APHA
Back to Annual Meeting Page
 
American Public Health Association
133rd Annual Meeting & Exposition
December 10-14, 2005
Philadelphia, PA
APHA 2005
 
4340.0: Tuesday, December 13, 2005 - 5:10 PM

Abstract #120593

Honest Confidence Intervals for Causal Effects in the Presence of Many Potential Confounders

James M. Robins, MD, Eric Tchetgen, and Lingling Li. Biostatistics, Harvard School of Public Health, 9999, Boston, MA 99999, 617-432-0205, robins@hsph.harvard.edu

Suppose given data on a dichotomous treatment T, a response Y, and a high (say 50) dimensional vector of pre-treatment covariates X, one wishes to estimate the effect of T on Y and is willing to assume no unmeasured confounding (ignorability given X) . Recently doubly robust estimators have become popular. These give valid confidence intervals for the treatment effect of length proportional to one over the square root of the sample size if either (but not necessarily both ) of (i) a ‘working' outcome regression (OR) model for the regression of Y on T and X, or (ii) a ‘working' propensity score (PS) model for the probability of treatment given X are correct.. However even a doubly robust estimator has the following problem. We get invalid confidence intervals if both models are wrong; furthermore with high dimensional X, due to lack of power, we cannot effectively test for misspecification of the working models. We describe more honest confidence intervals that are robust to mild to moderate misspecification of both working models. The price for this honesty is that the width of our intervals will be much greater, reflecting our true uncertainty. The new, more honest, intervals are based on higher dimensional U-statistics.

(the talk is based on joint work with Aad van der Vaart)

Learning Objectives: By the end of the talk, attendees will be able to

Presenting author's disclosure statement:

I wish to disclose that I have NO financial interests or other relationship with the manufactures of commercial products, suppliers of commercial services or commercial supporters.

Annual Statistical Methodology Session: Marginal Structural Models for Causal Inference

The 133rd Annual Meeting & Exposition (December 10-14, 2005) of APHA