Nber Working Paper Series External and Internal Validity of a Geographic Quasi-experiment Embedded in Cluster-randomized Experiment

Abstract

from internal validity. The point estimates for school enrollment and work-in-home are attenuated relative to experiment 1. Descriptive statistics suggests that the sample in experiment 2 is better-off than that of experiment 1, given higher rates of electricity use, asset ownership, and other income proxies. This plausibly explains the attenuated effects, since the literature on conditional cash transfers finds smaller effects in when children are less poor (Fiszbein and Schady, 2009; Galiani and McEwan, 2013). Third, we assess internal validity by comparing the unbiased estimates from experiment 2 to those of the GQE (noting again that both include the same treatment group but different control groups). Particularly for school enrollment, the GQE estimates are attenuated relative those of experiment 2. It suggests that imbalance in unobserved variables results in downward biases in the GQE enrollment estimates. This is perhaps consistent with the higher proportion of indigenous children in the GQE treatment group, relative to its quasi-experimental control group. In summary, we find that the GQE cannot fully replicate the policy-relevant experimental benchmark in Galiani and McEwan (2013) for reasons related to both validity concerns. Based on these results, we make two concrete recommendations. First, it is essential that researchers using a geographic design carefully assess treatment-control balance on a wide range of observed covariates that are plausibly correlated with dependent variables (echoing the recommendations

24 Figures and Tables

Cite this paper

@inproceedings{Galiani2016NberWP, title={Nber Working Paper Series External and Internal Validity of a Geographic Quasi-experiment Embedded in Cluster-randomized Experiment}, author={Sebastian Galiani and Patrick J. McEwan and Brian Quistorff and Matias D. Cattaneo and J. Carlos Escanciano and Luke J. Keele}, year={2016} }