To take one example, they studied interventions aiming to increase the attendance of teachers in the developing world. They found that half of the interventions studied had no effect whatsoever. Even once those were excluded, the best three were over ten times more effective than the worst intervention.

Moreover, these differences are hard to predict ahead of time. Most social interventions that end up being evaluated were originally supported by experts and governments, were executed on a wide scale and are widely thought to work. But when tested rigorously, they turn out not to.