Statistical Sampling in Wage-and-Hour Defense: Building a Duran-Compliant Framework

Duran v. U.S. Bank National Assn. (2014) 59 Cal.4th 1 is frequently cited for the proposition that statistical sampling is disfavored. That is not what it says. Sampling is a legitimate tool; the decision addresses the conditions under which using it satisfies due process.

Those conditions are that the sample must be representative of the population it is used to describe, the methodology must be statistically sound, and — the requirement that does the most work in practice — the defendant must retain a meaningful opportunity to contest liability as to individual claims. Statistical methods cannot be deployed in a way that bars the presentation of valid defenses.

The trial plan in Duran failed on all three counts, and the numbers are worth carrying because they make the standard concrete. The class was two hundred sixty people. The court picked twenty names for the representative group without a pilot study, though both sides’ experts had proposed one, and added the two named plaintiffs. Twenty-one witnesses ultimately testified — nineteen class members and the two named plaintiffs — for a two-hundred-sixty-person class. From that testimony the plaintiffs’ expert derived an average of 11.87 overtime hours per week with a margin of error of plus or minus 5.14 hours, a relative margin of error of 43.3 percent. The trial court entered judgment on it anyway: $8,953,832 in overtime restitution, and $14,959,565 with prejudgment interest.

The sample was not random, and how it stopped being random is the part worth studying. The court drew the initial names at random; its later rulings compromised the draw. The two named plaintiffs sat in the group, having been chosen by class counsel to replace four earlier representatives, and the court then refused to hear from ten people who would otherwise have been in the sample — four former named plaintiffs, four members who opted out, one whose work habits differed, and one who failed to appear. The resulting sample, the Supreme Court concluded, “was not random, but appeared to be biased in plaintiffs’ favor.” Dropping the named plaintiffs would not have cured it; that would have pushed the margin of error to 47 percent.

Duran does not leave the standard abstract. It names a published California case in which sampling was used successfully to try a wage and hour class action, and the contrast is the most useful passage in the opinion for anyone building a protocol.

In Bell v. Farmers Insurance Exchange (2004) 115 Cal.App.4th 715, both sides’ experts worked for months on a mutually acceptable sampling plan. They began with a pilot sample of fifty depositions to measure variability in the population, and from that estimated that a margin of error of about one hour per week could be achieved with a sample of two hundred eighty-six. The parties ultimately deposed two hundred ninety-five and brought the margin of error below an hour a week. The two sides’ calculations of total overtime owed came out, as Duran put it, “virtually identical.”

Set against that, the Duran court’s description of its own trial plan is damning. It “chose a sample size that would be convenient and manageable. The same could be said of a sample of one. Yet convenience alone cannot justify procedures that substantially curtail the parties’ ability to litigate their case.”

The affirmative instruction follows directly, and it is the sentence to quote when the request is structure rather than dismissal: with input from the parties’ experts, the court must determine that a chosen sample size is statistically appropriate and capable of producing valid results within a reasonable margin of error. Two things about that are worth noticing. It contemplates expert input as a condition of a valid plan rather than as a courtesy, which is the strongest available basis for insisting on participation at the design stage. And it locates the defect in the plan rather than in the claim — which is exactly the posture that survives Estrada.

The vulnerabilities cluster in four places, and they are worth understanding as a defense checklist because each of them is a separate ground.

Population definition. The sample universe must match the group whose claims are being adjudicated. Universes that quietly include or exclude a category — seasonal workers, employees at a location with different practices, employees who separated during the period — produce a rate that describes a different population than the one being paid.

Selection. Randomness is a technical property, not an adjective. Samples drawn from currently employed workers, or from employees who responded to an outreach, or from a list ordered by some characteristic correlated with the violation, are not random, and the resulting bias has a direction that can usually be identified.

Violation definition. This is where the largest distortions occur, because it is a legal question dressed as a data question. Whether a twenty-nine minute meal counts as a violation, whether a meal taken late but taken counts, whether a break the employee chose to shorten counts — each answer changes the rate substantially, and each is contestable under the governing standard rather than fixed by the data.

Uncertainty. A point estimate without its confidence interval is a rhetorical device. A forty-five percent rate with a twelve point interval means the true value may be a third or may be more than half, and a model built on the point estimate has silently adopted the least favorable reading available.

The defense position is strongest when it is not opposition but participation. Every methodological choice made without defense input is a choice that will have to be attacked later at greater cost and with less credibility.

That means engaging on the sample universe before it is drawn, on the selection method before it is executed, and above all on the violation definition before the coding begins. A protocol agreed in advance — even one agreed only in part — converts later disputes from credibility contests into applications of a written standard.

It also means running the defense’s own analysis on the same data. Where the records support it, the actual violation rate is almost always lower than the blanket rate asserted in the demand, and the gap between the two is the most persuasive number available in the case. That analysis is also what makes the premium-payment defense visible: violations for which a premium was already paid are frequently still counted in the plaintiff’s rate.

Duran was a class action, and the constitutional injury it describes is framed in class terms: a procedural device may aggregate claims, but it may not be used to abridge a party’s substantive rights. A PAGA action has no class, no certification, and no predominance inquiry, which raises a fair question about how much of Duran survives the translation.

The answer is that the defendant’s interest survives intact while the mechanism changes. Estrada v. Royalty Carpet Mills (2024) 15 Cal.5th 582 held that trial courts have no inherent authority to strike a PAGA claim as unmanageable, and it declined to import class manageability requirements into a statute that contains none. But it left the case-management toolkit fully in place — limiting witnesses and evidence, permitting representative testimony, surveys, and statistical analysis — and it expressly reserved whether case management could ever so abridge the right to present a defense that due process would support striking the claim.

That reservation is where Duran now lives on the PAGA side. The argument is not that a representative claim is too hard to try, which Estrada forecloses. It is that a particular trial plan leaves no route by which the employer’s evidence could affect the outcome — the same defect Duran identified, raised against the plan rather than against the claim.

Practically, that changes what defense counsel asks for. The relief sought is not dismissal but structure: a defined sample universe, an agreed coding protocol, a stated mechanism for individualized defenses, and where none of that can be constructed, a narrower claim under section 2699(p). Framed that way the request is one a court can grant without deciding anything Estrada forbids.

Of everything in Duran, the requirement that statistical methods cannot bar the presentation of valid defenses is the one worth building the argument around, because it is the hardest for a plaintiff to engineer away.

Any representative methodology has to answer what happens when the employer has an individualized defense as to a subset of the population — a signed on-duty meal agreement, a documented waiver, a premium already paid, a supervisor who applied a different practice at one location. If the trial plan provides no mechanism for presenting that evidence, the plan is defective regardless of how sound the underlying statistics are.

That argument connects directly to scope. Where individualized defenses are numerous enough that no workable mechanism exists, the conclusion is not a better sampling protocol but a narrower claim — which is the manageability argument under section 2699(p), reached from the evidentiary side.