MerQuR+Statistics Glossary

Statistics Glossary → Satterthwaite and Kenward-Roger degrees-of-freedom approximations

Satterthwaite and Kenward-Roger degrees-of-freedom approximationsconcept

Methods that compute approximate (often non-integer) degrees of freedom for t and F tests when the variance of an estimate is a combination of several variance components.

In classical t and F tests the degrees of freedom follow directly from the number of observations and parameters. When the variance of an estimate is a weighted sum of different sources, as with unequal group variances or random effects in mixed models, this is no longer possible. The Satterthwaite approximation matches the distribution of this combined variance to the chi-square distribution that fits it best and uses that distribution's degrees of freedom. The Welch-Satterthwaite formula in Welch's t-test is the best-known example.

The Kenward-Roger method goes one step further in mixed models: besides approximating the degrees of freedom, it inflates the standard errors of the fixed effects to account for the fact that the variance components are themselves estimated. This brings the Type I error rate closer to its nominal level, especially when there are few clusters or subjects. With many clusters the three options (Satterthwaite, Kenward-Roger and the asymptotic Wald z) give similar results; with few clusters Wald z is too optimistic.

An everyday example

An education study measures the effect of a new teaching method in six schools, each with several classes of pupils. Although there are 600 pupils, only six schools carry the between-school differences, so the test of the method should rest on far fewer than 600 degrees of freedom. The Satterthwaite approximation might put this at 4.3 and compute the p-value accordingly.

In MerQur

In Linear Mixed Models, the 'Fixed-effect df method' option: Satterthwaite t (default), Kenward-Roger or Wald z. The Nested and Crossed Mixed Models give the fixed-effect tests and EMMEANS pairwise comparisons in their REML section with Satterthwaite degrees of freedom, and the fixed-effects table of Variance Components Analysis also uses Satterthwaite degrees of freedom. The Welch solution of the Independent t-Test uses Welch-Satterthwaite degrees of freedom, and the Intraclass Correlation Coefficient uses Satterthwaite degrees of freedom for the confidence interval and F test under absolute agreement.

Analyses where this concept appears

Independent-Samples t-TestLinear Mixed Models (LMM)Nested Mixed Model (Nested LMM, Multi-Level Hierarchical Mixed Model)Crossed Mixed Model (Crossed LMM, Crossed Random Effects Mixed Model)Intraclass Correlation Coefficient (ICC)Variance Components Analysis (VARCOMP)

Related concepts

Degrees of freedomFixed and random effectsWelch correction