Implement DynamicDML

[moprescu]

• DynamicDML with all the OrthoLearner functionality
• API tests similar to the DML ones

[kbattocchi]

Here are some comments - please address what you're able to.

[vsyrgkanis]

Could you also add a snippet of code in the README, as it seems a substantive addition to advertise!

[vsyrgkanis]

The text that lies under the table in the summary results is good for a static setting, but might not be the best for the summary results displayed in the dynamic setting. So we might want to not show that text here in the summary

[vsyrgkanis]

In the plot of effects in the notebook, is theta_0 the effect from treatments in period 0 or the effect of treatment at lag 0 from the outcome. Similarly for the rest of the params.

[vsyrgkanis]

In the notebook when you say: "number of panels" do you mean "number of units"

[vsyrgkanis]

The results in the notebook, even with 5000 units are very poor! We need to investigate this before we push:

[vsyrgkanis]

Results are almost identical when n_periods =1
DynamicDML

LinearDML

So there is something when we have multiple periods that gets messed up

[vsyrgkanis]

Oh!! It was just the parameters of lassocv you set in the notebook that mess thigns up. So just leae the default tol and the default alpha grid. That improves performance significantly

and also matches the LinearDML.

Also I would say we might want to reverse the parameters to mean what you are descrbing and not the lag version that was in the prototype.

So the (T0)_0 should mean the effect of period 0 on final outcome and (T0)_m the effect of final period.

Currrently it's the reverse and that messes also the "effect(X, T0, T1)"
here one needs to give T0 and T1 in the reverse order too.

[moprescu]

Oh!! It was just the parameters of lassocv you set in the notebook that mess thigns up. So just leae the default tol and the default alpha grid. That improves performance significantly

and also matches the LinearDML.

Also I would say we might want to reverse the parameters to mean what you are descrbing and not the lag version that was in the prototype.

So the (T0)_0 should mean the effect of period 0 on final outcome and (T0)_m the effect of final period.

Currrently it's the reverse and that messes also the "effect(X, T0, T1)"
here one needs to give T0 and T1 in the reverse order too.

Fixed!