MerQuR+Statistics Glossary

Statistics Glossary → Adaptive Gauss-Hermite quadrature (AGQ) and Laplace approximation

Adaptive Gauss-Hermite quadrature (AGQ) and Laplace approximationconcept

Methods that numerically compute the likelihood integral over the random effects in generalized linear mixed models, which has no closed-form solution.

In logistic or Poisson mixed models, each cluster's contribution to the likelihood is found by averaging over all possible values of the random effect (an integral), and this integral has no closed formula. Gauss-Hermite quadrature approximates it by a weighted sum of values at specific nodes chosen for the normal distribution. The adaptive version centres the nodes around each cluster's own estimate of its random effect, achieving high accuracy with few nodes.

Adaptive quadrature with a single node equals the Laplace approximation; it is fast but can give biased variance components and coefficients for binary data with few observations per cluster. As the number of nodes grows, the result approaches the exact likelihood, but so does the computing time; because the number of nodes grows exponentially with the number of random effects, quadrature is usually limited to a single scalar random effect. If the results do not change noticeably when nodes are added, the approximation is adequate.

An everyday example

To find the area under a curve, imagine measuring its height not everywhere but at a few well-chosen points and adding them up with suitable weights. If the points are placed where the curve is highest, even 5-10 measurements give a very accurate answer; a rough estimate from a single measurement is like the Laplace approximation.

In MerQur

In Generalized Linear Mixed Models the maximum-likelihood route of the 'Method' option uses adaptive Gauss-Hermite quadrature (AGQ), with the 'nAGQ' option (number of quadrature nodes; 1 = Laplace); in Bayesian ANOVA the posterior parameter summaries under 'Additional output' are computed deterministically with Gauss-Hermite.

Analyses where this concept appears

Generalized Linear Mixed Models (GLMM)Bayesian ANOVA

Related concepts

Maximum likelihood (ML) estimationFixed and random effects