Hi Nick, the way to think about it is that if you omit a variable, you're implicitly including the variable and restricting its coefficient (in any regression) to 0, unless that relationship is picked up by other variables you do include.

Gary

Gary King - Albert J. Weatherhead III University Professor - Director, IQSS - Harvard University

GaryKing.org - King@Harvard.edu - @KingGary - 617-500-7570 - Assistant: 617-495-9271

On Mon, Jul 13, 2020 at 7:28 PM Nick Eubank <nick@nickeubank.com> wrote:

Hi All,

I'm working with panel data (state-years) and a difference-in-difference design. I'm looking to fill some values in the dependent variable time series.

My main specification is thus just two-way fixed effects (state FEs and year FEs) and a treatment variable.

I've gotten the best results modeling the missing data by running a local polynomial regression (a la Honaker and King 2010) for each state, and then using those predicted values as the predictor in Amelia. (i.e. run I run a local polynomial of my DV against time for state S, fill those values in to my predictor variable for state S, repeat for all states)

Everything I've read says I should definitely include all variables I plan to use in my analysis in Amelia, but I worry about failing to meet multi-variate normality conditions with the FEs, and running the model with and without them, I'm not sure they're adding much.

Do I need to include all those FEs (i.e. will I introduce some weird bias in my subsequent analysis if I don't)? And if I do, is there anything I can do to deal with them definitely not being multi-variate normal (or do I not need to worry about that)?

Thanks!

Nick