The method is well established, but it is also one of the parts of a case most often challenged on appeal, because an extrapolated demand stands or falls on its documentation. This guide is for special investigations units (SIUs), payment integrity teams, and Medicaid program integrity staff. It covers when SVRS is used, how to build the frame, how to size and draw the sample, how the 90% lower bound works, and where appeals push back.
When SVRS is used
Medicare contractor reviews. For Medicare, the rules on sampling and extrapolation are in Chapter 8, section 8.4 of the CMS Medicare Program Integrity Manual (Pub. 100-08). Under Social Security Act §1893(f)(3), a Medicare contractor may not use extrapolation to set an overpayment unless there is a sustained or high level of payment error, or documented educational intervention has failed to correct the error. That determination is not subject to administrative or judicial review. PIM §8.4.1.4 lists ways to show it, including high error rates compared with similar providers, a history of non-compliance, law enforcement information, and OIG audits.
OIG audits and self-disclosures. HHS-OIG's free RAT-STATS software is the primary statistical tool for OIG's Office of Audit Services, and many providers use it too. OIG's Health Care Fraud Self-Disclosure Protocol, updated October 15, 2026, allows damages to be estimated from a random sample of at least 100 items, using the mean point estimate rather than the lower bound.
Medicaid program integrity. Under 42 CFR Part 455, Subpart A, state Medicaid agencies must run preliminary and full investigations, refer suspected provider fraud to the Medicaid Fraud Control Unit, and resolve cases, which may include seeking recovery of payments. Under Subpart F, Medicaid Recovery Audit Contractors identify and recoup overpayments, and providers keep their appeal rights under state law. Sampling rules vary by state, but state program integrity units, Medicaid RACs, and MFCUs regularly project overpayments and face the same questions about method. See how these steps map to a case workflow on our Medicaid program integrity page.
Commercial and managed care SIUs. Many plans adopt the CMS approach in their provider contracts because it is familiar and easier to defend than a custom one.
Step 1: Define the universe, sampling unit, and frame
Under PIM §8.4.3.2, the universe is the target population that contains all claims or claim lines potentially under review. The sampling frame is the list of every sampling unit that can actually be selected. The sampling unit may be a claim line, a claim, a cluster of claims for one patient, a cluster of claims for one treatment day, or any other unit that fits the issue under review.
Good frames are narrow and explicit: limit them to the CPT or HCPCS codes tied to the issue, fix the review period, and exclude previously reviewed or unpaid lines. Write down every exclusion and why. A frame that doesn't match the population the overpayment is projected to will be raised on appeal.
Strata. If paid amounts vary widely, stratifying by paid-amount band or service type can improve precision. PIM §8.4.1.3 asks the statistician to assess the paid-amount distribution before choosing a design, and §8.4.7.1 requires the provider explanation to define the strata and allocations. OIG's Self-Disclosure Protocol expects simple random sampling unless a stratified or multistage design is shown to be needed.
Step 2: Determine the sample size
PIM §8.4.4.3 says no single minimum sample size fits every situation, and contractors may consider available resources. A challenge that only says "the sample is too small," without engaging the methodology actually used, is without merit.
In practice, teams size a sample using a few inputs:
- the confidence level (90% is the common choice for overpayment work);
- the precision or margin of error they want; and
- the expected error rate or the variability of paid amounts.
If the frame is small compared with the sample, a finite population correction is applied. RAT-STATS, SAS, and SPSS are all named in PIM §8.4.4.2 as acceptable packages. Whatever tool you use, record every input and the size it produced, so a reviewer can repeat the calculation.
PIM §8.4.1.5 also requires written approval of the methodology by a statistician or equivalent expert. Plan for that sign-off.
Step 3: Draw the sample with a recorded seed
Under PIM §8.4.2, the sample must be a probability sample. Every unit in the frame must have a known chance of selection, and that chance must be greater than zero. The manual also says that when a probability sample is properly executed, claims that it is "not statistically valid" cannot legitimately be made.
PIM §8.4.4.2 requires the contractor to identify the random number source and the program or algorithm used, record the seed value when a computer algorithm is used, and document every selection step so someone else can replicate the draw. The same frame plus the same seed produces the same sample.
Keep a copy of the frame exactly as it was at the draw. A frame rebuilt later from a live claims system may have changed, and the sample can no longer be reproduced.
Step 4: Review each sampled unit
Clinical and coding reviewers decide whether each sampled unit was paid correctly. PIM §8.4.4.4.4 requires worksheets showing the claim number, line item, amount paid, audited value, amount overpaid, and reason for disallowance. Under PIM §8.4.5, units with no records provided count as overpayments, underpayments count as negative overpayments, and a unit is set to zero when liability is waived (§1879) or the provider is without fault (§1870).
Step 5: Extrapolate with Mean Per Unit and the 90% lower bound
The most common estimator is Mean Per Unit: the average overpayment per sampled unit, multiplied by the number of units in the frame, gives the point estimate. The standard error then sets the confidence interval around it.
PIM §8.4.5 states that, in most situations, the lower limit of a one-sided 90% confidence interval should be used as the amount demanded from the provider. CMS calls this conservative: it works to the provider's advantage because the demand is very likely less than the true overpayment. The point estimate may be demanded when high precision is achieved, and a lower bound above the total paid in the frame is reduced to the total paid.
The OIG Self-Disclosure Protocol differs: it uses the mean point estimate. Know which framework applies before reporting a number.
Common challenges in provider appeals
PIM §8.4.1.1 says an appeal of the sampling methodology must show an actual error in the method that affects the overpayment amount. In practice, appeals focus on:
- Frame mismatches. Claims outside the period or issue, or a projection applied to a different population than the one sampled.
- Missing seed or selection records. A draw that cannot be replicated is hard to defend.
- Undocumented sample size or a missing statistician approval.
- Inconsistent claim decisions. Decisions without a reason code or rationale, or that differ for similar claims.
- Overturned claims not carried through. If an appeal reverses some claim decisions but upholds the method, the estimate is recomputed (PIM §8.4.9.2). If the method is found invalid, §8.4.9.1 sets out options such as correcting errors or starting a new review.
- Gaps in the work papers. PIM §8.4.4.5 requires documentation sufficient for future replication or validation by an administrative or judicial body. That includes the approved methodology, the universe, the frame, and the worksheets.
Most of these come down to record keeping: keep the parameters, seed, frame, decisions, and worksheets together with the case.
How AEGIS ISD Medical Review supports SVRS
AEGIS ISD Medical Review includes an SVRS module that runs inside the investigation workflow rather than as a separate calculator:
- Universe from the case. Claims are imported into a Case Management case as claim universes, and the frame can be narrowed by CPT or HCPCS codes. The universe size is recorded with the sample.
- Documented sample size. Cochran's formula with finite population correction, at 90%, 95%, or 99% confidence. Every input and the resulting size appear in the SVRS methodology PDF.
- Seeded selection. A seeded Fisher-Yates shuffle draws a simple random sample. The seed and the selected claims are saved with the sample, so the draw can be reproduced from the same claim universe.
- Line-level review with QA. Reviewers record a decision, reason code, and rationale per claim line, with QA Analyst verification before final determination and appeals tracked by level.
- Mean Per Unit extrapolation. AEGIS calculates the point estimate, standard error, and lower and upper confidence bounds, and highlights the lower bound as the conservative recovery target. Probe, nested, and member-universe samples use a one-sided 90% lower bound based on Student's t distribution. Each saved extrapolation worksheet is kept as an append-only run.
- Audit-ready exports. Teams can generate the SVRS methodology PDF, the extrapolation PDF, and a claims workbook, and save them to the case documents.
SVRS follows HHS-OIG RAT-STATS conventions for parameters, seed disclosure, sample-size derivation, and Mean Per Unit output, so auditors will recognize the documentation. It does not replace a qualified statistician's approval.
If your team is choosing which issues to address before payment and which to handle through post-payment review and extrapolation, read prepay vs postpay FWA detection. To see how sampling fits SIU casework, visit AEGIS ISD for SIUs.
Frequently asked questions
What makes a random sample "statistically valid"?
It must be a probability sample, meaning every unit in the sampling frame has a known chance of selection greater than zero. The draw must also be documented well enough, including the random number source and seed, that someone else can replicate it. CMS describes this standard in Program Integrity Manual chapter 8, section 8.4.2.
Why do Medicare overpayment demands use the lower bound of the 90% confidence interval?
Program Integrity Manual §8.4.5 says the lower limit of a one-sided 90% confidence interval should be used in most situations. It accounts for sampling uncertainty and produces a demand that is very likely less than the true overpayment.
Is RAT-STATS required?
No. OIG says it does not require RAT-STATS, though its Self-Disclosure Protocol strongly recommends it. The Program Integrity Manual names RAT-STATS, SAS, and SPSS as acceptable packages. What matters is documenting the tool, the algorithm, and the seed.
What is the minimum sample size?
CMS does not set one for Medicare contractor reviews. OIG's Self-Disclosure Protocol requires at least 100 items.
Sources
- CMS, Medicare Program Integrity Manual, Chapter 8 (Pub. 100-08, Rev. 13762, issued 05-27-26)
- Social Security Act §1893 (42 U.S.C. 1395ddd)
- HHS-OIG, RAT-STATS Statistical Software
- HHS-OIG, Health Care Fraud Self-Disclosure Protocol (updated Oct. 15, 2026)
- 42 CFR Part 455, Subpart A
- 42 CFR Part 455, Subpart F
See it on your cases
Request a demo to walk a sample from claim universe to extrapolation in AEGIS ISD.
Related: AEGIS SVRS · AEGIS ISD Medical Review · All resources