Tuning the EM‐test for finite mixture models

There has been rapid progress in developing effective and easy-to-use tests of the order of a finite mixture model. The EM-test is the latest to join the rank. It has a relatively simple limiting distribution and enjoys broad applicability. Based on asymptotic theory, the P-value of the EM-test is approximated via its limiting distribution. The built-in tuning parameter has an important influence on the approximation precision. Thus, choosing an appropriate value for this parameter is important for fully realizing the advantages of the EM-test. In this article, we develop a novel computer-experiment approach to address this issue. Through designed experiments, we derive a number of empirical formulas for the tuning parameter. Extensive validation simulation shows that these formulas work well in terms of providing accurate type I errors. The Canadian Journal of Statistics 39: 389–404; 2011 © 2011 Statistical Society of Canada Il y a eu de rapides progres dans le developpement de tests efficaces et faciles a utiliser pour l'ordre d'un modele de melange fini. Le test EM est le dernier de cette serie. Sa distribution limite est relativement simple et il beneficie d'une grande applicabilite. En se basant sur la theorie asymptotique, le seuil du test EM est approxime en utilisant sa distribution limite. Le parametre de reglage incorpore a une grande importance sur la precision de l'approximation. Ainsi, choisir une valeur appropriee pour ce parametre est important pour tirer pleinement avantage du test EM. Dans cet article, nous developpons une nouvelle approche utilisant l'experimentation par ordinateur pour aborder cette question. A l'aide d'experiences planifiees, nous obtenons certaines equations empiriques pour le parametre de reglage. Une simulation de validation de grande envergure montre que ces equations permettent bien d'atteindre les erreurs de type I de facon precise. La revue canadienne de statistique 39:389–404;2011 © 2011 Societe statistique du Canada

[1]  Pengfei Li,et al.  Testing the Order of a Finite Mixture , 2010 .

[2]  G. Schwarz Estimating the Dimension of a Model , 1978 .

[3]  J. Kalbfleisch,et al.  A modified likelihood ratio test for homogeneity in finite mixture models , 2001 .

[4]  E. Lehmann Testing Statistical Hypotheses , 1960 .

[5]  Jiahua Chen,et al.  The likelihood ratio test for homogeneity in finite mixture models , 2001 .

[6]  Jiayang Sun,et al.  Testing Homogeneity in a Mixture Distribution via the L2 Distance Between Competing Models , 2004 .

[7]  Y. Shao,et al.  Asymptotics for likelihood ratio tests under loss of identifiability , 2003 .

[8]  Pengfei Li,et al.  Non-finite Fisher information and homogeneity: an EM approach , 2009 .

[9]  Jiayang Sun,et al.  Testing homogeneity in discrete mixtures , 2008 .

[10]  John D. Kalbfleisch,et al.  Testing for a finite mixture model with two components , 2004 .

[11]  Bernard Garel,et al.  Asymptotic theory of the likelihood ratio test for the identification of a mixture , 2005 .

[12]  H. Akaike,et al.  Information Theory and an Extension of the Maximum Likelihood Principle , 1973 .

[13]  E. Gassiat,et al.  Testing the order of a model using locally conic parametrization : population mixtures and stationary ARMA processes , 1999 .

[14]  Jiahua Chen,et al.  Hypothesis test for normal mixture models: The EM approach , 2009, 0908.3428.

[15]  Jiahua Chen Penalized likelihood‐ratio test for finite mixture models with multinomial observations , 1998 .

[16]  Eric S. Lander,et al.  Asymptotic distribution of the likelihood ratio test that a mixture of two binomials is a single binomial , 1995 .