Effects of Conditioning on Instrumental Variables on Bias and Variance of Effect Estimates

• Activity Number: 88
• Type: Contributed
• Date/Time: Sunday, July 31, 2011 : 4:00 PM to 5:50 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Jessica Amelia Myers*+ and Joshua Gagne and Jeremy Rassen and Krista Hubrechts and Kenneth Rothman and Sebastian Schneeweiss and Robert Glynn
• Companies: Brigham and Women's Hospital and Brigham and Women's Hospital and Brigham and Women's Hospital and Brigham and Women's Hospital and Boston University and Brigham and Women's Hospital and Brigham and Women's Hospital
• Keywords: Instrumental variables ; Bias amplification ; Variable selection ; Unobserved confounding ; Unmeasured confounding ; Confounder adjustment

Abstract

Recent literature has reported evidence that conditioning on an instrumental variable (IV) can increase both bias and standard error of exposure effect estimates. Although these findings have obvious implications in cases of known IVs, their meaning remains unclear in the more common scenario where it is uncertain if a measured covariate meets the criteria for an IV. We present two simulation studies designed to gain insight into the problem of conditioning on IVs in routine epidemiologic practice. The simulations explore the effects on bias, variance, and mean squared error of conditioning on IVs, near-IVs (variables that are weakly associated with outcome), and confounders. The results indicate that estimates of exposure effect conditional on a true IV or near-IV may have larger bias and standard error than the unconditional estimate. However, in most common epidemiologic analyses, the potential for increases in bias and variance due to conditioning on an IV is small compared with the total estimation error. Therefore, minimizing unmeasured confounding should be the priority when selecting variables to be used in adjustment, even at the risk of conditioning on IVs.