Multiply robust

[sami6mz]

Multiply_robust estimator is fully implemented according Tchetgen Tchetgen 2012 and Huber 2016.
For linear generated data (even with x multidimensional), relative error is around 5% even for indirect effects, which is a noteworthy performance.
Docstring was changed, code was made readable, and .ravel() input was rectified (solving #34 and making #14 to progress).

Test code :

from med_bench.src.get_simulated_data import simulate_data
from med_bench.src.benchmark_mediation import *
data = simulate_data(1000, default_rng(4), False, False, 5, 1, 7, "binary", 0.5, 0.5, 0.5, 0.5)
x = data[0]
t = data[1]
m = data[2]
y = data[3]
effects = np.array(data[4:9])

effects_chap = multiply_robust_efficient(y, t, m, x)[0:5]
error = abs((effects_chap - effects) / effects)
print(effects)
print(error)

print(multiply_robust_efficient(y.ravel(), t.ravel(), m, x) == multiply_robust_efficient(y, t, m, x))

[sami6mz]

my bad, beb5f26 should be named solving #34

[judithabk6]

not sure this is the wanted behavior, the last returned value should have the same meaning in all methods, so the number of discarded observations by trimming, not an alternance of clipped and trimmed examples. WDYT @sami6mz @bthirion?

[bthirion]

+1 for consistency

[bthirion]

More importantly, there should be 6 output arguments, not one list.

[sami6mz]

not sure this is the wanted behavior, the last returned value should have the same meaning in all methods, so the number of discarded observations by trimming, not an alternance of clipped and trimmed examples. WDYT @sami6mz @bthirion?

Then should we keep clipping? Or replace clipping by trimming?

[judithabk6]

keep clipping but not return the number of clipped examples. And potentially open an issue to implement trimming in this method. And another issue to implement clipping in all methods and return 6 results instead of 5 in that case

[sami6mz]

I ended up removing clipping count with df600a9 + putting the clipping code in issue #39

[sami6mz]

keep clipping but not return the number of clipped examples. And potentially open an issue to implement trimming in this method. And another issue to implement clipping in all methods and return 6 results instead of 5 in that case

Let's just implement trimming for now, and remove clipping

[bthirion]

Thx for the great job !

[bthirion]

is there any reason to sue this patterns. Arguments should be passed explicitly.

[judithabk6]

you mean to use args directly rather than renaming it variables?

[judithabk6]

actually you can specify a list of variables between which you want to compute interaction terms. Do you think this should be done differently?

[bthirion]

+1 for consistency

[bthirion]

More importantly, there should be 6 output arguments, not one list.