R: Control of Mixed Model Fitting

lmerControl {lme4}

R Documentation

Control of Mixed Model Fitting

Description

Construct control structures for mixed model fitting. All arguments have defaults, and can be grouped into

general control parameters, most importantly optimizer, further restart_edge, etc;
model- or data-checking specifications, in short “checking options”, such as check.nobs.vs.rankZ, or check.rankX (currently not for nlmerControl);
all the parameters to be passed to the optimizer, e.g., maximal number of iterations, passed via the optCtrl list argument.

Usage

lmerControl(optimizer = "nloptwrap",
				    
    restart_edge = TRUE,
    boundary.tol = 1e-5,
    calc.derivs = TRUE,
    use.last.params = FALSE,
    sparseX = FALSE,
    standardize.X = FALSE,
    ## input checking options
    check.nobs.vs.rankZ = "ignore",
    check.nobs.vs.nlev = "stop",
    check.nlev.gtreq.5 = "ignore",
    check.nlev.gtr.1 = "stop",
    check.nobs.vs.nRE= "stop",
    check.rankX = c("message+drop.cols", "silent.drop.cols", "warn+drop.cols",
                    "stop.deficient", "ignore"),
    check.scaleX = c("warning","stop","silent.rescale",
                     "message+rescale","warn+rescale","ignore"),
    check.formula.LHS = "stop",
    ## convergence checking options
    check.conv.grad     = .makeCC("warning", tol = 2e-3, relTol = NULL),
    check.conv.singular = .makeCC(action = "message", tol = formals(isSingular)$tol),
    check.conv.hess     = .makeCC(action = "warning", tol = 1e-6),
    ## optimizer args
    optCtrl = list(),
    mod.type = "lmer"
)

glmerControl(optimizer = c("bobyqa", "Nelder_Mead"),
    restart_edge = FALSE,
    boundary.tol = 1e-5,
    calc.derivs = TRUE,
    use.last.params = FALSE,
    sparseX = FALSE,
    standardize.X = FALSE,
    ## input checking options
    check.nobs.vs.rankZ = "ignore",
    check.nobs.vs.nlev = "stop",
    check.nlev.gtreq.5 = "ignore",
    check.nlev.gtr.1 = "stop",
    check.nobs.vs.nRE= "stop",
    check.rankX = c("message+drop.cols", "silent.drop.cols", "warn+drop.cols",
                    "stop.deficient", "ignore"),
    check.scaleX = c("warning","stop","silent.rescale",
                     "message+rescale","warn+rescale","ignore"),
    check.formula.LHS = "stop",
    ## convergence checking options
    check.conv.grad     = .makeCC("warning", tol = 2e-3, relTol = NULL),
    check.conv.singular = .makeCC(action = "message", tol = formals(isSingular)$tol),
    check.conv.hess     = .makeCC(action = "warning", tol = 1e-6),
    ## optimizer args
    optCtrl = list(),
    mod.type = "glmer",
    tolPwrss = 1e-7,
    compDev = TRUE,
    nAGQ0initStep = TRUE,
    check.response.not.const = "stop"
 )

nlmerControl(optimizer = "Nelder_Mead", tolPwrss = 1e-10,
             optCtrl = list())

.makeCC(action, tol, relTol, ...)

Arguments

`optimizer`	character - name of optimizing function(s). A `character` vector or list of functions: length 1 for `lmer` or `glmer`, possibly length 2 for `glmer`). Built-in optimizers are `"Nelder_Mead"`, `"bobyqa"` (from the minqa package), `"nlminbwrap"` (using base R's `nlminb`) and the default for `lmerControl()`, `"nloptwrap"`. Any minimizing function that allows box constraints can be used provided that it (1) takes input parameters `fn` (function to be optimized), `par` (starting parameter values), `lower` and `upper` (parameter bounds) and `control` (control parameters, passed through from the `control` argument) and (2) returns a list with (at least) elements `par` (best-fit parameters), `fval` (best-fit function value), `conv` (convergence code, equal to zero for successful convergence) and (optionally) `message` (informational message, or explanation of convergence failure). Special provisions are made for `bobyqa`, `Nelder_Mead`, and optimizers wrapped in the optimx package; to use the optimx optimizers (including `L-BFGS-B` from base `optim` and `nlminb`), pass the `method` argument to `optim` in the `optCtrl` argument (you may need to load the optimx package manually using `library(optimx)`). For `glmer`, if `length(optimizer)==2`, the first element will be used for the preliminary (random effects parameters only) optimization, while the second will be used for the final (random effects plus fixed effect parameters) phase. See `modular` for more information on these two phases. If `optimizer` is `NULL` (at present for `lmer` only), all of the model structures will be set up, but no optimization will be done (e.g. parameters will all be returned as `NA`).
`calc.derivs`	logical - compute gradient and Hessian of nonlinear optimization solution?
`use.last.params`	logical - should the last value of the parameters evaluated (`TRUE`), rather than the value of the parameters corresponding to the minimum deviance, be returned? This is a "backward bug-compatibility" option; use `TRUE` only when trying to match previous results.
`sparseX`	logical - should a sparse model matrix be used for the fixed-effects terms? Currently inactive.
`restart_edge`	logical - should the optimizer attempt a restart when it finds a solution at the boundary (i.e. zero random-effect variances or perfect +/-1 correlations)? (Currently only implemented for `lmerControl`.)
`boundary.tol`	numeric - within what distance of a boundary should the boundary be checked for a better fit? (Set to zero to disable boundary checking.)
`tolPwrss`	numeric scalar - the tolerance for declaring convergence in the penalized iteratively weighted residual sum-of-squares step.
`compDev`	logical scalar - should compiled code be used for the deviance evaluation during the optimization of the parameter estimates?
`nAGQ0initStep`	Run an initial optimization phase with `nAGQ = 0`. While the initial optimization usually provides a good starting point for subsequent fitting (thus increasing overall computational speed), setting this option to `FALSE` can be useful in cases where the initial phase results in bad fixed-effect estimates (seen most often in binomial models with `link="cloglog"` and offsets).
`check.nlev.gtreq.5`	character - rules for checking whether all random effects have >= 5 levels. See `action`.
`check.nlev.gtr.1`	character - rules for checking whether all random effects have > 1 level. See `action`.
`check.nobs.vs.rankZ`	character - rules for checking whether the number of observations is greater than (or greater than or equal to) the rank of the random effects design matrix (Z), usually necessary for identifiable variances. As for `action`, with the addition of `"warningSmall"` and `"stopSmall"`, which run the test only if the dimensions of `Z` are < 1e6. `nobs > rank(Z)` will be tested for LMMs and GLMMs with estimated scale parameters; `nobs >= rank(Z)` will be tested for GLMMs with fixed scale parameter. The rank test is done using the `method="qr"` option of the `rankMatrix` function.
`check.nobs.vs.nlev`	character - rules for checking whether the number of observations is less than (or less than or equal to) the number of levels of every grouping factor, usually necessary for identifiable variances. As for `action`. `nobs<nlevels` will be tested for LMMs and GLMMs with estimated scale parameters; `nobs<=nlevels` will be tested for GLMMs with fixed scale parameter.
`check.nobs.vs.nRE`	character - rules for checking whether the number of observations is greater than (or greater than or equal to) the number of random-effects levels for each term, usually necessary for identifiable variances. As for `check.nobs.vs.nlev`.
`check.conv.grad`	rules for checking the gradient of the deviance function for convergence. A list as returned by `.makeCC`, or a character string with only the action.
`check.conv.singular`	rules for checking for a singular fit, i.e. one where some parameters are on the boundary of the feasible space (for example, random effects variances equal to 0 or correlations between random effects equal to +/- 1.0); as for `check.conv.grad` above. The default is to use `isSingular(.., tol = *)`'s default.
`check.conv.hess`	rules for checking the Hessian of the deviance function for convergence.; as for `check.conv.grad` above.
`check.rankX`	character - specifying if `rankMatrix(X)` should be compared with `ncol(X)` and if columns from the design matrix should possibly be dropped to ensure that it has full rank. Sometimes needed to make the model identifiable. The options can be abbreviated; the three `"*.drop.cols"` options all do drop columns, `"stop.deficient"` gives an error when the rank is smaller than the number of columns where `"ignore"` does no rank computation, and will typically lead to less easily understandable errors, later.
`check.scaleX`	character - check for problematic scaling of columns of fixed-effect model matrix, e.g. parameters measured on very different scales.
`check.formula.LHS`	check whether specified formula has a left-hand side. Primarily for internal use within `simulate.merMod`; use at your own risk as it may allow the generation of unstable `merMod` objects
`check.response.not.const`	character - check that the response is not constant.
`optCtrl`	a `list` of additional arguments to be passed to the nonlinear optimizer (see `Nelder_Mead`, `bobyqa`). In particular, both `Nelder_Mead` and `bobyqa` use `maxfun` to specify the maximum number of function evaluations they will try before giving up - in contrast to `optim` and `optimx`-wrapped optimizers, which use `maxit`. (Also see `convergence` for details of stopping tolerances for different optimizers.) Note: All of `lmer()`, `glmer()` and `nlmer()` have an optional integer argument `verbose` which you should raise (to a positive value) in order to get diagnostic console output about the optimization progress.
`action`	character - generic choices for the severity level of any test, with possible values "ignore": skip the test. "warning": warn if test fails. "message": print a message if test fails. "stop": throw an error if test fails.
`tol`	(numeric) tolerance for checking the gradient, scaled relative to the curvature (i.e., testing the gradient on a scale defined by its Wald standard deviation)
`relTol`	(numeric) tolerance for the gradient, scaled relative to the magnitude of the estimated coefficient
`mod.type`	model type (for internal use)
`standardize.X`	scale columns of X matrix? (not yet implemented)
`...`	other elements to include in check specification

Details

Note that (only!) the pre-fitting “checking options” (i.e., all those starting with "check." but not including the convergence checks ("check.conv.*") or rank-checking ("check.rank*") options) may also be set globally via options. In that case, (g)lmerControl will use them rather than the default values, but will not override values that are passed as explicit arguments.

For example, options(lmerControl=list(check.nobs.vs.rankZ = "ignore")) will suppress warnings that the number of observations is less than the rank of the random effects model matrix Z.

Value

The *Control functions return a list (inheriting from class "merControl") containing

general control parameters, such as optimizer, restart_edge;
(currently not for nlmerControl:) "checkControl", a list of data-checking specifications, e.g., check.nobs.vs.rankZ;
parameters to be passed to the optimizer, i.e., the optCtrl list, which may contain maxiter.

.makeCC returns a list containing the check specification (action, tolerance, and optionally relative tolerance).

Examples

str(lmerControl())
str(glmerControl())
## fit with default algorithm [nloptr version of BOBYQA] ...
fm0 <- lmer(Reaction ~ Days +   ( 1 | Subject), sleepstudy)
fm1 <- lmer(Reaction ~ Days + (Days | Subject), sleepstudy)
## or with "bobyqa" (default 2013 - 2019-02) ...
fm1_bobyqa <- update(fm1, control = lmerControl(optimizer="bobyqa"))
## or with "Nelder_Mead" (the default till 2013) ...
fm1_NMead <- update(fm1, control = lmerControl(optimizer="Nelder_Mead"))
## or with the nlminb function used in older (<1.0) versions of lme4;
## this will usually replicate older results
if (require(optimx)) {
    fm1_nlminb <- update(fm1,
                         control = lmerControl(optimizer= "optimx",
                                               optCtrl  = list(method="nlminb")))
    ## The other option here is method="L-BFGS-B".
}

## Or we can wrap base::optim():
optimwrap <- function(fn,par,lower,upper,control=list(),
                      ...) {
    if (is.null(control$method)) stop("must specify method in optCtrl")
    method <- control$method
    control$method <- NULL
    ## "Brent" requires finite upper values (lower bound will always
    ##  be zero in this case)
    if (method=="Brent") upper <- pmin(1e4,upper)
    res <- optim(par=par, fn=fn, lower=lower,upper=upper,
                 control=control,method=method,...)
    with(res, list(par  = par,
                   fval = value,
                   feval= counts[1],
                   conv = convergence,
                   message = message))
}
fm0_brent <- update(fm0,
                    control = lmerControl(optimizer = "optimwrap",
                                          optCtrl = list(method="Brent")))

## You can also use functions (in addition to the lmerControl() default "NLOPT_BOBYQA")
## from the 'nloptr' package, see also  '?nloptwrap' :
if (require(nloptr)) {
    fm1_nloptr_NM <- update(fm1, control=lmerControl(optimizer="nloptwrap",
                                      optCtrl=list(algorithm="NLOPT_LN_NELDERMEAD")))
    fm1_nloptr_COBYLA <- update(fm1, control=lmerControl(optimizer="nloptwrap",
                                      optCtrl=list(algorithm="NLOPT_LN_COBYLA",
                                                   xtol_rel=1e-6,
                                                   xtol_abs=1e-10,
                                                   ftol_abs=1e-10)))
}
## other algorithm options include NLOPT_LN_SBPLX

[Package lme4 version 1.1-35.5 Index]