Online Program

Return to main conference page

All Times ET

Program is Subject to Change

Thursday, June 17
Thu, Jun 17, 1:30 PM - 3:30 PM
TBD
Modernization Efforts in Establishment Statistics 2

Explicit Parameterizations for Linking Establishments and Entities under Sparse Classifications (308059)

Darcy Steeg Morris, Census Bureau Center for Statistical Research and Methodology 
*Yves Thibaudeau, Census Bureau Center for Statistical Research and Methodology 

Keywords: record linkage, log-linear models, zero cells

Positing log-linear models for linking and resolving entities is common (see Winkler 1988, Fellegi Sunter 1969). However, zero cells in the associated classification of the comparison patterns may reduce the number of dimensions of the parameter space and complicate the estimation of the model. This may occur when attempting to link establishments or other entities when the entity characteristics involved in the pattern comparisons are corolated and more finely parameterized models may be advantageous (Xu et al. 2019, Thibaudeau 1993). Faced with such difficulties, we propose parameterizations tailored for log-linear models in presence of likelihood zeros. These parameterizations and the associated treatment of the likelihood build on the proposals of Fienberg and Rinaldo (2012) who introduce "extended maximum likelihood estimators". Once adequately formulated these parameterizations can be easily estimated and lead to partial tests of model validity.