Effect sizes change from v0.2.5 to v0.3.0
timothysandhu opened this issue · 13 comments
Hi Joses,
I previously posted in the Google group - am reposting the issue here.
I am just wondering if I’m missing something with the updated version of the package.
I am using estimation statistics to analyse the difference in scores between before and after an intervention for two groups (case and control)
Back in July (using version 0.2.5) I ran a series of multi-paired plots using the v2_code attached below (adapted from the vignettes helpfully provided on the package page).
I returned to the analysis late last year, and updated to version 0.3.0, and I used the v3_code attached (again this was adapted from the provided vignettes).
However, the results provided very different paired mean difference plots, despite the Tufte slopegraphs being identical. I have verified this more recently by running the above code on the same machine and installing either version 0.2.5 or 0.3.0.
As far as I can see, the only difference in the code is the use of the mean_diff function in the code used for version 0.3.0. Are there any additional arguments I need to provide that would correct for this difference, or am I missing something?
I have dug a little deeper, and it seems that the mean differences and the mean of the bootstraps are the same across versions (the plots seem to show this too) - but the variance/shape of the bootstrap distribution and hence the confidence intervals are different in v.0.3.0.
I noticed that the bootstrap is now done using boot not simpleboot, could my difference be related to that?
I would be really grateful for any support with this,
Thanks so much,
Best wishes,
Tim
Hi Joses,
Hope you're well!
Just wondering if there had been any update on this. We are in the final stages of a manuscript and would love to be able to include these results.
Do let me know if you need any more information about the issue.
Thanks again for taking a look.
Best,
Tim
It looks to me like the default ylims for your dataset are different between 0.2.5 and 0.3.0.
Could you try
plot(multi.two.group.paired.mean_diff,
effsize.ylim = c(-0.5, 1))
to normalise the way plots from either 0.2.5 or 0.3.0 turn out?
Plots attached from both versions with the above effsize.ylim specification.
Thanks!
OK, looks like a bug, let me get back to you!
Thanks so much Joses! Let me know if you need anything else that would help.
Hi Joses,
Hope you're well!
Just wondering if there's anything else you need from me that might help with the bug?
Thanks so much again for taking a look.
Best,
Tim
Thanks for the update Joses - it looks as if the issues are the same. I will see if I can find anything in the boot package for groups with unequal N. Cheers!
Hi Joses, thanks for letting me know about this update. Should I edit the effectsize.R code manually, or should I wait for v.3.1 to be releaed (apologies, I'm not very Github literate!). Cheers.
Hi @timothysandhu ,
We're not quite ready to ship v0.4 yet, but you can install the dev version branch v0.3.9999 with
devtools::install_github("ACCLAB/dabestr", ref = "v0.3.9999")
and let us know if you face any issues!
Hi! Thanks for that. I've run through all my datasets and the differences between confidence intervals in v0.2.5 and v0.3.9999 only show up at the second decimal place (usually +- 0.02). Whereas v0.3.0 produced wildly different results. Thanks to you all for getting to the bottom of this! :)
All thanks to @FellowFish !