Does compound symmetry imply sphericity?

Yes, but sphericity does not require compound symmetry.

Is a sphericity test needed for a two-level within-subject effect?

No, because there is only one difference variance.

Do degree-of-freedom corrections change the raw observations?

No; they adjust the reference inference.

AI Glossary / Statistics and probability / Advanced

Sphericity

Definition

Sphericity is a covariance condition used in the conventional univariate analysis of repeated measures. For a single within-subject factor, it means that the population variances of the differences between every pair of repeated conditions are equal. It concerns the variability of within-person differences, not equality of condition means. Sphericity is less restrictive than compound symmetry, which requires equal variances and equal pairwise covariances among the repeated measurements.

What it means

Repeated observations on the same person are related. The uncertainty in a comparison depends on both the variability at each occasion and the covariance between them. If every pairwise difference has the same variance, a common error structure supports the usual uncorrected repeated-measures F test under its other assumptions. If some differences are much more variable than others, that reference test can have an incorrect false-positive rate. The issue is about the covariance model, not whether the observed average profile looks smooth.

How it works

Identify the within-subject factor and the particular effect being tested. Examine the covariance structure and, where relevant, a sphericity diagnostic such as Mauchly's test. A nonsignificant diagnostic does not prove the population condition. For suitable univariate analyses, Greenhouse-Geisser or Huynh-Feldt adjustments multiply the numerator and denominator degrees of freedom by an estimated epsilon before assessing the F statistic. Other approaches include appropriately specified multivariate or mixed models. These alternatives have their own assumptions and sample-size needs rather than providing an automatic remedy.

Three practical examples

Example 1: A training study compares the same participants' scores across three occasions. Example 2: An interface experiment measures completion time for each participant under four display conditions. Example 3: A repeated AI-assisted task study examines a within-person response across several levels of assistance. In each case, differences between conditions can have unequal variances even when the individual condition variances look similar. The sampling unit remains the participant or other independently sampled unit, not each repeated measurement treated in isolation.

Think of it this way

Think of measuring how much each person changes between several pairs of checkpoints. Sphericity asks whether those different kinds of change are equally variable across people. It does not ask whether people have equal starting points or whether the average changes are the same. The comparison concerns the spread of paired differences, which is why looking only at separate condition standard deviations can miss the issue.

Worked example

Consider three repeated measurements with population variances two, three and four. Let their pairwise covariances be 1.5 for conditions one and two, two for one and three, and 2.5 for two and three. Using Var(A minus B) = Var(A) + Var(B) minus twice Cov(A,B), the three difference variances are 2 + 3 minus 3 = 2; 2 + 4 minus 4 = 2; and 3 + 4 minus 5 = 2. This valid positive-definite covariance matrix satisfies sphericity even though its individual variances and covariances are unequal, so it lacks compound symmetry. By contrast, if all individual variances are four and the covariances are one, zero and one for the same pairs, the difference variances are six, eight and six. Equal individual variances alone therefore do not establish sphericity.

Illustrative case study

In this fictional case study, Ashford Learning evaluates the same employees after several training sessions. The analyst initially compares separate score variances and overlooks the covariance between sessions. A review instead examines the within-person difference structure and reports an appropriate corrected repeated-measures analysis alongside estimated mean differences. The team explains the correction and the model assumptions without claiming that the data have been transformed into a spherical population. Managers focus on the size and uncertainty of the training pattern.

Common misconceptions and mistakes

Common mistakes include defining sphericity as equal means, equating it with compound symmetry, and interpreting a nonsignificant Mauchly test as certification. With only two repeated levels there is only one pairwise difference, so this equality condition is automatically satisfied for that effect. More complex within-subject designs require attention to the specific contrasts and interactions. Corrections do not fix missing-data bias, an inappropriate outcome model or a failure to respect independent sampling units. A mixed model also needs a defensible covariance specification.

Business use

Businesses using repeated evaluations should preserve participant identifiers and plan the covariance analysis before looking for favorable significance results. Report the chosen method, relevant degrees of freedom and any epsilon correction, while keeping the substantive effect estimates visible. Repeated data can improve precision, but their dependence cannot be ignored. For irregular timing or incomplete repeated records, a suitable alternative model may better match the design; that decision should reflect the data structure and assumptions rather than a desire to avoid one diagnostic.

Check your understanding

What are the three difference variances in the first example?

Answer: All are two.

Why is that matrix not compound symmetric?

Answer: Its individual variances and pairwise covariances are unequal.

What are the difference variances in the second example?

Answer: Six, eight and six.

What extra information beyond individual variances is needed to compute them?

Answer: The covariance for each pair.

Related terms

Covariance matrix
Compound symmetry
Related terms include repeated-measures ANOVA
Mauchly's test
Greenhouse-Geisser correction
Huynh-Feldt correction and within-subject contrast

Learn next

Learn next how paired-difference variance depends on covariance and how orthonormal contrasts express the sphericity condition. Compare corrected univariate, multivariate and mixed-model analyses under their respective assumptions. Then examine how incomplete observations and irregular measurement times affect the choice of an appropriate repeated-measures model.

Sources

Sources: SAS, Repeated Measures Analysis of Variance, https://support.sas.com/documentation/cdl/en/statug/63347/HTML/default/statug_glm_sect037.htm; LeBlanc, Introduction to Repeated-Measures, https://support.sas.com/resources/papers/proceedings-archive/SUGI94/Sugi-94-237%20LeBlanc.pdf. These primary implementation and methodological materials explain the covariance condition and corrections. The covariance examples are original and independently checked; the case is fictional. Additional accessible guidance: University of Texas at Austin Statistical Services, Repeated Measures ANOVA Using SAS PROC GLM, reproduced with permission by UCLA, https://stats.oarc.ucla.edu/sas/library/sas-libraryrepeated-measures-anova-using-sas-proc-glm/, explains orthonormal-contrast sphericity and epsilon adjustments.

Previous: Spherical contact distribution function Next: Spike-and-slab distribution