How to use RAT-STATS for fraud, waste, and abuse investigations

The Office of Inspector General’s RAT-STATS (Regulatory Audit Tool – Statistics) is a free statistical sampling package used by auditors and investigators across Medicare, Medicaid, and other federal health programs to estimate overpayments and to support findings of fraud, waste, and abuse. When a reviewer cannot examine every claim in a universe, RAT-STATS draws a defensible random sample, and then extrapolates the findings from that sample to the whole population.

The software supports simple random sampling, stratified random sampling, and cluster sampling, and can extrapolate using either a difference or ratio estimator. Outputs include a point estimate of total overpayment and a one-sided lower confidence bound, the latter being what OIG typically demands back from a provider.

Step 1: Define the universe

Before touching RAT-STATS, you need a clean universe — the complete population of claims you intend to audit. This is typically pulled from a claims system or data warehouse and should represent a specific provider, procedure, time period, and payer.

At minimum, the universe file should include:

• A unique identifier for each sampling unit
• The paid amount for each unit
• Any stratification variables (e.g., dollar tier)

Export the universe as a fixed-width .txt file or use the RAT-STATS import utility.

Step 2: Choose the sampling unit

This is the most consequential decision in the sampling design. The three main options are claims, claim lines, and patients.

Sample by claim

Use claim-level sampling when each claim represents a distinct episode of care and your review question applies to the claim as a whole — for example, whether a home health episode was medically necessary, or whether a diagnosis-related group (DRG) was coded correctly. The paid amount per sampling unit is the full claim payment.

Sample by claim line

Use claim-line sampling when claims routinely contain multiple service lines and the error is likely to be isolated to specific procedure codes billed within a claim. This is common in professional fee audits and durable medical equipment reviews, where one claim might bill a dozen HCPCS codes and only one or two are problematic. Sampling by line exposes more of the universe to review without necessarily requiring full chart pulls.

Sample by patient

Use patient-level sampling when the audit question requires evaluating the full medical record — medical necessity for a pattern of services, or suspected services that were never provided at all. Grouping claims by patient reduces logistical complexity. Instead of receiving hundreds of disconnected records, reviewers receive a single chart per sampled patient. The paid amount per unit is the total amount paid for that patient across all claims in the universe.

A good rule of thumb: if your medical reviewers will need to evaluate context across multiple dates of service, sample by patient. If each claim or line stands alone, sample by claim or line.

Step 3: Stratify the universe

Stratified sampling divides the universe into non-overlapping groups called strata and draws a separate sample from each. For overpayment work, stratify by dollar amount. This is nearly always more efficient than simple random sampling because high-dollar claims — which have the most impact on the point estimate — get reviewed at a higher rate.

A common approach:

Stratum	Description	Review rate
Certainty	Claims above a high-dollar threshold	100%
High	Upper middle dollar range	~50 sample units
Low	Remaining claims	~50 sample units

The certainty stratum is reviewed in full, which removes the largest claims from the extrapolation and reduces variance. RAT-STATS handles the math for each stratum separately and combines them into a single estimate.

There is no single correct number of strata. The OIG’s A Practical Guide to Sampling recommends two to three strata for most audits. Beyond three, the additional precision rarely justifies the added complexity.

Step 4: Determine the sample size

In RAT-STATS, open the Sample Size module and enter:

• Confidence level: 90 percent is standard for OIG work; use 95% if the findings will face litigation.
• Precision: The acceptable margin of error as a percentage of the estimated total. A common target is ±25 percent relative precision, though tighter is better.
• Expected error rate: Your best estimate of the proportion of claims with an overpayment. If unknown, use 50 percent, which maximizes the required sample size and is the most conservative assumption.
• Population size: The total number of units in the stratum.

RAT-STATS will return the minimum number of units to sample from each stratum. For most strata in health insurance audits, this will fall somewhere between 30 and 100 units. If the required sample size exceeds the stratum population, review the entire stratum.

Step 5: Draw the sample

Use the Random Numbers module in RAT-STATS to generate the sample. Enter the population size, sample size, and a random seed. RAT-STATS produces a list of record numbers that correspond to rows in your universe file.

Keep the random seed and the universe file unchanged. Both will be needed to defend the sample selection if the findings are ever challenged.

Step 6: Conduct medical review

Each sampled claim, line, or patient record goes to a clinical reviewer — typically a registered nurse, certified professional coder, or physician, depending on the review type.

For each unit, the reviewer determines:

• Paid amount: what was actually reimbursed (already in your data)
• Correct amount: what should have been paid under the applicable coverage rules
• Overpayment: the difference, if any

The result is a per-unit overpayment amount. Units with no error get a zero. Partial overpayments are possible when, for example, a service was covered but billed at the wrong rate.

Document the reason for each overpayment. If the findings are later extrapolated and the provider appeals, reviewers will need to defend each determination individually.

Step 7: Extrapolate the findings

Open the Variable Appraisal module in RAT-STATS. Enter the paid amount and overpayment amount for each sampled unit. RAT-STATS will calculate:

• Point estimate: the projected total overpayment across the entire universe
• Lower bound: the lower limit of the one-sided confidence interval at your chosen confidence level

The OIG’s longstanding practice is to demand the lower bound, not the point estimate. This protects providers from overestimates caused by sampling variance. A 90 percent lower bound means there is a 90 percent probability the true overpayment is at or above the demanded amount.

For a stratified sample, RAT-STATS combines stratum-level estimates using the appropriate weights. The certainty stratum overpayment is added directly, with no sampling error, because it was reviewed in full.

A note on defending the extrapolation

Providers have the right to appeal extrapolated overpayment demands. Common challenges include:

• Universe errors: claims that should not have been included (wrong provider, wrong period, already adjusted)
• Sampling methodology: incorrect stratification or sample selection
• Review errors: incorrect medical determinations on individual claims

The best defense against all three is documentation. Maintain the universe file, the random seed, the sample selection, and each reviewer’s determination with supporting rationale. RAT-STATS itself is well-accepted in administrative law proceedings because it follows OIG-published methodology.

---

Contact Charadata for help designing and executing statistical sampling for fraud, waste, and abuse investigations.