Comparing Estimation Methods for Categorical Marginal Models
Categorical marginal models are flexible models for modelling dependent or clustered categorical data which do not involve any specific assumptions about the nature of the dependencies. Categorical marginal models are used for different purposes, including hypothesis testing, assessing model fit, and regression problems. Two different estimation methods are used to estimate marginal models: maximum likelihood (ML) and generalized estimating equations (GEE). We explored three different cases to find out to what extent the two types of estimation methods are appropriate for investigating different types of research questions. The results suggest that ML may be preferred for assessing model fit because GEE has limited fit indices, whereas both methods can be used to assess the effect of independent factors in regression. Moreover, ML is asymptotically efficient, while GEE loses efficiency when the working correlation matrix is not correctly specified. However, for parameter estimation in regression GEE is easier to apply from a computational perspective.
The authors would like to thank Klaas Sijtsma for commenting on earlier versions of the manuscript.
- Bergsma WP, Van der Ark LA (2013) cmm: An R-package for categorical marginal models (Version 0.7) [computer software]. http://cran.r-project.org/web/packages/cmm/index.html. Accessed 24 Feb 2013
- Bergsma WP, Croon MA, Hagenaars JA (2009) Marginal models: for dependent, clustered, and longitudinal categorical data. Springer, New YorkGoogle Scholar
- Kesteloot H, Geboers J, Joossens JV (1989) On the within-population relationship between nutrition and serum lipids, the birnh study. Eur Heart J 10:196–202Google Scholar
- Lipsitz S, Fitzmaurice G (2009) Generalized estimating equations for longitudinal data analysis. In: Fitzmaurice G, Davidian M, Verbeke G, Molenberghs G (eds) Longitudinal data analysis. Chapman & Hall/CRC, Boca Raton, pp 43–78Google Scholar
- Nunnally JC (1978) Psychometric theory. McGraw-Hill, New YorkGoogle Scholar
- Pawitan Y (2001) In all likelihood: statistical modelling and inference using likelihood. Clarendon Press, OxfordGoogle Scholar
- Skrondal A, Rabe-Hesketh S (2004) Generalized latent variable modeling: multilevel, longitudinal, and structural equation models. Chapman & Hall/CRC, Boca RatonGoogle Scholar
- Van der Ark LA, Bergsma WP, Croon MA (2013) Augmented empirical likelihood estimation of categorical marginal models for large sparse contingency tables (under review)Google Scholar
- Van der Veen G (1992) Principes in praktijk: CNV-leden over collectieve acties [Principles into practice. Labour union members on means of political pressure]. J.H. Kok, KampenGoogle Scholar
- Yan J, Højsgaard S, Halekoh U (2012) geepack: Generalized Estimating Equation package. (Version 1.1-6) [computer software]. http://cran.r-project.org/web/packages/geepack/index.html. Accessed 13 Dec 2013