Skip to content

Metafor and forest(); not showing 'ilab' and text

15 messages · Marco Colagrossi, Michael Dewey, Viechtbauer Wolfgang (STAT) +1 more

#
Hello folks,

I have a couple of issues with the metafor package, specifically with
the forest graphs.
I am currently conducting a Meta-Analysis in economics throughout the
metafor package.

My meta-analysis has the specific of having different cases from
single studies, and this proven to be challenging especially when
trying to plot graphically the results I'm obtaining.

Here's the code:

forest(pc, var, ci95m, ci95p, slab = authoryear, psize=1, subset=(pub==1),
       ilab = cbind(ys, f_dim, SIMdv, SIMiv),
       ilab.xpos = c(-9.5, -8, -6, -4.5), cex = 0.75)
par(font=2)
      text(c(-9.5,-8,-6,-4.5), 26, c("Years", "Firm(s) Dimension", "DV", "IV"))
      text(-16,                26, "Author(s) and Year",     pos=4)
      text(6,                  26, "Observed Outcome [95% CI]", pos=2)
par(op)

'pc' is the 'effect size', 'var' the variance, 'ci95m & ci95p' the CI,
'pub' if the paper has been published or not. the pub subset was the
first idea I had in order to split my sample that otherwise would have
been to big. The issue with this solution is that forest() displays
only the slap argument and the forest with the confidence interval,
completely ignoring the lab argument and the text I'm trying to add.
Moreover, the graph is showed correctly only within the zoom in
Rstudio but if I save it it is showed as enclosed.

What I'm doing wrong? I tried both to look at the package
documentation and online but I can't figure it out.

Moreover, how would you suggest to handle (graphically) the
multiple-cases-per-study thing? It's a 'good' way to average the cases
among different studies in the graphs?
In my meta-analysis I'm using a multilevel model as shown in
Gelman-Hill but graphically (and in tables) I'm struggling.

Thanks for your help and patience
#
Dear Marco

Comments inline
On 24/08/2015 15:03, Marco Colagrossi wrote:
At this point I think you meant to close the call to forest with another 
) as the subsequent calls to text are further commands and not internal 
to the call of forest.
For that to have worked you probably meant to go
op <- par() somewhere earlier
Sorry, do not use Rstudio myself
Are you looking for rma.mv perhaps?

  
    
  
#
I cannot reproduce the issue with 'ilab' not being shown when using 'subset'. My guess is that the values for 'ilab.xpos' specified are actually outside of the plotting region. After you have drawn the forest plot, try:

par("usr")[1:2]

to see what the default limits actually are. Then use 'xlim' to adjust the limits to your taste. And then use appropriate values for 'ilab.xpos', so they are inside those limits.
Nothing was enclosed (or it was stripped).
Maybe add some space between groupings (i.e., studies). The example given here can provide some clues how one could go about this: http://www.metafor-project.org/doku.php/plots:forest_plot_with_subgroups But drawing a plot like this requires a lot of hand-tweaking.

Best,
Wolfgang
#
I tried to upload the file once again. I tweaked it a bit, now my code is:

forest(pc, var, ci95m, ci95p, slab = authoryear, psize=1, subset=(pub==1),
       xlim = c(-16, 6),
       ilab = cbind(SIMdv, SIMiv),
       ilab.xpos = c(-7.5, -5.5), cex = 0.75)
op <- par(cex=.75, font=2)
      text(c(-7.5, -5.5), 54, c("DV", "IV"))
      text(-16,                54, "Author(s) and Year",     pos=4)
      text(6,                  54, "Outcome [95% CI]", pos=2)
par(op)

I managed to show both the Ilab argument and the text above. I still
have 3 issues:
- now the forest plot is too narrow - that is, pretty unreadable;
- I cannot still export it properly, as shown in the enclosed .png
- SIMdv, SIMiv are shown as number while on mine .csv are actually
text variable.

regarding the rma.mv package, I set it up this way (preliminarily)

multi <- rma.mv(pc, var, random = ~ 1 | author, data=codebook)

I'm trying to compare the results with this equation, which is what -
I think, correct me if I'm wrong -  in econometrics we call
author-fixed effect, that is, model which are constant across
individuals (the random\fix notation is a bit tricky):

author_fix <- rma(pc, var, mods = ~ I(author), data=codebook, method="ML")

What I was wondering if that the two equation above mentioned also
correct for heteroskedasticity which I need since my studies have
different sample and specifications.

Thanks for your help, your patience and your time, and many
compliments for the package, is guiding me through the use of R for
the first time - as you might have guessed.

Marco


On 24 August 2015 at 16:50, Viechtbauer Wolfgang (STAT)
<wolfgang.viechtbauer at maastrichtuniversity.nl> wrote:
-------------- next part --------------
A non-text attachment was scrubbed...
Name: Rplot.png
Type: image/png
Size: 105241 bytes
Desc: not available
URL: <https://stat.ethz.ch/pipermail/r-help/attachments/20150824/7e75b138/attachment.png>
#
Hello Marco

Comments in line again
On 24/08/2015 18:49, Marco Colagrossi wrote:
You need to re-read Wolfgang's advice again. The forest function tells 
you what values of xlim it used and you can then adjust them to suit. 
This will take a few attempts in my experience.
It looked correctly exported to me. One comment, do you really need the 
complete citation of each study? Most of the forest plots I see as a 
reviewer just use the first author name and the year. This would 
potentially give you a lot more space.
I will leave this one to Wolfgang to answer.

  
    
  
#
Further comments in line as well.
Also, if you have lots of outcomes, you may need to increase the height of the plotting device to make everything fit (or you need to reduce the font size even further, but things will become illegible eventually).
Those variables are apparently coded as factors, so use data.frame() instead of cbind() to avoid the coercion to integer codes.
I cannot comment on model choices. But yes, the functions properly account for the fact that the sampling variances are heteroskedastic.
#
Thanks again for your help. I'm sorry to bother you but I don't get
how to widen the forest plot; if I try to change the values of xlim or
the ilab.xpos values the width of the forest plot region does not
change, but only moves on the graphs. What I'm I missing?


forest(pc, var, ci95m, ci95p, slab = authoryear, psize=1, subset=(pub==1),
       xlim = c(-16, 6),
       ilab = data.frame(SIMdv, SIMiv),
       ilab.xpos = c(-7.5, -5.5), cex = 0.75)
op <- par(cex=.75, font=2)
      text(c(-7.5, -5.5), 54, c("DV", "IV"))
      text(-16,                54, "Author(s) and Year",     pos=4)
      text(6,                  54, "Outcome [95% CI]", pos=2)
par(op)
[1] -16   6

On 25 August 2015 at 15:54, Viechtbauer Wolfgang (STAT)
<wolfgang.viechtbauer at maastrichtuniversity.nl> wrote:
#
The 'xlim' argument does not change the actual width of the plotting device. For that, you need to use the 'width' argument with whatever device you are actually using. You can then use the 'xlim' argument to create appropriate spacing to the left/right of the part of the plot that shows the estimates and their CIs. Within that space, you can then add additional columns with the 'ilab' argument. It's up to you to find an appropriate combination of plotting device width, character/symbol expansion factor ('cex' argument), 'xlim' values, and 'ilab.xpos' values to create a nice looking plot that has no overlapping text and no excessive white space. An example is this:

http://www.metafor-project.org/doku.php/plots:forest_plot_with_subgroups 

Note that it took me dozens of iterations to create that plot. You just have to start experimenting.

Best,
Wolfgang
#
I think I've not explained myself well. When I say "the width of the
forest plot" I mean the region above the observed outcome, the
"actual" forest plot, not the plot as a whole. Even if I change values
for Xlim, cex or ilab.xpos the width of that particular region within
the plot doesn't change.

Best,

Marco

On 25 August 2015 at 18:11, Viechtbauer Wolfgang (STAT)
<wolfgang.viechtbauer at maastrichtuniversity.nl> wrote:
-------------- next part --------------
A non-text attachment was scrubbed...
Name: Rplot01.png
Type: image/png
Size: 119435 bytes
Desc: not available
URL: <https://stat.ethz.ch/pipermail/r-help/attachments/20150825/5b8eac03/attachment.png>
#
Dear Marco

When you change xlim it increases the width of the forest plot in the 
sense you describe. It does not push your text out of the way to make 
space for it but instead overprints it. You may like to use alim to 
truncate your confidence interval whiskers to fit within the space you 
see or make your labels shorter.
On 25/08/2015 17:25, Marco Colagrossi wrote:

  
    
  
5 days later
#
Thanks for your help,

I got the mistake I was making and I managed to find a solution
regarding those graphs; I don't want to abuse of your patience but I
have three further questions:

1. Always regarding the forest plots, it is possible to make a
cross-subset? I try to explain my self better; I have one dummy
variable called pub and another variable called SIMiv that can take
the values of "share", "loan", "number" and "duration". How can I
subset my sample so that the forest shows only (for example) studies
when the dummy takes the value of 1 and the SIMiv variable takes the
values of "share" and "loan"?
Something like this:
forest(pc, var, ci95m, ci95p, slab = authoryear2, psize=1,
subset=(pub==1, SIMiv=("share", "loan", "duration"))

2. I have few doubts regarding the multilevel modeling;
    rma.mv(pc, var, random = ~ 1 | author, data=codebook)
   if I'm correct this should be a multilevel model nested at "author"
level; what I cannot understand If it is a varying intercept
(Y=A+BjX), a varying slope (Y=Aj+BX) or a varying intercept&slope
model (Y=Aj+BjX). Are there the formulas for it somewhere? So far I
only found the formulas for the estimators included in the metafor
package.

3. metareg1 <- rma.mv(pc, var, random = ~ 1 | author, mods = ~ pub +
SIMiv, data=codebook)
Again, if I'm correct this should be a multilevel meta regression
(correct me if I'm wrong); I have the same doubts as before.

Thank you again

Marco
On 25 August 2015 at 19:24, Michael Dewey <lists at dewey.myzen.co.uk> wrote:
#
Comments in line
On 31/08/2015 16:08, Marco Colagrossi wrote:
Do you not want something like
(pub == 1) & (SIMIv %in% c("share", "loan", "duration"))
I think it a random intercept but Wolfgang may correct me there.

  
    
  
#
The solution that you proposed works perfectly, thank you very much.

I'll wait for Wolfgang answer as I'm having few doubts about the models.

Thanks
On 31 August 2015 at 18:34, Michael Dewey <lists at dewey.myzen.co.uk> wrote:
#
Have you read help(rma.mv)? It describes in detail what "random = ~ 1 | author" does. Also, I think you may find some of these useful:

http://www.metafor-project.org/doku.php/analyses#multivariate_multilevel_meta-analysis_models

Especially: http://www.metafor-project.org/doku.php/analyses:konstantopoulos2011

Using "random = ~ 1 | author" is likely to be insufficient. You also need to add random effects at the observation level.

Best,
Wolfgang
#
Dear Wolfgang,

Kindly please i have an issue with R code could you please help me.

Best Regards

On Mon, Aug 31, 2015 at 6:24 PM, Viechtbauer Wolfgang (STAT)-2 [via R] <
ml-node+s789695n4711682h19 at n4.nabble.com> wrote:

            
--
View this message in context: http://r.789695.n4.nabble.com/Metafor-and-forest-not-showing-ilab-and-text-tp4711432p4711683.html
Sent from the R help mailing list archive at Nabble.com.