Why is FLSD divided by two in ezPlot?

[Viktoria]

Hi!

Thanks a lot for a great package! It's extremely helpful.

I have a quick question. Looking through the code I noticed that the
calculated FLSD is divided by two when plotting the bars around
points. Was your intention to make the bars 1 SD around the mean (a
less conventional way) by default?

Thanks a lot!

--
Victoria

[Mike Lawrence]

With a half FLSD on either side, when the bars for two points overlap,
the points would be considered non-different and when they fail to
overlap they would be considered different.

[Viktoria]

Hi Mike,

Thank you very much for your prompt response!

It tried to specify the vector of alternative errors, but it seems
that it only takes in one number. Is there a special syntax for that?

I wonder if there is a reference for the statement about FLSD and
overlap?

Again. thanks a lot!

--
Victoria

[Thom]

I've written about extending this approach to within-subjects CI paper
in a BRM paper
and my book. A quick summary of blog posts about it is here.

http://psychologicalstatistics.blogspot.co.uk/2012/03/graphing-between-subject-confidence.html

I'd forgotten that you'd included FLSDs in ez. When I get time I'll
update the book blog post.

It should be fairly easy to add the within-subject CIs I propose to
ez ...

Thom

[Mike Lawrence]

Frankly, I'm reluctant to do much regarding implementing new within-Ss
error bars in ez because it's a tricky endeavour (e.g. Loftus & Masson
got it wrong, Cousineau got it wrong, etc) and I feel that this whole
issue is easily obviated by moving to mixed effects modelling (where
you also get a more powerful, more robust and more nuanced inferential
tool to boot) and generating model predictions after setting the
intercept variance to zero (as is the default behaviour of ezPredict).
The latter approach usually creates 95% CIs on the raw data that yield
inferences that should generally correspond with those one would make
if presented with the proper 95% CI on the difference score. Of
course, this is only for parsimony of presentation, and one should
always compute the difference score to ensure the inference is sound.

[Mike Lawrence]

That said, Thom, the source for ez is hosted on github, so if you
could always branch, implement your methods and submit a pull request,
in which case I'd add a note in the ezPlot/ezStats documentation
regarding your contribution.

[Henrik Singmann]

Hi guys,

I just want to add my two cents to the debate. Volker Franz who is a new professor at my old university Hamburg recently published a correction/update to the Loftus & Masson within-subject confidence intervals paper together with Loftus:

Franz, V., & Loftus, G. (2012). Standard errors and confidence intervals in within-subjects designs: Generalizing Loftus and Masson (1994) and avoiding the biases of alternative accounts. Psychonomic Bulletin & Review, 1–10. doi:10.3758/s13423-012-0230-1

You can get it here: http://webapp6.rrz.uni-hamburg.de/allpsy/vf/lit/2012_Franz_Loftus.pdf

Furthermore, my function that avoids plotting error bars but indeed plots the complete raw data in the background, called raw means plots, is now part of the plotrix package on CRAN (thanks to Jim Lemon). A short introduction/example can be found on my homepage: http://www.psychologie.uni-freiburg.de/Members/singmann/R/rm.plot
Note that the function was renamed to raw.means.plot and is (as the plotrix update was uploaded to CRAN yesterday and is not yet available) also available from r-forge: install.packages("raw.means.plot", repos="http://R-Forge.R-project.org")

Best,
Henrik

2012/4/29 Mike Lawrence <Mike.L...@dal.ca>

[Thom]

I like the idea of plotting raw data along with other plots. More
generally, there
isn't a single solution to plotting WSCIs because there are nearly
always
several quantities of interest. For accuracy you could plot the CIs
for the
differences directly - but this can be confusing with many means and
obscures
patterns among the means themselves (e.g., trends or changes in
variance).

I would sacrifice accuracy to get a plot that is sufficiently accurate
for
exploratory work and/or summarizing the main patterns among the
differences.

Thom

On May 1, 1:33 pm, Henrik Singmann <singm...@googlemail.com> wrote:
> Hi guys,
>
> I just want to add my two cents to the debate. Volker Franz who is a new
> professor at my old university Hamburg recently published a
> correction/update to the Loftus & Masson within-subject confidence
> intervals paper together with Loftus:
>
> Franz, V., & Loftus, G. (2012). Standard errors and confidence intervals in
> within-subjects designs: Generalizing Loftus and Masson (1994) and avoiding
> the biases of alternative accounts. Psychonomic Bulletin & Review, 1–10.
> doi:10.3758/s13423-012-0230-1
>
> You can get it here:http://webapp6.rrz.uni-hamburg.de/allpsy/vf/lit/2012_Franz_Loftus.pdf
>
> Furthermore, my function that avoids plotting error bars but indeed plots
> the complete raw data in the background, called raw means plots, is now
> part of the plotrix package on CRAN (thanks to Jim Lemon). A short
> introduction/example can be found on my homepage:http://www.psychologie.uni-freiburg.de/Members/singmann/R/rm.plot
> Note that the function was renamed to raw.means.plot and is (as the plotrix
> update was uploaded to CRAN yesterday and is not yet available) also

> available from r-forge: * install.packages("raw.means.plot", repos="http://R-Forge.R-project.org")
> *
> Best,
> Henrik
>
> 2012/4/29 Mike Lawrence <Mike.Lawre...@dal.ca>

>
>
>
>
>
>
>
> > That said, Thom, the source for ez is hosted on github, so if you
> > could always branch, implement your methods and submit a pull request,
> > in which case I'd add a note in the ezPlot/ezStats documentation
> > regarding your contribution.
>

> > On Sun, Apr 29, 2012 at 12:28 PM, Mike Lawrence <Mike.Lawre...@dal.ca>

> > wrote:
> > > Frankly, I'm reluctant to do much regarding implementing new within-Ss
> > > error bars in ez because it's a tricky endeavour (e.g. Loftus & Masson
> > > got it wrong, Cousineau got it wrong, etc) and I feel that this whole
> > > issue is easily obviated by moving to mixed effects modelling (where
> > > you also get a more powerful, more robust and more nuanced inferential
> > > tool to boot) and generating model predictions after setting the
> > > intercept variance to zero (as is the default behaviour of ezPredict).
> > > The latter approach usually creates 95% CIs on the raw data that yield
> > > inferences that should generally correspond with those one would make
> > > if presented with the proper 95% CI on the difference score. Of
> > > course, this is only for parsimony of presentation, and one should
> > > always compute the difference score to ensure the inference is sound.
>

> > > On Thu, Apr 26, 2012 at 7:33 AM, Thom <thomas.bagu...@ntu.ac.uk> wrote:
>
> > >> I've written about extending this approach to within-subjects CI paper
> > >> in a BRM paper
> > >> and my book. A quick summary of blog posts about it is here.
>

> >http://psychologicalstatistics.blogspot.co.uk/2012/03/graphing-betwee...