What percentage difference measures, and when it is the right measure
Percentage difference expresses the gap between two numbers as a share of their average. Use it when the two numbers have equal standing — when there is no earlier value, no correct value, and no reason to treat one as the reference.
That situation is more common than it sounds. Two aliquots of the same sample run through the same instrument. A part measured with callipers and again with a micrometer. Two contractors quoting the same job. Two sensors on the same duct. In every case, calling one of them “the base” would be an arbitrary choice that changes the answer, and the choice would be doing work the data does not support.
Analytical chemistry gives the measure its own name and its own acceptance criteria. Relative percent difference, RPD, is exactly this formula applied to a pair of duplicate results, and quality-control programmes set limits on it — EPA SW-846 methods, for example, specify RPD control limits for laboratory duplicates as part of the method's quality-assurance requirements. A pair failing its RPD limit triggers a re-analysis, so this is not an academic quantity.
The measure has one structural weakness, and it comes from the denominator. If the two values are of opposite sign, their mean lies between them and can be arbitrarily close to zero, which sends the percentage difference to arbitrarily large values. If they are exactly equal and opposite, the mean is zero and the quantity does not exist at all. Percentage difference is a tool for two numbers on the same side of zero and of broadly comparable size.
The formula, and how it relates to percent change
Take the size of the gap, divide by the average of the two values, multiply by 100:
% difference = 100 × |A − B| ÷ |(A + B)/2|
The numerator is an absolute value, so the result is never negative. That is deliberate. A signed result would need a direction, and a direction needs a reference, which is precisely what this measure declines to choose. If you need direction, you have a baseline, and you want percent change instead.
The two measures share a numerator and differ only in their denominator, which makes their relationship easy to state exactly. Writing Δ = |A − B|, percent change from A is 100Δ/|A| and percentage difference is 100Δ/|m| where m is the mean. Their ratio is therefore
% difference ÷ % change from A = |A| ÷ |m|
So percentage difference is smaller than the percent change measured from A exactly when |m| > |A|, which for two positive numbers means B > A. Measure the change from the larger of the two values and percentage difference comes out larger; measure it from the smaller and percentage difference comes out smaller. It always sits between the two possible percent changes, which is another way of saying it is the compromise between two arbitrary choices of base.
One consequence surprises people: the percentage difference between two values can exceed 100%. It reaches exactly 100% when one value is three times the other, because then the gap is 2x and the mean is 2x. Beyond a 3:1 ratio it keeps climbing, approaching 200% as one value approaches zero. Percent change has no such ceiling either, but its behaviour is different, and comparing the two on the same pair of numbers is the fastest way to see which one your reader will misread.
Worked example: duplicate laboratory results
A laboratory runs a sample in duplicate for lead by ICP-MS and reports 2.9 mg/kg and 3.0 mg/kg. The method's quality-control limit for duplicates is an RPD of 20%. Do the results pass?
- Absolute difference. |2.9 − 3.0| = 0.1 mg/kg.
- Mean. (2.9 + 3.0) ÷ 2 = 5.9 ÷ 2 = 2.95 mg/kg.
- Ratio. 0.1 ÷ 2.95 = 0.0338983. Check it back: 2.95 × 0.0338983 = 0.1000, so the division is right.
- Percentage difference. 100 × 0.0338983 = 3.39%.
3.39% against a 20% limit is a comfortable pass, and the pair would be accepted without re-analysis.
Now see what a baseline choice would have done. Percent change from 2.9 to 3.0 is 0.1 ÷ 2.9 = 3.4483%. Percent change from 3.0 to 2.9 is −0.1 ÷ 3.0 = −3.3333%. The percentage difference of 3.3898% lies between the magnitudes 3.4483% and 3.3333%, as the algebra above requires. Here the three numbers are close enough that nothing hangs on the choice.
Widen the gap and the choice starts to matter. Compare 1 and 3: the percentage difference is 2 ÷ 2 = 100%, percent change from 1 to 3 is 200%, and percent change from 3 to 1 is −66.67%. Three defensible descriptions of the same pair, spanning a factor of three. When you report a percentage comparison of two dissimilar numbers, you have to say which one you divided by, because the reader cannot recover it from the percentage alone.
What counts as an acceptable percentage difference
There is no universal threshold, because the acceptable gap depends on what produced the two numbers. Three benchmarks are worth having.
A published control limit. Where one exists, it wins. Environmental and clinical laboratories work to method-specific RPD limits, typically stated in the method itself; anything from 10% to 30% is common depending on the analyte, the matrix and how close the result is to the detection limit. Duplicates near the detection limit are usually judged on an absolute criterion instead, because RPD becomes unstable as both values approach zero.
Note the direction of that last point carefully: as both values shrink towards zero with a fixed absolute gap between them, the mean shrinks too, so the percentage difference grows. A 0.1 unit gap is 3.4% at a mean of 2.95 and 20% at a mean of 0.5. Nothing about the measurement got worse; the denominator got smaller.
The combined precision of the two methods. If instrument A repeats to ±2% and instrument B to ±3%, a 4% difference between them is unremarkable. Expect roughly the quadrature sum of the two relative precisions, and treat anything much beyond that as evidence of a real bias rather than of scatter.
The decision the number feeds. A 5% difference between two structural load estimates and a 5% difference between two paint-colour readings are not the same event. Ask what changes at the threshold you are approaching, and set the limit from that rather than from a round number.
When a percentage difference is large, look at the absolute difference before concluding anything. Percentage measures are compression: they throw away the scale on purpose. A 40% difference between 0.002 and 0.003 and a 40% difference between 2,000 and 3,000 are the same number and rarely the same problem.
Percentage difference against percent change for the same pairs
| A | B | Mean | Percentage difference | % change A→B | % change B→A |
|---|---|---|---|---|---|
| 100 | 101 | 100.5 | 0.9950% | 1.0000% | −0.9901% |
| 100 | 105 | 102.5 | 4.8780% | 5.0000% | −4.7619% |
| 100 | 110 | 105 | 9.5238% | 10.0000% | −9.0909% |
| 100 | 125 | 112.5 | 22.2222% | 25.0000% | −20.0000% |
| 100 | 150 | 125 | 40.0000% | 50.0000% | −33.3333% |
| 100 | 200 | 150 | 66.6667% | 100.0000% | −50.0000% |
| 100 | 300 | 200 | 100.0000% | 200.0000% | −66.6667% |
The percentage difference always lies between the magnitudes of the two percent changes, because its denominator lies between the two possible bases. The 100:300 row is where percentage difference reaches exactly 100%.
Ways this measure gets misused
- Using it when there is a baseline. Sales this year against sales last year is a percent change question. Dividing by the mean of the two years describes nothing anyone wants to know.
- Using it when one value is authoritative. A measurement against a certified reference value is percent error, which divides by the accepted value. Dividing by the mean drags the reference towards your own result.
- Applying it across zero. With values of opposite sign the mean can be close to zero and the percentage explodes. The calculator warns about this, and the honest answer is usually to report the absolute difference.
- Comparing quantities in different units. The formula is dimensionless only because the units cancel. 3 mg/L against 3 µg/L is a factor of 1,000, not a 0% difference.
- Reporting a percentage difference near the detection limit. Both values are then dominated by noise and the mean is small, so the measure is at its least stable exactly where laboratories most often want a number.
- Averaging percentage differences across pairs. The mean of several RPDs is not the RPD of the pooled data. If you need a summary of many duplicate pairs, use a pooled relative standard deviation.
Choosing between the percentage comparisons
Four measures sit in this family and the choice between them is entirely a choice of denominator.
Percentage difference divides by the mean. Symmetric, non-negative, no baseline required. Duplicate samples, inter-instrument comparisons, two independent estimates.
Percent change divides by the earlier or original value and keeps its sign. Time series, before-and-after, growth rates.
Percent error divides by the accepted value. One value is correct by assumption, the other is being judged against it.
Percentage of a whole divides by a total of which the value is a component. Shares, compositions, market splits.
Two habits make percentage comparisons far more readable. First, always publish the absolute difference beside the percentage, because the percentage alone cannot tell a reader whether the gap matters. Second, name the denominator in words — “3.4% of the mean of the pair” rather than just “3.4% different” — since the same pair supports several different percentages and only the denominator distinguishes them.
Where the comparison is between more than two values, none of these scale up gracefully. A relative standard deviation, which is the standard deviation divided by the mean, generalises the idea properly to a set of replicates, and it reduces to a fixed multiple of the percentage difference for a pair: for two values the sample standard deviation is |A − B|/√2, so the relative standard deviation is the percentage difference divided by √2, or about 70.7% of it. If you are moving from duplicates to triplicates, that is the quantity to switch to.