Plackett-Luce (PL) models are used as a sensitivity analysis complementary to the critical difference plots. Each dataset acts as a “rater” providing a complete ranking over all models (“items”).
KM and NEL baselines are excluded as they are uninformative for model comparison.
We refer to the PlackettLuce package for more details.
Full Model
Subgroup Analysis
Separate PL models are fit for subgroups based on dataset characteristics. Likelihood ratio tests compare the combined log-likelihood of subgroup-specific models against the full model.
PH Assumption Violation
LR test p-value: < 2.22e-16
LR test p-value: 5.7229e-12
Censoring Proportion
LR test p-value: 0.78673
LR test p-value: 0.001039
Sample Size to Feature Ratio (n/p)
LR test p-value: 0.44064
LR test p-value: 0.04657
PL Trees
PL trees use model-based recursive partitioning to automatically detect dataset subgroups where model rankings differ significantly.
Error in chol.default(meat) :
the leading minor of order 16 is not positive
Error in chol.default(meat) :
the leading minor of order 14 is not positive