Bland–Altman — Worked example

Example data

Suppose we measured 8 subjects with two devices:

A: 102  98 110 105  99 101 107 103
B: 100  97 111 106 101 100 108 102

Step-by-step

1. Compute paired differences (d = B − A) and paired means (m = (A+B)/2).
2. Bias = mean(d). This is the average systematic offset between methods.
3. SD of differences = sample SD(d).
4. 95% limits of agreement = bias ± 1.96·SD(d).
5. Interpretation: Are the LoA narrow enough for your application?

Intermediate results (illustration)

Differences d: −2 −1 +1 +1 +2 −1 +1 −1

• bias = mean(d) ≈ −0.125 (slightly lower B vs A on average)
• SD(d) ≈ 1.46
• LoA ≈ −0.125 ± 1.96·1.46 → [−2.99, 2.74]

Try it in the tool

Open with these values

FAQ

What does Bland–Altman measure?

It assesses agreement between two methods by analyzing differences and limits of agreement, not correlation.

Why not just use correlation?

Correlation can be high even when one method is biased. Bland–Altman focuses on agreement in the original units.

What is an “acceptable” LoA?

It depends on domain tolerance: clinical, engineering, and lab contexts have different acceptable error ranges.