Zoom Logo

Causal Inference Workshop Day 1 (Rubin) - Shared screen with speaker view
Scott Cunningham
08:06
Hi everyone!
Scott Cunningham
08:14
Welcome to the Northwestern causal inference workshop -- the main event!
Scott Cunningham
08:28
I'm going to be the speaker's teaching assistant this week in chat
Scott Cunningham
09:06
I encourage everyone to ask me questions or ask each other questions. More than likely DR. Rubin won't be answering questions while the talk is happening
Scott Cunningham
09:34
I'm not sure exactly about Q&A -- this is the first year the conference has done Zoom, so bear with us as we figure out what's going on
Scott Cunningham
09:39
WE haven't started yet
Scott Cunningham
09:49
WE are not yet started
Scott Cunningham
10:03
WE are waiting on Dr. Rubin to get here, and Bernie will soon speak
Scott Cunningham
10:11
They are setting up the audio visual at the moment
Scott Cunningham
10:36
I wanted to tell you I'm going to be recording the sessions and then we will post this to you so you can access it.
Scott Cunningham
17:46
Hold on
Scott Cunningham
18:00
Thank you
Scott Cunningham
18:03
They're working on it now
Scott Cunningham
18:44
I’ll be here to fix whatever.
Scott Cunningham
22:38
I am trying to find this word document to share
Scott Cunningham
26:00
https://github.com/Mixtape-Sessions/Causal-Inference-1
Scott Cunningham
26:08
mixtape.scunning.com
Scott Cunningham
26:28
The first is a link to my workshops and there are "Labs” for you with solutions
Scott Cunningham
28:48
https://github.com/Mixtape-Sessions/Difference-in-Differences
Scott Cunningham
29:39
This is my diff-in-diff and synth workshop and it will provide some opportunities for labs. But I will organize this more for us as I find the speakers code
Scott Cunningham
33:14
For those on the zoom, my job is to basically be here for you guys. Ask me any questions, and I will try to *guess* (!) at what is going on.
Scott Cunningham
35:33
(Do everyone hae these slides?)
Scott Cunningham
37:05
Yes let me find the slides again for all of us. Just wrote Bernie
Scott Cunningham
38:09
Let me ask abt getting close up
Scott Cunningham
38:31
Hold on let me find the AV guy after I get these slides for yall
Scott Cunningham
41:22
Sorry still waiting on Bernie to check his email — he's across the room
Scott Cunningham
42:49
I asked for a closeup of the speaker
Scott Cunningham
43:48
He's going to move the camera closer during the break
Scott Cunningham
44:14
https://nuwildcat.sharepoint.com/sites/LAW-Shared-Folders/Shared%20Documents/Forms/AllItems.aspx?id=%2Fsites%2FLAW%2DShared%2DFolders%2FShared%20Documents%2FBBlack%5FShared%2FCausal%20Inference%2F2022%20Slides%20and%20Materials&p=true&ga=1
Scott Cunningham
44:28
There you go! Now we're cooking with fire!
Scott Cunningham
46:09
https://www.dropbox.com/sh/aaxwxii2oa2v2g2/AAANzdSwTAlkyO4OK8lAEWx_a?dl=0
Scott Cunningham
46:14
Here's a dropbox link I just made
Scott Cunningham
46:19
I'll start populating it with materials
Scott Cunningham
46:25
But Don's stuff is up now
Scott Cunningham
47:20
Okay just to be safe I put those files on the dropbox for us too
Scott Cunningham
58:14
How is the volume veryone?
Scott Cunningham
58:16
everyone?
Scott Cunningham
02:20:01
SUTVA is violated probably a lot in observational work
Scott Cunningham
02:20:11
Study estimates returns to college for instance
Scott Cunningham
02:20:22
Many different types of colleges, different majors
Scott Cunningham
02:21:04
The "No hidden variation in treatment" is as Bernie said a separate assumption within SUTVA than the no interference part
Scott Cunningham
02:22:35
This is a common style of reasoning among econoetricians and statistiicans -- start with the question.
Scott Cunningham
02:22:47
Separate the question or the estimand fro the procedure you'll use to estimate it
Scott Cunningham
02:35:46
I think he's going to say this
Scott Cunningham
02:35:53
It's not that there's constant treatment effects
Scott Cunningham
02:36:01
It's that you and me got the same 100mg of aspirin
Scott Cunningham
02:36:28
this is what i meant by differeing quality of college above
Scott Cunningham
02:40:11
I think what he means here is each perso gets a discrete price
Scott Cunningham
02:40:16
but we have 10000 people
Scott Cunningham
02:40:22
and they each get different discrete prices
Scott Cunningham
02:40:33
so it's continuous in the sense that there’s a lot of prices
Scott Cunningham
02:40:44
but he's saying the treatmetn assignment is a particular amount per unit
Scott Cunningham
02:44:03
Sorry - he went faster than I followed
Scott Cunningham
02:44:43
Continuous prices happen in datasets all the time obviously and we use them to try an destimate demand
Scott Cunningham
02:44:52
But no one is being "assigned a continuous price"
Scott Cunningham
02:44:58
They are being assignd a specific price
Scott Cunningham
02:45:03
Another person a different price
Scott Cunningham
02:47:30
i asked abotu violations of sutva on platforms
Scott Cunningham
02:47:49
doing ab tests on websites with social media and interactions
Scott Cunningham
02:47:59
he said that he thinks there's a lot of naivete
Scott Cunningham
02:48:13
i could'nt understand the speaker
Scott Cunningham
02:48:22
i'll tell don at break to repeat the questions
Scott Cunningham
02:49:00
just emailed bernie
Scott Cunningham
02:50:04
Can you hear Bernie?
Scott Cunningham
02:50:32
Okay I bet it's bc he has a mic
Scott Cunningham
02:51:00
Let me talk to the AV people abt this.
Scott Cunningham
02:52:42
hold on
Scott Cunningham
02:52:46
school example
Scott Cunningham
02:52:52
let me get don to repeat the question
Scott Cunningham
02:53:50
i'm going to have to pull him aside. apologize.
Scott Cunningham
02:54:02
i'll get this fixed though
Scott Cunningham
02:57:55
Someone is asking abt spillovers
Scott Cunningham
02:58:06
What is a spillover
Scott Cunningham
02:58:10
Is tehre a way to never get it?
Scott Cunningham
02:58:17
Interference
Scott Cunningham
02:58:43
In economics we call it an externalitiy
Scott Cunningham
03:02:03
Spillover, interference -- he just means that the SUTVA assumption requires that a potential outcome is stable in the sense that when another peron takes an action, my potential outcome doesn't change.
Scott Cunningham
03:02:39
that was funny rubin saying imbens doesn't know what an externality is. I suspect our recent nobel winner in economics knows what an externality is though… 🙂
Scott Cunningham
03:10:27
Same unit i I think
Scott Cunningham
03:10:41
I think i is the unit index notation
Scott Cunningham
03:11:49
This is subtle what he's saying
Scott Cunningham
03:11:56
This is what he is calling no manipulation
Scott Cunningham
03:12:40
Perceptions of sex vs sex. Interesting
Scott Cunningham
03:13:39
Someone asked in DM can we compare the effect of treatment for the same units and then takes the average
Scott Cunningham
03:13:47
What happens when it's not possible for cancer treatment
Scott Cunningham
03:14:15
I think if I am following here that he is still describing the parameters. You can't compare the same units to one another in this part of the talk
Scott Cunningham
03:14:32
He is rather saying that the defined treatment is W and you either are treated W=1 or not W=0
Scott Cunningham
03:14:40
But the causal effect is Y1i - Y0i
Scott Cunningham
03:15:40
I'm not 100% sure what normatively he is suggesting. I think he is saying though that the randomization solves this problem (I think we are still in that part of th eleture)
Scott Cunningham
03:16:33
I think maybe this is going slower than it may seem. I
Scott Cunningham
03:16:49
I use the "perfect doctor" in my book and lectures all the time.
Scott Cunningham
03:17:00
Basically he's going to describe sorting
Scott Cunningham
03:17:08
People are going to "sort” into the better treatment
Scott Cunningham
03:17:25
"better treatment for each patient"
Scott Cunningham
03:18:58
Okay so he's saying that with sorting, we draw wrong conclusions bc we are using realized data outcomes (Y) but we need the potential outcomes (Y0,Y1)
Scott Cunningham
03:19:19
You're going to see here where selection bias comes from with raw correlations using this kind of example
Scott Cunningham
03:20:10
This is the source of bias in the sciences involving humans who optimize
Scott Cunningham
03:28:41
"In your book, you said both were wrong bc there was no control"
Scott Cunningham
03:28:45
And Rubin said "exactly right"
Scott Cunningham
03:31:24
Okay we are going to stop here.
Scott Cunningham
03:31:35
I'm going to bring up to the speakers abt repeating the audience questions.
Scott Cunningham
03:32:00
Can you hear bernie?
Scott Cunningham
03:32:05
Hes' speaking in a mic
Scott Cunningham
03:32:11
but i'm wondering if mics reach zoom
Scott Cunningham
03:32:28
observed outcome onto a treatment indicator
Scott Cunningham
03:33:23
“regression is dangerous!”
Scott Cunningham
04:42:15
I emailed bernie for this document fyi
Scott Cunningham
04:46:01
WE will discuss common support I'm almost positive tomorrow with Pedro Santanna
Scott Cunningham
04:49:58
There weren't any lunch talks
Scott Cunningham
04:50:05
Just people eating lunch and mingling
Scott Cunningham
04:50:17
Yes I'll get the word document also and post it
Scott Cunningham
04:50:20
I emailed bernie
Scott Cunningham
04:51:44
FYI, at 2:00 (20 min) I have to run to the Apple Store
Scott Cunningham
04:51:47
I'll hurry back
Scott Cunningham
04:52:04
(I woke up and my phone was on the “black screen of death”)
Scott Cunningham
04:53:52
Someone asked “May I ask where behavorial economics studies fall in these discussions?”
Scott Cunningham
04:54:22
As a group, that community has largely adopted the experimental design, but tend to be I think working inside lab settings
Scott Cunningham
04:54:36
That's my sense — I tend to see a lot of lab experimentalist being trained behavioralists
Scott Cunningham
04:55:12
but otherwise all of tehse principles are relevant to them as it is for anyone
Scott Cunningham
04:56:36
Someone writes "what' the cons of having unbalanced dataset? For example, in propensity score matching, I matched 1 treated with 3 controls”
Scott Cunningham
04:57:03
If selection is conditional on covariates, then the imbalance is a source of selection bias
Scott Cunningham
04:57:19
It does not necessarily guarantee you have seletion bias
Scott Cunningham
04:57:39
The covariates imbalanced must be those covariates that are needed to achiee conditiional independence.
Scott Cunningham
04:57:57
I tend to think in terms of causal graphs
Scott Cunningham
04:58:12
So for me this is more straightforward as I map the imbalance of *necessary* covariates to confounding
Scott Cunningham
04:58:59
https://mixtape.scunning.com/03-directed_acyclical_graphs
Scott Cunningham
04:59:07
I discuss this here
Scott Cunningham
04:59:08
https://mixtape.scunning.com/05-matching_and_subclassification
Scott Cunningham
04:59:53
he's addressing it here
Scott Cunningham
05:01:21
the imbalance in age in other words
Scott Cunningham
05:01:27
this is getting into the selection mechanism
Scott Cunningham
05:07:06
how do you deal with old people having more time to smoke
Scott Cunningham
05:07:11
migrating from cigerettes to cigars
Scott Cunningham
05:07:20
that was the question
Scott Cunningham
05:07:29
he said basically cochran ignored it
Scott Cunningham
05:09:53
Bernie said he uploaded the document
Scott Cunningham
05:09:55
But I did too
Scott Cunningham
05:09:56
https://www.dropbox.com/sh/aaxwxii2oa2v2g2/AAANzdSwTAlkyO4OK8lAEWx_a?dl=0
Scott Cunningham
05:14:51
We have a new link on the workshop materials
Scott Cunningham
05:14:57
Let me find it again
Scott Cunningham
05:15:52
https://nuwildcat.sharepoint.com/sites/LAW-Shared-Folders/Shared%20Documents/Forms/AllItems.aspx?id=%2Fsites%2FLAW%2DShared%2DFolders%2FShared%20Documents%2FBBlack%5FShared%2FCausal%20Inference%2F2022%20Slides%20and%20Materials&p=true&ga=1
Scott Cunningham
05:16:23
Oay I have to run. I'll hurrry back!
Scott Cunningham
07:29:54
apologies everyone