mhglm.control function

Auxiliary for Controlling Moment Heirarchical GLM Fitting

Auxiliary function for mhglm fitting. Typically only used internally by mhglm.fit, but may be used to construct a control argument to either function.

mhglm.control(standardize = TRUE, steps = 1, parallel = FALSE, diagcov = FALSE,
        fit.method = "firthglm.fit", fixef.rank.warn = FALSE, cov.rank.warn = FALSE,
        cov.psd.warn = TRUE, fit.control = list(...), ...)

mhglm_ml.control(standardize = FALSE, steps = 1, parallel = FALSE, diagcov = FALSE,
              fit.method = "firthglm.fit", fixef.rank.warn = FALSE, cov.rank.warn = FALSE,
              cov.psd.warn = FALSE, fit.control = list(...), ...)

Arguments

• standardize: logitcal indicating if predictors should be standardized before moment-based fitted
• steps: number of refinement steps
• parallel: fit the group-specific estimates in parallel rather than sequentially
• diagcov: estimate random effect covairance matrix with diagonal approximation
• fit.method: method for obtaining group-specific effect estimates
• fixef.rank.warn: if TRUE, print warnings when fixef is unidentifiable
• cov.rank.warn: if TRUE, print warnings when covariance matrix is unidentifiable
• cov.psd.warn: if TRUE, print warnings when moment based estimates of covariance matrix is not positive semi-definite
• fit.control: control parameters for fit.method
• ...: arguments to be used to form the fit.control argument if it is not supplied directly.

Details

Setting standardize = TRUE ensures that the procedure is equivariant, and generally leads to better estimation performance. Right now standardize = TRUE is not allowed for mhglm_ml.

The steps argument gives the number of refinement steps for the moment based parameters. In each step, the previous fixed effect and random effect covariance matrix estimates are used to weight the subpopulation-specific effect estimates. In principle, higher values of steps could lead to more accurate estimates, but in simulations, the differences are negligible.

Returns

A list with components named as the arguments.

See Also

mhglm.fit, the fitting procedure used by mhglm.

firthglm.fit, the default subpopulation-specific fitting method.

Examples

library(lme4) # for cbpp data

# The default fitting method uses Firth's bias-corrected estimates
(gm.firth <- mhglm(cbind(incidence, size - incidence) ~ period + (1 | herd),
                   data = cbpp, family = binomial,
                   control=mhglm.control(fit.method="firthglm.fit")))

# Using maximum likelihood estimates is less reliable
(gm.ml <- mhglm(cbind(incidence, size - incidence) ~ period + (1 | herd),
                data = cbpp, family = binomial,
                control=mhglm.control(fit.method="glm.fit")))