Methods

Units of analysis

The joint impact assessment in CC1 and the risk assessment in CC4 will be formulated at two levels of analysis: Countries as defined by Gleditsch & Ward [1] and Weidmann et al. [2], and the PRIO-GRID cell according to Tollefsen et al. [3]. For these, the aim throughout the programme will be to provide estimates of impact across all units globally, using techniques to handle missing data and to use estimates developed in the WPs based on samples as points of departure for estimating effects across entire conflict zones. In the outcome-specific work packages,we will adapt units of analysis according to data availability. For instance, we will link surveys, which typically do not have universal coverage, to geographic point coordinates or administrative regions.

Data

In several parts of the programme, we will seek to quantify the universal, average (although condi- tional) impact of conflict on the various outcomes. This requires that we define a metric to compare the intensity of the violence that is causing the impacts, with uniform coverage and strict adhe- sions to definitions that do not vary over time or space. The leading source for such data is the Uppsala Conflict Data programme [4], that have collected the number of recorded fatalities in all armed conflicts since 1989 that caused at least 25 battle-related deaths. The UCDP-GED dataset [5], that records all events within these conflicts with precise indications of where and when vio- lence occurred as well as estimates of how many were killed in direct violence. The data-collection component will collaborate with R2 and R3.

To measure the outcomes, we will make use of data from the World Bank Indicators, from the Global Burden of Disease project, the WHO, and survey data from the CE-DAT [6] and EM-DAT databases [7] as well as from LSMS [7], DHS [7], MICS [10], and LiTS [11]. Where feasible, we will use metrics and sources that allow analyzing impact separately by gender and by social group. In survey data, such characteristics are often included in aggregates or individual records, or can be inferred from the geographical location of the respondents.

Handling missing and incomplete data

The programme will systematically handle missing and incomplete data. To compensate for uncer- tainty about the exact location of some conflict events, we will relate the impact to the probability distribution over conflict locations. This is computed using multiple imputation [12]. Other vari- ables used in the programme, either as outcomes or as core conditional variables, also contain missing observations. This is particularly the case for survey data [13]. To avoid biased parameter estimates and/or standard errors [14], we will use multiple-imputation techniques (e.g.the Amelia II package).

Meta-studies

The programme will summarize the existing empirical research on the impact of conflict through a meta-analysis – the statistical analysis of published research findings on a given hypothesis or empirical effect [15, 16–18]. Following these studies, we will i) define systematic criteria for in- cluding/excluding the studies and estimated effects; ii) do a broad systematic search of the literature to identify all studies that meet the criteria; iii) collect textual information and its conversion to quantitative features characterizing the studies; iv) analyze statistically the empirical findings in the selected sample.

Modeling and estimating exposure to conflict across time and Space

Earlier studies [e.g. 19] typically estimate impact by in effect limiting exposure to the area within 50km of UCDP events. To provide a better model, both as input to modeling in other parts of the programme and as a deliverable in itself, we will work out a set of conflict exposure models. As a first step, the impact zone will be modeled as circles around the conflict events and functions that let impact decay over time, adding the size of local populations. We will then explore several extensions to such models – parameterize them and use statistically estimate precisely how the impact of conflict is reduced with a longer distance to conflict events in space and time. To identify the effect, we will use the observed outcomes treated in the WPs described below, along with a model for expected outcomea in a location in the absence of conflict. We will also consider more advanced concepts of distance, taking for instance into account that the impact of conflict may fail to cross international borders and be more likely to move through areas with contiguous settlements than across sparsely populated regions. We will explore network models or more complicated models of linkages between locations, and validate using out-of-sample evaluation of predictive performance. Finally, taking the probabilistic hazard estimates derived in the early-warning model as estimates, we will explore the importance of latent (unobserved) armed conflict.

Constructing a vulnerability index

CC3 sets out to construct a composite Vulnerability Index (VI) [20] at both the national and the sub-national level. Adapting the model developed by Eriksson et al. [21], the VI will be math- ematically derived by aggregating a combination of variables that represent different dimensions of communities’ vulnerability [22]. The main sub-components will be institutional, social and economic vulnerability/resilience, and community capital [23].

First, we will identify the relevant and robust variables to represent each sub-dimension of vul- nerability using the data collected in the programme. Next, the variable scores in each sub-index will be averaged to obtain an individual score for each sub-dimension of vulnerability (e.g. eco- nomic), and further aggregated to get the final VI [22, 24]. We will perform extensive sensitivity and robustness tests to evaluate how methodological choices influence the scores [25], including validating against existing indices such as the Social Vulnerability Index [26] and the Predictive Indicator of Vulnerability [27].

The VI will be used together with the early-warning system (CC4) to provide impact assessments for conflict and non-conflict scenarios. Hence, the VI will serve as both an ex-ante counterfactual and as an assessment of the conflict impact on societal and community vulnerability. As a final step, the Vulnerability assessment will be used to assess the risk of natural disasters and climatic shocks in conflict-affected countries according to IPCC’s definition.

To compute risk, we will integrate the conflict-scenario vulnerability assessment, with the esti- mates of hazard and exposure. We will collect data on the severity of climatic shocks and natural disasters to give an accurate representation of hazards, whereas demographic data including inhab- ited area, population size and distribution at both national and individual level will be included to proxy exposure. This will enable us to provide a thorough assessment of compound risks [28], induced by the combination of conflict exposure and natural disasters.

Bayesian models

We will use Bayesian models to estimate malnutrition, vaccination coverage, and excess mortality. These are useful in the sparse-data situation in conflict settings, as they allow effectively accounting for all sources of variability, incorporate past knowledge from previous studies (‘priors’), handling of missing data, and generate an easily interpretable direct posterior probabilistic value [29].

Verbal autopsy

To estimate causes of death (CoD) in conflict settings, the programme will adapt verbal autopsy (VA) approaches to conflict settings. These build on face-to-face interviews with a close relative of the deceased, collecting information on signs, symptoms, and circumstances leading to death, which is then interpreted into CoD. The increasing use of VA to collect routine CoD data in stable, low- income settings has led to many versions of the instrument such as the WHO VA 2016. While these tools are commonly used in low- and middle-income countries, their use in humanitarian settings with severe challenges in terms of security and access is novel and therefore, must be simplified, adapted and tested. We will test in a field experiment a reformulated VA tool that can: i) be used in complex emergency settings and ii) add and validate data to survey tools in conflict settings. We will compare VA results in a refugee camp with hospital records to validate an adapted instrument.

Handling estimation uncertainty

With incomplete input data and heavy reliance on various estimation techniques, it is necessary to account for the uncertainty in the results. This is often ignored in the literature. The study of Cooper et al. [30, SI, p. 8], for instance, ignores uncertainty completely. We will make use of a suite of techniques to cover various sources of uncertainty. Where available standard pack- ages do not provide such estimates, we will use bootstrapping techniques [31] to assess statistical uncertainty. We will make use of model ensembles to decrease the uncertainty about model specifi- cations, and apply simulation approaches from the early-warning system to bind the various sources of uncertainty together.

Risk assessment: forecasting and scenario simulation

To complete our conflict risk assessment, we will combine the estimates of impacts of observed conflict, of exposure, and vulnerability with modeling of the probability of conflicts of varying intensity. This is necessary for two reasons. First, according to the ‘early warning, early action’ doctrine prevention is better than mitigation [32]. Moreover, even when prevention is impossible, an estimate of the likely future scale and duration of the conflict is important. The risk assessment has to be forward-looking and take uncertainty into account. To do this, we will extend the VIEWS system [33, 34] to have global coverage, to forecast the number of fatalities in addition to the dichotomous presence/non-presence of conflict, and leverage the news-reading component described in WP4 to strengthen the warning of new conflicts [35].

The estimated probability distributions over conflict severity will be combined with the estimates of impact, exposure, and vulnerability to provide maps over the range of likely humanitarian effects of conflict.

The early-warning system will be maintained as a live system that produces a monthly as- sessment of the risk of humanitarian disasters. The system can also be used to develop scenario assessments (CC5).

References

1. Gleditsch, K. S. & Ward, M. D. A Revised List of Independent States since the Congress of Vienna. International Interactions 25, 393–413 (1999).
2. Weidmann, N. B., Dorussen, H. & Gleditsch, K. S. The Geography of the International System: The CShapes Dataset. International Interactions 36, 86–106 (2010).
3. Tollefsen, A. F., Strand, H. & Buhaug, H. PRIO-GRID: A unified spatial data structure. Journal of Peace Research 49, 363–374 (2012).
4. Pettersson, T. & Öberg, M. Organized violence, 1989–2019. Journal of Peace Research 57, 597–613 (2020).
5. Sundberg, R. & Melander, E. Introducing the UCDP Georeferenced Event Dataset. Journal of Peace Research 50, 523–532 (2013).
6. Altare, C. & Guha-Sapir, D. The Complex Emergency Database: A Global Repository of Small-Scale Surveys on Nutrition, Health and Mortality. PLoS ONE 9 (ed Welte, A.) e109022 (Oct. 21, 2014).
7. Université catholique de Louvain (UCL) - CRED, D. Guha-Sapir. EM-DAT: The Emergency Events Database Brussels, Belgium, www.emdat.be. 2020.
8. The World Bank, Center for Data Development. Living Standards Measurement Study Wash- ington DC, https://www.worldbank.org/en/programs/lsms. 2020.
9. The DHS Program website. Demographic and Health Surveys Funded by USAID, url://www.dhsprogram.com. 2018.
10. UNICEF. Multiple Indicator Cluster Surveys New York, USA, https://mics.unicef.org. 2021.
11. European Bank for Reconstruction and Development. Life in Transition Survey (LITS) Lon- don, UK, https://www.ebrd.com/what-we-do/economic-research- and-data/data/lits.html. 2016.
12. Croicu, M. & Hegre, H. A Fast Spatial Multiple Imputation Procedure for Imprecise Armed Conflict Events. Typescript, Uppsala University. 2018.
13. Manzi, G. et al. Modelling bias in combining small area prevalence estimates from multiple surveys. Journal of the Royal Statistical Society: Series A (Statistics in Society) 174, 31–50 (2011).
14. Allison, P. D. Missing Data in The SAGE handbook of quantitative methods in psychology (eds Millsap, R. E. & Maydeu-Olivares, A.) (Sage Publications, 2009).
15. Bauer, M. et al. Can War Foster Cooperation? Journal of Economic Perspectives 30, 249–74 (2016).
16. Stanley, T. D. & Doucouliagos, H. Meta-regression analysis in economics and business 186 pp. (Routledge, New York, 2012).
17. Stanley, T. D. & Doucouliagos, H. Neither fixed nor random: weighted least squares meta- regression: Weighted Least Squares Meta-Regression Analysis. Research Synthesis Methods 8, 19–42 (Mar. 2017).
18. Vesco, P., Dasgupta, S., Cian, E. D. & Carraro, C. Natural resources and conflict: A meta- analysis of the empirical literature. Ecological Economics 172, 106633 (2020).
19. Wagner, Z. et al. Armed conflict and child mortality in Africa: a geospatial analysis. The Lancet 392, 857 –865 (2018).
20. Diffenbaugh, N. S., Giorgi, F., Raymond, L. & Bi, X. Indicators of 21st century socioclimatic exposure. Proceedings of the National Academy of Sciences 104, 20195–20198 (Dec. 18, 2007).
21. Eriksson, A., Ohlsén, Y. K., Garfield, R. & von Schreeb, J. Who Is Worst Off? Developing a Severity-scoring Model of Complex Emergency Affected Countries in Order to Ensure Needs Based Funding. eng. PLoS currents 7, ecurrents.dis.8e7fb95c7df19c5a9ba56584d6aa2c59 (Nov. 2015).
22. Handbook on constructing composite indicators: methodology and user guide (ed OECD) OCLC: ocn244969711 (OECD, Paris, 2008). 158 pp.
23. Cutter, S. L., Burton, C. G. & Emrich, C. T. Disaster Resilience Indicators for Benchmarking Baseline Conditions. Journal of Homeland Security and Emergency Management 7 (2010).
24. Chen, T.-Y. An outcome-oriented approach to multicriteria decision analysis with intuition- istic fuzzy optimistic/pessimistic operators. Expert Systems with Applications 37, 7762 –7774 (2010).
25. Marzi, S. et al. Constructing a comprehensive disaster resilience index: The case of Italy.PLOS ONE 14 (ed Linkov, I.) e0221585 (Sept. 16, 2019).
26. Cutter, S. L., Boruff, B. J. & Shirley, W. L. Social Vulnerability to Environmental Hazards*.Social Science Quarterly 84, 242–261 (2003).
27. Adger, W. et al. New Indicatorsof Vulnerability and Adaptive Capacity Tyndall Centre for Climate Change Research. Technical Report. Norwich, UK: 2004.
28. Zscheischler, J. et al. Future climate risk from compound events. Nature Climate Change 8, 469–477 (2018).
29. Dunson, D. B. Commentary: Practical Advantages of Bayesian Analysis of Epidemiologic Data. American Journal of Epidemiology 153, 1222–1226 (June 2001).
30. Cooper, M. W. et al. Mapping the effects of drought on child stunting. Proceedings of the National Academy of Sciences 116, 17219–17224 (2019).
31. Efron, B. & Tibshirani, R. The Bootstrap Method for Assessing Statistical Accuracy. Behav- iormetrika 12, 1–35 (1985).
32. World Bank Group & United Nations. Pathways for Peace: Inclusive Approaches to Prevent- ing Violent Conflict. Main Messages and Emerging Policy Directions (International Bank for Reconstruction and Development/The World Bank, 2017).
33. Hegre, H. et al. ViEWS: A political Violence Early Warning System. Journal of Peace Re- search 56, 155–174 (2019).
34. Hegre, H. et al. ViEWS2020: Revising and evaluating the ViEWS political Violence Early- Warning System. Journal of Peace Research In press (2021).
35. Mueller, H. & Rauh, C. Reading Between the Lines: Prediction of Political Violence Using Newspaper Text. American Political Science Review 112, 358–375 (2018).