563 – Multiple Imputation for Measurement Errors and Other Structured Patterns of Missing Data
Multiple Imputation in Data Fusion: Making Better Assumptions Than Conditional Independence
Volker Bosch
GfK SE
Philipp Gaffert
GfK SE
The missing pattern of data fusion implies that the variables that are specific to the data sets are never jointly observed. When applying standard imputation techniques, independence conditioned on the common variables is implicitly assumed. In general, however, this assumption does not hold; consequently, the estimated correlations between the fused specific variables are usually biased toward zero. We argue that in the absence of further information, a correlation lying well within the bounds of the conditional independence assumption (CIA) and one specific measurement error model is a significantly more sensible assumption. This argument is derived from a simple trivariate model and empirically supported by data from various fields.