R help - Aug 2008 - Lattice: problem using panel.superpose and panel.groups

Hi.  I'm embarking on my first attempt at creating my own panel  
function for lattice graphics, and despite all of my online research  
and pouring through the documentation, I cannot figure out how to  
solve my particular problem.  Hopefully, a generous fellow R user can  
help.

I have some data that is split into two groups:  some "actual" data,  
and some simulated data, generated from several different models.  The  
actual data come from two different datasets (calibration and  
holdout), and the simulations were calibrated on each data set under  
the various models.

What I want to do is create a boxplot on the simulated data, and  
superimpose a line representing the actual data.  This plot would  
condition on dataset and model.  While the simulated data various  
across model and dataset, the actual data varies only across dataset  
and is common for all models.

My approach was to use panel.superpose, and create a panel function  
that, depending on the group.number, calls either panel.bwplot or  
panel.lines.  Like this:

panel.custom.plot <- function (..., group.number) {

	if (group.number == 1) {
		panel.bwplot(...)
	} else {
		panel.lines(...)
	}
}

obj <- bwplot(p ~ as.factor(count) | dataset + model, data = data.frame,
	panel = panel.superpose,
	groups = group,
        panel.groups = panel.custom.plot
}


When I do this, I see the bwplot, but I cannot see the lines.  It is  
as though the data isn't being passed through.

Can someone help me get started with this, and instruct me a little  
bit more on how arguments are passed from trellis functions to the  
panel functions?

I did consider a second approach, where I don't use panel.superpose or  
groups at all, but pass the actual.calib and actual.hold datasets as  
parameters to a panel function.  This panel function would call  
panel.bwplot, and then, depending on the value of dataset, add lines  
for either actual.calib or actual.hold.  But I do not know how to let  
the panel function know what level of a conditioning variable is being  
plotted in a particular panel.

Thanks in advance for any help you can provide,

Michael



Michael Braun
Homer A. Burnell Career Development Professor
	and Assistant Professor of Management Science (Marketing Group)
MIT Sloan School of Management
One Amherst St., E40-169
Cambridge, MA 02139
(617) 253-3436
braunm at mit.edu

Michael Braun <braunm <at> MIT.EDU> writes:
> 
> I have some data that is split into two groups:  some "actual"
data,
> and some simulated data, generated from several different models.  The  
> actual data come from two different datasets (calibration and  
> holdout), and the simulations were calibrated on each data set under  
> the various models.
> 
> What I want to do is create a boxplot on the simulated data, and  
> superimpose a line representing the actual data.  This plot would  
> condition on dataset and model.  While the simulated data various  
> across model and dataset, the actual data varies only across dataset  
> and is common for all models.
> 
Michael,

to understand you problem, I started creating a data set that simulates the
data, but I got stuck with with your intended meaning of "groups.

df =   
data.frame(p=rnorm(100),count=rpois(100,5),dataset=c("A","B"),
  model=rep(c("M1","M2"),each=50),groups= dont' know)

Could you please (as the posting guide says) supply a self-running example that
shows the problem? 

You named your data.frame "data.frame"; it could work, but better
don't use
common things like "c" or "data.frame" as variable name. 

Dieter

Dieter:

Thank you for your response.  As you requested, I created a self- 
running example, pasted below.  It may be a little wordier than I  
would like, but it runs.
The data I am creating assumes that there are two datasets ("calib"  
and "hold"), with actual values actual.calib and actual.hold.  I then
simulate data for three models, A, B and C.  I want to plot a 3 x 2  
lattice that has bwplots for each model-dataset combination in each  
panel.  On top of that, I want to superimpose a line that plots the  
actual data for calib in all three panels in the calib column (and the  
same for hold).

The data.frame sim.data is just the simulated data, and actual.data is  
just the actual data.  all.data combines them both, and the group  
variable identifies whether a record is from "sim" or
"actual."  Note
that there are 50 simulations, but only one actual, for each model- 
dataset combination.  The presentation in all.data is kind of  
wasteful, since I replicated the same actual data for all 3 models.   
So, ideally, I would work with sim.data and actual.data separately.   
But I can use all.data if necessary.

I apologize that I was not more claear in my earlier post, but  
hopefully this gives you enough information.  Thanks again for helping  
with this.

Michael



library(lattice)

# creating datasets


models <- c("A","B","C")
datasets <- c("calib","hold")
counts <- 1:10
n.sims <- 50
sim.data <- NULL
all.data <- NULL
actual.calib <- seq(0,5,length=10)
actual.hold <- seq(5,0,length=10)
cols <-
c("model","dataset","count","value","group")

actual.data <-
as.data.frame(cbind(c(rep("calib",10),rep("hold",10)),
                                     
c(1:10,1:10),c(actual.calib,actual.hold)))

set.names <- c(rep("calib",n.sims),rep("hold",n.sims))

for (mod in models) {
   for (y in counts) {

     x.calib <- rnorm(n=n.sims,mean=actual.calib[y],sd=1)
     x.hold <- rnorm(n=n.sims,mean=actual.hold[y],sd=1)

     x <- as.data.frame(cbind(rep(mod,2*n.sims), set.names, y,
                              c(x.calib,x.hold),rep("sim",2*n.sims)))

     colnames(x) <- cols
     sim.data <- rbind(sim.data, x)
   }

   rep.actual <- as.data.frame(cbind(rep(mod,20),
                                     actual.data,rep("actual",20)))
   colnames(rep.actual) <- cols
   all.data <- rbind(all.data,sim.data,rep.actual)
}

colnames(actual.data) <-
c("dataset","count","value")

#  for some reason, value is encoded as a factor.
#  This changes value to numeric.
#  (is there a better way to do this?)

sim.data[,"value"] <-
as.numeric(levels(sim.data[,"value"])
[sim.data[,"value"]])
all.data[,"value"] <-
as.numeric(levels(all.data[,"value"])
[all.data[,"value"]])
actual.data[,"value"] <-
as.numeric(levels(actual.data[,"value"])
[actual.data[,"value"]])


# at this point, there are three data frames worth considering:
# sim.data - simulated data
# actual.data - actual data
# all.data - sim.data, and the actual.data replicated for each model


# create panel function

panel.ppc.plot <- function(...,group.number) {

   if (group.number==1) {
     panel.bwplot(...)
   } else {

     panel.lines(...)
   }
}

# create and plot lattice object

obj <- bwplot(as.numeric(value) ~ as.factor(count) | dataset + model,
               data = all.data,
               panel = panel.superpose,
               groups=group,
               panel.groups = panel.ppc.plot
               )

plot(obj)
>
>
>
> ------------------------------
>
> Message: 119
> Date: Sat, 16 Aug 2008 08:14:36 +0000 (UTC)
> From: Dieter Menne <dieter.menne at menne-biomed.de>
> Subject: Re: [R] Lattice:  problem using panel.superpose and
> 	panel.groups
> To: r-help at stat.math.ethz.ch
> Message-ID: <loom.20080816T080831-746 at post.gmane.org>
> Content-Type: text/plain; charset=us-ascii
>
> Michael Braun <braunm <at> MIT.EDU> writes:
>
>>
>> I have some data that is split into two groups:  some
"actual" data,
>> and some simulated data, generated from several different models.   
>> The
>> actual data come from two different datasets (calibration and
>> holdout), and the simulations were calibrated on each data set under
>> the various models.
>>
>> What I want to do is create a boxplot on the simulated data, and
>> superimpose a line representing the actual data.  This plot would
>> condition on dataset and model.  While the simulated data various
>> across model and dataset, the actual data varies only across dataset
>> and is common for all models.
>>
> Michael,
>
> to understand you problem, I started creating a data set that  
> simulates the
> data, but I got stuck with with your intended meaning of "groups.
>
> df =   
data.frame(p=rnorm(100),count=rpois(100,5),dataset=c("A","B"),
>  model=rep(c("M1","M2"),each=50),groups= dont'
know)
>
> Could you please (as the posting guide says) supply a self-running  
> example that
> shows the problem?
>
> You named your data.frame "data.frame"; it could work, but better
> don't use
> common things like "c" or "data.frame" as variable
name.
>
> Dieter
>
>

Michael Braun <braunm <at> MIT.EDU> writes:
> 
> Dieter:
> 
> Thank you for your response.  As you requested, I created a self- 
> running example, pasted below.  It may be a little wordier than I  
> would like, but it runs.
.. Details removed
> 
> panel.ppc.plot <- function(...,group.number) {
> 
>    if (group.number==1) {
>      panel.bwplot(...)
>    } else {
> 
>      panel.lines(...)
>    }
> }
Trellis graphics are a bit like hash functions: you can be close to the 
target, but get a far-off result. I admit that I do not know why 
panel.lines does not work in the above example, but  

panel.polygon(...)

works in your special case of ordered data. More generally, I would suggest 
to use

panel.xyplot(x,y,type="l")

or, if you want the ... notation, 

panel.xyplot(...)

but then you have to set type="l" in your bwplot calling function.

Dieter

Le dimanche 17 ao?t 2008 ? 09:36 +0000, Dieter Menne a ?crit :

[ Snip .. ]
> Trellis graphics are a bit like hash functions: you can be close to the 
> target, but get a far-off result.
Nice candidate for a fortune() entry ...

					Emmanuel Charpentier