Randomized interventions and "real" treatment effects : A cautionary tale and an example

Abstract The experimental approach has revolutionized development economics. Nonetheless, randomization cannot do everything. We discuss challenges to RCTs, paying special attention to internal validity. Randomized interventions in social sciences are not double-blind and do not, in general, hold all relevant covariates constant. Treated and untreated subjects adjust their behavior in response to treatment status. Disentangling the treatment effect into its behavioral component and the direct effect of the intervention is difficult, and implies a return to the toolkit of observational studies. This is illustrated using improved seed distribution in African farming. While standard RCTs found large treatment effects, double-blind RCTs revealed that a large share of this impact is due to farmers allocating extra effort and their best plots to the cultivation of new seeds.