Perfect correlation, different readings
Consider four fictional measurements from device A: 60, 80, 100, and 120. Device B gives 70, 90, 110, and 130 for the same observations. Every increase in A is matched by the same increase in B, so the Pearson correlation is exactly 1. Yet B is ten units higher every time.
If someone switched devices, the recorded value would jump even though the underlying condition had not changed. A statement that the devices are “perfectly correlated” would be mathematically correct and practically incomplete. The missing information is whether a ten-unit discrepancy is acceptable for the purpose of the measurement.
What correlation is responding to
Correlation summarizes a pattern of association across observations. It is affected by the range of values being compared. A study spanning a broad range can produce a strong correlation even when discrepancies would be consequential within a narrower range relevant to a particular user.
Imagine comparing height measurements across children and adults rather than among adults of similar height. Much of the variation now comes from large differences between people. A method might preserve that ordering while still disagreeing enough to complicate the smaller comparisons someone actually cares about. The correlation and the intended use are describing different scales of variation.
That is why a product claim based on a single correlation coefficient leaves several questions unanswered: how far apart the readings were, whether the differences changed at higher values, and which people or conditions were tested. A large sample can estimate the correlation precisely without resolving these other questions.
Differences reveal the offset and the spread
An agreement analysis examines paired differences. In the first example, B minus A is always ten. The average difference reveals the offset, while the absence of variation in those differences makes this invented case unusually simple. Real measurements commonly vary around an average discrepancy.
Consider a second fictional comparison with differences of minus 20, minus 10, 10, and 20 units. The average difference is zero, but no pair agrees exactly. Positive and negative discrepancies cancel. Describing the methods as unbiased on average would not mean that individual readings can be exchanged without consequence.
Both examples show why an average and a spread provide different information. The first has an offset with no variation in the discrepancy. The second has variation without an average offset. A real comparison can have both, and either may matter for the decision being made.
What a Bland–Altman plot adds
A common approach plots the difference between paired measurements against their average. It can reveal whether disagreement becomes larger at higher values or whether a small group of observations behaves differently. Conventional limits of agreement are often calculated as the mean difference plus and minus 1.96 standard deviations of the differences, when the relevant assumptions are appropriate.
Those limits describe the distribution of differences; they are not automatically acceptable limits for use. They also differ from a confidence interval around the average difference. The latter describes uncertainty in the estimated mean, while limits of agreement address variation among paired measurements. Both can have uncertainty of their own.
Repeated readings from the same people, nonconstant variability, and non-normal differences can require a different analysis. Applying a familiar formula does not remove the need to account for the study design. The plot is useful partly because it can make problems visible that a single summary conceals.
Acceptable for which decision?
A discrepancy might be tolerable for displaying a broad activity trend and unacceptable near a threshold used for a consequential decision. There is no universally good limit of agreement independent of units and purpose. An acceptable range is more interpretable when justified before inspecting the results rather than chosen afterward to fit them.
For a hypothetical tracker, daily totals could agree reasonably well while minute-by-minute estimates differ substantially. Validation of the daily total would not settle the minute-level question. Similarly, testing a device at rest would not establish its performance during movement. The time scale and conditions belong in the interpretation.
Agreement is not the same as truth
Two devices might share an incorrect assumption or similar calibration problems and therefore agree with each other. A reference method can help evaluate accuracy, but it too has uncertainty and a defined scope. Agreement with a consumer device is a different claim from agreement with an appropriate reference measurement.
NIST also distinguishes repeatability under the same conditions from reproducibility under changed conditions. A device can give very consistent readings without being accurate. Together, these distinctions make a comparison easier to understand: consistency, association, closeness, and accuracy each describe something useful, but none should silently stand in for all the others.
