How We Used Open Data to Identify Investor-Owned Single-Family Rental Properties
Lessons Learned from Exploring the Landscape of Landlords in the Twin Cities
In North Minneapolis, renters living in single-family homes owned by real estate investment trusts and private equity firms reported poor housing conditions, unexpected fees, and unresponsive management staff. Many of the tenants are also preparing to fight eviction filings from these owners that they worry will follow the end of state and city eviction moratoriums passed during the COVID-19 pandemic. These tenants are not alone. Having an investor or corporation as a landlord has become increasingly common, with nearly 40 percent of all rental units nationwide owned by anonymous shell entities.
Researchers, housing advocates, and politicians aiming to promote equitable housing and economic recovery are concerned that the growth of speculative investment in single-family homes will exacerbate racial inequity in the housing market for both homeowners and tenants. Our team partnered with organizations in the Twin Cities to conduct research that would support their agendas for informed and equitable housing policy decisionmaking. The Twin Cities already has the country’s largest racial homeownership gap and a rental market that is unaffordable for low-income renters and renters of color.
These investors often form anonymous shell entities, like limited liability companies (LLCs) and limited partnerships (LPs), making it nearly impossible to track down who is behind these investments. The lack of legally mandated transparent ownership data has implications for housing research and policy. It is difficult to discern the individuals behind these single-family rental purchases and to determine where and to what extent this investment occurs at the local level.
Our partners at the Alliance, the Family Housing Fund, and the Center for Economic Inclusion wanted property- and neighborhood-level trends to advocate for equitable and targeted policy solutions. We worked with researchers at the Center for Urban and Regional Affairs (CURA) in Minneapolis to use publicly available county assessment data to identify LLC and corporate single-family rental owners in Hennepin and Ramsey Counties in 2005, 2010, 2015, and 2020.
We created new key ownership variables, including “form of ownership,” “owner location,” and “portfolio size.” We found that the share of corporate-owned single-family rentals more than tripled between 2005 and 2020, from 4 percent to 14 percent. We demonstrated that these investor-owned single-family rentals were disproportionately located in neighborhoods with higher shares of residents of color and low-income residents.
We outline key lessons below in hopes that they might be useful to researchers, advocates, and organizations interested in understanding their landscape of landlords.
Lessons learned from attempting to disentangle the web of corporate landlords in the Twin Cities:
1. Partner with local leaders to create analyses that support ongoing research and organizing efforts
Renters, housing organizers and advocates, and local researchers are housing experts. They understand their local housing markets, the landscape of landlords, and the equity implications of shifting trends, so we partnered with CURA, a local research organization. Its input and support were vital to making our analysis relevant to research and advocacy in the Twin Cities region. Our teams created the key search terms list and code we used to text-mine the assessment data. For example, CURA suggested including the name of the nonprofit developer Urban Homeworks in the nonprofit class, excluding properties owned by utility companies and those owned by private education institutions like the University of Minnesota, all seen in the ownership class keyword code below. These iterative conversations helped flag key investor-owners in the region that our initial text mining would have missed. The conversations also minimized the possibility of accidentally classifying owners like local nonprofits or public housing authorities as investors.
Below is the R script we used to examine property records from both counties to identify and categorize investor-owners of single-family homes.
CURA and our partners were also interested in understanding where these landlords were located in relation to their property holdings. Their researchers had heard anecdotally that corporate landlord absenteeism made repair requests and positive landlord-tenant relationships difficult. We used geocoded taxpayer addresses using Urban’s geocoder to get an approximate location for each parcel owner and to create a clean and unified owner address field. Property tax records often have data entry errors, such as misspellings, and the many departments involved in the data collection might not be compatible or consistent. Geocoding our taxpayer data thus reduced the number of unique owner addresses of single-family rentals in 2020 from 520,000 to 476,000.
Geocoding the taxpayer addresses helped us highlight transfers of wealth, in the form of rent, from low-income neighborhoods and neighborhoods of color to investors outside the region. But our partners also emphasized that local ownership did not equate to equitable economic growth. We used this feedback to create a subanalysis that evaluated intraregional ownership trends and found that 25 percent of local single-family rental owners lived in high-income majority-white neighborhoods, while only 16 percent lived in neighborhoods with high shares of low-income residents or residents of color. This knowledge-sharing partnership made our analysis more relevant to local concerns about inequitable capital flows.
1. Identify gaps and supplementary data sources early on
We found gaps in data quality and completeness in the parcel assessment data between counties and over time for important variables, like those that specified parcel unit count. There was also no variable for designating whether a parcel was a rental property. This meant we had to use parcel assessment data to identify institutional owners and to determine whether a parcel was even a single-family home or a single-family rental property. We initially considered increasing the data accuracy by shrinking the geographic scope to cover only a point-in-time snapshot in Minneapolis or St. Paul, but our partners emphasized the importance of understanding changes in investor ownership over time and particularly in first-ring suburbs.
Conversations with the Hennepin and Ramsey county assessor and property tax division offices gave us background on tax exemptions in the data, like homesteading, that we used to determine whether a parcel was likely a rental property. We then worked with CURA to create a proxy variable for designating properties as rentals using information about owner and taxpayer addresses, the homestead tax exemption, and land-use designations, which allowed us to generally know whether a parcel was an owner-occupied single-family home or a likely rental. We refined the accuracy of this proxy by testing it against a rental property license dataset in Minneapolis and rental registration data in St. Paul and by comparing our estimates with those produced by the US Census Bureau.
2. Use clustering and fuzzy matching to make manual cleaning less painful
The most difficult task was trying to disentangle the web of LLCs, LPs, and real estate investment trusts. Investor owners frequently create these shell entities that make understanding the true size and scope of their portfolio difficult. For example, Invitation Homes, formerly owned by Blackstone, had multiple owner and taxpayer names (e.g., “IH3 Property Minnesota LP,” “IH2 Property Illinois LP,” and “IH4 Property Minnesota LP”) associated with a single taxpayer address in Dallas. Manually cleaning ownership information for tens of thousands of parcels was not feasible and would have taken a lot of time, so we used other string manipulation approaches to determine the extent of the portfolio of these investor owners.
After conducting some baseline cleaning, we used fuzzy matching to reduce the number of slight variations between the recorded owner name and taxpayer name in the assessor data. We used the Jaro-Winkler (JW) method in the R stringdist package to produce a measure between 0 and 1 for each owner and taxpayer name pair, with 0 being an exact match and 1 meaning there is no similarity between the names. Jaro-Winkler is commonly used to compare short strings, like names, where scores are based on character similarity, spacing, and distance, along with weighting for differences at the beginning and end of strings. We then manually created JW score cutoffs for landlord names in each year by eyeing which score seemed to be a conservative cutoff between different spellings and different owners.
You can see an example of similar owner name and taxpayer names, along with JW scores, in in the table below. Our team would have used matches with JW scores less than 0.216667, highlighted in red below, for our conservative cutoff, even though other owners and taxpayers were likely the same.
A part of the R code we used to fuzzy-match is below.
Then, we used a basic unsupervised cluster method to further decrease the number of names within our new “clean_name” variable. We started by creating a distance matrix using the Jaro-Winkler edit distance metric. Then, we used an unsupervised machine learning algorithm called hierarchical clustering. This algorithm started by treating each clean name as its own individual cluster and gradually combined increasingly similar names into larger clusters until it reached our predefined number of clusters. We shifted our cluster count (K in the code below) until we saw that the names within each cluster were likely to be a single owner.
The R script example below was done using a base data table of landlord names for all likely rental properties in both counties in a single year and allowed us to reduce the number of unique landlord names from 11,089 to 10,500.
As a last step, we looked through the data and created a few additional manual cleaning rules, like grouping by taxpayer address, to reduce the number of variations and to more accurately understand how many properties were in an owner’s portfolio. This was important because we created a cutoff for “corporate ownership” that excluded small-scale owners with fewer than three total properties.
This preliminary analysis gave us a broad understanding of where and when investor owners bought single-family rentals, but there were limitations in our data cleaning and analysis. Our text mining and clustering methods could not be tested for accuracy because it was an unsupervised method. Our team is sure there were missing keywords in our word search, owner name spelling variations, and missing data that likely made our estimates of the scope and scale of investor ownership conservative. We also know that our lack of a concrete rental property designation variable limits what we can say about this phenomenon. We hope others can refine our methods or potentially pair these data with other sources, like rental licenses, eviction records, and actual rent data, to better understand the actual implications and outcomes for renters in these properties.
Advocates and policymakers need data like these to make more informed, equitable, and targeted housing policies. Housing policies, like taxes or caps on single-family rental portfolios of a certain size or rent control based on total properties owned, are only as effective as the data used to enforce them. Until clear enunciation of property ownership is legally required to be publicly accessible, we hope our preliminary techniques and approaches can help others looking to understand their landscape of landlords. While there were limitations, especially our use of publicly available assessment data,[1] we hope this can build on efforts to increase rental property ownership transparency for renters. Renters deserve to know who is responsible and should be held accountable for their properties.
[1] NOTICE: The Geographic Information System (GIS) Data to which this notice is attached are made available pursuant to the Minnesota Government Data Practices Act (Minnesota Statutes Chapter 13). THE GIS DATA ARE PROVIDED TO YOU AS IS AND WITHOUT ANY WARRANTY AS TO THEIR PERFORMANCE, MERCHANTABILITY, OR FITNESS FOR ANY PARTICULAR PURPOSE. The GIS Data were developed by the counties of Anoka, Carver, Dakota, Hennepin, Ramsey, Scott and Washington and the Metropolitan Council for their own internal business purposes. These organizations do not represent or warrant that the GIS Data or the data documentation are error-free, complete, current, or accurate. You are responsible for any consequences resulting from your use of the GIS Data or your reliance on the GIS Data. You should consult the data documentation for this particular GIS Data to determine the limitations of the GIS Data and the precision with which the GIS Data may depict distance, direction, location, or other geographic features. If you transmit or provide the GIS Data (or any portion of it) to another user, it is recommended that the GIS Data include a copy of this disclaimer and metadata.
WARRANTY AS TO THEIR PERFORMANCE, MERCHANTABILITY, OR FITNESS FOR ANY PARTICULAR PURPOSE. The GIS Data were developed by the counties of Anoka, Carver, Dakota, Hennepin, Ramsey, Scott and Washington and the Metropolitan Council for their own internal business purposes. These organizations do not represent or warrant that the GIS Data or the data documentation are error-free, complete, current, or accurate. You are responsible for any consequences resulting from your use of the GIS Data or your reliance on the GIS Data. You should consult the data documentation for this particular GIS Data to determine the limitations of the GIS Data and the precision with which the GIS Data may depict distance, direction, location, or other geographic features. If you transmit or provide the GIS Data (or any portion of it) to another user, it is recommended that the GIS Data include a copy of this disclaimer and metadata.
