The Department of Transitional Assistance (DTA) is seeking to better understand churn in SNAP recertification so that no participant is ever surprised by a lapse in food assistance. Recently, we have partnered with DTA on a data-driven project to identify actionable insights that will improve their administration of nutrition programs within the Commonwealth. Working together the team has been able to deliver descriptive, predictive and prescriptive analytic practices to help improve business processes around SNAP and begin to reduce churn.
What is SNAP Churn?
SNAP stands for Supplemental Nutrition Assistance Program (also formerly known as food stamps), a federal aid program administered at the state level that offers nutrition assistance to millions of eligible, low-income individuals and families. In Massachusetts, the Department of Transitional Assistance (DTA) manages the SNAP program.
Every month in Massachusetts approximately 50,000 SNAP cases are due for recertification, essentially signing up for a renewal of their benefits. While most participants renew their benefits on time, a significant number of people “churn” — meaning they let their benefit eligibility expire and then rejoin the program again soon after. This isn’t ideal for anyone involved. The individual might not realize their benefit has lapsed and has to deal with the anxiety and financial insecurity associated with those lost benefits. At the same time, that churn costs the DTA — a recent federal study determined the administrative costs associated with churn to be $80 for each instance.¹
DTA built complex database queries and compiled information on over 1.7 million SNAP client recertification events during 2016, 2017, and 2018. Our data scientists then used open-source tools (like Python and R) to analyze this data to better understand churn. The project’s aim is to identify the characteristics of churners and use machine learning tools and processes to predict churn behavior. All in all, the goal is to target individuals for early intervention strategies to reduce churn behavior. So far, we’ve made a few important steps toward that goal, first by defining the cohort of churners that could be targeted for intervention.
Using Data to Define Churn
Historically, a churner has been defined as a SNAP household that experiences a break in participation of four months or less¹. You can see in the graph below that the bulk of churners return to SNAP within a few days of their benefit lapse.
This shows that churners behave in a very predictable declining pattern. The “half-life” of churn, or time at which 50% of churners return, is 18 days after case closure. The next half-life, or time at which 75% of churners return, is 140 days after case closure.
Combining the insights from this dataset with domain knowledge about SNAP, and its participants, drew out a new theory that churn is composed of three different behaviors:
- Late Recertifiers: Very early churners that are aware of the recertification deadline but are late in renewing their benefits.
- Pausers: Churners who are intentionally and voluntarily “pausing” their participation in SNAP (shown in the tail of the distribution shown in Figure 1).
- True Churners: Churners who were unaware of recertification and discovered their benefit had lapsed when they tried to use it.
Our theory is that these three behaviors are embedded in the combined data as shown in Figure 2 below.
Fitting a Statistical Model to the Data
To test our theory we took a closer look at the shape of the decay curve plotted in Figure 1. If our theory is valid, then the distribution of “true churners” should be the dominant behavior after the “late recertifiers” group has shrunk and before the “pausers” start to return to SNAP. We can fit an exponential distribution model, which is commonly used to model the time between events, to this portion of the dataset and see a result that matches our expectations, seen below in the green dotted line.
The fitted distribution matches the churn data from ~3 days though ~21 days. At both ends of the data, the model fit degrades. While not conclusive, this result does support our theory that three different populations make up the SNAP churn dataset. At three days it is reasonable to think that most of the “late recertifiers” have returned to SNAP. Likewise, at 23 days most people who rely on SNAP will have tried to use the benefit and realized it was no longer active, while “pausers” will start returning to the program.
Exploring the SNAP recertification dataset has led us to a refined definition of churn. We employed statistical models to help justify theories about churn behavior and add additional context. This new definition will be carried forward in the churn study and will help provide the focus for different intervention strategies. For example, web-based renewal tools can be targeted to help reduce the impact of late recertification, while communication strategies can be developed to focus on individuals who were unaware that their benefits had lapsed. We’re looking forward to continuing to support the DTA’s data initiatives and helping them continue to improve constituents’ experiences.
A big thank you to Andrew Wheeler, Michael Cole and Hussen Mohammed from the Massachusetts Department of Transitional Assistance as well as Kris Johnson from Massachusetts Digital Service who were key contributors to the work highlighted in this post.
¹Mills, Gregory, Tracy Vericker, Heather Koball, Kye Lippold, Laura Wheaton, Sam Elkin. Understanding the Rates, Causes, and Costs of Churning in the Supplemental Nutrition Assistance Program (SNAP) ‐ Final Report. Prepared by Urban Institute for the US Department of Agriculture, Food and Nutrition Service, November 2014.