Method · v1.4
Method for the coach effect estimator v1.4
How much of a swimmer's improvement is associated with a named programme, once age, event and prior trajectory are accounted for — and the two things this estimate can't tell apart from a coach's own effect.
What it does
For a named programme, the estimator compares the improvement of the swimmers training there against the improvement the development curve already expects for those swimmers’ ages and events. The remainder is the estimated programme effect, in seconds per season, with a 95% interval. A negative figure is faster than the cohort expectation.
Fitted on 41 214 swimmer-seasons across 3 844 programmes, 31 federations, 2019–2026.
What it can’t separate
Two things the estimator can’t tell apart from a genuine coaching effect:
- Recruitment. A programme that selects swimmers who were already improving quickly will show a programme effect whether or not its coaching adds anything, because the estimator holds prior improvement rate but cannot hold selection on trajectories it never observed.
- A coach change inside the fitting window. Where a programme changed head coach during the period being estimated, the published figure is for the programme, not the individual coach, and the method note on the programme’s own page states the date of any change inside the window.
Below the publication threshold
Below 40 swimmer-seasons in the window, the estimate is compatible with almost any value and Swimetrica publishes none — a stated absence in its place, naming the count that fell short.
Where it is weak
The estimator’s interval is calibrated the same way the adjustment’s is (method: how the intervals are calibrated). It is weakest for programmes near the 40-swimmer-season threshold, where a handful of transfers in or out of the programme can move the estimate by more than its own interval width.
Version history
- v1.4, 2026-08-20 — publication threshold raised from 24 to 40 swimmer-seasons after a backtest showed intervals below 24 covered at 71% against a nominal 95%. 118 programmes that had published estimates under v1.3 fell below the new threshold and now carry a stated absence instead.
- v1.3, 2026-04-01 — first version to hold prior improvement rate as a covariate rather than only age and event.