In their Letter, Dong et al.1 claimed to have found evidence for a natural limit to human longevity, based on analysis of yearly maximum reported age at death (MRAD) values over the period 1968–2006 (see figure 2a in ref. 1). After splitting the period into two ranges (1968–1994 and 1995–2006), the authors fitted linear regressions to each group separately and concluded that the MRAD no longer increased significantly after 1995, and instead slightly decreased by 0.28 years each year1. There are strong statistical grounds to question the validity of their conclusions, which we illustrate by an analogous treatment of historical data on athletic performance. There is a Reply to this Comment by Dong, X. et al. Nature546, http://dx.doi.org/10.1038/nature22789 (2017).
Whether or not human lifespan is subject to an intrinsic limit has long been discussed. Current understanding of the biology of ageing points firmly away from any idea that the end of life is itself genetically programmed2. Ageing results from progressive accumulation of multiple forms of damage, many of which arise as inevitable by-products of the molecular and cellular processes that underpin living systems. As time goes by, the burden of damage grows and the risk of dying increases. In this sense, the extremes of human longevity resemble the records for physical performance in other spheres, such as sport.
In 1968, Bob Beamon set a male long-jump record that would endure for more than 20 years (see Fig. 1, which is based on records from the Olympics (https://www.olympic.org/olympic-results), and the International Association of Athletics Federations World Championships (https://www.iaaf.org/competitions/iaaf-world-championships)). In 1991, Beamon’s record was exceeded by Mike Powell, whose long-jump record still stands. From Fig. 1 it might seem that after the world record in 1991, a decline in long-jump performances set in. If we partition the data in two groups, before and after 1991, and apply linear regression followed by two-tailed statistical testing, an argument is provided for rejecting the null hypothesis that there is no change, thereby pointing to an improvement in performance up to 1991 and deterioration thereafter. But when we apply a similar analysis to the whole of the data since 1960, we find a statistically significant increase in the winning long-jump distances over time.
There is marked similarity between the patterns seen in our Fig. 1 and figure 2a of ref. 1. We make the comparison not because there is any specific reason to pick long-jump records as being intrinsically important among other datasets of a comparable nature, but simply to illustrate the fundamental nature of the statistical reasoning that has been applied, we think questionably, to the MRAD values. It is the analytical principle that matters, not the case to which it has been applied. Just as we should be cautious to accept our first analysis of the long-jump data suggesting a decline, especially when most other athletic sports have shown progressive improvement, we believe that on a question of such perennial interest and importance as human longevity, a more careful approach is called for before claiming evidence of a limit.
First, Dong et al.1 committed a basic statistical error by using the same dataset both to propose the hypothesis that there has been a change in the trends of human longevity occurring around 1995 and also to test it. It is well known that such a procedure leads to a false assessment of statistical significance.
Second, the idea of a set limit to human longevity is not strongly supported by what is being discovered about the biology of ageing. The continuing increase in human life expectancy that has occurred over recent decades was unforeseen. It provides evidence for greater malleability of human ageing than was originally thought. In the light of this evidence, we suggest the claim that the maximum human lifespan is fixed is improbable a priori, and therefore, even in the absence of statistical error, should be regarded with appropriate caution.
We conclude that it is premature to make inference based on data from 1995–2006, spanning barely a decade, and too soon to know what might be revealed by MRAD values in the decades to come. Improvements in underlying survival probabilities at old age show no deceleration, stall or decline. In technical terms, the authors have committed a statistical ‘error of the third kind’ by correctly rejecting the null hypothesis of no change but erroneously inferring that there is a limit to, or even a decline in, maximum lifespan, while in reality there is the strong possibility of a continuing increase3, 4, 5.