Which disputes reach trial

Why do the cases that reach trial look nothing like the disputes people actually have?

If you wanted to know what goes wrong between people — how often landlords cheat, how often doctors err, how the law is actually applied — the trial record looks like the obvious place to read it. Disputes go to court; courts hold trials; trials are written down.

But the funnel is brutal. In the US federal courts, civil trials fell from about eleven and a half percent of case dispositions in 1962 to under two percent by 2002, and further since. So a trial is not a sample of disputes. It is what survives a long sequence of decisions to stop.

And here is the part that should make you uneasy. If the survivors were a random few in a hundred, we could still read them and multiply. Ask yourself whether anything about that sequence looks random.

Work backwards from the end. Two parties are in front of a trial date. Both know a trial is expensive, slow, and public. Both can settle instead. When would a rational pair refuse the cheaper option?

Suppose both sides look at the same case and both privately conclude the plaintiff will probably win, and roughly by how much. Then there is an obvious deal: some figure near that expected outcome, minus the costs both save. Both prefer it. They settle.

Now suppose they disagree — the plaintiff is confident of winning, the defendant equally confident of a defence. Each thinks the other's offer is far too mean. There is no number both will accept, so the case goes on.

Notice what that gives us. Cases do not proceed to trial because they are important, or egregious, or typical. They proceed because the parties could not agree about how they would come out. Trials are what is left when prediction fails.

That is the core of the selection hypothesis set out by George Priest and Benjamin Klein in 1984, and it has a striking corollary. If settlement removes the cases whose outcome is obvious, what remains at trial is disproportionately the close ones — the cases sitting near the decision boundary, where reasonable people genuinely differ. And the closer a case is to the boundary, the nearer to a coin flip its result should be. Priest and Klein's strong prediction was that plaintiff win rates at trial should tend toward fifty percent regardless of how the underlying legal standard is set.

Now here is the honest part. That prediction has been tested for four decades and it does not hold cleanly. Observed win rates vary widely by case type — medical malpractice plaintiffs, for instance, win at trial only around a fifth to a quarter of the time. The reasons offered are that the model assumes symmetric stakes and symmetric information, and real disputes violate both: a repeat-player insurer with better data, a defendant protecting a precedent across a thousand future cases, or a plaintiff who cannot afford to wait, will all trade differently from the model's parties. What survives the criticism is not the fifty percent number but the mechanism — that trials are a selected set, chosen by disagreement — and that part is not seriously contested.

Then follow the consequence, which is the payoff. Suppose a legislature tightens the standard of care so that far more conduct counts as negligent. What happens to the plaintiff win rate at trial? Very plausibly nothing at all. The newly-clear cases get settled, because now both sides can predict them. The rule changed enormously; the visible statistic did not move. Reading the trial record as a thermometer would have told you the reform failed.

Three refinements.

First, the settlement stage is only the last filter. Long before it, most grievances never become claims at all: the sociological work on the dispute pyramid describes the sequence in which an injury must first be perceived, then attributed to someone, then voiced as a claim, before a lawyer is anywhere near it. Each of those stages filters non-randomly too — by resources, information and confidence — so the case population was already unrepresentative before the settlement logic touched it.

Second, resist the misreading that selection means trial outcomes are meaningless. They are perfectly meaningful about the tried cases. The error is only in the inference outward — from tried cases to all disputes, or from a win rate to the state of the law.

Third, the mechanism cuts both ways for reformers. If you want to know whether a rule change worked, the trial record is close to the worst available evidence, because the reform's clearest successes are exactly the cases that will now never be tried. Settlement amounts, filing rates and claim rates are noisier data but far less filtered.

• Asymmetric stakes, and how a defendant litigating for precedent behaves differently from one litigating for money.
• Whether the same selection logic explains which cases produce published appellate opinions — a second filter on top of this one.
• Measuring legal change through settlement and claim data instead of trial outcomes.

Every term the collection defines is gathered in the glossary.