Open StatClinic →
Home› Blog› Bland-Altman Analysis
📈 Method Comparison

Bland-Altman Analysis in Medical Research:
Complete Method Comparison Guide

🕑 26 min read 📅 July 2026 ✅ Peer-reviewed content 📚 3700+ words
S
StatClinic Editorial Team Statistical content for medical researchers and clinicians
A new portable pulse oximeter shows a Pearson correlation of r = 0.97 with a standard hospital device across 80 patients. The manufacturer publishes this as evidence that the two devices are "highly interchangeable." A hospital procurement committee is about to approve a large purchase when a biostatistician points out the problem: a correlation of 0.97 tells you the two devices rank patients in the same order — nothing more. Two devices that consistently differ by 8 percentage points in SpO₂ readings will still show r = 0.97. In a patient hovering at the critical 88% threshold for oxygen supplementation, this 8-point systematic difference is the difference between treatment and no treatment. The question the hospital actually needs answered is: "How closely do the two devices agree in actual SpO₂ units?" This is the question that Bland-Altman analysis was created to answer.

What Is Bland-Altman Analysis?

Bland-Altman analysis — also called the method of differences or the difference plot — is a graphical and statistical technique for evaluating the agreement between two quantitative measurement methods applied to the same subjects. It was introduced by Martin Bland and Douglas Altman in their landmark 1986 paper in The Lancet, which has since become one of the most cited papers in medical literature, appearing in tens of thousands of subsequent studies.

The method directly addresses a specific clinical question: "Can these two measurement methods be used interchangeably for the same clinical purpose?" A new automated blood pressure device needs to agree with an auscultatory reference; a point-of-care HbA1c test needs to agree with a laboratory analyser; an ultrasound tumour volume estimate needs to agree with CT volumetry. In all these cases, the clinical audience needs to know not whether the two methods correlate, but whether their differences are clinically tolerable.

Bland-Altman analysis achieves this by plotting, for each subject, the difference between the two methods on the y-axis against the mean of the two methods on the x-axis — then calculating the mean bias (systematic offset between methods) and the 95% Limits of Agreement (the range within which 95% of future individual differences are expected to fall).

Why Correlation Cannot Answer the Method Comparison Question

The core methodological argument made by Bland and Altman in 1986 was that correlation is the wrong tool for method comparison — and this argument remains just as relevant forty years later, as papers using Pearson's r or Spearman's ρ to validate measurement methods continue to be published and criticised in peer review.

Correlation measures the strength of a linear relationship — whether two variables move together in the same direction. It does not measure agreement. Three specific limitations make correlation fundamentally unsuitable for method comparison:

  1. Insensitive to systematic bias: If Device A reads 10 mmHg higher than Device B for every patient, the correlation is still perfect (r = 1.00) because the rank ordering is preserved. Yet the devices are clearly not interchangeable for clinical use.
  2. Inflated by sample heterogeneity: Pearson's r increases automatically when the range of values in the sample is wide. A method comparison study with patients ranging from hypertensive to hypotensive will produce a higher r than the same study in a homogeneous sample — even if the measurement errors are identical in both. The r is measuring the sample range as much as the method agreement.
  3. No clinical units: A correlation of r = 0.92 gives no information about how many mmHg, mg/dL, or degrees the two methods differ in practice. It is dimensionless and cannot support clinical decisions about interchangeability.
The paired t-test is equally inadequate: A paired t-test on the differences (Method A − Method B) tests only whether the mean difference is significantly different from zero. A non-significant result does not mean good agreement — it means the mean bias is not detectable, but individual differences can still be enormous. Two methods with mean bias = 0 but individual differences ranging from −30 to +30 units would pass a t-test while being completely useless for patient care.

Constructing the Bland-Altman Plot

For n subjects, each measured once by Method A and once by Method B, the Bland-Altman procedure involves four steps:

  1. Compute the difference for each subject: dᵢ = Aᵢ − Bᵢ
  2. Compute the mean for each subject: mᵢ = (Aᵢ + Bᵢ) / 2
  3. Plot dᵢ on the y-axis against mᵢ on the x-axis
  4. Draw three horizontal reference lines: the mean difference (bias), and the upper and lower 95% Limits of Agreement

Bland-Altman Plot Structure

Mean bias +1.96 SD −1.96 SD Mean of two methods Difference (A − B)
Red dashed lines = upper and lower 95% Limits of Agreement (±1.96 SD). Purple solid line = mean bias. Red dot = outlier outside LoA (expected ~5%).
Mean bias (d̄)
95% Limits of Agreement (d̄ ± 1.96 SD)
Zero difference reference

The Core Calculations

The three key statistics from Bland-Altman analysis are the mean bias, the SD of differences, and the 95% Limits of Agreement:

d̄ = ∑(Aᵢ − Bᵢ) / n
Mean bias: systematic average difference between the two methods
SDd = √[∑(dᵢ − d̄)² / (n−1)]
Standard deviation of the individual differences
LoA = d̄ ± 1.96 × SDd
95% Limits of Agreement: the interval expected to contain 95% of all future pairwise differences
d̄ = mean of differences (positive = Method A reads higher)
SDd = standard deviation of all individual differences
n = number of subjects (each measured by both methods)
1.96 = z-score for 95% coverage (use t-statistic for n < 30)

Confidence Intervals for the Limits of Agreement

The LoA are estimates from a sample and have their own uncertainty. Bland and Altman (1986) provided formulas for the 95% confidence intervals around each limit:

These CIs should be plotted as horizontal error bars on the LoA lines in the Bland-Altman figure and reported numerically in the results section. Wide CIs around the LoA — which occur in small samples — indicate substantial uncertainty about the true limits and should be interpreted cautiously.

Interpreting the Bland-Altman Plot

What to Look for — Four Key Questions

  1. Is the mean bias clinically acceptable? A non-zero mean bias indicates Method A consistently reads higher or lower than Method B. Whether this is clinically acceptable depends on the context — a +2 mmHg bias in blood pressure is generally tolerable; a +0.5% bias in HbA1c straddles a diabetes management decision threshold.
  2. Are the Limits of Agreement within the pre-specified maximum acceptable difference? Both the upper and lower LoA must fall within the pre-specified clinically acceptable range for the methods to be considered interchangeable. If either LoA falls outside the acceptable range, the methods are not interchangeable — even if the mean bias is small.
  3. Is there proportional bias? Scan the scatter of points across the x-axis. If the points fan out or converge (differences grow larger or smaller as the mean increases), there is proportional bias — the two methods agree better in one part of the range than another. Formally test this with linear regression of dᵢ on mᵢ.
  4. Are approximately 5% of points outside the LoA? By definition, ~5% of differences should fall outside the 95% LoA. If substantially more than 5% fall outside, the LoA may be too narrow (violating the normality assumption) and the differences should be checked for outliers or the normality of the differences should be formally tested.

Proportional Bias in Detail

Proportional bias occurs when the discrepancy between two methods is not constant but scales with the magnitude of the measurement. This is common in analytes spanning a wide range (e.g., troponin from near-zero to very high, tumour diameter from 5 mm to 80 mm) where one method's measurement error grows proportionally with the true value.

Detection: compute Pearson's correlation between the differences (dᵢ) and the means (mᵢ). A statistically significant correlation (p < 0.05) confirms proportional bias. When present, the fixed LoA from standard analysis are invalid — the LoA are too wide at one end and too narrow at the other. Solutions include:

Clinical Examples

1
Radiology: MRI vs CT Measurement of Liver Volume
A hepatology centre evaluates whether MRI-based liver volumetry can replace CT volumetry for pre-operative surgical planning in living-donor liver transplantation. Sixty consecutive donors undergo both CT and MRI on the same day. CT volumetry is the established method; MRI is the proposed alternative, offering no ionising radiation.
60
Subjects (n)
1,284 mL
Mean CT volume
1,271 mL
Mean MRI volume
±150 mL
Pre-specified MAD
Step 1 — Compute differences and means for each subject d_i = CT_i − MRI_i (positive = CT reads higher) m_i = (CT_i + MRI_i) / 2 Step 2 — Summary statistics of differences Mean bias (d̄) = +13.2 mL (CT reads 13.2 mL higher on average) SD of differences = 58.4 mL n = 60 Step 3 — 95% Limits of Agreement Upper LoA = 13.2 + 1.96 × 58.4 = 13.2 + 114.5 = +127.7 mL Lower LoA = 13.2 − 1.96 × 58.4 = 13.2 − 114.5 = −101.3 mL Step 4 — 95% CI for mean bias SE_mean = 58.4 / √60 = 7.54 mL 95% CI: 13.2 ± 1.96 × 7.54 = +0.4 to +26.0 mL Step 5 — Proportional bias test Pearson r (differences vs means) = 0.08, p = 0.54 → no proportional bias
+13.2 mLMean bias (95% CI: 0.4–26.0)
+127.7 mLUpper LoA
−101.3 mLLower LoA
Limits of Agreement: −101.3 to +127.7 mL (MAD pre-specified: ±150 mL) — Methods are interchangeable
Interpretation: Both LoA (−101.3 and +127.7 mL) fall within the pre-specified ±150 mL maximum acceptable difference for liver transplant planning, where a volume error of up to 150 mL is deemed clinically tolerable given the surgical margin. The mean bias of +13.2 mL (CT reads 13.2 mL higher) is statistically significant (95% CI excludes zero) but clinically negligible as a percentage of mean liver volume (~1%). MRI can be accepted as a radiation-free alternative to CT for liver volumetry in this centre's protocol. Note: The conclusion depends entirely on the ±150 mL threshold, which must be pre-specified and justified — not selected after seeing the results.
2
Laboratory Medicine: Point-of-Care vs Laboratory HbA1c
A diabetes outpatient clinic evaluates whether a point-of-care (POC) HbA1c analyser can replace laboratory HPLC measurement for routine monitoring, avoiding the 3-day laboratory turnaround. Ninety-five consecutive patients with type 2 diabetes have blood drawn simultaneously for both methods. The maximum acceptable difference is pre-specified as ±0.5% HbA1c (a clinically meaningful threshold around the 7.0% treatment target).
95
Subjects (n)
7.8%
Mean HPLC HbA1c
7.6%
Mean POC HbA1c
±0.5%
Pre-specified MAD
Bland-Altman statistics Mean bias (d̄) = HPLC − POC = +0.24% (HPLC reads higher) SD of differences = 0.38% 95% Limits of Agreement Upper LoA = 0.24 + 1.96 × 0.38 = +0.98% Lower LoA = 0.24 − 1.96 × 0.38 = −0.50% Proportional bias test r(differences vs means) = 0.31, p = 0.002 → significant proportional bias detected Regression: d_i = −1.84 + 0.27 × m_i (differences grow with HbA1c level) Implication at key thresholds At HbA1c = 6.5% (diagnosis threshold): predicted bias = 0.12% — acceptable At HbA1c = 10%: predicted bias = 0.86% — exceeds ±0.5% MAD
+0.24%Mean bias
+0.98%Upper LoA ❌ exceeds MAD
−0.50%Lower LoA (borderline)
Upper LoA (+0.98%) exceeds the ±0.5% MAD — Methods are NOT interchangeable across the full range
Interpretation: The upper LoA of +0.98% exceeds the pre-specified ±0.5% threshold, and proportional bias analysis reveals the divergence is worse at high HbA1c values — precisely where clinical decisions about insulin intensification are made. The POC device may be acceptable for monitoring well-controlled patients (HbA1c < 8.0%) but should not replace laboratory HPLC for patients with poor glycaemic control, where differences of nearly 1% could lead to inappropriate treatment decisions. This example demonstrates how Bland-Altman analysis can reveal a method's valid range of use even when the overall agreement appears reasonable at first glance.
3
Device Comparison: Wearable vs Sphygmomanometer Blood Pressure
A cardiovascular research group validates a new wrist-worn wearable blood pressure device against a validated upper-arm oscillometric sphygmomanometer in 85 hypertensive outpatients. Both devices measure systolic blood pressure (SBP) three times; the mean of three readings is used for each method. The IEEE/AAMI validation standard requires LoA within ±15 mmHg and mean bias within ±5 mmHg.
85
Subjects (n)
145.2
Mean wearable SBP (mmHg)
142.6
Mean sphygmo SBP (mmHg)
±15 mmHg
IEEE/AAMI MAD
Bland-Altman statistics Mean bias = Wearable − Sphygmo = +2.6 mmHg (within ±5 mmHg criterion) SD of differences = 9.1 mmHg 95% Limits of Agreement Upper LoA = 2.6 + 1.96 × 9.1 = +20.4 mmHg ← exceeds ±15 mmHg criterion Lower LoA = 2.6 − 1.96 × 9.1 = −15.2 mmHg ← borderline 95% CI for LoA (Bland & Altman formula) CI width for LoA = ±t₀.₀₅ × 9.1 × √(3/85) = ±1.99 × 9.1 × 0.188 = ±3.40 Upper LoA 95% CI: 20.4 ± 3.4 = [17.0, 23.8] — entire CI outside ±15 criterion Lower LoA 95% CI: −15.2 ± 3.4 = [−18.6, −11.8] Proportional bias r(differences vs means) = 0.12, p = 0.28 → no proportional bias
+2.6 mmHgMean bias ✓ within ±5
+20.4 mmHgUpper LoA ❌ exceeds ±15
−15.2 mmHgLower LoA (borderline)
Upper LoA (+20.4 mmHg) exceeds IEEE/AAMI ±15 mmHg standard — Device does NOT meet validation criteria
Interpretation: Despite an acceptable mean bias of +2.6 mmHg, the wearable device fails the IEEE/AAMI validation standard because the upper LoA of +20.4 mmHg indicates that for some patients the wearable will read up to 20 mmHg higher than the reference standard. This is a critical finding: a single measurement could overestimate a patient's blood pressure by 20 mmHg, potentially triggering unnecessary antihypertensive escalation. The device's mean accuracy is good — the problem is the spread of individual errors, which is too large for clinical use. This example underscores why the LoA, not the mean bias, is the primary criterion for method comparison.

✓ When Two Methods CAN Be Used Interchangeably

Both the upper AND lower Limits of Agreement fall within the pre-specified maximum acceptable difference. The mean bias is clinically negligible or within an accepted standard. No systematic proportional bias is detected across the measurement range.

✗ When Two Methods CANNOT Be Used Interchangeably

Either or both LoA exceed the pre-specified acceptable range. Significant proportional bias is present — methods agree at some values but diverge dangerously at others. Mean bias is acceptable but individual variation (SD of differences) is too large.

Extensions of the Bland-Altman Method

Repeated Measurements Per Subject

Standard Bland-Altman assumes one measurement per subject per method. When multiple measurements per subject are available (e.g., three readings by each method), the simple approach of randomly selecting one pair per subject wastes data. Bland and Altman (1999) described an extension for repeated measurements that uses within-subject means and variance components. The R package MethComp implements this directly with the BA.est() function. SPSS does not natively support this extension; a mixed-model approach is required.

Comparing Three or More Methods

When three methods are compared (e.g., CT, MRI, and ultrasound for the same measurement), separate pairwise Bland-Altman analyses are conducted for each pair of methods. There is no direct three-method extension of the Bland-Altman plot — each comparison remains pairwise. Adjust for multiple comparisons if the primary endpoint is whether any pair meets the acceptability criterion.

Log Transformation for Ratio Data

For measurements that span orders of magnitude (e.g., serum troponin, viral load, tumour marker levels), the proportional nature of measurement error is expected. Log-transform both measurements before Bland-Altman analysis. The differences on the log scale represent log-ratios, and the LoA on the log scale, when back-transformed, give the ratio Limits of Agreement: the new method is expected to read between X times lower and Y times higher than the reference.

Thesis and Research Reporting Recommendations

Bland-Altman analysis is a complete statistical method that requires specific elements in every section of a manuscript or thesis chapter:

Methods Section

State: (1) the study design (cross-sectional simultaneous comparison, paired test-retest, or crossover); (2) whether each method was applied by blinded assessors; (3) the order of method application and whether order was randomised; (4) the pre-specified maximum acceptable difference (MAD) with justification; and (5) whether proportional bias will be assessed and how. The MAD must be stated in the Methods section — defining it after seeing the results is an unacceptable form of outcome reporting bias.

Results Section

Model Reporting Paragraph
"Bland-Altman analysis was used to assess agreement between POC HbA1c and laboratory HPLC measurements. The mean bias was +0.24% (95% CI: +0.16 to +0.32%), indicating that the POC device reads systematically lower than HPLC. The 95% Limits of Agreement were −0.50% to +0.98% (95% CI for upper LoA: +0.83 to +1.13%; 95% CI for lower LoA: −0.65 to −0.35%). The upper LoA of +0.98% exceeded the pre-specified maximum acceptable difference of ±0.5%. Proportional bias was identified (r = 0.31 between differences and means, p = 0.002), with differences increasing by 0.27% for each 1% increase in mean HbA1c. On this basis, the POC device cannot be considered a suitable replacement for laboratory HPLC across the full HbA1c range in this patient population."

Always include the Bland-Altman plot as a figure with: individual data points, horizontal reference line at zero, horizontal line at the mean bias, dashed lines at both LoA, and optional 95% CI shaded bands around the LoA. If proportional bias was detected, add the regression line through the scatter.

Discussion Section

Interpret the clinical meaning of the LoA in terms that matter to clinicians. State clearly whether the two methods are interchangeable for their intended purpose, and if not, describe what modifications to the measurement protocol or instrument calibration would be required to achieve interchangeability.

Common Mistakes Researchers Make

Mistake 1: Using Pearson's r as a Method Comparison Statistic

Despite four decades of guidance from Bland and Altman, studies continue to report Pearson's r (or Spearman's ρ) as the primary evidence of method agreement. Correlation cannot detect systematic bias, is inflated by sample heterogeneity, and provides no clinically interpretable information about the magnitude of disagreement.

Fix: Replace Pearson's r with Bland-Altman analysis for any method comparison study. Pearson's r may be reported as supplementary information but must never be the primary agreement statistic.

Mistake 2: Defining the Maximum Acceptable Difference After Seeing Results

Post-hoc MAD selection — choosing a threshold that the observed LoA happens to meet — is one of the most common and consequential forms of publication bias in method comparison research. It produces false validation of inadequate measurement tools.

Fix: Pre-specify the MAD in the study protocol, registered on a clinical trials registry where possible. Justify the MAD from clinical guidelines, biological variation databases (e.g., Westgard's desirable specifications), or published validation standards for the specific measurement type.

Mistake 3: Ignoring Proportional Bias

Many Bland-Altman studies report the LoA without checking for proportional bias. When proportional bias is present (error increases or decreases across the measurement range), the constant LoA from standard analysis are misleading — they are too wide at one end of the range and too narrow at the other.

Fix: Always include a regression of differences on means in your analysis and report the Pearson correlation coefficient and p-value. If proportional bias is confirmed, use regression-based LoA or log-transformed analysis rather than fixed LoA.

Mistake 4: Using the Paired t-Test Alone to Confirm Agreement

A non-significant paired t-test on the differences (p > 0.05) is sometimes cited as evidence that two methods agree. This only confirms that the mean bias is not distinguishable from zero — it gives no information about the spread of individual differences, which is what actually determines interchangeability in practice.

Fix: Report the paired t-test result as additional information about the mean bias, but make the LoA the primary analysis. Emphasise that a non-significant t-test combined with wide LoA means the two methods are not interchangeable despite similar means.

Mistake 5: Reporting the Plot Without Confidence Intervals on the LoA

Many published Bland-Altman plots show only the mean bias and LoA lines without confidence intervals. In studies with fewer than 100 subjects, the uncertainty around the LoA can be substantial — a lower LoA of −9.2 mmHg might have a 95% CI of −13.5 to −4.9, spanning both "within criterion" and "outside criterion."

Fix: Always compute and report the 95% CIs for both LoA (formula: LoA ± t × SD_d × √(3/n)). Add these as error bars or shaded bands in the Bland-Altman figure. The decision about interchangeability should use the outer bound of the CI rather than the point estimate.

Mistake 6: Using Multiple Measurements Per Subject Without Adjustment

When subjects contribute more than one measurement pair to the analysis (e.g., three readings per device), treating all pairs as independent violates the independence assumption, artificially narrows the SD of differences, and produces LoA that are too narrow — falsely suggesting better agreement than actually exists.

Fix: Use the extended Bland-Altman method for repeated measurements (Bland & Altman 1999) or use only one randomly selected measurement pair per subject. The R package MethComp handles the extended method automatically.

Mistake 7: Plotting Method A vs Method B Instead of Differences vs Means

Some researchers plot Method A on the x-axis and Method B on the y-axis (a standard scatterplot) and call it a Bland-Altman analysis. This is not Bland-Altman — it is a scatterplot. The power of Bland-Altman lies specifically in plotting differences against means, which makes systematic patterns, proportional bias, and outliers visually apparent in a way that a scatterplot cannot.

Fix: The y-axis must always be the difference (A − B) and the x-axis must be the mean ((A + B) / 2). Only then does the Bland-Altman analysis deliver its diagnostic value. Software such as SPSS, R (blandr or MethComp packages), or GraphPad Prism can produce the correct plot automatically.

Scientific Reporting Standards

The following guidelines govern Bland-Altman reporting in high-quality clinical journals and thesis submissions:

Practical Recommendations for Medical Researchers

Define Your MAD From Literature, Not Your Data

Search for published clinical guidelines, validation standards, or biological variation data for your specific analyte or measurement. For HbA1c, consult the IFCC biological variation database. For blood pressure, consult IEEE/AAMI SP10. For imaging measurements, check the applicable RECIST or organ-specific guidelines. Your MAD should be evidence-based and documented before any data are collected.

Test Normality of Differences Before Applying Standard LoA

The 1.96 SD Limits of Agreement assume the differences are normally distributed. Test this with a Shapiro-Wilk test or a QQ plot of the differences. If the differences are right-skewed or bimodal, the LoA will not correctly capture the 95th percentile. In this case, use log transformation or bootstrapped LoA from the empirical distribution.

Randomise the Order of Method Application

If the order of measurement matters (patient fatigue, learning effects, environmental changes), randomise whether Method A or B is applied first in each subject. This prevents order effects from masquerading as systematic bias between methods. Document the randomisation procedure in your methods section.

Aim for at Least 50 Subjects (Ideally 100)

With n < 30, the 95% CIs around the LoA are very wide and the analysis is too imprecise for clinical validation decisions. Bland and Altman originally recommended approximately 100 subjects. For regulatory submissions (device CE marking, FDA clearance), standards typically require 85–100 subjects minimum.

Include a Clinical Context Statement

Always interpret the LoA in clinical terms — not just in statistical terms. "The 95% Limits of Agreement were −15.2 to +20.4 mmHg" should be followed by "This means the wearable device could read up to 20 mmHg higher than the reference in any given patient, which could lead to unnecessary antihypertensive treatment in borderline hypertensive individuals."

Report Both ICC and Bland-Altman for Complete Characterisation

ICC tells you the proportion of variance explained by true subject differences (a dimensionless reliability coefficient). Bland-Altman tells you the actual magnitude of measurement errors in clinical units. They answer different questions and together give a complete picture of measurement performance. Report both for any continuous reliability or agreement study.

Frequently Asked Questions

What is Bland-Altman analysis? +
Bland-Altman analysis is a graphical and statistical method for comparing two quantitative measurement techniques applied to the same subjects. Introduced by Bland and Altman in 1986, it plots the difference between two methods against their mean, then calculates the 95% Limits of Agreement (mean difference ± 1.96 × SD of differences). The primary purpose is to determine whether two methods are clinically interchangeable — a question that correlation analysis cannot answer. It is among the most cited statistical methods in biomedical research.
Why can't I use Pearson's correlation for method comparison? +
Pearson's r measures whether two variables rank subjects in the same order — it detects linear association, not agreement. Two methods that consistently differ by 15 units will show r = 1.00 despite being clinically useless as interchangeable tools. Pearson's r is also inflated by sample heterogeneity (wide range of values) and provides no information about the magnitude of disagreement in clinically meaningful units. Bland-Altman analysis directly quantifies how large the differences between methods are and whether they are clinically tolerable.
What are Limits of Agreement? +
The 95% Limits of Agreement (LoA) define the interval expected to contain 95% of all future pairwise differences between two methods. Calculated as mean difference ± 1.96 × SD of differences, they tell clinicians: "In 95% of individual measurements, Method A will read between X and Y units relative to Method B." For the two methods to be clinically interchangeable, both the upper and lower LoA must fall within the pre-specified maximum acceptable difference. If either LoA exceeds the acceptable range, the methods are not interchangeable.
How do I decide whether the LoA are acceptable? +
There is no statistical threshold for acceptability — it is a clinical judgement that must be pre-specified before data collection. Before your study, define the maximum acceptable difference (MAD) for your specific measurement context based on clinical guidelines, validation standards (e.g., IEEE/AAMI for blood pressure: ±15 mmHg), or biological variation databases. If both LoA fall within ±MAD, the methods are interchangeable. If either LoA exceeds ±MAD, they are not. Never define the MAD after seeing the results — this is outcome reporting bias.
What is proportional bias and how do I detect it? +
Proportional bias occurs when the disagreement between two methods is not constant but increases (or decreases) with the magnitude of the measurement. It is detected by computing the Pearson correlation between the differences (y-axis values) and the means (x-axis values) in the Bland-Altman plot. A statistically significant correlation (p < 0.05) confirms proportional bias. When present, the standard fixed Limits of Agreement are invalid and regression-based LoA or log transformation should be used instead.
How many subjects do I need? +
Bland and Altman originally recommended approximately 100 subjects for a reliable study. A minimum of 30–50 subjects is sometimes used for preliminary studies, but the 95% CIs around the LoA are very wide below 50 subjects. Regulatory device validation standards (IEEE/AAMI, CLSI EP09) typically require 85–100 subjects. Formal sample size can be calculated from the expected SD of differences and the desired CI width around the LoA using published formulas.
What should be on the x-axis of the plot? +
The x-axis shows the mean of the two methods for each subject: (Method A + Method B) / 2. Bland and Altman recommended using the mean because it is the best estimate of the true value when neither method is a gold standard. If one method is a well-established reference standard, it is also acceptable to plot the reference method on the x-axis, which makes interpretation more intuitive. Always state which was used in the figure legend and methods section.
What is the difference between Bland-Altman analysis and ICC? +
ICC is a dimensionless reliability coefficient (0–1) expressing how much of total variance is signal vs noise — it answers "how reproducible is this measurement?" Bland-Altman analysis quantifies the actual magnitude of disagreement in clinical units and answers "are these two methods clinically interchangeable?" For a complete method comparison study, many researchers report both: ICC characterises the degree of reliability, while Bland-Altman shows the clinical magnitude and pattern of errors.
How do I report Bland-Altman in a thesis or paper? +
Report: (1) mean bias with 95% CI; (2) SD of differences; (3) both 95% Limits of Agreement with 95% CIs; (4) the pre-specified MAD and whether the LoA meet it; (5) proportional bias assessment (r between differences and means, p-value); (6) the Bland-Altman plot as a figure with mean bias and LoA lines labelled. Conclude explicitly whether the two methods are or are not clinically interchangeable based on the pre-specified criterion.
Can I use Bland-Altman with repeated measurements per subject? +
Standard Bland-Altman assumes one measurement pair per subject. Using multiple pairs per subject without adjustment violates independence and artificially narrows the LoA. For repeated measurements, use the extended Bland-Altman method described by Bland and Altman (1999) using within-subject variance components, implemented in the R package MethComp. Alternatively, use the mean of multiple measurements per subject as a single data point — though this reduces the apparent SD of differences and may not reflect single-measurement clinical use.

Run Bland-Altman Analysis Online

Use StatClinic's Bland-Altman calculator to compute mean bias, Limits of Agreement, and 95% confidence intervals from your paired measurements. Get a publication-ready plot and complete numerical output formatted for thesis and journal submission.

Open Bland-Altman Calculator →