Table 3:

Simulation results for estimators under different approaches to estimation of the nuisance functions

	Bias	RMSE	ASE	ESE	CLD	Coverage
G-computation
True	0.000	0.017	0.017	0.017	0.065	93.5%
Main-effects	−0.023	0.029	0.017	0.018	0.067	72.3%
Machine learning	0.026	0.031	0.015	0.017	0.058	56.5%
IPW
True	0.007	0.025	0.025	0.024	0.097	94.9%
Main-effects	−0.022	0.032	0.023	0.023	0.091	86.6%
Machine learning	0.010	0.023	0.023	0.021	0.090	94.8%
AIPW
True	0.000	0.021	0.020	0.021	0.077	93.9%
Main-effects	−0.016	0.026	0.020	0.020	0.076	84.4%
Machine learning	0.004	0.020	0.017	0.019	0.066	91.3%
TMLE
True	0.000	0.021	0.020	0.021	0.077	93.6%
Main-effects	−0.017	0.025	0.019	0.018	0.075	84.9%
Machine learning	−0.002	0.020	0.017	0.020	0.065	89.5%
DC-AIPW
True	0.000	0.021	0.022	0.021	0.085	95.2%
Main-effects	−0.015	0.026	0.027	0.022	0.106	92.4%
Machine learning	−0.001	0.020	0.021	0.020	0.082	95.6%
DC-TMLE
True	0.001	0.020	0.021	0.020	0.084	95.8%
Main-effects	−0.018	0.025	0.024	0.018	0.094	91.4%
Machine learning	0.000	0.020	0.020	0.020	0.079	95.2%

RMSE: root mean squared error, ASE: average standard error, ESE: empirical standard error, CLD: confidence limit difference, Coverage: 95% confidence limit coverage of the true value

IPW: inverse probability of treatment weighted estimator, AIPW: augmented inverse probability weighted estimator, TMLE: targeted maximum likelihood estimator, DC-AIPW: double cross-fit AIPW, DC-TMLE: double cross-fit TMLE.

True: correct model specification. Main-effects: all variables were assumed to be linearly related to the outcome and no interaction terms were included in the model. Machine learning: super-learner with 10-fold cross-validation including empirical mean, main-effects logistic regression without regularization, generalized additive models, random forest, and a neural network.