Beginning in October 2027, the Supplemental Nutrition Assistance Program (otherwise known as SNAP and formerly “food stamps”) is going to be significantly more expensive for states. Currently, the program is fully guaranteed by the federal government, so aside from some administrative costs this amounts to free money for the states.
Now with the passage of the One Big Beautiful Bill Act, states will have to pay a portion of their SNAP benefits themselves. This change alone is potentially devastating to state governments. SNAP is a massive entitlement program which means that anyone who is eligible for benefits may claim them. This means that 1) some states might not be able to afford paying for any portion of their total SNAP benefits (especially in the face of balanced budget requirements) and 2) states wont know how much spending they are on the hook for until after the fact.
Those two problems are significant enough that it might be enough to cause states to restrict eligibility or cancel the program altogether, but the actual implementation that states have to contend with stands to make matters even more complicated.
According to the new legislation, a state’s SNAP error rate will determine what percentage of the total costs they end up paying for each year.
What is the SNAP error rate?
The SNAP error rate is the percentage of total SNAP benefit payments that are estimated to have been incorrectly overpaid or underpaid to participating households. This is calculated by taking a sample of households that receive benefits, and auditing them to see if they received the correct amount of benefits each month.
There are two points about this calculation that are crucial to understand. First, this is about total error at the household level. Overpaying one household and underpaying another don’t cancel out in the aggregate estimate, they both count towards a higher error rate. The second point is that this is calculated on a random sample of households, and is subject to sampling variation.
That second point is especially important because the new law uses this error rate to determine how much states will have to pay. Beginning in 2027, states with an error rate below 6 percent will continue to not have to contribute state funds toward SNAP benefits. States with error rates between 6 and 8 percent will pay 5 percent of benefits, states between 8 and 10 percent will pay 10 percent, and states with error rates of 10 percent or higher will generally pay 15 percent.*
How will this impact states?
A Brookings study published last month looked into state SNAP error rates and measured how much they vary from year to year. In short, they found that these cost-sharing thresholds are far too small to account for the natural variability of the error rate estimate. This is a clear example of why it is foolish to build policies around point estimates. Uncertainty that is inherent whenever we look at a sample makes it so states have basically no way to plan for how much their SNAP program will cost year to year.
Brookings estimates that the average standard error of a state’s annual SNAP error rate is about 1.1 percentage points. That means a typical 95 percent confidence interval spans roughly 4.4 percentage points. But the cost-sharing thresholds between 6, 8, and 10 percent are only two percentage points apart.
To put that into perspective, imagine that a state has an estimated error rate of 8.5 percent. Under the new policy, that state would have to pay 10 percent of its SNAP benefits. But the statistical uncertainty surrounding that estimate means the state cannot reliably be distinguished from one with an error rate of 7.1 percent or 10.9 percent. Those three estimates would result in the state paying 5, 10, or 15 percent of its SNAP benefits, respectively, which could amount to hundreds of millions of dollars of differences in state required spending.
When looking at the historical data, Brookings found that 43 states had error rates that moved between at least two of the cost-sharing categories between 2022 and 2025. Fourteen states moved between three categories, and North Carolina moved between all four. Only five states (Idaho, South Dakota, Vermont, Wisconsin, and Wyoming) had error rates consistently below 6 percent over that period.
The potential dollar amounts involved are enormous. Using 2023 data, Brookings estimates that California, for example, could face anywhere from no state contribution to more than $2 billion in SNAP costs simply because of the statistical uncertainty surrounding its estimated error rate.
This creates a strange situation where states can make substantial investments in improving their SNAP administration without knowing whether those improvements will actually reduce their financial responsibility. A state’s error rate can change substantially from one year to the next simply because a different set of households happened to be selected for review.
The underlying issue is not that statistical sampling is inherently bad. Sampling is an essential tool for estimating characteristics of large populations, and every estimate based on a sample comes with some degree of uncertainty. The problem is using a noisy estimate to make a high-stakes financial decision without accounting for that uncertainty.
States are required to balance their budgets, which makes unpredictable expenses particularly difficult to manage. A state cannot simply assume that its SNAP costs will be the same next year, because its required contribution could change by hundreds of millions of dollars based largely on statistical noise.
If the goal of this policy is to encourage states to administer SNAP more efficiently, there are better ways to accomplish that goal. Making state governments responsible for billions of dollars in benefits based on an estimate that can change substantially because of random sampling variation does not create a particularly strong incentive for better administration. Instead, it creates uncertainty for state budgets and, potentially, for the people who rely on SNAP.
* There is a temporary exception for states with particularly high error rates during the first two years of the policy.
