Pilot project combines federal and state datasets to find SNAP fraud

The Department of Agriculture’s Food and Nutrition Services is piloting the use of predictive analytics to crack down on fraud in the Supplemental Nutrition Assistance Program.

FNS has partnered with the IT vendor SAS to leverage its analytical software, which applies anomaly detection, predictive modeling and link analysis to a dataset including both federal and state information.

SNAP fraud rates are just over 1 percent, said Jon Lemon, systems engineer with SAS Federal. Yet while this proves FNS efficiency, the seemingly reassuring statistic can be quite deceiving.

“That's 1 percent of a multi-billion dollar a year program,” he said "We're still talking tens of millions, if not hundreds of millions of dollars, that is going to fraudulent activity every year."

Using sophisticated analytics to knock down the 1 percent to half that or even lower, could save taxpayers nearly $40 million to $60 million a year  and offer more benefits to citizens who need them.  

The key to the analysis is combining both federal and state datasets.  Usually, the federal government tracks the SNAP benefit retailers, while state governments track the recipients.  In a pilot with seven states, FNS used SAS analytics to merge federal and state datasets and then analyze state recipient data to help uncover fraud on the federal end of the program.

One corrupt retailer is usually associated with 30 to 40 fraudulent recipients, so tracking the activity of a dishonest recipient who travels unusually far to find a retailer can point to a crooked store.

“Being able to combine those pieces of data is really effective, so that’s why we've been working with FNS and the various states and retailers out there to uncover these anomalies,” Lemon said.

According to SAS, the software is installed at the FNS facility where the data is housed. Once as much data as possible is gathered, cleaned and integrated, a four-step analytical process is applied that allows SAS to assess results along the way to strengthen the accuracy of its findings.

Applying a set of rules outlining eligibility for benefits is the first line of defense and helps to initially prevent fraudsters from getting electronic benefits cards.

SAS then applies anomaly detection to the data by building “peer groups,” with information on age, geographic location or income status that shows what the average beneficiary group looks like. Against that baseline, “good” and “bad” candidates stand out.  If, for example, the average person in a certain peer group receives $400 worth of benefits in a month and another in the same group receives $5,000, that anomaly is flagged and that individual is considered more of a risk and possibly fraudulent.

The predictive modeling technique helps SAS and FNS prevent the money from going out to fraudsters rather than having to “pay and chase” – or recover funds once they’ve been distributed. By tracking behavior over time, the system can predict behavior like “stair-stepping,” in which recipients  take small incremental steps to increase their benefits. Behaviors like this that have been linked to fraudulent activity before are scored as a higher risk and brought to an investigator’s attention.

The final technique, link analysis, creates a web of connected information about a particular beneficiary. This could include data on family, business and banking as well as relationships with FNS. If someone closely associated to an individual is guilty of fraudulent activity, or the store an individual frequents is turning over SNAP benefits illegally, the beneficiary is flagged as suspicious.  

All four of these predictive analysis techniques are used together to build a stronger case around an fraudulent individual, retailer or network.

When SAS uncovers retailer fraud against the federal government, it takes the information straight to FNS, which then decides how to proceed. When recipient fraud against a state is discovered, SAS can give the state a list of names, evidence and risk probability scores.  And though it is in the state’s best interest to act, it is not required to do so.

“Over the years, FNS has developed best practices for using technology and analytics to combat fraud, waste and abuse in its program,” said Karen Terrell, vice president of SAS Federal. “With this project, FNS is taking a strong leadership role, using what they’ve learned to help the states go after bad actors.”

Though the pilot phase is soon coming to an end, Lemon hopes to expand the program to all 50 states and apply the same techniques to other assistance programs. He said the Women, Infants and Children program, Social Security, Medicare and Medicaid all could benefit from similar predictive analytical techniques.  

Strict privacy laws and policies that restrict sharing data between certain parts of the federal government might prevent some analysis, Lemon said, but comparing the Social Security Administration’s master death list to SNAP program recipient lists, for example, could be greatly beneficial.

“We’re trying to talk about how we can combine these datasets,” he said, “and I think the government recognizes the benefit behind it.”