|
Activity Number:
|
309
- Advances in Causal Inference
|
|
Type:
|
Contributed
|
|
Date/Time:
|
Tuesday, July 30, 2019 : 8:30 AM to 10:20 AM
|
|
Sponsor:
|
Section on Statistics in Epidemiology
|
|
Abstract #307018
|
Presentation
|
|
Title:
|
On the Robustness of Doubly Robust Estimators in Causal Inference
|
|
Author(s):
|
Weicong Lyu* and Peter Steiner
|
|
Companies:
|
University of Wisconsin-Madison and University of Wisconsin
|
|
Keywords:
|
Doubly robustness;
Causal inference;
Propensity scores
|
|
Abstract:
|
Doubly robust estimators (DREs) are widely used in causal inference because they are unbiased whenever either treatment selection or the outcome is correctly modeled. DREs rely on an implicit but in practice rarely noticed condition: The predictors in the selection and outcome model must be identical. Doubly robustness is not guaranteed if the two models use different predictors, even if one of the two models is correctly specified. This paper reviews two classes of commonly used DREs and uses theory and simulations to demonstrate under which conditions doubly robustness breaks down despite a correctly specified selection or outcome model. The first class of DREs relies on independently estimated selection and outcome models (Bang & Robins, 2005) and is in general biased whenever the selection and outcome models use different predictors. DREs of the second class use inverse-probability-of-treatment weights to fit the outcome model (Schafer & Kang, 2008) and are biased only when a collider variable is the reason for a misspecified model. The paper also addresses how statistical software packages implement DREs and concludes with suggestions for practice.
|
Authors who are presenting talks have a * after their name.
