This page provides methodological interpretation of aggregate measurement output only. It does not decide whether two methods are safe or interchangeable for an individual.
Read the three horizontal references
| Element | Meaning |
|---|---|
| Mean difference | Average signed bias under the chosen subtraction order |
| Upper limit | Mean difference + 1.96 SD of differences under the usual normal-difference convention |
| Lower limit | Mean difference − 1.96 SD of differences |
Reversing method order reverses all difference signs. The 1.96 construction is a method-specific convention with assumptions, not a universal acceptability boundary. Whether the limits are narrow enough is a domain decision defined before inspecting the plot.1
Patterns beyond the lines
A slope suggests proportional bias: disagreement changes with measurement magnitude. A funnel suggests nonconstant difference variance. Isolated points may be data issues or valid extremes. If repeated measurements come from the same subjects, ordinary independent-observation limits are inappropriate; Bland and Altman provide extensions for repeated data.2
Do not use a correlation coefficient as a substitute. Correlation measures association and can be high when one method is systematically offset. Scatter plots show association geometry, while forest plots summarize study effects rather than method agreement.
A defensible reporting sentence
“Method A minus Method B had mean difference 2.1 units, with 95% limits of agreement −7.4 to 11.6. The plot showed increasing spread at larger means, so constant limits were not treated as adequate without further modeling.”