search

joshuaschwab / ltmle

Longitudinal Targeted Maximum Likelihood Estimation package
http://joshuaschwab.github.io/ltmle/
23 stars 16 forks source link

cum.g weight truncation scheme and untruncated weight diagnostics #10

Open osofr opened 10 years ago

osofr commented 10 years ago

Hi Josh,

1) Is the weight truncation (on predicted g) performed separately for each type of AC node or is it being applied to the joint probability of all interventions at a particular time point? For example, given predicted probabilities for treatment node, A, and two censoring nodes, C1, C2, at particular time-point, I can truncate each of them, with minimum of the joint P(A,C1,C2) being equal to (0.01)^3 (for default gbounds) or instead, I could truncate the joint probability of (A,C1,C2), with minimum now equal to 0.01. From what I can see in the code the truncation is being applied separately to each node, am I correct?

2) Is it possible to add optional output with untruncated cum.g, along with the usual truncated cum.g? We use untruncated cumulative g as a diagnostic tool for practical positivity violations, presumably this could also be relevant to other users. Current ltmle implementation requires us to ru-run the entire ltmle algorithm to get untruncated g, it would save us a great deal of time if we could get both truncated and untruncated cum.g at once... This could be a part of some "diagnostic" mode for ltmle()...

Thanks a lot! Best, Oleg

joshuaschwab commented 10 years ago

Hi Oleg,

1) Truncation of g is applied to the joint probability - see ltmle:::CalcCumG 2) That seems like a good idea, I'll add it to the next version. It may be a while before that's released, though. In the meanwhile, you could probably get it from the fit$g object returned from ltmle, although that could be a bit of a pain. Alternatively, you could hack it in there in ltmle:::EstimateG

  cum.g <- CalcCumG(gmat, gbounds)   cum.g.unbounded <<- CalcCumG(gmat, c(0, 1))  #add this line to set a global variable with the unbounded cum g   return(list(cum.g=cum.g, fit=fit))

Josh

From: osofr notifications@github.com To: joshuaschwab/ltmle ltmle@noreply.github.com Sent: Friday, May 16, 2014 3:10 PM Subject: [ltmle] cum.g weight truncation scheme and untruncated weight diagnostics (#10)

Hi Josh, 1) Is the weight truncation (on predicted g) performed separately for each type of AC node or is it being applied to the joint probability of all interventions at a particular time point? For example, given predicted probabilities for treatment node, A, and two censoring nodes, C1, C2, at particular time-point, I can truncate each of them, with minimum of the joint P(A,C1,C2) being equal to (0.01)^3 (for default gbounds) or instead, I could truncate the joint probability of (A,C1,C2), with minimum now equal to 0.01. From what I can see in the code the truncation is being applied separately to each node, am I correct? 2) Is it possible to add optional output with untruncated cum.g, along with the usual truncated cum.g? We use untruncated cumulative g as a diagnostic tool for practical positivity violations, presumably this could also be relevant to other users. Current ltmle implementation requires us to ru-run the entire ltmle algorithm to get untruncated g, it would save us a great deal of time if we could get both truncated and untruncated cum.g at once... This could be a part of some "diagnostic" model for ltmle()... Thanks a lot! Best, Oleg — Reply to this email directly or view it on GitHub.