asreml.options {asreml} | R Documentation |
Set less frequently used asreml()
options.
asreml.options(...)
... |
Arguments in the form |
The following settings can be altered:
ai.loadings=0
Controls modification to AI updates of
loadings in extended factor-analytic (fa
) models. After
ASReml calculates updates for variance parameters, it checks
whether the updates are reasonable and sometimes reduces them over
and above any step.size
shrinkage. The extra shrinkage has
two levels. Loadings that change sign are restricted to doubling in
magnitude, and if the average change in magnitude of loadings is
greater than 10-fold, they are all shrunk. Unless the user
specifies constraints, ASReml sets them and rotates the loadings
each iteration. When ai.loadings
i is specified
(default i=-1 specifies no action), it also prevents AI
updates of some loadings during the first i iterations. For
f > 1 factors, only the last factor is estimated (conditional
on the earlier ones) in the first f-1 iterations. Then pairs,
including the last, are estimated until iteration i.
ai.penalty=10
The algorithm for updating loadings in
factor analytic models has been improved. The original update
procedure sometimes produced unreasonable updates, or exhibited
drift. The present strategy modifys the average information matrix
by increasing the diagonal elements pertaining to loadings by a
percentage, p. The default is to start with p = 10%
and reduce it by 1 or 2% each iteration down to 1%. If the
starting values are poor, 10% may not be a suficient initial
retardation. If it appears the updates are unreasonable, the value
of p is increased by 10%. The default is p=10%. After
the penalty has reduced to 1%, it is further reduced to
0.2%. ai.penalty
can be set to 0 if desired.
ai.sing=FALSE
Force continuation if a singularity is detected in the average information matrix.
aodev=FALSE
If TRUE
, return an analyais of deviance.
aom=FALSE
If TRUE
, return standardized
conditional residuals and standardized conditional BLUPs
in the aom
component of the asreml
object.
Cfixed=FALSE
If TRUE
, return the computed part of the
C^{-1} matrix in component Cfixed
; the default is FALSE
.
The inverse coefficient matrix is fully formed for terms in the dense
set.
colourise=TRUE
If TRUE
(the default) the header text produced
by functions such as wald
and predict
will be displayed in a different
colour if supported by the output terminal device.
Csparse=~NULL
If a formula is specified, return the
computed part of C^{-1} for those terms given in the
formula. asreml
does not compute the whole of C^{-1},
only that which is sufficient to calculate the REML solution.
debug=FALSE
Return internal data structures.
dense=~NULL
Include the equation(s) for the term(s) in the formula in the dense set. This results in faster processing if the term is associated with a known dense inverse relationship matrix.
design=FALSE
If TRUE
, return the design matrix
in component design
of the asreml
object.
drop.unused.levels=TRUE
If TRUE
(the default),
levels of simple factors that do not appear in the data are dropped.
eqorder=3
Set the algorithm used for ordering the
mixed-model equations prior to solution. eqorder
=-1
processes the equations in user order; generally this will
run much slower, if at all in real time for large analyses.
extra = n
Forces another mod(n, maxit)
iterations
after apparent convergence; the default is n=0
.
fail="hard"
If "hard"
(the default) fatal
errors will terminate execution, otherwise if "soft"
such
conditions will be reported as warnings, allowing simulation runs,
for example, to continue. In both cases the converge
component
of the asreml
object will be set to FALSE
and the results
will be erroneous.
fixgammas=FALSE
If TRUE
, all variance
parameters are constrained to be fixed at their starting values.
font.scale=1.0
Scale axis text and labels (relative to
the asreml
default settings) in the graphs generated by
plot.asreml()
.
gammaPar=FALSE
If TRUE
(the default is FALSE
),
single section models will be fitted using the gamma
parameterization
irrespective of whether the residual
formula specifies a correlation
or variance model. The default behaviour for single section models is to fit
on the gamma
scale if the residual
formula specifies a correlation
structure, and on the sigma
scale if the residual
formula
specifies a variance structure.
glmminloop=1
Sets the number of inner iterations performed in an iteratively weighted least squares analysis. These estimate the effects in the linear model for the current set of variance parameters; outer iterations are the AI updates to the variance parameters. The default is to perform 4 inner iterations in the first round and 2 in subsequent rounds of the outer iteration. Set to 2 or more to increase the number of inner iterations.
grid=TRUE
A logical vector of length 1 or
length(design.points)
(see predict)
controlling the expansion of coordinates for 2 dimensional
kriging. For a given term, the coordinates for prediction in 2
dimensions (x,y) are given as a list of two vectors or a two
column matrix component of design.points
. If TRUE
,
the coordinates are expanded to form an (x,y) grid of all
possible combinations, otherwise the columns of the matrix and are
taken in parallel.
keep.order=FALSE
If TRUE
, the order of terms
in the fixed
formula is retained. Set to TRUE
if the
special model function and()
is present.
knots=50
The default number of knot points for spline
terms. For a variate x
, the number of knot points is
min(length(unique(x)), knots).
maxit=13
Maximum number of iterations.
nsppoints=21
Influences the number of points used
when predicting splines and polynomials. The design matrix
generated by the pol(x)
and spl(x)
functions are
modified to include extra rows for points used in prediction. The
range of x is divided by nsppoints-1 to give a step size
i. For each point p in x, a predict point is
inserted at p+i if there is no data value in the interval
[p, p+1.1i]. nsppoints
is ignored if the
predict.asreml()
argument design.points
is set (or the
design.points
component of the predict
list argument
to asreml()
is not empty). This process also affects the
number of levels identified by dev(x)
.
oscillate=TRUE
Test for oscillating log-likelihood (default=TRUE).
pxem=1
(PX)EM update strategy for unstructured (US)
variance models when Average Information updates cause them to be
non-positive definite (see uspd
). Valid values are:
pxem | Action |
1 | standard EM + 10 local EM steps |
2 | standard EM + 10 local PXEM steps |
3 | standard EM + 10 local EM steps^* |
4 | standard EM + 10 local PXEM steps^* |
5 | standard EM only |
6 | single local PXEM |
7 | standard EM + 1 local EM step |
8 | standard EM + 1 local PXEM step |
^*Options 3 and 4 cause all US structures to be updated by (PX)EM if any particular one requires EM updates.
pworkspace="128mb"
Sets the workspace needed by the
predict()
method; follows the same convention as
workspace
. Ignored if the predict
argument to
asreml()
is not set. Note that the total workspace used for
prediction is workspace
+pworkspace
.
random.order="noeff"
Reorder terms in the
random
and sparse
formulae in increasing order of
number of effects. This is almost always desirable, especially if
the stratum variance decomposition is required. Other options are
"user"
to retain the order given, or "R"
for the
default R rules.
rotate.fa=FALSE
If FALSE
(the default),
asreml()
initially constrains the first k-1 loadings
for higher order (k>1) factors in factor analytic models to
zero. If constraints are not set for factor analytic models with
more than one factor, asreml()
will set them internally and
rotate the loadings each iteration (rotate.fa=TRUE
). This
option also modifies the action of update.Gcon
such that
rotation, if specified, is applied on an update.
scale=1.0
Overall scale parameter.
score=FALSE
If TRUE
, the score vector is
returned in a component score
of the asreml
object; the default is FALSE
.
spline.scale=-1
When forming a design matrix for a
spl()
term, a standardised scale is used. Setting
spline.scale = 1
forces asreml
to use the scale of
the variable. The default (-1) is recommended in most cases.
spline.step=list(spl=10000,dev=10000,pol=10000)
A list
with components named spl
, dev
and pol
specifying the resolution for spline deviations and polynomial
functions, respectively. Points closer together than
1/spline.step
of the range will be treated as a single
point.
step.size=0.316
Update shrinkage factor, reduces the update step sizes of the variance parameters. The step size is incremented each iteration to a maximum of 1.0.
tol=c(0,0)
A vector of length two that modifies the
sensitivity of asreml
to detect singularities in the mixed
model equations. This is intended for the rare occasions when
singularities are detected after the first iteration.
Normally a singularity is declared if the adjusted sum of squares
of a covariable is less than e, or less than the uncorrected
sum of squares \times e, where e=10^{-8} in the first
iteration and 10^{-10} thereafter. If tol=c(a,b)
,
e is scaled by 10^a for the the first iteration, and
10^b for subsequent iterations. Once a singularity is
detected, the corresponding equation is dropped (forced to be zero)
in subsequent iterations. If the problem of later singularities
arises because of the low coefficient of variation of a covariable,
it may be advisable to centre and rescale the covariable. If the
degrees of freedom are correct in the first iteration, the problem
lies with the variance parameters and a different variance model
(or constraint) is needed.
trace=TRUE
Report convergence monitoring in the console.
update.Gcon=TRUE
Update the constraint status of
variance parameters in the G.param
list component on
termination; this may influence subsequent updates to the model fit
(the default is TRUE
).
update.Rcon=TRUE
Update the constraint status of
variance parameters in the R.param
list component on
termination; this may influence subsequent updates to the model fit
(the default is TRUE
).
update.step.size=0.316
Update shrinkage factor to use
in a call to update()
. Ignored if set to the default (0.316)
or if step.size
is explicitly specified on the update()
call; otherwise the shrinkage factor is set to update.step.size
in the call to asreml()
constructed by update()
.
uspd=TRUE
If TRUE, (the default) set the boundary constraint for
each parameter in unstructured variance models to "P"
. Under
these conditions, asreml
checks whether the updated matrix
is positive definite; if not, the average information update
is replaced with an EM update (see pxem
).
workspace="128mb"
Sets the workspace for the core REML
routines in the form of a number optionally followed directly by a
valid measurement unit. Valid units are kb
, mb
or
gb
; if no units are given then the value is interpreted as
double-precision words (groups of 8 bytes).