Direct estimation of the mean outcome on treatment amidst early treatment termination
Several authors have investigated the challenges of statistical analyses  and inference amidst  early treatment termination, including a loss of efficiency in randomized  controlled trials and its connection to dynamic regimes in observational  studies.  Popular estimation strategies for causal estimands in dynamic regimes    lend themselves to studies where treatment is assigned at a finite  number of points; the extension to continuous treatment assignment   is non-trivial and can lead to other challenges depending on the dynamic regime  under investigation.  We re-examine  this problem from a different perspective and propose a new direct estimator  for the mean outcome of a target treatment length policy  that does not model the propensity score. Because this strategy does not include a model for treatment selection or assignment,  the estimator works well in both discrete and continuous time and avoids finite sample bias associated with squeezing continuous time data into intervals. We show how the competition of treatment assignment and terminating event through  time leads to a type of competing risks problem and, hence, concepts  from survival analysis.  The technique is exemplified through small  sample numerical studies and the analysis of two real data sets.

Bio: Brent Johnson earned a MS in Biostatistics from the UMN in 1997, worked for 2 years at Hopkins, went back to school and earned his PhD in Statistics from NCSU in 2003.  He worked as a postdoc in Biostatistics at UNC-CH 2003-2006 and started at Emory in 2006.  He is now Associate Professor of Biostatistics at Emory University.