Uncertainty due to model choice in variant Creutzfeldt-Jakob disease projections

For some statistical applications, uncertainty due to unverifiable assumptions can be much greater than that arising from the random variability that is quantified by conventional confidence intervals. The case of projecting possible maximum numbers of eventual deaths due to variant Creutzfeldt-Jakob disease in the United Kingdom provides an extreme example of this phenomenon. The need for parametric extrapolation of the incubation distribution, along with non-identifiability of the number of infected persons, makes assumptions very influential. Several publications in leading science journals gave upper bounds that were 100-fold to 20 000-fold lower than projections from other plausible models given here that fit the data about as well. The crucial assumption for projections is how the risk of death increases with time since infection, and exponential growth is an obvious choice for pessimistic models. Parametric extrapolation from the generalized lambda, generalized F, or lognormal distributions produced upper bounds much lower than exponential extrapolation (i.e. assuming a Gompertz distribution, which fit the data up to early 2002 quite well). Had the publications considered such possibilities, they would have reached much weaker conclusions and been less suitable for leading general science journals. The scientific publication process may have inherent disincentives for thorough assessment of uncertainty.