What 99% Accurate Really Means on a Body Composition Scale

Episode Summary

A client sent Mikki a link to a very flash Withings bioelectrical impedance scale — NZ$819, marketed as 99% as accurate as a DEXA scan — and asked the obvious question: is it worth the money? In this solo episode Mikki goes and reads the validation document behind that claim, and finds the problem isn't really the scale at all. It's a statistical sleight of hand that turns up constantly in health tech marketing: reporting a correlation and calling it accuracy.

Mikki explains the difference in plain terms, with a worked example where a scale reads exactly half the true value every single time and still produces a perfect correlation of 1.0. She then sets the company's own document against three pieces of independent research — a January 2026 systematic review of BIA against a four-compartment model, Siedler and colleagues' 2023 comparison of fifteen BIA devices, and a June 2026 study putting the InBody H20N up against DEXA — to show what honest reporting of a body composition device actually looks like. She finishes with the four readouts on these scales that aren't worth your attention, a practical protocol for getting genuine information out of the one you already own, and four questions worth asking of any health tech accuracy claim you meet in the next few years.

Key Topics

• Correlation is not accuracy. The 99% claim rests on a correlation of 0.99 for total fat mass across 80 people. Mikki accepts the correlation is real, and shows with a worked example why it says nothing about whether any individual reading is right.
• What you actually want is agreement. Mean difference and limits of agreement, via a Bland-Altman analysis — the standard approach since the mid-1980s, and one the company's document didn't use.
• What the validation document left out: mean difference from DEXA, limits of agreement, individual prediction error, the age and sex breakdown, the body fat range, whether hydration and fasting were standardised, and whether the validation sample was independent of the data used to build the algorithm. Mikki could find no peer-reviewed publication of it, and notes total body water was validated against a different device entirely.
• Better hardware doesn't settle the evidence question. Mikki is clear the Withings unit is well built — eight electrodes including a handheld bar, multiple frequencies, current through arms, trunk and legs rather than up one leg and down the other. That is a separate question from whether the accuracy claim stands up.
• The independent evidence. A January 2026 systematic review of 12 studies found mean bias for body fat percentage ranging from 3.5 percentage points too low to 4.4 too high, with limits of agreement typically spanning 15 to 20 percentage points, and concluded BIA was not equivalent to the four-compartment reference.
• Reliable does not mean accurate. Across 15 devices, precision error was 0 to 0.5 percentage points — these scales are very consistent with themselves. A device can be reliably wrong, which is why direction of change can still be informative even when the absolute number isn't. The longitudinal results were more mixed: some devices tracked change reasonably, others didn't.
• What honest reporting looks like. In 55 adults with obesity, the InBody H20N sat 1.64 percentage points from DEXA on body fat, with limits of agreement of −3.4 to +6.7 — good across a group, much looser for the individual standing on it. Skeletal muscle mass was worse, overestimated by 5.66 kg on average. Less impressive than 99%, and far more useful.
• Four readouts not worth watching: visceral fat (a proprietary index on an arbitrary scale, not a measurement of visceral adipose tissue like CT or MRI); bone mass (not bone mineral density, and no guide to osteoporosis risk — DEXA is the answer there); basal metabolic rate (calculated from an equation, not measured — that needs indirect calorimetry); and body water (better on a multi-frequency device, still modelled).
• How to use the scale you have. Standardise everything you can control, because hydration, glycogen and cycle phase move these numbers more than real change does over a short window. Measure three or four times a week, average weekly, read four-week trends, and triangulate against weight, waist, clothing fit, training numbers and photos. Swap "my body fat is 23.7%" for "my body fat estimate has trended down over the last 10 weeks".
• The transferable point. InBody's own product page claims over 98% agreement with DEXA — near enough the same move. When you're told a number is 99% accurate, ask what that means, what it was compared against, who paid for the comparison, and how far off an individual reading can be.

Chapters

00:00 The $819 scale and the 99% claim
02:15 Why correlation is not accuracy
04:27 Better hardware, missing evidence
06:48 What the independent research shows
09:10 InBody versus DEXA: the real numbers
11:53 Readouts not worth watching
13:27 How to use a scale well
15:40 Questions to ask of health tech

Guest / Resources

This is a solo episode — no guest.

Studies discussed

• The Validity of Bioelectrical Impedance Analysis Compared to a Four-Compartment Model in Healthy Adults: A Systematic Review — Journal of Functional Morphology and Kinesiology, 31 January 2026. Twelve studies; mean bias for body fat percentage −3.5 to +4.4 percentage points; limits of agreement typically 15–20 percentage points. https://doi.org/10.3390/jfmk11010065
• Siedler MR et al. Asse