Are random effects simply random errors?

No. A group effect is shared by observations in that group, while residual errors describe additional observation-level variation.

Can fixed and random effects coexist?

Yes; mixed models include both.

Does more data always estimate group variance well?

Many observations in very few groups may still provide limited information about between-group variability.

AI Glossary / Statistics and probability / Intermediate

Random model Anova

Definition

A random model ANOVA, more commonly called a random-effects analysis of variance, treats the effects of factor levels as draws from a population of possible effects. Its central purpose often includes estimating how much variability arises between levels, such as batches or operators, rather than only comparing the particular levels observed. The supplied label is retained, while the standard random-effects terminology is used here. Random effects describe a statistical model; they are not synonymous with random assignment of treatments.

What it means

In a fixed-effects analysis, interest may center on the specific treatments or settings included in the study. In a random-effects analysis, the observed levels represent a broader population, and their variability is itself an object of interest.

For example, a quality team may sample production batches to understand how much future output varies from batch to batch. The model separates variation shared by observations from the same batch from variation among observations within a batch. This separation also explains why readings from one batch are not generally independent when considered without conditioning on that batch's effect.

How it works

A simple one-way random-effects model is Y_ij = mu + a_i + e_ij. Here mu is the overall population mean, a_i is the random effect for level i, and e_ij is the residual for observation j within that level. The usual introductory model assumes independent, mean-zero level effects with variance sigma_a squared and independent, mean-zero residuals with variance sigma_e squared, with the two sources independent of each other.

In a balanced design with m observations per level, the expected within-level mean square is sigma_e squared. The expected between-level mean square is sigma_e squared + m times sigma_a squared. Equating these expectations to observed mean squares gives classical method-of-moments estimates.

Normality assumptions support familiar exact F procedures in the simple model. More complex designs need the correct expected mean squares or a suitable mixed-model analysis. Restricted maximum likelihood is another common estimation approach. Unequal group sizes, crossed factors, nested factors, and unequal residual variances require additional care rather than automatic reuse of the balanced formula.

Three practical examples

1. A manufacturer samples batches to estimate the contribution of batch differences to total product variation.

2. A laboratory samples operators from a relevant workforce and studies how much measurement variability is associated with operator differences.

3. An education study models variation among sampled schools while recognizing that observations from the same school share a common component.

Think of it this way

Think of several bowls of soup drawn from different cooking batches. Tasting several spoonfuls from each bowl helps distinguish variation within a bowl from differences between bowls. Taking many spoonfuls from one bowl cannot reveal how variable the entire population of batches is.

Worked example

Consider an original balanced teaching example with five sampled batches and three observations per batch. Suppose the ANOVA summaries give between-batch mean square MS_B = 18 and within-batch mean square MS_W = 6. These are stipulated summaries, not reported measurements from a real process.

The estimated residual variance is 6. The classical estimate of batch variance is (18 - 6)/3 = 4. Estimated total variance for one observation is therefore 4 + 6 = 10. The proportion attributed to the batch component is 4/10 = 0.40, or 40%.

Under this model, the intraclass correlation for two observations from the same batch is also 0.40: their shared covariance is the batch variance, while each has total variance 10. The conventional F statistic for testing zero batch variance in this simple balanced normal model is 18/6 = 3, with 4 and 10 degrees of freedom. No significance conclusion is asserted without the chosen reference calculation.

The variance of a three-observation batch average is 4 + 6/3 = 6. Averaging reduces the residual contribution, but the shared batch effect remains. This is why more readings within the same batch are not equivalent to sampling additional independent batches.

Illustrative case study

In a fictional food-packaging company, a team collects several measurements from each of a small number of production batches. An initial report treats every measurement as an independent representation of future production.

A statistician fits a batch random-effects model and explains the shared source of variation. The team expands the number of sampled batches and retains enough within-batch readings to estimate repeatability. Management uses the variance components to decide whether process changes should target batch consistency or measurement-level variability. The model does not identify the physical cause of either component without further investigation.

Common misconceptions and mistakes

Do not infer that a factor is random solely because treatments were randomly assigned. The modeling question concerns what the levels represent and which population the inference targets. Nor does declaring an effect random make a convenience sample representative of all possible levels.

A negative unconstrained method-of-moments variance estimate can occur when the between-level mean square is smaller than the within-level mean square. A population variance cannot be negative; boundary handling or an alternative estimation method must be documented. Few sampled levels can leave variance components highly uncertain even when many observations are collected within each level. More complicated ANOVA designs can require different test denominators.

Business use

Random-effects ANOVA helps organizations identify where variability sits and plan efficient measurement. It supports repeatability studies, supplier monitoring, and forecasts involving new groups. Report the population of levels, sampling process, variance estimates, and their uncertainty. Treat variance components as statistical descriptions that guide investigation, not automatic explanations of why a process varies.

Check your understanding

In a balanced design with four readings per level, MS_B = 22 and MS_W = 10. What is the classical level-variance estimate?

Answer: It is (22 - 10)/4 = 3, under the stated one-way model.

Related terms

Variance components
ANOVA
random effect
fixed effect
mixed model
variance component
intraclass correlation
nested design
restricted maximum likelihood

Learn next

Learn one-way ANOVA and expected mean squares, then study variance components and mixed models. Compare adding more groups with adding more observations within existing groups.

Sources

NIST/SEMATECH, What are variance components?: https://itl.nist.gov/div898/handbook/prc/section4/prc44.htm

David Howell, Statistical Methods for Psychology glossary: https://www.uvm.edu/~statdhtx/methods8/Glossary/Glossary.html

Both primary educational sources were read. The balanced numerical example and fictional case are original.

Previous: Random marginal totals Next: Random network distillation