All simulators
Good & Bad Controls
Eight structures from confounder to M-bias: bias and standard error of the treatment coefficient with and without the control.
Day 53: Adjustment Sets & Bad ControlsBad: amplifies the bias from U.
bias without / with Z
0.208 / 0.390
SE without / with Z
0.031 / 0.041
Exact population regressions from linear models (unit-variance disturbances), with 95% intervals for a sample of n. U nodes (dashed) are unobserved. Compare the coefficient on X without and with Z against the true total effect.