Abstract:
|
We explore different strategies for combining identifying assumptions to handle nonignorable missing data. In the first strategy, the main idea is to use a sequential identification procedure, whereby we specify potentially different missingness mechanisms for different blocks of variables. In its most general form, this strategy consists in expanding the observed-data distribution sequentially by identifying parts of the full-data distribution associated with blocks of variables, one block at a time. In the second procedure we exploit the availability of auxiliary marginal information on the distribution of some of the variables recorded in the sample, which allow us to identify missingness mechanisms where the nonresponse for these variables may directly depend on the value of the variables.
|