Estimating Fixed Effects (The Effect, Videos on Causality, Ep 46)

Please visit www.theeffectbook.net to read The Effect online for free, or find links to purchase a physical copy or ebook.
The Effect is a book about research design and causal inference. How can we use data to learn about the world? How can we answer questions about whether X causes Y even if we can't run a randomized experiment? The book covers these things and plenty more. These videos are meant to accompany the book, although they can also be viewed on their own.
This video relates to material found in Chapter 16 of the book.
A version of this video without background music can be found here: • Estimating Fixed Effec...
So we've got the notion that we can control for some unmeasured variables that sit on back doors by controlling for individual (using fixed effects). But how can we actually do that? This video talks about how to estimate a model with fixed effects.

[@qinghuafeng1705]

Because of your videos, I think I understand fixed effects now. I appreciate your excellent explanations and your replies for our questions!

[@rosch10]

Hi Nick, maybe you can help: Is there a problem in running a logit fixed effect model for my paneldata where my dependent variable is representing change (e.g. increase or decrease in worries about something) and my independet variable of interest is also a binary variable representing the happening of an event (e.g. losing a job/family member/ a particular amount of income)?

[@NickHuntingtonKlein]

Nothing wrong with that I don't think. Of course you'd want to use a logit Fe estimator designed for it rather than, say, adding a categorical control.

[@Tee_Cee_]

EXCELLENT resource! I do have a question though... Does the TWFE equation presented in this video (Yit = βi + βt + β1Xit + ε) allow me to separately control for/distinguish between entity and time effects? Or is that equation treating both effects as part of the overall model parameters? The book reads as through temporal effects are parallel for all entities in my panel dataset, but what if the entities have different things occurring in the same year amongst them? BC of that, I'm considering Yit = α + βXit + γi + δt + ϵit, but I'm interested in your take on how best to handle such. 👀

[@NickHuntingtonKlein]

The time effects that TWFE models are specifically the time effects that are shared across all units. So if different entities have different things occurring in the same year, TWFE won't control that away. But it doesn't want to - the reason you'd use TWFE is if you think "the different things occurring in the same year across entities" *is* your variation of interest that you want to drive your estimation of the coefficient on the Xs. It's worth thinking carefully on whether you actually want to distinguish those - if you have one observation per period per entity, then allowing time effects to vary completely across entities swallows up all your degrees of freedom and there's no data left to estimate anything with. You would need to restrict it in *some* way. Also a bit confused by your two equations - those appear to be identical (other than whether you drop an i/t reference group), just written with different notation.

[@Tee_Cee_]

I'm moving too fast & copied my alternate equation. But TY for your insight! IT HELPS!!! @@NickHuntingtonKlein