The Denominator Problem in Statistical Evidence

Why counts of bad events can mislead in litigation unless lawyers ask what population, exposure, or opportunity produced the count.
Author

James G. Scott

Why Counts Can Mislead in Litigation

In many cases, the first number to appear is a count. A regulator identifies 1,200 denied insurance claims. A plaintiff points to 87 customer complaints about a product. A government audit lists 4,600 allegedly improper bills. Counts like this have force because they look concrete, like a definitive tally of misconduct or misery.

But the count is only the legally sexy half of a rate. The missing half is the denominator: the population of opportunities, transactions, claims, products, patients, employees, or applications from which the count arose. Without that denominator, it may be impossible to determine whether the count is alarming or thoroughly ordinary.

This is a recurring problem in litigation. Parties argue about the number of bad events without first establishing the number of chances for those events to occur.

The Denominator Behind the Count

It is tempting to treat the denominator as bookkeeping. Count the bad events, count the total events. Divide one into the other and call it a day.

But product-liability cases show why the missing denominator can be harder to define than it first appears. Suppose a plaintiff identifies fifty battery fires involving a consumer device. The number sounds catastrophic, and in human terms it may be. But the statistical inference depends on the product population and the relevant exposure. Fifty fires among five thousand units suggests a very different risk profile from fifty fires among fifty million units. The same count may also carry different meaning depending on whether the devices were used once a week or every day, charged overnight or intermittently, paired with an approved charger or a third-party accessory, and drawn from the full production run or a particular manufacturing period.

That is why “units sold” is not always the right denominator. The relevant denominator may be charging sessions, device-years of use, units exposed to a particular condition, or units manufactured with a particular component. A denominator should correspond to the opportunity for the alleged failure to occur. Otherwise the analysis may compare fifty events to a population that was never meaningfully at risk.

Apples in the numerator, oranges in the denominator

Even a correctly computed rate can mislead if the numerator and denominator are not aligned. The numerator counts events with one definition; the denominator must count opportunities under the same definition.

Suppose a consumer class action challenges a subscription service’s cancellation process. The plaintiffs count complaints from customers who could not cancel online. The company divides by all subscribers. That denominator may be too large if many subscribers never attempted cancellation. The legally relevant question may concern the rate of failed cancellation among customers who tried to cancel, not among everyone who ever subscribed.

Now reverse the problem. Plaintiffs divide the complaints by the number of customers who contacted customer service about cancellation. That denominator may be too small if many customers successfully canceled online and never contacted support. The rate may then describe the experience of customers already having trouble, not the experience of canceling customers generally.

The litigation lesson

A count may be enough to justify investigation or support discovery. But when the count is used to quantify risk, prove systemic conduct, or estimate damages, the denominator becomes part of the proof.

The basic question about any count is: out of what? That question leads quickly to others. What exactly was counted in the numerator? What opportunities were included in the denominator? Are the numerator and denominator measuring events and opportunities according to the same definition? Would a different denominator change the legal meaning of the rate? These are far from technical afterthoughts. They determine whether the count supports the inference being drawn from it. A pile of adverse examples can be powerful evidence, but only after the court knows the population from which the pile was drawn.