HAC Standard Errors

[s3alfisc]

See equations (9) and (10) in Various Versatile Variances. See also this set of lecture notes.

Implement utility functions _vcov_hc() and _vcov_hac() following the R sandwich package.

Allow arguments "hac" for pf.feols() and pf.fepois().

Likely, everything needs to be implemented in numba for performance reasons; the estimators are hard to vectorize.

[s3alfisc]

Hi @damandhaliwal, if you are looking for a challenge, I think this would be an interesting issue. It covers some econometrics, API design, and optimization (via numba). Would be a good issue to get started on via a pair programming session!

[damandhaliwal]

Hi Alex, sign me up! I can get started on this later in the week and get a good session in over the weekend. I'll read the notes you attached and the R sandwich package documentation.

[s3alfisc]

Awesome! I also have some code lying around in a branch, will share it with you

[s3alfisc]

So, for the API design I made the following comments - I thought it would be a good idea to add a new argument to the estimation functions, vcov_kwargs, that would allow users to pass information required for the panel HAC estimation:

    vcov : Union[VcovTypeOptions, dict[str, str]]
        Type of variance-covariance matrix for inference. Options include "iid",
        "hetero", "HC1", "HC2", "HC3", "NW" for Newey-West HAC standard errors,
        "DK" for Driscoll-Kraay HAC standard errors, or a dictionary for CRV1/CRV3 inference.
        Note that NW and DK require to pass additional keyword arguments via the `vcov_kwargs` argument.

    vcov_kwargs: dict[str, Any]
        Additional keyword arguments to pass to the vcov function. These keywoards include
        "lags" for the number of lags to use in the Newey-West (NW) and Driscoll-Kraay (DK) HAC standard errors.
        "time_id" for the time ID used for NW and DK standard errors, and "panel_id" for the panel
        identifier used for NW and DK standard errors.

Then, we'd have to add def _nw_meat and _dk_meat to vcov_utils.py. These would implement the meat matrix for the Newey West and Driskool-Kray vcov's. Here we can lazily use for loops and after we're done, JIT compile them via numba.

Lag lengths can either be specified by the user, or we default to the standard heuristic.

Maybe we can start with Driskol-Kraay, which is well explained in here - this is actually the best explainer of DK I have seen.

Some minor adjustments to Feols.vcov & tests and that would be it! =)

[s3alfisc]

Btw, at second thought, I actually think this is quite a challenging feature, but I'm here to help! =)

[damandhaliwal]

Thanks for sharing Alex. Does look challenging but I'm on it. I'll just start with DK in the paper you linked and then follow up with sandwich does it in R. Hoping to get a solid understanding this weekend.