Fig. 4.

Comparison of adjusted 2SLS IV model estimates from 200 simulated data sets. Different strengths of confounding are given across the x-axis. Two different strengths of IV α₁ = 0.1 (weak) α₁ = 0.5 (strong) and sample sizes (N = 20,000, 200,000) are given across the y-axis; dashed blue horizontal line is the true causal treatment effect β₁ = 3.0, error bars are 95% confidence intervals for the treatment effect estimates for each simulated data set. Black error bars indicate a significant effect estimate, red error bars indicate a non-significant effect estimate