How Useful are Uncertainty Bounds? Some Recent Theory With an Application to Rubin's Causal Model

Abstract

Techniques of data fusion are applied to merge multiple data sets that do not contain exactly the same variables, to a thorough data set. The common distributions and correlations of variables that were not observed simultaneously can not be identified unambiguously. Traditional matching algorithms are based on the assumption of conditional independence of non-simultaneously observed variables. As this assumption can not be tested without additional information, we suggest taking into account this deficit of information in the matching algorithm. For this purpose, various datasets with different correlation structures are generated by regression-based algorithms. Subsequently, any desired statistic analysis can be carried out, with the robustness of results being testable against different correlation structures. The procedure is also appropriate for impact evaluations of political interventions, as long as not only mean treatment effects are to be estimated. (IAB)

Cite article

Rässler, S. & Kiesl, H. (2009): How useful are uncertainty bounds? Some recent theory with an application to Rubin's causal model. In: International Statistical Institute (2009): Proceedings of the ISI 2009 (CD-Rom), p. 1-17.