MOSS performs ensemble machine learning and Targeted Maximum Likelihood (TMLE) to estimate the counter-factual marginal survival functions, while non-parametrically adjusting for measured confounding. TMLE approach is employed to create a doubly robust and semi-parametrically efficient estimator. Simultaneous confidence bands of the entire curve is also available for inference. User can specify what kind of static intervention on treatment (exposure).
The following comparable methods are also included in the package for you to easily compare methods: - Inverse censoring probability weighted (IPCW) - Locally efficient one-step estimator (estimating equation methods)
install.packages('MOSS')
devtools::install_github('wilsoncai1992/MOSS')- To see all available package documentation:
?MOSS
help(package = 'MOSS')The data input of all methods in the package should be an R
data.frame in the following survival long data format:
# ID W A T.tilde delta
# 1 1 0 0 95 1
# 2 2 1 1 1 0
# 3 3 0 0 215 1
# 4 4 1 1 15 1
# 5 5 0 0 73 1
# 6 6 0 0 15 1- perform SuperLearner fit of the conditional survival function of
failure event, conditional survival function of censoring event,
propensity scores (
initial_sl_fit) - perform TMLE adjustment of the conditional survival fit
(
MOSS_hazard) - simultaneous confidence band (
compute_simultaneous_ci)
To cite MOSS in publications, please use:
Cai W, van der Laan MJ (2019+). One-step TMLE for time-to-event outcomes. Working paper.
This software is distributed under the GPL-2 license.
Feedback, bug reports (and fixes!), and feature requests are welcome; file issues or seek support here.