[Lme4-commits] r1690 - pkg/lme4/R

[Douglas Bates]

On Thu, Mar 22, 2012 at 2:08 PM,  <noreply at r-forge.wu-wien.ac.at> wrote:
> Author: bbolker
> Date: 2012-03-22 20:08:47 +0100 (Thu, 22 Mar 2012)
> New Revision: 1690
>
> Modified:
>   pkg/lme4/R/lmer.R
> Log:
>
>   fix typo/bug (still passing X to mkRespMod, even though that's commented out currently)
>   if all(par==0) and optimizer=="bobyqa", push initial parameters away from the boundary
>         (fix for problem with refitML() in SASmixed)
>
>
> Modified: pkg/lme4/R/lmer.R
> ===================================================================
> --- pkg/lme4/R/lmer.R   2012-03-22 18:49:37 UTC (rev 1689)
> +++ pkg/lme4/R/lmer.R   2012-03-22 19:08:47 UTC (rev 1690)
> @@ -321,7 +321,7 @@
>     parent.env(rho) <- parent.frame()
>     rho$pp <- do.call(merPredD$new, c(reTrms[c("Zt","theta","Lambdat","Lind")], n=nrow(X), list(X=X)))
>                                        # response module
> -    rho$resp <- mkRespMod(fr, family=family)
> +    rho$resp <- mkRespMod(fr, family=family) ## , X=X)
>                                        # initial step from working response
>     if (compDev) {
>        .Call(glmerWrkIter, rho$pp$ptr(), rho$resp$ptr())
> @@ -1977,6 +1977,7 @@
>              if (is.null(control$xt)) control$xt <- control$xst*5e-4
>              if (!is.null(rho)) rho$control <- control
>          })
> +    if (optimizer=="bobyqa" && all(par==0)) par[] <- 0.001  ## minor kluge

In the longer term the appropriate check is something like any(par == lower)

>     arglist <- list(fn=fn, par=par, lower=lower, upper=upper, control=control)
>   ## optimx: must pass method in control (?) because 'method' was previously
>   ## used in lme4 to specify REML vs ML
>
> _______________________________________________
> Lme4-commits mailing list
> Lme4-commits at lists.r-forge.r-project.org
> https://lists.r-forge.r-project.org/cgi-bin/mailman/listinfo/lme4-commits