Modelos quase-experimentais na avaliação de impactos econômicos de desastres sociotécnicos

Abstract

Causal inference aims to assess the effect of some potential cause on some outcome. In this sense, the quasi-experimental models Propensity Score Matching (PSM) and synthetic control stand out, which are non-experimental evaluation techniques. These quasi-experimental models use information from a set of units that did not participate in a treatment to identify what would have happened to the participating units if they had not participated in the intervention. Thus, in this work, quasi-experimental models, PSM and synthetic control were applied, with the objective of evaluating the economic impact of the Fundão dam failure in the region of Mariana-MG. These models were also applied using cluster analysis to use control municipalities similar to the treated units and analyze the differences in the results obtained. The analyzed data were taken from different sources and are socioeconomic indicators that characterize the municipalities affected by the Fundão dam failure. The results obtained by the PSM, to assess any effect on the real GDP per capita of the 35 impacted municipalities, did not reveal any significant average treatment effect in the years 2016, 2017 and 2018. When analyzing the results obtained in the PSM with the cluster analysis, it was found a greater difference between the means of the treated and control groups, but it was also not significant at the 5% significance level. The synthetic control was also used to evaluate the impact of the Fundão dam failure on the real GDP per capita in the municipality of Mariana, in which the synthetic trajectories of the response variable were traced with and without cluster analysis. The results obtained in the synthetic control, without cluster analysis, did not prove to be a good counterfactual for the treated unit, municipality of Mariana. The synthetic control without the grouping considered all the municipalities in Minas Gerais that were not impacted by the socio-technical disaster as potential controls and this, in addition to affecting the estimates, led to a computational time of more than 6 hours. When considering cluster analysis in the synthetic control, the counterfactual obtained had a satisfactory fit and reduced computational time by more than 600 times.