Predicting Evictions in New York City
New York City is one of the most socioeconomically stratified and geographically segregated cities in the United States. As such, it is not uncommon for more vulnerable populations, such as low income communities or communities of color, to find themselves at the mercy of those with greater access to resources and power in our city – there is no clearer example than in New York City’s countless evictions and failure to provide accessible and affordable housing.
New York’s residents are deeply stratified on the basis of race and class, and evictions for the city’s most vulnerable are not uncommon. Unfortunately, evictions and a lack of access to stable housing can result in a ripple effect through other aspects of a person’s life, resulting in job loss, food insecurity, or poor educational outcomes. I ultimately decided to develop a housing vulnerability prediction model for the Legal Aid Society, an organization in New York City that offers pro bono legal services to low-income New Yorkers across all five boroughs. This model was developed in ArcGIS Pro because the Legal Aid Society is relatively well resourced.
The goal was to build a model that could predict which ZIP codes contained households that were at the greatest risk for eviction. The independent variables were to include: race, household income, and proximity to other evictions; the dependent variable would have been the likelihood of an eviction filing. In order to present the data, I hoped to create a storyboard highlighting the highest-risk neighborhoods and key demographic information. This model would hopefully be used by Legal Aid to improve their canvassing and outreach efforts in high-risk neighborhoods. Those facing eviction are frequently low-income people of color, oftentimes with Limited English Proficiency – access to adequate legal representation allows these individuals to have a fighting chance. According to the Center for American Progress, in most eviction cases, 90 percent of landlords have legal representation, as opposed to only 10 percent of tenants. Providing free legal services helps level the playing field for those who may lack access to the necessary funding or resources to protect themselves in court.
By building this model, my hope was to better understand both the factors that fuel evictions and to provide targeted support to individuals in these vulnerable areas. This model can be thought of as a basic template that can be customized based on the availability of data and the unique factors of each city or town. As this was an exercise in experimentation, my regression equation was relatively simple:
Data Sources
Much of the data used for this model was sourced from the U.S. Census Bureau’s 2020 Census and the City of New York’s open data repository. Unlike many other metropolitan areas in the United States, New York makes a significant amount of data available for use by the public – including information on evictions, youth programs, and even a “Squirrel Census.” While the data do have significant gaps, including datasets where entire neighborhoods or ZIP codes are missing, the accessibility and variety of information here far exceeds that of nearly every other American city.
Walking Through the Process
It is worth noting that the model in its current iteration is not perfect – while it does run, it generates a few errors that I have not been able to rectify at the time of writing this report. However, this has been an excellent educational experience and I can confidently share several cautionary tales for those trying to build similar models for their respective cities.
Step 1: Obtaining, Cleaning, and Preparing the Data
First, I obtained shapefiles from NYC’s data repository outlining the five boroughs, the neighborhood boundaries within each borough, and the outline for each ZIP code. In order to obtain more granular data for analysis, I downloaded both the TIGER shapefiles from the U.S. Census Bureau alongside data on demographics, income, and other housing characteristics by Census Tract. All datasets were given a preliminary cleaning in Python and exported as Excel files before being uploaded to GIS, where relevant columns were converted into compatible data types for future joins.
Next, I cleaned and imported New York City’s eviction data. This dataset provides the date and address of each eviction filing in New York for the past few years, which provided me the opportunity to geocode each location and get a better sense of the spread of evictions throughout the city. Unfortunately, it was impossible to glean any geospatial insights from this point layer because there had been so many evictions it was difficult to tell where, exactly, they were concentrated. Using this point layer, I summarized the number of evictions by ZIP code after joining it to the neighborhood shapefile. I similarly summarized the demographic and income data from the Census to prepare it for the model.
Step 2: Try, and Fail, to Test a Basic Training Model. Realize, for the First Time, the Limitations of Your Data
Falling victim to my own hubris, I assumed that the prepared data would simply allow me to “plug and play” when developing a Forest-Based Classification and Regression Model. This assumption would prove to be incorrect for two reasons: (1) All the information needed to be stored in the same feature class or table, as the model is incapable of pulling from multiple classes, and (2) If there are any null values in your data, regardless of whether or not you are using that variable in your model, the model will not run.
Unfortunately, ArcGIS does not necessarily provide a simple or straightforward way to clear all null values. A find and replace approach requires the user to sort through each individual column, which would be quite difficult with the number of columns I had hoped to incorporate in the model. One could simply select and delete rows containing null values, but this would likely skew our results significantly and exclude numerous ZIP codes from our analysis. Ultimately, the most sensible approach was to export the data from ArcGIS to Excel (using the Feature Class to Excel tool), calculate the average value for each column containing null values, and replace those values with the average. Though not an ideal approach, replacing null values with the average prevents the data from becoming too skewed, which is what would happen if we were to replace those nulls with zeroes instead.
After manually cleaning the data, it was imported back into ArcGIS and joined with the shapefile. It was then that I noticed something unusual: after multiple joins, there appeared to be duplicate rows despite ZIP codes matching perfectly in all tables pre-join. This turned out to be an unusual quirk of the way ZIP codes are laid out in New York City: some ZIP codes span multiple boroughs, and were thus spanning multiple rows where the demographic information matched, but the boroughs were logged differently. I used the summary statistics tool and remove identical features tool to ensure we were not double-counting certain areas.
Step 3: Try, and Fail, to Create a Proximity-Based Analysis. Realize, for the Second Time, the Limitations of Your Data
I had assumed that proximity to other evictions may be indicative of the likelihood of evictions occurring for a specific building or neighborhood. I attempted two analyses that would help quantify where evictions were most frequent: (1) A hot spot analysis and (2) Utilizing a radial basis function to weight high-eviction areas more heavily. A radial basis function essentially creates a flat “sheet” to represent your geographic area, with each point on the sheet corresponding to a point on the map. For every input at a particular point – say, a documented eviction – the function raises the “sheet” at that point, ultimately creating a landscape where locations with a greater number of inputs are “taller.”
Unfortunately, as the datasets I was working with were relatively small, ArcGIS was unable to perform either analysis. In fact, I was so unsuccessful, that attempting to use the radial basis function prompted an error message to contact Esri support for additional assistance.
Rather than use a proximity based analysis, I ultimately decided to use the proportion of evictions to residents by ZIP code as the model’s location-based variable.
Step 4: Run the Model and Assess Preliminary Findings
As mentioned at the beginning of this report, the model does run, but with a few errors. This is likely due to the relatively small sample size available: I ultimately only had around 300 rows of data to feed the model, which severely limits its training ability. Nevertheless, I ran the model using 10,000 trees to see if we could glean any preliminary findings from what was available.
The model identified areas in the South Bronx and Eastern Brooklyn, such as Bedford-Stuyvesant, as being particularly high-risk areas for eviction. This finding is not terribly surprising based on the variables we have fed the model: while evictions are extremely common across the city (with the exception of Staten Island), these areas tend to have residents with the lowest incomes and large Black populations relative to other neighborhoods.
Areas of particular note include: Brownsville, East New York, Rugby-Remsen Village, Canarsie, Bedford-Stuyvesant, East Flatbush-Farragut, Bushwick, Mount Hope, Highbridge, East Tremont, and Concourse Village. I would therefore recommend that the Legal Aid Society, should they use this model to improve their outreach efforts, focus their energy on these areas.
Reflection and Next Steps
It is worth noting the limitations of this model as it currently stands; while I had hoped to build something useful, it currently requires quite a bit of work to live up to its full potential.
First, as the model is currently running with errors, it is impossible to generate a report outlining the R-squared values of my independent variables. This is, again, likely due in part to the small datasets that I worked with for this project – as the city continues to collect data, the model can be more accurately trained. Unfortunately, as this report cannot be generated, I can only assume a causal relationship, or at least a correlated relationship, between the variables I currently have and the output variable.
Next, should I continue to develop this model, I would like to include additional variables that I feel are lacking from the current model. As discussed, I would strongly prefer to include a proximity-based analysis in the future, but would also like to include additional information about median home value, rates of foreclosure in the area, the proportion of renters to home owners, average rent prices, and average mortgage by ZIP code. Despite the robustness of NYC’s Open Data repository, it does not currently house any of this information. I had hoped to use an API that Zillow had initially created and shared publicly for this information, but it has become defunct in recent years. Were there additional time to develop this model, I would develop a Python script to manually scrape the data from the website itself.
Finally, I would like to reflect on the need for improved data infrastructure. The first iteration of this project was geared towards the city of Atlanta, as can be seen in my initial project proposal – this task proved virtually impossible, as the city has not collected data on evictions, or other housing-related metrics, in several years. While the New York model will hopefully serve as a useful template for other cities, these models simply cannot run without data to feed it.
What New York has is an exceptional resource – but it shouldn’t be. ArcGIS and other geospatial technology can be incredibly useful in helping us locate vulnerable communities so we can develop more targeted policies and provide immediate aid wherever and whenever necessary. However, the absence of data threatens our ability to do that work effectively – we cannot support people if we don’t know where they are. My greatest takeaway from this project, aside from familiarizing myself with the many hoops it takes to build a model, is how much we need improved access to data. New York’s current model is just the first step.
Final Process Log
The model, alongside the cleaned data, can be found in the Final Project .zip file.