JASP Random Forest - Variable Importance Scores

gabriellej

Hello!

I am using Random Forest on JASP for a multiclass classification problem, mainly to obtain variable importance scores. I also ran the same analysis in RStudio using the ranger package, with which I am more familiar. But for some reason, I obtain very different variable importance rankings between JASP and ranger (both for Mean Decrease in Accuracy and Total Increase in Node Purity), and the differences are stable across runs. I am using the same dataset and predictors in both analyses.

I am wondering whether JASP uses different defaults or a different implementation of Random Forest than ranger, and whether this could explain the discrepancy. Does anyone know what might cause such differences, or how I could check whether the model settings are truly equivalent across the two programs?

Thanks in advance for the help!

Gabrielle

EJ

I'll bring this to the attention of our experts!