How these numbers are made

The question is not how long will I stay independent. That question cannot be answered, and §2.6 shows the arithmetic of why. The question is the inverse:

Pick an age you want to still be looking after yourself at. Find the level of some capacity below which people stop managing alone. Work back through the rate that capacity declines with age. What comes out is a number you can measure today.

There is one such number per row, and no total. Nine measurements, seven of which can be scored against a published reference, and no composite score — no published method exists for combining these tests, and no weights in the literature would justify one.

Two measures on this panel have a real sampled distribution behind them, and only two: aerobic capacity and gait speed. For those, the published source gives five percentile anchors — the 5th, 25th, 50th, 75th, 95th — at each of several age bands. Everything the engine does with a percentile is interpolation between those five numbers.

An anchor age is the integer midpoint of its published band. A study reporting "40–49" becomes an anchor at 45, because that is what the band's mean is an estimate of.

Between two published bands, each of the five anchors moves linearly:

anchor(age) = y₀ + (age − a₀) · (y₁ − y₀) / (a₁ − a₀)

Below the youngest published band the values are held flat rather than extended downwards. Nothing in these tables supports a claim about a 20-year-old.

This is the part that is genuinely a modelling choice, and it is the part to be most suspicious of. The gait table stops at 85 and the aerobic table at 85; a target age of 85 or 90 is past the end of the published data.

An earlier version of this engine extended each percentile along its own slope, and that was wrong. Fitted separately, the curves cross at age 94 for men — the 95th percentile falls below the 75th. The visible symptom was a table in which being independent to 95 required less capacity than being independent to 90.

What replaced it: one common proportional decay, taken from the median's last published segment and applied to all five anchors together, so they scale as a group and cannot cross.

rate  = (median at top band / median at previous band) ^ (1 / years between)
anchor(age > top) = anchor(top) · rate ^ (age − top)

The rate is derived from each table, never hard-coded. On the aerobic table it works out at 1.51% a year for men and 1.24% a year for women. Those two figures were computed by this page calling the same function the engine calls, so they cannot drift away from the model.

No author validated this extension and none endorsed it. Real decline accelerates after 80, so a projection to 85 or 90 is, if anything, optimistic. Wherever a number depends on the extension, the result page says so on the row.

Both directions are linear interpolation between the bracketing anchors. Outside the outermost anchors the engine continues the outermost segment but never reports beyond the 0.5th or 99.5th percentile, because nothing outside the published 5th–95th was observed. Claiming a 99.9th percentile from a table whose highest published anchor is the 95th is precision the data does not contain.

Two lookups, and no assumption about your personal trajectory:

p*     = the percentile at which the threshold sits at your TARGET age
number = the value at p* at your age TODAY

In words: find the percentile track that lands exactly on the independence threshold at the age you picked, then read that track's value at your age now. It assumes you hold your position in the distribution as you age — which is an assumption, and §5 lists it first.

The obvious product is the other one: take your capacity, project it forward, report the age at which it crosses the threshold. That version was built and then thrown away, because its error bars swallow it.

• Measured on a good watch rather than in a lab, the crossing age moves about eight years.
• Estimated rather than measured, it moves seventeen.
• Hold the measurement fixed and change only the assumed rate of decline, and it moves more than twenty.

The inverse has neither problem. No decline assumption enters it, and no measurement error propagates through a projection, because the target is computed from published reference data alone — your own measurement is only compared against it. The number is well determined. The age is not.

Percentiles are rounded with ties going down, not to nearest. A man of 47 walking at 1.50 m/s lands at exactly 62.5, because 1.50 is the midpoint of the 50th (1.43) and the 75th (1.57) at that age. He is reported at the 62nd, not the 63rd.

It is the right rule on its own terms. On an exact tie, report the lower rank: overstating where somebody sits in a distribution is the one error this product cannot afford.

Push-ups. The published figures are fitness-category boundaries — excellent, above average, average, below average — not a sampled distribution. A percentile drawn through four boundaries is a curve somebody chose, presented as a measurement. The row shows the published figures and where you sit among them, and stops there. Its table also stops at 65, so values clamp at the oldest published band rather than being extended.

Everything else with a threshold. One-leg stance, sitting-rising, five-times sit-to-stand and heart-rate recovery are scored pass or fail against a published bar, not ranked.

The female gait percentiles are partly derived, and the row says so. The published table gives women's 50th and 95th but not the 5th, 25th and 75th. Those three are reconstructed assuming normality, by a method checked by reproducing the men's published percentiles — where the real answers exist — to within 0.010 m/s. That is a defensible method and it is still a derivation, so it is labelled on the row rather than printed to look identical to the men's column.

The bar each row is measured against. Every one of these is a soft zone, not a cliff. Nobody becomes dependent at a specific decimal.

On 18 September 2026 the thirty references that had never been looked up were checked against their records. Three were wrong. The instructive one was the push-up citation, which read Med Sci Sports Exerc 2000;32(9):1564–1569.

No such paper. The real one is Payne N, Gledhill N, Katzmarzyk PT, Jamnik VK and Keir PJ, Canadian musculoskeletal fitness norms, in Can J Appl Physiol 2000;25(6):430–442 — whose abstract confirms the 571 participants, 312 women and 259 men aged 15 to 69, that this site quotes.

The finding was right and the locator was invented. That is worth stating plainly, because it is the error that survives every read-through and dies on the first lookup. It was live on the result page. If you had decided to check this site's work, it is where you would have started.

Also corrected: a paper credited to the wrong first author, and a citation using a paper's online-publication year where the rest of the page uses the print year.

2 references could not be resolved at all and are marked in §6. They are not cited by any number on this site and should not be relied on by anyone else until someone reads the articles.

1. It assumes you hold your percentile as you age. Real people drift, in both directions.
2. It uses a population cross-section to stand in for one person's decline. Cohort effects and survivor bias both bend that curve, and the oldest bands are flattered — the people who lived to 85 to be measured are not a random sample.
3. Past the oldest published band the curve is modelled, not measured (§2.3), and real decline accelerates after 80, so the projection is optimistic there.
4. Every reference table comes from a population that is not you. The aerobic percentiles are from clinical exercise laboratories. The push-up figures are Canadian adults measured in 2000.
5. Self-measurement is noisy. Hand timing and self-scoring will not match a trained assessor. Treat your own trend over time as more meaningful than any single session.
6. Nearly all of this evidence is observational. These tests predict outcomes. They do not establish that changing your score changes your outcome.

Read the output as a direction and a magnitude. Not a date.

43 references. The DOI is the citation of record — a publisher URL rots and a DOI does not. Entries marked unresolved could not be matched to a record and are shown rather than quietly dropped.

What this site is not

This site ran no study. It collected no data. It validated nothing. Everything above is published work by other people, read and arranged so that one number can be computed from it. The evidence is theirs. Only the arithmetic is mine.

One part is mine and deserves more suspicion than the rest. The percentile curves are extended past the oldest age band any cited study published. No author validated that extension and none endorsed it. It is a modelling choice, made because a target age has to reach years the reference data does not cover.

The people who actually did the work

Behind every threshold on this page is a cohort: people who agreed to be measured, repeatedly, for decades, and then agreed to be followed until they died. The authors got papers and careers. The participants got nothing.

• Dunedin has followed about a thousand people born in New Zealand in 1972 and 1973 for more than fifty years. They keep coming back.
• PURE enrolled roughly 140,000 adults across seventeen countries.
• CLINIMEX is two Rio cardiologists who tested patients for thirty years and counted who was still alive.
• FRIEND is the only reason an aerobic percentile exists at all.
• EPESE produced the physical performance battery most of this field still uses.

Cohort attributions here are stated because the associations are well established. "Well established" is not "verified against the methods section", and that check is still outstanding.