A high correlation answers a narrower question
No. A correlation coefficient describes how strongly two sets of values vary together. It does not say that the methods return the same value for the same observation. Bland and Altman made this distinction in their original method-comparison paper; NIST’s statistical guidance likewise uses paired differences to assess agreement.
Consider a deliberately invented example, not a device study: method B always reports A plus two arbitrary units. The two lists rise and fall together perfectly, yet every pair differs by two units. Whether that difference matters depends on what the measurement is for. This toy example makes no claim about a real sensor, person or clinical threshold.
How far apart are paired values, and is that gap acceptable for this use?
What the original paper actually demonstrated
Bland and Altman illustrated the statistical problem using peak-expiratory-flow readings from 17 colleagues and family members, selected to span a wide range. Each person had readings with two meters. The authors explicitly said this convenience sample was not representative of a defined population and did not provide evidence that the meters were comparable.
In the example, the two methods had a strong reported correlation while the paired differences were much wider than that one statistic revealed. We use the example to explain the analysis, not to judge either meter today. The author-hosted reproduction incorporates numerical corrections to the 1986 print article; we rely on that corrected text and avoid transplanting its example numbers into a current product claim.
Look at differences, then their limits
For each paired observation, subtract one method’s value from the other and plot that difference against the pair’s mean. This can reveal a consistent offset, widening spread or a pattern at higher values that a correlation headline hides. A mean difference summarizes average bias. Estimated limits of agreement describe the spread of individual differences under the method’s assumptions.
NIST explains that conventional limits use the mean difference and the standard deviation of differences. A small dataset can make those estimated limits uncertain; skewed or changing differences can make a simple summary inappropriate. Neither a graph nor a formula alone decides whether the difference is acceptable. The intended use and a justified tolerance have to supply that judgment.
Association → paired difference → acceptable tolerance for the intended use.
Do not turn a method lesson into a device verdict
The original 17-person convenience sample was selected as a teaching example, not a representative test of wearables, software or today’s meters. NIST’s page is statistical documentation, not a clinical trial. The invented offset example is illustration only. None of these sources rank products, validate a health dashboard or explain an individual’s changing readings.
When you see “highly correlated,” look for paired observations, the difference plot, the size and uncertainty of agreement limits, the tested population and conditions, and a tolerance tied to the decision the method is supposed to support. If these are absent, treat interchangeability as unproven rather than assuming the methods agree.