Abstract:
|
In statistical data fusion, ignorability and independence assumptions are often assumed to ensure identifiability since there inevitably exists potential outcomes. However, these assumptions do not necessarily hold in real data analysis. Therefore, to model a more precise relationship between observed and unobserved variables, we need to consider nonignorable missing mechanisms, which often cannot be tested by original data alone. Hirano et al. (2001, Econometrica) provided sufficient conditions for the identifiability in a two-period panel study where there is nonignorable attrition, while additional moment information is available. Focusing on data fusion, we develop their results to relax ignorability and independence assumptions, and provide additional conditions for identifiability. We also discuss estimation methods. Our approach can not only contribute to theoretical understandings of this field but also be helpful for practitioners who want to combine data sets so that a moment of combined data corresponds to a target population.
|