[R-sig-ME] Questions about porting pedigreemm to the new lme4

[Joehanes, Roby (NIH/NHLBI) [F]]

Hi all:

Just a follow up: It seems that some changes need to be made in the new lme4 so that the pedigreemm can be modified to use it. They are:
1. To allow the return of all necessary internal variables required for invoking the optimizer (especially rho plus devfun in one call), without invoking the optimizer itself. Right now I did two calls, the first to get the devfun, the second to get the rest. However, I cannot obtain rho in all cases.
#lmerc$doFit <- FALSE # call lmer without pedigree and with doFit = FALSE
l... <- list(...)
lmerc$devFunOnly <- TRUE
devFun <- eval(lmerc, parent.frame())
lmerc$devFunOnly <- FALSE
lmf <- eval(lmerc, parent.frame())

2. Some partitioning of Zt that allows duplicate level names for multiple random factors. That is, if random factors 1 and 2 (let's call them RF1 and RF2) share the same level-string, my current hack will not work. Perhaps someone can propose a better way. My current hack is as follows:
Zt <- getME(lmf, "Zt")
Zt <- Zt[rownames(Zt) %in% levels(getME(lmf, "flist")[[tn]]), ]
(The original was: Zt <- lmf$FL$trms[[tn]]$Zt)

Since I cannot get rho, I cannot realize the lmer_finalize or glmer_finalize. Any help is appreciated. I am stuck right now since I have installed the lme4 SVN version, but the pedigreemm package still assumes the CRAN version.

Thank you,
Roby


On Apr 2, 2012, at 4:24 PM, Joehanes, Roby (NIH/NHLBI) [F] wrote:

Hi Ben and Andrzej:

Thank you for the info. I am trying to get pedigreemm to work with the new lme4 and I am now stuck on how to translate the following line:
lmf$FL$trms[[tn]]$Zt <- lmf$FL$trms[[tn]]$A <- relfac[[i]] %*% Zt

Because the matrix A is defined in getME method as:
PR$Lambdat %*% PR$Zt

I am still trying to figure the code out. I think we can do away with the A matrix since it is apparently not essential to the computation. I think the crux of the pedigreemm call to the underlying lme4 is in the following short lines:

   pnms <- names(pedigree)
   stopifnot(all(pnms %in% names(getME(lmf, "flist"))))
   asgn <- attr(getME(lmf, "flist"), "assign")
   for (i in seq_along(pedigree)) {
       tn <- which(match(pnms[i], names(getME(lmf, "flist"))) == asgn)
       if (length(tn) > 1)
           stop("a pedigree factor must be associated with only one r.e. term")
       Zt <- getME(lmf, "Zt")
       relfac[[i]] <- relfactor(pedigree[[i]], rownames(Zt))
       lmf$FL$trms[[tn]]$Zt <- lmf$FL$trms[[tn]]$A <- relfac[[i]] %*% Zt
   }
   ans <- do.call(if (!is.null(lmf$glmFit)) lme4:::glmer_finalize else lme4:::lmer_finalize, lmf)
   ans <- new("pedigreemm", relfac = relfac, ans)

Any help is appreciated.

Thanks,
Roby


On Apr 2, 2012, at 4:17 PM, Ben Bolker wrote:

Andrzej T Galecki 1 <agalecki at ...> writes:

Ben prepared getME(} function to extract various components of the model
fit
obtained using lme4.0 and new lme4.

Try
getME(lmf,"Zt")
getME(lmf,"A")
getME(lmf,"flist")

Andrzej

On Mon, 2 Apr 2012 14:27:20 -0400, "Joehanes, Roby (NIH/NHLBI) [F]"
<roby.joehanes at ...> wrote:
Hi:

I am trying to port pedigreemm to use the new lme4. However, the new lme4
is somewhat opaque to me and hence I need help. Specifically, how can I
extract the following terms:
lmf$FL$trms[[tn]]$Zt
lmf$FL$trms[[tn]]$A

The information lies within lmf <at> pp object, but it is too opaque to me to
fish anything out.

I understand that lmf$FL$fl is lmf <at> flist and that lmer_finalize calls
should be replaced with optwrap and mkMerMod. Nevertheless, I think that
the lmer or glmer could use some refactoring to ease the calls.

I appreciate any help on these.


 A couple of questions:

1. how broadly are you using A?  Off-list, Doug Bates has commented to
me that defining A is relatively straightforward for LMMs but not necessarily
as simple (nor uniquely defined!) for [GN]LMMs -- for the time being, I
am quite likely to add code that *disallows* getME(.,"A") except for LMMs --
I can imagine that the bulk of pedigreemm use is for LMMs, but I suppose
that people are using it for [GN]LMMs as welll ...

2. I can't quite tell from your e-mail -- would it be useful to have
the stuff between the first optwrap() and the last mkMerMod abstracted
into a single (exposed) function?  For [GN]LMMs it should probably
also allow control of whether the preliminary "nAGQ=0" optimization
is done or not ...

(All comments are mine alone, not necessarily speaking for all lme4
authors ...)

Ben Bolker

_______________________________________________
R-sig-mixed-models at r-project.org<mailto:R-sig-mixed-models at r-project.org> mailing list
https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models