PFLPhysiological Fitness Landscape
Terrain

Stability from within-person drift

Do people drift back to a level over the years? We followed the same people across visits in the United States, England and China.

What we found

The main open question

As allostatic load builds, the efficiency with which mitochondria turn fuel into usable energy falls. Is that fall just one more marker of load, or a measurable mechanism that drives the loss of stability and the move into disease? If it is the mechanism, the landscape can predict when adaptation is close to failing, not only draw a new picture of allostatic load. A slow pull back to a settling point, as found here, is what a well with spare energy looks like; a well losing that energy should pull back ever more slowly before it gives way. See the open problem.

See it for yourself

Pick a cohort, a measure, sex and an age band. Each chart follows the same people from one visit to the next.

Which way people drift

Each dot is a group of people who started at a similar level; the line is their average change per year once measurement noise is removed (dashed: before). Arrows show which way each group moves. Where the line crosses zero going down, people settle; a filled circle there passes every check.

The same drift as a landscape

The same drift drawn as a landscape: people roll downhill, so valleys are where they settle and ridges are where they would be pushed apart. It shows how people move, not mortality risk.

The landscape at each age

One cross-section per age band on a shared scale, to see whether a valley moves, deepens or flattens with age. The dotted line joins settling points that pass the checks or are candidates.

How long it takes to settle

The relaxation time is one over the restoring rate at a settling crossing: the time for a displacement to shrink by a factor of e. Only age bands with a settling crossing are shown; gaps are not filled in. The release has no interval for this rate.

How sure the crossings are

Each mark is a crossing in one age band, faded by its support: the share of resamples of the drift that reproduce it. For ELSA and CHARLS the lower row repeats the analysis without the visit pairs flagged for a laboratory or protocol change, so a point that survives both is a finding, one that appears only with them was dropped, and one that appears only without them is a candidate.

How stiff is each well?

A settling point works like a spring: the further people are from it, the harder they are pulled back. The half-life is the time for half of a displacement to fade. There are two very different springs: a slow one, seen between visits years apart, and a fast one, seen in the minutes after a stress test.

The slow pull, from the ELSA and CHARLS visits; each dot is the median over sex and age groups. The pull toward each person's own level is at least this strong: lasting differences between people make it look weaker.

The fast pull: recovery after a laboratory stress test or exercise, in open wearable studies. A different mechanism, on a different clock.

Pushed, then let go: what treatment says about the pull

Drift at rest is one way to measure the pull toward a settling point. A treatment is another: a drug pushes a marker away from where it sits, and stopping the drug shows how fast it goes back. If both measure the same spring, a change made by treatment should fade as slowly after stopping as a difference fades at rest. A first test, on two markers that trials have measured for decades: systolic blood pressure and LDL cholesterol.

The push. Systolic blood pressure: the fall from one drug at standard dose, by how far a person starts above or below the trials' average (placebo-controlled trials pooled). LDL cholesterol: the fall from a standard-dose statin and from high-dose atorvastatin. The trials report each fall from one average starting level (the filled dots); the lines assume the same percentage at every start.

Two clocks, on a log scale. Days until a treatment-made change is largely undone after stopping, against years for half of a difference to fade at rest in ELSA. The stopping times are upper bounds from small studies; the rest half-life is a median over sex and age groups and counts lasting differences between people as part of the pull.

A first deliverable, on two markers only. Every number comes from a published abstract and is quoted in the data release next to its source; the quotes are checked by machine against the abstract, not yet read against each full paper. The stopping studies are small, and one is of a single drug class. This is not a claim that treatment is futile: it lowers the marker while it is taken. It says the change in the marker lasts about as long as the treatment does; whether any lowering of risk outlasts it is a separate question these data do not answer.

Sources, as quoted in the data release:

What shape is the valley?

Does the pull weaken toward an edge?

If a valley flattens toward one edge, the pull back toward the centre fades there, which is what a coming tipping point would look like. We compared the pull in the outer tenths of each measure with the pull in its middle. This comparison does not depend on how noisy the measure is. In simulated cohorts it found planted feedback every time and raised few false alarms.

HbA1c softens at its upper edge, roughly where prediabetes begins, but the signal comes from England and is not seen in China or the United States, and within England it rests on two age-and-sex groups: drop either and the pooled call is lost. Treatment started between visits would pull the top down, so it cannot explain this. The HOMA-IR result rests on one group in one cohort. No measure runs away, and blood pressure, lipids, grip and BMI show no softening. This is a lead to test, not a finding. Cases where the edge looks steeper are not interpreted, because height, which has no dynamics, shows them too.

Where do placebo arms drift?

Each placebo or sham arm in a registered trial is a group measured twice with no active treatment. Across many trials that start at different levels, these arms show the level they drift toward and how fast.

These are trial populations, often selected for disease, and most keep their usual care, so this is not a healthy set point or truly untreated drift. Most trials last months, which makes the level, the pull and any placebo response hard to separate. Only blood pressure gave a level that held steady across subsets of trials. The other measures gave no usable or believable level and are not shown.

Do treatments work harder further from the middle?

For each drug class we compared each active arm with its own trial's placebo arm. We then placed the arm's starting value in the distribution for people of its age and sex. A flat line suggests a treatment moves the valley. A line that falls suggests it deepens the valley, working harder the further out people start.

These lines compare trials, not people, so a slope may reflect dose or who each trial enrolled. An arm's average is being placed against the spread of individuals. Effects are measured against general-population spread, so they look large in very selected groups. Most active arms were one-off compounds and could not be grouped. These are not treatment recommendations.

How a result is checked

Measurement noise alone makes people measured high seem to fall and people measured low seem to rise, which looks like a pull toward the middle. Each cohort's drift is corrected for that before anything is claimed. A crossing is reported only if it survives the correction that works hardest against it, reappears in most resamples of the data, and the lowest and highest starting groups clearly move in the directions it implies.

How much of one reading lasts

With three or more visits per person, the part of a reading that reflects a lasting difference between people can be measured; the rest is day-to-day variation. The correction uses it.

Laboratory changes

If a laboratory changes its test between visits, readings can shrink toward the middle and look like a pull. Visit pairs where a blood marker's spread changed more than physical measures ever do, or where ELSA records a change of assay, are flagged and the analysis is rerun without them. A point found both ways is a finding; one found only with the flagged pairs was dropped; one found only without them is a candidate.

United States: Health and Retirement Study

Two visits per person, about four years apart. No crossing passes every check.

  • No mountain tops. In no measure, sex or age band are people pushed away from a value, so these everyday measures show no tipping point over four-year steps.
  • Blood pressure and pulse drift back toward the group average. The pull is larger than the noise between readings at one visit, but with two visits per person it cannot be told apart from ordinary day-to-day and month-to-month variation, so it is not claimed as a stable point.
  • Grip and peak flow settle below their averages, which may be a floor or the weakest leaving the study; this does not pass the noise checks either.
  • Body size is measured once per visit, and the two timed walks per visit were not extracted for walking speed, so their noise cannot be removed and nothing is claimed for them.

England: English Longitudinal Study of Ageing

The English Longitudinal Study of Ageing measured the same people at nurse visits about four years apart. With three or more visits per person, ordinary day-to-day and month-to-month variation can be told apart from lasting differences between people (the three-wave method of Heise), which two visits cannot do. The correction for regression to the mean then uses that long-term reliability instead of a bound.

  • No mountain tops in ELSA either: no unstable point passes the checks.
  • Three stable points pass every check, including the laboratory check below: diastolic blood pressure in women in their sixties, pulse in women in their seventies and haemoglobin in women in their fifties. Diastolic pressure sits at the group's own average; pulse and haemoglobin sit away from it.
  • Two points for insulin-like growth factor and one for vitamin D were dropped. They came from changes of laboratory method, not from the body.

China: CHARLS

The China Health and Retirement Longitudinal Study measured the same people at three visits about two years apart, on the same protocol each time. Three visits give the same long-term reliability as in ELSA, so the correction for regression to the mean uses it. Blood pressure and pulse are the mean of the second and third readings, and grip is the best trial.

No crossing in CHARLS passes every check.

One shared pull, or personal set points?

Does moving away from a held level predict death?

Exploratory. People whose value moved away from the held level between two visits, against people who moved toward it, adjusted for age, sex, the starting distance and the mortality curve at the later value. The same model measured against the group's own average is the control: if both predict death equally, the result says nothing about attractors. Deaths are family-reported in exit interviews; there is no registry linkage.