Setting base period in DoubleMLDIDMulti

[oskasf]

Thanks for your work on this package and apologies if this is a silly question / I've missed something simple.

I'm wondering whether it's possible to specify the base period in DoubleMLDIDMulti. I have a multi period diff in diff with a non-staggered treatment timing.

Currently when I plot the effects the first period in the data is excluded, but I'm wondering if the functionality from the did package in R is possible. Here we can specify a universal base period, which normalizes the coefficient before treatment to 0 (see here). This seems to be the standard in running event studies currently.

Is this possible or would I need to modify something? Thanks in advance!

[SvenKlaassen]

Thank you for the feedback @oskasf!
Currently, there is no convenience feature which directly sets this for pre-treatment periods, but I think this would be a great addition.
In principle, this should be possible to "just" include the right set of gt_combinations, see Example Notebook.

[SvenKlaassen]

If I understand the documentation correctly, the universal base period in the did package, basically computes all ATTs with respect to one specific base period (see documentation which chooses g-anticipation-1 as base period).

This depends on the group, e.g. for the Example the group first treated in '2025-04' and base period '2025-03' this would correspond to the following periods

group = np.datetime64('2025-04')
base_period = np.datetime64('2025-03')

gt_dict = {
    "gt_combinations": [
        (group,
         base_period,
         np.datetime64('2025-01')),
        (group,
         base_period,
         np.datetime64('2025-02')),
        (group,
         base_period,
         np.datetime64('2025-04')),
        (group,
         base_period,
         np.datetime64('2025-05')),
    ]
}
dml_obj_universal = DoubleMLDIDMulti(dml_data, **(default_args| gt_dict))
dml_obj_universal.fit()

Remark that for all pre-treatment periods the base-period is actually after the evaluation period. This is currently not possible as the currently release does not allow for this specification, raising an error:

ValueError: The pre-treatment period must be before the evaluation period. Got t_value_pre 2025-03 and t_value_eval 2025-01.

We will create an issue for this, but it may take a while.
To disable the check you can comment out the following lines

doubleml-for-py/doubleml/did/utils/_did_utils.py

Lines 87 to 91 in 9df0711

	if t_value_pre > t_value_eval:
	raise ValueError(
	"The pre-treatment period must be before the evaluation period. "
	f"Got t_value_pre {t_value_pre} and t_value_eval {t_value_eval}."
	)

Then you should be able to run

dml_obj_universal.fit()
dml_obj_universal.summary

When plotting the effects you should also obtain nearly the desired plot but without the zero in the desired base period (which you can add to the figure manually):

dml_obj_universal.bootstrap(n_rep_boot=5000)
fig, axes =dml_obj_universal.plot_effects()
axes[0].scatter(base_period, 0, color='black', marker='o', zorder=5, label='Base Period')

This should be close to the implementation of the did package, which also basically sets a set of gt-combinations relative the base period (see https://github.com/bcallaway11/did/blob/bfeecac64ec431ceb6bb7360ba7765a97ac8803f/R/compute.att_gt.R#L89-L92 and https://github.com/bcallaway11/did/blob/bfeecac64ec431ceb6bb7360ba7765a97ac8803f/R/compute.att_gt.R#L177-L182).

[oskasf]

Thanks so much for your detailed response. This looks great as it should match the did package. It would be good to get an issue created for this! For now I think I will just make the change you suggested in a fork. Thanks so much for your help.

[SvenKlaassen]

This is the corresponding Issue #340

[SvenKlaassen]

Hello @oskasf,

I just wanted to mention that the new option is now available in version 0.10.1, see example

[oskasf]

That's great thanks so much for your help with this! Really appreciate it.