Dynamic panel data models

[nilshg]

As discussed on Slack, I think dynamic panel data models would be a great enhancement for this package.

Some things off the top of my head that this would require:

• Time/observation aware tables (for difference and lag operators)
• GMM implemenentation (there's one in mcreel/econometrics.jl but I guess we'd want one in this package?)
• API decision - how much "hand holding" do we want to do? E.g. Stata requires the user to specify which variables are instrumented or not and then auto-generates a set of instruments, which are customisable (e.g. using levels or differences, restricting lags); another option would be to offer some "out of the box" estimators e.g. Arellano-Bond, Blundell-Bover etc. where the correct settings for instruments and lags used are pre-canned.

The first two points on this list probably also require thinking about whether they should live elsewhere, given they are more widely applicable than just dynamic panel data models.

[Nosferican]

Dynamic panel models are not my expertise, but as the tools become available we could try patching something together. Any update on the ecosystem developments?

[nilshg]

I haven't seen anything, but also haven't been following very closely. There's a new project coming up though which will probably require Arellano-Bond though, so I might be revisiting this soon.

I actually preserved our old discussion about this on Slack by keeping a draft message in the thread, maybe I'll move that over to Zulip when I get back to this and we can continue the discussion there (and maybe rope in others who are interested)?

[Nosferican]

Sure. I am also paying attention to https://github.com/JuliaArrays/ShiftedArrays.jl/pull/37 which might make it a bit easier to work with panel data as StatsModels supports time-aware operations.