We use cookies to distinguish you from other users and to provide you with a better experience on our websites. Close this message to accept cookies or find out how to manage your cookie settings.
Online ordering is currently unavailable due to technical issues. We apologise for any delays responding to customers while we resolve this. For further updates please visit our website: https://www.cambridge.org/news-and-insights/technical-incident
Joel A. Middleton (UC Berkeley), Mark A. Scott (NYU), Ronli Diakow (New York City Department of Education) and Jennifer L. Hill (NYU), "Bias Amplification and Bias Unmasking", Political Analysis 24/3
Selection committee: Patrick Brandt (UT Dallas, chair), Devin Caughey (MIT), Sunshine Hillygus (Duke) and Michael Alvarez (Cal Tech, ex officio)
Citation:
Middleton et al. offer new insights into the age-old problem of omitted variable bias. Specifically, they tackle the vexing question of how best to minimize bias in observational studies, given the near certainty that at least some confounders will be unobserved. They offer both theoretical and empirical contributions. Conceptually, they provide a novel decomposition of omitted variable bias into three components: (1) bias due to unobserved confounders, (2) bias due to the exclusion of observed covariates from the conditioning set, and (3) bias due to the inclusion of observed covariates, which can amplify existing biases. They then formally analyze the conditions under which covariate inclusion results in more bias than covariate exclusion. They show that inclusion can increase bias both through bias amplification and through what they label "bias unmasking", where conditioning on a covariate removes "good" bias that countervails unobserved confounders. Contrary to the existing emphasis on instruments, Middleton et al. demonstrate that even covariate sets that explain substantial variation in the outcome can cause bias amplification. They emphasize the potential for group fixed effects, often seen as a benign robustness strategy, to increase bias through both amplification and unmasking. Finally, Middleton et al. develop tools for sensitivity analysis to help applied researchers reason through these problems. These methodological contributions are cleverly and lucidly illustrated through constructed observational placebo studies, which demonstrate the striking increases in bias that can result from the inclusion of fixed effects. In sum, Middleton et al. advance a rapidly developing literature on several fronts, in ways that should interest both applied political scientists and statisticians across fields.