A nonlinear Gauss-Seidel algorithm for inference about GLMM (Q5943412)

The following problem is considered. Suppose \(y\) is a vector of observations, \(\beta\) is a vector of fixed effects, \(\alpha\) is a vector of random effects and \(\theta\) is a vector of variance components. \(\alpha\) and \(\beta\) are unknown. Let \(f(y,\alpha |\beta,\theta)\) be the joint density of \(y\) and \(\alpha\) given \(\beta\) and \(\theta.\) The author considers estimates \(\hat \alpha\) and \(\hat \beta\), such that \(\hat \alpha = \hat \alpha (y, \theta)\), \(\hat \beta = \hat \beta (y, \theta)\) jointly maximize \(f(y,\alpha |\beta,\theta).\) He observes that \(f(y,\alpha |\beta,\theta) = f(y |\beta,\theta)f(\alpha |y,\beta\theta)\) and proposes to maximize by two steps. In the first step one finds \(\tilde\alpha = \tilde\alpha(\beta)\) that maximizes \(f(y,\alpha |\beta\theta)\) for a fixed \(\beta.\) In the second step one finds \(\hat\beta\) that maximizes \(f(y,{\bar\alpha} |\beta\theta)\) and let \(\hat\alpha = \tilde\alpha (\hat\beta).\) Generalized linear mixed models (GLMM) are considered in more details. For GLMM the author proposes a recursive algorithm to solve the computational problems. He proves that this algorithm converges and considers a numerical example. The numerical example shows the superiority of the proposed algorithm over the standard Newton-Raphson procedure when the number of random effects is large.