@@saurabhmahajan4459 not by itself, no, but having some programming knowledge will be expected in a lot of Economics jobs

@@pedrogustavo6823 see the chapter, which covers this. But basically, it's the exact same process, just with different start and end points.

This can work, but only in very specific circumstances. If, for example, there's actually another variable Z here that causes both drugs and lifespan, then controlling for drugs opens up another back door via collider bias (wine -> drugs <- Z -> lifespan) and you aren't getting that direct effect you want.

@@NickHuntingtonKlein That makes a lot of sense, thanks a lot! After reading this reply, having read the chapter and watching the episodes till #20 so far, I con affirm that colliders will be my worst enemy.

@@nushinqobil8670 I'm afraid I don't, sorry

@@wangguan1548 as I use the term, event studies are for cases where you are not using a comparison group. So you don't need a comparison group for this, although conversely this approach is only appropriate when you don't need one.

@@NickHuntingtonKlein Sorry, I may misunderstand, is it (event study) similar to interrupted time series, if without the comparison groups in this context?

@@wangguan1548 correct. I recommend checking out the event study chapter in the book. www.theeffectbook.net/ch-EventStudies.html

@@NickHuntingtonKlein Thanks, very clear

@@wangguan1548 yes, no reason why not. RDD often is used with cross sectional data.

@@NickHuntingtonKlein Thanks for confirming

@@wangguan1548 this works mathematically and is called an Interrupted Time Series. However it is usually less plausible that time is a good running variable that satisfies all the assumptions you need, so this is usually considered lower quality than other RDD studies.

@@NickHuntingtonKlein Thanks, so ITS could be considered as an exceptional case for event study

@@mohakraitani6228 csdid has a matching step for applying (time-constant, measured pre-treatment) covariates. So it will do it automatically for you.

@@eserhaneser I don't know much about it

@@NickHuntingtonKlein Thanks for the quick response! I’m looking for the most current, widely accepted method to evaluate the impact of a law that goes into effect in various states at different times while remaining inactive in others. I’m working on a paper, and I’m hearing a lot of conflicting advice. From my review of the literature, it seems like researchers often select a method that demonstrates the desired effect, but I want to ensure I’m using a sound, unbiased approach. Could you recommend a rigorous method for this type of analysis?

@@eserhaneser unfortunately the answer is that different methods rely on different assumptions, and fail or are weak in different circumstances. So there isn't a single best option.

I go into extensions of DID in the DID chapter of my book ( theeffectbook.net/ch-DifferenceinDifference.html ). I don't cover DiDC explicitly but I do cover DIDID which is only a small step away. Basically in DiDC you have an RDD that you think will only go into effect at a specific time. So interact your regular RDD specification with a before/after indicator, and the change in the RDD effect from before to after is your DiDC effect. Sometimes you do not just a cutoff that becomes active at a certain time but also for a certain group. Then you interact your RDD specification with both before/after, affected/unaffected group, and hte interaction between the two (basically a DID specification). The coefficient on the interaction between the DID interaction term and the RDD effect shows how the RDD effect change from before to after for the treated group relative to how it changes for the control group.

@@NickHuntingtonKlein Thank you so much for the explanation, Nick!! I'm a final year PhD candidate in Health Economics at Stellenbosch University. I've used your book extensively and it's fantastic! ;-)

Yes, typically this treats minimum wage *increases* as the relevant policy event delivered in a staggered fashion. Although I suppose you could also do some sort of continuous-treatment DID strategy to account for different levels or increase-sizes.

@@donoiskandar6820 that paper predates the new methods by a long ways so it is not taking the same approach. Cohort variation is different from over-time variation so I'd need to read the paper to see if it needs the new methods.

@NickHuntingtonKlein I see, I wonder whether Duflo's cohort approach is kind of substitutable with the staggered DID since in Duflo (2001; 2004) each school had a different time of school construction program implementation (depending on the region), so I guess in essence both allow estimating the impact of treatment at different times. But I have no clue pros and cons of both techniques. have been looking for comparison across the internet but haven't had any luck so far :')

@donoiskandar6820 in that case duflo would probably benefit from an update using the new methods (or a bacon decomposition showing that the twfe didn't bias things much)

The iv_robust R2 calculation uses a slightly different approach concerning how the IVs and their predictions are incorporated. In either case, R2 doesn't mean much in IV analyses and can be mostly ignored.

The adjustment for autocorrelation is different. You can use vcovHAC or NeweyWest from the sandwich package, or NW from fixest. See this section of my book theeffectbook.net/ch-StatisticalAdjustment.html#your-standard-errors-are-probably-wrong

@@mohammadrezamehrpour7724 if the group is a collider, then splitting the sample or something like that would introduce collider bias you'd need to fix.

this is a program-specific thing; it doesn't work in all Windows programs. Usually if it's in the text editor you're using, alt-click is the way to do it.

In R the Callaway and Sant'Anna method can be performed using the did package, or the Wooldridge method can be peformed using the etwfe package.

@@NickHuntingtonKlein Thank you so much for your reply and all those amazing videos. I tried both packages no success so far. Best

@@RightAIopen That's odd, those packages are for the exact situation you described. Why didn't they work? (Unless you mean that your treatments turn on *and then off* at different times for different groups).

​@@NickHuntingtonKlein They do, but even if I consider forever treated I have another problem. The DID report shows as an effect of the treatment the "group" effect. However this groups are not the cities with companies, they are the cities treated in the same year. Although I specify that the ID is the city group I can not extract the city effect. The grouping of cities treated together doesn't hold much relevance for my study - the focus is more on the individual cities or companies themselves.

​@RightAIopen oh I see, easy fix then. Just do a separate DID estimation for each treated city, dropping all the other cities already treated at the time. You'll be dropping a lot of data for each estimation so your standard errors will be bigger but this design works (and you're inherently doing something noisier by getting individual effects, so bigger SEs make sense). Or if you have enough pre-treatment data, synthetic control instead of DID would be better.

@@schafer4935 no, beta-1 would not be what you'd rest. In that model (assuming you've estimated it using only pre treatment data) you'd want to look at beta-3. This would test equality of (linear) prior trends.

@@remogurtner6907 I haven't used it much, but I believe sunab already adds the proper fixed effects for anything included. So your period fixed effects aren't necessary. The firm ones might be though. I'd check the example code in help(sunab)

@@NickHuntingtonKlein thank you for the quick reply!

You could use an empirical test to assess the validity assumption *under some theoretical assumptions that cannot be verified*. For example, you could use a test like that to check the validity of Z *if you are very certain that Z -> X -> Y is the only possible path violating validity*. However, note that this assumes no other possible pathway like Z->X2->Y, or that X is potentially a collider - that test also doesn't work if the path Z->X<-X3->Y is on the diagram. So you can use an empirical test to *support* some of the assumptions you're making (like if cor(X, Z) is near 0 then your assumption that the path Z->X->Y is not on the diagram is more plausible, supporting your validity assumption), but since validity is inherently a causal assumption and not a predictive one, you can't test it using data alone.

relevant section of the book theeffectbook.net/ch-InstrumentalVariables.html#checking-the-instrumental-variables-assumptions

There are two reasons to add controls in 2SLS: 1. To improve predictive precision and reduce standard errors 2. To support the validity assumption for the instrument. Often, we might think there is an open back door (or front door!) between the instrument and the outcome that does not go through the treatment of interest. In these cases we'd want to close those pathways down by controlling for variables.

@@NickHuntingtonKlein Interesting! For #2, it's kind of ironic that IV, although promoted as a clever way to get around selection on observables, needs the unconfoundedness assumption too, which is in the realm of selection on observables.

@@abduislam23 Yep! The idea is that the IV is *easier* to establish conditional unconfoundedness for than your treatment. i.e. maybe you can observe and measure all the confounders for your IV but not for your treatment. You still need conditional unconfoundedness.