We used simulated cohort data sets to investigate the effect of varying the sample size for subsample and 2-sample IV estimators on power, precision, and bias. For each simulated scenario, we generated 10,000 data sets with 10,000 observations on 4 variables: a genetic susceptibility score used as the IV (G), an exposure (X) influenced by G, an outcome (Y) influenced by X, and a confounding variable (U), assumed to be unmeasured, with effects on both X and Y. G and U were generated randomly from a standard normal distribution. X was also a randomly generated standard normal variable with linear effects exerted by G and U: (1)
