Statistical evidence in discrimination cases is how plaintiffs turn thousands of individual hiring, pay, and promotion decisions into a measurable pattern a court can evaluate. Direct proof of discriminatory intent is rare, so federal courts have accepted a handful of mathematical methods, each with its own strengths and weak spots, for showing that outcomes for a protected group deviate from what a fair process would produce. The methods themselves matter, but so do the rules about which comparison group to use, how large the sample needs to be, and whether the expert presenting the numbers can get past the courtroom door.
Two Kinds of Claims, Two Uses of the Numbers
Statistics do different work depending on what the plaintiff is trying to prove. In a disparate treatment case, the allegation is intentional discrimination, and aggregate data is used to show a “pattern or practice” that individual anecdotes alone couldn’t establish. The Supreme Court approved this approach in International Brotherhood of Teamsters v. United States, finding that a stark statistical gap between minority and white employees in desirable jobs, together with testimony from individual workers, was enough to prove intentional discrimination across an organization.1Legal Information Institute. International Brotherhood of Teamsters v. United States
Disparate impact claims use statistics differently. Intent drops out of the picture; the question is whether a facially neutral policy, such as a written exam or a minimum education requirement, screens out a protected group at a significantly higher rate. Title VII puts the initial burden on the plaintiff to identify a specific employment practice causing the disparity. If the plaintiff clears that hurdle, the employer must prove the practice is job-related and consistent with business necessity, and even then the plaintiff can still win by showing a less discriminatory alternative would serve the same business purpose. A plaintiff cannot simply point to a workforce imbalance; the specific practice must be named. The one exception is where the components of the employer’s decision-making are so entangled they can’t be analyzed separately, in which case the whole process is treated as a single practice.2Office of the Law Revision Counsel. 42 USC 2000e-2 – Unlawful Employment Practices
The Four-Fifths Rule
The simplest screening tool comes from the EEOC’s Uniform Guidelines on Employee Selection Procedures. Under 29 C.F.R. ยง 1607.4, if the selection rate for any race, sex, or ethnic group falls below 80% of the rate for the highest-selected group, federal enforcement agencies will generally treat that as evidence of adverse impact.3eCFR. 29 CFR 1607.4 – Information on Impact
The math is straightforward. If an employer interviews 100 men and hires 20, the selection rate for men is 20%. If it interviews 100 women and hires 12, the rate for women is 12%. Dividing 12% by 20% gives 60%, which falls below the 80% threshold and triggers the rule. Hiring 16 women would put the ratio at exactly 80% and clear it.
A four-fifths violation is a red flag, not a finding of discrimination. And federal appellate courts have repeatedly warned that the rule is unreliable with small applicant pools, where a handful of decisions can produce lopsided ratios through random variation alone. For serious litigation, courts generally want more rigorous proof.
Standard Deviation and Statistical Significance
The method courts prefer when real weight is riding on the numbers measures how far observed outcomes depart from what a nondiscriminatory process would produce, expressed in standard deviations. In Castaneda v. Partida, the Supreme Court said that a gap of two or three standard deviations between expected and observed results would make the hypothesis of race-neutral decision-making “suspect to a social scientist.” The Court imported that same standard into Title VII law in Hazelwood School District v. United States, and it has been the benchmark ever since.4Justia Law. Hazelwood School District v. United States, 433 US 299
In probability terms, a two-standard-deviation gap corresponds roughly to a 5% chance that the result occurred by random variation; three standard deviations drops that probability to about 0.3%. Statisticians call this the p-value, and by convention in both science and law, a p-value at or below 0.05 is treated as statistically significant.5National Center for Biotechnology Information. When the Alpha is the Omega – P-Values, Substantial Evidence, and the 0.05 Standard at FDA
Sample size does a lot of the heavy lifting here. In a large workforce, even a modest disparity can clear the two-standard-deviation bar because random variation has less room to explain things. In a small group, a much larger gap is needed before chance can be ruled out. This is why a department of 15 people rarely produces usable statistical proof regardless of how skewed the numbers look, and why class actions covering thousands of employees generate the strongest cases.
Multiple Regression Analysis
Regression is the most powerful statistical tool in discrimination litigation because it can isolate the effect of a protected characteristic after accounting for legitimate explanations. A raw $10,000 pay gap between men and women at a company proves little on its own; it ignores job title, tenure, education, and performance. A regression model holds those variables constant and asks whether a gap based on sex or race survives once the legitimate factors are removed.
The Supreme Court addressed regression directly in Bazemore v. Friday. The lower court had rejected the plaintiffs’ regression because it did not include every variable that might affect salary. The Supreme Court reversed, holding that omitting variables makes a regression less persuasive but not inadmissible, so long as the model accounts for the “major factors.”6Justia Law. Bazemore v. Friday, 478 US 385 A plaintiff’s model doesn’t need to be perfect, but it does need to capture the variables that obviously matter. Leaving out job title in a pay case, for example, would almost certainly be fatal.
The defense side of Bazemore matters just as much. An employer can attack a plaintiff’s regression by showing that adding a key omitted variable makes the disparity disappear. Employers also run their own regressions and reach opposite conclusions by choosing different variables or defining the workforce differently. Many discrimination trials come down to this contest of competing models, with the judge or jury deciding whose statistical story is more credible.
Choosing the Right Comparison Group
No statistical method produces meaningful results if the comparison group is wrong. In Hazelwood, the Supreme Court held that a school district’s teaching staff had to be compared to the pool of qualified teachers in the relevant labor market, not the general population of the surrounding area.4Justia Law. Hazelwood School District v. United States, 433 US 299 In Wards Cove Packing Co. v. Atonio, the Court rejected a comparison between the racial makeup of cannery workers and non-cannery workers at the same company, reaffirming that the right benchmark is the racial composition of qualified people in the relevant labor market.7Justia Law. Wards Cove Packing Co. v. Atonio, 490 US 642
The principle is intuitive. Comparing a hospital’s surgeon demographics to the general city population ignores the fact that surgeons need years of specialized training; the right pool is board-certified surgeons in the geographic area where the hospital recruits. An over-inclusive pool makes almost any employer look discriminatory. An under-inclusive one can hide real exclusion.
Applicant Flow vs. Labor Market Data
Two methods compete for defining the comparison group. Applicant flow analysis looks only at people who actually applied and compares selection rates among them. Some courts have called this the preferred method because it reflects real-world interest in the position.8The University of Chicago Law Review. Statistical Evidence in Discrimination Cases – Standards and Methods of Proof Labor market analysis compares the employer’s workforce to the broader pool of qualified workers in the area, applicant or not.
Applicant flow data has a notorious weakness: if an employer’s reputation discourages minorities from applying, the applicant pool is already skewed and selection rates within it will look clean. The worst discriminators can end up looking blameless. Labor market data avoids that trap but opens its own fights about geographic boundaries and required qualifications. Strong cases typically use both.
When Statistics Need a Human Voice
Numbers alone can prove discrimination when they are extreme, but most cases combine the data with the experiences of real people. EEOC guidance describes individual testimony as what “brings the cold numbers convincingly to life.”9U.S. Equal Employment Opportunity Commission. Section 15 Race and Color Discrimination A regression showing a persistent pay gap is persuasive; the same regression paired with testimony from women told their husbands earned enough is devastating.
The one scenario where statistics regularly carry the full load is the “inexorable zero,” a term courts use when a protected group is entirely absent from a workforce or job category over a sustained period despite being present in the qualified labor pool. Short of that kind of total exclusion, expect courts to want both the math and the stories.
Getting the Expert Past Daubert
Statistical evidence almost always reaches a jury through an expert witness, which means it first has to clear an admissibility threshold. Under Daubert v. Merrell Dow Pharmaceuticals, federal judges act as gatekeepers for expert testimony, screening it for reliability and relevance before letting the jury hear it.10Legal Information Institute. Daubert v. Merrell Dow Pharmaceuticals, 509 US 579
Federal Rule of Evidence 702 codifies the standard. The proponent must show it is more likely than not that the expert’s knowledge will help the jury, that the testimony rests on sufficient facts or data, that the methods are reliable, and that the expert applied those methods reliably to the case.11Legal Information Institute. Rule 702 – Testimony by Expert Witnesses For a statistician, that means judges look at whether the methodology has been tested and peer-reviewed, whether it has a known error rate, and whether it is generally accepted in the field.
Daubert challenges are a standard defense move. Employers routinely file pretrial motions to exclude the plaintiff’s statistical expert, arguing that the model used the wrong comparison group, omitted key variables, or applied an inappropriate methodology. If the judge agrees and strikes the expert, the plaintiff often has no case left.
How Employers Fight the Numbers
Defense teams rarely rely on a single line of attack. The common strategies target the analysis itself rather than the underlying raw data.
- Wrong comparison group. The employer argues the plaintiff’s labor pool is too broad or too narrow, inflating the expected representation of a protected group. Wards Cove gave employers significant ammunition by requiring plaintiffs to compare against the qualified population rather than the general population.7Justia Law. Wards Cove Packing Co. v. Atonio, 490 US 642
- Omitted variables. The employer identifies a factor the plaintiff’s regression failed to account for, such as shift preference, part-time status, or geographic location. Under Bazemore, omitted variables go to weight rather than admissibility, but an employer who can show that adding a key variable eliminates the disparity effectively neutralizes the plaintiff’s model.6Justia Law. Bazemore v. Friday, 478 US 385
- Small sample size. The employer argues the numbers are too small to support reliable conclusions. Courts have rejected both the four-fifths rule and standard deviation analysis in cases built on tiny datasets.
- Business necessity. Even where the statistics show adverse impact, the employer can justify the challenged practice by proving it is job-related and consistent with business necessity.2Office of the Law Revision Counsel. 42 USC 2000e-2 – Unlawful Employment Practices
- Competing analysis. The employer presents its own regression or statistical model reaching a different conclusion, leaving the factfinder to pick between dueling experts.
The business necessity defense has an important limit. It cannot be used against a claim of intentional discrimination. If the plaintiff is alleging deliberate disparate treatment rather than a facially neutral policy with disparate effects, showing a business purpose for the practice does not help.2Office of the Law Revision Counsel. 42 USC 2000e-2 – Unlawful Employment Practices