The story of civilizations is usually told as a life cycle: rise, maturity, decadence, fall. If that were literally true, an old empire would face a higher yearly risk of ending than a young one. Samuel Arbesman tested it in 2011 on Rein Taagepera's list of 41 ancient empires. He found the opposite of a life cycle: lifespans fit an exponential distribution, the signature of a risk that does not change with age, with a mean of about 220 years.
Forty-one empires is a small sample. Seshat's Cliopatria dataset (CC BY 4.0) now maps 1,587 named polities from 3400 BCE to 2024. We ran the same question on it in civ-patterns, using standard survival analysis that handles states still alive today.
Big empires really are memoryless. For the 224 polities that ever held a million square kilometres, the Weibull shape parameter is 0.91, with a 95% interval of 0.82 to 1.02. A value of 1 means a constant risk, so Arbesman's result survives a sample five times larger. Their yearly risk of ending sits between 0.55% and 1.2% at every age we can measure (the highest, after age 400, rests on 14 endings).
Small and new states die young. Across all 1,543 polities, the risk is 1.1% a year in the first 25 years. It then settles near 0.6% a year, about a 46% chance of ending in any given century, and stays there until age 400. Only the few that pass 800 years show a lower rate. There is infant mortality, then a long flat plateau, and no sign of old age.
Size at birth does not protect. Polities that started out larger did not last longer. Each tenfold increase in initial territory raised the hazard by 6% (interval 0 to 13%).
The era matters more than the age. We asked a forecasting question. Using only data before 1600, 1700, 1800 or 1900, could a duration model tell which of the states then on the map would still be there a century later? Mostly it could not. Within each cutoff, the models ranked survivors little better than a coin flip (AUC 0.46 to 0.61). What changed was the whole world's turnover rate. Of the states on the map in 1600, 35% were gone by 1700. Of those in 1800 and 1900, 78% and 79% were gone a century later: colonisation, the unifications of Germany and Italy, two world wars and decolonisation.
Applied to today's 161 mapped states, the age-only model expects about half to survive under the same name to 2124. The backtest says to read that as a wide range: about 65% if the coming century is as calm as the 17th, about 20% if it churns like the 19th or 20th.
Caveats. Cliopatria ends a polity when its name changes, so "the end of a state" here includes renamings and successions, not only collapses. Before 1000 BCE it maps a snapshot only every century or so. And a model of age alone ignores everything a historian would look at. Code, data handling and every number are in the project README. Produced with AI assistance (Claude).