Fig. 4
Scenario B: Comparison of bias of treatment effects between models with and without adjusting for near-IV (X3B) varying its relationship with unmeasured confounder (X1). Beta (β1 in X3Bi ~ N (β1X1, 1)) ranged from 0.01 to 0.3. Y-axis shows bias which was calculated by subtracting true values of marginal expectations obtained in a large (N = 10,000,000) sample from g-computation estimates across the 10,000 datasets in each situation. Biases of [Y1, 1] – E [Y0, 0] were multiplied by − 1 to provide intuitive information on the gap from the true estimates because the true estimates of [Y1, 1] – E [Y0, 0] were negative