mrshravan commited on
Commit
68744ef
·
verified ·
1 Parent(s): f06bc4b

Upload benchmarks/cate_results.json with huggingface_hub

Browse files
Files changed (1) hide show
  1. benchmarks/cate_results.json +1236 -0
benchmarks/cate_results.json ADDED
@@ -0,0 +1,1236 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "n_datasets": 100,
3
+ "overall": {
4
+ "rmse_mean": 1.0950021263211966,
5
+ "rmse_std": 0.6657068416286972,
6
+ "bias_mean": 0.07855330407386646,
7
+ "mae_mean": 0.9142323684692383,
8
+ "direction_acc_mean": 0.46335505171120217,
9
+ "correlation_mean": 0.005576916516092281,
10
+ "rank_correlation_mean": 0.01190094221366825
11
+ },
12
+ "per_regime": {
13
+ "confounded": {
14
+ "n": 33,
15
+ "rmse_mean": 0.9907208999449556,
16
+ "rank_corr_mean": -0.01544867591731753
17
+ },
18
+ "direct": {
19
+ "n": 36,
20
+ "rmse_mean": 1.1825751289725304,
21
+ "rank_corr_mean": -0.002596587489964132
22
+ },
23
+ "feedback": {
24
+ "n": 12,
25
+ "rmse_mean": 1.152058593928814,
26
+ "rank_corr_mean": -0.0031580659268785082
27
+ },
28
+ "mediated": {
29
+ "n": 19,
30
+ "rmse_mean": 1.0741586928304874,
31
+ "rank_corr_mean": 0.07479106291761933
32
+ }
33
+ },
34
+ "per_dataset": [
35
+ {
36
+ "dataset": 0,
37
+ "regime": "direct",
38
+ "is_null": false,
39
+ "rmse": 0.6617948412895203,
40
+ "bias": -0.5320540070533752,
41
+ "mae": 0.5779643058776855,
42
+ "true_cate_mean": 0.2862856090068817,
43
+ "direction_acc": 0.48,
44
+ "correlation": 0.25372428725295815,
45
+ "rank_correlation": -0.013221330533263333
46
+ },
47
+ {
48
+ "dataset": 1,
49
+ "regime": "direct",
50
+ "is_null": false,
51
+ "rmse": 0.6788373589515686,
52
+ "bias": -0.4441879987716675,
53
+ "mae": 0.5643178224563599,
54
+ "true_cate_mean": 0.29235100746154785,
55
+ "direction_acc": 0.49454545454545457,
56
+ "correlation": 0.045813056094384234,
57
+ "rank_correlation": 0.08733594611432224
58
+ },
59
+ {
60
+ "dataset": 2,
61
+ "regime": "confounded",
62
+ "is_null": true,
63
+ "rmse": 0.11162853986024857,
64
+ "bias": 0.11050385236740112,
65
+ "mae": 0.11050385236740112,
66
+ "true_cate_mean": 0.0,
67
+ "direction_acc": null,
68
+ "correlation": null,
69
+ "rank_correlation": null
70
+ },
71
+ {
72
+ "dataset": 3,
73
+ "regime": "feedback",
74
+ "is_null": false,
75
+ "rmse": 1.8240662813186646,
76
+ "bias": 1.5448065996170044,
77
+ "mae": 1.6062787771224976,
78
+ "true_cate_mean": 0.6935684084892273,
79
+ "direction_acc": 0.342,
80
+ "correlation": 0.1394818308242272,
81
+ "rank_correlation": 0.10960681523907226
82
+ },
83
+ {
84
+ "dataset": 4,
85
+ "regime": "confounded",
86
+ "is_null": false,
87
+ "rmse": 1.0922218561172485,
88
+ "bias": 0.7621322274208069,
89
+ "mae": 0.9323183298110962,
90
+ "true_cate_mean": 0.4948004484176636,
91
+ "direction_acc": 0.3466666666666667,
92
+ "correlation": -0.039573780222467554,
93
+ "rank_correlation": 0.14561536459385424
94
+ },
95
+ {
96
+ "dataset": 5,
97
+ "regime": "confounded",
98
+ "is_null": true,
99
+ "rmse": 0.536484956741333,
100
+ "bias": 0.5364429354667664,
101
+ "mae": 0.5364429354667664,
102
+ "true_cate_mean": 0.0,
103
+ "direction_acc": null,
104
+ "correlation": null,
105
+ "rank_correlation": null
106
+ },
107
+ {
108
+ "dataset": 6,
109
+ "regime": "direct",
110
+ "is_null": false,
111
+ "rmse": 0.28288498520851135,
112
+ "bias": 0.0183744914829731,
113
+ "mae": 0.16790835559368134,
114
+ "true_cate_mean": 0.1600653976202011,
115
+ "direction_acc": 0.49,
116
+ "correlation": -0.03533444045606569,
117
+ "rank_correlation": -0.11254573089067903
118
+ },
119
+ {
120
+ "dataset": 7,
121
+ "regime": "direct",
122
+ "is_null": false,
123
+ "rmse": 2.680650234222412,
124
+ "bias": -1.6684149503707886,
125
+ "mae": 2.175900936126709,
126
+ "true_cate_mean": 1.6444482803344727,
127
+ "direction_acc": 0.5,
128
+ "correlation": -0.1678807402168779,
129
+ "rank_correlation": 0.08995746658174382
130
+ },
131
+ {
132
+ "dataset": 8,
133
+ "regime": "confounded",
134
+ "is_null": false,
135
+ "rmse": 1.6310144662857056,
136
+ "bias": -0.9647163152694702,
137
+ "mae": 1.2830296754837036,
138
+ "true_cate_mean": 0.9816451072692871,
139
+ "direction_acc": 0.5413333333333333,
140
+ "correlation": 0.2299950451217524,
141
+ "rank_correlation": -0.12699502915551797
142
+ },
143
+ {
144
+ "dataset": 9,
145
+ "regime": "mediated",
146
+ "is_null": false,
147
+ "rmse": 0.7033092975616455,
148
+ "bias": -0.3257269561290741,
149
+ "mae": 0.4938548505306244,
150
+ "true_cate_mean": 0.34903624653816223,
151
+ "direction_acc": 0.492,
152
+ "correlation": 0.0028060751132965947,
153
+ "rank_correlation": 0.014819114819114819
154
+ },
155
+ {
156
+ "dataset": 10,
157
+ "regime": "direct",
158
+ "is_null": false,
159
+ "rmse": 0.48507022857666016,
160
+ "bias": -0.4210164248943329,
161
+ "mae": 0.44296810030937195,
162
+ "true_cate_mean": 0.07615800946950912,
163
+ "direction_acc": 0.5,
164
+ "correlation": -0.010054047534218156,
165
+ "rank_correlation": 0.06500035354238533
166
+ },
167
+ {
168
+ "dataset": 11,
169
+ "regime": "confounded",
170
+ "is_null": true,
171
+ "rmse": 0.623291015625,
172
+ "bias": 0.6228503584861755,
173
+ "mae": 0.6228503584861755,
174
+ "true_cate_mean": 0.0,
175
+ "direction_acc": null,
176
+ "correlation": null,
177
+ "rank_correlation": null
178
+ },
179
+ {
180
+ "dataset": 12,
181
+ "regime": "direct",
182
+ "is_null": false,
183
+ "rmse": 2.746073007583618,
184
+ "bias": 2.11612606048584,
185
+ "mae": 2.2534735202789307,
186
+ "true_cate_mean": 1.3518986701965332,
187
+ "direction_acc": 0.5,
188
+ "correlation": -0.22870013630468136,
189
+ "rank_correlation": -0.022097022425900167
190
+ },
191
+ {
192
+ "dataset": 13,
193
+ "regime": "feedback",
194
+ "is_null": false,
195
+ "rmse": 0.5141034126281738,
196
+ "bias": -0.26295584440231323,
197
+ "mae": 0.3845289945602417,
198
+ "true_cate_mean": 0.2553296387195587,
199
+ "direction_acc": 0.4942857142857143,
200
+ "correlation": 0.03561391070018756,
201
+ "rank_correlation": 0.010580493736403511
202
+ },
203
+ {
204
+ "dataset": 14,
205
+ "regime": "direct",
206
+ "is_null": false,
207
+ "rmse": 2.3931431770324707,
208
+ "bias": -2.3208019733428955,
209
+ "mae": 2.32643723487854,
210
+ "true_cate_mean": 0.29879188537597656,
211
+ "direction_acc": 0.5163636363636364,
212
+ "correlation": 0.2000924816563767,
213
+ "rank_correlation": 0.18435959699989268
214
+ },
215
+ {
216
+ "dataset": 15,
217
+ "regime": "direct",
218
+ "is_null": false,
219
+ "rmse": 0.9207562804222107,
220
+ "bias": 0.39409393072128296,
221
+ "mae": 0.6890666484832764,
222
+ "true_cate_mean": 0.6267673373222351,
223
+ "direction_acc": 0.43125,
224
+ "correlation": -0.20191348975634052,
225
+ "rank_correlation": 0.21338323171372772
226
+ },
227
+ {
228
+ "dataset": 16,
229
+ "regime": "mediated",
230
+ "is_null": false,
231
+ "rmse": 1.0763795375823975,
232
+ "bias": 0.3026221990585327,
233
+ "mae": 0.8172667026519775,
234
+ "true_cate_mean": 0.7511969208717346,
235
+ "direction_acc": 0.4025,
236
+ "correlation": 0.02162201075841969,
237
+ "rank_correlation": 0.30320164152944123
238
+ },
239
+ {
240
+ "dataset": 17,
241
+ "regime": "confounded",
242
+ "is_null": true,
243
+ "rmse": 0.8667718172073364,
244
+ "bias": 0.19768454134464264,
245
+ "mae": 0.6907031536102295,
246
+ "true_cate_mean": 0.6773681044578552,
247
+ "direction_acc": 0.48,
248
+ "correlation": -0.11169212266319788,
249
+ "rank_correlation": 0.2834917415907876
250
+ },
251
+ {
252
+ "dataset": 18,
253
+ "regime": "mediated",
254
+ "is_null": false,
255
+ "rmse": 1.6441996097564697,
256
+ "bias": -0.4325984716415405,
257
+ "mae": 1.3819541931152344,
258
+ "true_cate_mean": 1.3594996929168701,
259
+ "direction_acc": 0.46,
260
+ "correlation": 0.17386228001283585,
261
+ "rank_correlation": -0.019429824758340798
262
+ },
263
+ {
264
+ "dataset": 19,
265
+ "regime": "direct",
266
+ "is_null": false,
267
+ "rmse": 0.9828081727027893,
268
+ "bias": -0.8013883233070374,
269
+ "mae": 0.8635560870170593,
270
+ "true_cate_mean": 0.3381716012954712,
271
+ "direction_acc": 0.49272727272727274,
272
+ "correlation": 0.017970058255186973,
273
+ "rank_correlation": 0.07130040753602428
274
+ },
275
+ {
276
+ "dataset": 20,
277
+ "regime": "confounded",
278
+ "is_null": false,
279
+ "rmse": 1.1153393983840942,
280
+ "bias": 0.8920561075210571,
281
+ "mae": 0.9412471652030945,
282
+ "true_cate_mean": 0.46289828419685364,
283
+ "direction_acc": 0.432,
284
+ "correlation": -0.04849502511918369,
285
+ "rank_correlation": 0.20644204582215164
286
+ },
287
+ {
288
+ "dataset": 21,
289
+ "regime": "confounded",
290
+ "is_null": false,
291
+ "rmse": 0.7101827263832092,
292
+ "bias": 0.5805736780166626,
293
+ "mae": 0.63060462474823,
294
+ "true_cate_mean": 0.265241414308548,
295
+ "direction_acc": 0.5047058823529412,
296
+ "correlation": -0.04775305459758306,
297
+ "rank_correlation": -0.17117018197950784
298
+ },
299
+ {
300
+ "dataset": 22,
301
+ "regime": "direct",
302
+ "is_null": false,
303
+ "rmse": 1.1525394916534424,
304
+ "bias": 0.25128981471061707,
305
+ "mae": 0.8543605804443359,
306
+ "true_cate_mean": 0.8255916833877563,
307
+ "direction_acc": 0.342,
308
+ "correlation": -0.29114347207924324,
309
+ "rank_correlation": -0.03410472806570497
310
+ },
311
+ {
312
+ "dataset": 23,
313
+ "regime": "confounded",
314
+ "is_null": true,
315
+ "rmse": 0.284614622592926,
316
+ "bias": -0.2843656539916992,
317
+ "mae": 0.2843656539916992,
318
+ "true_cate_mean": 0.0,
319
+ "direction_acc": null,
320
+ "correlation": null,
321
+ "rank_correlation": null
322
+ },
323
+ {
324
+ "dataset": 24,
325
+ "regime": "confounded",
326
+ "is_null": true,
327
+ "rmse": 0.5636901259422302,
328
+ "bias": -0.5634097456932068,
329
+ "mae": 0.5634097456932068,
330
+ "true_cate_mean": 0.0,
331
+ "direction_acc": null,
332
+ "correlation": null,
333
+ "rank_correlation": null
334
+ },
335
+ {
336
+ "dataset": 25,
337
+ "regime": "confounded",
338
+ "is_null": false,
339
+ "rmse": 1.1715936660766602,
340
+ "bias": -0.267246276140213,
341
+ "mae": 0.9231908917427063,
342
+ "true_cate_mean": 0.8973104357719421,
343
+ "direction_acc": 0.5288888888888889,
344
+ "correlation": -0.02616444245555424,
345
+ "rank_correlation": 0.054066255958465745
346
+ },
347
+ {
348
+ "dataset": 26,
349
+ "regime": "feedback",
350
+ "is_null": false,
351
+ "rmse": 0.9438083171844482,
352
+ "bias": 0.6170626282691956,
353
+ "mae": 0.7767817378044128,
354
+ "true_cate_mean": 0.48403799533843994,
355
+ "direction_acc": 0.448,
356
+ "correlation": 0.16503029274249034,
357
+ "rank_correlation": 0.12191156422766432
358
+ },
359
+ {
360
+ "dataset": 27,
361
+ "regime": "mediated",
362
+ "is_null": false,
363
+ "rmse": 1.1129872798919678,
364
+ "bias": -0.14711788296699524,
365
+ "mae": 0.8506404161453247,
366
+ "true_cate_mean": 0.8597917556762695,
367
+ "direction_acc": 0.5485714285714286,
368
+ "correlation": 0.12058365617422274,
369
+ "rank_correlation": -0.20792218952009622
370
+ },
371
+ {
372
+ "dataset": 28,
373
+ "regime": "direct",
374
+ "is_null": false,
375
+ "rmse": 0.7927359342575073,
376
+ "bias": -0.6411092281341553,
377
+ "mae": 0.7015825510025024,
378
+ "true_cate_mean": 0.24323248863220215,
379
+ "direction_acc": 0.52,
380
+ "correlation": -0.18338806165322882,
381
+ "rank_correlation": 0.12846956493189915
382
+ },
383
+ {
384
+ "dataset": 29,
385
+ "regime": "direct",
386
+ "is_null": false,
387
+ "rmse": 0.8758434057235718,
388
+ "bias": 0.6014345288276672,
389
+ "mae": 0.7171859741210938,
390
+ "true_cate_mean": 0.43472763895988464,
391
+ "direction_acc": 0.48,
392
+ "correlation": -0.19069777671602842,
393
+ "rank_correlation": 0.3301995644250856
394
+ },
395
+ {
396
+ "dataset": 30,
397
+ "regime": "direct",
398
+ "is_null": false,
399
+ "rmse": 1.5130974054336548,
400
+ "bias": 1.2606492042541504,
401
+ "mae": 1.3535507917404175,
402
+ "true_cate_mean": 0.6040079593658447,
403
+ "direction_acc": 0.31666666666666665,
404
+ "correlation": -0.23176214936231737,
405
+ "rank_correlation": -0.11630048848063379
406
+ },
407
+ {
408
+ "dataset": 31,
409
+ "regime": "confounded",
410
+ "is_null": true,
411
+ "rmse": 0.36047711968421936,
412
+ "bias": 0.3603271543979645,
413
+ "mae": 0.3603271543979645,
414
+ "true_cate_mean": 0.0,
415
+ "direction_acc": null,
416
+ "correlation": null,
417
+ "rank_correlation": null
418
+ },
419
+ {
420
+ "dataset": 32,
421
+ "regime": "confounded",
422
+ "is_null": true,
423
+ "rmse": 0.5993546843528748,
424
+ "bias": -0.5989521145820618,
425
+ "mae": 0.5989521145820618,
426
+ "true_cate_mean": 0.0,
427
+ "direction_acc": null,
428
+ "correlation": null,
429
+ "rank_correlation": null
430
+ },
431
+ {
432
+ "dataset": 33,
433
+ "regime": "direct",
434
+ "is_null": false,
435
+ "rmse": 0.44596976041793823,
436
+ "bias": -0.05933057516813278,
437
+ "mae": 0.3397771418094635,
438
+ "true_cate_mean": 0.32470467686653137,
439
+ "direction_acc": 0.4666666666666667,
440
+ "correlation": -0.18873303171129766,
441
+ "rank_correlation": -0.2780444886499582
442
+ },
443
+ {
444
+ "dataset": 34,
445
+ "regime": "feedback",
446
+ "is_null": false,
447
+ "rmse": 1.2407569885253906,
448
+ "bias": 1.1606225967407227,
449
+ "mae": 1.1622587442398071,
450
+ "true_cate_mean": 0.3097666800022125,
451
+ "direction_acc": 0.4106666666666667,
452
+ "correlation": 0.23092054061833645,
453
+ "rank_correlation": -0.11857129549208995
454
+ },
455
+ {
456
+ "dataset": 35,
457
+ "regime": "mediated",
458
+ "is_null": false,
459
+ "rmse": 1.232722520828247,
460
+ "bias": 0.5017558336257935,
461
+ "mae": 0.9297564625740051,
462
+ "true_cate_mean": 0.7608705759048462,
463
+ "direction_acc": 0.3888888888888889,
464
+ "correlation": -0.17626619547913722,
465
+ "rank_correlation": 0.1823410140025066
466
+ },
467
+ {
468
+ "dataset": 36,
469
+ "regime": "direct",
470
+ "is_null": false,
471
+ "rmse": 0.9090337157249451,
472
+ "bias": 0.5197310447692871,
473
+ "mae": 0.7401201725006104,
474
+ "true_cate_mean": 0.5816068649291992,
475
+ "direction_acc": 0.48,
476
+ "correlation": 0.16809013209400478,
477
+ "rank_correlation": 0.08288970424089236
478
+ },
479
+ {
480
+ "dataset": 37,
481
+ "regime": "confounded",
482
+ "is_null": false,
483
+ "rmse": 0.8660591840744019,
484
+ "bias": -0.5686516761779785,
485
+ "mae": 0.7244287133216858,
486
+ "true_cate_mean": 0.4071798026561737,
487
+ "direction_acc": 0.32266666666666666,
488
+ "correlation": -0.02132912010248223,
489
+ "rank_correlation": -0.1237706537010838
490
+ },
491
+ {
492
+ "dataset": 38,
493
+ "regime": "confounded",
494
+ "is_null": true,
495
+ "rmse": 1.106910228729248,
496
+ "bias": -1.1068511009216309,
497
+ "mae": 1.1068511009216309,
498
+ "true_cate_mean": 0.0,
499
+ "direction_acc": null,
500
+ "correlation": null,
501
+ "rank_correlation": null
502
+ },
503
+ {
504
+ "dataset": 39,
505
+ "regime": "confounded",
506
+ "is_null": true,
507
+ "rmse": 0.567194402217865,
508
+ "bias": 0.5667347311973572,
509
+ "mae": 0.5667347311973572,
510
+ "true_cate_mean": 0.0,
511
+ "direction_acc": null,
512
+ "correlation": null,
513
+ "rank_correlation": null
514
+ },
515
+ {
516
+ "dataset": 40,
517
+ "regime": "direct",
518
+ "is_null": false,
519
+ "rmse": 1.397670030593872,
520
+ "bias": -0.17690709233283997,
521
+ "mae": 1.090164303779602,
522
+ "true_cate_mean": 1.064383625984192,
523
+ "direction_acc": 0.4581818181818182,
524
+ "correlation": -0.027744084231596827,
525
+ "rank_correlation": 0.04980398856250664
526
+ },
527
+ {
528
+ "dataset": 41,
529
+ "regime": "direct",
530
+ "is_null": false,
531
+ "rmse": 0.32633817195892334,
532
+ "bias": 0.18273623287677765,
533
+ "mae": 0.2651352882385254,
534
+ "true_cate_mean": 0.1770557463169098,
535
+ "direction_acc": 0.3333333333333333,
536
+ "correlation": -0.3910663940272414,
537
+ "rank_correlation": 0.1159205414120186
538
+ },
539
+ {
540
+ "dataset": 42,
541
+ "regime": "mediated",
542
+ "is_null": false,
543
+ "rmse": 0.300677627325058,
544
+ "bias": 0.17716442048549652,
545
+ "mae": 0.22806189954280853,
546
+ "true_cate_mean": 0.10198188573122025,
547
+ "direction_acc": 0.5,
548
+ "correlation": -0.10747974986435516,
549
+ "rank_correlation": -0.22780317758974925
550
+ },
551
+ {
552
+ "dataset": 43,
553
+ "regime": "mediated",
554
+ "is_null": false,
555
+ "rmse": 0.9151133298873901,
556
+ "bias": -0.2879578471183777,
557
+ "mae": 0.7501578330993652,
558
+ "true_cate_mean": 0.688305139541626,
559
+ "direction_acc": 0.54,
560
+ "correlation": -0.2809260275364567,
561
+ "rank_correlation": 0.21723999249988937
562
+ },
563
+ {
564
+ "dataset": 44,
565
+ "regime": "feedback",
566
+ "is_null": false,
567
+ "rmse": 0.22314691543579102,
568
+ "bias": 0.0460340715944767,
569
+ "mae": 0.1160632073879242,
570
+ "true_cate_mean": 0.08742111921310425,
571
+ "direction_acc": 0.5035294117647059,
572
+ "correlation": -0.14868787398196923,
573
+ "rank_correlation": -0.12195820340235766
574
+ },
575
+ {
576
+ "dataset": 45,
577
+ "regime": "confounded",
578
+ "is_null": true,
579
+ "rmse": 1.2274088859558105,
580
+ "bias": -0.5336847901344299,
581
+ "mae": 1.002920389175415,
582
+ "true_cate_mean": 0.844397783279419,
583
+ "direction_acc": 0.455,
584
+ "correlation": 0.1636622976655014,
585
+ "rank_correlation": 0.09016541873240035
586
+ },
587
+ {
588
+ "dataset": 46,
589
+ "regime": "feedback",
590
+ "is_null": false,
591
+ "rmse": 2.275334596633911,
592
+ "bias": -2.143026828765869,
593
+ "mae": 2.149407386779785,
594
+ "true_cate_mean": 0.4541521966457367,
595
+ "direction_acc": 0.3247058823529412,
596
+ "correlation": -0.18683309824562103,
597
+ "rank_correlation": -0.12124576306599065
598
+ },
599
+ {
600
+ "dataset": 47,
601
+ "regime": "confounded",
602
+ "is_null": true,
603
+ "rmse": 2.1218416690826416,
604
+ "bias": 0.2755686938762665,
605
+ "mae": 1.7084693908691406,
606
+ "true_cate_mean": 1.692691445350647,
607
+ "direction_acc": 0.49523809523809526,
608
+ "correlation": 0.015541130776727134,
609
+ "rank_correlation": -0.19914015233797178
610
+ },
611
+ {
612
+ "dataset": 48,
613
+ "regime": "confounded",
614
+ "is_null": true,
615
+ "rmse": 0.5085961222648621,
616
+ "bias": -0.3013857901096344,
617
+ "mae": 0.39317184686660767,
618
+ "true_cate_mean": 0.2506243884563446,
619
+ "direction_acc": 0.5,
620
+ "correlation": -0.046895797307984814,
621
+ "rank_correlation": 0.08331121757723443
622
+ },
623
+ {
624
+ "dataset": 49,
625
+ "regime": "mediated",
626
+ "is_null": false,
627
+ "rmse": 1.136450171470642,
628
+ "bias": -0.9775269627571106,
629
+ "mae": 1.002151370048523,
630
+ "true_cate_mean": 0.43590861558914185,
631
+ "direction_acc": 0.3054545454545455,
632
+ "correlation": 0.22040005152069553,
633
+ "rank_correlation": 0.09248555535855589
634
+ },
635
+ {
636
+ "dataset": 50,
637
+ "regime": "direct",
638
+ "is_null": false,
639
+ "rmse": 1.2079155445098877,
640
+ "bias": -0.28232020139694214,
641
+ "mae": 0.9482648968696594,
642
+ "true_cate_mean": 0.8484445214271545,
643
+ "direction_acc": 0.31666666666666665,
644
+ "correlation": 0.2525981085714635,
645
+ "rank_correlation": -0.24126202602928648
646
+ },
647
+ {
648
+ "dataset": 51,
649
+ "regime": "confounded",
650
+ "is_null": false,
651
+ "rmse": 2.4825713634490967,
652
+ "bias": -2.017218589782715,
653
+ "mae": 2.119798183441162,
654
+ "true_cate_mean": 1.081558346748352,
655
+ "direction_acc": 0.36538461538461536,
656
+ "correlation": 0.20578901497956456,
657
+ "rank_correlation": 0.048410740886873034
658
+ },
659
+ {
660
+ "dataset": 52,
661
+ "regime": "confounded",
662
+ "is_null": true,
663
+ "rmse": 0.515188455581665,
664
+ "bias": -0.5148139595985413,
665
+ "mae": 0.5148139595985413,
666
+ "true_cate_mean": 0.0,
667
+ "direction_acc": null,
668
+ "correlation": null,
669
+ "rank_correlation": null
670
+ },
671
+ {
672
+ "dataset": 53,
673
+ "regime": "direct",
674
+ "is_null": false,
675
+ "rmse": 2.076636791229248,
676
+ "bias": 0.3256375789642334,
677
+ "mae": 1.7001655101776123,
678
+ "true_cate_mean": 1.6919690370559692,
679
+ "direction_acc": 0.52,
680
+ "correlation": -0.17944557834155747,
681
+ "rank_correlation": 0.007736749317661583
682
+ },
683
+ {
684
+ "dataset": 54,
685
+ "regime": "confounded",
686
+ "is_null": true,
687
+ "rmse": 0.5558907985687256,
688
+ "bias": -0.5556618571281433,
689
+ "mae": 0.5556618571281433,
690
+ "true_cate_mean": 0.0,
691
+ "direction_acc": null,
692
+ "correlation": null,
693
+ "rank_correlation": null
694
+ },
695
+ {
696
+ "dataset": 55,
697
+ "regime": "feedback",
698
+ "is_null": false,
699
+ "rmse": 1.5055180788040161,
700
+ "bias": 1.1131011247634888,
701
+ "mae": 1.2696088552474976,
702
+ "true_cate_mean": 0.827399492263794,
703
+ "direction_acc": 0.40551724137931033,
704
+ "correlation": -0.6211552657274656,
705
+ "rank_correlation": -0.010728745823194226
706
+ },
707
+ {
708
+ "dataset": 56,
709
+ "regime": "direct",
710
+ "is_null": false,
711
+ "rmse": 1.284661889076233,
712
+ "bias": 1.10772705078125,
713
+ "mae": 1.155983567237854,
714
+ "true_cate_mean": 0.43856048583984375,
715
+ "direction_acc": 0.36,
716
+ "correlation": -0.37397681292261553,
717
+ "rank_correlation": 0.3061012881330508
718
+ },
719
+ {
720
+ "dataset": 57,
721
+ "regime": "confounded",
722
+ "is_null": true,
723
+ "rmse": 0.3467826247215271,
724
+ "bias": 0.34630560874938965,
725
+ "mae": 0.34630560874938965,
726
+ "true_cate_mean": 0.0,
727
+ "direction_acc": null,
728
+ "correlation": null,
729
+ "rank_correlation": null
730
+ },
731
+ {
732
+ "dataset": 58,
733
+ "regime": "mediated",
734
+ "is_null": false,
735
+ "rmse": 1.622272253036499,
736
+ "bias": 1.0999506711959839,
737
+ "mae": 1.3339359760284424,
738
+ "true_cate_mean": 0.9427119493484497,
739
+ "direction_acc": 0.4909090909090909,
740
+ "correlation": 0.44890804207893,
741
+ "rank_correlation": 0.09728233865164497
742
+ },
743
+ {
744
+ "dataset": 59,
745
+ "regime": "direct",
746
+ "is_null": false,
747
+ "rmse": 1.5152853727340698,
748
+ "bias": 0.499358594417572,
749
+ "mae": 1.175520658493042,
750
+ "true_cate_mean": 1.1380959749221802,
751
+ "direction_acc": 0.48,
752
+ "correlation": -0.24521228921978372,
753
+ "rank_correlation": 0.2244690819962583
754
+ },
755
+ {
756
+ "dataset": 60,
757
+ "regime": "direct",
758
+ "is_null": false,
759
+ "rmse": 1.7777525186538696,
760
+ "bias": 1.5777971744537354,
761
+ "mae": 1.5809075832366943,
762
+ "true_cate_mean": 0.5989304780960083,
763
+ "direction_acc": 0.5993333333333334,
764
+ "correlation": 0.3262938393343619,
765
+ "rank_correlation": 0.10222567296391448
766
+ },
767
+ {
768
+ "dataset": 61,
769
+ "regime": "mediated",
770
+ "is_null": false,
771
+ "rmse": 0.901781439781189,
772
+ "bias": -0.45496928691864014,
773
+ "mae": 0.6888560056686401,
774
+ "true_cate_mean": 0.5056042671203613,
775
+ "direction_acc": 0.504,
776
+ "correlation": 0.08854087309046303,
777
+ "rank_correlation": 0.00858445735131762
778
+ },
779
+ {
780
+ "dataset": 62,
781
+ "regime": "direct",
782
+ "is_null": false,
783
+ "rmse": 0.5744344592094421,
784
+ "bias": -0.39384299516677856,
785
+ "mae": 0.47346794605255127,
786
+ "true_cate_mean": 0.2462013214826584,
787
+ "direction_acc": 0.5,
788
+ "correlation": -0.06313499897379056,
789
+ "rank_correlation": -0.043311700740811906
790
+ },
791
+ {
792
+ "dataset": 63,
793
+ "regime": "direct",
794
+ "is_null": false,
795
+ "rmse": 0.6153081655502319,
796
+ "bias": -0.45722103118896484,
797
+ "mae": 0.5229002833366394,
798
+ "true_cate_mean": 0.27910929918289185,
799
+ "direction_acc": 0.4307692307692308,
800
+ "correlation": 0.24816406335789198,
801
+ "rank_correlation": -0.2790294698879036
802
+ },
803
+ {
804
+ "dataset": 64,
805
+ "regime": "feedback",
806
+ "is_null": false,
807
+ "rmse": 0.58504319190979,
808
+ "bias": 0.18742384016513824,
809
+ "mae": 0.4442776143550873,
810
+ "true_cate_mean": 0.39694955945014954,
811
+ "direction_acc": 0.3757894736842105,
812
+ "correlation": -0.09886001498025465,
813
+ "rank_correlation": -0.08739310797966363
814
+ },
815
+ {
816
+ "dataset": 65,
817
+ "regime": "direct",
818
+ "is_null": false,
819
+ "rmse": 3.4095096588134766,
820
+ "bias": -2.770055055618286,
821
+ "mae": 2.8267109394073486,
822
+ "true_cate_mean": 1.610695481300354,
823
+ "direction_acc": 0.56,
824
+ "correlation": -0.18748488284281228,
825
+ "rank_correlation": 0.07160997213150568
826
+ },
827
+ {
828
+ "dataset": 66,
829
+ "regime": "feedback",
830
+ "is_null": false,
831
+ "rmse": 0.62506103515625,
832
+ "bias": 0.03470136597752571,
833
+ "mae": 0.4052111804485321,
834
+ "true_cate_mean": 0.4166032373905182,
835
+ "direction_acc": 0.7023529411764706,
836
+ "correlation": 0.5342147056531474,
837
+ "rank_correlation": 0.11980041820534171
838
+ },
839
+ {
840
+ "dataset": 67,
841
+ "regime": "mediated",
842
+ "is_null": false,
843
+ "rmse": 1.0732353925704956,
844
+ "bias": 0.3430785536766052,
845
+ "mae": 0.7885157465934753,
846
+ "true_cate_mean": 0.7630085349082947,
847
+ "direction_acc": 0.6,
848
+ "correlation": 0.24381759555469237,
849
+ "rank_correlation": 0.04924841104049069
850
+ },
851
+ {
852
+ "dataset": 68,
853
+ "regime": "mediated",
854
+ "is_null": false,
855
+ "rmse": 1.775130033493042,
856
+ "bias": -0.359058141708374,
857
+ "mae": 1.2157050371170044,
858
+ "true_cate_mean": 1.1729272603988647,
859
+ "direction_acc": 0.368,
860
+ "correlation": -0.1723621324545148,
861
+ "rank_correlation": -0.14805488550674573
862
+ },
863
+ {
864
+ "dataset": 69,
865
+ "regime": "mediated",
866
+ "is_null": false,
867
+ "rmse": 0.803101658821106,
868
+ "bias": -0.5771546363830566,
869
+ "mae": 0.6553364992141724,
870
+ "true_cate_mean": 0.42683160305023193,
871
+ "direction_acc": 0.34363636363636363,
872
+ "correlation": 0.3219465597595731,
873
+ "rank_correlation": -0.025317077928059546
874
+ },
875
+ {
876
+ "dataset": 70,
877
+ "regime": "feedback",
878
+ "is_null": false,
879
+ "rmse": 2.03717303276062,
880
+ "bias": 1.2718799114227295,
881
+ "mae": 1.5802956819534302,
882
+ "true_cate_mean": 1.2098674774169922,
883
+ "direction_acc": 0.41904761904761906,
884
+ "correlation": -0.07142280222828572,
885
+ "rank_correlation": 0.15982717479417302
886
+ },
887
+ {
888
+ "dataset": 71,
889
+ "regime": "direct",
890
+ "is_null": false,
891
+ "rmse": 0.5316846966743469,
892
+ "bias": -0.04687440022826195,
893
+ "mae": 0.4009736478328705,
894
+ "true_cate_mean": 0.3978419005870819,
895
+ "direction_acc": 0.42,
896
+ "correlation": -0.5371065415667979,
897
+ "rank_correlation": -0.3416207348128312
898
+ },
899
+ {
900
+ "dataset": 72,
901
+ "regime": "direct",
902
+ "is_null": false,
903
+ "rmse": 2.3474178314208984,
904
+ "bias": -1.7128716707229614,
905
+ "mae": 2.0255486965179443,
906
+ "true_cate_mean": 1.0809729099273682,
907
+ "direction_acc": 0.5,
908
+ "correlation": 0.1609626546823143,
909
+ "rank_correlation": -0.057417526861409565
910
+ },
911
+ {
912
+ "dataset": 73,
913
+ "regime": "mediated",
914
+ "is_null": false,
915
+ "rmse": 1.2348747253417969,
916
+ "bias": 0.9530985951423645,
917
+ "mae": 1.0505603551864624,
918
+ "true_cate_mean": 0.5775202512741089,
919
+ "direction_acc": 0.42428571428571427,
920
+ "correlation": 0.21626558893008752,
921
+ "rank_correlation": -0.1008894862928908
922
+ },
923
+ {
924
+ "dataset": 74,
925
+ "regime": "confounded",
926
+ "is_null": false,
927
+ "rmse": 0.9250388145446777,
928
+ "bias": -0.06805108487606049,
929
+ "mae": 0.6749475002288818,
930
+ "true_cate_mean": 0.6547553539276123,
931
+ "direction_acc": 0.432,
932
+ "correlation": -0.05035273265587565,
933
+ "rank_correlation": -0.09166850536803188
934
+ },
935
+ {
936
+ "dataset": 75,
937
+ "regime": "direct",
938
+ "is_null": false,
939
+ "rmse": 0.43760359287261963,
940
+ "bias": -0.3078922927379608,
941
+ "mae": 0.3614242672920227,
942
+ "true_cate_mean": 0.16797016561031342,
943
+ "direction_acc": 0.56,
944
+ "correlation": 0.4032750406146064,
945
+ "rank_correlation": 0.21623448978564253
946
+ },
947
+ {
948
+ "dataset": 76,
949
+ "regime": "direct",
950
+ "is_null": false,
951
+ "rmse": 1.1499524116516113,
952
+ "bias": 0.4531259536743164,
953
+ "mae": 0.9202916622161865,
954
+ "true_cate_mean": 0.8663648366928101,
955
+ "direction_acc": 0.48,
956
+ "correlation": 0.4413121782975764,
957
+ "rank_correlation": -0.005817486030229543
958
+ },
959
+ {
960
+ "dataset": 77,
961
+ "regime": "mediated",
962
+ "is_null": false,
963
+ "rmse": 1.927869439125061,
964
+ "bias": 1.3129929304122925,
965
+ "mae": 1.5304144620895386,
966
+ "true_cate_mean": 0.9578075408935547,
967
+ "direction_acc": 0.28,
968
+ "correlation": 0.06463246526968648,
969
+ "rank_correlation": 0.3882779229272937
970
+ },
971
+ {
972
+ "dataset": 78,
973
+ "regime": "confounded",
974
+ "is_null": false,
975
+ "rmse": 2.3135221004486084,
976
+ "bias": 1.7471965551376343,
977
+ "mae": 1.9356516599655151,
978
+ "true_cate_mean": 1.2314894199371338,
979
+ "direction_acc": 0.5353846153846153,
980
+ "correlation": 0.025232074834987992,
981
+ "rank_correlation": 0.05381236256328642
982
+ },
983
+ {
984
+ "dataset": 79,
985
+ "regime": "direct",
986
+ "is_null": false,
987
+ "rmse": 0.355951726436615,
988
+ "bias": -0.277770459651947,
989
+ "mae": 0.31757596135139465,
990
+ "true_cate_mean": 0.09696255624294281,
991
+ "direction_acc": 0.52,
992
+ "correlation": -0.16052925852927646,
993
+ "rank_correlation": -0.20993693280273243
994
+ },
995
+ {
996
+ "dataset": 80,
997
+ "regime": "confounded",
998
+ "is_null": false,
999
+ "rmse": 2.063884735107422,
1000
+ "bias": 0.029220758005976677,
1001
+ "mae": 1.7094248533248901,
1002
+ "true_cate_mean": 1.718315601348877,
1003
+ "direction_acc": 0.52,
1004
+ "correlation": 0.04545089140896247,
1005
+ "rank_correlation": 0.08280352535157581
1006
+ },
1007
+ {
1008
+ "dataset": 81,
1009
+ "regime": "feedback",
1010
+ "is_null": false,
1011
+ "rmse": 1.6594241857528687,
1012
+ "bias": 1.4915045499801636,
1013
+ "mae": 1.5116161108016968,
1014
+ "true_cate_mean": 0.5303452014923096,
1015
+ "direction_acc": 0.46,
1016
+ "correlation": -0.7555004292205885,
1017
+ "rank_correlation": -0.13013765914872294
1018
+ },
1019
+ {
1020
+ "dataset": 82,
1021
+ "regime": "direct",
1022
+ "is_null": false,
1023
+ "rmse": 0.5617979764938354,
1024
+ "bias": 0.002507680794224143,
1025
+ "mae": 0.39343899488449097,
1026
+ "true_cate_mean": 0.38851553201675415,
1027
+ "direction_acc": 0.4342857142857143,
1028
+ "correlation": -0.17480939647861074,
1029
+ "rank_correlation": -0.22579062452987952
1030
+ },
1031
+ {
1032
+ "dataset": 83,
1033
+ "regime": "confounded",
1034
+ "is_null": true,
1035
+ "rmse": 1.2210314273834229,
1036
+ "bias": -1.0801631212234497,
1037
+ "mae": 1.0967117547988892,
1038
+ "true_cate_mean": 0.40355435013771057,
1039
+ "direction_acc": 0.435,
1040
+ "correlation": 0.00020017295213392016,
1041
+ "rank_correlation": -0.41299273303759443
1042
+ },
1043
+ {
1044
+ "dataset": 84,
1045
+ "regime": "mediated",
1046
+ "is_null": false,
1047
+ "rmse": 1.2496973276138306,
1048
+ "bias": 0.98516845703125,
1049
+ "mae": 1.1042251586914062,
1050
+ "true_cate_mean": 0.4158540368080139,
1051
+ "direction_acc": 0.5109090909090909,
1052
+ "correlation": -0.10009513241217136,
1053
+ "rank_correlation": 0.11150905682933887
1054
+ },
1055
+ {
1056
+ "dataset": 85,
1057
+ "regime": "direct",
1058
+ "is_null": false,
1059
+ "rmse": 1.6278619766235352,
1060
+ "bias": 0.48378050327301025,
1061
+ "mae": 1.288899540901184,
1062
+ "true_cate_mean": 1.2557249069213867,
1063
+ "direction_acc": 0.46,
1064
+ "correlation": 0.6976954485818548,
1065
+ "rank_correlation": -0.03012473068437639
1066
+ },
1067
+ {
1068
+ "dataset": 86,
1069
+ "regime": "confounded",
1070
+ "is_null": true,
1071
+ "rmse": 0.7647475600242615,
1072
+ "bias": 0.764339804649353,
1073
+ "mae": 0.764339804649353,
1074
+ "true_cate_mean": 0.0,
1075
+ "direction_acc": null,
1076
+ "correlation": null,
1077
+ "rank_correlation": null
1078
+ },
1079
+ {
1080
+ "dataset": 87,
1081
+ "regime": "mediated",
1082
+ "is_null": false,
1083
+ "rmse": 0.9813193082809448,
1084
+ "bias": 0.6778125166893005,
1085
+ "mae": 0.7999677062034607,
1086
+ "true_cate_mean": 0.557606041431427,
1087
+ "direction_acc": 0.392,
1088
+ "correlation": -0.06099537249596694,
1089
+ "rank_correlation": 0.2942086037337185
1090
+ },
1091
+ {
1092
+ "dataset": 88,
1093
+ "regime": "direct",
1094
+ "is_null": false,
1095
+ "rmse": 0.9802929162979126,
1096
+ "bias": 0.725418210029602,
1097
+ "mae": 0.8183322548866272,
1098
+ "true_cate_mean": 0.4953778088092804,
1099
+ "direction_acc": 0.5169230769230769,
1100
+ "correlation": -0.10241924131202237,
1101
+ "rank_correlation": 0.05587145240861302
1102
+ },
1103
+ {
1104
+ "dataset": 89,
1105
+ "regime": "mediated",
1106
+ "is_null": false,
1107
+ "rmse": 0.21357114613056183,
1108
+ "bias": -0.036262623965740204,
1109
+ "mae": 0.13350236415863037,
1110
+ "true_cate_mean": 0.1193106546998024,
1111
+ "direction_acc": 0.54,
1112
+ "correlation": -0.33443272683742586,
1113
+ "rank_correlation": 0.22478835031013308
1114
+ },
1115
+ {
1116
+ "dataset": 90,
1117
+ "regime": "direct",
1118
+ "is_null": false,
1119
+ "rmse": 0.5598409175872803,
1120
+ "bias": 0.04221957176923752,
1121
+ "mae": 0.24207328259944916,
1122
+ "true_cate_mean": 0.21683551371097565,
1123
+ "direction_acc": 0.4991304347826087,
1124
+ "correlation": -0.022785294109784745,
1125
+ "rank_correlation": -0.12554913134530837
1126
+ },
1127
+ {
1128
+ "dataset": 91,
1129
+ "regime": "feedback",
1130
+ "is_null": false,
1131
+ "rmse": 0.3912670910358429,
1132
+ "bias": 0.23992693424224854,
1133
+ "mae": 0.31413596868515015,
1134
+ "true_cate_mean": 0.1853516548871994,
1135
+ "direction_acc": 0.34285714285714286,
1136
+ "correlation": 0.04385112267618765,
1137
+ "rank_correlation": 0.030411517586822118
1138
+ },
1139
+ {
1140
+ "dataset": 92,
1141
+ "regime": "mediated",
1142
+ "is_null": false,
1143
+ "rmse": 0.5043230652809143,
1144
+ "bias": -0.1925269514322281,
1145
+ "mae": 0.40458598732948303,
1146
+ "true_cate_mean": 0.35438355803489685,
1147
+ "direction_acc": 0.6355555555555555,
1148
+ "correlation": 0.4543439804631641,
1149
+ "rank_correlation": 0.16646037797720417
1150
+ },
1151
+ {
1152
+ "dataset": 93,
1153
+ "regime": "direct",
1154
+ "is_null": false,
1155
+ "rmse": 0.8582991361618042,
1156
+ "bias": 0.7085089087486267,
1157
+ "mae": 0.7516292333602905,
1158
+ "true_cate_mean": 0.3462170362472534,
1159
+ "direction_acc": 0.46,
1160
+ "correlation": 0.10202824107807708,
1161
+ "rank_correlation": -0.27922083922083923
1162
+ },
1163
+ {
1164
+ "dataset": 94,
1165
+ "regime": "confounded",
1166
+ "is_null": true,
1167
+ "rmse": 1.7765039205551147,
1168
+ "bias": 1.2695926427841187,
1169
+ "mae": 1.3785573244094849,
1170
+ "true_cate_mean": 1.0134458541870117,
1171
+ "direction_acc": 0.6,
1172
+ "correlation": 0.2171717151189476,
1173
+ "rank_correlation": -0.1849107078547139
1174
+ },
1175
+ {
1176
+ "dataset": 95,
1177
+ "regime": "confounded",
1178
+ "is_null": true,
1179
+ "rmse": 0.24133837223052979,
1180
+ "bias": 0.24062727391719818,
1181
+ "mae": 0.24062727391719818,
1182
+ "true_cate_mean": 0.0,
1183
+ "direction_acc": null,
1184
+ "correlation": null,
1185
+ "rank_correlation": null
1186
+ },
1187
+ {
1188
+ "dataset": 96,
1189
+ "regime": "confounded",
1190
+ "is_null": false,
1191
+ "rmse": 2.0764267444610596,
1192
+ "bias": 1.7063145637512207,
1193
+ "mae": 1.757581353187561,
1194
+ "true_cate_mean": 0.9683166146278381,
1195
+ "direction_acc": 0.48,
1196
+ "correlation": 0.16983647195980414,
1197
+ "rank_correlation": -0.24412860918194604
1198
+ },
1199
+ {
1200
+ "dataset": 97,
1201
+ "regime": "confounded",
1202
+ "is_null": true,
1203
+ "rmse": 0.4481605887413025,
1204
+ "bias": 0.4473162293434143,
1205
+ "mae": 0.4473162293434143,
1206
+ "true_cate_mean": 0.0,
1207
+ "direction_acc": null,
1208
+ "correlation": null,
1209
+ "rank_correlation": null
1210
+ },
1211
+ {
1212
+ "dataset": 98,
1213
+ "regime": "direct",
1214
+ "is_null": false,
1215
+ "rmse": 1.455250859260559,
1216
+ "bias": 0.837905764579773,
1217
+ "mae": 1.1487460136413574,
1218
+ "true_cate_mean": 0.9252989292144775,
1219
+ "direction_acc": 0.5356521739130434,
1220
+ "correlation": 0.20474727816951477,
1221
+ "rank_correlation": -0.0809512304441057
1222
+ },
1223
+ {
1224
+ "dataset": 99,
1225
+ "regime": "confounded",
1226
+ "is_null": false,
1227
+ "rmse": 0.898026704788208,
1228
+ "bias": -0.8334742188453674,
1229
+ "mae": 0.8447396159172058,
1230
+ "true_cate_mean": 0.17805933952331543,
1231
+ "direction_acc": 0.5009523809523809,
1232
+ "correlation": 0.054145654002977515,
1233
+ "rank_correlation": 0.2285817330280228
1234
+ }
1235
+ ]
1236
+ }