Embedding a Nonparametric Weighting Scheme in Latent Class Regression Procedure to Evaluate Risk Factors for Multiple Pathogens of Diseases

• Activity Number: 315
• Type: Contributed
• Date/Time: Tuesday, August 11, 2015 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Nong Shang*
• Companies: CDC
• Keywords: etiology ; latent class ; risk factor ; kernel density ; imputation ; Gibbs sampling

Abstract

Latent class analysis approach has been applied to estimate prevalence rate of a specific pathogen based on a few laboratory tests for the same pathogen. However, the problem becomes much more complicated when the whole spectrum of the etiology (with often more than 20 pathogens) of a disease needs to be estimated, with only one single test for each of the pathogens. Further, the etiology prevalence as well as the background carriage rates of the pathogens might be associated with some risk factors. Regular latent class regression approach would fail for our problem as the number of different combinations of testing results often barely surpass the number of parameters in the model. We proposed an approach that embedded a non-parametric weighting scheme to weight imputed latent classes into the regular Gibbs procedure. The parameters of Dirichlet (for pathogen etiology prevalence) and Beta (for background carriage rates) distributions at each data point in the domain of the risk factors are estimated by the weights. The approach does not only produce robust estimations of pathogen prevalence rates, but also provides graphical tools to examine risk factors from various perspective.