Using Zero-inflated Regression and the Homophily Principle to Model Migration for Population Projections

Abstract

Sub-national population projections have become an essential component of policy relevant environmental assessment research. This dissertation develops 96 unique migration models by combining spatial variables, regression model types, and data sources. Overall performance is assessed with out-of-sample validation for each model using five error metrics. I find that a zero-inflated negative binomial model using modified versions of Stouffer’s Intervening Opportunities and Competing Migrants calculations yielded the best overall results. Model performance was improved by conceptualizing migration according to the homophily principle and partitioning model inputs by race. After incorporating projections of births, deaths, and immigration, this new model was used to simulate migration and project county population. I find that this new model performs well compared to existing projections and I define new quantitative benchmarks for evaluating population projections going forward.