Open psolymos opened 7 years ago
Write tmpfile for model directly to avoid multiple thread clashes.
Deal with possibly different species subsets in object
and ynew
:
ynew
but not in object
: dropobject
but not in ynew
: 0 value is information, thus setting it to 0 is not correct. Need to subset JAGS input data to contain the intersect of species.This is best achieved by a subset method for the opticut/uncertainty model classes: carefully subset species (modifying other attributes etc. as needed). This is useful and safe option.
dcoptions
used to avoid overwriting model file when running in parallel -- needs testingsubset
method added and used in ipredict.opticut
multicut
) modeled after opticut
and optilevels
(all dist, because it is already used in rankComb
)ipredict.default
for multicut
objects (binomial and poisson cases only).\dontrun
or implement fast version:
predict
method for opticut and multicut objects (with args: object, xnew, gnew; all dist types)predict
method for multicut objects (with args: object, xnew, gnew; all dist types)ipredict
: analytic should implement additive loglik (binomial and poisson cases only)Needed methods for multicut class:
Not sure if uncertainty and related methods need to be implemented as well. It is not needed for ipredict, stratification is fixed, and bootstrap is kind of trivial in that case.
Need to test all the dist cases (compare coef with getMLE output)
Most of the prep work done for milestone v1.0. Here is the list of remaining items for milestone v1.1:
ipredict
methods for opticut and multicut objects implementing IP.ipower
referring to [posterior] indicator power).Need to figure out final object structures, print/summary/plot methods, write documentation and tests. Otherwise the multiclass branch has the new functionality in /extras
.
kappa stat needs to be fully fledged out, also considering prior weights.
Create new branch (
inverse-prediction
) and implement the following functions:ipredict
generic and methods (default, opticut): currently calledcalibrate
, butipredict
should better reflect the intention. This will only run a single instance with any ynew/xnew arguments.calibrate
method to implement LOO. Implement comparison based on multiclass measures similarly to AIC but meybe only for 2 objects: LOTO (leave one taxon out) vs. reference (all taxa LOO)