You are here
Imputation of Potential Outcomes to Estimate Treatment Effects of Small Groups
The German Federal Employment Agency is using a comprehensive evaluation system to measure the efficiency of the large variety of job-training programs. Because the assignment to any of the possible training programs, i.e., treatments, typically is not based on an explicit randomized process, corrections have to be done before drawing causal inferences. Basically, we try to answer the following type of question for each job-training program and each trained person: If a person had not been job trained, how much time would have passed until the person found employment? Based on the framework of Rubin’s Causal Model, we multiply impute these missing potential outcomes to get estimates of the individual causal effects. Since this is a massive imputation task, innovative and complex algorithms, as well as automation of the whole process, are required. In this talk, we describe our approach, the imputation routines, and present some results.
Authors of this paper are Susanne Rässler and Donald B. Rubin
Susanne Rässler, Prof. Dr. rer. pol., holds the Chair of Statistics and Econometrics at the Otto-Friedrich-University of Bamberg and is speaker of the methods group of NEPS (National Educational Panel Studies). Her special interests are methods for handling missing data, multiple imputation, Bayesian methods, and matching techniques for causal analysis as well as marketing research. She is associate editor of the Journal of Official Statistics (JOS) and the AStA - Wirtschafts- und Sozialstatistisches Archiv. She is working closely together with Prof. Donald B. Rubin from Harvard University.