Bias and Bias Correction in Multisite Instrumental Variables Analysis of Heterogeneous Mediator Effects

• Published in: Journal of educational and behavioral statistics 2014-02, Vol.39 (1), p.53-86
• Main Authors: Reardon, Sean F., Unlu, Fatih, Zhu, Pei, Bloom, Howard S.
• Format: Article
• Language: English
• Subjects: Analytical estimating; Class Size; Computation; Covariance; Education; Estimation bias; Estimation methods; Estimators; Federal Programs; Grade 1; Grade 2; Instrumental variables; Instrumental variables estimation; Kindergarten; Least Squares Statistics; Outcome Measures; Reading Programs; Sampling; Sampling bias; Standard error; Statistical Analysis; Statistical Bias; Tennessee
• ISSN: 1076-9986; 1935-1054
• DOI: 10.3102/1076998613512525

We explore the use of instrumental variables (IV) analysis with a multisite randomized trial to estimate the effect of a mediating variable on an outcome in cases where it can be assumed that the observed mediator is the only mechanism linking treatment assignment to outcomes, an assumption known in the IV literature as the exclusion restriction. We use a random-coefficient IV model that allows both the impact of program assignment on the mediator (compliance with assignment) and the impact of the mediator on the outcome (the mediator effect) to vary across sites and to covary with one another. This extension of conventional fixed-coefficient IV analysis illuminates a potential bias in IV analysis which Reardon and Raudenbush refer to as "compliance-effect covariance bias." We first derive an expression for this bias and then use simulations to investigate the sampling variance of the conventional fixed-coefficient two-stage least squares (2SLS) estimator in the presence of varying (and covarying) compliance and treatment effects. We next develop two alternate IV estimators that are less susceptible to compliance-effect covariance bias. We compare the bias, sampling variance, and root mean squared error of these "bias-corrected IV estimators" to those of 2SLS and ordinary least squares (OLS). We find that, when the first-stage F-statistic exceeds 10 (a commonly used threshold for instrument strength), the bias-corrected estimators typically perform better than 2SLS or OLS. In the last part of the article, we use both the new estimators and 2SLS to reanalyze data from two large multisite studies.