A comparison of methods to assess selective disappearance and quantify ageing

1. Age-dependent change in traits at the population-level level can diverge from the within-individual ageing trajectory due to selective disappearance, the biased removal or death of individuals with certain phenotypes. 2. Commonly used methods to assess, and account for, selective disappearance have been developed for relatively simple settings. As such, we currently lack a clear understanding of how well these methods correct for bias when selective disappearance itself is age-dependent, or of their accuracy in quantifying ageing under biologically realistic systems. 3. We use simulations and analytical solutions to quantify and compare how well existing methods (which we term: “additive covariate”, “mean-centring”, and “decomposition”) recapitulate the true ageing trajectory under a range of sampling designs and organismal life-histories. Using two empirical datasets, we further illustrate how the inferred ageing pattern can depend on the specific statistical model used. 4. We show that when selective disappearance is driven by the removal of individuals with distinct ageing trajectories, the decomposition method’s estimate of ageing diverges substantially from the true underlying ageing pattern. Here, the additive covariate and mean-centring methods, which typically only account for age-independent selective disappearance based on an individual’s average quality, are also biased. Moreover, mean-centring based models and the decomposition method perform especially poorly when sampling is sparse. 5. Our study quantitatively and systematically synthesizes when and why typically used methods to quantify ageing might lead to inaccurate conclusions. We suggest that a model where age of last sampling is included in interaction with age should be used to quantify ageing and selective disappearance. This model best recapitulates the true ageing trajectory irrespective of organismal life-history, the type of selective disappearance, or missingness in data.