This material is published in: R. Neugebauer, M.J. Silverberg, M.J. van der Laan (2011). "Individualized Antiretroviral Initiation Rules." In M.J. van der Laan and S. Rose, Targeted Learning: Causal Inference for Observational and Experimental Data, Chapter 26. New York, Springer.


Targeted Maximum Likelihood Learning (TMLL) has been proposed as a general estimation methodology that can, in particular, be applied to draw causal inferences based on marginal structural modeling with observational data using either a point treatment approach (all confounders are assumed not to be affected by the exposure(s) of interest) or a longitudinal data approach (some confounders may be affected by one of the exposures of interest). While formal development of TMLL has included road maps for applications in longitudinal data approaches, real-life implementations have been restricted to studies based on a point treatment approach. In this article, we illustrate the application of TMLL using a longitudinal approach to investigate the comparative effectiveness in delaying onset of AIDS-defining cancers of two clinical guidelines regarding "when to start" combination antiretroviral therapy based on a patient's evolving CD4 count level. The analysis is based on observational data from the Kaiser Permanente electronic medical record to fit a non-parametric dynamic marginal structural model with the so-called inverse probability of action weighted-reduced data-targeted minimum loss-based estimator (IPAW-R-TMLE). The estimator is developed using formal results from previous, theoretical articles on TMLL before providing details on its implementation for this analysis.



Included in

Biostatistics Commons