Latent Class Predictions for Subsequent Analysis

• Activity Number: 111
• Type: Contributed
• Date/Time: Monday, August 3, 2009 : 8:30 AM to 10:20 AM
• Sponsor: Biometrics Section
• Author(s): Janne Petersen*+ and Karen Bandeen-Roche and Klaus G. Larsen and Ove Andersen and Esben Budtz-Jørgensen
• Companies: Copenhagen University Hospital, Hvidovre and Johns Hopkins Bloomberg School of Public Health and Lundbeck and Copenhagen University Hospital, Hvidovre and University of Copenhagen
• Keywords: three-step analysis ; latent class analysis ; latent profile analysis ; latent-class assignment

Abstract

Mixture models are used to study syndromes because a single marker cannot characterize these. The model separates patients into homogenous groups, latent classes, based on a set of markers. Our aim is to predict the latent class variable for use in subsequent analysis. The prediction method should depend on whether the latent variable is an independent or a dependent variable. For the latter case, we have developed a new prediction method that unlike traditional methods yields consistent estimates of the covariate effects in a subsequent regression analysis. When the prediction is a covariate, we show that the posterior class probability has nice properties. Both of the recommended predictors are multidimensional, and they are developed both for discrete and continuous markers. The methods are illustrated in data on HIV-associated lipodystrophy syndrome, a syndrome of fat redistribution.