|
Activity Number:
|
413
- Recent Advances in Statistical Modeling and Machine Learning for Official Statistics and Survey Methodology
|
|
Type:
|
Topic Contributed
|
|
Date/Time:
|
Wednesday, August 10, 2022 : 10:30 AM to 12:20 PM
|
|
Sponsor:
|
Survey Research Methods Section
|
|
Abstract #322626
|
|
|
Title:
|
Hierarchical Bayesian Mixed Effect Models for Spatially Correlated Areal Multi-Distributional Survey Data When Covariates Are Measured with Error and Are Spatially Correlated
|
|
Author(s):
|
Saikat Nandy* and Scott H. Holan and Jonathan R Bradley and Christopher K. Wikle
|
|
Companies:
|
University of Missouri and University of Missouri/U.S. Census Bureau and Florida State University and University of Missouri
|
|
Keywords:
|
American Community Survey;
Basis Function Expansion;
Bayesian Hierarchical model;
Generalized Transformation Model;
Measurement error;
Multi-distributional Response
|
|
Abstract:
|
We introduce hierarchical Bayesian models for predicting high-dimensional tabular survey data which can be distributed from more than one class of distributions (e.g. Gaussian, Poisson, Binomial trials, etc.). We adopt a Bayesian implementation of a Hierarchical Generalized Transformation (HGT) model design to deal with the non-conjugacy of data models for non-Gaussian responses when estimated using a Latent Gaussian Process (LGP) model. We model spatially dependent covariates using a conditionally autoregressive model in a mixed effect setting that also accounts for sampling error variance in the covariates. Survey data is prone to high levels of sampling error and ignoring these errors in the estimation of the parameters leads to biased estimators. The aim is to incorporate a classical measurement error component in a LGP framework to model multiple response type tabular survey data. The proposed models are high dimensional and employ the notion of basis function expansion as effective approach to dimension reduction. The HGT component lends flexibility to incorporate multiple type response datasets under a unified framework, without any major changes in the latent process model.
|
Authors who are presenting talks have a * after their name.
