Causality: Difference-in-Differences

The tenth video in a series on causality. This video covers one of the most widely-used causal inference methods: Difference-in-Differences (or Diff-in-Diff, or DID, or DD...). See more information about the class at nickchk.com/econ305.html.

[@aldersgrens]

Very helpful video. Thank you!

[@siddharthkaushal5194]

thank you, great video!! very clear and very crisp :)

[@johannesh1741]

Very helpful, many thanks!

[@hamayoonshah1990]

Love your teaching

[@tannerhoughton6233]

Very helpful!!

[@Mostusernamesaretakn]

Excellent video! I enjoy the way you explain metrics. Will you eventually cover event study methods?

[@NickHuntingtonKlein]

Thank you! And yep! I cover event studies in my book ( theeffectbook.net ), both "finance-style" where there's just one time series and no control, and "econ-style" which I call "dynamic difference-in-differences." I'll be making a series of videos for the book at some point, and there will be event study videos then.

[@Mostusernamesaretakn]

@@NickHuntingtonKlein Thanks for the reply, Nick! Glad to hear there will be forthcoming videos and yes, I am familiar with your book - it’s an excellent resource and I am looking forward to getting a physical copy when it’s available!

[@prashant0104]

Excellent video! Can we use matching method (instead of fixed effects as you mention) to control for differences in the treatment and control group in addition to applying dif-in-dif?

[@NickHuntingtonKlein]

The way matching and DID are generally combined is to do both - match to get a control group, and then do fixed effects as normal with that group. Or, alternatively there is synthetic control which is a related method that's more like what you're describing

[@5wyujg789ikgmhgy77i5]

Hello Nick! Really informative video) It would be really cool to see a video on 'dinamic' DID. Curently I'm reading your book 'The Effect", and there are several things in chapter 18.2.4 I can't understend clearly: (1) when we are creating the centered time variable, what are the values of these variable for control group? (For treanment group they are smthing like -2,-1,0,1,2,3 etc), (2) then, after we created a bunch of dummie variables for these centered timings (for example 6 dummies for our -2,-1,0,1,2,3) and multiply each of these dummies by the current treatment variable, aren,t the dummies for pretreated periods (-2,-1,0) just wiped from the model as we multiply them by 0 ectually? It,s kind of wierd this time: then I don't understend somthing in econometrics I just watch your videos and everithing becomes clear)

[@NickHuntingtonKlein]

Glad you like the videos! The dynamic did video is coming eventually when I get to that chapter in my Effect video series. For now, check out lost-stats.github.io for info on this

[@5wyujg789ikgmhgy77i5]

@@NickHuntingtonKlein thank you!

[@5wyujg789ikgmhgy77i5]

Yep, it was really helpful, as for (2) issue: we shoul multiply by dummie for treatment group, not courrent treatment), now it makes sense and it is like I thought it to be. Maybe I've been missing something as English is not my native language, but now it is clear. Many thanks:0)

[@bakther]

Wonderful explanation as usual. If there is no way to have special group as a control, is it possible to use DID or I must think about other identification strategies. For instance, assume that I am interested to evaluate effectiveness of policy using all listed companies in certain country in which I can not split equities into control or treatment or use across border sample.

[@NickHuntingtonKlein]

Thanks! Yeah you can't do DID without a control group (otherwise it's just "D"). You would need a different design at that point, perhaps an event study if you think the assumptions for that are plausible.

[@bakther]

​@@NickHuntingtonKlein Thank you Dr. Nick. Actually, I thought the event design but it seems plausible for corporate announcements not policy evaluation. I thought about the regression discontinuity using cutoff point (using relevant scale) to distinguish groups, or using instrumental variable assuming there is only way to affect treatment variable.

[@NickHuntingtonKlein]

@@bakther if there's a cutoff to which firms got treated by the policy, RDD is the best bet. Otherwise event study may be the best you can do

[@adnanbakather4384]

@@NickHuntingtonKlein Thank you Nick for your comment.

[@mariaaguilera1414]

You are saving me!! Thank you! Is there any problem with putting the variable after different for each individual? i.e. different treatment timing? And how could i also control apart from fixed individual effects, an industry-by-year cohort fixed effect?

[@NickHuntingtonKlein]

Do you mean different individuals got treated at different times, or there was a single treatment time but you want to see the dynamic effect? For the former, fixed effects doesn't work properly, try something like the did package in R with the Callaway and Sant'anna estimator. For the latter, interacting treatment group with time-to-treatment dummies will do it, like in Autor 2003.

[@mariaaguilera1414]

@@NickHuntingtonKlein Yes, individuals got treatment at different years. Previous studies have done fixed effects and fixed year by industry effects with dummy variable Post and Treated. I will take a look at the did package then, thank you very much!!

[@NickHuntingtonKlein]

@@mariaaguilera1414 yes, it has been common for a long time to do varied-treatment-timing with fixed effects. It's only been in the past few years that we learned it doesn't work.

[@mariaaguilera1414]

@@NickHuntingtonKlein I am not being able to to create the "first treatment variable" .. and I need to submit my thesis today. I have a "post" variable indicating when it got post treatment... Would that be ok? Sorry to bother you again..

[@NickHuntingtonKlein]

@@mariaaguilera1414 you're submitting your thesis today? Just go ahead and do the outdated fixed effects method. You won't have time to read Callaway and Sant'anna and describe it in your paper anyway. Just add fixed effects for individual, fixed effects for time period, and a "currently treated" variable equal to 1 if you're in a treated group and it's currently after treatment starts, , and cluster your standard errors at the group level. feols(y~treated|group+time) from the fixest package will do this. Then in your paper cite Goodman-Bacon (2018) on DID and point out that this estimation method may lead to a bias

[@jake_gifford]

Can you apply a t-test to test whether the difference is significant?

[@NickHuntingtonKlein]

Sort of but it's not just a regular two group t test. A t test of the coefficient on the interaction term in the regression model (as specified in the chapter/videos) is better

[@rajat1770]

Very Helpful content! Can you please provide "mariel.RData"? It will be a great help :) Thanks :)

[@NickHuntingtonKlein]

It's available here github.com/NickCH-K/introcausality/blob/master/Lectures/mariel.RData