Index of Deep Disadvantage: Technical Documentation

Summary

Index of Deep Disadvantage: Technical Documentation by Silvia Robles, Jasmine Simington and H. Luke Shaefer of the University of Michigan, accompanying the project Understanding Communities of Deep Disadvantage. It states that the authors constructed a multidimensional Index of Deep Disadvantage for all counties and the 500 largest cities in the U.S., drawing on census and administrative data. It states that the index covers three domains: income, using poverty and deep poverty rates; health, using life expectancy and low birth weight; and social mobility. It states that the authors used principal component analysis to weight the variables and ranked communities on a continuum of disadvantage.

Summary drafted by a model from the document's text below and checked by script against that text before publication. It is a navigation aid, not a reading of what the document proves. Where AI is used

Issuer: Source documents
Document type: Index of Deep Disadvantage: Technical Documentation
Case: Index of Deep Disadvantage: Technical Documentation

Cite this document

“Index of Deep Disadvantage: Technical Documentation.” Original at poverty.umich.edu. Archived at Pandemic Darlings, https://pandemicdarlings.com/source-documents/h-hern-ndez-finch-m/, SHA-256 1a34f7861b1892b0b2365bb3eb23ed1784a7c1a13f4b40994b170f0b9809f316.

Full text

Index of Deep Disadvantage: Technical Documentation

Silvia Robles
Jasmine Simington
H. Luke Shaefer
University of Michigan

In Understanding Communities of Deep Disadvantage, we use methods refined through
Kathryn Edin and Luke Shaefer’s book $2.00 a Day, combining big data with systematic,
in-depth qualitative interviews and ethnographic observations to better understand the history
and contemporary dynamics in a subset of communities of deep disadvantage. We seek to paint a
vivid portrait of the lived-experiences of poor individuals and families in these places, and
explore the views and activities of key stakeholders, including politicians, civil servants, faith
leaders, and representatives of the business community. From these in-depth conversations, as
well as our observations of community events, we seek to uncover community factors that may
drive the sharp disparities documented in our big data work.

The first step in this project was to build a measure of “disadvantage” that could be applied
consistently to communities all over the country. To this end, we constructed a unique,
multidimensional Index of Deep Disadvantage (IDD) for all counties and the 500 largest cities in
the U.S. We drew on census and administrative data to examine vulnerability in three
interconnected domains of high salience to Americans: 1) income, using poverty and deep
poverty rates that are official metrics of well-being for the nation; 2) health, using life
expectancy and low birth weight, both of which are deeply connected to well-being over the life
course; and 3) social mobility, using new estimates for counties and cities. We used principal
component analysis to weight these variables (standardized for comparison). Then, we ranked
communities on a continuum of disadvantage. We tested the robustness of our findings through
inclusion of numerous other factors related to disadvantage. The data we used for this exercise
can be accessed here. We invite others to utilize it to build their own index.

Sources of Data

Income: We include two related measures of income poverty: the share of community residents
with cash incomes falling below poverty, and the share with cash incomes falling below 50% of
poverty, often referred to as deep poverty. The estimates for counties and cities come from the
2017 American Community Survey, 5-year estimates. Poverty is defined according to the Office
of Management and Budget (OMB), and historical poverty thresholds are available on the U.S.
Census Bureau website. The amount of cash income designated as falling below the poverty
level varies by household size, but not by geographic area, and is adjusted for inflation using the
Consumer Price Index.

Health: The index also takes into account life expectancy at birth, from the 2019 RWJF County
Health Rankings and the City Health Dashboard. Life expectancy was estimated using small area
methods from deidentified death records from the National Center for Health Statistics (NCHS),
and population counts from the U.S. Census Bureau, NCHS, and the Human Mortality Database.
For counties, life expectancy is measured over the time period of 2015-2017. For cities, life
expectancy is a six-year average for the time period 2010-2015.

The measure for low birth weight represents the share of live births weighing less than 2500
grams in each county or city. Information on low birth weight is available from the 2019 RWJF
County Health Rankings and the City Health Dashboard, building from data housed at the
National Center for Health Statistics Natality Files. The RWJF County Health Ranking
documentation notes that “these data are submitted by the vital registration systems operated in
the jurisdictions legally responsible for registering vital events (i.e. births, deaths, marriages,
divorces, and fetal deaths). Missing values are reported for all counties where fewer than 20
births were considered low birth weight.” For cities, low birthweight is a three year average for
the time period 2015-2017. For counties, the measure captures the time period of 2011-2017.
Life expectancy and low birth weight together constitute the health domain for the indicator.

Social mobility: Data on adult income mobility by county are publicly available from Chetty
and Hendren (2018). These observational estimates represent the mean household income rank
for children whose parents were at the 25th percentile of the national income distribution.
Income is measured as mean earnings in 2014-15 for individuals at age 26 and household income
is defined as the sum of own and spouse’s income. By focusing on people whose parents were in
the 25th percentile of the national income rankings, the measure characterizes intergenerational
mobility specifically for low-income populations who may be more vulnerable to variation in
local conditions relative to high-income populations who can use their own resources to insulate
themselves from differences in community resources. Chetty and Hendren (2018) note that these
observational mobility estimates actually represent a large portion of the causal effect of living in
a given county on whether a child moved up or down in the national income distribution relative
to the average 26-year old. To estimate mobility for the top 500 cities, the index uses the same
mean household income rank variable from Opportunity Insights, but starting at the tract level,
aggregates up to the city. We verified the legitimacy of the tract to city aggregation by
aggregating from tract to county, then comparing the aggregated tract-to-county estimates with
the publicly available county estimates. Less than 1% of counties had an aggregate
tract-to-county measure that was one standard deviation away from the county estimate, giving
us confidence that the tract to city aggregation is sound.

Geographic unit of analysis
Defining “communities” is challenging. In our case, the primary objective is to compare
communities across a set of consistent metrics, and so counties and cities become our primary
units of analysis. To our knowledge, this is the first such index to compare both counties and
cities on the same outcomes. The reason for doing this is to identify cities of deep disadvantage
situated in counties where the non-city population has more resources. For instance, Wayne
county, Michigan does not appear among the nation’s most disadvantaged places according to
our index, but the City of Detroit, within Wayne County, does.

One might argue that some cities with extremely high rates of inequality will not appear on our
index, such as New York City or Los Angeles, because the disadvantage in these places is
shrouded by their proximity to areas of affluence. Yet the same argument could be made for all
counties. We welcome ideas and debate about what the right unit of community is. And we
would argue that while some might make the case that there are places missing from the IDD’s
list of communities of deepest disadvantage, the ones that do appear on our list are clearly so,
given that they emerge even despite all these factors.

A novel aspect of our index, in fact, is that it ranks cities and counties together, with the goal of
understanding the differences in vulnerability across place. The analysis excludes a handful of
cities and counties for various reasons, most commonly:
(1) The county was a U.S. territory, and not a part of the continental U.S., Hawaii or Alaska; or
(2) The city was not one of the top 500 most populated cities; or
(3) The city and county were spatially equivalent units, therefore we excluded the county. Note
in these cases that we find that the city and county estimates are virtually identical

The full list of geographic areas excluded, and the reason why, can be found in Appendix B.

Urban, Rural, and Tribal Lands
We classify counties as urban or rural using definitions from the National Center for Health
Statistics, which offers more nuance than census definitions. Definitions drawn from this source
are below:
Urban counties include: (1) Counties in MSAs of 1 million or more population that: contain the
entire population of the largest principal city of the MSA; or have their entire population
contained in the largest principal city of the MSA; or contain at least 250,000 inhabitants
of any principal city of the MSA; (2) Other counties in MSAs of 1 million or more
population but that did not qualify as large central metro counties; (3) Counties in MSAs
of populations of 250,000 to 999,999; and (4) all other counties in metropolitan MSAs.
Rural counties are all other non-urban counties, including counties that are in
micropolitan statistical areas or nonmetropolitan counties that did not qualify as
micropolitan.

Furthermore, counties that include any form of federally-recognized tribal land are flagged as
such, using standard census and OMB definitions.
According to the census, the various types of Tribal land include: (1) reservations—areas with
boundaries established by treaty, statute, and/or executive or court order. Reservations
can cross state boundaries; (2) Tribal trust land—real property, held in trust by the
Federal Government, that is associated with a specific American Indian reservation or
tribe, or, in some cases, individual American Indians. Tribal trust land may be located
inside or outside of a Reservation; (3) Tribal jurisdiction statistical areas—land
associated with Federally recognized tribes in Oklahoma that no longer have a
reservation; (4) Tribal designated statistical areas (TDSAs)— geographic entities
delineated by Federally and State recognized tribes that have no reservation or trust
lands. TDSAs consist of territory that contains the American Indian population over
which a Federally recognized tribe has jurisdiction or territory within which a
State-recognized tribe provides benefits and services to its members; and (5) Alaska
Native Regional Corporations—There are 12 ANRCs that cover the entire state of
Alaska.
Methodology
The IDD is the first principal component from a principal component analysis (PCA) of the five
features. Principal component analysis is a technique often chosen to reduce the dimensions
being considered in prediction-style analysis. This made it a natural choice to summarize the
information carried across the five characteristics we chose. The five factors, while highly
correlated (see Table A1), also highlight different aspects of the sources and consequences of
deep disadvantage. PCA yields a weighted average of the five features where the weightings are
chosen to capture as much of the variation in the observed data as possible.

The first step was to gather the variables. Table A2 lists the source and a URL for each piece of
information included in the index. PCA is sensitive to the magnitude of each feature included.
Therefore, after merging all five factors into one dataset, each feature was normalized by
subtracting its mean and dividing by its standard deviation.

Missing data was a challenge for this procedure. The first approach to reduce the potential for
bias due to missing information was to choose the county and city as the units of analysis. Many
metrics are available nation-wide at the county (or city) level, but are not measured at a more
fine-grained unit of analysis such as census tracts. In addition, the analysis focused on variables
that were available for a large fraction of counties in the U.S. If a county was missing any of the
features, the principal component analysis was completed with the remaining features. The
analysis examined data only from mainland U.S. counties, Alaska, and Hawaii. U.S. territories
were excluded.

The first principal component represents over sixty percent of the variation (or information) in
the data. The weights on each variable are fairly even, with a slightly higher loading on the share
of community residents in poverty.

Sensitivity Analyses
The project undertook two types of sensitivity analysis. The first checked for stability of the
measure to the addition of different or additional characteristics. The second checked for face
validity by examining whether the most disadvantaged communities (those with the lowest index
rankings) displayed compound disadvantage. It would be less credible for a county to be ranked
as extremely disadvantaged if most of the measures were not demonstrating disadvantage.

The results of testing for stability are available here. The first tab lists the top 50 most
disadvantaged counties according to seven different specifications of the index. Each
specification adds one or more additional feature to the core index. The additional features were
either county arrest rates from the National Archive of Criminal Justice Data, the share of adult
county residents with a high school diploma or less education from the ACS, the share of adult
county residents who are underemployed (defined as percent working part-time or less than 34
hours per week) or combinations of these three additional features. The rankings are remarkably
stable. The second tab describes the original rankings of any county that broke into the top 50
most disadvantaged counties under the different specifications. A maximum of seven counties
move into the top 50 in any single alternative specification, and across all alternative
specifications, 10 unique counties break into the top 50. Only four of the ten have substantially
different rankings in the core index. Overall, we conclude that the index is generally stable to the
addition of other metrics.

Defining compounded disadvantage
To check for face validity of the index we define a measure of compound disadvantage where a
place must rank in the bottom quartile of at least one characteristic in two of the three areas of
focus. The areas of focus are poverty, health (life expectancy and low birth weight), and social
mobility.

One of the outputs of the analysis is a ranked list of the 300 most disadvantaged localities. To
qualify as a city or county in the top 300, localities had to meet the compound disadvantage
criteria above. In the original ranking, 299 of 300 places met this criteria. When a place did not
fit the definition of compounded disadvantage, it was bumped down in the ranking and replaced
by the next highest ranked geography that did.

We consider the Index of Deep Disadvantage to be the start of a conversation. We welcome
feedback. Visit the map that shows the results of this work, and the underlying dataset can also
be found on the same page.

Appendix A. Supplemental Tables
Table A1: Correlation between any two components of the index
Mobility Poverty Deep Life Low Birth
Poverty Expectancy Weight

Mobility 1.00

% Pov. -0.53 1.00

% Deep -0.52 0.94 1.00
Pov.

Life Exp 0.45 -0.68 -0.56 1.00

Low -0.59 0.56 0.51 -0.67 1.00
Birth Wt.

Table A2: List of component features of the index with links to their sources
Feature Data Source
Chetty and Hendren. estimates Opportunity Insights

Life expectancy Robert Wood Johnson Foundation County
Health Rankings (RWJ); City Health
Dashboard
Percent of births that are low birth Robert Wood Johnson Foundation County
weight Health Rankings (RWJ); City Health
Dashboard
Percent of residents below the poverty American Community Survey (ACS)
line

Percent of residents living in deep American Community Survey (ACS)
poverty
Appendix B. Geographic Sample
Dropped County Reason for dropping

All 78 Puerto Rican Dropped b/c U.S. territory
counties

Hoonah-Angoon Boundaries have changed and has been annexed to other counties/
Census Area, Alaska functions as a part of other counties

Prince of Wales-Hyder Boundaries have changed and has been annexed to other counties/
Census Area, Alaska functions as a part of other counties

Wade Hampton Census Boundaries have changed and has been annexed to other counties/
Area, Alaska functions as a part of other counties

Wrangell – Petersburg Boundaries have changed and has been annexed to other counties/
Census Area, Alaska functions as a part of other counties

Kalawao County, Does not function like other counties: it has a very a small
Hawaii population & it is a judicial district of Maui County ,which includes
the rest of the island of Molokaʻi. Kalawao county has no elected
government

Bedford, Virginia Bedford county added an independent city in 2013 that was
previously separate. So , we have a record for an independent
Bedford City at the FIPS level and for Bedford County at the FIPS
level. SOLUTION: remove Bedford City and just use the FIPS code
that represents the whole (Bedford) county

Clifton Forge, Virginia The city of Clifton Forge (FIPS=51560) was incorporated into
Allegheny County (FIPS=51005)

Anchorage, Alaska In Alaska, municipalities, rather than counties, are the primary legal
divisions. The municipality of Anchorage, Alaska is considered a
consolidated city-county (shows up as both an independent city and
county). Because of the overlap, this county is dropped.

Denver, Colorado Denver is in our top 500 cities && is considered a consolidated
city-county (shows up as both an independent city and county).
Because of the overlap, the county is dropped.

Honolulu, Hawaii Honolulu is in our top 500 cities && is considered a consolidated
city-county (shows up as both an independent city and county).
However, all Honolulu data from the ACS and RWJ were
county-level metrics, so the city is dropped and the county is
retained.
Orleans Parish, New Orleans is in our top 500 cities && is considered a
Louisiana consolidated city-county (shows up as both the independent city of
New Orleans and Orleans Parish). Because of the overlap, the
county is dropped.

San Francisco, San Francisco is in our top 500 cities && is considered a
California consolidated city-county (shows up as both the independent city of
San Francisco and San Francisco County). Because of the overlap,
the county is dropped.

Philadelphia County, Philadelphia is in our top 500 cities && is considered a consolidated
Pennsylvania city-county (shows up as both the independent city of Philadelphia
and Philadelphia County). Because of the overlap, the county is
dropped.

New York City, New NYC is complex--each borough is considered its own county, but
York together the five boroughs represent New York City. is in our top
500 cities && because of the overlap, the county is dropped.

Washington, D.C. D.C. is in our top 500 cities && is considered a consolidated
municipal government, and the county appears to be comprised
primarily of D.C. the city an da few unincorporated areas. Because
of the overlap, the county is dropped.

Baltimore, Maryland Baltimore is in our top 500 cities && is considered a consolidated
city-county (shows up as both the independent city of Baltimore and
Baltimore County). Because of the overlap, the county is dropped.

Durham, North Durham is in our top 500 cities && is considered a consolidated
Carolina city-county (shows up as both the independent city of Durham and
Durham County). Because of the overlap, the county is dropped.

St. Louis, Missouri St. Louis is in our top 500 cities && is considered a consolidated
city-county (shows up as both the independent city of St. Louis and
St. Louis County). Because of the overlap, the county is dropped.

Alexandria, Virginia This is a consolidated city. The city is in our top 500 cities, and also
shows up as a county. Because of the overlap, the county is
dropped.

Chesapeake, Virginia This is a consolidated city. The city is in our top 500 cities, and also
shows up as a county. Because of the overlap, the county is
dropped.
Hampton, Virginia This is a consolidated city. The city is in our top 500 cities, and also
shows up as a county. Because of the overlap, the county is
dropped.

Lynchburg, Virginia This is a consolidated city. The city is in our top 500 cities, and also
shows up as a county. Because of the overlap, the county is
dropped.

Norfolk, Virginia This is a consolidated city. The city is in our top 500 cities, and also
shows up as a county. Because of the overlap, the county is
dropped.

Portsmouth, Virginia This is a consolidated city. The city is in our top 500 cities, and also
shows up as a county. Because of the overlap, the county is
dropped.

Richmond, Virginia This is a consolidated city. The city is in our top 500 cities, and also
shows up as a county. Because of the overlap, the county is
dropped.

Roanoke, Virginia This is a consolidated city. The city is in our top 500 cities, and also
shows up as a county. Because of the overlap, the county is
dropped.

Suffolk, Virginia This is a consolidated city. The city is in our top 500 cities, and also
shows up as a county. Because of the overlap, the county is
dropped.

Virginia Beach, This is a consolidated city. The city is in our top 500 cities, and also
Virginia shows up as a county. Because of the overlap, the county is
dropped.

File and source

File: h-hern-ndez-finch-m.pdf
Size: 140,400 bytes
SHA-256: 1a34f7861b1892b0b2365bb3eb23ed1784a7c1a13f4b40994b170f0b9809f316
Our copy: h-hern-ndez-finch-m.pdf
Original: poverty.umich.edu