Table 3.

ITE estimation errors (lower is better) at varying complexities for self-interpretable models.

Modality →			Tabular
Dataset →			IHDP		TCGA		IHDP-C
Method ↓	Trees	Depth	${ϵ}_{\text{ATE}}$(In-sample)	${ϵ}_{\text{ATE}}$(Out-of-sample)	${ϵ}_{\text{ATE}}$(In-sample)	${ϵ}_{\text{ATE}}$ (Out-of-sample)	AMSE
Decision Tree	-	6	0.693±0.028	0.613±0.045	0.200±0.012	0.202±0.012	21.773±0.190
	-	100	0.638±0.031	0.549±0.052	0.441±0.004	0.445±0.004	23.382±0.342
Random Forest	1	6	0.801±0.039	0.666±0.055	19.214±0.163	19.195±0.163	21.576±0.185
	1	100	0.734±0.041	0.653±0.056	0.536±0.011	0.538±0.012	33.285±0.940
	10	100	0.684±0.033	0.676±0.034	0.536±0.011	0.538±0.012	38.299±0.841
NAM	-	-	0.260±0.031	0.250±0.032	-	-	24.706±0.756
ENRL	1	6	4.104±1.060	3.759±0.087	10.938±2.019	10.942±2.019	24.720±0.985
	1	100	4.094±0.032	4.099±0.107	10.938±2.019	10.942±2.019	24.900 ± 0.470
Causal Forest	1	6	0.144±0.019	0.275±0.035	-	-	-
	1	100	0.151±0.019	0.278±0.033	-	-	-
	100	max	0.124±0.015	0.230±0.031	-	-	-
BART	1	-	1.335±0.159	1.132±0.125	230.74±0.312	236.81±0.531	12.063±0.410
	N	-	0.232±0.039	0.284±0.036	-	-	4.323±0.342
DISCRET (ours)	-	6	0.089±0.040	0.150±0.034	0.076±0.019	0.098±0.007	0.801±0.165
TransTEE + DISCRET (ours)*	-	-	0.082±0.009	0.120±0.014	0.058±0.010	0.055±0.009	0.102±0.007

We bold the smallest estimation error for each dataset, and underline the second smallest one. Results in the first row for each method are duplicated from Table 2. For BART, we set $N=200$ for IHDP, and $N=10$ for TCGA and IHDP-C due to large feature number of features in the latter. We show that DISCRET outperforms self-interpretable models and has simpler rules regardless of the model complexity used. Asterisk (*) indicates model is not self-interpretable.