Sample Attrition Bias in Randomized Experiments: A Tale of Two Surveys

The randomized trial literature has helped to renew the fields of microeconometric policy evaluation by emphasizing identification issues raised by endogenous program participation. Measurement and attrition issues have perhaps received less attention. This paper analyzes the dramatic impact of sample attrition in a large job search experiment. We take advantage of two independent surveys on the same initial sample of 8, 000 persons. The first one is a long telephone survey that had a strikingly low and unbalanced response rate of about 50%. The second one is a combination of administrative data and a short telephone survey targeted at those leaving the unemployment registers; this enriched data source has a balanced and much higher response rate (about 80%). With naive estimates that neglect non responses, these two sources yield puzzlingly different results. Using the enriched administrative data as benchmark, we find evidence that estimates from the long telephone survey lack external and internal validity. We turn to existing methods to bound the effects in the presence of sample selection; we extend them to the context of randomization with imperfect compliance. The bounds obtained from the two surveys are compatible but those from the long telephone survey are somewhat uninformative. We conclude on the consequences for data collection strategies.