swh:1:snp:13ceaf33a0cf03b54a9c83ec2c853f3c4226f6da
Tip revision: 1fe1daaec4e0e4abab833a89e3d7729279ad0c18 authored by Ben Bolker on 16 January 2012, 00:00:00 UTC
version 1.0.4.1
version 1.0.4.1
Tip revision: 1fe1daa
TODO
BUGS/ISSUES:
* change the way 'better fit found' is handled in profiles.
with try_harder, push through anyway ... otherwise stop
WITH AN ERROR and give the user a utility function for how to proceed?
*Don't* want to return an object with a different structure --
maybe attributes?
* is there a way to (optionally?) save the environment of the
call so that fits could be profiled after being removed from
their original environments? (maybe fixed?)
* consider "data-absent" flag for big data sets?
* remove "quad" method, replace with confint.default [NO]
* move calcslice from emdbook and make it more robust/general
(different signatures: mle2 fits, numeric vectors)
* prettier multi-parameter profiles a la Bates/lme4
* ggplot2 profiles?
* fix confint plot to use linear interpolation when non-monotonic (done)
* pass parameters through from confint to profile (done?)
* nobs() methods
* filter use.ginv properly
* fix gradient with profile
* work on vignette: ask MM/PD/BDR about mismatch between confint (approx) and profile plot (backspline) ... ?
* DISABLE boundary warning when profiling ...
* try to do something about rescaling when hessian is problematic?
* add ginv for problematic hessians? (done)
* implement 'subset' argument
* check problems with optimizer="optimize" in tests/parscale.R
* allow ICtab etc. to recognize optim(), optimize() fits
(ASSUMING that the function was a NLL)
* add optimizer() as alternative optimizer [DONE]
* fix par mfrow resetting glitch on plotting profile
* prettier profile plots (with lattice)? xyplot for profiles?
* make sure numeric deriv modifications are working
* incorporate optimx (done)? minpack.lm?
* proper initialization of intercept-less parameter()
entries
* plot methods/fortify, a la ggplot2?
* add deviance() method [need S3-to-S4 conversion]
* make sure subset arg is really working!
* spurious error on mismatched parameter names
* spurious warnings in 1-parameter conf int: [FIXED]
library(bbmle)
m1 <- mle2(10~dbinom(prob=p,size=15),start=list(p=0.67))
c1 <- confint(m1)
* do one-parameter profiles with optimize??
* use numDeriv library hessian() function instead of nlme::fdHess?
(or use nlme::fdHess to avoid loading whole package?)
[DONE]
* turn off Hessian calculation for profile fits??
[maybe DONE by virtue of previous fix]
* should print warning immediately if convergence fails
* some weird stuff with returned fit from found-better-fit profile --
treating profiled value as fixed ...
* replace approx() in confint() with backspline?
general solution for non-monotonic profiles?
BUG: order of parameters matters for L-BFGS-B
(fixed)
adjusting parameter vectors for lower, upper, parscale, ...
when some params are fixed ...
sort out names BS -- when can we safely remove names?
TO DO:
model-averaging?
more documentation -- especially S4 methods!
especially:
profile
plot profile
confint
catch/interpret more error messages?
(try to filter last.warning?)
add DIC to IC tabs?
lmer?
WISHLIST:
start as FUNCTION (i.e., self-start)
analytic derivatives
relist
subset
plot.predict
drop1, add1, etc.
link functions ("identity","log", "logit", etc.)
delta method standard error calcs
tranformations on LHS of formula (i.e. use
link rather than inverse-link function? only
possible if link is known and invertible:
inverse
log
logit (qlogis)
probit (qnorm)
etc.
clean up/argue about data handling: closures etc. etc. etc...
document argument handling:
start must be a named vector or a named list
[OR?? inherit from parnames(minuslogl)?]
if start is not a list (i.e. a numeric vector)
set vecpar TRUE
convert start to a list
if missing and default.start is TRUE use formals(minuslogl)