R Govys April: Predictive Latent-Class Modeling and Imputation with bigLC

Joseph L. Schafer, Senior Mathematical Statistician for Analytic Modeling., U.S. Census Bureau will present on bigLC a new R package for describing multivariate data through latent-class (LC) models with an emphasis on prediction and imputation. Unlike traditional LC analysis, the classes are not intended to tell a story about real-world phenotypes; rather, their main purpose is to make the model flexible enough to detect and conform to unforeseen features in the data (bimodal distributions, nonlinear relationships, interactions, etc.) without extra input by the user. Items in a bigLC model may be categorical, continuous, semicontinuous, or integer counts, and the data may have arbitrary rates and patterns of missing values. bigLC also accommodates multilevel structures where observational units are nested within larger units (e.g., persons within households), and missing values may be present at any level. The user selects a distributional family for each item, sets the maximum number of classes for each level, and proceeds with the model fit. Fitting methods include an EM algorithm for maximum-likelihood estimation and a Bayesian Markov chain Monte Carlo procedures. Utilities are provided for prediction, imputation, and synthetic data generation. The package was presented at R Govys previously and has been expanded and enhanced, based on comments from this group.

Register: https://amstat.zoom.us/webinar/register/WN_lsnhWUleSDG5mIr38BIzIA