Sean Reardon
Published: 2011
Total Pages: 7
Get eBook
The proposed paper studies the bias in the two-stage least squares, or 2SLS, estimator that is caused by the compliance-effect covariance (hereafter, the compliance-effect bias). It starts by deriving the formula for the bias in an infinite sample (i.e., in the absence of finite sample bias) under different circumstances. Specifically, it considers the following cases: (1) A single site study with one mediator; (2) A multiple site study with one mediator; and (3) A multiple site study with multiple mediators. The formulas demonstrate how the magnitude of the compliance-effect bias varies with different parameters (e.g., compliance-effect correlation, mean and variance of the compliance and effect) in infinite samples. However, as the situation under consideration gets more complicated, the bias formula quickly becomes intractable. The second part of the paper, therefore, uses simulations to demonstrate the relationship between the compliance-effect bias and various parameters, as well as the behavior of the estimated 2SLS standard errors. Furthermore, the simulation exercise assesses how the compliance-effect bias interacts with the finite sample bias when the analysis sample is small or when the instrument is weak. The paper also uses simulations to compare the properties of the 2SLS estimator with those of the ordinary least squares (OLS) estimator in the presence of the compliance-effect bias, the finite sample bias, and the omitted variable bias. (Contains 6 footnotes.).