In November I gave a seminar for the Canadian Network for Observational Drug Effect Studies (CNODES) entitled

"Targeted Learning for the Estimation of Drug Safety and Effectiveness: Getting Better Answers by Asking Better Questions."

​The video presentation synced with the overhead slides is now online on the CNODES website.

In this talk I demonstrate the importance of taking a marginal approach to the estimation of the effects of medications. I focus on the "Targeted Learning" roadmap (van der Laan and Rose, 2011) which is used to construct robust estimators of causal effects.

Photo credit: Christina Esteban Photography, http://cepstudio.com

Our article, entitled 
"Double robust and efficient estimation of a prognostic model for events in the presence of dependent censoring"

has been published in the journal Biostatistics and is now available on Pub Med at http://www.ncbi.nlm.nih.gov/pubmed/26224070 .

This article presents two simple double robust estimators of a mean ("marginal") model when subjects are longitudinally censored. In our example, we were interested in modeling the probability of negative health outcomes in subjects with HIV under the scenario where all subjects remained on treatment that successfully controlled viral load. Subjects were artificially censored when they did not continue on treatment, which therefore required modeling adjustment.

In an effort to make double robust computation more accessible, we provide code for fitting similar models (with example usage). Thanks to the referees and the journal editors for encouraging open access code!

A new article entitled "The causal inference paradigm for network meta-analysis with implications for feasibility and practice" is currently available (pre-peer review) on arXiv stat: 
http://arxiv.org/abs/1506.01583

Network meta-analysis involves the aggregation of study results in order to contrast the effects of different treatments for a common condition. While standard meta-analysis only summarizes over one pair of treatment contrasts, network meta-analysis can combine a multitude of different treatments. 

In this article, we define a nonparametric "effect of interest" in a network meta-analysis over heterogeneous underlying populations. We then develop a set of restrictions that allow for estimation of this effect. We propose several estimators and compare their realistic small-sample performance in a simulation study. Finally, we demonstrate this approach in a real data example to evaluate the relative effectiveness of antibiotics on methicillin-resistant Staphylococcus aureus (MRSA) infection.