Zoom Logo

Main Workshop Day 3 (SantAnna DiD) - Shared screen with speaker view
Qinzheng Xu
21:05
Hi, we cannot hear anything
Hyunwoo Jung
21:42
Yes
Yizhi Zhu
21:45
Yes
Jiawei Li
21:45
Now yes
XIANGYU WANG
21:52
Now Yes
Xijin Huang
21:53
Now Yes
Boyu Shen
21:58
Yes
Praveen Bhagawan
22:01
Now Yes
Scott Cunningham
29:07
Good morning everyone
Scott Cunningham
29:13
I apologize for being late
Scott Cunningham
29:27
Here’s what we will do today
Scott Cunningham
29:35
1) I will answer questions I can answer
Scott Cunningham
29:55
2) I will ask on your behalf to Pedro also (not instead but also)
Scott Cunningham
30:51
3) we asked Pedro to try and remember to use the cursor. Problem is he has a particular presentation style and is very dynamic and stands out front. So he kind of gets lost in the energy of it but I will try to help as much as I can
DANDAN GU
32:10
Thank you, sounds good
Amit Mehra
32:34
Thanks - much appreciated!
Yeosong Yoon
32:43
Thank you!
Adam Wu
32:56
Thank you!
Hongyu Fu
33:23
Thanks, sounds good
Xijin Huang
33:24
Thanks!
Naoka Carey
33:41
For those of us not in economics, do you have any suggested resources for cross-walking some of the the terms to the way they are described in other fields (for me, that's psychology).
Scott Cunningham
34:08
I’m not sure these slides are available yet
Scott Cunningham
34:14
I’ll find fnem
Scott Cunningham
34:41
Naoka my book may be helpful.
Scott Cunningham
34:50
Mixtape.scunning.com
Naoka Carey
34:53
Thanks!!
Scott Cunningham
35:09
Mainly bc it’s kind of a less technical style of writing (but sometimes still technical)
Scott Cunningham
35:51
The cross walk to other disciplines — that’s definitely an important jigsaw puzzle piece that I haven’t yet seen fully developed. A lot of my popular writing tries to do it but I haven’t with psychology
Yeosong Yoon
36:41
Scott's book is really helpful to understand causal
Scott Cunningham
40:22
I seriously think Pedro is one of the smartest people I’ve ever met. We talk every day on slack bc of a revise and resubmit we are never apparently going to resubmit. I think he has a photographic memory or something. He remembers everything!
Scott Cunningham
40:36
I think today will be a treat
Scott Cunningham
41:28
Pedro is as you guys may know a leader in diff in diff. He and I think real similar bc I basically have studied his papers so closely but I’m kind of more common sense guy so I’ll give my answers and try to help
Naoka Carey
42:54
He has a wonderful way of assuming we are all as smart as him and can follow (alas, that may not be the case for me!)
Scott Cunningham
43:23
Just moved Pedro’s new materials to the Dropbox
Scott Cunningham
43:55
https://www.dropbox.com/sh/aaxwxii2oa2v2g2/AAANzdSwTAlkyO4OK8lAEWx_a?dl=0
Scott Cunningham
44:12
It’s Pedro’s code problem sets slides and simulation results
Scott Cunningham
46:37
Can you guys hear okay?
Laura Nicolae
46:48
Yes
Hyunwoo Jung
46:49
yes 🙂
Hoa Vo
46:50
Yes
Boyu Shen
46:50
Yes
Khalid Ibrahim
46:51
Yes
Xijin Huang
46:54
That's much better!
Boyu Shen
46:59
It better than yesterday
Jiawei Li
47:09
I would recommend giving a read to the survey paper by Imbens and Wooldridge (2009):Recent Developments in the Econometrics of Program Evaluation. This paper contains a nice summary of the things Pedro covered yesterday and today, and many more.
Alice Malmberg
47:16
FYI, the Dropbox folder for today looks empty
Scott Cunningham
47:41
Alice — I bet the stuff I just uploaded is this material
Scott Cunningham
47:45
Let me move it over
Naoka Carey
47:55
Today
Naoka Carey
48:03
Slides are in the "Slides" folder in Day 2
Scott Cunningham
48:14
Ok I just moved over new material
Alice Malmberg
48:20
Thank you!
Scott Cunningham
48:26
Check day 3 now under slides
Jiawei Li
49:38
That paper I mentioned above is also the core reference paper when I was taking a class with Rucio Titiunik, our speaker tomorrow
Scott Cunningham
49:51
That’s a great great paper I agree
Scott Cunningham
50:03
I’ll find it and put it in our shared folder
Scott Cunningham
51:45
I just created a sub folder called scott readings and put the paper in there that Jiawei mentioned
Farah Mejri
55:35
Thank you
Scott Cunningham
56:18
I’mHearing audio has improved. He’s wearing a mic so that must be the difference
Scott Cunningham
56:24
I’ll pass along that this helped
Scott Cunningham
01:25:50
Feel feee to ask me any questions you think I can answer
Scott Cunningham
01:26:06
Any little things like notation or certain words
XIANGYU WANG
01:26:18
Could u please let us know what is the question?
Scott Cunningham
01:26:42
Sorry — I forgot to pay attention. It was sometjifn abt notation. Let me remind Pedro to repeat question
XIANGYU WANG
01:26:52
Thank you!!
Xiaoying Yu
01:27:03
why one uses d and the other uses g as group indication in equation?
Scott Cunningham
01:27:15
Hold on
Scott Cunningham
01:28:07
Let me find it
Scott Cunningham
01:29:00
Hm, that g is new to me. I may have heard him say it was a typo but let me study it
Scott Cunningham
01:29:39
I think it’s supposed to be d and d is either treatment or control group
Scott Cunningham
01:29:51
It appears to be the only place it appears
Alix Walkup
01:29:57
I believe he said it was a typo in this slide. In later slides he switches to g to represent "group"
DANDAN GU
01:31:40
May I ask what “PT”?
Dongning Shao
01:32:03
parallel trend
DANDAN GU
01:32:13
I see
DANDAN GU
01:32:24
thanks!
Syed Waleed Mehmood Wasti
01:32:37
Is it important to show the point at which the treatment takes place, like event studies?
Scott Cunningham
01:35:12
ParallelTrend
Scott Cunningham
01:35:49
Yes — you need to know when treatment takes place bc diff in diff uses a pre treatment observation as calculation and it must be untreated .
Scott Cunningham
01:36:05
Matching methods and controls
Scott Cunningham
01:36:15
He’s asking abt that
Syed Waleed Mehmood Wasti
01:36:52
Okay thank you. I was particularly asking about graphs. Like in this one, we do not know exactly when the treatment started.
Naoka Carey
01:37:51
Happy to defer this question if I should wait - but wondering if it's possible to use geographic clustering or random effects or levels (e.g. if level of treatment is clustered geographically) in a DID - or do you really have to use random effects models for that?
Scott Cunningham
01:40:00
Triple difference relies on a different assumption
Scott Cunningham
01:40:16
It’s a parallel gap of the within group
Jin Miao
01:41:34
If treatment occurs at different times, do we need to assume that the earlier adoptions at one place do not affect later adoptions elsewhere?
Praveen Bhagawan
01:42:00
Should we use cross-sectional PSM or Panel PSM? I have seen papers where researchers use both methods. Which is better?
Scott Cunningham
01:42:21
Old versus young within state would need to evolve similar in counterfactual for treated states as they do in control for triple diff
XIANGYU WANG
01:43:55
Maybe I missed this. What is the connection with matching and DID?
Jin Miao
01:44:25
Matching is used to construct a better control group
XIANGYU WANG
01:44:35
Thank u!
Scott Cunningham
01:45:11
If I miss a question please repost
Ramaa Chitale
01:45:15
Can we look at subgroups within a DiD or will the PT assumption fall apart? Do I have to test parallel trends for each subgroup I choose?
Scott Cunningham
01:45:43
Looking at subgroups in what sense? As a triple diff?
Ramaa Chitale
01:46:34
I guess I want to know treatment effect for example urban vs. rural
Ramaa Chitale
01:46:56
but from this graph it looks like DiD is average treatment effect for everyone
Ramaa Chitale
01:47:02
am I understanding this right?
Scott Cunningham
01:47:05
It is here
Scott Cunningham
01:47:12
He’s averaging Y
Scott Cunningham
01:47:18
Over the population
Praveen Bhagawan
01:47:29
Should we use cross-sectional PSM or Panel PSM? I have seen papers where researchers use both methods. Which is better?
Ramaa Chitale
01:47:40
okay - is that a challenge that it glosses over some population nuances?
Scott Cunningham
01:47:55
The parameter is defined as an ATT
Scott Cunningham
01:48:16
So it’s the average over everyone. If it’s a repeated cross section, some additional assumptions have to be introduced
Scott Cunningham
01:48:22
Stationarity specifically
Ramaa Chitale
01:50:19
okay thank you
Dongning Shao
01:50:30
will these assumptions about repeated cross-section be covered in this workshop?
Scott Cunningham
01:51:01
Probably
Scott Cunningham
01:51:12
When he covers doubly robust did I think he will
Dongning Shao
01:51:40
ok thank you
XIANGYU WANG
01:55:43
what does “recovering” mean on slide Page #13?
Scott Cunningham
01:56:31
It means estimating
Scott Cunningham
02:11:32
We can define post treatment for the controlBecause we are only working with a single treatment date
Scott Cunningham
02:11:46
When we have staggered rollout with multiple treatment dates things change and that’s undefined
Scott Cunningham
02:12:04
But if the treatment happens in 2004, we can define post treatment always for anyone — it’s after 2004
Scott Cunningham
02:12:35
For fixing time varying covariates to be invariant, Pedro may be anticipating what he’s going to say momentarily
Scott Cunningham
02:13:03
Basically, he’s going to incorporate covariates not using regression but using a doubly robust model that estimates propensity scores at baseline
Scott Cunningham
02:13:18
As well as an outcome regression (for the doubly part)
Naoka Carey
02:21:19
Is the ATU just the same thing as the parallel trend for the untreated?
XIANGYU WANG
02:24:06
Not sure why we jumped to this SUTVA directly. What's the connection then?
Scott Cunningham
02:26:44
I’m stuck in a phone call
Scott Cunningham
02:27:19
I’ll be right back but I think SUTVA is an assumption bc you need a pre treatment untreated observation
Scott Cunningham
02:27:40
I’ll ask all these questions asap so apologies
Alix Walkup
02:27:52
Is there a way to test the no-anticipation assumption?
XIANGYU WANG
02:28:42
Thank you! I like your explanations.
Scott Cunningham
02:35:04
No anticipation
Scott Cunningham
02:35:07
Let’s be concrete
Farah Mejri
02:35:11
the difference between slide 30 and 31 is that we replaced Yi,t=1 (1) with Yi,t=1 (0) because before the event happened we expected the same potential outcome for both treated and untreated right?
Scott Cunningham
02:35:19
Say the state says they’re going to raise taxes
Hyunjoo Oh
02:35:25
@Scott About no-anticipation assumption (in ppt p.27), does this means that corporate decisions (like pre- and post-M&A) may not work as treatment if it's a strategic choice of firms (i.e., treatment is not announced in advanced, but firms consider its effect and select themselves into it)?
Scott Cunningham
02:35:31
But that announcement happens today and taxes are next year
Scott Cunningham
02:35:49
So the question is whether the treatment is the announcement or the actual raising taxes
Scott Cunningham
02:36:12
You need no anticipation to mean “they don’t change behavior when they learn abt the annoucnemen
Scott Cunningham
02:36:18
Learning isn’t a problem iow
Scott Cunningham
02:36:30
It’s if the learning changes behavior.
Scott Cunningham
02:36:47
You could test it sort by using announcement dates as the treatment instead of effective dates
Alix Walkup
02:38:31
Thank you
XIANGYU WANG
02:38:50
Why does the column Di change from slide #30 to #31?
Yitong Hu
02:42:40
Is there any cases that we can do the DID without parallel trends assumption? I saw some papers using only data from two units (which means there is no parallel trends evaluation) I'm not sure if they are wrong
Scott Cunningham
02:46:21
Pedro pretty much groups diff in diff AS parallel trends based assumptions
Scott Cunningham
02:46:36
If you lose PT, then you go in other directions than diff in diff
Yitong Hu
02:47:18
could you explain "go in other directions"
Scott Cunningham
02:47:31
Like use synthetic control, matrix completion, panel data
Scott Cunningham
02:47:46
You may have to bring some structure too
Scott Cunningham
02:48:12
Diff in diff in my mind is simple bc it uses the change in Y0 for control group to impute the missing Y0 for treated
Scott Cunningham
02:48:17
But that require parallel
Yitong Hu
02:48:25
so without PT, we can not name anything DID
Scott Cunningham
02:48:25
If you don’t, then it’ll be biased
Xiaoying Yu
02:49:38
Are the PT assumption usually met in real life problems?
Scott Cunningham
02:49:43
It depends a little on who you ask
Scott Cunningham
02:49:56
I personally find it convenient to group diff in diff with PT and Pedro too
Scott Cunningham
02:50:27
PT assumption is a particular kind of behavioral situation and Pedro has a new paper describing a range of selection that is consistent with it
Yitong Hu
02:55:08
could you share the same of the Pedro's new paper
Scott Cunningham
02:55:13
Yes
Scott Cunningham
02:55:16
I’ll dig it up
Scott Cunningham
02:55:23
There’s two new ones and I’ll post both
Yitong Hu
02:56:03
Thank you!
Scott Cunningham
02:56:39
https://arxiv.org/pdf/2203.09001.pdf
Scott Cunningham
02:57:06
https://arxiv.org/pdf/2207.06564.pdf
Hyunjoo Oh
03:01:15
What does "Y(2)" of E[ Y(2) | D=1 ] means on the 4th equation on p.40? I missed this part.
Yitong Hu
03:11:24
will he introduce the difference between comparative interrupted time series analysis and DID? Or maybe could you talk about your thoughts on comparative interrupted time series analysis? Thanks!
Sohyeon Hwang
04:44:00
Is anyone else not getting audio?
Monika Avila Marquez
04:44:12
I hear all
Hyunwoo Jung
04:44:14
No problem with audio
Sohyeon Hwang
04:44:31
Got it thanks, I'll try joining again
Farah Mejri
04:44:35
Me too I hear all
Ran Barzilay
04:45:43
Why is minus 3 the omitted year and not minus 1?
Jiawei Li
04:45:57
Did Scott have a lunch talk just now? I think I may have missed it
Hyunwoo Jung
04:46:16
No he didn’t
Ran Barzilay
04:46:36
Wouldn't that change the interpretation? meaning - prices were lower in the three years prior
Hyunwoo Jung
04:46:42
He said he will have a lunch talk on Friday
Jiawei Li
04:46:42
Thanks
Xiaoying Yu
05:09:18
Shall we cluster the subject level with repeated measures on time which has pre and post measures from the same subject?
Xiaoyong Zheng
05:12:52
How about two-way clustering for both unit and time?
Hyunjoo Oh
05:23:11
Where can we find "PS1.pdf"?
Zeping Pan
05:24:52
it is in “Pedro additional materia”, “Problem Sets” folder
Hyunjoo Oh
05:27:08
@Zeping Thank you!
Laura Nicolae
05:27:39
Not related to the current topic, but how is it that the DD papers that study the effects of minimum wage changes on employment don’t violate SUTVA since people can move between states?
Cindy Chen
05:49:09
is conditional ATT estimator based on CPT assumption going to be the same as triple DiD estimator if there is only one covariate (ie the third difference)?
Monika Avila Marquez
06:36:07
Probably I missed it, we need a common support assumption?
Monika Avila Marquez
06:41:28
Yes, I found it.
Jin Miao
07:28:09
How should we deal with the case that the same treatment groups get treated multiple times repeatedly?
DANDAN GU
07:30:01
In Problem Set2, after repeating the procedures with Michigan Sample, should we compare the difference between Kentuchky and Michigan?
DANDAN GU
07:32:20
Thanks!
Jiawei Li
07:41:42
All these notations here are standard in the staggered adoption literature, so Pedro is not doing something for nothing