Data
Bands on the curve
Percentile curves are estimates, and the confidence band around each one widens toward the tails.
Guides on Data: Meta-analysis statistics, start to finish, Reading a size study critically, How a size study is built, Reading a nomogram without fooling yourself, The Veale meta-analysis, read properly
A confidence band around a percentile curve shows how far the curve might move if the study were repeated with a different sample. It is narrowest near the middle, where the curve is anchored by the mean, and widest in the tails, where it depends on the spread as well. A chart that prints a single clean line is hiding that uncertainty, not removing it.
What the band is a band around
A nomogram turns a measurement into a percentile using two estimated quantities: a mean and a standard deviation. Veale et al. (2015) built theirs from weighted means and pooled standard deviations across the contributing studies. Both are estimates from samples, so both could have come out slightly different had different men been measured.
The percentile curve is a function of those two numbers. Wherever the estimates move, the curve moves with them, and a band is the range of curves that the data cannot rule out. This post assumes you know what the curve is, and what a nomogram is covers it if you do not.
Two sources of uncertainty, with different reach
The mean and the standard deviation each shift the curve, and in different ways.
An error in the mean slides the entire curve left or right by the same amount. A mean that is off by 0.1 cm moves the figure for every percentile by 0.1 cm, in the middle and in the tails alike.
An error in the standard deviation stretches or squeezes the curve around its centre. The figure at any point is the mean plus a number of SDs, so an error in the SD is multiplied by that number. In the middle, at a z-score of zero, it has no effect at all. At two SDs out, an SD that is wrong by 0.1 cm moves the figure by 0.2 cm. At three SDs out, by 0.3 cm.
| Position on the curve | Z-score | Effect of an SD error of 0.1 cm |
|---|---|---|
| Median | 0 | 0 cm |
| About 84th percentile | 1 | 0.1 cm |
| About 98th percentile | 2 | 0.2 cm |
| About 99.9th percentile | 3 | 0.3 cm |
These rows are arithmetic, not data from a study: they show how a fixed error in the spread grows with distance from the centre. Add the error in the mean, which is the same everywhere, and the band is thinnest at the middle and fans out toward both ends.
Why the tails are worse than the arithmetic alone
The fan-out above assumes the shape of the curve is right. A nomogram built on a normal distribution extends that shape into the tails, where few men are measured and the assumption is tested least. The sample that estimates a percentile in the tails is, by definition, a sliver of the sample: only a small share of men sit past the 95th percentile, so any direct estimate of that region rests on very few observations.
That is a separate point from the one in why the tails of the distribution are so thin, which is about how rare extreme values are, not about how well they are estimated. It also connects to whether the distribution is normal: if the real tails differ from the assumed ones, the band describes only part of the uncertainty, because it takes the shape for granted.
What the band does not include
A confidence band reflects sampling error, the randomness of which men were drawn. It does not reflect everything that can make a curve wrong.
- Method differences between studies, which can shift a whole study's mean by more than its interval.
- Selection, such as which men attend the clinics that supply the data.
- A reference sample that does not match the person using it.
Pooled data add a further layer, because studies that disagree more than chance allows widen the true uncertainty beyond what a simple formula gives, which is the subject of heterogeneity in pooled studies. A band that looks tight can still be centred on the wrong thing.
How to read a figure against the curve
For a percentile read off a nomogram, the useful habit is to take the reading with its own uncertainty and ask how much of the percentile range it covers. Near the middle, a centimetre of reading error moves a percentile a lot, as why one centimetre swings your percentile shows, but the curve itself is well estimated there. In the tails, a small error in the reading moves the percentile less, but the curve itself is less certain, and so the two effects trade places.
The practical advice is to quote a percentile as a range, and to say the tails are less certain than the middle. A figure reported as "somewhere around the 90s" is more honest than "the 96th". A trustworthy percentile calculator says which study and state it assumes and shows a range, which is the same idea applied to a tool.
A note on neighbouring questions
Questions about how a photo-scoring tool arrives at its output are not about uncertainty in a sample mean at all. How an automated scoring tool works is for that site to describe, and the tools hub of a score-based rater covers what is on offer there. A human judge gives an impression with its own variability, and Rate Cock produces a rating of an image, a different kind of quantity with no percentile curve in the sense used here.
For a measured length, the curve is an estimate with a band, and the band widens where the data thin out. That is the whole of the point, and it is a reason to hold a tail percentile loosely and a median one more firmly.