Methodology Report
v.1.1.0 (5/7/18)
1
Table of Contents
Table of Contents 1
I.
Introduction 2
I.1.
How to Cite the Data and Methodology Report 3
II.
Data Collection 4
II.1.
Court Records 4
II.2.
Data Acquisition 4
III.
Data Cleaning 9
III.1.
Reformatting the Data 10
III.2.
Dating an Eviction Case 11
III.3.
Commercial Cases 12
III.4.
Name Standardization 13
III.5.
Preparing Addresses for Geocoding 16
III.6.
Separating Street Address from Apartment Designation 18
III.7.
City, State, and Zip Code Cleaning 18
III.8.
Geocoding and Linking to Census Geographies 21
III.9.
Address Standardization 24
III.10.
Duplicate Records 26
III.11.
Serial Cases 26
IV.
Estimating the Prevalence of Eviction 28
IV.1.
Classifying Eviction 28
IV.2.
Estimating Case Volume Fluctuation 28
IV.3.
Imputation 32
IV.4.
Aggregation 33
IV.5.
External Data Sources 34
IV.6.
Denominator 35
V.
Validation of Estimates 36
V.1.
State Data Merges 36
V.2.
State-Reported County Statistics 37
V.3.
Low and High Rates 39
VI.
Future Research 43
2
I.
Introduction
In recent years, renters’ housing costs have far outpaced their incomes, driving a nationwide
affordability crisis. Current data from the American Housing Survey show that most poor renting
families spend at least 50 percent of their income on housing costs.
1
Under these conditions,
millions of Americans today are at risk of losing their homes through eviction.
An eviction occurs when a landlord forcibly expels a tenant from a residence. While the majority
of evictions are attributed to nonpayment of rent, landlords may evict tenants for a variety of
other reasons, including property damage, nuisance complaints, or lease violations. A formal
eviction occurs when a landlord carries out an eviction through the court system. Conversely, an
informal eviction occurs when a landlord executes an eviction without initiating a legal process.
For example, a landlord may offer a buyout or perform an illegal lock-out. Until recently, little
was known about the prevalence, causes, and consequences of eviction.
The Eviction Lab at Princeton University has collected, cleaned, geocoded, aggregated, and
publicized all recorded court-ordered evictions that occurred between 2000 and 2016 in the
United States. This data set consists of 82,935,981 million court records related to eviction cases
in the United States between 2000 and 2016, gleaned from multiple sources. It is the most
comprehensive data set of evictions in America to date.
These data allow us to estimate the national prevalence of court-ordered eviction, and to compare
eviction rates among states, counties, cities, and neighborhoods. We can observe eviction trends
over time and across geography, and researchers can link these data to other sources of
information.
The Eviction Lab at Princeton University is directed by Matthew Desmond and designed by
Ashley Gromis, Lavar Edmonds, James Hendrickson, Katie Krywokulski, Lillian Leung, and
Adam Porton, with research assistance from Henry Gomory, Samarth Gupta, Chase Hommeyer,
Hyojin Lee, Sejin Park, Gillian Slee, and Haley Zeng. The Eviction Lab is funded by the JPB,
Gates, and Ford Foundations as well as the Chan Zuckerberg Initiative. More information is
found at www.evictionlab.org.
The following report details the Eviction Lab’s methodology. For further information, please
email [email protected].
1
Matthew Desmond, “Unaffordable America: Poverty, Housing, and Eviction,” Fast Focus: Institute for Research
on Poverty 22 (2015) 1-6.
3
I.1.
How to Cite the Data and Methodology Report
Citations for academic publications and research reports:
Data: Matthew Desmond, Ashley Gromis, Lavar Edmonds, James Hendrickson, Katie
Krywokulski, Lillian Leung, and Adam Porton. Eviction Lab National Database: Version
1.0. Princeton: Princeton University, 2018, www.evictionlab.org.
Methodology Report: Matthew Desmond, Ashley Gromis, Lavar Edmonds, James
Hendrickson, Katie Krywokulski, Lillian Leung, and Adam Porton. Eviction Lab
Methodology Report: Version 1.0. Princeton: Princeton University, 2018,
www.evictionlab.org/methods.
Citations for media outlets, policy briefs, and online resources:
Eviction Lab, Princeton University, www.evictionlab.org.
4
II.
Data Collection
II.1.
Court Records
In the United States, eviction cases are typically heard in civil courts at the county-level.
Although information contained in court records varies between jurisdictions, most records
include the date the eviction case was opened, the names and addresses of plaintiffs (landlords)
and defendants (tenants), the resolution of the case, and the money judgment (if any). Depending
on the jurisdiction, an eviction may be classified as a “forcible detainer,” “forcible entry and
detainer,” or “unlawful detainer.”
Courts resolve most eviction cases in one of three ways:
(1) an eviction judgment is rendered, ordering the defendants to vacate a premise by a
specific date;
(2) the case is dismissed or ruled in favor of the defendant, allowing the tenant to remain
in the residence;
(3) a mediated agreement, often called a “settlement” or “stipulation,” is established
between a landlord and a tenant. In a mediated agreement, the tenant must meet certain
terms or adhere to a negotiated payment schedule. If the terms are met, the eviction is
dismissed; if they are not, an eviction judgment can be rendered.
Table 1 provides an example of the information included in a typical eviction record.
Table 1. Sample Eviction Record
Case Number
Defendant
Defendant
Plaintiff
Plaintiff
Street City State Zip
Filing
Judgment Outcome Judgment
First
Name
Last
Name
First
Name
Last
Name
Date
Date
Awarded
2016NYBRO-01 Jane Doe John Adams 12
Main
Street
New
York
NY 12345 1/1/2016 4/1/2016 Find for
Plaintiff
$700
II.2.
Data Acquisition
5
The Eviction Lab created a centralized repository of U.S. eviction records by acquiring data from
states and counties and purchasing records from two independent data acquisition companies.
We first contacted court clerks in all 50 states, requesting details about eviction records in their
jurisdictions.
2
Depending on the staffing, the use of electronic data storage, and the
responsiveness of individual courts, the time from initial contact to completion of our request
was between one month (North Carolina) and ten months (North Dakota).
We collected eviction records en masse where they were available. As a result, we made requests
of court systems for bulk reports. Bulk reports provided a corpus of a court’s available eviction
case information, viz., names, addresses, dates of eviction filings, and outcomes. Hundreds of
thousands of records can be retrieved through a single bulk record request, streamlining data
collection. The Freedom of Information Act (FOIA) extends the right to access eviction records;
however, local policy barriers and resource limitations pose significant challenges to data
acquisition. Some states, such as Illinois, enforce policies preventing bulk data retrieval.
Through bulk requests, we obtained 12,870,070 individual-level, court-ordered eviction records
from 13 states. Table 2 represents the number of records we received from each state and the
range of years for which we have data.
The Eviction Lab also purchased data sets of public eviction records from two companies:
LexisNexis Risk Solutions (LexisNexis)
3
and American Information Research Services Inc.
(AIRS)
4
. We obtained 66,512,547 eviction records from LexisNexis, which included all states
except North and South Dakota. LexisNexis’ cases occurred between February 1970 and October
2017. AIRS provided 10,606,467 eviction records from six states: Arizona, California, New
Jersey, Nevada, Ohio, and Oregon. AIRS’s cases occurred between January 1988 and March
2018.
Table 2. Records Collected from State Courts
State Years Number of
Observations
2
Not all records are collectible. Some eviction records may be sealed. For instance, California established policies
allowing tenants to block public access to prior eviction court records. Wisconsin seals cases ending in a dismissal
after two years, pulling them out of their database and destroying the case information.
3
LexisNexis records were last updated on September 29th, 2017.
4
AIRS data were last updated on March 9th, 2018.
6
Alabama
1,148,022
Connecticut
377,095
Hawaii
28,955
Iowa
374,419
Indiana
366,246
Minnesota
414,393
Missouri
534,775
Nebraska
857,687
New Jersey
1,638,451
Oregon
709,467
South Carolina
2,591,938
Pennsylvania
2,769,941
Virginia
1,058,681
Note: Pennsylvania data come from a combination of two data
sets: Pennsylvania State Court data for all counties between
2006-2016, and Philadelphia County Court data from 1969-
2016.
As of April 2018, the Eviction Lab has collected a set of 82,935,981 million court records related
to eviction cases in the United States between 2000 and 2016. Not all records and cases were
suitable for inclusion in our analysis. Duplicate records and non-meaningful records (e.g., blank
records, test records) were removed. Table 3 shows a summary of the number of records we had
in our database after dropping unsuitable cases. Numbers are reported between 2000 and 2016,
as this is the timeframe we used for analysis and produced on the map.
7
Table 3. Relevant Records by Data Source (2000-2016)
Data Source
Number of Records
LexisNexis
63,677,688
State Court Data
10,895,619
AIRS
8,362,674
To retrieve these records, the companies utilized collection methods similar to our own (e.g.,
automated record collection, bulk record collection). In addition, individual records were
collected in person from courts where bulk and online collection were not available. Examining
paper court records granted access to name and address information available on both the
petition and the final judgment, as well as allowed them to determine the outcome of the case.
This included any money judgments related to landlord/tenant cases, dismissals, and instances
where a tenant was ordered to vacate the premises. Record information was then consolidated
through data entry. Not all available eviction-related records can be collected in person, given the
time and resource-intensive nature of the practice. However, these data represent the most
comprehensive attempts to collect court-ordered eviction records nationwide.
Our data include all landlord/tenant-related cases. Some landlord/tenant cases do not involve the
return of a property to the landlord. For instance, some cases involve a disputed security deposit.
To determine which cases resulted in eviction, the companies drew on court procedures,
developing criteria for determining whether an eviction took place. Because the data come from
independent sources, they contain additional information apart from the standard collected from
states. Some states indicate whether the case involved possession of property. If the possession
indicator was present, the case was classified as an eviction. Other states had more parsimonious
case categorizations, indicating whether a money judgment, eviction, or dismissal took place.
Our team also gathered aggregate state-reported county-level data from 27 states, New York
City, and the District of Columbia.
5
We attempted to collect reports of the number of
5
These states include: Alaska, Arkansas, California, Colorado, Delaware, Florida, Georgia, Hawaii, Maine,
Maryland, Massachusetts, Michigan, Missouri, Nebraska, Nevada, New Jersey, New Mexico, North Carolina, North
Dakota, Ohio, Pennsylvania, South Dakota, Texas, Utah, Vermont, Virginia, and Wyoming.
8
landlord/tenant cases filed in counties within a state each year.
6
If the reports were not available,
then our team made individual requests to each state. Many states did not maintain reports or
retain landlord/tenant case information housed at the state-level.
We converted any PDFs or workbooks into CSV files, then wrote a Python program to rearrange
and extract relevant eviction count information. We accomplished this by searching for any
information in a report pertaining to an eviction-related case type and outputting it into a separate
spreadsheet. In Ohio, Massachusetts, and Alaska, filings were reported by courts rather than
counties. Our team classified all reported filings into counties (based on courthouse location) to
maintain standard geographic units for comparison. Using a list of courts provided by
LexisNexis, we were able to identify which county a court belonged to based on indicators in
this list. If we could not determine this information from the list, we classified the court into a
county based on the address of the courthouse.
7
Only a handful of states maintained data on how landlord/tenant cases ended, but all had
information on number of cases opened and/or filed in a given year.
8
All data from these reports
were aggregated together to create a uniform output, which can be found in the data download
section of our website. The file lists the year reported, the state and county, and the number of
cases reported each year in the report.
9
Table 4 serves as an example output.
Table 4. Example Aggregate Output
Year
State
County
State-Reported Cases
2004
WI
Adams
24
2004
WI
Ashland
10
2004
WI
Barron
500
2005
WI
Adams
35
6
Some states did not report case volume as “landlord/tenant.” The name for these types of cases varied by
jurisdiction. We sought to collect all case types that fell under the relevant umbrella but did not include foreclosures.
7
In some smaller areas, court jurisdictions may cross county lines. Our team was unable to determine the precise
coverage area of each individual court, so we relied on the list of courts, as well as the seat of the county courthouse,
to determine in which county to report the filings.
8
Some states reported case volume by the fiscal year. Because we had no way to accurately determine the
distribution of cases over the year, we report the year as the first calendar year in the reported fiscal year.
9
For North and South Dakota, we received lists of case numbers for landlord tenant cases, which county the case
pertained to, and filing dates. In these states we summed the number of cases in each year and reported the
information in the same format.
9
2005 WI Ashland 7
2005 WI Barron 425
The Eviction Lab collected records from a variety of sources, with many overlaps in jurisdiction.
Each source differed in its ability to collect and report eviction cases. For instance, had we
reported only state-reported case filings we would have been unable to determine the eviction
rate, or where evictions occurred within a county. Our primary objectives with our data and map
were to promote comparability between areas over time, and to achieve geographic specificity
when reporting. To this end, we used the most nationally comprehensive data source available,
which is LexisNexis.
There are a handful of exceptions. Between 2007 and 2016 in Alaska, Arkansas, North Dakota,
South Dakota, and Pennsylvania, we lacked sufficient individual-level eviction records to create
estimates of eviction filings or evictions. Given this circumstance, we chose to supplement our
statewide gaps with data from external sources. For Alaska, Arkansas, North Dakota, and South
Dakota, we derived estimates from state-reported volume of landlord/tenant case filings. In
Pennsylvania between 2007 and 2016, we included the individual-level state court data.
In the following sections, the counts found in all tables are based on individual-level data that we
used to produce the estimates found on the website. Our data includes nationwide LexisNexis
records between 2000-2016 – excluding the four states listed above – and Pennsylvania state
court data between 2007-2016.
III.
Data Cleaning
The eviction court records in our data set were created for administrative purposes rather than
research analysis. Accordingly, the quality of information and level of detail included in these
records varied substantially across municipalities. The Eviction Lab’s data represent nearly every
jurisdiction in the country, so it was essential to standardize each record to create accurate
measures for comparison. It was similarly important to correct errors that occurred during data
entry or collection. For example, we eliminated instances where court clerks created a duplicate
record of the same eviction case and corrected clear instances of typing errors in addresses. This
required thorough investigation and analysis of the data.
We used regular expressions in many of our data cleaning procedures. Regular expressions are
sets of characters that can be used to search for a specified pattern within text. They rely on
wildcard characters and other conventions that specify a character’s location within a string of
10
text and/or how often a character appears in a string of text. Using regular expressions allowed
us to detect irregularities in text variables and to collect and rearrange pieces of information.
We implemented regular expressions in Stata, predominantly using the standard set of regular
expressions commands:
regexm() to find a specific pattern of text
regexr() to find a specific pattern of text and replace it with another chunk of text
regexs() to save a specific chunk of text identified using regexm()
To allow an expanded set of special and wildcard characters, we also used their unicode string
equivalents: ustrregexm(), ustrregexra(), and ustrregexs().
III.1.
Reformatting the Data
Renting households are our main unit of our analysis. We focus on households, rather than
individuals, because not all household members are listed as defendants when eviction cases are
filed in court. Court records do not reveal how many individuals, particularly children or other
household members not listed on a rental lease, are affected by an eviction. In our data, each
entry records an action between a defendant and a plaintiff associated with an eviction case. We
received some data in full case form and some as individual actions within cases. Depending on
when records were collected and our data source, we could see an eviction filing and an eviction
judgment for the same defendant as two separate entries sharing the same case number and court
identifier. Multiple entries could therefore be recorded under the same court identifier and case
number, when there were multiple actions between the same plaintiff(s) and defendant(s),
multiple defendants listed on the case, or a combination of both. We reformatted the data so that
each record contained all information on defendants and action types pertaining to the case. In
this format, each case represented one household. Table 5a below presents an example of data
format upon receipt, and Table 5b presents our data format.
Table 5a. Example Format of Original Data
Case
Number
Court
Number
Action
Type
Defendant
First Name
Defendant Plaintiff Plaintiff
Last Name First Name Last Name Street
City
State
Zip
123
456
Filing
Jane
Doe
John
Adams
12 Main
Street
Albany
NY
12345
123
456
Filing
John
Doe
John
Adams
12 Main
Street
Albany
NY
12345
123
456
Judgment
Jane
Doe
John
Adams
12 Main
Street
Albany
NY
12345
11
123 456 Judgment John Doe John
Adams
12 Main
Street Albany NY 12345
Table 5b. Format of Reshaped Data
Case
Number
Court
Number
Number of
Filings
Number of
Judgments
Number of
Defendants
Number of
Plaintiffs Street City State Zip
12 Main
123 456 2 2 2 1 Street Albany NY 12345
Similarly, we reformatted state court data so that each entry corresponded to a single case, each
uniquely identified by a court identifier and case number.
10
We included 38,564,127 total unique
cases in our analysis between 2000 and 2016.
Not all unique cases were included in the analysis. Cases were excluded for three reasons. First,
we excluded cases that duplicated information contained within another case already present in
the data. Second, we excluded cases with commercial defendants. Commercial defendants
indicate that the eviction is being carried out against a business rather than renting household.
Identification of commercial defendants is discussed in more detail below. Finally, we excluded
cases that occurred before 2000 or after 2016. Case collection was more consistent after 2000,
and 2016 is the most recent year for which we have complete data.
III.2.
Dating an Eviction Case
Nearly every action on an eviction case was associated with a date in the data provided by
LexisNexis.
11
When the action was a new case filing, this date reflected the filing date. When the
action was the dismissal of the case, this date reflected the dismissal date. When the action was
the result of a judgment, such as awarding rent owed to a landlord or issuing a writ of restitution
for the property, the judgment date was recorded. The Pennsylvania court data only included
10
The state court data we received from Pennsylvania (this excludes data from Philadelphia County) were missing
several key variables, including defendant names, plaintiff names, case numbers, and filing years. As a result, we
had to determine which observations corresponded with which unique cases. Many of the cases in these data
included a unique judgment number that associates with a unique case. For cases without a judgment number, we
generated a unique case identifier using a combination of relevant variables: county name, court office code,
judgment disposition date, judgment disposition, and claim amount. We then assigned this identifier to cases as a
proxy for case numbers.
11
There were 68 cases in the data that were missing action dates. These cases were excluded from the analysis.
12
judgment dates for cases.
12
There were no records of when the case was initially filed. The
Philadelphia court data included both filing and judgment dates.
Of the 38,564,127 cases in the analytic sample, 35.1% of cases did not include a record of the
initial filing date, while 38.1% of cases did not include information about what occurred between
the landlord and tenant after the case was filed (i.e., no judgment or dismissal information was
present). To calculate yearly filing and eviction rates, each case was assigned a fixed date of
record that was used as a timestamp. We assigned the date of record using the earliest action on
a case. In approximately 65% of cases this was the filing date. In the remaining cases, it was the
date of the first judgment or dismissal. The date of record was used to to aggregate cases over
time. For consistency, cases are assigned to the year of their date of record. For example, if a
case showed a filing date of November 30, 2010, and a first judgment date of January 5, 2011,
that case is recorded as having occurred in 2010.
LexisNexis data also contained release dates that were populated if the entry recorded an action
involving a “judgment release.” A judgment release is the official notice from the court that a
judgment has been fulfilled. If a tenant is ordered to leave the property at the conclusion of an
eviction case, a landlord may return to court to file a release confirming that the tenant vacated
the property. The action dates associated with release entries reflect the date of the judgment the
release satisfies. For consistency, when the only action present on a case was a judgment release,
the action date, rather than the release date, was used to assign the date of record.
Finally, a case may also include a “vacated judgment.” A vacated judgment indicates that a
previous judgment in the case was overturned. For these records, the action date corresponds to
the judgment being overturned, and the release date indicates when that judgment was vacated. If
an action date on a judgment corresponds to an action date of a vacated judgment that occurred
later in the same case, the original judgment was marked as vacated. Vacated eviction judgments
were not counted as evictions.
III.3.
Commercial Cases
Our data include evictions filed against households (residential cases) and businesses
(commercial cases). A case is marked “commercial” if at least one defendant was identified as a
commercial entity. Because our research examines eviction at the household-level, we excluded
commercial cases from our estimates. We did not distinguish between individual and commercial
plaintiffs, since both individual and commercial landlords evict residential households.
12
Fewer than 1% of total cases (N = 822 cases) had multiple judgments dates associated with the same case. When
this occurred, we selected the latest judgment date, it being most affiliated with the final outcome of the case.
13
We identified commercial cases in our data using two methods. First, where available, we used
variables provided in the raw data that identified commercial cases. This indicator was available
in 97.3% of cases in our data. We developed a master list of keywords to identify commercial
cases in data without an existing indicator. This list includes common words clearly and
consistently associated with businesses:
● Legal terms for businesses: e.g., Inc., Corporation, LLC
● Types of organizations: e.g., Church, School, College, Trust, Foundation
● Types of services: e.g., Laundry, Barbershop, Dentist, Financial
● Types of products: e.g., Pizza, Toys, Tires, Books, Goods
We added keywords to the master list as we encountered new ones. We then used regular
expressions to determine whether defendant names included any of the keywords from the
master list. We checked all data against the latest version of the master list each time it was
updated.
We validated this second method using regular expressions by testing it on records that already
had a commercial indicator. We found that the regular expressions procedure detected
approximately 75% of the cases marked by the commercial indicator (where comparison was
possible). This discrepancy was often due to situations in which the record appeared to list only
the owner name without an accompanying business name. We ultimately relied on the
commercial indicator and supplemented identification of commercial cases via the regular
expressions method where the indicator was not available (0.58% of cases in our data).
13
In total,
we identified and excluded 597,327 commercial cases in our data set; this amounted to 1.55% of
all eviction cases.
III.4.
Name Standardization
Some defendant and plaintiff names appeared multiple times within eviction court records.
However, the same name may not have always appeared in the same format. This can occur for
several reasons. Sometimes, only first and last names appeared on a record; other times, middle
names and initials were included. Additionally, entry errors are always present in administrative
data. Identifying name variations was essential for determining the number of defendants and
plaintiffs associated with each case and for identifying duplicate cases. To accomplish this, we
developed a protocol for creating a standardized version of names.
Plaintiff Names
13
We were unable to perform this procedure for Pennsylvania Court Data, which did not contain defendant names.
14
Plaintiffs listed as businesses can have non-uniform name entries. For example, a plaintiff may
appear in the data as both “SUNNYSIDE APARTMENTS” and “SUNNYSIDE APARTMENTS
LLC.” Accordingly, we removed common abbreviations for commercial entities, e.g., “LLC,”
“INC,” and “CORP,” using regular expressions. We also removed common symbols and words
that may have otherwise generated multiple entries for the same plaintiffs (e.g., “&”, “AND”).
We also standardized spellings and abbreviations for other words that were common to many
plaintiff names. For example, the word “MANAGEMENT” may appear as “MGMT”,
“MANAGMNT”, or “MANGAMENT.” For these cases, we created flexible search strings using
regular expressions to find such variations and replace them with “MANAGEMENT.” The
following is an example:
regexr(plaintiff_name, "(^| )(MANAGE[MENT ]+|M[GMN][GMN]?T)(
|$)", “MANAGEMENT”).
Plaintiff names were then segmented into separate pieces using spaces (“ ”) as the delimiter. All
name pieces were compared within the same street address and then within the same case
number. If all name segments of a plaintiff name were contained within another plaintiff name
appearing at the same street address or within the same case, these entries were assigned the
same standardized name. Public Housing Authorities and University Regents were particularly
likely to enter the data in many versions in areas with concentrations of subsidized or university
housing, respectively. Local Housing Authorities were identified by searching for the occurrence
of both “HOUS” and “AUTH” in plaintiff names; University Regents were identified using
“REGENTS.” Both of these entities were then standardized within the same cases and addresses.
Finally, plaintiff names were compared within cases and addresses for minor variations in
spelling. We calculated the Levenshtein distance between each pairwise combination of plaintiff
names. Levenshtein distance represents the number of edits necessary to make one string match a
second string. Edits can be deletions, insertions, or substitutions of any character that appears in
the string. We used a threshold of two or fewer edits to determine that two versions of a name
matched and should be assigned the same standardized name.
14
Defendant Names
We also searched for common spelling and format variations to check for duplicates and ensure
consistency across defendant names. Initially our data had defendant names contained in two
variables, one for the full defendant name and another for the name suffix (e.g., Jr, Sr, III).
Defendant names were formatted as LASTNAME, FIRSTNAME MI (if present). We separated
14
We implemented the additional requirement that names be at least seven characters long to be determined a
match. This requirement was set to ensure that very short plaintiff names were not erroneously linked.
15
the names into three components: first name, middle name or initial (if present), and last name.
We kept the name suffix separate.
15
Like plaintiff names, we compared each pairwise combination of defendant names within the
same case numbers and street addresses. The following rules were used to determine if two
variations of name represented the same defendant with the same case number:
1. Exact match of first name, middle name, and name suffix, Levenshtein edit distance of
two or fewer characters in the last name, when the last name was longer than two
characters. This rule was used to catch minor spelling differences in last name.
2. Exact match of middle name, last name, and name suffix, Levenshtein edit distance of
one character or less in the first name, when the first name was longer than one character.
This rule was used to catch minor spelling differences in first names.
3. Exact match of middle name, last name, and name suffix, one version of the first name is
contained within the other. This rule was used to catch nicknames or initials.
4. Exact match of first name and name suffix, one or less character Levenshtein edit
distance in the last name and the two middle name fields start with the same letter. This
identifies cases in which the full middle name was used instead of just the middle initial.
5. Exact match of first name and name suffix, the middle and last names are combined in
one instance and separated in the other. This rule catches differences that occur when a
second last name is entered in court records.
6. Exact match of first name and name suffix, one character or less Levenshtein edit
distance in the last names and Levenshtein edit distance of one character in the middle
name, when the middle name is longer than one character. This catches minor spelling
variations in middle name.
7. Exact match of first name and name suffix, one or less character Levenshtein edit
distance in the last name and one version missing the middle name. Sometimes the
middle name is entered for a defendant; other times it is not.
These same rules were used to compare defendant names within the same street address, with the
exception that Levenshtein edit distance between last names could only be one character or less,
15
There were very few instances in which part of a defendant’s first name was erroneously split into the middle
name field due to a presence of a space. These were corrected by comparing concatenated first and middle names
within cases. If the concatenated first and middle names matched another first name on the same case (and the last
names and name suffixes matched), the first name and middle name fields were rejoined.
16
rather than two characters or fewer. This change accounted for the increased likelihood of two
distinct defendants with similar names located at the same address over time as compared to
being contained in the same case.
Each defendant name within the same case number or street address was compared twice.
Comparing names twice allowed for standardized versions of names to be updated in sequence,
and was more precise than creating less restrictive standards for matching two versions of a
name. For example, the following three names may appear at “123 MAIN STREET,
ANYWHERE, US”: “BUDDY HOLLY”, “BUDDY HOLY”, and “BUDDDY HOLY.” In the
first iteration, “BUDDY HOLY” will be standardized to “BUDDY HOLLY” and “BUDDDY
HOLY” will be standardized to “BUDDY HOLY”. On the second pass through, the standardized
name for “BUDDDY HOLY” will be updated once more from “BUDDY HOLY” to “BUDDY
HOLLY,” resulting in standardization across names within addresses. In total, 2.7% of names in
our data were updated during these standardization procedures.
III.5.
Preparing Addresses for Geocoding
To pinpoint the location of an eviction, we geocoded the defendant address associated with each
court case in our data. Cleaning addresses was essential to the success rate of our geocoding
process. Geocoding involves matching street addresses against a standardized data set of street
addresses and locations then assigning latitudes and longitudes to each address. Extraneous
information (e.g., name of apartment complex, unit number), inconsistent formatting, and
spelling errors can impede geocoding.
Some cases listed multiple addresses for defendants. In some cases, different addresses were
associated with different defendants on the case. In other cases, multiple addresses were listed
for the same defendant. This may occur if new addresses were entered for defendants after they
left the landlord’s property but the case was still ongoing (e.g., the defendant may still be ordered
to pay the landlord back rent). In 1.8% of cases from LexisNexis, a single eviction case
contained multiple defendant addresses.
For this reason, an address associated with a new case filing, forcible detainer judgment
(judgment for restitution of premises to the landlord), or case dismissal is most likely to
correspond to the property from which the landlord sought to evict the tenant. In the LexisNexis
data, 67.7% of case addresses were associated with filing records, 14.6% were associated with
forcible detainer judgments, and 3.4% were associated with case dismissals. Only 14.3% of
addresses were taken from entries associated with monetary judgments. In 0.4% of cases, no
address was present for the defendant(s).
17
If multiple addresses were listed, we used the following criteria in the order presented to select
the representative address for a case:
1. We selected the address associated with the original eviction filing, if present.
2. We selected the address associated with the forcible detainer judgment, if present.
3. We selected the address associated with the earliest case action, with preference given to
non-missing, in-state addresses. If multiple addresses were associated with the earliest
case action (and were non-missing and in-state), we selected the most common address.
If further delineation was necessary, we gave priority to the addresses that were able to be
geocoded.
We used this protocol to select the address with the highest likelihood of representing the
location where the eviction occurred.
Due to anonymized records and lack of filing dates, we performed slightly different procedures
in the Pennsylvania court data (except in Philadelphia County). Within a case with multiple
addresses, we prioritized in-state non-missing addresses and randomly selected an address from
these candidates. This procedure affected 1.7% of the total cases. In the Philadelphia state court
data, only one address is associated with each case, so a procedure for selecting addresses was
not needed.
A primary goal of the cleaning process was to ensure addresses were accurately split into five
distinct fields: street, apartment, city, state, and 5-digit zip code. Some data arrived already split
into these fields; other data were often incorrectly or incompletely split. We standardized all
address data by capitalizing alphabetic characters, removing trailing and leading spaces, and
reducing spaces within a string to one space each (e.g., “123 Main St” became “123 MAIN
ST”).
We employed regular expressions to search for two capital letters followed by a 5-digit numeric
sequence to detect cases where city/state/zip data was housed in the street or unit fields. We then
employed additional regular expressions to place text within the appropriate field. Additionally,
we used regular expressions to search for the presence of numbers to check the integrity of the
city field. Numbers indicated either a nonsensical value or the presence of street, unit, or zip code
information within the city field. We also removed punctuation, with the exception of
apostrophes. We ensured that all state values consisted solely of two letters (e.g., AK, AL) and
that all letter combinations matched a legitimate U.S. state abbreviation. Cases that did not, but
where the true state could be identified (e.g., a transposition of letters), were spot corrected. Zip
code cleaning consisted of verifying that the zip code field contained either 5 digits in a row or 5
18
digits followed by a dash and 4 additional digits. Non-numeric information was removed from
this field. We left partial zip codes unchanged and shortened 9-digit zip codes to 5 digits.
III.6.
Separating Street Address from Apartment Designation
We employed regular expressions to remove unit information from the street field and place it in
the unit field. We searched for unit designators, such as “unit,” “apartment,” “room,” “suite.”
Below is an example of a regular expression used for this purpose:
regexm(street,"(APA?R?TM?E?N?T?|GARAGE|SPACE|SU?I?TES?|RO?O
?M|NO\.?|TRAILER) (# ?|NO\.? ?|NUMBER ?)?([0-9A-Z\-]+)”).
We also corrected cases where the data populating the street and unit fields were flipped, i.e.,
valid street address information ended up in the unit field or vice versa. We again used regular
expressions to search for cases where the unit field began with a set of numbers followed by
alphabetic characters and the street field did not begin with a numeric value. This pattern usually
indicated the need for a street-unit swap. We performed spot checks and additional customization
to refine the accuracy of the code.
Next, we used regular expressions to search for Post Office Box (PO Box) information within
any address field. We removed PO Box numbers and placed them in a separate field to prevent
erroneous matches to numeric street names during geocoding.
Additionally, we employed regular expressions to standardize street field abbreviations.
Specifically, we replaced abbreviated versions of the following street designations with the full
word: avenue, boulevard, circle, country road, court, canyon, drive, expressway, glen, highway,
lane, manor, place, parkway, road, rural route, state route, state highway, state route, state
road, square, street, terrace, trail, valley, way. We also expanded numeric suffixes like 1
st
and
2
nd
to read First and Second.
III.7.
City, State, and Zip Code Cleaning
We cleaned city names, state abbreviations, and five-digit zip codes associated with eviction
court records via an automated comparison to a standardized listing of city names and state
abbreviations associated with U.S. zip codes.
16
16
We did not apply this procedure on defendant addresses in the Philadelphia Court data. The original data we
received had already undergone an address-cleaning procedure, in which addresses were cleaned and normalized
19
We conducted this standardized comparison listing using two sources: Zip Codes To Go and Zip
Code Tabulation Areas. Zip Codes To Go (www.zipcodestogo.com) maintains an up-to-date
listing of all zip codes used by the United States Postal Service, along with the preferred city
name and state abbreviation associated with each zip code. The U.S. Census provides Zip Code
Tabulation Areas (ZCTAs) that roughly align with USPS zip codes.
17
We combined the listings
of city names, state abbreviations, and zip codes from both sources and created a listing of valid
city name, state abbreviation, and zip code combinations in the United States.
We then compared each city name, state abbreviation, and zip code in the eviction records to the
standardized listing. We used the MatchIt program in Stata to find the best standardized match
for each record. MatchIt is a user-written program that allows inexact (or “fuzzy”) matching of
two sets of string (or text) variables. We used bigram vectorial decomposition distance to match
combined city name, state abbreviation and zip code variables to the standardized listing. Bigram
vectorial decomposition creates a similarity measure for two strings by breaking each string into
two-character pieces (bigrams) and then calculating how many of those bigrams are shared by
both strings. This procedure allowed us to correct data-entry errors in city names, state
abbreviations, and zip codes. The bigram vectorial decomposition algorithm created a similarity
score for each set of strings in each of the court records and standardized listings.
After running the matching algorithm, we retained the standardized city, state, and zip code
listing with the highest similarity score for each eviction record. We retained all standardized
candidates if multiple matches tied with the highest similarity score. The highest possible
similarity score was 1. A score of 1 indicated that a city name, state abbreviation, and zip code
from an address in the eviction records perfectly matched a city, state, and zip listing in the
standardized data set. Matches with a similarity score of 1 (95.6% of all defendant addresses)
were marked as correct matches.
Because the matches between the eviction record addresses and the standardized listing of cities,
states, and zip codes were inexact, some of the suggested standardized matches with similarity
scores below 1 were incorrect. We used the following criteria to classify as correct standardized
listing matches with similarities scores below 1:
using the Google Maps Geocoding API. Nearly all of the cases (99%) were assigned a recognizable address.
Because of the high success rate in this address-cleaning process, we did not apply the city, state, and zipcode
cleaning procedures in the Philadelphia Court data.
17
ZCTAs do not align perfectly with USPS postal codes. ZCTAs are used for estimation of demographic statistics,
as opposed to location identification for mail delivery and reflect other U.S. Census geographical boundaries (e.g.,
Census blocks) that are not considered in USPS assignment or realignment. ZCTAs may also include historical zip
codes as well as secondary city names associated with a zip code in some areas.
20
1. A bigram string similarity score greater than 70% with a perfect match for the city name
and zip code. In these cases, only the state abbreviation in the eviction record differed.
2. A bigram string similarity score greater than 70% with a perfect match for the state
abbreviation and zip code. These matches represented misspellings in city names.
18
3. A bigram string similarity score greater than 70% with a perfect zip code match. These
matches represented minor misspellings in city names and incorrect state abbreviations.
4. A bigram string similarity score greater than 75% with a perfect match for city name,
state abbreviation, and the final two digits of the zip code but a mismatch of the first digit
of the zip code. Here, the only unmatched character was the first digit of the zip code.
Owing to assignment of zip codes across the United States, first digits of zip codes are
typically consistent with states.
5. A bigram string similarity score greater than 75% with a perfect match for city name,
state abbreviation, and all but the second digit of the zip code.
6. A bigram string similarity score greater than 75% with a perfect match for city name,
state abbreviation, and all but the third digit of the zip code. In some places, this may
have represented a re-categorization of zip codes in an area over time.
7. A bigram string similarity score greater than 75% with a perfect match for city name,
state abbreviation, and all but the fourth digit of the zip code.
8. A bigram string similarity score greater than 75%, with a perfect match for city name,
state abbreviation, with the second and third digits of the zip codes flipped.
9. A bigram string similarity score greater than 75%, with a perfect match for city name,
state abbreviation, with the third and fourth digits of the zip codes flipped.
10. A bigram string similarity score greater than 75%, with a perfect match for city name,
state abbreviation, with the fourth and fifth digits of the zip codes flipped.
If the matches between the address records and standardized listing were marked as correct, we
updated the cleaned address fields from the eviction records with the standardized cities, state
abbreviations, and zip codes. City names, state abbreviations, and zip codes were updated for
18
We performed initial checks during the matching process to ensure that a 70% similarity score provided a
reasonable basis for this assumption. This score is conservative and misses some of the more egregious errors in
spelling of city names and abbreviations.
21
1.75% of defendant addresses using these criteria. Standardized matches that did not meet any of
the above criteria were not used to update the eviction address variables (2.70% of addresses).
19
We developed the above criteria to maximize the number of true errors corrected without
introducing changes that would distort the original address. We tested these thresholds on
address data from LexisNexis records in three states in different regions of the county – Alaska,
California, and Georgia – to ensure that they met our standard of correcting without distorting
the data.
III.8.
Geocoding and Linking to Census Geographies
After we completed our address cleaning process, we submitted the files to Environmental
Systems Research Institute (ESRI), for geocoding and linking to U.S. Census geographies.
Geocoding assigned latitude (Y) and longitude (X) coordinates to addresses. These coordinates
were based on spatial reference points in a locator file of U.S. street addresses. When possible,
each address was linked to a standardized address in the locator file with the highest possible
match score. The higher the match score, the greater the similarity between the text of the
original address and that of the official address included in the locator file.
Addresses can be matched to latitude and longitude coordinates at varying levels of geographic
specificity. The level of geographic specificity to which an address can be linked depends on
how much information is included in the address, how closely it matches an address included in
the locator file, and how accurately the locator file represents all possible addresses within a
geographic area. We undertook the address cleaning steps described above to maximize the
number of usable pieces of information in each address. ESRI used the most up-to-date locator
files (published in 2016) to geocode all defendant addresses in our data set (N = 64,700,554).
20
The overall accuracy of locator files tends to increase over time, as U.S. spatial mapping is
continuously improved.
Each address was matched to coordinates at the most specific level of geography possible. ESRI
used an iterative geocoding process, designed to achieve the most specific match possible for an
address. First, match attempts were made at the point address-level. If no standard address in the
locator file satisfied the minimum match score, a match was attempted at the street address-level.
19
We also compared city names and states abbreviations in cases where a match could not be found for city, state,
and zip code. We used this measure to confirm the existence of a city within a state. Of the addresses that were not
matched to a standardized city name, state, and zip code, we verified that 72.2% registered a valid city and state
combination.
20
An alternative strategy would have been to use versions of locator files that were published closer to the date of an
eviction record. We choose to use only the 2016 locator files for all addresses, however, after discussing the
improvement of accuracy in locator files over time directly with ESRI.
22
If no street address-level match was possible, a match was attempted at the street name-level,
and so on.
Each level of geography has a minimum similarity score that must be met to assign a match.
Table 6 lists the levels of geographic specificity, from most to least specific, along with the
minimum required match score.
Table 6. Levels of Geographic Specificity in ESRI Geocoding
Match Type Description Min. Match
Score
Point Address The full street address, including house number, was
matched with a standard address appearing in the locator
file.
93%
Street Address The street address was placed within a valid range of
house numbers on the street. The range of house numbers
was narrow enough to place an address on the correct side
of a street within a typical street block.
85%
Street Name The house number could not be matched to a specific
point on a street and coordinates were only be assigned for
a central point (centroid) along the street.
85%
Postal Only the zip code was used to assign coordinates; these
coordinates reflect only the central point (centroid) of the
zip code.
100%
Administrative
Place
Only the place name, usually a city or a municipality
name, was used to assign coordinates; these coordinates
reflect only the central point (centroid) of the
administrative unit.
96%
The level of geographic specificity of a geocoded address is important because it affects our
ability to count the number of evictions that occurred within small areal units, such as Census
block groups. Geocodes at the point- and street address-levels are assigned coordinates that
represent the exact or a very close approximation of the location of a street address, rather than a
23
centroid point associated with the street name, zip code, or other administrative boundary. At the
point- and street-levels of specificity, we can be sure that an assigned Census block group
reflects that of the address. As shown in Table 7, of the 64,700,554 records with available
addresses, 93.4% were geocoded at either the point address or street address-level.
After an address was assigned latitude and longitude coordinates, its location could be pinpointed
on a map and linked to other geographical areal units, such as Census tracts. We assigned each
address both Census block group FIPS codes and Census place FIPS codes using Shapefiles
provided by the U.S. Census Bureau. We assigned addresses to the Census block group or place
that contained their latitudinal and longitudinal coordinates. We chose to use Census block group
as the smallest areal unit because it is the lowest geographic level at which the U.S. Census
reports income data. Census tract and county FIPS identifiers can then be directly inferred from
Census block group FIPS codes.
21
For each geocoded address, then, we were able to potentially
identify the Census block group, tract, place, county, and state where the eviction occurred. This
geographic information allowed us to link eviction rates to demographic characteristics in a local
area.
The 93.4% of addresses geocoded at the point- or street-level were assigned Census geography
identifiers based on the above procedures. We reassigned the 6.61% of addresses that could not
be geocoded at the point- or street address-levels to Census block groups using spatial
imputation. Addresses that geocoded to out-of-state locations were also re-assigned to Census
block groups within the county where the case was heard.
22
This occurred in 26,403 cases. These
addresses were aggregated within counties then randomly assigned to a Census block group
within the county using weighted probability proportional to its share of renting households
within the county. Weighted probabilities were adjusted to reflect the number of eviction cases
already positively pinpointed to households in that block group in the same year. Addresses were
assigned iteratively and the weighted probabilities were updated after each assignment.
Table 7. Geocoding Outcomes
Geocoding
Outcome
Match Types Number of
Records
Percent of
Total
Address coordinates Point, Street 60,413,361 93.7%
21
Census blocks boundaries are nested within Census tracts, which are nested within counties. Place FIPS codes had
to be assigned separately as their geographical boundaries do not perfectly align with block groups, tracts, or
counties.
22
Out-of-state cases were reassigned in this way as the address listed on the court record likely did not represent the
location which the eviction was threatened or occurred. This address could have instead represented an alternative
residence for the defendant or the new residence a defendant moved to after being evicted.
24
Centroid coordinates
Street Name, Postal,
Administrative Places
4,277,614
6.6%
Unmatched
N/A
9,579
<0.1%
No Address present
N/A
160,043
0.2%
III.9.
Address Standardization
Even after separating apartment designations and cleaning city names, state abbreviations, and
zip codes, some addresses still appeared in multiple variations in the data. To aid in identifying
duplicated records and cases reappearing for the same tenant at the same address, multiple
versions of the same address were assigned a standardized form within cases with the same
defendant. This occurred when one or more of the following four patterns were observed.
1. The street address, state, and zip code of a defendant’s address matched, but one or more
entries had a different city name. In many of these cases, otherwise identical address
records listed different neighboring city names. For example, “123 MAIN STREET,
HENRICO, VA 55555” and “123 MAIN STREET, RICHMOND, VA 55555.” When this
was observed, all entries were updated to share the same city name. In choosing this
standardized city name, preference was given according to the following ordered criteria
(e.g., if not a, then b; if not b, then c):
a. We selected the city name associated with the address version with the most
geographically specific level of geocoding.
b. We selected the city name associated with the address version that was able to be
geocoded at any level of geography.
c. We selected the city name associated with the address version that was verified as
a valid city name, state, and zip code combination in the city, state, and zip code
cleaning.
d. We selected the city name associated with the most common version of the
address.
2. The street address, state, and city name of a defendant’s address matched, but one or
more entries had a different zip code. As with variations in city names, this commonly
occurred in bordering zip codes, and all otherwise similar addresses were updated to
25
share the same zip code. The same ordered criteria was used to assign the consistent zip
code.
a. We selected the zip code associated with the address version with the most
geographically specific level of geocoding.
b. We selected the zip code associated with the address version that was able to be
geocoded at any level of geography.
c. We selected the zip code associated with the address version that was verified as a
valid city name, state, and zip code combination in the city, state, and zip code
cleaning.
d. We selected the zip code associated with the most common version of the address.
3. The apartment designation in an address was appended to the street number: e.g., “123
MAIN STREET, APT 2B, ANYWHERE, US 55555” to “1232B MAIN STREET,
ANYWHERE, US 55555.” We identified cases these by searching for instances in which
both the street and apartment numbers associated with one address appeared within the
street number portion of another address within the same defendant name. It was
important to identify these cases because addresses with apartment designations were
often more difficult to geocode. We updated incorrect versions using those with the
correct separation of address components.
4. We synchronized multiple addresses with a shared format if multiple addresses shared the
same street number and had street names that had either (1) a Levenshtein edit distance of
one character or less or (2) the containment of one street name within the other
(e.g.,“MAIN” and “MAINE”) within the same defendant name. For the updated version
of the address, ordered preference was given via the following criteria:
a. We selected the address with the most specific level of geocoding.
b. We selected the address able to be geocoded.
c. We selected the most common version of the address.
As addresses were updated according to these criteria, they were assigned the geocoded latitude
and longitude coordinates and Census geography identifiers consistent with the version of the
address used to do the updating.
As a final step, we created a standardized set of address variables – street address, city, state
abbreviation, and zip code – for all address records. Those able to be assigned to point- or street-
level locations were updated using the matched address provided by the ESRI geocode. For
addresses geocoded at centroid points within streets, cities, or zip codes, we used the cleaned and
26
reconciled address variables to populate these fields. We updated 3.3% of the address records in
our data during this standardization process.
23
III.10.
Duplicate Records
We excluded duplicate records in our estimates to avoid over-counting. First, we dropped perfect
duplicates (scenarios when the values of all variables were the same).
24
Then, we excluded
duplicate records when all key variables – full defendant name, street address, city, state, zip,
action date, action type, and judgment outcome and amount – contained the same value.
25
We
considered these cases to reflect data-entry errors in which multiple records referred to the same
case.
Some of these issues stemmed from human error, while others were the result of differing
systems implemented by county courts throughout the state. To standardize case numbers, we
used regular expressions, removing extra white spaces in the numeric entries. Table 8 lists the
numbers, as explained above, that were dropped or excluded from our data. These records were
neither used nor included in our estimates.
Table 8. Duplicates Dropped in All Eviction Records
Duplicates Dropped Duplicates Marked and Excluded
8,590 697,188
III.11.
Serial Cases
A landlord may file multiple evictions against the same household. These “serial cases” differ
from duplicates in that they reflect multiple (real) eviction filings by a landlord on the same
23
We were unable to conduct the first four checks in this section on the Pennsylvania court data due to lack of
defendant names. We were able to create standardized address fields using the geocoded matched addresses
geocoded at the point- and street-level. The cleaned address fields were used in the standardized variables for those
addresses only able to be geocoded at areal centroids.
24
We were unable to fully perform these steps with Pennsylvania court data, which did not include defendant
names, plaintiff names, or case numbers. We were only able to look for perfect (not near perfect) duplicates in these
data.
25
We cannot conduct these procedures for records missing defendant last name, street name, and/or anonymous
defendants.
27
tenant. However, given that evicted tenants are forcibly removed from their homes, multiple
repeated filings against within a short period of time suggest the tenant did not in fact vacate the
premises. As a result, including all such cases would have overestimated the actual number of
evictions. We first observed serial cases in eviction data received from South Carolina. To ensure
these cases were not the result of administrative error, we spoke with court clerks and several
legal aid representatives from South Carolina Legal Services and South Carolina Appleseed
Legal Justice Center, all of whom confirmed the prevalence of serial cases.
26
To properly identify and categorize serial cases for the purpose of counting evictions, we first
standardized names and addresses using the criteria discussed above and sorted observations by
defendant name, address, and action date. Next, we grouped cases sharing the same defendant
name and address. These groupings of cases were categorized as one set of serial cases. We were
unable to identify serial cases if names were missing or anonymized (e.g., Jane Doe, Unknown
Tenant), or if an address was not listed for a tenant. The number of cases that were identified as
the final case of a series of cases with the same defendant at the same property was 5,461,151.
Next, we referred to the outcome of the most recent case within a set of serial cases to determine
whether an eviction appeared to have taken place. If it did, we counted one eviction for the
whole set of serial cases. If the last case did not result in an eviction outcome, no eviction was
recorded for the set of serial cases. If a set of serial cases included multiple eviction judgments
and the last case in the set resulted in such a judgment, we counted that as a single eviction. If a
set of serial cases included multiple eviction judgments and the last case in the set did not result
in an eviction judgment, then no eviction was recorded for that set. Because serial cases involved
multiple (real) filings, each filing was recorded in our final count of active eviction cases in a
jurisdiction.
For example, if a landlord took the same tenant at the same address to court three times in three
consecutive months – receiving an eviction judgment, case dismissal, then another eviction
judgment (in that order) – we would record three filings and one eviction. If a landlord took the
same tenant at the same address to court three times in three consecutive months – receiving an
eviction judgment, a case dismissal, and another case dismissal (in that order) – we would record
three filings and zero evictions.
26
We identified serial cases in all data except Pennsylvania’s court data (excluding Philadelphia County), as the
state’s bulk data collection rules prevented us from accessing defendant and plaintiff names.
28
IV.
Estimating the Prevalence of Eviction
IV.1.
Classifying Eviction
In our data, we classified a case outcome as an eviction if we could see that it contained a record
of either a money or forcible detainer judgment for the plaintiff.
27
A money judgment awards the
plaintiff some amount of money that is owed by the defendant.
28
A forcible detainer judgment
grants possession of the property to the landlord. As discussed in Section III.2, judgments that
were later vacated were not counted as evictions. Evictions in New Jersey and Philadelphia were
defined differently, as detailed below.
We employed additional information on case outcomes in New Jersey using data from AIRS.
The AIRS data contained individual-level court records for eviction cases, including judgment
information. There was significant overlap between the cases contained in the AIRS data and
cases contained in our data. We used court identifier, case number, and defendant name to
perform a one-to-one merge linking our data to the AIRS data. We were able to update case
outcomes in 3,431,883 records, representing 88.4% of records in New Jersey for the 2000-2016
period. Judgment data from AIRS indicated which party (plaintiff or defendant) a judgment
favored and whether restitution of the premises was part of the judgment.
Philadelphia court data came with its own set of outcome codes. In that data, we counted a
judgment as an eviction outcome if any of the following took place: (1) the defendant lost the
case in a court trial; (2) the defendant lost the case by default; or (3) there was a stipulation and a
writ of possession was granted as part of the judgment.
IV.2.
Estimating Case Volume Fluctuation
Sometimes, recorded eviction data are not uniform across space or time. In particular, there were
areas where the volume of cases collected was substantially lower than expected, based on
demographic variables and state court-reported volume. For this reason, not all of the records
27
We did not separate settlements and stipulations from money or forcible detainer judgements.
28
There may be situations in which a tenant is the plaintiff and the landlord is the defendant in a landlord/tenant
dispute. For example, a tenant may sue their landlord for harassment and be award a monetary settlement. We
cannot directly observe which party is the tenant and which in the landlord in our data. We checked for any
situations in which a plaintiff listed on the initial case filing later became a defendant in the same case by looking for
instances in which the parties appeared in the reverse order in any subsequent case actions. We only identified a
handful of situations in which party names appeared in opposite fields, and these appeared to be data irregularities
rather than a reversal of plaintiffs and defendants (e.g., the same name was listed as both the defendant and the
plaintiff on a case).
29
collected for the 2000-2016 period were included on the Eviction Lab map. Our goal was to
provide the most information possible about the spatial locations of eviction in America without
presenting statistics that misrepresented the picture of eviction in a given area.
To help determine where coverage of eviction records was most consistent, we compared our
data to eviction case volume statistics received directly from local and state courts. We received
these data at the county-level for 27 states, New York City, and the District of Columbia, as
discussed in Section II.2. We first calculated expected case rates and case volumes from year-to-
year in the state-reported court data. We then compared these numbers with our case counts at
the county-level. We used the county where the case was filed in this comparison, not the county
where the premise address was geocoded.
We assigned each county to a broad housing market to better understand expected rates and
changes in eviction case volume year-to-year in local areas. We used three different sets of
characteristics to categorize housing markets: (1) Urban or Rural; (2) Northern or Southern; and
(3) High- or Low-Renter Household Population.
Urban or Rural: We distinguished urban and rural counties using the U.S. Census
Metropolitan Statistical Area (MSA) delineation files. For years 2007 or before, we used
the 2007 MSA delineation; for years after 2007, we used the 2015 MSA delineation.
29
A
county was marked “rural” if it was not included in any MSA, or if it was an “Outlying”
county in an MSA.
Northern or Southern: After recognizing regional effects in housing market behavior
(particularly in eviction rates), we assigned a dummy variable to distinguish Northern and
Southern states.
High- or Low-Renter Household Population: We designated counties having “high-
renter populations” if its number of occupied renter households was at or above the
nationwide median. “Low-renter populations” were those with renter household
populations below the median.
We calculated the annual expected percentage increase and expected percentage decrease in case
volume for each housing market using state-reported statistics. We compared these values to the
year-to-year percentage change in case volume in our data.
29
U.S. Census Bureau (2007). Core Based Statistical Areas (CBSAs) And Combined Statistical Areas (CSAs), July
2007; U.S. Census Bureau (2015). Core Based Statistical Areas (CBSAs), Metropolitan Divisions, And Combined
Statistical Areas (CSAs), July 2015.
30
We created four indicators based on these comparisons. The first marked the first year when data
were available in a county. The second marked the last year data was available in a county. The
third indicator marked years when case volume was substantially lower than the preceding year.
The fourth marked years when case volume was substantially higher than the preceding year.
It is difficult to accurately distinguish between expected and unexpected changes when the
number of cases is small. Accordingly, we did not include indicators or adjust case volume in
counties that (1) were in the first quartile of renting households in the United States or (2) either
had an average of fewer than 50 eviction cases per year or never had more than 100 eviction
cases in any year.
All remaining counties were evaluated with the four indicators described above to estimate how
case volume fluctuated over time.
We marked a year as having an unexpected increase in the case volume if it met all the following
criteria:
1. The case volume increased by two or more standard deviations from the average
observed increase in the same housing market.
2. The case volume increased 100% or more from the previous year.
3. There were at least ten cases recorded in the county in the previous year.
We marked a year as having an unexpected decrease in the case volume if it met all the
following criteria:
1. The case volume decreased by two or more standard deviations from the average
observed decrease in the same housing market.
2. The case volume had a decrease of at least 50% from the previous year.
3. There were at least ten cases recorded in the county in the previous year.
After marking an unexpected decrease, we also noted whether the data ever recovered. A
recovery occurred when the case volume reached at least 75% of the volume in the year before
the unexpected decrease occurred.
Marking First Available Year
31
All counties began evaluation in 2000. We then sequentially excluded years according to the
following set of criteria:
1. There was an unexpected increase in case volume the next year.
2. The data was low, relative to the average case volume. This was measured as an average
case volume less than 75% of the standard deviation of case volume over time in that
county.
3. A year recorded zero cases, and the following year recorded at least 10 cases.
4. A year recorded fewer than 10 cases, and the following year recorded zero cases. This
was used to catch instances in our data when a handful of cases was collected in a year,
but the long-term trend in the beginning years reflected a lack of overall coverage.
5. There was a 1,000% or greater increase in case volume in a future year. This detected
counties that had very low case volume in early years, followed by a later increase in case
volume that could only be attributed to a significant lack of complete collection coverage
in the earlier years.
Marking Last Available Year
Under a similar strategy, we began a second evaluation in 2016 and worked backwards to mark
sequentially previous years as unavailable according to the following criteria:
1.
There was an unexpected decrease in case volume compared to the previous year, and it
was not a result of a temporary spike in case volume.
30
2.
Case volume was less than 75% of the standard deviation of average case volume over
time in a given county.
3.
There were no cases recorded in that county.
4.
We observed fewer than 10 cases with zero cases in the preceding year. Again, this was
used to catch lack of overall coverage in later years.
30
A “temporary spike” in case volume was identified as an unexpected increase immediately followed by an
unexpected decrease.
32
5.
Case volume decreased by 90% or more in a preceding year. We used this to detect
counties with more consistent coverage in early years followed by incomplete coverage
in later years.
IV.3.
Imputation
Once the above trends were identified, we marked each year within a county to indicate whether
it represented a valid data point. Years before the first available year for data within a county
were not included in our data. Likewise, years after the last available year of data within a county
were excluded. One exception was made for data in 2016. If a county had available data for 2015
but was missing data in 2016, we pulled the value from 2015 forward to estimate projected case
volume in that county (N = 93). In these cases, we anticipate updating the 2016 numbers during
the next data delivery in 2018. These areas likely reflect counties that have adequate collection of
cases but were not fully updated when the data were delivered in September 2017.
31
After determining first and last available years, we marked years within the available period that
had inconsistent case volumes, relative to other years in that county, and thus were unlikely to
represent valid data points. We marked unavailable years sequentially until case volume
recovered.
Within counties, years that were not marked as valid data points between the first and last years
of availability were imputed if there were no more than two consecutive years of missing data.
When only one year of data was missing within a county between two years of valid data, the
case volume was imputed using the average of the preceding and following years. When two
consecutive years of data were missing, we linearly interpolated between the last known and
reliable value and the next known and reliable value. A total of 318 counties had at least one year
of data imputed at some point in the 2000-2016 period in accordance with these two procedures.
Some counties had data imputed in multiple years, resulting in a total of 519 imputed data points.
When three or more consecutive years of data were missing, we did not impute the data; thus,
these years were not included on the map.
Counties had to have at least two consecutive years of valid data to be included. Counties that
did not meet this requirement are shown as missing. In total, 171 counties are displayed as
missing for all years. The counties and years marked as missing are displayed as "Unavailable"
on the map. Four states – Alaska, Arkansas, South Dakota, and North Dakota – did not have
consistent data coverage at any point in the 2000-2016 period. For these states, we substituted
data on case volume from the state-reported court statistics, as discussed in Section II.2. No
counties in these states were included in the imputation procedure.
31
There is often a lag time between when a case is filed within a court and when it can be recorded.
33
For counties with data adjusted during the imputation, additional cases were assigned Census
block group FIPS codes using weighted random assignment. Here, we followed the re-
assignment procedure as described in Section III.8. Each Census block group within a county
was assigned a weighted probability consistent with its share of renter households. This weighted
probability was updated to adjust for the number of cases successfully geocoded in the block
group. Additional cases were then assigned randomly using these weights. Cases were assigned
to block groups iteratively. After a case was assigned, the block group weights were re-estimated
to account for this new assignment and then the next case was assigned. We followed this
procedure until all adjusted cases were assigned to block groups within the county.
IV.4.
Aggregation
Eviction cases in our data were aggregated yearly in Census block groups, Census tracts,
counties, Census places, states, and nationally across the United States. Once the data were
cleaned, each case was marked as representing an active eviction case (filing) and/or an eviction
judgment in the year that corresponded with the date of record assigned to the case. Cases
contributed to the count of active cases so long as (1) the property at dispute was residential
(rather than commercial) and (2) the case did not duplicate a previous case already represented in
the data. Cases were additionally counted as resulting in evictions if the case included one of the
judgments related to eviction criteria described above, and the judgment was not vacated at a
later date. If a case was included in the eviction count, it was also represented in the pool of
active eviction cases. There are no situations in which a case is marked as an eviction but not as
an active eviction case pending that year. There are many instances, however, in which an active
case did not contain a record of eviction.
32
As noted above, judgments on cases that recurred for
the same defendant at the same property were not counted as evictions unless they appeared as
the final action for a tenant at that property.
Although our data span 17 years, we standardized all geography to reflect 2010 Census
boundaries. This allows us to compare the prevalence of eviction cases and judgments within the
same area over time. As the Census geography boundaries are redrawn after each decennial
Census, particularly at the block group- and tract-levels, we could not compare these areas
directly without a standard set of geographical boundaries. For this reason, we had all geocoded
eviction cases matched to 2010 Census block group and place FIPS codes.
Aggregation then involved counting the number of active cases and eviction judgments within
each level of geography. Cases and judgments were first summed within Census block groups
32
Some of our data sources prioritized cases that ended in a civil judgment. For example, in some areas, dismissals
were not recorded, which deflated the total filings numbers. When this happens, filing and eviction rates are similar.
34
using Census FIPS codes, our smallest areal unit of geography. Counts were then aggregated into
Census tracts, as blocks groups are contained within Census tracts. Counts in Census tracts were
then summed within counties and counts within counties were aggregated to states. As a final
step, all counts within states were collected into an estimate of the number of cases and evictions
that occurred nationally each year between 2000 and 2016.
For aggregation we used the county derived from the geocoded address, rather than the county
representing the court the case was heard in. We used the geocoded county as this was the best
representation of the local area in which the case occurred. As mentioned in Section III.8, cases
that geocoded outside of the state the court was located in were reassigned to a Census block
group located in the county corresponding to where the case was heard.
Census places were aggregated separately, as their boundaries do not directly correspond to those
of Census block groups, Census tracts, or counties. However, Census place boundaries
correspond to Census block boundaries. Accordingly, after cases were assigned to a Census
block group, they were allocated proportionally into Census blocks within that block group using
household population as the weighting variable. We obtained allocation factors for each 2010 US
Census block group from the Missouri Census Data Center. After cases were allocated to Census
blocks, they were aggregated into 2010 Census places.
33
For this reason, the counts of renter
households, eviction cases, and evictions for places are often non-whole numbers (e.g. 4.5
evictions). On the map, these numbers are displayed as rounded whole numbers (e.g. 2.3 is
displayed as 3 and 4.5 is displayed as 5). All values less than 1 but greater than 0 are displayed as
1. The filing and eviction rates are calculated with the original, non-rounded values. We chose to
calculate the rates from the non-rounded values in order to not inflate our estimates. The original,
non-rounded values of the eviction case and eviction counts are available in the downloadable
files.
IV.5.
External Data Sources
33
Our estimates for Pennsylvania evictions and filings were derived from multiple sources: LexisNexis (2000-2006)
and state court data from 2007 onward, when Pennsylvania limited automated, bulk data collection. Because state
court data was anonymized and did not include case numbers (as explained above), we could not check for serial nor
duplicate cases in these data. Therefore, we used the available LexisNexis data to inform our post-2006 eviction
estimates. We calculated the average filing-to-eviction ratio per block group, based on LexisNexis’ Pennsylvania
records between 2000 and 2006; for block groups that were in the state court data, but not in the LexisNexis data, we
used the average filing-to-eviction ratio at the tract-level. We then multiplied this ratio to the state court data’s
aggregate case count, rounding down the estimates to the nearest integer, to get post-2006 eviction estimates at the
block group-level. We applied these procedures specifically for Pennsylvania because we had sufficient data from
LexisNexis – verified using published court statistics – to infer eviction counts in the state court data.
35
As mentioned previously, Alaska, Arkansas, North Dakota, and South Dakota estimates are
derived from sources outside of LexisNexis. To address this, our team compared the volume in
our individual-level data against the state-reported county statistics to determine an appropriate
adjustment to the county statistics volume.
34
We calculated the ratio of county-level case filings
found in LexisNexis data to the county-level case filings reported by the state. Then, using the
regions used to create the markets described in Section IV.2, we found that in the North, on
average, our estimates of eviction case filings were 72% of the volume reported by the state. In
the South, this figure was 71%. For estimation purposes, Alaska, North Dakota, and South
Dakota were classified as Northern, while Arkansas was categorized as Southern.
As a result, our estimates are the multiplication of the state-reported county case volume by its
corresponding market adjustment. For example, if Adams County, North Dakota reported 100
landlord tenant cases in 2016, our estimate of case volume would be 72 eviction filings that year.
Estimates found in the data download section and on the Eviction Lab map include these
adjustments.
IV.6.
Denominator
To calculate eviction and filing rates, we divided the number of evictions and filings by the
number of renter-occupied households in that area. We used estimates of renter-occupied
households from the 2000 and 2010 U.S. Censuses and ESRI Business Analyst 2016. We first
created renter-household estimates for Census block groups and then aggregated these numbers
to create estimates for Census tracts, counties, and states.
The boundaries of 2000 and 2010 Census block groups are not directly comparable. For this
reason, we used the Time Series Tables that aggregate 2000 Census demographics into 2010
Census geographies available from the National Historical Geographic Information System
(NHGIS).
35
We then created yearly estimates by using linear interpolation between the 2000 and
2010 data points. We used this same procedure to create yearly estimates between the 2010
Census and ESRI Business Analyst 2016.
36
34
While Pennsylvania state court data are used in our estimates for 2007-2016, we did not institute the same
penalties. Using court reported statistics for Pennsylvania, we found that the proportion of cases to reported filings
was comparable for 2000-2006 using LexisNexis and 2007-2016 using state court data.
35
Steven Manson, Jonathan Schroeder, David Van Riper, and Steven Ruggles. IPUMS National Historical
Geographic Information System: Version 12.0 [Database]. Minneapolis: University of Minnesota. 2017.
http://doi.org/10.18128/D050.V12.0
36
ESRI Business Analyst 2016 was standardized to 2014 Census geography, which is consistent with the 2010
Census at the block group-level.
36
To estimate the number of renter-households in Census places, we had to create separate
estimates. This time, we started with renter-households estimates from the 2000 Census at the
2010 Census block-level from NHGIS. Then we aggregated 2010 blocks to 2010 Census places.
We obtained 2010 Census place renter-household estimates directly from the Census. We
allocated the 2016 block group-level estimates from ESRI Business Analyst down to the block
group-level, again using files available from Missouri Census Data Center. These estimates were
aggregated into 2010 Census places as well. We used linear interpolation to create yearly
estimates between 2000 and 2010 and between 2010 and 2016.
V.
Validation of Estimates
We validated our data by gathering additional information on evictions, both at the individual-
and aggregate-levels. This included two validation methods at the individual-level and one at the
aggregate-level.
V.1.
State Data Merges
We merged and compared LexisNexis data with state court data as a validation measure at the
individual-level. These included records from Alabama, Hawaii, Iowa, Minnesota, Missouri,
Nebraska, South Carolina, Pennsylvania, and Virginia.
37
We used case numbers and court numbers, the combination of which uniquely identifies each
case, to merge LexisNexis data with state court data. We considered a case “matched” when two
or more of the following variables corresponded in the LexisNexis data and state court data:
defendant names, plaintiff names, court dates (filing dates, disposition dates, and/or judgment
dates) and defendant addresses. According to these criteria, over 99% of LexisNexis cases
accurately represent cases in the state court data.
Using the state court data, we also wanted to compare the outcomes found on the cases in
LexisNexis data to another source of data. We attempted to parse the text of the outcome and
case information to classify cases in our state court data. However, case disposition meanings
lack uniformity across states and sources. In these cases, we searched for additional information
contained in our data. For example, Nebraska, Philadelphia, and Virginia case records contained
fields indicating whether a writ of restitution was executed. A writ of restitution authorizes a
landlord to forcibly remove a tenant from the disputed property following a judgment for the
plaintiff. If a writ of restitution was executed, we considered the case to have resulted in eviction.
37
We did not perform this check in Pennsylvania because the date ranges of data available in both LexisNexis and
State Court Data did not overlap. We were only able to do so with Philadelphia County, PA, because we received
Philadelphia case records from a separate source.
37
Because information on writs was available in Nebraska and Virginia, we chose to compare data
from these two states against LexisNexis data. Nebraska court data specify whether a judgment
includes “Restitution of Premises,” and Virginia court data contain information on whether a
possession was at stake or if a writ of possession was ever issued.
Our comparisons showed that in Nebraska, 97.9% of cases marked as having an eviction
outcome in LexisNexis would also have been marked as such in the court data; in Virginia,
93.9% of cases marked as an eviction outcome in LexisNexis were marked as such in the court
data. In Nebraska, non-eviction outcomes in the LexisNexis data (0.02% were dismissals and
8.43% did not have further outcomes collected) were marked as evictions in the state data; in
Virginia, non-eviction outcomes in the LexisNexis data (0.23% were dismissals and 11.8% did
not have further outcomes collected) were marked as evictions in the state data.
V.2.
State-Reported County Statistics
To validate our estimates of eviction case volume, we compared our counts directly to another
source: the state-reported county-level statistics on eviction filings. In some jurisdictions, certain
counties were not reported in the state aggregate report; if we did not have state-reported
statistics for these areas, they were not included in our calculation.
38
Table 10 represents the average ratio of the number of cases within our data to the number of
cases reported by the state over all years with state-reported statistics at the state-level. This table
does not include comparisons for states where we substituted in the state-reported court statistics
to create our estimates of eviction filings. These ratios can vary within state and across years. For
instance, Massachusetts’ case volume in our data increases over time, so in later years our
estimates of eviction case filings are much closer to the reported volume by the state.
It is also important to note that we can only compare case volume for years and counties in
which we have additional data sources. In some states, while we believe we see low counts in
certain jurisdictions, we have no external numbers to validate. Data collection is ongoing – we
will continue to gather more resources to help us validate our national estimates of eviction
filings.
Table 9. Ratio of Aggregated Individual-Level Cases to County-Level Cases
38
Missouri lists out separate case types related to landlord/tenant. While two case types – rent and possession and
unlawful detainer – were reported across all jurisdictions, landlord/tenant cases were not reported uniformly. We
compared case volume for what we could see. As a result, we have more individual cases than are technically
publically reported by the state. This results in the proportion for Missouri being higher than 100%.
38
Region
Ratio
California
0.42
Colorado
0.65
District of Columbia
0.57
Delaware
0.81
Florida
0.83
Georgia
0.85
Hawaii
0.07
Massachusetts
0.60
Maryland
0.13
Maine
0.70
Michigan
0.72
Missouri
1.08
North Carolina
0.82
Nebraska
0.84
New Jersey
0.92
New Mexico
0.86
Nevada
0.74
New York City
0.37
39
Ohio
0.94
Pennsylvania
0.94
Texas
0.86
Utah
0.81
Virginia
0.90
Vermont
0.05
Wyoming
0.25
Note: For the 23 states listed above, the District of Columbia, and New York City, we compared the case volume of
aggregated individual-level eviction cases to case volumes reported by the states at the county-level. For example,
ratio of aggregated individual-level cases to county-level cases in Ohio is 0.94, meaning that there were 94
individual-level cases for every 100 reported by the state. New York City represents the five counties housed in New
York City: Kings, Queens, Bronx, Richmond, and New York.
V.3.
Low and High Rates
Low Rates
As noted in Section V.2 and on the Eviction Lab map, we have indicated states with
underestimated eviction counts. Some places have naturally low eviction rates because there are
few renters living there or because there are very few evictions. Still, some states and the District
of Columbia have underestimated eviction counts for a variety of other reasons.
In New York, records are often kept as “abstracted judgments,” meaning they are only in the
public record if the plaintiff/landlord pays to have them placed there. That, plus the amount of
town and village courts in the state, makes collection difficult.
In California, many cases that end in eviction are sealed and therefore not accessible by the
general public. Moreover, it can be difficult to collect data from California as a whole, owing to
restrictions on the number of records one can collect.
In most jurisdictions, the eviction process starts with an out-of-court notice delivered to a tenant,
but in Maryland the process begins with an eviction filed in court. This means Maryland has a
very high case volume. Because of that, it is difficult to collect data from Maryland, with the
40
exception of Prince George’s county where we have consistency in case volume over time.
(Thus, Maryland is in the ironic position of having an inflated eviction filing rate and an
underestimated eviction rate.)
In New Jersey, while the number of eviction cases was collected reliably, information about the
outcomes of those cases were not readily accessible. As a result, we believe our eviction rate
there is an underestimate.
In Kentucky, Louisiana, Tennessee, and Texas, there is robust data coverage in urban centers,
but there are other, more rural areas that are missing data, in light of collection difficulties.
After comparing our estimates to external data sources, we found that Hawaii, Connecticut,
Vermont, Wyoming, and the District of Columbia have low counts in the Eviction Lab map and
raw data, due either to the remoteness of some areas or to data collection difficulties. We also
suspect that the numbers in Arizona, Idaho, New Hampshire, and Washington may be too low.
To identify low case rates, we created an indicator that factored in renter households and external
statistics.
39
We also marked counties with low, or underestimated, eviction counts on our map. We used the
individual-level data collected directly from state courts (Section II.2.) and the state-reported
county statistics (Section V.2.) as a comparative external source of data. We assessed the breath
of coverage in counties each year by looking at the difference in the number of reported eviction
cases in our data compared to the external source. We applied the following criteria to determine
when case volume in a given county was low:
1.
The proportion of external cases represented in our data was less than 85%.
40
2.
The absolute difference in the number of cases between our data and state or county court
reported data was greater than 32 cases.
41
39
We generated external case rates using external court statistics, then found the tenth percentile of the external case
rates in each renter household quartile each year. We marked our case rate estimate in that year as “too low” if it fell
below the tenth percentile of the corresponding quartile. We placed a flag for a state if 25% or more counties in the
state were identified as “too low” between 2009 and 2016.
40
The difference in number of cases between our data and the external data source was right skewed (i.e. most
counties had low to moderate differences in case volume while a few outliers had more extreme differences). For
this reason, the average level of coverage across counties appeared to be less than 65% of cases. We then restricted
our comparison to only those counties that were within 50% coverage of the external case counts to account for this.
The median value for coverage after imposing these restrictions was 85%.
41
This represented the 75
th
percentile in the absolute difference in the number of cases in our data and the external
data when our coverage was greater than or equal to 85% of external data.
41
Counties that met both these criteria were marked as having a low case count on our map. In
many instances, the counties marked as having low case counts are located in the states discussed
above. However, coverage across counties also varies within states. In some cases, comparisons
to the external data showed that some counties were well represented in our data, even if the state
overall appeared to have low coverage. In other instances, we discovered that evictions appeared
to be underrepresented in a few counties within states that were well represented overall in the
data.
We used an alternative strategy to mark counties with low eviction case volume when direct,
external sources of data from state or county courts were not available. We started by calculating
“external” eviction case rates in all the counties were external data was available. We then
grouped these counties into quartiles based on the number of renting households. In each
quartile, we examined the distribution of external case rates to determine what a typical eviction
case rate should look like in counties with comparable numbers of renting households. We used
these distributions for the case volume comparison in the counties for which no direct, external
source of data was available. We identified counties with a case rate in the 20
th
percentile or
below of the external case rates within their renter household quartile each year as having a low
eviction case count.
42
Since it is possible that few evictions occurred in counties with low
numbers of renter-occupied households, we did not mark counties in the lowest (first) quartile of
renting households. Table 10 lists the percentage of counties in a given year that were either
flagged as low in our data or marked as “Unavailable” on our map. We also marked areas
embedded in counties (e.g., Census tracts, block groups) if their corresponding counties received
flags.
43
However, we only validated low counts at the county and state levels. Therefore, Census
tracts and block groups that received a flag due to their location in low count counties may not in
fact be undercounted.
Table 10. Percentage of Counties Marked as Having a Low Count of Cases or Unavailable Data, by Year
Year Not Marked Low Unavailable Total
2000
47.12
16.00
36.88
100
2001
57.46
18.80
23.74
100
42
We chose the 20
th
percentile threshold by calculating the median case rate in the counties that were marked low
when external data was available for direct comparison. The median case rate in those counties was consistent with
the 20
th
percentile of external case rates.
43
We also assigned flags to Census Designated Places (CDPs). Unlike Census tracts and block groups, CDPs are not
always fully contained within counties. Therefore, we used area allocation factor files from the Missouri Census
Data Center to assign CDPs (or parts of CDPs) to counties. When a CDP was contained entirely within one county,
as was the case for 96.6% of CDPs, we added a low count flag to the CDP when we had also added one to the
county. We added low count flags to the remaining 3.4% of CDPs if 15% or more of the CDP’s area was contained
within a low count county. For example, if a CDP’s area was split evenly among three counties (i.e., approximately
33% of the CDP’s area was contained in each county) and we marked one of those counties with a low count
indicator, then we also marked the CDP as having a low case count.
42
2002
63.25
17.72
19.03
100
2003
63.98
18.29
17.72
100
2004
64.91
17.37
17.72
100
2005
64.91
17.79
17.31
100
2006
65.22
17.50
17.28
100
2007
60.90
21.22
17.88
100
2008
62.04
21.13
16.83
100
2009
64.52
19.15
16.32
100
2010
65.45
21.60
12.95
100
2011
65.03
21.99
12.98
100
2012
65.51
21.57
12.92
100
2013
64.84
22.14
13.01
100
2014
63.98
22.34
13.68
100
2015
66.47
18.84
14.70
100
2016 63.12 23.10 13.78 100
Average 62.87 19.80 17.34 100
High Rates
The Eviction Lab map also indicates Census tracts and block groups with filing and eviction
rates in the top 1%. Filing and eviction rates are calculated by dividing the number of filings or
evictions in an area by the number of renter homes. That means that very high rates could be the
result of (1) a large number of evictions or filings or (2) a small number of renter homes in an
area.
High eviction and filing rates should be interpreted with caution, for a few reasons. First, they
may be driven by small denominators found in areas with few or underestimated renter homes.
For instance, some Census tracts incorporate both industrial zones and a small residential
section. If the residential section had 40 renter homes, and 20 households were evicted
residential section that would generate an eviction rate of 50%. Here, the eviction rate is high
because the number of renter homes in the area is small. A residential neighborhood surrounding
a college with 10 estimated renter homes and 20 evictions would report an eviction rate of 200%.
In this case, the high eviction rate is explained by the fact that the Census often undercounts
seasonal renters (like college students) in its estimate of the number of renter homes in an area.
In most jurisdictions, the eviction process starts with an out-of-court notice delivered to a tenant.
In Maryland, however, the process begins with an eviction filed in court. Many landlords file
against their tenants every month, resulting in a very high case volume. Here, the number of
filings is inflated because of unique court procedures, resulting in a filing high rate, but not
necessarily a similarly high eviction rate.
43
VI.
Future Research
The Eviction Lab’s efforts have produced the most comprehensive data set of evictions in
America to date. We have published aggregated eviction data for years 2000 to 2016 at the
Census block group-, tract-, place-, county-, and state-level. Our data are publicly-available at
www.evictionlab.org.
The Eviction Lab will continue our endeavor to provide the most comprehensive data on
evictions in America and we invite you to stay tuned for updates to our data set by signing up for
our email list at www.evictionlab.org. We will continue to update our data set and this
Methodology Report as new data are collected and as we improve upon our methods. At our
website, researchers interested in merging data at the individual-level can also submit a Data
Merge Application.
If you have eviction data we do not or if you use our data, please contact us and tell us about
your findings by emailing [email protected]. Fully understanding America’s housing
crisis, including the dynamics of eviction, requires a collective and interdisciplinary endeavor.
v.1.1.0 (5/7/18)