improve bootMer, profile confidence interval (cosmetics)

[bbolker]

• confint(.,"profile") [default] returns a good set of values (fixed effects, RE standard deviations and correlations, and sigma for LMMs). It should also return sigma if useSc is TRUE, and should return more informative labels (at the moment they're .sigNN).
• confint(.,"boot") has better labeling (based on the names of getME(.,"theta")), but returns theta instead of standard deviations and correlations, and doesn't return sigma.

These can't be fixed without breaking backward compatibility (although we could add a forward/backward-compatibility switch ...), and we are now in feature freeze ...

[bbolker]

maybe add attributes to bootMer/confint objects that specify how many sims and how many successful sims they embody? (It can be reconstructed from bootMer objects but not from confint objects)

[mmaechler]

Ben Bolker notifications@github.com
on Sun, 21 Jul 2013 14:01:42 -0700 writes:

> maybe add attributes to `bootMer`/`confint` objects that specify how many sims and how many *successful* sims they embody? (It can be reconstructed from `bootMer` objects but not from `confint` objects)

That seems very reasonable to me (and is back compatible).
I'm always emphasizing returning convergence "statistics"
and in the print.() and print.summary.() methods also print
that information.

Martin

[bbolker]

opinions: I could probably fix the initial stuff (not just afterthought of adding the attributes) [names and structure of profile/bootMer objects] before release. It would be breaking feature freeze but would also mean we would have less compatibility-breaking stuff to do after release. Opinions which would be the lesser evil?

[mmaechler]

Ben Bolker notifications@github.com
on Mon, 22 Jul 2013 18:47:34 -0700 writes:

> opinions: I *could* probably fix the initial stuff (not
> just afterthought of adding the attributes) [names and
> structure of profile/bootMer objects] before release.  It
> would be breaking feature freeze but would also mean we
> would have less compatibility-breaking stuff to do *after*
> release.  Opinions which would be the lesser evil?

breaking it now. Chances are pretty high that this currently
only affects less than a dozen people.

After it's on CRAN this might affect many more, notably those
who get that version of lme4 and stick to it "accidentally"
(instead of update.pacakges() regularly).

But: is "attributes" really the right way?
Without having looked at the code (which file ?),
I'd thought that the result of profile() is of S3 class
'profile' which is a list + .. and so I'd rather have list
components than attributes.

Martin

[bbolker]

To recap, there are three issues here:

• adding information about number of sims/successful sims (and reporting)
• changing the information that bootMer computes
• changing the naming of bootMer and profile objects

I propose doing the first two and implementing the third as follows:

• make getME(.,"theta") names backward compatible
• make profile() names backward compatible (i.e. a default signames=TRUE option)
• make confint(.,method="boot") names match the new standard, since probably nobody's using this at the moment anyway ...

Comments?

[bbolker]

Re previous comment: the information computed and naming have changed. I haven't added the information about number of sims/successful sims (it might?? be a good idea to allow storing the raw bootstrap results as an attribute, too ...) -- these can definitely be downgraded to "wishlist" items, as they can be done later with only minimal compatibility issues.