Full Citation
Title: Using Machine Learning to Estimate Racially Disaggregated Wealth Data at the Local Level
Citation Type: Miscellaneous
Publication Year: 2023
ISBN:
ISSN:
DOI:
NSFID:
PMCID:
PMID:
Abstract: Household wealth data at the local level are generally not widely available, especially statistics disaggregated by race and ethnicity. Using machine learning, we estimate net worth and emergency savings data at the local, city, state, and national levels. We also disaggregate our estimates by racial and ethnic groups at the city, state, and national levels. We use machine-learning models trained on the Survey of Income and Program Participation (SIPP), a survey with detailed wealth information but too few observations for local estimates, to estimate emergency savings and net worth using the American Community Survey (ACS), which allows for state and local estimation. The imputations allow us to estimate the proportion of households with more than $2,000 in emergency savings and median net worth at the household level. We then aggregate household-level data to different geographic levels. Household wealth is a safety net. It protects families from unexpected expenses such as replacing a water heater and income shocks such as a jobless spell. Household wealth is also a springboard. It gives families the opportunities to invest in wealth-building opportunities like starting a small business or moving to an area with more opportunity. Understanding wealth is necessary for uncovering the barriers to wealth generation and designing policies that unlock opportunity for everyone. Two datasets, the Federal Reserve Board’s Survey of Consumer Finances and the US Census Bureau’s Survey of Income and Program Participation (SIPP), have detailed information about wealth and are typically used for national or state estimates (Bhutta et al. 2020). Unfortunately, little information is widely available about wealth at the local level, especially disaggregated by different racial and ethnic groups. This is because household surveys with adequate sample sizes for local estimation, such as the American Community Survey (ACS), do not include detailed information about wealth. We employ machine learning to leverage a smaller nationally representative survey with detailed questions about household wealth to impute household wealth onto the ACS, which allows for small area estimates. We released this novel dataset with the Urban Institute’s “Financial Health and Wealth Dashboard” in 2022. In particular, we use machine learning to estimate median net worth and the proportion of households with at least $2,000 emergency savings at the local level defined as Public Use Microdata Areas (PUMAs). We provide financial measures not only for a city, but also for subregions (Public Use Microdata Areas, or PUMAS) within many cities and for different racial and ethnic groups. City leaders and policymakers are able to use the dashboard to identify wealth disparities within cities. Because household wealth measures for subregions within cities are not typically available in public survey data, we use a machine-learning method to fill the gap in the existing survey data. We use household measures including net worth and emergency savings to better understand households’ wealth. Liquid assets—the sum of assets in checking accounts, stocks, bonds, and other liquid savings accounts—can indicate households’ resilience and ability to bounce back from financial shocks. Net worth—total assets minus total debt—can provide an overview of households’ economic well-being and their ability to pursue new opportunities. Overall, the 2022 financial health dashboard provides firsthand data and analysis to city leaders and policymakers for a better understanding of financial health, financial resilience, wealth, and debt within local regions in their cities and counties. They will also be able to compare their cities and counties with neighboring cities and counties. This works builds on previous asset and wealth imputation research. Using 2014 data, the Prosperity Now Asset Scorecard estimated the asset poverty rate, liquid asset poverty rate, and share of households with liquid assets for many cities in the US. The Federal Reserve Bank of “St. Louis’s Real State of Family Wealth”2 provides national-level average wealth for various demographic groups overtime. Using 2013 data, Ratcliffe and colleagues (2017) estimated the shares of unbanked and underbanked populations for different PUMAs in New York City. This project advances this literature by (1) looking at a broader set of asset outcomes; (2) estimating assets at the PUMA level throughout the country; (3) using more recent data than the literature; and (4) disaggregating the estimates by race and ethnicity when possible. Because PUMA boundaries are often smaller than city boundaries, our analysis sheds light on wealth disparities within cities.
User Submitted?: No
Authors: Williams, Aaron R; Zhong, Mingli; Braga, Breno
Publisher:
Data Collections: IPUMS USA
Topics: Housing and Segregation, Population Data Science
Countries: