The Power of Unbiased Recursive Partitioning: A Unifying View of CTree, MOB, and GUIDE (2019)

by Lisa Schlosser and Torsten Hothorn and Achim Zeileis

Replication materials

The simulation results can be reproduced by the function calls in the file
simulation.R

The corresponding figures can be reproduced by the function calls in the file
figures.R

The already calculated simulation results are provided in the folder results containing three .rda files:

• sim_stump.rda
• sim_3way.rda
• sim_tree.rda

If results should not be recalculated for producing figures, the .rda files need to be exported from the folder results to the main folder simulation, otherwise the results are recalculated when executing figures.R.

Functions applied in the simulation are provided in the source files

• tests_guideprune.R
• guidelike.R
• ccprune.R

The file tests_guideprune.R contains the data generating proceses and the evaluation functions providing the general framework to apply all tests considered in the study and obtain results already prepared for illustration.

The file guidelike.R contains the testing functions, i.e., the modifications of CTree and MOB and the reimplementation of GUIDE also allowing for modifications.

The file ccprune.R contains the reimplementation of cost-complexity pruning.

For the partykit package a development version (1.2-4.2) which is based on version 1.2-4 is applied. Contrary to version 1.2-4 (as currently available on CRAN) this development version provides M-fluctuation tests. Therefore, the applied development version 1.2-4.2 is provided in the folder simulation. (For vignettes please see partykit package version 1.2-4. or https://CRAN.R-project.org/package=partykit)

Computation time for full reproduction of results (on our machines, using 3 kernels): 296 hours and 20 min consisting of approximately:

• 21 h for the “stump” scenario
• 20 min for the full factorial analysis of the building blocks in the “stump” scenario
• 275 h for the “tree” scenario.