I might eventually think that maxLik or even optimx might deserve to be
a suggests as well as only one or the other is really needed, but I use
them both all the time, so probably not.
ML.all estimates all the different pfuncs available on a full dataset or
one split by a factor in the dataset. ufuncs will still
have to be passed as an argument, as well as with/without context. This
function brings in one of my other packages "ctools" to implement
parallel estimation of the split dataset. If you're not using ctools,
check it out, its super useful for these kind of situations.
This function takes a list which contains a list of estimated models.
The main use of this function will be for a yet written, ML.all
function which will split a dataset by a factor and estimate all models
on the split data. This ML.dataframe function will allow for all this
data to be easily condenced into a data.frame object. Many users might
find this a useful way to share the results of estimates with users of
Stata and other statistics softwares.
Some people don't use contextual utility. I suppose at some point I'll
incorporate more exotic stochastic functions, and probably utility
functions as well.
Noticed that some insturments were bombing out with errors saying that
the function couldn't even evaluate once. Did some investigating, found
that uvec indicies were for some reason not grabbing the instrument
unless thew were transposed, but also the choice vectors stuff wasn't
pulling in the right choice data. Not by a long shot. Found that this
could be corrected by actually using the c_index vector that had been
defined but unused, I don't know why it wasn't. So much, Do not use the
immediately previous commit.
If you want to revert behind this commit, use 966dfaf if you need MSL
to estimate properly. Note this won't give you the non-linear transforms
from nlWaldTest.
The basepars were essentially being replaced by the last dem * the
last dempar. This totally erased the effect of the main par. This was
replace so dem*dempar is added to basepar.
Also kkt2 was decreased to make it harder to converge.
These functions are all non-linear, but I've been using linerar
transformations to recover standard errors from the hessian. Woops...
The nlWaldTest package provides a way of retrieving the non-linear
confidence intervals, from which the SEs can be backed out. It also
provides a function to test parameters against critical values, which
is tremendously useful when trying to classify a subject as EUT or RDU.
options() can be set for arguments that already have a default. E.g.
options(MSL.HH = 150) can be set to have all MSL functions use 150
halton draws instead of the default 100.
Removing Covar-version 1) because I never use it, 2) becuase it allows
me to use the same setup function I use with the RDU functions, which
never had covar-versions. This drops a fair amount of code and creates
very similar MSL functions, so things are now easier to maintain.