The Relative Performance of Collaborative Double Robust Targeted Maximum Likelihood Estimators

• Activity Number: 84
• Type: Contributed
• Date/Time: Sunday, August 1, 2010 : 4:00 PM to 5:50 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Kristin E. Porter*+ and Susan Gruber and Jasjeet Sekhon and Mark J. Van der Laan
• Companies: University of California, Berkeley and University of California, Berkeley and University of California, Berkeley and University of California, Berkeley
• Keywords: double robust ; targeted ; collaborative ; maximum likelihood

Abstract

We add to the current debate on the relative performance of double robust (DR) estimators, particularly in the context of sparse data. We compare a new class of collaborative targeted maximum likelihood estimators (C-DR-TMLE) to DR estimators in the literature. C-DR-TMLE have the unique feature of using likelihood-based cross validation to select a judicious model of the treatment mechanism. This refines the TMLE fluctuation of the initial estimate of the conditional distribution of the outcome, resulting in wished bias reduction under weaker conditions and with smaller increases in variance. Moreover, by penalizing the log-likelihood, C-DR-TMLE are even more targeted when data are sparse, avoiding hazards many methods typically cannot overcome. We illustrate the strong performance of C-DR-TMLE with simulations estimating causal effects and a population mean from incomplete data.