The AI Journal

A population that suffered terribly can leave skeletons that look healthier than a population that suffered mildly. This is not a rhetorical flourish; it follows from how skeletal lesions form. Enamel hypoplasia, cribra orbitalia, periosteal reaction on the tibia — these are markers of physiological stress severe enough to leave a mark on bone but survivable enough that the individual lived on afterward. A population under lethal, fast-acting stress will show few such lesions, because its members die before a lesion has time to form; a population under milder, chronic stress will show many, because its members survive long enough to accumulate them. Read naively, the first population looks like the healthy one. James Wood, George Milner, Henry Harpending and Kenneth Weiss set this problem out in Current Anthropology in 1992 under the name that stuck, the osteological paradox, and it reoriented three decades of paleopathology around a single, uncomfortable admission: skeletal samples are not a random draw from a past population’s health. They are a record of who a particular level of stress killed, filtered again by who a particular level of frailty let survive long enough to be buried in the sample at all. The stakes were not abstract. Whole debates about whether a shift in subsistence, a period of urban crowding, or a conquest made a population sicker had been argued from lesion frequencies read the naive way, and the paradox implied that some of those arguments could have their sign wrong.

The paper is usually narrated as a discipline’s overdue reckoning with the limits of its own evidence — the moment bioarchaeology grew up. That narration skips a detail sitting in the paper’s own apparatus. The concept doing the actual work, heterogeneous frailty — the idea that a population is not uniformly susceptible to death but contains individuals of differing, unobserved risk, and that mortality acts on this hidden variation by selectively removing the frailest first — was not new in 1992. It had been formalised thirteen years earlier, in a different field, answering a different question. James Vaupel, Kenneth Manton and Eric Stallard, writing in Demography in 1979, showed that standard life-table methods, applied to a population heterogeneous in frailty, systematically distort mortality estimates: they overstate life expectancy, understate the pace of individual ageing, and make cohorts look more alike than they are, because the observed survivors at any age are always a self-selected, hardier remnant of who started out. Vaupel and his successors built the mathematics — gamma-distributed frailty, hazard functions conditioned on latent risk — to correct for this. Wood and his co-authors did not discover that skeletal samples suffer from selection on hidden frailty. They recognised that skeletal samples suffer from the same selection problem formal demography had already named and partly solved, and they imported the diagnosis without yet importing the fix.

The gap between diagnosis and fix is the part worth dwelling on, because it was not brief. Vaupel’s frailty models existed as usable mathematics from 1979. Wood et al.’s 1992 paper cited them, by way of Vaupel and Anatoli Yashin’s 1985 work on heterogeneity’s ruses, as the conceptual source of the paradox — but the 1992 paper is, by its own design, a diagnosis rather than a method: it shows why naive skeletal samples mislead, not how to build an estimator that corrects for it on real, poorly aged, incomplete remains. The applied fix took another eight years to reach dissertation form, in Bethany Usher’s 2000 multistate model of health and mortality, built on the Tirup cemetery sample, which treated individuals as moving between observable states — no detectable lesion, detectable lesion, dead — at age-specific hazard rates that could differ between the two living states. Not until 2008 did the model meet a case dramatic enough to make its output legible outside specialist circles: Sharon DeWitte and James Wood’s study of the Black Death cemetery at East Smithfield, which used Usher’s framework to ask whether plague mortality in 1349 discriminated by preexisting frailty. It did. Individuals already carrying skeletal stress markers before the epidemic reached London faced a measurably higher risk of dying in it than individuals without them — a finding only extractable because the analysis modelled selection on frailty explicitly, rather than reading lesion frequency as a direct health signal the way pre-1992 paleopathology had.

Twenty-nine years, then, separate the demographic tool from its first fully worked archaeological application — sixteen years separate the archaeological diagnosis from that application. Bioarchaeology spent most of that interval treating the paradox as a problem about evidence and inference in the abstract: symposium papers, review articles, position statements about what skeletal samples can and cannot license a researcher to claim. Lori Wright and Christine Yoder’s 2003 review of the field’s response catalogues this literature and is itself mostly philosophical in register — a taxonomy of what the paradox does and does not undermine, running through age estimation, sex estimation, growth disruption and diet reconstruction in turn, rather than an estimator that would let a working paleopathologist actually correct a lesion count for selection bias. The tools to move past cataloguing had been sitting in demography journals the whole time, uncited by most of the paleopathology literature that spent the 1990s debating what the paradox meant rather than what to do about it.

The obvious objection is that this is not a fair comparison, because the transplant is harder than it looks. Vital registration data, the material Vaupel’s models were built on, comes with known ages, known birth cohorts, and exposure time measured directly. A skeletal sample offers none of these outright: age at death is itself an estimate, reconstructed from proxies with their own uncertainty, and that estimation error compounds with the frailty-selection problem instead of sitting outside it. Wright and Yoder are explicit that this is what stalled a straightforward import — a hazards model built for clean demographic data does not transfer cleanly onto material where the covariate of interest, age, is measured with error large enough to swamp the signal being sought. This is a real constraint, and it explains why the gap between 1992 and Usher’s dissertation was not simple negligence. But it is a constraint on execution, not on the availability of the underlying idea, and Sharon DeWitte and Christopher Stojanowski’s 2015 retrospective, marking the paradox’s own anniversary, still finds bioarchaeologists debating first principles — what a lesion can license a researcher to infer, in the abstract — that formal demography had already resolved for its own data thirty-six years earlier, in a form waiting only to be adapted rather than invented. A field can be right that its data are harder than another field’s and still be slow, for reasons that have nothing to do with the data, to go looking in that other field for the tool it already needed.