Vocational training for demobilized ex-combatants with disabilities in Rwanda

ABSTRACT

Disability-inclusive development is receiving growing attention as a pressing international development issue. Disability-inclusive development is especially urgent and complicated in post-conflict countries. This paper examines the impacts of vocational training on economic empowerment and social reintegration among demobilised ex-combatants with disabilities in Rwanda. This is the first quasi-experimental study on vocational training for disabled ex-combatants. Exploiting the variation in the timing of training uptake within the same training course, we employ a pipeline approach in the following three steps: (1) trimming to guarantee common support within courses, (2) exact matching on key covariates within courses, and (3) regression controlling for covariates within courses based on the matched sample. The results show that the training greatly increased not only employment and earnings, but also trainees’ reintegration into the family and community. The results are robust to potential omitted variable bias and attrition bias according to a coefficient stability test and bound analysis, respectively. Our findings suggest a significant potential of vocational training for disabled ex-combatants in disarmament, demobilisation, and reintegration programmes. Our study exemplifies the utility of a credibly designed pipeline approach, which can be applied in a wide range of development projects in practice.

KEYWORDS:

• Disability
• ex-combatants
• vocational training
• pipeline approach
• Rwanda

Disclosure statement

Kana Takio was an evaluation officer at Japan International Cooperation Agency (JICA). Kengo Igei and Keitaro Aoyagi were paid as an evaluation consultant and Yoshito Takasaki was paid as a technical advisor to the Evaluation Department at JICA. All four of us were paid to conduct a neutral and scientific evaluation of the project studied in this paper, “ Skills Training for the Reintegration of Demobilized Soldiers with Disabilities.“ While JICA had the right to review the manuscript, JICA did not have the right to request changes in the analysis. We have no other interests to disclose.

Notes

1. The project targeted ex-combatants with disabilities regardless of their army group––RDF, the Forces Armées Rwandaises (the former government army), and private armed groups––formed according to ethnicity. With no data about social cohesion across ethnic groups, we cannot examine the training impacts on ethnic reintegration.

2. The project survey data were supplemented by the data of demobilised ex-combatants from the RDRC.

3. The remaining 295 enrolled trainees graduated before the baseline survey or after the follow-up survey, if they graduated. We exclude the former trainees from the enrolment sample because (1) we cannot construct their corresponding comparison group, or (2) even if we can do so for some of them, they are not in the batch consecutive to the comparison group and thus are trimmed from the sample as discussed below. The trainees who graduated after the follow-up survey were not covered by the baseline survey.

4. Other reasons for attrition in the treatment group include dropout and death. Attrition in the comparison group was only due to refusal of the survey.

5. For education level and type of disability, we consider missingness as a separate category.

6. In the project, all trainees but one in the comparison group were male. This unique female trainee remains in the trimmed sample, but not in the matched sample.

7. The weight for the treatment group is equal to one, and the weight for the comparison group is equal to $m_C/m_T\times m_T^s/m_C^s$, where $m_T$ and $m_C$ are the total number of matched observations of the treatment and comparison groups, respectively, and $m_T^s$ and $m_C^s$ are the number of matched observations of the treatment and comparison groups, respectively, in stratum s.

8. These 11 types of courses consist of brick/block laying and concreting at three centres; welding at two centres; plumbing and pipe fitting at two centres; and electronics, cooking, carpentry and joinery, and tailoring at one centre each (Appendix, Table A1). The remaining 16 types include courses other than these 11 types, such as computing, shoemaking, and leather craft. The location of the four training centres in the matched sample is different from the remaining six centres. In particular, whereas none of the former four centres are located in Kigali, the national capital, four out of the latter six centres are located there.

9. When computing standard errors, we adopt Webb’s (2014) six-point distribution as the auxiliary distribution and cluster bootstrap errors by training centre, which is a cluster with the smallest number, following MacKinnon, Nielsen, and Webb (2020). We do 1000 bootstrap replications.

10. Marital status includes a category for missingness. We combine Forces Armées Rwandaises and armed group into one group due to the small number of observations of the latter. Exact matching based on this dummy variable yields the same matching results.

11. This correction has an advantage over the conventional Bonferroni or Holm correction in its ability to correctly reject a false null hypothesis because it allows the dependence among the p-values or the test statistics from multiple hypothesis testing with resampling and step-down procedure.

12. This is based on the matched sample of 80 observations, which is slightly different from the matched sample of 74 observations with nonmissing employment type among the employed, based on which the histogram of employment status (panel A, Figure 2) was generated.

13. With the small number of observations for each course (Appendix, Table A1), it is infeasible to examine potentially heterogeneous impacts across the courses. For most courses, the treatment means of most outcomes are greater than the corresponding comparison means.

14. Oster (2019) suggests $R_{max}=1.3\tilde{R}$ as a lower bound as well as $\delta$=1, which indicates that the degree of selection on unobservables is no greater than that on observables. In our setting, this assumption seems to be appropriate because the exact matching should limit the difference in unobservables between the treatment and comparison groups. We also consider $\delta$=−1, corresponding to the case where the direction of selection on unobservables is opposite to that on observables.

15. The trimming proportion is defined as $q=\left(q_C-q_T\right)/q_C$, where $q_C$ and $q_T$ are the rates of nonattrited trainees in the comparison and treatment groups, respectively.

16. The number of observations for the dummy for being treated preferably by family is slightly smaller than that in the original analysis because a few strata that lose common support due to missing values in this outcome were dropped. For all outcomes, all attritors are in the treatment group.

Additional information

Funding

This work was supported by the JSPS [25257106,19H00590]; the University of Tokyo.

Notes on contributors

Kengo Igei

Kengo Igei is a Project Researcher in the Graduate School of Economics at the University of Tokyo. He also works as a consultant in the field of impact evaluation and evidence-informed decision making in the public policy sector at Metrics Work Consultants Inc. His research focuses on disability and development.

Kana Takio

Kana Takio is a former master's student at Harvard T. H. Chan School of Public Health. Her research addressed health systems and policy in East Africa. She now works as a consultant in the field of public-private partnership for global health.

Keitaro Aoyagi

Keitaro Aoyagi is a Representative Director at Metrics Work Consultants Inc. He specializes in program evaluation and evidence-based policy making, especially in the field of international development.

Yoshito Takasaki

Yoshito Takasaki is a Professor in the Graduate School of Economics at the University of Tokyo. His research interests include poverty, disability, conflict, and labor.