Essays on program evaluation methods with applications on the economic costs of conflict

Abstract

This thesis studies the economic cost of armed conflict using treatment effect methodology. It initially focuses on obtaining an updated measure of the average effect of armed conflict. I first estimate the probability of a given country engaging in an armed conflict at each time period, i.e. the propensity score. I then use this propensity score estimation to estimate the average treatment effect on of conflict on those countries that experienced an armed conflict episode, by combining regression-adjustment and propensity score weighting tools. I find that the average annual effect of conflict is around 7% of real GDP per capita. This result seems robust to different model specifications. I also show that the model is an unrestricted version of the seminal model presented by Collier (1999).I continue by measuring the economic cost of conflict from another angle: an application of the synthetic control method (Abadie and Gardeazabal, 2003). I apply the synthetic control method to as many armed conflicts as possible (from the 1950s until 2010) in order to obtain a distribution of the effect of conflict both on real GDP per capita and Investment. This distribution is stratified by type of conflict, intensity, and continent. Results show that the effect of conflict is very heterogeneous even when accounting for those three characteristics, and that average effect measures are consequently not very representative of the whole issue. I also consider another program evaluation method, the panel data approach by Hsiao et al (2012), which can also be applied to measure the cost of a specific conflict; and see how it compares to the synthetic control approach. I begin this section with an empirical application, and confirm that each method gives a different measure of the effect of this intervention. In order to explain this difference, I then carry out a simulation. I find that the synthetic control method results in a higher Mean Squared Error and a lower bias than the panel data approach on average, though the results vary depending on the parameter considered. Furthermore, the estimated effect appears to be sensitive to a change in the donor pool in the case of the Panel Data approach. This paper also proposes a modification of the panel data approach method, allowing a more automatic selection of the controls used to build the counterfactual. I finalize the dissertation with a statistical software to aid other researchers in the application of the panel data approach.