Every statistical sample should tell a story.

The question is whether an IRS auditor can follow it.

A defensible, statistical sampling R&D study is not built by simply selecting a random sample. It is built by creating a logical chain of evidence that begins with the company's operations and ends with a mathematically supportable estimate of qualified research expenditures.

That chain should be easy to follow.

The first step is understanding how the business actually operates.

For engineering firms, the natural sampling unit is often the project. Modern ERP systems frequently contain project-level labor, expenses, and engineering activity that allow qualified costs to be identified and sampled directly. In these cases, the objective is to refine the project population, remove clearly non-qualified or immaterial records, identify high-value projects for complete review, and statistically sample the remaining population before extrapolating the results.

Not every company has that level of detail.

Software companies and technology organizations sometimes track payroll but not employee time by project. In those situations, the methodology shifts. Departments and job titles are analyzed to identify personnel likely performing qualified research. Employees are randomly sampled and interviewed using structured time surveys that allocate their effort across projects, with each survey reconciling to 100 percent of the employee's qualified time. Payroll provides the wage base while the general ledger is analyzed separately to identify qualified supplies, contract research, and computer rental expenses.

Although the mechanics differ, the objective remains exactly the same.

Every dollar included in the credit should be traceable back to supporting documentation.

Interviews with engineering managers, project leaders, and technical personnel provide the operational context. Project documentation explains the uncertainty being resolved and the work performed. Payroll records, ERP data, and accounting records provide the financial support. Together, they create a consistent narrative that explains not only what was claimed, but why it qualifies.

The statistical component is equally important.

Revenue Procedure 2011-42 provides multiple acceptable estimators for extrapolating sampled results to an entire population. The appropriate estimator depends on the characteristics of the data—not personal preference. When multiple estimators satisfy the statistical assumptions, the procedure directs taxpayers to use the estimator with the smallest standard error. Relative precision then determines whether the estimate must be adjusted before reporting.

This is where many studies become difficult to defend.

A mathematically correct extrapolation cannot overcome a poorly defined population. Likewise, excellent interviews cannot compensate for unsupported statistical assumptions. Every component of the study must reinforce the others.

Ultimately, an IRS audit is not evaluating a spreadsheet.

It is evaluating whether an independent reviewer can understand how qualified activities were identified, why the sample represents the population, how the final estimate was calculated and if there is supporting documentation to justify what and how it was claimed.

When those questions can be answered clearly— with documentation, statistical mathematics, and a methodology that follows established guidance— the statistical sample becomes far more than a calculation.

It becomes a defensible story.