Using randomised experiments and structural models for 'scaling up': evidence from the PROGRESA evaluation

The evaluation of welfare programs and more generally government or international organisma interventions is often posed as a one off question, in that evaluators ask whether a specific intervention achieves a specific objective in a specific situation. However, recently, the more general questions of whether results from a given study can be used to predict the the effect of different interventions in, possibly, different contexts has received a considerable amount of attention. The usefulness of such an exercise, if successful, is obvious. The ability to extrapolate success stories and avoid failures in different situation would obviously be highly desirable. Unfortunately a rigorous and successful extrapolation is extremely difficult. In this paper we discuss the issues involved with the evaluation of social interventions and with the attempts at 'scaling them up'. In particular we discuss the realtive merits of non-parametric evaluation strategies that rely on (possibly experimental) exogenous variation to estimate the impact effects and of more structural approaches. The difference between the two appraoches is particularly relevant when one comes to the issue of 'extrapolation' and 'scaling up'. One could consider two types of extrapolation: (i) Predict the effects of a program that is different from the one that was evaluated; (ii) predict the effect of exporting an existing program from a context where is was evaluated to a different one. In this paper we focus on the latter problem. After discussing the conceptual and technical issues, we apply the ideas we discuss to the results from the evluation of PROGRESA, a large welfare program in Mexico, for which a randomized evaluation sample is available and has been extensively studied. In particular, we divide the seven Mexican states in which the evaluation was carried out into two groups and check to what extent the results in one group can be extrapolated to the others.