Direct implementation of unbalanced panel

[bcallaway11]

Currently, with unbalanced panel, we play a trick and run it through the repeated cross sections code and then fix the standard errors at the end. But it would be nice to have a specific implementation for unbalanced panel:

• This would likely be much faster
• There are issues related to time-varying covariates in this case, see: use with unbalanced panels #76
• Perhaps the sorts of underlying missing-at-random type assumptions that rationalize unbalanced panel could be weakened if we provide a direct implementation.

This one is probably a lot of work though!

[bcallaway11]

Perhaps one idea worth thinking about going forward is to "locally" balance. For example, if the user wants event studies that go from 3 periods before the treatment to 3 periods after, then we could get a balanced panel for the relevant periods only (this is broadly similar to the idea of "clean controls"), or perhaps we could even more locally balance by just using balanced observations w.r.t. that (g,t). Currently, we either "globally" balance the panel or do not balance at all.