On diversification curves and patience
10 April 2026 · methods, birth–death models
A rate shift is a hypothesis, not a discovery. The reversible-jump machinery will happily hand you three shifts in a record that contains one genuine change and two sampling artifacts, and the posterior will look confident about all three. The discipline is in what you do next: check the chain, check the ESS, check whether the shift times move when you change the preservation model, and then — hardest of all — check whether you would still believe the shift if it were in a clade you did not care about.
I have been re-running the Paleozoic fish analyses with the time-dependent preservation model, and the familiar Late Devonian extinction signal softens, as it should, because the Frasnian–Famennian record is also a record of facies change. Softening is not disappearing. The association with environmental covariates survives, but the wording has to survive with it: consistent with, covaries with, not "driven by". The difference between those phrases is the difference between a result and a story.
Patience enters twice. First in the MCMC itself — the chains need the generations they need, and thinning is not a substitute for running longer. Second in interpretation: a curve through deep time is a summary of summaries, and every layer of smoothing is a decision that should be written down where a reviewer can find it. My rule this year is that every figure ships with its binning and interpolation choices in the caption or the supplement, no exceptions.