search

nhejazi / biotmle

:package: :microscope: R/biotmle: Targeted Learning with Moderated Statistics for Biomarker Discovery
https://code.nimahejazi.org/biotmle/
Other
4 stars 2 forks source link

Outcomes method for biomarkers #30

Closed nhejazi closed 5 years ago

nhejazi commented 7 years ago

It appears that the idea of considering the impact of biomarkers on some kind of outcome measures was rather ill-conceived -- that is, with the current use of tmle::tmle for assessing the impact of biomarkers (on some kind of outcome via the ATE) requires that the expression measures associated with the biomarkers be discretized.

Generally speaking, there does not exist a reasonable way to discretize expression measures -- there is great disagreement in the genomics / high-dimensional biology literature -- and any estimates based on discretized biomarker expression values are likely to be un-meaningful scientifically.

nhejazi commented 7 years ago

For now, it's best to remove the outcome option in the main biomarkertmle function. In future, we can consider ways to approach this -- with the most obvious idea being to estimate a different parameter entirely (than the ATE), perhaps using the tmle.npvi package.

nhejazi commented 6 years ago

It seems that the implementation of a TMLE for a target parameter interpretable as the counterfactual value of the outcome of interest under a posited shift of the observed treatment value, available in the txshift R package, could be useful for this purpose.

nhejazi commented 5 years ago

At the present, implementing such an extension of the methodology appears to be outside the scope of this R package.

nhejazi commented 5 years ago

Outside the scope of this package. An equivalent shrinkage-based variance estimator should likely be implemented separately in other packages so as to better make use of each of their (likely very different) APIs.