Are variance components always directly observable?

No; they are estimated model parameters.

Do repeated observations within a group share covariance?

In a random-intercept model, yes.

Should standard deviations simply be added to obtain total spread?

No; independent variance components add, and the square root gives total standard deviation.

AI Glossary / Statistics and probability / Intermediate

Variance components

Definition

Variance components are parameters that represent distinct sources of random variation in a statistical model. In a simple random-intercept model, total outcome variance can be separated into variation between groups and variation within groups. More complex models can include several random factors and their covariance structure. The components are model-based quantities, not automatically causal explanations of every observed difference, and their interpretation depends on how the observations and groups were sampled.

What it means

Repeated measurements from the same group often resemble one another because they share conditions or a latent group effect. A variance-components model separates that shared variation from residual variation among observations within the group. This helps explain dependence and plan data collection. It can also show why collecting many measurements from only a few groups may provide less information about a population mean than the same number spread across more independent groups. The decomposition depends on the assumed random-effects structure.

How it works

For a one-way model, write Yij = mu + bi + eij. Assume group effects bi have mean zero and variance tau squared, residuals eij have mean zero and variance sigma squared, and all distinct group effects and residuals are mutually independent. Then an individual outcome has variance tau squared plus sigma squared, and two outcomes in the same group have covariance tau squared. In a balanced design with m observations per group, expected within-group mean square is sigma squared and expected between-group mean square is sigma squared plus m times tau squared. Other designs commonly use likelihood-based methods such as restricted maximum likelihood.

Three practical examples

Example 1: A measurement study separates variation between machines from variation among repeated readings on a machine. Example 2: An AI evaluation separates differences among tasks from variation across repeated runs on the same task. Example 3: A service study includes organization and team random effects. These models require a design that can distinguish the proposed sources. If factors are completely confounded, a sophisticated fitting method cannot manufacture the missing information.

Think of it this way

Think of comparing several gardens, each containing several plants. Some variation occurs because gardens differ, and some because plants within the same garden differ. Repeated plants help describe the within-garden component, while multiple gardens are needed to learn about the between-garden component. The labels organize variation under a model; they do not automatically reveal which physical cause produced it.

Worked example

Suppose a balanced random-intercept study has five observations in each of several independent groups. Its ANOVA between-group mean square is 24 and within-group mean square is four. The method-of-moments residual variance estimate is four, and the between-group variance estimate is (24 minus 4)/5 = 4. The estimated variance of an individual observation is eight. The intraclass correlation is 4/(4 + 4) = 0.5. For a mean of five observations within one randomly selected group, the variance is four plus four divided by five, equal to 4.8, because the shared group effect does not average away within that group. These calculations assume the stated balanced independent-component model; they do not supply confidence intervals for the estimated components.

Illustrative case study

In this fictional case study, Magnolia Evaluation compares an AI model across benchmark tasks with repeated runs per task. Analysts find substantial variation between tasks as well as run-to-run variation. They explain that adding many more runs to the existing tasks reduces only part of the uncertainty about broader task performance. The team therefore considers sampling additional tasks. Its report defines the task population and avoids treating the estimated task variance as a complete measure of every deployment difference.

Common misconceptions and mistakes

Common mistakes include adding component standard deviations instead of variances, treating random effects as observed causes, and assuming repeated observations are independent. Another mistake is interpreting a negative method-of-moments estimate as a physically negative variance. Such estimates can arise from sampling variation; constrained methods or explicit boundary handling are needed. Estimates near zero can still be uncertain. Random slopes and correlated random effects require covariance terms, so the simple sum of two independent components does not cover every mixed model.

Business use

Businesses use variance components to improve measurement systems, allocate replication and understand clustered data. Ask whether the design includes enough independent groups and whether the model separates the sources of interest. Distinguish the variance of an individual outcome from the variance of an average or prediction. Report uncertainty and the assumed grouping structure. A component that is large relative to another may guide investigation, but deciding how to reduce it usually requires additional operational or experimental evidence.

Check your understanding

What is the estimated between-group variance in the example?

Answer: Four.

What is the estimated individual-observation variance?

Answer: Eight.

What is the intraclass correlation?

Answer: 0.5.

Why is the variance of a five-observation group mean 4.8 rather than 8/5?

Answer: The shared group effect remains common to all five observations and does not average down like independent residual noise.

Related terms

Related terms include random effects
mixed model
intraclass correlation
restricted maximum likelihood
nested design and covariance

Learn next

Learn next how expected mean squares lead to simple component estimates and how restricted maximum likelihood handles more complex designs. Compare nested and crossed factors, then study how random slopes change covariance between observations. Use these ideas to plan replication at the level that matters for the target estimate.

Sources

Sources: NIST Dataplot, One Way ANOVA, https://itl.nist.gov/div898/software/dataplot/refman1/auxillar/onewayan.htm; NIST/SEMATECH, The one-way ANOVA model and assumptions, https://www.itl.nist.gov/div898/handbook/prc/section4/prc422.htm. These primary references describe random effects and component estimation. The balanced mean-square, correlation and group-mean calculations are original and independently checked; the AI evaluation case is fictional.

Previous: Variance Next: Variance estimation