MerQuR+Statistics Glossary

Statistics Glossary → Variance components

Variance componentsconcept

The shares into which the total variability of a measurement is divided among different sources modelled as random effects (e.g. subjects, raters, groups, error).

In a random-effects model each source is assumed to come from a distribution with mean zero and variance σ². In a study of schools, for example, total variance is split into variance between schools (σ²_school) and variance between pupils within schools (σ²_e). These components are estimated by the ANOVA (expected mean squares) method or, more commonly today, by REML (restricted maximum likelihood). Their ratios underlie important measures such as the intraclass correlation (ICC = σ²_school / (σ²_school + σ²_e)) and heritability (h²) in genetics.

Variance components cannot be negative by definition; the ANOVA method can produce negative estimates in small samples, whereas REML pulls them to the boundary at zero. Testing whether a component is zero is a test on the boundary of the parameter space, so standard Wald or likelihood-ratio tests are conservative. With few groups (e.g. 5 schools), the higher-level component is estimated imprecisely, and confidence intervals should be built on the log scale or by profile likelihood. Unlike fixed effects, the levels are regarded as a sample from a population, and conclusions are generalized to that population.

An everyday example

A laboratory wonders where the uncertainty in blood-sugar measurements comes from. If 70% of the total variance comes from genuine differences between people, 20% from the day of measurement and 10% from the fluctuation of repeated measurements on the same day, the first step in improving the process is to look into why the days differ.

In MerQur

Variance Components Analysis (VARCOMP) is devoted to this topic. The 'REML section' gives the σ² components in the Nested and Crossed Mixed Models, the 'ANOVA / variance components table' output gives the two-way components in ICC, the 'Subject (repeated measures)' option gives σ²_b + σ²_w in Bland-Altman analysis, and 'Profile CI for variance components' gives their confidence intervals in Hierarchical Bayesian Regression. They are also reported in Linear Mixed Models and GLMM.

Analyses where this concept appears

Bland-Altman Agreement AnalysisLinear Mixed Models (LMM)Nested Mixed Model (Nested LMM, Multi-Level Hierarchical Mixed Model)Crossed Mixed Model (Crossed LMM, Crossed Random Effects Mixed Model)Generalized Linear Mixed Models (GLMM)Intraclass Correlation Coefficient (ICC)Variance Components Analysis (VARCOMP)Hierarchical Bayesian Regression

Related concepts

Fixed and random effectsRestricted maximum likelihood (REML)Intraclass correlation (ICC)