Draft

GATEs

GATE Coverage

The simulations are based on the the make_heterogeneous_data-DGP with \(500\) observations. The groups are defined based on the first covariate, analogously to the GATE IRM Example, but rely on LightGBM to estimate nuisance elements (due to time constraints).

The non-uniform results (coverage, ci length and bias) refer to averaged values over all groups (point-wise confidende intervals).

DoubleML Version                        0.9.0
Script                   irm_gate_coverage.py
Date                      2024-09-09 08:26:21
Total Runtime (seconds)            3764.03605
Python Version                         3.12.5
Table 1: Coverage for 95.0%-Confidence Interval over 1000 Repetitions
Learner g Learner m Bias CI Length Coverage Uniform CI Length Uniform Coverage
LGBM LGBM 0.485 2.575 0.977 5.085 1.000
LGBM Logistic Regression 0.089 0.465 0.961 0.920 0.998
Lasso LGBM 0.497 2.440 0.959 4.820 1.000
Lasso Logistic Regression 0.090 0.478 0.960 0.940 0.999
Table 2: Coverage for 90.0%-Confidence Interval over 1000 Repetitions
Learner g Learner m Bias CI Length Coverage Uniform CI Length Uniform Coverage
LGBM LGBM 0.485 2.161 0.941 5.064 1.000
LGBM Logistic Regression 0.089 0.390 0.916 0.919 0.997
Lasso LGBM 0.497 2.048 0.904 4.807 1.000
Lasso Logistic Regression 0.090 0.401 0.917 0.942 0.999