Nonparametric Bayes models for mixed-scale longitudinal surveys

TL;DR

This paper introduces a novel nonparametric Bayesian method for modeling mixed-scale longitudinal survey data, effectively handling complex survey designs, missing data, and dynamic associations.

Contribution

It presents a new nonparametric Bayesian approach that incorporates survey weights and models mixed-scale longitudinal data with dynamic latent factors.

Findings

Method performs well in simulation studies.

Successfully applied to adolescent health survey data.

Adjusts for survey design biases in posterior inference.

Abstract

Modeling and computation for multivariate longitudinal surveys have proven challenging, particularly when data are not all continuous and Gaussian but contain discrete measurements. In many social science surveys, study participants are selected via complex survey designs such as stratified random sampling, leading to discrepancies between the sample and population, which are further compounded by missing data and loss to follow up. Survey weights are typically constructed to address these issues, but it is not clear how to include them in models. Motivated by data on sexual development, we propose a novel nonparametric approach for mixed-scale longitudinal data in surveys. In the proposed approach, the mixed-scale multivariate response is expressed through an underlying continuous variable with dynamic latent factors inducing time-varying associations. Bias from the survey design is adjusted for in posterior computation relying on a Markov chain Monte Carlo algorithm. The approach is assessed in simulation studies, and applied to the National Longitudinal Study of Adolescent to Adult Health.