Targeted Learning with Moderated Statistics for Biomarker Discovery

Authors: Nima Hejazi, Mark van der Laan, and Alan Hubbard

The biotmle R package facilitates biomarker discovery through a generalization of the moderated t-statistic (Smyth 2004) that extends the procedure to locally efficient estimators of asymptotically linear target parameters (Tsiatis 2007). The set of methods implemented modify targeted maximum likelihood (TML) estimators of statistical (or causal) target parameters (e.g., average treatment effect) to apply variance moderation to the standard variance estimator based on the efficient influence function (EIF) of the target parameter (van der Laan and Rose 2011, 2018). By performing a moderated hypothesis test that pools the individual probe-specific EIF-based variance estimates, a robust variance estimator is constructed, which stabilizes the standard error estimates and improves the performance of such estimators both in smaller samples and in settings where the EIF is poorly estimated. The resultant procedure allows for the construction of conservative hypothesis tests that reduce the false discovery rate and/or the family-wise error rate (Hejazi, van der Laan, and Hubbard 2021). Improvements upon prior TML-based approaches to biomarker discovery (e.g., Bembom et al. (2009)) include both the moderated variance estimator as well as the use of conservative reference distributions for the corresponding moderated test statistics (e.g., logistic distribution), inspired by tail bounds based on concentration inequalities (Rosenblum and van der Laan 2009); the latter prove critical for obtaining robust inference when the finite-sample distribution of the estimator deviates from normality.

For standard use, install from Bioconductor using BiocManager:

if (!requireNamespace("BiocManager", quietly=TRUE)) {
  install.packages("BiocManager")
}
BiocManager::install("biotmle")

To contribute, install the bleeding-edge development ver