Treatment effect estimation under data from complex sample surveys.

Abstract

In public policy evaluation two populations ( a treatment group and a control) are compared in terms of an intervention e ect. Those interventions can be government aids, grants, subsidy, a training program. The goal of public policy evaluation (PPE) is to determine the magnitude of the impact of the intervention, in order to take decisions about the design of the program, to evaluate the program e ciency in terms of cost-bene t and the ful lment of the program objectives. Program evaluation typically based on process evaluation and outcome evaluation, rst case o er possibility to assess if the public policy as implemented as planned, validating whether objectives were met, target population was blanketed with the program and determine success of the program. Other one diagnose whether program e ect generate a change in population behaviour, Impact revealed by PPE call of special attention since resource allocation may be a ected directly as mentioned (?) doing that great programs receive a better resources contribution meanwhile bad projects could be deleted or corrected. On the case, indirectly resource allocation not always due to technical criteria, refer to political interests. As a result, PPE brings an important feature released with counter factual measuring comparing observe program results in cases where program is present and the results would have been observed in their absence. Traditionally the surveys carry on in public policy evaluation studies come from a complex surveys such as strati ed and multi-stage designs, nevertheless, practitioners usually ignores the complex survey design in statistical analysis process. That means that the sample weights are ignored and the conclusions from this studies can not be extend to the population. In this work, we will use the proper estimators for the parameters and the variance in order to extrapolate our results to the population and to get a uncertainty measures with respect to the estimator accurate. Unbiased estimators will be implement in this work. Under simulation scenarios we will show the consequences of ignoring the complex survey design in terms of the estimator bias. A design e ect (de ) approach will be used to determine sample size for di erent scenarios, the power of a test approach to compute sample size will describe in this work. A comparison between di erent sampling design will be show (classic inference approach, simple random sample without replacement, simple random strati ed sample and two-stage sample.