Reading bias in the fossil record

It is fashionable to treat sampling bias as a nuisance parameter — something to be estimated, corrected, and then forgotten. I think this gets the epistemology backwards. Collection effort is not noise laid over a clean signal; it is part of the process that produced the record we have. The Devonian fish curve is a curve of fishes and of quarries, of monographs, of which formations happened to lie near nineteenth-century rail lines.

Specimen-level databases change what is possible here. When every occurrence keeps its voucher, its locality, and its literature source, you can ask questions that aggregated richness curves silently forbid: which intervals are rich because many animals lived then, and which are rich because one productive lens was worked for forty years? The two patterns look identical at the stage level and completely different at the collection level.

None of this means diversification models are useless. It means their outputs should be read as statements about the sampled record, with the sampling model carrying as much interpretive weight as the birth–death model. When a preservation model and a rate model trade off against each other — and they do, constantly — the honest report is the trade-off, not whichever corner of it produces the prettier curve.

Practical note to self: before the next covariate analysis, tabulate collection counts per bin alongside richness, and refuse to interpret any rate shift that sits exactly on a sampling cliff. The environment did not necessarily do anything at that boundary. A railway might have.