MerQuR+Statistics Glossary

Statistics Glossary → Covariance (correlation) structure (UN, VC, AR(1), exchangeable)

Covariance (correlation) structure (UN, VC, AR(1), exchangeable)concept

The model assumption that specifies which pattern the variances and correlations among observations from the same unit follow in repeated or clustered measurements.

Measurements taken on the same patient at different times are not independent, and one must choose how the model represents this dependence. The independence structure sets all correlations to zero. Compound symmetry (exchangeable, CS) assumes the same correlation between every pair of measurements. In AR(1) the correlation decays geometrically with the time gap (ρ between adjacent measurements, ρ² two steps apart, and so on). The Toeplitz structure allows a separate correlation for each distance, and the unstructured (UN) form leaves all variances and covariances free. For random effects, the VC (diagonal) structure treats the effects as uncorrelated and UN as correlated.

More flexible structures need more parameters: UN estimates t(t + 1) / 2 parameters for t time points and can cause convergence problems in studies with few subjects. Choosing between models with the same fixed effects by AIC or BIC is a common approach. In mixed models a wrong structure can distort the standard errors and tests of the fixed effects. In GEE, by contrast, the robust (sandwich) standard errors remain consistent even if the 'working correlation' is wrong; the right structure only improves efficiency.

An everyday example

A child's height is measured every year. Heights at ages 6 and 7 are very closely linked, those at 6 and 12 more loosely. This is well represented by a structure such as AR(1), in which correlation decays with distance. For pupils in the same class, however, there is no 'order' among their scores, so an exchangeable structure with equal correlation for every pair is more natural.

In MerQur

In Linear Mixed Models, the 'Random-effects covariance' (UN or VC), 'Residual covariance structure (R / REPEATED)' (σ²I, AR(1), CS, UN, TOEP, CSH, ARH(1)) and 'Compare R structures by AIC/BIC' options; in Generalized Estimating Equations, 'Correlation structure' and 'M-dependent lag'. In Latent Profile Analysis, the 'Variance-covariance model' (EEI, VVI, EEE, VVV) sets the constraints on variances and covariances across profiles.

Analyses where this concept appears

Variable Clustering (VARCLUS)Linear Mixed Models (LMM)Generalized Estimating Equations (GEE)Latent Profile Analysis (LPA)

Related concepts

Random slopeInformation criteria (AIC, AICc, BIC)