Some early references are Fay (1986 JASA), Baker & Laird (1988 JASA) and Park & Brown (1994 JASA). See also chapter 11 of my book with Rubin. There is a limited discussion of PM models for categorical data in Little (1993 JASA). There are a number of papers on nonignorable models for longitudinal categorical data, e.g. Molenberghs, Kenward & Lesaffre (1997 Biometrika).
In one sense there is no difference between PM and selection models in the categorical setting: if you write a loglinear model for the joint distribution of Y (the outcomes) and M (the md pattern), then this implies both selection and pattern-mixture models, depending on which way you factor, that is Y and M given Y (selection) or M and Y given M (PM). This arises from the "nonparametric" character of the multinomial distribution and is in contrast to (say) normal models, where the factorization yield different distributional assumptions. The main issue with either factorization is how to deal with the lack of identifiability, that is, what assumptions one is willing to make to identify the model. Rod Little On Thu, 14 Dec 2000 [email protected] wrote: > What is the best appoach to deal with categorical data subject to > nonignorable nonresponse: selection models or mixture models? If there is > no approach, how can be determined which model is the best approach to the > data? > > > ___________________________________________________________________________________ Roderick Little Chair, Department of Biostatistics (734) 936-1003 U-M School of Public Health Fax: (734) 763-2215 M4208 SPH II [email protected] 1420 Washington Hgts http://www.sph.umich.edu/~rlittle/ Ann Arbor, MI 48109-2029
