Models: what the map assumes and how it could fail

Exposure with memory. Each stressor fades at its own rate: an acute infection over days, a bereavement over days, financial strain and caregiving over a few years.

Demand. The sum of each stressor's remembered exposure, weighted by its potency, plus pairwise terms. Positive pairs are synergies (financial strain with discrimination); negative ones are buffers (social support with caregiving). Earlier exposure can also sensitise the response to a later one.

Load grows slowly with any demand, much faster when demand exceeds capacity (the damage term is quadratic in the excess), and is repaired over time.

Capacity relaxes toward a baseline that falls with age, is worn down by load, and is built up by moderate, recoverable demand (hormesis). Erosion is proportional to remaining capacity, so capacity is never driven to zero and always recovers once the stressor ends.

Scar. A fraction of overload damage becomes a slow offset on baseline capacity (S) that heals over about 20 years, larger for overload in childhood. Everyday demand leaves no scar.

dL/dt = αD + β[D − C]₊² − ρL
dC/dt = κ(C₀ − S − C) − δL·C/C₀ + η h(D)
dS/dt = σβ[D − C]₊² − S/τ_S

Capacity sets how deep a person's basin is on the landscape. Chronic stress makes it shallower, and recovery takes years after heavy strain. That is consistent with studies suggesting that adult economic strain is largely recoverable once it ends, and that stress-driven rises in biological age can reverse. What lasts is a scar, strongest after early-life adversity.

A permanent collapse ("tipping") is offered as a hypothesis in the Model menu. It needs erosion that does not slow as capacity runs out, plus everyday demand inside the range where two states exist. Evidence for true tipping points in people is weak, so it is not the default. Either way, the model predicts a warning sign: recovery from a small challenge slows as capacity is worn down.

Load and capacity are not measured directly. They are read through biomarkers on several timescales: cortisol and heart-rate variability (fast), CRP and HbA1c (medium), epigenetic pace of ageing and grip strength (slow). In simulated people, the repair and erosion rates are recovered almost exactly from a large overload, but are biased by about 30% from short episodes. Challenge tests and natural experiments that produce real overload are what pin them down.

This is open problem 4. If you work on stress physiology, life-course epidemiology or state-space models, get in touch.

Events shorter than half a year stay out of the chronic pile-up term. The first-marathon training year sits on that boundary. The same stressor can also build capacity when the dose is moderate and there is time to recover: growth or scarring.

• Control: whether a person can act.
• Dose: moderate challenge versus demand above capacity.
• Recovery: growth happens between episodes.
• Predictability: known stress sensitises less.
• Social context: support buffers and threat amplifies.
• Appraisal, timing and meaning-making alter persistence and response.

Each run is compared with the same model without the stressor: growth raises the capacity set point, resilience returns it, scarring lowers it, and collapse leaves capacity in a depleted state.

A challenge can leave a capacity marker higher once recovery has had time to work. Each line is a published result: higher than the same adults' baseline, or higher than a control who did not get that dose.

• Twelve trained cyclists. Ten days of exercise near 50% VO2max in 40°C heat, versus the same exercise at 13°C. Cool VO2max rose 5% and hot VO2max rose 8%. Time-trial work rose 6% in the cool and 8% in the heat. Lactate-threshold power rose 5% in both. Plasma volume rose 6.5%. Maximal cardiac output rose 9.1% in the cool and 4.5% in the heat. The cool-exercise control was unchanged. Lorenzo et al. 2010, PMID 20724560.
• Fourteen endurance-trained men. Five days of isothermic heat acclimation versus thermoneutral exercise. Resting heart rate fell 8 bpm and plasma volume rose 7.3%. Sweat loss rose by about 500 mL. Kidney injury markers were unchanged. Snape et al. 2026, PMID 42579663.
• Sixteen triathletes. Two 18-day live-high train-low camps. VO2max rose 4.9% in normobaric hypoxia and 3.2% in hypobaric hypoxia. Saugy et al. 2016, PMID 27148076.
• Meta-analysis, 62 studies, 1,332 people. Live-high train-low raised VO2max more than control in athletes (mean difference 2.17, 95% CI 0.77 to 3.56) and not clearly in non-athletes (3.1, −0.23 to 6.43). Kohlbrenner et al. 2026, PMID 42742930.
• BCG vaccination in healthy volunteers. Monocyte TNF and IL-1β responses to unrelated pathogens were about twofold higher, and still present at 3 months. Kleinnijenhuis et al. 2012, PMID 22988082.
• Pilot trial. A 90-minute exercise bout immediately after SARS-CoV-2 immunisation lengthened antibody durability in people without pre-existing immunity. Kothadiya et al. 2026, PMID 42502427.
• HERITAGE Family Study, 371 men and 465 women. Twenty weeks of endurance training raised mechanical efficiency 4–5%, with a similar rise in women and men. Tremblay, Bouchard et al. 2025, PMID 41162184.
• Meta-analysis of stopping exercise, 86 trials, 5,670 people. Lower waist-hip ratio, mean arterial pressure and fasting glucose were still present after withdrawal. Zhang et al. 2026, PMID 42558561.
• Meta-analysis, 12 studies, 859 team athletes. Blood-flow restriction added to resistance training increased hypertrophy (SMD 0.32) and strength (SMD 0.42). Jumping and sprinting did not change. Huang et al. 2026, PMID 42549104.

The simulator above does not reproduce these percentages. Acute events on their own clocks are on the stress response page.

Repeated laboratory stress tests, training studies and cohorts with dated events can test whether a moderate, controllable challenge speeds recovery from the next one. See open problem 6.

Each row takes one biomarker, an average US adult at 50, and the best lifestyle change and the best drug in the intervention catalogue for that marker. The catalogued effect is applied to the population mean and read off the risk of death curve: that gives the risk-of-death ratio for the move and the undiscounted life-years it would add if the whole curve were causal: the benefit of acting on the result.

It is not. Age- and sex-adjusted associations overstate what changing a marker achieves. So each marker carries a causal fraction κ by evidence status: 0.8 where Mendelian randomisation supports it, 0.5 where trials move the marker, 0.25 for strong observational evidence, 0.1 otherwise. These are judgements, following the priors in our value-of-information programme, and the "if fully causal" number is shown beside each adjusted one.

The break-even column asks the reverse question: at $100,000 per life-year, how much of the curve would have to be causal for the fee to be covered by the benefit of acting on the result? For every priced test it is a tiny fraction (the headline figure above). Whether the test itself adds that value is a separate question — see What the test itself is worth.

A biomarker can point to a large benefit and still be worth little as a test. If the favourable direction is the same for everyone — lower blood pressure, more fitness, less visceral fat — that step can be taken without measuring, and the result changes nothing. The value of a test is the value of the decisions it changes.

Our value-of-information analysis of 103 mortality-associated biomarkers across six US age-by-sex groups separates the two. Modelling a sustained one-standard-deviation shift, the typical marker carried a benefit of acting of more than a year of life (median 1.1–1.5 years by age band), but its expected value of perfect information — what knowing the result adds — was effectively zero (median ≤ 0.000002 years). Almost all of the information value (over 99.9%) sat in the 17 markers with a documented U-shaped association, where higher risk at both extremes means the result decides which way to move.

• Benefit of acting

  How much would moving this marker help?

  Lipid accumulation product (monotonic), men 40–59: 9.74 model years per person.
  Mean corpuscular volume (U-shaped), women 20–39: 3.32 model years per person.

• Value of the test

  Would knowing my value change what I do?

  Lipid accumulation product (monotonic), men 40–59: ≈ 0.00001 model years per person.
  Mean corpuscular volume (U-shaped), women 20–39: 3.32 model years per person.

Price rarely decides. At Medicare fees of a few dollars, a test whose information value is material clears its fee at any conventional value of a life-year ($50,000–150,000), and one whose value is near zero clears it at none. What matters more is whether the association is causal, and whether people would act without the test: when they would not, the value of testing approaches the full benefit of acting.

The dashboard's Value of Information tab reports the two quantities separately from the same calculation.

These are model results that treat observational associations as if they were causal and hold risk-of-death ratios constant over remaining life; they are not estimates of the effect of screening anyone. Details: value-of-information manuscript in preparation.

Set the test fee against the thing the result leads to. A year of the intervention ranges over four orders of magnitude, and the cheapest options sit at the top of the table.

1. $0Diet, exercise, sleep, alcohol and smoking changesno direct medical cost; time and food are not valued here
2. $13Generic atorvastatin 20 mg, dailyNADAC ≈ $0.03–0.04 per tablet, June 2026
3. $20Generic allopurinol 300 mg, dailyNADAC ≈ $0.056 per tablet, July 2026
4. $30Generic lisinopril + amlodipine, dailyNADAC ≈ $0.015–0.05 per tablet, June 2026
5. $40Generic metformin ER 500 mg, twice dailyNADAC ≈ $0.055 per tablet, June 2026
6. $782Medicare Diabetes Prevention Program, one yearmaximum CY2026 payment: 22 sessions at $27 plus weight-loss milestones
7. $2,364Empagliflozin (Jardiance)2026 Medicare negotiated price, $197 per 30 days; $573 list
8. $4,188Semaglutide (Wegovy / Ozempic), cashmanufacturer cash price $349 per month, from November 2025
9. $5,850Evolocumab (Repatha)list price after the 2019 cut; 2026 Part D per-syringe pricing implies about $7,300

The table above covers the 22 tests priced for this page, an earlier and smaller set than the 54 tests with a Medicare fee in the value-of-information analysis.

Three patterns fall out. First, where a cheap generic moves a marker in trials, the arithmetic is lopsided: a first-line blood-pressure combination costs about $30 a year for an expected third of a κ-adjusted year. Second, the newer agents cost a hundred times more: PCSK9 inhibitors and GLP-1 drugs give 0.7 to 0.9 κ-adjusted years on LDL, apoB and glucose at $4,000 to $6,000 a year, so the price of the treatment, not the test, decides the value. Third, the best lifestyle change often matches the drug: DASH with sodium reduction equals the drug combination on blood pressure, and higher fibre intake recovers about two-thirds of the GLP-1 gain on glucose. Some best moves were measured in patient groups, such as dialysis or nephrotic syndrome, and are named as such in the table; applied to an average adult they overstate the gain.

Full cost-effectiveness models of screening for single markers reach the same shape of answer. hs-CRP screening to target statins was estimated at $42,600 per QALY for 65-year-old men and $207,300 for 35-year-old women: same test, different prior risk. An HbA1c cut-off of 5.7% for diabetes screening came in under $50,000 per QALY in a CDC model, and the broadened US screening recommendation at about $29,000 per QALY. Combined cholesterol screening and treatment has been put at roughly $34,000 per QALY. In each case the fee of the test is a rounding error; the result turns on who is tested and what treatment follows.

Preventive drugs are prescribed on population averages, so most people treated gain little individually. In primary prevention, statins reduced absolute risk over the trials' follow-up by about 0.8 percentage points for all-cause death, 1.3 for myocardial infarction and 0.4 for stroke [Byrne 2022] — roughly 125, 77 and 250 people treated per event avoided. With an estimated 92 million US adults using statins in 2018–19 [Matyori 2023], even small improvements in targeting reach a very large population. Across the US system, overtreatment and low-value care are estimated at $76–101 billion a year [Shrank 2019].

The PFL hypothesis is that a person's position and trajectory on the landscape identify who is likely to benefit, who is likely to be harmed, and who can safely skip treatment — raising net benefit per person treated. This has not yet been tested. Our value-of-information results set the bar: for a single marker whose favourable direction is the same for everyone, deciding person by person adds almost nothing. PFL's added value therefore has to come from where the right move differs between people — U-shaped markers, interactions between markers, and trajectories over time — and that is what validation must show.

This is a screen, not a cost-effectiveness analysis: there is no discounting, quality adjustment, adherence, side effects or downstream care, and trial effects are applied to an average adult. A full decision model with value-of-information results is under way; get in touch if you hold cost or outcome data.