The Impact of Demographics and Income on Eviction Rates

Introduction

In an earlier post, we introduced the concept of impact charts and showed some examples illustrating the impact of race, ethnicity and income on eviction filing rates in DeKalb County, Georgia. If you are not familiar with impact charts, we recommend you go back and read the original post, then return here. Here is an example of an impact chart from that earlier post:

The charts we included in that post were selected more to instruct the reader in how to interpret impact charts than they were to reach deep conclusions on the dynamics of eviction in DeKalb County. In this post, we’ll introduce a new website, evl.datapinions.com, that allows readers to examine over 1,600 impact charts from counties all over the country.

Like our earlier impact charts, the impact charts at evl.datapinions.com were built using the impactchart project on top of a combination of eviction filing data from the Eviction Lab at Princeton University and demographic data on renters from the U.S. Census Bureau. In the initial version of the site, we used only the proprietary data the Eviction Lab collected from vendors.¹ We subsequently updated the site to use their dataset that includes both proprietary data and data they collected directly from courts and other governmental agencies.²

The web site includes charts for 479 counties across the country. (The earlier version with proprietary data only covered 168 counties.) Our criteria for including a county was that it had at least 200 records with data on filing rate in the eviction lab proprietary data from between 2009 and 2018. The actual amount of data included, what years it covers, and how much of the overall geography and population of the county is covered varies. Without examining the data coverage for any given county in more detail, we can’t make final conclusions on the dynamics of eviction in any given county, or what policy changes might change them.

Model Accuracy

For each of the counties, we include the r² score of a model with the same hyperparameters we used to generate the charts, but trained on the full data set for the county. This gives us some indication of whether the approach we are taking is capable of modeling the data for any given county in a reasonably accurate way. The scores cover a wide range, from nearly 1.0 down to negative scores that show that the machine learning model is less accurate that simply predicting the overall mean in all cases. In those cases, whatever variation exists in eviction filing rates has something to do with factors we did not consider at all. Machine learning models don’t work well in all cases, and this work demonstrates that.

But it is important to note that model accuracy is not actually the goal here. We’re trying to understand and explain where in the country there are structural factors that cause race and ethnicity to impact eviction filing rates and what those factors look like. When the r² score is below zero, which is was is some counties, that means our machine learning model failed to find any way of using race, ethnicity, and income of renters to predict eviction rate. This is actually a good thing, indicating there is no structural racism the model can find and exploit.

When the r² score is high, that’s when there is actually a problem. It means that the eviction filing rate can be predicted if we know the racial and ethnic makeup of a given census tract in the county we are modeling. We can then look at the impact charts for that county and see exactly what the structural issues are. We will cover this in detail in a post coming soon.

Data Coverage

Data on eviction rates comes from a data set of government records and proprietary data set produced by the Eviction Lab. This data set does not cover all counties in the U.S. and for the counties it covers it does not cover all years. Additionally, the Census Bureau adds, removes, or changes the boundaries of census tracts from time to time as the population shifts.

In order to make it more clear what data was used to produce each of the impact charts, we also provide a series of maps at the bottom of the page. There is one map of the county for each year we considered. Each small area in each map is a census tract. The green ones are the ones for which data was available. Data from all of the green tracts went into building the models
that produced the impact chart on the same page. Dollar figures were inflation adjusted to 2018 dollars.

Here are what the maps for Orange County, Florida look like:

We can see from all of the green that most of the county was covered, and coverage went up from 2009 to 2011. We can also see that in 2010, the Census Bureau increased the number of tracts in the county by splitting some existing tracts whose populations had grown.

The more green in the maps, the more representative our models are. To the extent that the times or places where data is missing are large and are correlated with circumstances of atypical eviction activity, the impact charts may be less accurate. Missing data for an entire year or two is less problematic than missing data systematically year after year for one part of the county.

Final Thoughts

We encourage readers to take some time to look at a variety of different impact charts on the page. We would especially like to hear feedback from readers with qualitative familiarity with any of the counties considered to hear whether the impact charts are or are not consistent with their observations.

Updates

Updated Nov. 5, 2023 with the launch of a new version of the site containing data coverage maps.

Updated Nov 15, 2023 with the launch of a new version containing all observed data, not just the proprietary data.

References

Gromis, Ashley, Ian Fellows, James R. Hendrickson, Lavar Edmonds, Lillian Leung, Adam Porton, and Matthew Desmond. Estimating Eviction Prevalence across the United States. Princeton University Eviction Lab. https://data-downloads.evictionlab.org/#estimating-eviction-prevalance-across-us/. Deposited May 13, 2022.
Desmond, Matthew, Ashley Gromis, Lavar Edmonds, James Hendrickson, Katie Krywokulski, Lillian Leung, and Adam Porton. “Eviction lab national database: Version 1.0.” https://data-downloads.evictionlab.org /#legacy-data.