In stevencarlislewalker/lme4ord: Generalized Linear Mixed Models with Flexible (Co)Variance Structures

Notes on glmer deviance functions

library("lme4")

parsedForm <- glFormula(cbind(incidence, size - incidence) ~ period + (1 | herd),
                        family = binomial, data = cbpp)

glmer models are usually fitted in two steps. The first step involves no integration (nAGQ = 0), and the second involves integration by either Laplace (nAGQ = 1) or adaptive Gaussian quadrature (nAGQ > 1).

(devfun0 <- do.call(mkGlmerDevfun, parsedForm))

(devfun1 <- updateGlmerDevfun(devfun0, parsedForm$reTrms))

as.list(body(devfun1)[c(3, 4, 8, 9)])

First stage deviance function

Note that there are several objects being used that do not appear to be created anywhere (e.g. lp0). Don't worry, they are actually in the environment of devfun0, which we can list using the following command.

ls(rho <- environment(devfun0))

Therefore, all of these objects are available to devfun0.

We will now step through each line, explaining what it does. The first line,

body(devfun0)[[2]]

updates the conditional mean of the response variable at the value of the linear predictor, lp0. Mathematically, the linear predictor is, $$ \mathbf\eta = \mathbf o + \mathbf X\mathbf\beta + \mathbf Z \mathbf\Lambda_\theta \mathbf u $$ The conditional mean is then, $$ \mathbf\mu = \text{link}(\mathbf\eta) $$ Note that the condition mean depends on the covariance parameters $\mathbf\theta$. Therefore, this update of the conditional mean is based on the old value of $\mathbf\theta$. The new value of $\mathbf\theta$ is updated on the next line:

body(devfun0)[[3]]

Next the penalized weighted residual sum of squares is calculated with this new value of $\mathbf\theta$:

body(devfun0)[[4]]

Finally the weights are updated (for safety?)

body(devfun0)[[5]]

---

Second stage deviance function

ls(rho <- environment(devfun1))

rho$baseOffset

body(devfun1)[[2]]

body(devfun1)[[3]]

body(devfun1)[[4]]

body(devfun1)[[5]]

body(devfun1)[[6]]

body(devfun1)[[7]]

stevencarlislewalker/lme4ord documentation built on May 30, 2019, 4:43 p.m.