Does compound symmetry require normal observations?

The covariance pattern itself does not, although particular estimation or testing methods may assume normality.

Can means change over time?

Yes; mean trends and covariance structure are separate aspects of the model.

Can the common correlation be negative?

Yes, within dimension-dependent bounds, although a simple random-intercept construction does not generate negative common covariance.

AI Glossary / Statistics and probability / Intermediate

Compound symmetry

Definition

Compound symmetry is a covariance structure in which all measurements have the same variance and every distinct pair has the same covariance. It is often used to model repeated observations within a person, machine, or other cluster. Because the variances are equal, the correlations between distinct measurements are also equal. The structure describes variability and dependence; it does not require the measurements to have equal means.

What it means

A covariance matrix places each measurement's variance on its diagonal and pairwise covariances elsewhere. Under compound symmetry, the diagonal contains one repeated value and every off-diagonal position contains another repeated value.
For repeated visits, this means the model treats the association between visits one and two as equal to the association between visits one and five. That can be a useful simplification when shared subject characteristics dominate the dependence. It can be inappropriate when nearby observations are much more alike than distant ones or when variability changes substantially over time.

How it works

Let there be m measurements, each with variance v and common pairwise correlation rho. Their covariance matrix can be written v[(1 - rho)I + rho J], where I is the identity matrix and J contains ones everywhere. Distinct measurements therefore have covariance v times rho.
For a positive-definite matrix, v must be positive and rho must lie strictly between -1/(m - 1) and 1. Boundary values produce singular matrices. Thus a correlation that is valid for two measurements may be invalid as a common correlation for many measurements.
A random-intercept model with independent, equal-variance residual errors generates a common positive covariance: the intercept variance. This is one useful construction, but general compound symmetry also permits some negative correlations that a nonnegative random-intercept variance alone cannot produce.
Fit the mean structure and covariance structure appropriately, inspect residual patterns, and consider plausible alternatives. Compound symmetry implies equal variances of pairwise differences, a repeated-measures sphericity condition. Sphericity is weaker; the two assumptions should not be treated as interchangeable. Estimation and inference also depend on sample size, missing-data assumptions, and the chosen statistical method.

Three practical examples

An analyst models four repeated satisfaction measurements from each customer with a shared within-customer correlation.
A quality team considers compound symmetry for repeated measurements from the same manufactured unit when measurement order has little relevance.
An AI evaluation team accounts for correlated scores collected from the same participant across several tasks, while checking whether the tasks have comparable residual variability.

Think of it this way

Think of several readings sharing a common background influence plus their own fluctuations. The common influence creates a similar connection between every pair. This picture explains a positive-correlation random-intercept construction, but it is only an analogy: the mathematical covariance structure has a wider range of permissible correlations.

Worked example

Consider an original example with three repeated measurements. Let each variance be 16 and each distinct-pair covariance be 4. The covariance matrix has rows (16, 4, 4), (4, 16, 4), and (4, 4, 16). Each pair has correlation 4/sqrt(16 times 16) = 0.25.
The variance of the mean of the three measurements is the sum of all nine matrix entries divided by 9: (3 times 16 + 6 times 4)/9 = 72/9 = 8. Its standard deviation is sqrt(8), approximately 2.828. Treating the readings as independent would instead give variance 16/3, approximately 5.333, understating variability of their average.
For any pair, the difference has variance 16 + 16 - 2 times 4 = 24. All pairs share this difference variance. The matrix eigenvalues are 24, 12, and 12, so it is positive definite.
For three measurements the correlation must exceed -0.5 for positive definiteness. Setting every pairwise correlation to -0.6 would fail, even though -0.6 is a possible correlation for a single pair.

Illustrative case study

In a fictional business, Cedar Service measures employee task performance during four sessions after a training program. A first analysis counts every session as an independent observation. A reviewer points out that repeated scores from the same employee may share stable influences.
The team fits a model that allows session means to differ and considers compound symmetry for within-employee residual dependence. It then examines whether variability grows across sessions and whether adjacent sessions are more strongly related. Those patterns could favor another covariance structure.
The team reports the selected model and sensitivity of the training comparison to reasonable alternatives. This fictional case illustrates model assessment, not evidence that compound symmetry is correct for all training data.

Common misconceptions and mistakes

Equal covariance does not mean zero covariance. Equal variances do not require equal means. A correlation matrix with equal off-diagonal entries is not enough to establish compound symmetry of the covariance matrix when variances differ. Compound symmetry is sufficient for repeated-measures sphericity but not necessary. A random intercept with independent homogeneous residuals represents only the nonnegative common-covariance case. More repeated observations do not automatically compensate for a poor dependence model.

Business use

Compound symmetry gives businesses a parsimonious way to analyze clustered or repeated measurements, using few covariance parameters. It can support comparisons of service quality, learning outcomes, or product performance while recognizing repeated observations from the same unit. Its value depends on whether the simplification is credible. Compare practical conclusions under alternative structures and avoid selecting a model solely because its name is familiar.

Check your understanding

If all variances are 9 and covariances are 3, what is the common correlation?

Answer: It is 1/3.

If adjacent visits correlate 0.8 but distant visits correlate 0.2, is compound symmetry an exact description?

Answer: No.

Why can ignoring positive dependence distort uncertainty?

Answer: The covariance terms contribute to the variance of an average and cannot simply be discarded.

Related terms

Covariance matrix
Correlation
Sphericity
repeated measures
random intercept
mixed-effects model
positive-definite matrix
autoregressive covariance
heterogeneous compound symmetry

Learn next

Learn how variances and covariances determine uncertainty in sums, averages, and contrasts. Then compare compound symmetry with autoregressive and unstructured covariance models. Practice keeping the mean model, covariance assumptions, and distributional assumptions distinct when interpreting repeated-measures analyses.

Sources

David Howell, Statistical Methods for Psychology glossary, Compound symmetry: https://www.uvm.edu/~statdhtx/methods8/Glossary/Glossary.html SAS, Constructing common covariance structures: https://blogs.sas.com/content/iml/2012/11/05/constructing-common-covariance-structures.html SAS/STAT documentation, RANDOM statement, covariance structure definitions; indexed excerpt consulted: https://support.sas.com/documentation/cdl/en/statug/66859/HTML/default/statug_hpmixed_syntax13.htm

Previous: Compound Events Next: Compound-term processing