Everyone Counts: Advanced Methods for Estimating Marginalized Populations
Open Access
- Author:
- Laga, Ian
- Graduate Program:
- Statistics
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- May 19, 2022
- Committee Members:
- Ashton Verdery, Outside Unit & Field Member
Xiaoyue Niu, Co-Chair & Dissertation Advisor
Le Bao, Co-Chair & Dissertation Advisor
Hyungsuk Tak, Major Field Member
Runze Li, Major Field Member
Ephraim Mont Hanks, Program Head/Chair - Keywords:
- HIV
Key Population
Bayesian Modeling
Small Area Estimation
Prevalence Mapping
Size Estimation
Aggregated Relational Data - Abstract:
- Population size estimation is important in a variety of applications. Ecologists estimate the size of wolf populations to monitor migration. Demographers estimate the number of immigrants in each census. Epidemiologists estimate the people living with HIV to allocate resources for treatment. In what follows, existing approaches to estimate population size are reviewed and several new models are proposed to estimate population sizes under various data types. The main portion of this dissertation consists of five chapters. In the first of these, we develop a model to estimate and map the number of female sex workers in a fine grid cell in Malawi from marked presence-only data. The proposed model uses spatially misaligned covariates to map female sex worker counts to the rest of Malawi. Next, we provide the first mapping of female sex worker size in all of sub-Saharan Africa. We perform a literature review of female sex worker size estimates in sub-Saharan Africa and use these estimates to map female sex worker prevalence in all sub-national and national areas. Our proposed Bayesian hierarchical model combines the data from the diverse countries with auxiliary covariates, controls for biases from the original direct size estimation method, and fills in areas with missing size estimates. Third, we provide a comprehensive and in-depth review of Network Scale-up Method (NSUM) models. The review provides researchers with an introduction to data collected from "How many X's do you know?" questions, delineates the process for selecting a model, outlines known NSUM biases, and discusses future research directions. Fourth, we propose a correlated NSUM model to provide more accurate population size estimates and better understand how individuals form social networks. The existing NSUM models have slowly improved parameter estimates, and our model adds to the development by including a correlation structure. Furthermore, our model reveals whether a respondent who knows someone in one subpopulation is more likely to know someone from another subpopulation. We propose a novel scaling procedure that accounts for shared NSUM biases amongst correlated subpopulations. Finally, we develop the first method to estimate and account for the degree ratio bias without collecting additional NSUM data. The degree ratio arises because members of certain subpopulation, on average, have either larger or smaller network sizes than the average network size of the entire population. This discrepancy in degrees biases size estimates. We correct for this bias by estimating the degree ratio directly by reusing the NSUM responses. We also propose scaling size estimates to better use the known subpopulation sizes, greatly improving size estimates with very little effort.