For many decades, Medicare has used a statistical sampling and extrapolation method to impose large recoupments on providers. For a hypothetical example, if a provider had 10,000 claims for $1M dollars, Medicare might sample 30 claims, find a 50% error rate, and seek a recoupment for 50% of that year's claims ($500,000).
There is a line in statute that extrapolation can only be used when error rates are high. However, I was involved in a case once where the initial rate was high (say, 50%) and the final error rate was low (5%) yet CMS pursued extrapolation because the initial (and no matter how wrong) error rate was high. I found this appalling. The law is at SSA 1893(e)(3).
I was also appalled in a case where CMS calculated a 90% confidence band using the simple standard deviation method - but standard deviation are parametric statistics (based on a "standard curve" or "Gaussian curve") and the data was even remotely normal (parametric). This update seems to at least take account of this problem - noting that in some data sets it may not be possible to use normal methods "due to the use of theoretical assumptions underlying standard statistical methods."
But the wording is funky - such principles are not "theoretical assumptions" (sic) but rather "mathematical principles." For example, Pi = circumference / r2, is not a theoretical assumption about pi, but the definition of pi.
CMS also admits that some methods could result in an "overpayment debt" that is more than originally paid - which makes visible the nonsense created by poor statistical methods.
Now in January 2023, CMS has released a heavily updated version of its statistical investigation guide, which is "Program Integrity Manual, Chapter 8". Find the PDF here:
I haven't done extra research, but I noted a law firm report of certain extrapolation changes made in 2018/2019 here. Medicare Compliance (trade journal) has an open access issue on CMS changes also, here.
See Q&A via AI at ChatGTP here.