The test says running’s oxygen cost fell by two per cent. This may be an interesting, useful difference, but the number alone does not reveal its uncertainty. A single short measurement differs from repeated, well-controlled trials consistently producing similar results. Here, ‘measurement noise’ does not mean a loud laboratory fan: we are discussing variability in the numbers obtained.
The question cannot be resolved through a two- or three-per-cent dividing line that applies to everyone. Reliability belongs to the method, people studied and circumstances. Through illustrative examples of our own, we separate changes in the instrument and runner, uncertainty in a group mean and individual result, and statistically detectable versus practically useful differences. A small result need therefore be neither immediately believed nor automatically dismissed.
First state what the two per cent is a percentage of
If an invented oxygen cost falls from two hundred to 196 millilitres per kilogram per kilometre, the change is minus two per cent relative to the starting two hundred. The same four-unit difference gives a different percentage against another baseline. Name the direction alongside the figure: lower cost at a given speed may be a favourable energetic change, but ‘two per cent better’ does not identify which indicator decreased.
Oxygen cost and energy cost are not interchangeable in every situation either. A rate per unit of time differs again from a per-kilometre figure. Interpret a percentage result only alongside definitions of the indicator and task. The same speed, a different time point, another surface or workload based on another reference can represent different comparisons. Only then is uncertainty worth discussing.
Nor can two per cent simply become a race-time change at the same proportion. An economy indicator is the result of a specified laboratory task, while a race includes further factors. If a report immediately promises a guaranteed marathon-minute gain alongside a cost difference, it makes two separate claims. Each would need its own evidence. A small improvement’s value need not be magnified through a guaranteed time result.
Variability can have several sources
In an imagined measurement, instrument setup, sampling and averaging may all influence the result. The runner’s condition and task performed are not mathematical constants either. Comparing two separate days introduces between-day differences too. Test–retest research therefore informs repeatability of the whole procedure, rather than only a single sensor’s technical accuracy stated on its box.
In Saunders and colleagues’ study, repeated treadmill tests in eleven elite male runners had a typical error for oxygen uptake of 2.4% in pooled data across the speeds examined. Variability in other physiological indicators differed. This reliability result belongs to that sample and protocol, rather than a certainty threshold applying to every laboratory and runner.
In Shaw and colleagues’ work separately analysing oxygen and energy costs, typical error for both indicators exceeded the chosen smallest worthwhile change. This highlights that a result may be of an interesting practical magnitude while a given repeated test struggles to distinguish it from variability. ‘The method is insufficiently sensitive’ is not the same as ‘no real change can exist’.
Enlarge imageTypical error is not a hard wall around reality
A repeatability indicator summarises a distribution or variability. It does not mean every measurement errs by exactly that amount, or that every smaller difference is false and every larger one certainly real. A decision also involves the full measurement model and the estimate’s uncertainty. Use percentage typical error as a question: how well can this protocol resolve the particular difference?
In a simple example of our own, the first condition’s unknown true value is one hundred units and the second’s is 98. The first measurement has a plus-one deviation and the second minus one, so we see 101 and 97: the difference looks larger. In another trial, the first receives minus one and the second plus one: both become 99. The same assumed true two-unit difference produced two observed outcomes.
We chose the example’s deviations; they were not established from real data. The point is that variability from two measurements enters the difference. If part of the deviation affects both conditions alike, it may partly cancel in a paired difference; if directions differ, it may increase uncertainty. Repeated paired measurements for the particular design therefore say more than general instrument accuracy alone.
Reliability during long running can also be studied separately
Repeatability of a short test while fresh does not automatically describe every longer task. If the research question concerns how economy changes towards the end of training, repeatability of that long task becomes important. Repeating the same procedure on several occasions can show how similarly indicators develop at particular time points. Seek reliability suited to the measurement question.
In Zanini, Folland and Blagrove’s 2025 research, fourteen well-trained male marathon runners completed two ninety-minute treadmill runs. Economy indicators were evaluated every fifteen minutes; typical error for oxygen and energy costs ranged from 0.7–1.4%. This differs from earlier short-test figures, but belongs to another protocol and sample. It does not establish that every longer measurement is necessarily more accurate.
Such a result is interesting for more than the number’s magnitude. It shows why reliability should be documented alongside changes over time. If a new intervention supposedly helps mainly at the eightieth minute, uncertainty at a nearby measurement point matters. An earlier general percentage from another task is insufficient to establish the new claim. Checking the relevant protocol strengthens the inference.
A group mean’s confidence interval is not a personal prediction band
A group’s mean difference can be estimated relatively precisely while individual results remain dispersed. A confidence interval expresses uncertainty in the chosen estimation procedure for the mean effect. It does not claim that ninety-five per cent of participants or new purchasers will have results in that same narrow range. Individual responses and estimation of a mean are different levels.
In a statistical illustration of our own, suppose the paired changes of one hundred independent people average minus two percentage points, with a standard deviation of four. The mean’s standard error is then four divided by the square root of one hundred, or 0.4. Under appropriate assumptions, a rough band formed from approximately twice the standard error would extend plus or minus 0.8 percentage points around the mean. This is a calculation demonstration, rather than a shoe’s actual research result.
Meanwhile, the participants’ standard deviation of four has not become 0.4. The latter is a component of uncertainty about the mean, rather than a substitute for differences between individuals. A narrow mean interval around minus two therefore cannot promise nearly minus two to everyone. Moreover, measurement variability may enter observed individual differences, so their spread is not purely biological ‘responsiveness’ either.
More measurement points do not always mean more independent participants
Calculating sixty one-second points from one runner’s breathing data does not mean studying sixty different people. The data come from the same person and task and may be related. More points may help averaging or describing a curve, but sample-size and uncertainty analysis must retain this structure. Many numbers alone do not replace an appropriate design.
The multi-visit shoe study by Bolliger and colleagues involved specified speeds, repeated conditions and different outcomes. When reading such a protocol, identify whether a difference is within repetitions or between people. A comfort score from several questions, a respiratory-gas average and the number of runners are three different data structures, rather than interchangeable sample sizes.
In practice, check the number of participants, trials per condition and whether analysis was paired. Then examine how missing or faulty data were handled. Prespecified exclusion rules help explain why a measurement was omitted from a summary. Results are transparent when they show not only the best-looking mean but also traceable data handling that led to it.
Enlarge imageDetectable and useful: two separate decisions
A statistical difference indicates what the data say about the difference examined within the selected analytical framework. Practical importance, however, relates to the particular goal. A tiny cost change may matter for peak performance, while another runner gives fit, comfortable use and price greater weight. A smaller p-value does not establish that personal order of importance for you.
Defining the ‘smallest worthwhile change’ may also reflect a chosen method and goal. Ask how it was set and to which group it applies. If measurement uncertainty exceeds it, the study may simply be less suited to resolving a small effect. This neither licenses selecting favourable numbers nor justifies denying every small improvement. It instead informs the precision needed in the next measurement or study.
In your own decision, write evidence and preference separately. For example: ‘the group mean showed a small cost difference, my personal benefit is uncertain, but the fit is better’. This is more meaningful than saying a more comfortable shoe will certainly make you two per cent faster. A concrete choice remains possible while acknowledging uncertainty; it need not claim more than the measurement.
Transparency begins with the method
Funding and conflict-of-interest declarations help explain the research setting. Check who supplied equipment, who participated in analysis and whether the method or data are adequately described. An industry relationship alone does not refute a result, while concealing it makes evaluation harder. Evidence strength still comes from question, design, measurement and analysis together.
To assess two per cent, ultimately seek five pieces of information: the precise indicator and condition, repeatability of that protocol, uncertainty in the difference, presentation of individual data and the practical aim. If these are missing, the figure may remain an interesting signal, but should not be treated as a certain personal benefit. When they are available, even a small effect can be discussed on a sound basis.
Good measurement does not remove variability; it makes it known and interpretable. A well-written result reports not only the improvement percentage, but the question and precision with which it was obtained. Two per cent then becomes neither a marketing slogan nor a dismissed trifle, but information usable for a particular decision, with its uncertainty acknowledged.
Sources
- Saunders and colleagues: repeatability of an economy test in elite runners.
- Shaw and colleagues: typical error and smallest worthwhile change.
- Zanini, Folland and Blagrove: repeatability during ninety-minute running.
- Bolliger and colleagues: repeated shoe comparisons across several conditions.
How Futószív articles are prepared: our editorial guidelines.


