Combining Random Forests and GLMs in R

[@decimus6]

Thanks a lot for this video ! It's awsome as usual. A question rised in my mind : RF are good to select predictors. Yet how does it tackle multicollinearity ? Is it possible that two or three good predictors are multicollinear? Thanks a lot.

[@damaranaidoo9855]

Hi there, I am looking into the RF analysis for my MSc thesis but I am struggling a bit, I am looking at the effect of various environmental variables (Predictor variables) on a binomial response variable (absence/presence), the data is non-linear, and is somewhat skewed, I have tried to run a glm and a gam analysis but both models are not good fits and are underfitting the data. Do you think a RF analysis with the glm would be more appropriate in this case?

[@nikidiogou4203]

Another very useful video for stats, thank you for all of them! The estimates from the rf using flexplot seem not to align with the variable importance score (vi). Shouldn't we have the same ranking of variables when we look at the estimates and when we look at the vi?

[@OnLyhereAlone]

Very informative as usual. The case for linear mixed models (LMM) brought me to your channel. Question; could random forest be used to determine variables to include in an LMM too? Thanks again for the great work you do.

[@QuantPsych]

Yes. There was a paper I saw recently that builds random forest atop mixed models: onlinelibrary.wiley.com/doi/abs/10.1002/sam.11505 Alternatively, I've averaged the scores within cluster, used RF to find variables, then used mixed models on those variables.

[@mohamedrefaat197]

Thanks for the quality content! I wonder what you mean by transportable at the beginning?

[@QuantPsych]

Have you checked out this video? kzbin.info/www/bejne/jKKudquQfJaWl6s I believe that explains what "transportable" means.

[@tatjanajak]

cforest() from party package takes a loooong time. But, when I try to use result of the randomForest() from randomForest package within the estimates(), I get the following error: Error in x$r.squared : $ operator is invalid for atomic vectors. I guess that the results of these two functions are different and only cforest() can be used within flexplot functions. I hope in the future you will introduce randomForest() into this whole process. I think it's worth it because cforest is just too memory/time consuming.

[@christoph3933]

What do you do in case of missing values? Do you recommend doing Multiple Imputation before?

[@QuantPsych]

Yes!

[@gimanibe]

Thanks for the videos you make. I learn a lot! Are this R script available somewhere?

[@QuantPsych]

The code in the video should work. If I find time, I'll put them in the description.

[@scottnelson7841]

no matter how many times I load the package and Library, I get this error message: Error in variable_dropout(explained_rf, type = "raw") : could not find function "variable_dropout". Any help?

[@francisolsson9728]

Can you use random forests using categorical and numeric variables?

[@tatjanajak]

@QuantPsych it seems you did not use GLMs but rather standard lm.

[@QuantPsych]

Possibly. I haven't watched this video for a while :) But I use to use GLM to refer to general linear models and GLIM to refer to general*ized* linear models. I switched the notation somewhat recently. I might have meant to refer to LMs instead of GLMs.

[@tatjanajak]

@@QuantPsych at 20:58 is where I believe the error occurs. It is really just a small mistake. It is really not a big deal, but the whole presentation is awesome as usual and the only thing to do is to say at 20:58 "I ment glm instead of lm".

[@woosterjeeves]

@@tatjanajak GLM is variously used to refer to General Linear Model (which is done using lm() in R), or the Generalized Linear Model, which is what you are referring to (which is done using glm() in R). So he is talking of General Linear Model. Hence not a "mistake".

[@tatjanajak]

@woosterjeeves thanks.

[@Martyr022]

still waiting on that paper!

[@QuantPsych]

Here it is! psyarxiv.com/ebsmr

[@Martyr022]

@@QuantPsych Huzzah! Thank you! Your channel has been super helpful!