Hoonhout, Pierre: Identification and EM-Estimation of Panel Data Models with Non-Ignorable Attrition and Refreshment Samples
World Conference Econometric Society, 2000, Seattle

Pierre Hoonhout, Vrije Universiteit Amsterdam
Geert Ridder, Johns Hopkins University
Identification and EM-Estimation of Panel Data Models with Non-Ignorable Attrition and Refreshment Samples
Session: C-13-20  Wednesday 16 August 2000  by Hoonhout, Pierre
The benefits of panel data are well-documented but missing data problems are often more severe. In particular, units that respond in the first wave may drop out of the panel after one or more periods of participation. This paper focuses on identification and Maximum Likelihood estimation of panel data models when the process that governs this so-called attrition is possibly non-ignorable. In that case, conventional estimation procedures are inconsistent. We derive a multi-period nonparametric identification result and propose estimation by an EM-algorithm that exploits the availability of refreshment samples, consisting of new units randomly drawn from the original population. This additional data source reduces the informational incompleteness of the unbalanced panel in case of non-ignorable attrition. The algorithm is stated in terms of a general population model. Issues related to specific standard panel data models are discussed seperately. Problems caused by partially observed time-varying covariates are addressed along the way.
Submitted paper full-text in .pdf


File created by Jurgen Doornik with eswc2000.ox on 2-01-2001