Assessing the properties of patient-specific treatment effect estimates from causal forest algorithms under essential heterogeneity

Table 4 Average Percentage Differences Between the Estimated Treatment Effects and True Treatment Effects from the Causal Forest Algorithm within the Generalized Random Forests Application (CFA-GRF) Under Partially Observed Heterogeneity Across Simulated Populations Which Differ by the Extent That Treatment Effect Influences Treatment Choice

Simulation	A	B	C	D	E	F
	Proportion of true (TE_i) influencing (ETE_i) at Treatment Choice – (K_j)^a	% of Treatment Choice Variation Explained by (TE_i)^b	% of Patients Overlapped^c	Average Percentage Difference Between True and Estimated Treatment Effects
		% of Treatment Choice Variation Explained by (TE_i)^b	% of Patients Overlapped^c	Full Population	Treated	Untreated
1	0	.0006	100	-1.16%	-1.08%	-1.27%
2	.10	.18	100	1.12%	-0.20%	2.44%
3	.20	1.4	100	4.12%	0.57%	8.02%
4	.30	5.3	100	6.44%	-0.36%	14.87%
5	.40	11.9	100	12.12%	0.72%	27.69%
6	.50	20.1	100	15.08%	0.27%	37.40%
7	.60	27.8	100	18.68%	0.58%	48.22%
8	.70	34.5	100	18.60%	-1.51%	53.42%
9	.80	39.8	100	24.36%	1.27%	66.03%
10	.90	44.3	100	22.00%	-1.75%	66.51%
11	1.00	48.0	100	25.44%	-1.03%	76.28%

^aThe proportion of patient-specific TE_i knowledge used by decision makers in simulation “j” in developing the expected treatment effect for patient “i” that is distinct from the population average treatment effect based on the equation ETE_i = K_j * (TE_i(X_1i,X_2i,X_3i,X_4i,X_5i,X_6i)—.25) + .25
^bThe percentage of treatment choice variation explained by TE_i using a linear probability model of treatment choice T_i on true TE_i using SAS PROC REG procedure with the SCORR1 option
^cPercentage of patients in sample with treatment propensity score greater than .05 and less than .95 when only X_1i, X_2i, X_3i, X_4i factors are specified in the propensity score equation

ISSN: 1471-2288