Best BMI Scale: What the Spec Sheet Is Really Worth

The short answer

This page will not tell you which brand to buy, because it has not measured any device, and a ranking written without measurements would be fiction. What it can do is tell you which numbers on the box change the answer and by how much, so that you can read any spec sheet and know what you are looking at.

• The BMI number is limited by your height entry, not the load cells.At 1.70 m and BMI 25, a ±0.1 kg tolerance moves BMI by 0.035; one centimetre of wrong height moves it by 0.294. Measure your height against a wall and type it in once, correctly.
• The body fat number is limited by the device's own constants, not by averaging. A fixed offset is the same every morning, so it cancels when you compare two of your own readings and never cancels when you compare against anything else.
• Buying past a point buys nothing.Once the device's body fat error falls below about 0.46 times your own day-to-day noise, you are within 10 percent of the best total error available to you, and every further improvement goes into your own protocol instead.
• The features worth paying for are the ones that reduce your error, not the ones that add digits. Correct height handling, consistent units, one decimal place, raw data export, and a current path that crosses the whole body.

If you only want one number: at 1.70 m, one centimetre of height error is worth 0.85 kg of body weight. Most scales are specified to 0.1 kg. You are comparing an eight-and-a-half kilogram problem to a one-kilogram solution.

Computed: what each printed spec is worth

A tolerance in kilograms means almost nothing to most people. Dividing by height squared turns it into BMI points, which is the unit the number is reported in. The table below is tolerance ÷ height², computed at seven heights and five tolerances.

Read down the ±0.1 kg column: the entire range of BMI uncertainty from a good weight sensor runs from 0.042 at 1.55 m to 0.029 at 1.85 m. That is the whole prize for buying a better load cell, and it is smaller than the third decimal place of the number the scale shows you.

Graduation is a separate and smaller effect. A display that steps in units of u rounds your true weight by a uniform amount of width u, whose standard deviation is u÷ √12. At 1.70 m, a 0.1 kg graduation carries a rounding standard deviation of 0.029 kg, worth 0.010 BMI — smaller than the ±0.1 kg tolerance itself. A 1 lb graduation carries 0.131 kg, worth 0.045 BMI. Graduation matters only once it is coarser than the tolerance, and on most devices it is not.

Computed: the height you typed is the biggest number in the room

Height enters BMI squared and in the denominator, so the derivative carries a factor of two: a relative error of ε in height produces a relative error of 2ε in BMI. One centimetre on a 1.70 m person is a 0.59 percent height error and therefore a 1.18 percent BMI error. The table below converts that into the two units that matter — BMI points, and the kilograms of body weight it would take to produce the same shift.

At 1.70 m and BMI 25 the effect is linear in the size of the mistake: half a centimetre costs 0.147 BMI and is worth 0.425 kg; one centimetre costs 0.294 and is worth 0.850 kg; two centimetres cost 0.588 and are worth 1.700 kg; three cost 0.882 and are worth 2.550 kg. Compare any of those to the tolerance table above. The comparison is the point of this page, so here it is directly:

One centimetre of height is worth between 7.5 and 13.0 times a ±0.1 kg weight tolerance, depending on how large you are. Even against a loose ±0.5 kg tolerance it is worth 1.5 to 2.6 times. There is no consumer scale on the market whose weight sensor compensates for a height entered from memory. This is the single highest-value thing you can fix, it costs nothing, and it is not a feature anybody sells.

Computed: how close are you to a category line

BMI is usually reported as a category, and a category has edges. If the combined error in your weight and height is wider than your distance to the nearest edge, the category itself is uncertain — which is a different and more useful question than whether the number is accurate. Combining the two errors as independent gives a required clearance of √(tol² + kgEquiv²). At 1.70 m, in kilograms of body weight:

Notice how flat the tolerance column is and how steep the height column is. Going from a ±0.5 kg sensor to a ±0.1 kg sensor changes the required clearance by a few hundred grams. Going from an exact height to a height that is one centimetre out changes it by roughly eight hundred grams to a kilogram. If your weight sits within about a kilogram of 18.5, 25, 30, 35 or 40 BMI, the honest statement is not which category you are in but that you are near a line, and the fix is a measured height.

The bands themselves are wide enough to make this manageable once the height is right: at 1.70 m the normal band spans 18.76 kg and each band above it spans 14.42 kg. One BMI point is 2.89 kg at that height, 2.56 kg at 1.60 m and 3.24 kg at 1.80 m.

Computed: the three levers inside the body fat number

Body fat is one minus fat-free mass over weight. Fat-free mass is not measured; it is built by dividing a measured quantity by an assumed constant. That structure produces a single rule: a one percent relative error anywhere in the chain costs (1 − your body fat) points — 0.90 points if you are at 10 percent, 0.80 at 20 percent, 0.70 at 30 percent. Leaner people pay more per percent of error, because there is more lean tissue for the error to land on. The levers:

Three things follow. First, the hydration constant is a finer lever than people expect: moving the assumed water fraction of fat-free tissue from 0.73 to 0.74 swings a 20 percent reading by 1.08 points, and two devices that disagree only about that constant will disagree by roughly two points before anything else is considered.

Second, height enters the body fat number too, and twice as hard as it enters BMI, because the impedance-to-water conversion uses height squared. One centimetre at 1.70 m moves a 20 percent reading by 0.94 points — three times the 0.294 BMI shift from the same centimetre.

Third, the leg-share lever is the largest of the three and it belongs only to foot-only devices. A foot-to-foot scale sends its current up one leg and down the other, then scales what it found up to a whole body using an assumed share. Under the pure-ratio form of that model, a 5 percent relative error in the assumed share is worth 4.00 points at 20 percent body fat. Real devices use regression equations fitted on populations rather than a bare ratio, so the effective sensitivity is smaller than this — treat 4.00 as the size of the lever, not as a measured error. It is still the reason a hand-to-foot path, which crosses the trunk as well as the limbs, is the structural upgrade rather than a marketing one, and the reason these devices behave worst on people whose legs are unrepresentative of their whole body.

Computed: how much a better scale can actually buy you

Two errors combine as the square root of the sum of squares. Call your own day-to-day noise u— the part that comes from your hydration, your last meal, the time of day — and the device's contribution d. The total is √(u² + d²), and the best you could ever reach by buying a perfect device is u.

Reading down the u = 1.5 column: a device contributing 3.0 points gives a total of 3.354, and a perfect device gives 1.500. The entire range of hardware quality is worth 55 percent of your total error. The rest is yours. Halving the device error, over and over:

The returns collapse. The first halving is worth 36.8 percent of your total error; by the time the device is down to half a point, halving it again is worth 3.8 percent. There is a clean place to stop. Solving √(u² + d²) ≤ 1.10·u gives d ≤ 0.4583·u: once the device error is below about 0.46 times your own noise, better hardware cannot move your total by more than 10 percent. If your day-to-day noise is 1.5 points, that threshold is 0.69 points; at 2.5 points of noise it is 1.15. The marginal rate makes the same point — at u = 1.5, each further point of device improvement returns 0.894 points of total when the device is at 3.0, but only 0.164 when the device is at 0.25.

Computed: averaging has a floor

Averaging n readings divides your own noise by √n. It does nothing at all to a device offset, because a constant is the same on every reading. That gives two different floors, and confusing them is the most common way people misread these devices.

• Comparing two of your own readings on the same scale. The offset cancels. Your floor is 1.96·√2·(u ÷ √n). With 1.5 points of daily noise, a single pair of readings can only resolve a 4.16 point change; a week of readings on each side brings that to 1.57 points.
• Comparing against another device, a chart or a target. The offset does not cancel. Your 95 percent band is 1.96·√(d² + u²/n), and it can never fall below 1.96·d. With a ±3.5 point device that floor is ±6.86 points, at any number of readings.

Readings column assumes 1.5 points of your own daily noise; with 2.5 points the ±0.5 row needs 114 readings at d = 0.1 and 250 at d= 0.2. The word “impossible” is literal: once the device offset exceeds the target divided by 1.96, no amount of data recovers it. This is the arithmetic answer to “should I weigh in every day?” — daily weighing is worth it for the trend, and worthless for the comparison against anything outside your own bathroom.

The decision rule: which spec matters for what you are doing

“Best” is a function of the job. Each row below names the one spec that decides the outcome, the threshold this page's arithmetic puts on it, and what you can safely ignore.

Put together, the shopping list is short and almost none of it is about accuracy claims. You want a device that makes you enter height once and correctly and stores it per user; that reports weight to one decimal in a unit you will keep using; that lets you export the raw history rather than locking it in an app; and whose current path crosses your whole body if you care about the body fat number at all. You do not need the tightest printed tolerance on the shelf, and per the tables above you cannot buy your way past your own hydration.

Setting one up so the numbers are usable

Most disappointment with these devices is a protocol problem rather than a hardware problem. This is the short version; the longer version on our smart scale page covers what the device measures at each step.

• Measure your height barefoot against a wall with something flat on your head, and enter it to the nearest half centimetre. Per the tables above this is worth more than every other setting combined.
• Put the scale on a hard, level floor and never move it. Carpet and uneven tiles change the load cell readings, and moving the scale changes them again.
• Weigh at the same point in your day — after the bathroom, before food and coffee is the most repeatable state most people have.
• Leave four hours between hard training or a sauna and a reading, and skip the morning after a very salty meal or a long flight.
• Record the reading every time and decide from the seven-day average. One reading resolves nothing that matters.
• Re-enter your height once a year. A stale height quietly corrupts both the BMI and the impedance conversion, and nothing on the device will tell you.

Where this page's arithmetic stops being true

• No product is named, ranked, reviewed or recommended here. This page has not measured any device. There are no prices, brands, model numbers or star ratings on it, by deliberate choice.
• The body fat model is a model.The hydration fraction of 0.73 and the height-squared impedance relationship are this page's working description of a class of devices. They are not any manufacturer's firmware, and no manufacturer publishes its constants.
• The leg-share lever is an upper bound. 4.00 points per 5 percent is what the pure-ratio form gives. Real devices use population-fitted regression equations, so the effective sensitivity is smaller. The number sizes the lever; it is not a measured error for any device.
• “Tolerance” is not defined consistently across the industry. This page treats the figure you enter as a bound comparable to a standard deviation and combines errors as independent. What a given box meant by its number may be something else entirely.
• BMI categories are WHO classifications for adults. They are screening categories, not diagnoses, and they do not apply to children, to pregnancy, or to people with unusual muscle mass. The class-boundary arithmetic says whether your category is certain; it says nothing about your health.
• The BMI-method body fat shown in the calculator is not independent. It is a population prediction equation evaluated on your BMI, age and sex, so agreement with your scale is not confirmation of either number.
• Your own noise figure is an estimate. The tables treat it as known. If you want it measured rather than assumed, weigh daily for two weeks under fixed conditions and take the standard deviation of the readings.
• None of this is medical advice. See the disclaimer.

Best BMI scale questions

Related tools

BMI calculator · BMI scale accuracy and what a smart scale measures · body fat calculator · measure body fat percentage · fat percentage across three methods · visceral fat calculator · how to measure body fat at home · body fat percentage chart

Not medical advice.Every figure on this page is arithmetic this page performs on the BMI definition and on a stated model of how impedance devices estimate fat-free mass. No device was measured, no product is ranked, and nothing here is a claim about any manufacturer's accuracy. See our disclaimer.