spaMM-conventions: spaMM conventions and differences from related fitting...
In spaMM: Mixed-Effect Models, with or without Spatial Random Effects

spaMM-conventions

R Documentation

spaMM conventions and differences from related fitting procedures

Description

input arguments are generally similar to those of glm and (g)lmer, in particular for the spaMM::fitme function, with the exception of the prior.weights argument, which is simply weights in the other packages. The name prior.weights seems more consistent, since e.g. glm returns its input weights as output prior.weights, while its output weights are instead the weights in the final iteration of an iteratively weighted least-square fit.

The \bolddefault likelihood target for dispersion parameters is restricted likelihood (REML estimation) for corrHLfit and (marginal) likelihood (ML estimation) for fitme. Model fits may provide restricted likelihood values(ReL) even if restricted likelihood is is not used as an objective function at any step in the analysis.

See good-practice for advice about the proper syntax of formula.

Computation times depend on control parameters given by spaMM.getOption("spaMM_tol") parameters (for iterative algorithms), and spaMM.getOption("nloptr") parameters for the default optimizer. Do not use spaMM.options() to control them globally, unless you know what you are doing. Rather control them locally by the control.HLfit argument to control spaMM_tol, and by the control arguments of corrHLfit and fitme to control nloptr. If nloptr$Xtol_rel is set above 5e-06, fitme will by default refit the fixed effects and dispersion parameters (but not other correlation parameters estimated by nloptr) by the iterative algorithm after nloptr convergence. Increasing nloptr$Xtol_rel value may therefore switches the bulk of computation time from the optimizer to the iterative algorithm, and may increase or decrease computation time depending on which algorithm is faster for a given input. Use control$refit if you wish to inhibit this, but note that by default it provides a rescue to a poor nloptr result due to a too large Xtol_rel.