[R-sig-ME] plot log transformed variable with Effects package

[Ben Bolker]

Ben Bolker <bbolker at ...> writes:

> 
>  <John.Morrongiello <at> ...> writes:
> 
> > 
> > Hi list
> > 
> > I'm having trouble plotting a mixed model that includes log transformed
> (or any transformed) terms.
> > 
> library(lme4)
> library(effects)
> data(cake, package="lme4")
> > 
> > ###this works
> fm1 <- lmer(angle ~ recipe * temperature
>  + (1|recipe:replicate), 
>   cake, REML = FALSE)
> 
> plot(Effect(c('recipe','temperature'), fm1))
> ##but this doesn't (log transformed angle)
> 
> fm2 <- lmer(log(angle) ~ recipe * temperature + (1|recipe:replicate), 
>  cake, REML = FALSE)
> 
> plot(Effect(c('recipe','temperature'), fm2))
>   This is probably something to take up with the maintainer of
> the 'effects' package, who in turn might have to consult the lme4
> maintainers.  The proximal problem occurs in 
> 
>   plot -> Effect -> Effect.merMod -> Effect.mer -> Effect ->
> mer.to.glm
> 
>   There is a 'data' object that appears to be coming from the
> model.frame() of the original object, but I haven't tracked its
> source down yet -- but the problem is that it has log(angle) rather
> than angle as a column ...
> 
> 

  with a bit more work, I have come up with a way to
hack the effects package to make this work.  The function below
needs to be defined, then the arguments of the effects:::mer.to.glm
function have to be changed from

function(mod, data=model.frame(mod))

to

function(mod, data=xdata(mod))

You can hack the package yourself and/or request the maintainer
to add this capability ...
                                    
## modeled after stats::expand.model.frame
## expand the model frame to include any variables present
## in the original 'data' object but missing from the model frame
## potentially fragile:
##  * depends on 'data' still being present in the original environment
##  * doesn't check for any potential mishaps
xdata <- function(model, envir=environment(formula(model))) {
    fr <- model.frame(model)
    data <- eval(getCall(model)$data,envir)
    ## find missing variables
    newvars <- setdiff(all.vars(formula(model)),names(fr))
    if (length(newvars)>0) {
        fr <- data.frame(fr,data[newvars],check.names=FALSE)
    }
    fr
}