Mixed Models Module

Everything related to the development of modules in jamovi

by jonathon » Mon Oct 23, 2017 10:39 am

righto. I'll push this to the library in the next day or so!

what if i say that it's a beta?

jonathon
User avatar
jonathon
 
Posts: 899
Joined: Fri Jan 27, 2017 10:04 am

by mcfanda@gmail.com » Mon Oct 23, 2017 8:07 pm

beta is better then alpha, I'd say
mcfanda@gmail.com
 
Posts: 143
Joined: Thu Mar 23, 2017 9:24 pm

by jonathon » Wed Oct 25, 2017 5:03 am

OK, it's in the jamovi library for windows and macOS, i'll add the linux version soon.
User avatar
jonathon
 
Posts: 899
Joined: Fri Jan 27, 2017 10:04 am

by snfraser » Wed Oct 25, 2017 3:31 pm

Thanks for this...installing it now. Time to dust off some data and give it a go.
-shawn
snfraser
 
Posts: 7
Joined: Thu May 25, 2017 7:48 pm

by mcfanda@gmail.com » Fri Oct 27, 2017 4:57 pm

Hi, just pushed a new version on github (https://github.com/mcfanda/gamlj), with
* the generalized linear model (logistic, poisson, multinomial)
* estimated marginal means (for all model families)
* confidence intervals (for all models families).
Any change to have a look at it, please let me know if it installs (people are having problems but they do not why?)
mcfanda@gmail.com
 
Posts: 143
Joined: Thu Mar 23, 2017 9:24 pm

by jonathon » Sat Oct 28, 2017 1:06 am

hey, i'll push out the new version(s) shortly.

(people are having problems but they do not why?)


they're having problems installing with jmvtools? there is an issue with viridisLite that could be at fault. jmvtools installs packages without their data (to reduce module size), but viridis imports data from viridisLite, it fails to find the data, and the installation of the dependency fails.

you can work around this by manually installing viridisLite into the build/R directory
User avatar
jonathon
 
Posts: 899
Joined: Fri Jan 27, 2017 10:04 am

by dropmann » Sat Oct 28, 2017 9:47 am

I just built gamlj 0.0.6 for windows and it worked fine. :)
dropmann
 
Posts: 57
Joined: Thu Feb 02, 2017 10:26 am

by Ravi » Sat Oct 28, 2017 11:59 am

This is really great stuff! Loving the generalized linear model. I'm working on the binomial logistic regression at the moment, but this more general approach to generalized linear models is pretty powerful too. I'll go through the analyses in the next couple of days and will give some feedback.

Btw, it also builds fine on Ubuntu Zesty (17.04)
User avatar
Ravi
 
Posts: 140
Joined: Sat Jan 28, 2017 11:18 am

by Ravi » Mon Oct 30, 2017 11:05 am

Ok, here my feedback:

    General comments:
  1. By adding suggests to the variable supplier you push the variable types that are not in here to the bottom of the list (in this case the nominal text variables), making it harder to find them. We usually only use this when certain types of variables are incompatible with the analysis. Because all three analyses can use nominal text variables I would just get rid of the suggests.

    GLM:
  2. Title is GLM, but this is also often used for Generalized Linear Models, so maybe just write it out so it's clear you are doing a General Linear Model here.
  3. I would show the "contrast coding" directly underneath the model coefficients so it's easy to do a simple look up
  4. If I add a variable to "Separate plots" in the "Plots" menu the effects plots dissapear
  5. The interaction terms can in an inconsistent order. See screenshot, where 3-way interaction in ANOVA is "school*x1*x2" while in model coefficients is "x1*x2*school1". I assume this has something to do with wanting the factor to be after the covariate in the model coefficients.
    screenshot1.png
    screenshot1.png (178.82 KiB) Viewed 1827 times


    Mixed Model:
  6. If you have a mixed model with only a cluster variable, the Fixed Effect ANOVA table is still shown even thought it doesn't have any rows. I would just hide the table if there are no fixed factors and/or covariates.
  7. Related to 6.: The intercept in the parameter estimates shows a dot for the 'df' and 'p' when there are no fixed effects. I would make this an NaN, plus I would add a footnote to these values instead of adding a more general note.
  8. I always show the main results first. So I'm definetly not an expert on mixed effect models, but I would think that the most important results are the Fixed Effects ANOVA and Parameter Estimates tables. So shouldn't they be on top?
  9. When I only add the intercept to the Random Coefficients I get the following table (I'm not sure why I'm not getting any output here.
    screenshot2.png
    screenshot2.png (54.91 KiB) Viewed 1827 times
  10. Getting an error message with the following setup:
    gamlj_error1.omv
    (11.96 KiB) Downloaded 90 times
  11. The contrast column in the estimates table does not update when you change the factor coding.

    Generalized Linear Models:
  12. I cannot add a nominal text variable to the dependent variable (while this should be possible when I want to do a logistic regression).
  13. Is the poisson model similar to the log-linear model? If so, how does it work? Does it use count data?
  14. The multinomial doesn't work yet. Maybe create a clearer error message, or just omit it until you have implemented it.
  15. What do the Sum of Squares do in the Model submenu?
User avatar
Ravi
 
Posts: 140
Joined: Sat Jan 28, 2017 11:18 am

by mcfanda@gmail.com » Tue Oct 31, 2017 9:02 am

Hi Ravi,
thanks a lot for the review. I'm preparing version 0.0.7 with all your suggestions implemented, fixes, and multinomial model. As regards the poisson model is basically the log-linear model with continuous and categorical independent variables. I called poisson rather than log-linear because SPSS and R called it so and because I have the feeling that users would expect the log-linear to have categorical independent variables (cross-tabs analysis).

Here is a quick reference to their equivalence (https://onlinecourses.science.psu.edu/stat504/node/216)

thanks again
mcfanda@gmail.com
 
Posts: 143
Joined: Thu Mar 23, 2017 9:24 pm

PreviousNext

Return to Module development

cron