Chapter 1. What Counts as the Target, What Counts as Evidence, Who Decides and What Reopens It

The texts answer the paper's questions in a particular order of confidence. They are most confident that ageing is a fit target, less confident about what would show it had been slowed, and least explicit about who should decide. On what would reopen a decision they are, by contrast, precise: the measurement papers in particular write their own conditions for being doubted. This chapter says how the texts were chosen and read, then takes the questions in turn.

1.1Data and method

The paper reads nine texts, chosen by one rule: every text saved for this study on the question of slowing ageing whose licence allows it to be quoted, that is, every work of the United States government (which is not subject to copyright) and every article published under a Creative Commons Attribution licence. The list, with each licence and the evidence for it, is kept beside the texts in a manifest. Four further documents that the study tried to retrieve (two pages of the National Institute on Aging, a licence service and the publisher's text of a review) could not be read from this session and are not used.

One text is a government work: the programme page of PROSPR (Proactive Solutions for Prolonging Resilience) at the Advanced Research Projects Agency for Health, ARPA-H, saved as text from the agency's site. Eight are articles read in full from PubMed Central, seven under CC BY 4.0 and one, Horvath's, under CC BY 2.0: those of Gems, Barzilai, Scott and others, Horvath, Belsky and others, Fahy and others, Mattison and others, and Waziry and others, introduced in the debate above or in the sections that follow. Six further works are cited only from their abstracts, as the bibliography lists them, and nothing is said of them beyond what an abstract says.

Each text was read for four things: what it takes the target to be (ageing itself, the diseases of age, or health across the life course); what it offers as evidence that the target has moved; who, if anyone, it names as deciding what counts as that evidence, and any period it sets; and any condition it writes down under which a measure or a finding should be looked at again. Every block quote is a passage of a saved text, word for word, and its source line names the file. Every number in the prose stands in a saved text and is listed, with its file and place, in a numbers file.

Three cautions govern the reading. First, several texts report experiments in animals or trials in people. Their findings are given in their authors' words and with their authors' limits; this paper does not combine them, weigh them against each other or draw any conclusion about whether any intervention works in people. Second, statements about the future, including hopes for a field, are given as their authors' statements. Third, the texts are not a sample of the literature. They are the texts this study could quote; what they show is how the question is framed in them, not where the field as a whole stands.

1.2What counts as the target

The texts converge on the claim that ageing can be aimed at, and they reach it from three directions. The first is biology. Gems builds his argument from what he sees in the worms he studies, and offers a definition that removes the line between ageing and disease:

On this definition there is nothing left over called normal ageing that medicine must respect; there are pathologies, some more lethal than others. The definition does not make ageing one thing. Gems's essay keeps the plural, and the next step he takes is practical rather than biological: he favours a broad definition of protection against ageing, one that counts protection against single pathologies of age, because it is more achievable in the short term. The question of who should act on his argument he answers in his conclusion, and the answer is the people who set priorities:

The second direction is the experiment in long-lived animals. Mattison and others compare the two studies of caloric restriction (a reduced intake of calories without malnutrition) in rhesus monkeys, begun at the National Institute on Aging in 1987 and at the University of Wisconsin in 1989. They close by naming the target in so many words:

The sentence is a claim about primates, and the article frames it as a concept the studies address, not as a finding about people. It is also, for this paper's question, a claim about targets rather than about evidence. That ageing presents a reasonable target says nothing about what a trial aimed at it should read.

The third direction is money and the public interest. Scott and others value longer and healthier life by what people would be willing to pay for it, and find, as quoted above, that targeting ageing offers potentially larger gains than eradicating single diseases. Their reasoning has a feature that bears directly on who decides: it ends not with a recommendation about any treatment but with a condition about access.

The argument values treatments that do not yet exist, on the condition that they would work as the model assumes. The authors say plainly that their dollar values are sensitive to how the model is calibrated, while the ranking of the cases is not. Read for this paper's purpose, the economic case shifts the question of who decides away from the laboratory: if the value is as large as the model says, the decision about what counts as evidence is also a decision about public spending and access.

The government text names the target in yet another way. The PROSPR page begins not with ageing but with the gap between length of life and health:

Healthspan is the target here, and ageing enters through the causes of age-related disease that the programme aims to understand and target. Barzilai, from the clinic, joins the two words that the federal page keeps apart. He asks the field to stop treating lifespan and healthspan as rivals and proposes a measure of success:

The four positions differ in what they name, but they share a structure. Each moves the target upstream of any single disease, and each, in doing so, makes the evidence harder to gather. A disease can be diagnosed within a trial. Health-adjusted years, the age-related increase in disease vulnerability, or the pathologies that increase with age, can only be counted over a long time. That is where the texts' second question begins.

1.3What counts as evidence

The texts offer four kinds of evidence that ageing has been slowed: survival and disease in animals; health-adjusted outcomes in people; a molecular clock that estimates how old a body is; and a molecular measure of how fast it is ageing. They disagree about which can be trusted, and the disagreement is sharpest between those who want hard outcomes and those who want a measure that moves within a trial.

The measurement papers start from a practical constraint. Belsky and others put it at the head of their abstract:

A surrogate endpoint is a measurement that stands in for the outcome a trial is really about, such as blood pressure for strokes. The case for one in ageing is made in their introduction: in trials of middle-aged adults, they write, traditional endpoints such as disease diagnosis or death are too far in the future to serve as outcomes. The same constraint appears from the opposite side in Barzilai's commentary, which reports the demand for hard outcomes in mammals and then concedes the clinical reality:

What Barzilai proposes in place of death as an endpoint is not a molecular measure but a set of human outcomes that map to health-adjusted survival: delayed multimorbidity, frailty, function and immune resilience, paired with replicable lifespan effects in mammals where feasible. The measurement papers propose something faster.

The first candidate is the epigenetic clock. Horvath's multi-tissue predictor estimates the age of a tissue from methylation at 353 sites, and his conclusion makes a careful forecast about its use:

The words arguably and may carry the claim. The paper proposes a model of what the clock measures, the cumulative work of a system that maintains the stability of the epigenome, and finds that it does not measure the number of cell divisions or cellular senescence. What it leaves open is stated in its final paragraph: whether DNA methylation age is only a marker of ageing or relates to an effector of it. A marker that is not an effector can move without anything that matters moving with it; this is the gap any surrogate must close.

Belsky and others build a different kind of instrument from a clock, and say why. Clocks are built by comparing older people with younger ones, and so mix the effects of ageing with those of having been born in a different decade. Their measure is built from people born in the same year, followed over twelve years, and is meant to read a rate rather than a state:

The difference between a clock and a speedometer is not a detail. A trial lasting a year or two asks whether something changed during that year or two. This paper reads a measure of accumulated age as slow to change, and a measure of rate as designed to register a change in rate. Which instrument is chosen therefore shapes what a trial can find, and the choice is made, in these texts, by the people who build the instruments.

The two trials in people among the texts show the choice at work. Fahy and others report a small pilot trial of a protocol intended to regenerate the thymus, the gland in which immune cells mature. Four epigenetic clocks, measured against each volunteer's own baseline, showed a lower epigenetic age after a year. The authors defend the use of clocks in terms that make the evidential claim explicit:

The passage concedes and claims in the same breath. It concedes that epigenetic age is not ageing; it claims that it is the best available measure and that this justifies its use on a practical timescale. The phrase practical timescale names the pressure that drives the whole literature: the measure is chosen because it can be read soon. The same discussion states what must follow before the result can be relied on, and that sentence belongs to Section 1.5.

Waziry and others report what happened when several such measures were applied to the same randomized trial. CALERIE assigned 220 adults without obesity to a prescribed 25% reduction in calories or to an unrestricted control diet for two years; the methylation analysis was done afterwards. The measures did not agree:

DunedinPACE is a pace-of-ageing measure, of the kind Belsky and others describe; PhenoAge and GrimAge are clocks built to predict mortality. One measure moved and two did not, in the same blood from the same people. The authors note that the measures were developed using different methods and reflect different models of ageing, and they open their own list of limits with a sentence that bears on every evidential claim in this literature:

Read together, the texts show the evidence question as a choice between instruments that answer different questions. Survival in animals is hard and slow and, in primates, contested. Health-adjusted outcomes in people are what matter and take decades. Clocks estimate how much ageing has accumulated and may not register change within a trial; a pace measure registers change and has not been shown to stand in for the outcomes people care about. No text claims that the choice has been made; each says which instrument it prefers and why.

1.4Who decides and by when

On who decides, the texts say less than on any other question, and what they say is mostly implicit. The researchers who build a measure propose it, validate it in their cohorts and state what further validation it needs; the trials that use it are approved by review boards and, in one text, by the drug regulator. Only the federal text addresses the regulator as the audience of a decision about endpoints, and only it sets a period.

The PROSPR page states the programme's method and its time frame in one sentence:

The sentence is a programme's aim, and the page conditions what would follow on success. It also names the period in which an assessment is to be possible, which no other text does: three years instead of decades. The same page lists a consensus document, written by a working group the agency convened, and says whom it is for:

Here the allocation is explicit, and it is divided. Experts convened by one agency recommend; the Food and Drug Administration, the United States drug regulator, receives the recommendations. The endpoint proposed is not a molecular measure but intrinsic capacity and its domains, a concept the page names and does not define. The document is described as outlining a stepwise validation plan toward potential future use, so even the recommending body places the decision in the future and makes it conditional on steps. The page does not say what the regulator has done with the recommendations, and this paper does not either. The meeting report of Justice and others, as its abstract describes it, shares the aim of further dialogue with regulators about which indications such therapies could be registered for.

The measurement papers place the first decisions with research bodies. Belsky and others record that a national advisory council has approved one design choice for development:

That approval concerns who may be enrolled, not what may be measured, but the two are linked: enrolling younger adults is what makes a fast surrogate necessary, because disease and death in them are furthest away. The decision about the measure then falls back to the scientific process of validation, which the authors describe in steps and which no institution in the texts is named as closing.

The trial in people that used clocks as evidence shows the regulator at the gate rather than at the end. Fahy and others report that their pilot was run under an investigational new drug application from the FDA and approved by an institutional review board, and that, in accordance with FDA guidelines for exploratory studies, it was not preregistered. The regulator's role in that text is to permit the study, not to accept its endpoint.

Where no regulator is named, the period is set by the biology and by the length of the studies. The monkey studies show how long the wait can be and how the question can shift during it. Mattison and others say what the studies were for:

Studies begun in the late 1980s to ask about health became, by the time of the 2017 comparison, a dispute about survival, which neither study had taken as its primary outcome. Waziry and others name the same constraint for people in one phrase: a barrier to trials, they write, is that intervention studies run for months or years, but human aging takes decades to cause disease. Scott and others, valuing what slowing ageing would be worth, make their use of one drug as an illustration wait on a trial that had not reported:

The sentence makes a pending trial the date on which the model's illustration may need to be revised. It is the clearest case in the texts of a calculation that names the event that would change it. Newman and others, as their abstract describes it, report the meeting whose discussion informed the proposed TAME study, and the trial frameworks of Justice and others share its aim; none of the texts this paper read reports TAME's results.

The pattern is consistent. Within the research texts, the holder of the decision about evidence is the research community, acting through validation studies and through the choices of the people who design trials; the deadline is the length of the biology. The one government text names the regulator as the recipient of expert advice and sets a programme period of three years. None of the texts says that any body has decided that ageing, or a measure of it, may serve as the basis for approving a treatment.

1.5What would reopen the decision

On this question the texts are unusually explicit, and the measurement papers most of all. They write down, in advance, what would have to be shown before their own measures could be trusted as surrogates, and so what would reopen any decision taken on them. Belsky and others give the condition in its fullest form:

The sentence has three conditions, and the third is the hard one. A measure must be associated with healthy life; it must change when something is done; and the change it shows must correspond to a change in healthy life. This paper reads the first two as shown in a cohort and a short trial, and the third as requiring exactly the long follow-up that the measure was meant to avoid. The sentence cites Prentice (1989) for the condition.

Waziry and others, three years later and with the measure applied to a randomized trial, restate the condition almost word for word:

and they say what their own trial cannot show:

They add two further conditions. Changes in DunedinPACE overlapped heavily between the two arms of the trial; the authors write that the effect sizes imply close to 90% overlap of trajectories between the groups. And the trial's volunteers, they say, do not represent the general population, so the effects may not generalise beyond them. Belsky and others had written the same condition about their own samples, which were mostly of white European descent: follow-up in more diverse samples, they write, is needed to establish cross-population validity. A decision taken on these measures would, on the texts' own terms, be reopened by follow-up of trial participants to disease and death, by results in other populations, and by evidence about whether the change a measure shows tracks the change in health.

The pilot trial writes its condition as replication:

The same discussion describes the study as a one-year pilot trial involving only 9 volunteers, compared with their own baselines. The authors' general prospects for the field are given in their words in the appendix; what their text sets as the condition for relying on the result is a larger study.

The monkey studies show a third kind of reopening: two studies of one question, with different answers, sent back to their designs. Mattison and others write of the contrasting survival results:

Their comparison names the differences that could explain the split: the age at which restriction began, whether the control animals were fed freely, the composition of the diet and the source of the animals. These differences, on their reading, may have contributed to the dissimilar outcomes, and some survival curves were still incomplete. Read as a condition for revisiting, the comparison says that a result about ageing in long-lived animals is open to revision by the design choices of the study that produced it, and that those choices are made decades before the result is known.

The economic model writes its own conditions in a different register. Scott and others say that their estimates abstract from inequalities in health and income and that their dollar values depend on calibration; they make the metformin illustration wait on TAME. Waziry and others close with a sentence addressed beyond the laboratory, to the people who plan:

The sentence does not tell planners what to decide. It names the assumption from which, in the authors' view, a forecast should start, and so the assumption whose failure would make a plan worth revisiting. Set beside the large values of the economic model, it marks the distance between what slowing ageing might be worth and what a trial has so far been able to show.