EJ
About
- Username
- EJ
- Joined
- Visits
- 2,597
- Last Active
- Roles
- Member, Administrator, Moderator
Comments
-
How different are the results? Most of ML is based on resampling algorithms so results are expected to differ -- but if they differ by a lot than this would be surprising to me. E.J.
-
So you want to divide by N instead of N-1, is that what you mean? If I'm not mistaken you can take the SD from JASP, multiply by [sqrt(1/N) / sqrt(1/(N-1))] and you should get the result you want. I did this very quickly though, so best to check! A…
-
The method for correction can be the same as that applied for post-hoc tests in ANOVA. See the thesis by Tim de Jong for details: https://psyarxiv.com/s56mk Cheers, E.J.
-
Hi Ana, I'm sure you can put everything in a single model, but I'd keep things simple and conduct two separate analyses. Cheers, E.J.
-
We'll soon (this week, hopefully) preprint a paper in which we outline a problem with the currently standard way of specifying RM models in BayesFactor.
-
Hi Marthapapa, I'm afraid that I have never used circular statistics myself. I think I can only advise you to talk with an expert for this particular analysis (it was a contributed module, so the other JASP team members are not experts either I thin…
-
Hi Marijn, This is not yet possible. Well, I guess with some categorical variables it would be an ANCOVA... Cheers, E.J.
-
Yes, so Richard confirms that it would just be an unbalanced design in that case, which the Bayesian model handles.
-
The BayesFactor models are in fact linear mixed models. Let me get confirmation from Richard, also on the missing data aspect.
-
Hi ARF, If you have many missing data you might want to look into missing data imputation. This is something that we still have to do in JASP, but R has several packages for this. I vaguely recall that data missing at random are not a problem for ou…
-
yes!
-
Your current first sentence may make a naive reader believe that the group-only model ia clearly the best, whereas the data support that model only weakly over the null model. E.J.
-
""Using a Bayesian RM ANOVA, the Bayes factor indicates that the data is best represented by the model that just included the group. " Yes but the evidence is really weak. E.J.
-
Dear Nuno, Largely fine; two remarks: " (prior probabilities of each model were equal to 0.5)" -- the BF does not depend on prior probabilities. "Additionally, post-hoc group comparisons exposed posterior odds of 20788.04, indicating …
-
Dear Akypar, I'll attend our expert to this. In general, predictive performance is always less impressive than goodness-of-fit (a Danish proverb goes "prediction is difficult, especially about the future" :-)). E.J.
-
Hi Chloe, I'll pass this on to our SEM expert but they will probably need some more detailed information... Cheers, E.J.
-
Dear lalladem, From the information you provide this seems reasonable. Of course there are many tests presented here, and this does mean you have a multiple comparisons problem (see https://link.springer.com/article/10.3758/s13423-015-0913-5); I did…
-
Hello Max, This code is from the BayesFactor package (I see we still have to produce a help file, but it does list the package). The key paper is by Gunel ands Dickey (also listed in the help file), and a modern version is here: https://link.springe…
-
Have you turned on "safe graphics" in the preferences? If that does not work, please post this on our GitHub page!
-
Great that it worked!
-
This is really strange. Could you attach or email the data? E.J.
-
These are a series of good suggestions -- a feature request on our GitHub page would be great! This is a very popular test so I'll bump the priority once I see the request. (could you make a single request per issue?) E.J.
-
I've tried this, and ticking "display density" changes "counts" (with numbers on the y-axis) to "density" and adds a smooth line. If this does not work for you, can you please create an issue on our Git Hub page?
-
Strange, I'll take a look. This does seem more of a bug report for our GitHub page...
-
This is very much on our radar!
-
Could you elaborate on what you mean with "questionable interpretability"?
-
(I mean, maybe we'll create one sooner than the 0.17 version)
-
No, we need to do the Mac install separately. We intend to have this functionality included in our 0.17 release (before the summer, hopefully). Maybe we'll create a Mac install separately.
-
The more complicated the test, the more involved all of this becomes. For instance, in the Pearson correlation we really have 5 parameters (correlation, plus two means and two variances). What also complicates matters is that we often conduct a test…
-
Because it is Descriptives, we use a uniform prior here. This should be clearly indicated in the help file at least; I'll make an issue on our GitHub page