood- affected households that do not experience conflict (Table 4). These results demonstrate statistical significance at 5%. However, due to the small number of observations and reliability limitations of self-reported data, this finding can be considered a correlation as a larger number of observations to assure causality linkages. Table 4. DiD for the log(Consumption expenditure) in 2012-2018 (second specification (a)) VARIABLES Log(Consumption expenditure) Treated -0.261** (0.133) Post Period -2.436*** (0.374) -1.522** (0.708) Observations 366 Number of HH 122 Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 Figure 4 and Table 5 present the common trend assumption and robustness check for the pre- time period (2012-2015), with 2015 onward acting as the Post Period variable. The common trend assumption is somewhat visually validated, supported by the result obtained for the Treatment variable, which is not significant. This result reinforces the correlation identified in Table 4. 9 Figure 4. Changes in log(consumption expenditure) due to conflict (second specification (a)) Note: The exogeneous shock is based on conflicts that have occurred from 1st January of 2016. This means 2015 is the baseline, whereby neither group (control and treatment) is conflict-affected. It does not mean that there was no conflict in Nigeria before 2016. The dataset is restricted to allow for a difference in difference regression analysis. Table 5. DiD for log(consumption expenditure) over 2012-2015 (second specification (a)) VARIABLES Log(consumption expenditure) Treated -0.0932 (0.149) Post Period 0.351** (0.137) -0.238 (0.211) Observations 244 Number of HH 122 Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 b. Satellite imagery-based identification The satellite imagery-based sample of flood-affected households results in a total of 211 observations per LSMS survey year, including 19 observations in the Treatment group and 192 observations in the Control group. In this scenario, distribution of consumption expenditure can be normalized by applying a log transformation using the few observations reported in the 2018 10 survey wave. Results can then be interpreted as percentages. Although there are few observations, the variation observed in the distribution is sufficient to apply the 5th-95th percentile assumption to outliers. However, the imbalance in the number of observations between the Treatment and Control groups calls for caution when it comes to interpretation. The results show that conflicts are associated with a decrease of up to 58% in households’ consumption expenditure compared to those unaffected by conflicts. These results demonstrate statistical significance at 10%, although due to the low number of observations, only point to a correlation rather than causality. Table 6. DiD over 2015-2018 for Consumption expenditure VARIABLES Log(consumption expenditure) Treated -1.150*** (0.169) Post Period -1.682*** (0.0875) -0.584* (0.344) Observations 422 Number of hhid 211 Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 Table 7 and figure 5 serve as a robustness check for the pre-time period (2012-2015), with 2015 acting at the Post Period variable. The result obtained for the Treatment variable is not significant, reinforcing the correlation identified in Table 6. 11 Figure 5. Changes in log(consumption expenditure) due to conflict for flood satellite based affected households Note: The exogeneous shock is based on conflicts that have occurred from 1st January of 2016. This means 2015 is the baseline, whereby neither group (control and treatment) is conflict-affected. It does not mean that there was no conflict in Nigeria before 2016. The dataset is restricted to allow for a difference in difference regression analysis. Table 7. DiD over 2012-2015 for log(consumption expenditure) VARIABLES Log(consumption expenditure) Treated -1.342*** (0.132) Post Period 0.0879** (0.0431) 0.192 (0.123) Observations 422 Number of HH 211 Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 Households affected by the 2012 flood (as identified by the satellite imagery of the 2012 flood) experienced a larger negative impact on their consumption expenses when also affected by conflict than households affected by a flood but not by conflict. Although lower than the drop 12 observed for households that self-reported having experienced a flood event, the 58% figure is higher than the loss of consumption expenditure estimated in the first specifications - where households suffering from only conflict saw their consumption expenditure decreasing by 16%. The large difference observed between the survey-based (a) and satellite imagery-based (b) specifications could be explained by the fact that the satellite-based specification accounts for a single flood event (based on the 2012 flood map) while the survey-based specification includes households that declare having experienced one to three flood events. These results thus do not account for flood intensity. However, despite the complexity encountered to evidence causal relationships in this study, the results obtained in all three specifications are all pointing to a single answer to this paper’s research question: conflict has a negative effect on households’ consumption expenditure, and this effect is likely greater when the household is also exposed to flood risk. Limitations The results presented above are highly significant and robust, particularly in the first specification, where a strong causal relationship is established. The second specification, while robust, is based on a smaller sample and indicates correlation rather than causation, due to data limitations. High attrition rates in household coverage across survey years reduce the number of available observations when obtaining a balanced dataset. Since GPS coordinates are only available for one survey year (2012), the balanced panel dataset is based on that survey, further restricting the number of observations. Additionally, discrepancies in questionnaire structures across survey years presented challenges to ensuring harmonization, constraining the analysis to a single reliable outcome variable: non-food-related consumption expenditure, as proxied by utilities expenses. Finally, identifying flood-affected households over the entire period further reduced the dataset, limiting causal inference possibilities. Despite these challenges however, this paper’s breakthrough findings highlight a large negative impact of conflict on households, and an even larger negative impact for households affected by both floods and conflicts. Comparing results across specifications underscores the persistent negative impact of conflict on populations, and the co-location of conflict and flood exposure presents significant implications when it comes to exacerbating household vulnerabilities. Discussion This study is one of the initial attempts to quantitatively estimate the impacts of conflicts and floods on households. While it faces challenges related to the observational data used, the study and its findings have significant implications for disaster risk management policies, investments, and knowledge, particularly in the case of Nigeria. Integrating conflict exposure dynamics into decision-making processes for disaster resilience, particularly at the state and regional levels, is crucial for effective DRM policy formulation and mainstreaming across the sectors. A better understanding of impacts channels by which conflict dynamics and actors’ relationships can affect resilience-building efforts is critical. In practice, this involves understanding the prevailing and dynamic conflict conditions, the contextual nuances of how and in what ways this may exacerbate the impact of disasters, and the reverberating impact on the resources available for disaster response – from the household to the state level. All of which significantly affect the social fabric that is crucial for recovery (see Table 8). Table 8. Summary results log(consumption expenditure) 13 1st specification -0.167** conflict-affected households 2nd specification (a) -1.522** conflict- and flood-affected households (survey based) 2nd specification (b) -0.584* conflict- and flood-affected households (satellite based) There is a need to improve data collection efforts to capture the intersectionality of conflicts and natural hazard-related disasters, including those impacted by climate change. This includes incorporating conflict-related variables into existing surveys and regional-national-local risk knowledge assessments, and enhancing collaboration between relevant agencies to ensure comprehensive data coverage, harmonizing methodologies, and coordinating data needs and availability at the country level. Within the World Bank, the LSMS survey, while comprehensive, lacks a specific focus on populations affected by disasters. The discontinuity across survey waves impedes accurate quantification of the impact of conflict-affected populations in areas at risk of flooding. Further refining of the survey would yield to a many differentiated uses of the datasets that could further provide granularity on the impacts of the disaster and conflict nexus on poverty; an agenda that will become ever more urgent as climate change impacts continue to manifest in FCV settings. The inquiry into the disparity in the manifestation of conflict and its impact on households' resilience to climate and disaster shocks raises several questions. For example, efforts to understand the underlying reasons for the observed differences in conflict manifestation are still needed. Identifying the specific aspects of conflict that most significantly undermine households' ability to withstand such shocks is crucial. These aspects may include displacement and migration, disruption of social networks, economic instability, and the erosion of institutional capacities, all of which can exacerbate household vulnerability. Further investigation is needed to answer these questions, as they are vital for enhancing the effectiveness of DRM strategies in conflict-affected areas, ensuring that they are both responsive and resilient to the complexities introduced by conflict dynamics. Conclusion This study aimed to estimate a causal relationship between conflict, flood exposure, and consumption expenditure. The paper used two primary data sources: i) three years of the LSMS data (2012, 2015, and 2018) from Nigeria, which include GPS localization of households and observational information; and ii) geolocated ACLED data for violent conflict events occurring from 2011 to 2023. Using a difference-in-difference approach, the paper investigated the economic impact of conflict on all households, and then specifically on flood-affected households. Results show a negative association between conflict and both consumption expenditure and income (see detailed income analysis in Annex A), despite several limitations related to the high attrition rates that reduced the number of observations, GPS coordinates availability, and discrepancies in questionnaire structures, which limited the causal inference possibilities. 14 Annex A1. Summary statistics As previously mentioned, we encountered issues with our variables of interest and the specifications we are using, such as the number of observations with ‘zero’ income, which prevents us from using logarithms for income. As such, it is deemed useful to provide descriptive statistics on this variable to facilitate interpretation of results. The tables below represent the income mean for each year over the whole dataset (first specification) as well as for the restricted dataset of self-reported flood-affected households (second specification). Table A1. Summary statistics of annual income (₦) for the first specification (whole dataset) Year Mean Std. Dev. Numb. Obs. 2012 3043.524 8902.957 1,135 2015 2393.083 13117.47 1,135 2018 8493.609 23919.36 1,135 Average 4643.405 16787.27 3,405 (total) Table A2. Summary statistics of annual income (₦) for the second specification (restricted dataset) Year Mean Std. Dev. Numb. Obs. 2012 1676.007 3616.121 122 2015 1329.572 3806.218 122 2018 4232.202 12468.69 122 Average 2412.593 7896.609 366 (total) At an aggregate, the annual income for the first specification (whole dataset) exceeds that of the restricted dataset by half. We thus can see that, on average, households declared an average income of 4,643.41 naira over the three years (2012-2015-2018) of our first specification, whereas the average annual income within the restricted panel data fell down to 2,412.59 naira. This difference might be attributed to the higher vulnerability and impoverishment of households affected by floods at least once during the observed time period, compared to households that were not affected by flood as well as those affected by floods (whole dataset). When looking at the yearly average income for the two datasets, we also observe that 2015 was the year where annual income was the lowest. This could be either explained by a specific conflict event 9 or may also come from specificities of the data. 10 As already mentioned, for 2015 and 2018 year we observe more than 70% of missing values or households 10 declaring earning 0 income. 15 A2. Robustness check using income as variable To ensure validity of the results, we proceed with a second analysis using the same methodology described in the data and methodology section above and replacing the consumption expenditure variable of the LSMS with income from the same dataset, in both specifications used in the main paper. As this paper build on previous research that focuses on a top-down approach to understanding impacts of disasters in conflict settings, using night light as a proxy for economic activity (Ben Bih et AL, 2024), income as a variable is indeed the closest proxy to economic activity, and thus to economic impacts of disasters in conflict settings, at the micro/household levels. However, due to data limitations, this proxy is used as a robustness check instead of a main indicator in the current paper: Due to the significant drop in income between 2012 and 2015, resulting in higher levels of reported $0 in the income category, this variable cannot be used as a main indicator. Results of the Overall dataset on income To ensure harmonization across the three survey years, the income variable is calculated as the average wage of the household’s primary and secondary livelihoods. This excludes income compensation originating from remittances and humanitarian assistance, which is available in 2012 and 2015 surveys but is not available in 2018. While this exclusion leads to a decrease in reported income, particularly in the 2015 and 2018 waves where over 70% of observations indicate 0 income, this approach helps maintain coherence in income computation across all the three waves (see Table A3). Table A3. Summary statistics for annual income (₦) for each wave and overall Number of Year Mean Standard Deviation Number of 0 Observations 2012 3,043.5 8,903.0 1,135 445 (39.2%) 2015 2,393.1 13,117.5 1,135 831 (73.21%) 2018 8,493.6 23,919.4 1,135 814 (71.71%) Average 4,643.4 16,787.3 3,405 (total) 2,090 (61.4%) When taking income as the dependent variable (see table A5), our analysis includes 419 observations in the treated groups and 716 observations in the control group, amounting to 991 observations available for each wave of LSMS survey. Our treatment variable is highly significant. However, due to the prevalence of 0 values in the datasets, and although a logarithmic transformation would help normalize income distribution, we keep the dependent variable (income) in its original form instead of using a logarithmic transformation. 11 Using a log- transformation is helpful in case the data is continuous and we have low number of observations taking ‘0’ since it is going to skew the distribution of the income. However, the important number of 0 and if we were applying a log-transformation to income we could distort the statistical inference, which in turn would bias our results. Our findings indicate that conflicts lead to an average decrease of 4105 Naira in household income. While this decrease might initially seem modest, it represents more than the annual 11 https://academic.oup.com/qje/advance-article-abstract/doi/10.1093/qje/qjad054/7473710?login=false 16 average income in 2015 (2393.083 Naira) and accounts for one half of the average income in 2018 (8493.609 Naira), as highlighted in Table A3. Table A4. DiD over Income for the 2012-2018 time period (1) VARIABLES Income treated_group -1,310*** (476.3) postperiod 7,312*** (950.4) Treatment -4,105*** (1,214) Gender FE -2,780*** (547.7) Observations 3,405 Number of hhid 1,135 Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 Figure A1. Changes in income due to the occurrence of conflict (first specification) Note : the exogeneous shock is related with conflicts that are occurring and accounting from 1st January of 2016 such that 2015 is the baseline where the two groups (control and treated) are not conflicts affected. It does not mean before 2016 there was no conflict in Nigeria. We are restricted our dataset to be able to apply the difference in difference. 17 Table A5. DiD over Income before conflict (2012-2015) (1) VARIABLES Income treated_group -1,103* (596.8) postperiod -456.6 (607.3) Treatment -502.4 (806.3) Gender FE -1,897*** (341.7) Observations 2,270 Number of hhid 1,135 Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 In table A4 and figure A1, we show the common trend robustness check for the pre-time period. Regarding the income variable over the whole dataset the common trend seems to be visually checked (figure 3). We can see that the treatment variable is also not significative for the pre-time period. 18 Results of the restricted data on income When income serves as the dependent variable, we observe 52 observations in the treated groups and 70 observations within the control group, resulting in a total of 122 observations available for each LSMS survey wave. Our treatment variable demonstrates high significance. Results shows that households impacted by conflicts in a flood setting experience an average income decrease of 4,912 naira compared to households affected by a flood but not conflict from 2012 to 2018. When we compare these findings with the previous results, where floods were not considered we note a larger negative impact when households are affected by both floods and conflicts in the Nigeria context. Referring to the summary statistics table 1, the adverse effects of conflicts on households are further exacerbated when they coincide with flood events. The income loss of 4,912 naira is important given the average annual income in 2015 (1,329,572 naira) and in 2018 (4,232,202 naira) 12 as noted in table A2. Table A6. DiD over annual income for 2012-2018 time period (1) VARIABLES Income treated_group -1,787 (1,115) postperiod 5,042*** (926.7) Treatment -4,912*** (1,410) Gender FE -3,564*** (1,365) Observations 366 Number of hhid 122 Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 12 Note that the value of Naira has diminished significantly since 2015 19 Figure A2. Changes in income due to the occurrence of conflict (second specification) Note: the exogeneous shock is related with conflicts that are occurring and accounting from 1st January of 2016 such that 2015 is the baseline where the two groups (control and treated) are not conflicts affected. It does not mean before 2016 there was no conflict in Nigeria. We are restricted our dataset to be able to apply the difference in difference. Table A7. DiD over Income for 2012-2015 time period (1) VARIABLES Income treated_group -1,629** (663.9) postperiod -272.5 (285.1) Treatment -48.96 (433.9) Gender FE -1,295* (752.7) Observations 244 Number of hhid 122 Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 In table A7 and figure A2, we show the common trend robustness check for the pre-time period. When it comes to the income variable, the common trend can be visually checked (figure A2) and confirmed by Table A7. We can see that the treatment variable is also not significative for the pre- time period. 20 Table A8. Summary results of income specifications Income 1st specification 4105*** conflict-affected population 2nd specification 4912*** conflict- and flood-affected population 21