No Bang for Buck

Abstract

Between 1989 and 2022, an estimated 3.4 million people died as a result of intrastate conflict, with 76% of those deaths occurring in low-income countries, making it a leading cause of humanitarian suffering and economic decline. Official Development Assistance (ODA) is frequently used to promote stability and development in conflict-affected states, yet the extent to which it actually reduces violence remains difficult to quantify. Numerous studies show cases in which ODA worsened conflict rather than reducing it, yet stopping ODA for the countries that need it most seems cruel. This thesis investigates the association between ODA and intrastate conflict-related casualties in recipient countries using a data-driven approach.

The analysis draws on custom-made a panel dataset of ODA-eligible countries spanning 1989 to 2022, combining ODA data from the OECD's Creditor Reporting System, conflict casualty data from the Uppsala Conflict Data Programme's Event Dataset, and a broad set of country-level indicators derived from three established frameworks: the OECD States of Fragility Framework, the World Bank Country Policy and Institutional Assessment, and the Composite Index of Absorptive Capacity. Nine distinct forms of ODA are examined, including total, bilateral, multilateral, and purpose-specific categories targeting construction, government and civil society, and conflict, peace and security.

Two data analysis methods are used. A Dynamic Panel Generalised Method of Moments (GMM) model estimates which forms of ODA and country-level indicators are statistically associated with changes in conflict deaths, while accounting for endogeneity, the problem that countries experiencing more conflict also tend to receive more ODA, making standard correlations misleading. This long-established and widely used method in the social sciences is supplemented by a Causal Forest model. This moves beyond average effects, which are criticised for yielding conclusions that are too general to estimate heterogeneous treatment effects. These heterogeneous treatment effects identify the conditions under which different forms of ODA are associated with increases or decreases in casualties.

After combining and analysing over 5,000 country-year observations across 156 countries, 34 years, and 76 country-level indicators, both methods converge on the same overarching finding: ODA, on average, does not reliably reduce intrastate conflict deaths. The main predictor of conflict deaths in any given year is the level of deaths in the previous year, by far. This reflects the self-reinforcing nature of intrastate violence, as the literature calls it, the conflict trap. None of the nine forms of ODA consistently breaks this cycle. In the Panel GMM, only multilateral ODA reaches statistical significance among the ODA variables, but rather than being associated with fewer conflict deaths, it is associated with more.

The causal forest average treatment effects are small in magnitude and cluster near zero for most ODA types. These averages, however, conceal the heterogeneity. The newer Causal Forest technique identifies four main conditions that shape ODA's relationship with conflict. First, political stability is the most consistently important factor: below a threshold of approximately −0.2 on the World Bank's Political Stability and Absence of Violence/Terrorism index, ODA effects are mostly harmful and unpredictable, while above this threshold they become more muted and stable. This pattern is consistent with the mechanisms of predation and sabotage identified in the literature: in politically unstable contexts, armed groups are more likely to seize ODA resources or attack providers perceived as partisan. Second, while aid fragmentation is known to limit developmental returns, it appears to reduce conflict: countries receiving more, smaller projects experience more predictable and less harmful ODA effects, likely because dispersed aid is a less attractive target for predation or sabotage. This may be due to dispersed ODA presenting less attractive targets for predation than large, concentrated inflows. Third, the choice of delivery channel matters in fragile settings: bilateral ODA is associated with higher casualties in politically unstable countries, whereas multilateral ODA is not, possibly because multilateral channels are less susceptible to geopolitical perceptions and provide greater stability against sudden aid shocks. Fourth, the purpose of ODA matters: the impact of purpose-specific ODA is substantially larger than for aggregate ODA, though the results for ODA with the purposes of conflict, peace and security, and construction are inconclusive due to the reactive nature of their allocation, which makes causal identification difficult.

Five policy implications follow from these findings.
First, since ODA in high-conflict environments carries substantial risk of worsening violence through predation and sabotage, detailed local assessments should precede any disbursement decision, as country-level indicators alone are insufficient to determine whether aid is safe to deliver. These assessments should cover whether aid resources could be targeted for predation, whether delivery could be perceived as politically aligned with a conflict party, and whether the security situation permits effective oversight.
Second, since political stability is the most important moderator of ODA's effects, it should be treated as a precondition for scaling up bilateral ODA: for most ODA-recipient countries, the political conditions needed for bilateral ODA to be absorbed safely have not yet been met, meaning that increasing volumes prematurely risks fuelling rather than reducing conflict.
Third, since bilateral ODA is associated with increased casualties in politically unstable settings while multilateral ODA is not, aid in fragile contexts should be channelled through multilateral institutions, which are less susceptible to geopolitical perceptions and more resilient to aid shocks.
Fourth, since conflict is persistent and ODA rarely breaks this cycle once violence is entrenched, prevention should be prioritised over de-escalation.
Fifth, since these findings are based on country-level aggregates that cannot capture subnational variation, local power dynamics, or cultural factors, they should inform but not replace context-specific, project-level analysis.

The thesis also provides policy recommendations tailored to three different levels of political stability and violence. For stable countries, the priority of ODA should be development rather than minimising conflict. For moderately stable countries, policymakers should assess whether ODA is the right tool, thoroughly research local circumstances, distribute ODA evenly, and consider multilateral aid channels. For unstable countries, policymakers should prioritise policy alternatives. If ODA is still chosen, they must conduct thorough local research first, opt for multilateral ODA rather than bilateral, focus on small projects, and target regions that are already relatively stable.

This thesis contributes to the aid effectiveness literature by moving beyond typical methods that do not account for endogeneity or average-country effects, which draw conclusions for all countries. By combining an established econometric method with a modern statistical approach that combines machine learning and causal inference to reveal the conditions under which foreign ODA may mitigate or exacerbate intrastate conflict. Future research should move toward subnational and project-level analysis and be complemented by qualitative methods, as local power dynamics, cultural norms, and community-level perceptions cannot be measured.