Statistics Expert Witness

Overview

About Statistics Expert Witnesses

A statistics expert decides whether a set of numbers can carry the inference a party wants to draw from them. The work is how the data were collected, how the sample was chosen, what the model controlled for, and whether the result separates a pattern from chance. This page covers statistical method. A forensic economist or damages expert may use the same tools to price a loss; the statistician tests whether the method behind the number holds up.

Sampling: When a Few Stand for Many

Sampling disputes start with two questions from the Reference Manual on Scientific Evidence, the guide published by the Federal Judicial Center and the National Academies: what method selected the units, and which of the selected units actually provided measurements ¹. The manual explains that probability sampling gives a sample that is representative of the sampling frame within the limits of chance ¹. It also warns that selection bias is acute when people volunteer their answers or when attorneys choose cases for trial, and that those who refuse to take part create nonresponse bias ¹.

Whether a sample can prove a claim for a group is a legal question with a statistical core. In Tyson Foods v. Bouaphakeo (2016), the Supreme Court refused to adopt a broad rule for or against representative evidence. One way to show a sample is permissible, the Court said, is that each class member could have relied on it in an individual action ². The Court also said that representative evidence that is "statistically inadequate or based on implausible assumptions" could not lead to a fair estimate ². In that case the employees worked in the same facility, did similar work and were paid under the same policy ².

So the expert reads the sampling plan, the frame, the selection records, the response rate and the variation inside the population, and says whether the average describes the people it is offered to describe.

Regression: What the Model Left Out

Regression fights turn on specification. In Bazemore v. Friday (1986), the Court said a regression need not include all measurable variables, and that "failure to include variables will affect the analysis' probativeness, not its admissibility" ³. It added a limit: some regressions are "so incomplete as to be inadmissible as irrelevant" ³. The manual explains the mechanism. An omitted variable may cause an included variable to be incorrectly credited with an effect the excluded one caused, and the concern is greatest when the omitted variable is correlated with both the outcome and the key explanatory variable ¹.

Three more questions recur:

• Significance. The Supreme Court has described a result more than "two or three standard deviations" from its expected value as suspect, and the manual notes those differences produce p-values of about 5% and 0.3% for a normally distributed statistic ¹
• Size. A significant result may show that something beyond random error is at work. It is not evidence that the something is legally or practically important ¹
• Multiple testing. The manual warns that no general solution exists for the typical case where analysts test and reject a variety of models before settling on one, and says those comparisons should be disclosed ¹

Class Certification: Does the Model Fit the Theory

At certification the question is fit. In Comcast v. Behrend (2013), a regression damages model assumed four theories of antitrust impact when only one survived. The Court held that a model offered as evidence of damages "must measure only those damages attributable to that theory" ⁴. In Wal-Mart v. Dukes (2011), the plaintiffs offered regressions at the regional and national level. The Court quoted the view that such disparities do not establish disparities at individual stores, and found the inference to a company-wide policy failed ⁵. The Court also said it doubted that Daubert did not apply at the certification stage ⁵.

In 2025 the Court took up whether a federal court may certify a damages class holding both injured and uninjured members, then dismissed the case without deciding the question ⁶. Rule 702, amended in 2023, requires the proponent to show it is more likely than not that the testimony meets the rule, and the committee note says each opinion must stay within the bounds of a reliable application of the method ⁷.

Who Is Qualified

No license is required to practice statistics. The manual says anyone with substantial training in and experience with regression and other statistical methods may qualify, that a doctoral degree in a discipline teaching statistics usually meets the preliminary test, and that an expert whose only tool is regression may be poorly placed to judge when another method fits ¹. The American Statistical Association offers voluntary accreditation, PStat, which calls for an advanced degree in statistics or a related quantitative field and at least five years of documented experience ⁸. It is a useful signal, not a requirement.

Neighbors do different work. A forensic economist or lost profits expert prices the harm. An actuary models insured risk. A survey expert designs and defends consumer surveys. A statistician can review any of their numbers, but those experts answer the merits question within their field.

Qualifications

Qualifications to Look For

• ✓Doctoral or master's degree in statistics, biostatistics, econometrics or a field that teaches applied statistics
• ✓American Statistical Association PStat accreditation where held; it is voluntary and no license exists
• ✓Practical experience designing samples and defending them, not only analyzing data handed over by counsel
• ✓Command of more than one method, so regression is chosen for a reason and not by default
• ✓Experience with the data systems at issue, such as payroll, claims or transaction records
• ✓Published or peer-reviewed work in applied statistics or the relevant field
• ✓Prior deposition or trial testimony, with reports that disclose the models tested and rejected

FAQ

Frequently Asked Questions