Duration, dispersion, and disposition in 9.1 million federal filings, 2014–2026
The median federal civil case resolves in well under a year. Duration risk concentrates in the tail: depending on case type, the slowest decile runs up to 5× the median — past eight years in antitrust — and much of that variation is associated with facts observable on filing day: case type, venue, and jury demand. We further document the waiting paradox: expected residual duration rises, not falls, with time already elapsed.
Every family resolves quickly at the median — most in under a year. The distinguishing statistic is the tail: patent's median case closes in roughly seven months, while its 90th percentile runs 4.8× longer. Reporting a median without its tail understates the risk a litigant, funder, or reserve actually carries.
“Patent cases are fast” and “patent cases take years” are both accurate. They describe different cases.
The typical patent case resolves in roughly seven months because 52% are dismissed and most of the remainder settle early. The contested minority that litigates on runs two to three years or more. The median describes the many; the tail describes the few that go the distance — and aggregate statistics that do not separate the two mislead in both directions.
Holding case type fixed, the fastest and slowest districts differ by months — and, for personal injury, by nearly four years. The personal-injury extreme is a multidistrict-litigation artifact (consolidated mass-tort dockets in E.D. La.), reported here because unlabeled venue statistics routinely absorb it unremarked.
Settlement is the default assumption for civil litigation, but it is not the most common ending everywhere. In patent, dismissal — not settlement — is the single most likely disposition; in social security appeals, judgment dominates. Priors calibrated on one case type transfer poorly to another.
The fitted survival model carries a single multiplicative jury-demand term — ×1.36 on expected duration, entering as a main effect, so the proportional gap is constant across case types by construction rather than separately estimated for each. We report it as what it is: a fitted association the model conditions on — jury-demanded cases differ from bench cases in ways beyond the demand itself, and no causal claim is made.
This is not a paradox in the formal sense: heavy-tailed duration distributions have rising residual life mechanically, and the pool of still-open cases shifts toward intrinsically slow matters as it ages. We do not apportion the rise between those mechanisms. As a contract case ages, its median remaining time climbs — the curve shown is the engine's conditional estimate, recalibrated against survivor cohorts and scored out of sample at 1, 2, and 3 years elapsed.
9,138,235 rows were ingested from the FJC Integrated Data Base. The benchmark pool comprises 2,824,382 cases from 2014–2026 filing cohorts. Durations are estimated by Kaplan–Meier, so pending cases contribute exposure time and long-running matters cannot vanish from the percentiles. Disposition shares are computed on cohorts old enough to be fully observed (2,439,427 resolutions).
The model-derived findings — the jury-demand multiplier (05) and residual-duration profile (06) — come from tertius-acta-4, a parametric survival model whose out-of-sample record is public: on 1,370,792 held-out cases it placed 77.3% inside its 80% prediction interval, with a 114-day median absolute error. The full record, including per-group calibration, is at usetertius.com/algorithm.
Source. The Federal Judicial Center's Integrated Data Base (IDB) records every civil case filed and terminated in U.S. district courts.
Validation. Cleaned durations are cross-checked against the Administrative Office of the U.S. Courts' published Table C-5 statistics (Spearman 0.86–0.88 by district).
Case types. Grouped from IDB nature-of-suit codes into the 13 families shown.
Scope. Timing and disposition only — never merits or damages.
All figures describe populations of filed cases. Associations — venue, jury demand — are conditioning variables in a fitted model, not causal effects of choosing differently.
Cases still open at the data cutoff are right-censored and retained via Kaplan–Meier, which assumes censoring is uninformative within cohort. Percentiles beyond observed follow-up are not reported.
The 13 families aggregate finer nature-of-suit codes; within-family heterogeneity (e.g., FLSA vs. discrimination inside employment) is not shown here.
District cells in mass-tort-heavy venues (E.D. La. personal injury) reflect multidistrict-litigation consolidation, not ordinary single-case dockets, and are flagged where shown.
The benchmark tables behind this note are public, citable, and queryable at the single-case level.