v2.0.3
User interface changes
-
Detection of initial conditions was rewritten to enable additional features
in the initial conditions (#322). The most important user-facing change is
that now arbitrary R expressions can be used when setting initial conditions
such astvCL <- log(c(2,3,4))
(#253) instead of simplytvCL <- log(3)
-
The function as.nlmixrBounds() now supports adding the columns that are
missing into the input data.frame. -
omega definitions can be correlation matrices (#338)
-
Can specify
keep=
anddrop=
in the nlmixr function to keep and
drop columns in nlmixr output. Can also specify
control=list(keep=,drop=)
ornlmixr(...,keep=,drop=)
to
keep/drop columns (#260)
focei
changes:
-
Uses RxODE to re-arrange the problem so it does not include
if/else
in the model (ie. un-branched code). This allows
sensitivities to be calculated in one pass saving time for multiple
endpoint models and models withif/else
in them. -
linCmt()
now uses solved systems instead of translating to ODEs.- Uses
RxODE
/stan
's math headers to calculate the sensitivities
of the super-positionedlinCmt()
solutions. - This uses the
advan
solutions and hence supports
support time-varying covariates.
- Uses
-
focei
now supports censoring in the same waymonolix
does, with
cens
andlimit
columns -
focei
now allowseta
s on dose-related modeled events like
alag
,f
, etc by finite difference sensitivities. -
focei
now supports 2 combined additive + proportional error
models;combined1
:trans(y) = trans(f) + (a+b*f^c)*err
combined2
:trans(y) = trans(f) + sqrt(a^2+b^2*f^(2c))*err
-
focei
etaNudge
parameters were changed to use quadrature points
covering 95% percent of a standard normal. -
With zero gradients, Gill differences are recomputed to try to find
a non-zero gradient. -
Now when running if a zero gradient is detected, reset the problem
(theta reset) and re-estimated withouterOpt="bobyqa"
-
Now when running a model where the last objective function is not
the minimum objective function, issue a warning and skip the
covariance step. (See Issue #403) -
focei
proportional and power models are more tolerant of 0
predictions in your data
SAEM changes
-
saem
fits now gracefully fall back to thefocei
likelihood when
they support files are no longer on the loaded disk -
saem
phi pile is now saved in theRxODE::rxTempDir()
which can
be customized to allow thephi
file to remain after R has exited -
saem
fits now can add info
,foce
andfocei
likelihood -
saem
fits now useliblsoda
by default and are multi-threaded when
running (controlled byRxODE
) -
saem
now supports time-varying covariates (like clock-time) -
saem
now supports 2 combined additive + proportional error models:combined1
:trans(y) = trans(f) + (a+b*f^c)*err
combined2
:trans(y) = trans(f) + sqrt(a^2+b^2*f^(2c))*err
-
saem
proportional and power models are more tolerant of 0
predictions in your data -
saem
now supports censoring a similar way asmonolix
does, with
cens
andlimit
columns -
The default of
saem
additive + proportional error has been
switched tocombined2
, which was thefocei
default, but you can
change this back withsaemControl(addProp="combined2")
. The
table results will likely be different because in the last release
thesaem
calculatedcombined1
and then used these coefficients
in thecombined2
focei problem.
nlme changes
-
nlme
will now support 2 combined additive + proportional error models (if the patched version of nlme is used) -
Can switch with
nlmeControl(addProp="combined1")
to use the combined1 type of error model
New Utilities
-
bootstrapFit
now calculates the bootstrap confidence bands and
(optionally) will compare with the theoretical chi-squared
distribution to help assess their adequacy. -
covarSearchAuto
now allows automatic forward/backward covariate
selection
General Changes
-
Added auto-completion of
nlmixr
object properties accessed by
$
. This works for major editors includingRstudio
,ESS
, and
Base R itself. -
Changed the way that Rstudio notebooks display
nlmixr
objects; It
should be more legible in Rstudio. -
Graphics have been revamped to show censoring (including adding
ggplot stat/geomgeom_cens
) as well as useRxODE
's ggplot theme
(rxTheme()
). Additionally time after dose is calculated astad
for allnlmixr
models -
Tables generation has been refactored;
npde
uses thearma
and
RxODE
random number generators which may change results. Also
the default ofties=TRUE
has been changed toties=FALSE
.
npde
calculations have been threaded withOpenMP
to speed up
the calculation as well. This refactoring was required to have the
dv
imputation betweencwres
andnpde
use the same method.
Thenpde
option now calculates the decorrelatednpd
as well, (which is
the recommended weighted residual; see Nguyen 2017)
Bug Fixes
- Aligned
saem
andfocei
additive + proportional error models, so
saem
additive+proportional
outputs will be different using the
correctfocei
method
Note this includes all the RxODE changes including dropping python.