Date of Award

2017

Embargo Period

8-1-2024

Document Type

Dissertation - MUSC Only

Degree Name

Doctor of Philosophy (PhD)

Department

Public Health Sciences

College

College of Graduate Studies

First Advisor

Mulugeta Gebregziabher

Second Advisor

Michael D. Sweat

Third Advisor

Patrick D. Mauldin

Fourth Advisor

Andrew B. Lawson

Fifth Advisor

Brian Neelon

Abstract

Multilevel complex survey data are obtained from study designs that involve multiple stages of sampling where sampling units are drawn at each stage. The features of such survey design often include clustering, stratification, multilevel sample selection, and unequal probability of selection of observations. Typically, specialized methods that account for these features are needed to estimate and make inference on parameters of interest. For example, multilevel models that account for sampling weights have become popular for the analysis of such type of data. Recently, multilevel pseudo-likelihood (MPL) methods with scaled weights are gaining popularity for the analysis of Gaussian and Binomial data. However, there are no studies that assess the performance of pseudo-likelihood and scaling methods on models for count data that are characterized by point mass at zero. The literature on Bayesian modeling of count data from complex surveys is also limited. Thus, we propose to develop and assess the performance of MPL and Bayesian methods for the analysis of count data from complex surveys under several scenarios of sampling weights. Another common issue that arises with complex surveys is aggregation of outcomes and covariates from lower level to higher level (eg. from individual level to household level). But, there are no studies that are developed for dealing with how to deal with the aggregation of sampling weights which is a subject of interest in this proposal. This work accomplished three aims: i) we developed a multilevel pseudo maximum likelihood estimate for count data from multilevel complex survey and assess its performance under several weight scaling approaches. ii) we developed a Bayesian approach for the analysis of count outcomes from complex survey comparing different weight approaches, iii) we developed and assessed an aggregate data model for weighted survey data, which allows for multilevel weight among disease rates across cluster. We apply the proposed analysis strategies of three aims to the real survey data, the multi-country data from DHS (Demographic heath survey) to demonstrate the methods.

Rights

All rights reserved. Copyright is held by the author.

Download

Share

COinS