Define a list of control parameters. Note that
the format of this input is likely to change more rapidly than that of
dsem
dsem_control(
nlminb_loops = 1,
newton_loops = 1,
trace = 0,
eval.max = 1000,
iter.max = 1000,
getsd = TRUE,
quiet = FALSE,
run_model = TRUE,
gmrf_parameterization = c("separable", "projection"),
constant_variance = c("conditional", "marginal", "diagonal"),
use_REML = TRUE,
profile = NULL,
parameters = NULL,
map = NULL,
getJointPrecision = FALSE,
extra_convergence_checks = TRUE,
lower = -Inf,
upper = Inf
)An S3 object of class "dsem_control" that specifies detailed model settings, allowing user specification while also specifying default values
Integer number of times to call nlminb.
Integer number of Newton steps to do after running
nlminb.
Parameter values are printed every `trace` iteration
for the outer optimizer. Passed to
`control` in nlminb.
Maximum number of evaluations of the objective function
allowed. Passed to `control` in nlminb.
Maximum number of iterations allowed. Passed to `control` in
nlminb.
Boolean indicating whether to call sdreport
Boolean indicating whether to run model printing messages to terminal or not;
Boolean indicating whether to estimate parameters (the default), or instead to return the model inputs and compiled TMB object without running;
Parameterization to use for the Gaussian Markov random field, where the default `separable` constructs a precision matrix that must be full rank, and the alternative `projection` constructs a full-rank and IID precision for variables over time, and then projects this using the inverse-cholesky of the precision, where this projection can be rank-deficient.
Whether to specify a constant conditional variance
\( \mathbf{\Gamma \Gamma}^t\) using the default constant_variance="conditional",
which results in a changing marginal variance
along the specified causal graph when lagged paths are present. Alternatively, the user can
specify a constant marginal variance using constant_variance="diagonal"
or constant_variance="marginal",
such that \( \mathbf{\Gamma}\) and \(\mathbf{I-P}\) are rescaled to achieve this constraint.
All options
are equivalent when the model includes no lags (only simultaneous effects) and
no covariances (no two-headed arrows). "diagonal" and "marginal"
are equivalent when the model includes no covariances. Given some exogenous covariance,
constant_variance = "diagonal" preserves the conditional correlation and has
changing conditional variance, while constant_variance = "marginal" has changing
conditional correlation along the causal graph.
Boolean indicating whether to treat non-variance fixed effects as random, either to motigate bias in estimated variance parameters or improve efficiency for parameter estimation given correlated fixed and random effects
Parameters to profile out of the likelihood (this subset will be appended to random with Laplace approximation disabled).
list of fixed and random effects, e.g., as constructed by dsem and then modified
by hand (only helpful for advanced users to change starting values or restart at intended values)
list of fixed and mirrored parameters, constructed by dsem by default but available
to override this default and then pass to MakeADFun
whether to get the joint precision matrix. Passed
to sdreport.
Boolean indicating whether to run extra checks on model convergence.
vectors of lower bounds, replicated to be as long as start and passed to nlminb.
If unspecified, all parameters are assumed to be unconstrained.
vectors of upper bounds, replicated to be as long as start and passed to nlminb.
If unspecified, all parameters are assumed to be unconstrained.