How the number is calculated

Most apps in this category ask you to trust a prediction. This page exists so you do not have to. Every number Expectancy shows comes from something written down here, and where the model makes a judgement call rather than following a published figure, it is said plainly rather than buried.

The baseline

Mortality comes from WHO Global Health Estimates 2019 abridged life tables, covering 185 countries plus 11 regional and income-group aggregates, by sex.

2019 rather than 2021, which is the most recent year WHO publishes, because a 2021 period table embeds pandemic excess mortality. A period life table is already a fiction: it asks how long someone would live if today’s mortality rates held forever. Building that fiction out of the worst year in a century would shorten every estimate for reasons that have nothing to do with the person reading it.

WHO publishes 19 age bands ending at “85+”. The engine needs single years of age up to 110, because risk factors multiply a hazard age by age. That conversion happens in three steps:

Constant force within each band, which reproduces the band’s average hazard exactly.
Monotone cubic smoothing through band midpoints in log space, rescaled inside each band so the integral still reproduces WHO’s published values exactly. Smooth and lossless, rather than one or the other.
A Kannisto logistic tail above 85, because old-age mortality decelerates rather than continuing to grow exponentially. The shape comes from each country’s own 70-to-84 slope, and the level is solved so the curve reproduces WHO’s published life expectancy at 85 exactly.

The build refuses to ship a data file unless that last figure matches WHO to within a millionth of a year for all 392 country-sex pairs, no country is dropped, and the round-trip through the app’s storage format moves life expectancy by less than 0.01 years.

Why your factors are not simply multiplied in

Two decisions carry most of the accuracy, and both are places where a naive implementation gets a plausible-looking wrong answer.

Population re-centring

A hazard ratio from a cohort study is relative to that cohort. A national life table already contains that country’s smokers, its obesity, its blood pressure. Multiplying the table by a smoker’s raw hazard ratio double counts: it adds a smoker’s excess on top of a baseline that already assumed some smoking.

So every factor declares the prevalence-weighted average of a reference population, and only your difference from it is applied.

The invariant this buys, and which the test suite holds exactly: a person whose every factor sits at the population mean gets back their country’s published life expectancy.

Saturating combination

Risk factors overlap. Someone who smokes, drinks heavily, never moves and sleeps badly does not lose the arithmetic sum of four independent penalties, because those four things share causal pathways and share the same person. The model combines them with a saturating rule calibrated against published multi-factor cohort data rather than multiplying them naively, which is why the simulator’s “everything at once” figure is smaller than adding up the list beneath it.

What this model is not

Not a diagnosis, and not medical advice. It is an actuarial estimate built from population statistics.
Not a prediction about you. The distribution is wide. The app shows the 10th-to-90th-percentile range beside the headline figure for exactly this reason, and the headline is a median rather than a mean, because a mean is not a date anyone can picture.
Not causal. Almost every input is observational. Where an association is probably not causal, such as long sleep or flossing, the app says so rather than advising you to change it.
Not equally confident everywhere. Every factor carries an evidence grade, and those grades are shown to you rather than averaged away.

The largest known limitation

Risk factors are re-centred against one reference distribution applied to all 185 countries. Smoking prevalence, average BMI and blood pressure genuinely differ between countries, so for a country far from that reference the re-centring is approximate. This is the single biggest known source of error in the model, and it is written here rather than left for someone to discover.

Reading the sources yourself

Every factor names its study inside the app, on that factor’s own screen, along with the grade for how strong the evidence is. Citations, author names and journal titles are deliberately left in English in every language the app ships in, because a paper is unfindable under a translated title.