Are repeated-measures designs always more powerful?

No; benefits depend on covariance, effect patterns, measurement quality and study burden.

Is compound symmetry required for every analysis?

No; alternative covariance structures and methods are available when supported by the design and data.

Does having equal variances at each occasion establish sphericity?

No; the variances of pairwise differences also depend on covariances, as the worked example demonstrates. With only two levels, the usual sphericity restriction is automatically satisfied.

AI Glossary / Statistics and probability / Intermediate

Repeatedmeasures designs

Definition

Repeated-measures designs collect several responses from each unit, such as measurements of the same participant across conditions or observations of the same machine over time. The design creates a structured set of linked observations rather than a collection of unrelated rows. This term names the same broad design family as repeated measures design. The plural wording does not imply a separate method; this entry emphasizes how within-unit covariance affects planning and analysis.

What it means

Repeated observations can help separate stable differences between units from changes within units. Their value depends partly on the pattern of association among measurements. Nearby time points may be strongly correlated, while distant ones may be less similar. Responses under different conditions can have another pattern entirely. Consequently, counting measurements is not enough to determine precision. The analysis must connect the scientific comparison to the units, conditions and covariance structure that generated the data.

How it works

Specify the units and within-unit factors, distinguish them from any between-unit treatment factors, and decide which contrasts answer the research question. Plan order, timing and missing-data handling. Preserve the repeated structure in the dataset. For analysis, use a method whose assumptions match the design, such as a mixed model with a defensible covariance pattern or a traditional repeated-measures ANOVA when its conditions are appropriate. Assess whether a simpler or richer covariance model is supported by the available number of units.

Three practical examples

Example 1: Participants rate three AI-generated explanations, with presentation order varied across participants. Example 2: Engineers measure output from each device at three temperatures, allowing for device-specific differences. Example 3: A team follows service quality within the same offices over several months. In each example, the repeated factor differs, and the appropriate covariance assumptions may differ too. Treating all these datasets as identical merely because each has three columns can produce misleading uncertainty estimates.

Think of it this way

Think of several musical notes played on the same instrument. They share characteristics of that instrument, even when the notes differ. Hearing more notes reveals useful information, but it does not create more independent instruments. The statistical task is to learn from both the shared characteristics and the differences without counting the shared information repeatedly as if it came from unrelated sources.

Worked example

Consider three measurements from one unit, each with variance four. Suppose adjacent measurements have covariance two and the first and third have covariance one. The covariance matrix has diagonal entries 4,4,4 and off-diagonal pairs 2,2,1. The variance of that unit's three-measurement average is the sum of all matrix entries divided by nine: (12 + 2 times 5)/9 = 22/9, approximately 2.444. Incorrectly assuming independence would give 12/9 = 1.333. The same structure gives variance four for an adjacent difference, because 4 + 4 - 2 times 2 = 4, but variance six for the first-to-third difference. These unequal difference variances show that sphericity does not hold for this stipulated three-occasion structure. The example illustrates covariance consequences; it does not estimate a covariance matrix from three observed values.

Illustrative case study

In this fictional case study, Riverbend Analytics compares three explanation styles for an AI assistant. Every participant reviews all styles, with order balanced across the study. Analysts retain participant identifiers and inspect the pattern of repeated responses before selecting a model. They report planned comparisons with uncertainty and consider multiplicity rather than searching every possible contrast for a favorable result. A sensitivity analysis examines whether reasonable covariance choices change the interpretation. The report distinguishes participant variation from repeated response variation.

Common misconceptions and mistakes

Common mistakes include assuming all pairwise correlations must be equal in every repeated-measures design, interpreting sphericity as merely equal individual variances, and ignoring covariance when averaging measurements. Compound symmetry is a stronger condition that can imply sphericity, but the concepts are not identical. A correction to an ANOVA reference distribution does not repair a poorly designed experiment or informative dropout. Mixed models relax some traditional restrictions but still depend on assumptions about the response, covariance and missingness.

Business use

Businesses use these designs when units are scarce, within-unit comparisons are informative or change itself is the outcome of interest. Planning should balance the number of independent units against the burden of additional observations per unit. AI usability studies need to account for participant learning and fatigue; device studies need to consider drift and carryover. Sample-size planning should reflect the intended contrast and plausible correlations, with sensitivity to uncertain assumptions rather than a row-count target alone.

Check your understanding

What is the variance of the three-measurement average in the example?

Answer: 22/9, approximately 2.444.

Why does the independence calculation underestimate it?

Answer: It omits positive covariances.

Are all pairwise difference variances equal?

Answer: No; adjacent differences have variance four, while the first-to-third difference has variance six.

Does this example provide evidence from a real study?

Answer: No; it is a stipulated covariance model used to explain the calculation.

Related terms

Repeated measures design
Covariance matrix
Compound symmetry
Sphericity
Related terms include within-subject design
autoregressive correlation
mixed-effects model and planned contrast

Learn next

Learn next how covariance affects sums, averages and differences. Compare compound symmetry with an autoregressive pattern, then study how planned contrasts target specific questions. Explore the assumptions behind univariate repeated-measures ANOVA and alternatives for incomplete or unequally spaced observations. Always connect the selected statistical model to the actual experimental unit and schedule.

Sources

Sources: University of Vermont, Statistical Methods for Psychology glossary, repeated-measures designs, https://www.uvm.edu/~statdhtx/methods8/Glossary/Glossary.html; Penn State STAT 502, Introduction to Repeated Measures, https://online.stat.psu.edu/stat502/Lesson11; Penn State STAT 505, Repeated Measures Analysis, https://online.stat.psu.edu/stat505/Lesson09. The covariance calculations and fictional explanation study are original illustrations of the explicitly stated design assumptions.

Previous: Repeated survey Next: Replacing eligibility trace