Stuck in a Fragility Trap: The Case of the Central African Republic Civil War

ABSTRACT

This study utilizes the synthetic control method to assess the economic consequences of the ongoing civil war in the Central African Republic since December 2012. Drawing on a donor pool of low-income and lower-middle-income countries, it constructs a synthetic counterfactual to depict the economic trajectory in the absence of conflict. The analysis reveals a significant decline in national gross domestic product (GDP) per capita, estimated between 45.3 percent and 47.8 percent over a decade of conflict, resulting in a cumulative GDP loss of US$29.7 billion to US$32.4 billion (purchasing power parity, PPP, adjusted). Two model specifications are employed, one using pre-treatment outcomes and the other integrating external covariates. Robustness checks support the findings, indicating a minimum 10-year decline of 35.3 percent in GDP per capita. Even considering the 2003 coup, this civil war has the most detrimental economic impact. The analysis remains robust when incorporating GDP data from remote sensing sources. These effects align with the fragility trap concept, portraying one of the highest economic impacts of civil conflict in terms of relative GDP per capita decline.

KEYWORDS:

• Central African Republic
• civil war
• fragility trap
• synthetic control method

JEL CLASSIFICATION:

• C31
• D74
• O11

Disclosure Statement

No potential conflict of interest was reported by the author(s).

Notes

1. We extend our sincere gratitude to Editor Dr. Khusrav Gaibulloev and the two anonymous referees for their invaluable comments and feedback, which greatly enhanced the manuscript both technically and theoretically. We would also like to express our warm thanks to Prof. Cruz Echevarría and Prof. Rok Spruk for generously sharing their code and offering insightful clarifications on the technical implementation of robustness checks methods related to the Synthetic Control Method. We use the swift_red scheme from schemepack package for charts on Stata designed by Asjad Naqvi (Naqvi 2022).

2. There were also the monetary shocks of the CFA franc, culminating the devaluation of 1994.

3. The United Nations Multidimensional Integrated Stabilization Mission in the Central African Republic (MINUSCA in French acronym) which started on April 10, 2014, to protect Central African Republic civilians under Chapter VII of the UN Charter. Since the end of September 2021, it has more than 15,000 troops, police, and civilian personnel on the ground, with Rwanda contributing the largest contingent of troops and police forces.

4. While officially contested by Russian officials, the economic, military, and political presence of the Wagner Group is extensively documented in the literature (e.g. Stanyard, Vircoulon, and Rademeyer 2023). Less documented and publicized is the presence of significant bilateral Rwandan troops not restricted by UN rules of engagement, as part of a broader economic and diplomatic partnership (International Crisis Group 2023). In both cases, it is alleged that CAR grants diplomatic support, forestry, mining concessions, and land for agricultural projects in exchange for security services, amid severe fiscal pressures (Doboš and Purton 2024; International Crisis Group 2023).

5. Available here: https://acleddata.com/.

6. As recalled in the Country Economic Memorandum (World Bank 2022, 20): ‘Rebels launched an unsuccessful coup attempt in May 2001 by storming strategic buildings in the capital city of Bangui. General Francois Bozizé, Patassé’s chief of the army and suspected to be behind the coup, was forced to flee the country and seek asylum in Chad. In March 2003, Patassé was overthrown by a surprise attack headed by Francois Bozizé, who suspended the constitution, dissolved the government, and organized new elections. […] Bozizé was elected president in May 2005, despite a fragile political context and disputes over the control of power, including several rebel attacks against the army. […] [M]ost of the reforms agreed upon […] were not implemented, and the inclusive dialogue was never established, leading to a rise in the number of zones controlled by armed groups. Moreover, the implementation of the demobilization, disarmament, and reintegration program for combatants in the northeast failed, with significant adverse effects on security, poverty, and the humanitarian situation.’

7. Appendix A to this paper provides more background and context of the CAR’s civil war.

8. The SCM has also been employed to study the economic impacts of natural disasters (Cavallo et al. 2013) and organized crime (Pinotti 2015), for instance.

9. The working paper version of this article included a dashboard of socio-economic indicators (i.e. the local level of NTLs, the value added of industry and manufacturing activities, and the human asset index), which are now removed from the analysis for at least three reasons: (i) this choice could be considered non-exhaustive and subjective to cherry-picking, (ii) some of these indicators could are now considered as external covariates when assessing the SCM approach on GDP, and (iii) the current version of the paper focuses on a significant number of relevant robustness checks suggested by peer-reviewers on the baseline specification (GDP), to ensure the strength of a genuine detrimental impact of CAR’s civil war.

10. More details in Methodology and identification strategy section.

11. The classification is directly available in the Online Appendix of Venables (2016) accessible on the website of the American Economic Association (AEA): https://www.aeaweb.org/articles?id=10.1257/jep.30.1.161.

12. We extend our gratitude to the two anonymous referees for their valuable suggestions regarding the robustness checks, which have enhanced the strength of our findings. The final two series of robustness checks are detailed in Appendix B and C of this paper.

13. Excluding CAR itself from the list of LICs.

14. The income classification and methodology are available online: https://blogs.worldbank.org/en/opendata/new-world-bank-group-country-classifications-income-level-fy24.

15. https://acleddata.com/analysis/.

16. See Tables A1 and A2, for more details.

17. In both cases, the nested option is used with the synth package on Stata. When used, the SCM embarks on a fully nested optimization procedure that searches among all (diagonal) positive semidefinite V-matrices and sets of W-weights for the best fitting convex combination of the control units.

18. For more details on how optimal weights ($w_r^{\ast }$) are solved, and the key role of the matrix V, see Echevarría and García-Enríquez (2019, 594-596) and Appendix B.

19. This indicator is used, for instance, in Matta, Appleton, and Bleaney (2019).

20. https://zenodo.org/records/7898409.

21. As expected, with the minor exception of Tajikistan, 99 percent of the synthetic is best reconstructed as a combination of LICs from sub-Saharan Africa (SSA). Although LICs are frequently affected, directly or indirectly, by armed conflicts of various magnitudes (Laville 2019; Laville and Mandon 2023), which may suggest a possible violation of the stable unit treatment value assumption (SUTVA) in case of conflicts’ spillovers, CAR lags significantly behind LMICs even before the war, rendering their inclusion in the synthetic control almost impossible. However, the qualitative results (i.e. the significant economic disaster involved by the CAR civil war) appear not sensitive to the exclusion of any individual country to the donor pool. Specifically, the exclusion of the Democratic Republic of Congo (DRC), a neighboring country to CAR, from the synthetic control, does not change our main insight, i.e. even in the most conservative specification, the relative magnitude of the impact of the civil war as a percentage of per capita GDP is found to be one of the highest in the literature. See the rest of the section for more discussion.

22. For a formal presentation of the quasi p-values see Cavallo et al. (2013) and Kešeljević and Spruk (2023b). We extend our gratitude to the first anonymous referee for recommending the use of quasi p-values and to Rok Spruk for verifying the code necessary to implement them through the synth_runner package on STATA proposed by Galiani and Quistroff (2017).

23. As in Kešeljević and Spruk (2023b), we drop, from the donor pool, the countries with pre-`treatment’ root mean squared prediction error (RMSPE) at least twice larger as CAR with the ‘pre_limit_mult(2)’ option with the synth_runner package.

24. We extend our gratitude to the first anonymous referee for recommending the use of in-time placebo robustness checks and to Rok Spruk for verifying the code necessary to implement them through the zandrews package on Stata.

25. This finding is comparatively more fragile than our baseline results as it heavily relies on the composition of the synthetic control (right panel of Figure 6); when keeping the baseline composition of the synthetic control (Table 1), a notable gap between the actual GDP per capita series and the synthetic one only emerges from 2013 onwards, i.e. after the civil war onset.

26. When comparing the baseline synthetic control (with the `treatment’ year in 2012) to the in-time synthetic control (with the `treatment’ year in 2003), we find that 35.7 percent of the gap in GDP per capita in 2013 is attributable to the 2003 coup. This represents a loss of US$286.2 per capita out of a total loss of US$801.4 per capita when accounting for both the coup and the onset of the civil war.

27. We extend our gratitude the first anonymous referee for recommending the use of leave-one out SCM and sparse SCM robustness checks.

28. Part of the explanation could be driven by the relatively strong growth experienced in Niger in 2022, attributed to a favorable agricultural season, according to the World Bank (https://www.worldbank.org/en/news/press-release/2023/06/19/niger-strong-agricultural-season-boosts-economic-rebound-in-2022). It could be also explained by the indirect exclusion of Ethiopia from this new synthetic control.

29. https://data.worldbank.org/indicator/NY.GDP.MKTP.KD.ZG?locations=ET (consulted on May 5, 2024). The results when removing Ethiopia alone are available upon direct request.

30. Specifically, Martínez (2022) has found that authoritarian regimes are prone to overstating yearly self-reported GDP growth due to a lack of effective checks and balances, and using remote sensing data (and more precisely NTLs) can help assess the reality of economic activity. Although GDP series are verified by international institutions, it is possible that the relatively robust economic performance observed under former President François Bozizé was overestimated. Specifically, GDP per capita grew by 25.5 percent under his reign, from USD963 per capita in 2003 to USD1,209 per capita in 2012. At the same time, the country was under scrutiny for moving towards authoritarianism, during this period, as indicated by its Polity2 score of −1 at the time, corresponding to a closed anocracy (Center for Systematic Peace 2022). Consequently, the subsequent drastic correction of GDP from 2013 could be partly artificial.

31. The variation in the composition of the donor pool can be obviously explained by a change of the outcome variable considered, but also by the change of the new pre-`treatment’ period used to fit the SCM (i.e. 2000-12).

32. In 2012, the CAR GDP per capita level was approximately US$1,209, in PPP, constant 2017 international USD, according to the WDI, and US$726, in PPP, constant 2005 international USD, according to Wang and Sun (2022). The disparity may arise from differences in the reference year used for computing constant PPP values, and/or mismeasurements in one or both data series.

33. Results are available upon direct request.

34. See the Methodology and Data sections for more details. To recap, we choose the specification from Matta, Appleton, and Bleaney (2019) mostly for two reasons. First, besides the average pre-`treatment’ outcomes, the external covariates can be easily clustered as three categories of predictors of the GDP dynamics, namely (i) demand aggregates (i.e. total consumption, GCF, exports, and imports) as a percentage of GDP; (ii) sectoral value added (i.e. agriculture, industry, services) as a percentage of GDP; and (iii) education and health components of the human capital (i.e. life expectancy at birth, and gross secondary school enrollment rate). Second, all these covariates comes from the same data source as the main outcome of interest, i.e. the WDI (World Bank 2024).

35. With the minor exception of Cote d’Ivoire, 97.8 percent of the synthetic is best reconstructed as a combination of LICs from SSA.

36. See Appendix B for a discussion related to the customization of the supply of the V-matrix. We extend our gratitude the first anonymous referee for suggesting the discussion on the V-matrix, as in Echevarría and García-Enríquez (2019), for instance.

37. When comparing the baseline synthetic control (with the `treatment’ year in 2012) to the in-time synthetic control (with the `treatment’ year in 2003), we find that 32.5 percent of the gap in GDP per capita in 2013 is attributable to the 2003 coup. This represents a loss of US$283.6 per capita out of a total loss of US$874.1 per capita when accounting for both the coup and the onset of the civil war.

38. Kešeljević and Spruk (2023a) also highlight that human development, particularly life expectancy at birth and infant mortality, as well as population density and the nation’s institutional quality framework, experienced substantial deterioration in Syria, from 2011-19, and are likely yet to fully recover.

39. Results are available upon direct request.

40. As noticed by Costalli, Moretti, and Pischedda (2017) there are remarkable variations in the economic impacts of civil wars across cases, so the average effect conceals a considerable dispersion.

41. It should be noted that Lopez and Wodon (2005) did not employ the SCM but instead relied on a methodology for identifying and correcting outliers in time series data. Therefore, their results are not directly comparable with ours. Furthermore, the country had already made significant progress in economic terms and human development, such as improvements in primary school enrollment and reductions in infant mortality just a decade after the genocide.

42. https://pulitzercenter.org/stories/how-worlds-deadliest-crises-go-unseen.

43. For more details, see Echevarría and García-Enríquez (2019, 594–596).

44. The income classification and methodology are available online: https://blogs.worldbank.org/en/opendata/new-world-bank-group-country-classifications-income-level-fy24.

45. The six LICs are Burkina Faso, Madagascar, Malawi, Mozambique, Tanzania, and Togo.

46. The classification is directly available in the Online Appendix of Venables (2016) accessible on the website of the American Economic Association (AEA): https://www.aeaweb.org/articles?id=10.1257/jep.30.1.161. According to Venables (2016), countries like CAR and Mozambique are considered ‘prospective natural resource exporting countries’ because they do not fully exploit the economic potential of their natural resources.