Abstract:
|
In non-experimental studies, conditioning on observable covariates is one way to try to reduce unobserved confounder bias. However, a developing literature has shown that there exist covariates which, when conditioned on, amplify bias due to unobservables. A special case of bias amplification, for example, occurs when conditioning on instruments. We describe this phenomenon in the context of a linear model. We then identify a new special case of bias amplification, in particular, when so-called fixed effects (indicators for groups) amplify bias. Bias can be amplified even when the fixed effects do not act as instruments and they absorb heterogeneity in (and are causally related to) the outcome. A method of visualizing the conditions under which fixed effects are bias amplifying is proposed and an expression for the bias is derived. Nonetheless, results suggest that researchers must bring substantive knowledge to bear when determining which covariates to include in a causal analysis. These results also suggest that advice that adding fixed effects ``might help but should not hurt" could be dangerous.
|
ASA Meetings Department
732 North Washington Street, Alexandria, VA 22314
(703) 684-1221 • meetings@amstat.org
Copyright © American Statistical Association.