Open brash6 opened 1 day ago
I may have missed some of the previous steps, but I would add the logic to infer such parameters automatically after the current refactoring. It's an optimization, not really a must-have imho.
@bthirion one of the first step included some legacy code from @houssamzenati, so we are rather reverting back to the current state
not a must-have but not so complicated I think?
@brash6 yes I agree with what you wrote, it is indeed what we said yesterday @bthirion we are indeed trying to implement automatic inference of such variants, the only parameter that will be let is ratio='propensities' or 'density' for the mediator m when it is only discrete only. Would you like to also enforce a single option in that case? we thought with judith that in that case it was beneficial to let the choice.
@houssamzenati It's clear on my side, so :
_procedure
variable)_ratio
variable to enable the choice between propensities and density when m
is discreteI agree with the 3 points
yes you can remove DML in the PR after the points are treated and also when cross fitting is implemented (is back actually) in the code.
@houssamzenati Inside the med_bench_prototype code, you implemented MultiplyRobust'
fit
function like this :We can see that there's two discrimination variables :
ratio
andprocedure
From my understanding about our discussion today, with some questions in the middle :
ratio='propensities'
scenario of the MultiplyRobust estimator, wherem
is continuous (and multidimensional ?).procedure
variable enables to specify, in the scenario wherem
is continuous, if we want to "discretize"m
, in order to make the estimation faster ?procedure
variableratio
variable and detect automatically if we are in the density (m is discrete) or propensities (m is continuous or discrete) scenario ?!fyi @judithabk6