, 23 tweets, 7 min read Read on Twitter
1/ It's #EconBookClub time again. Today, I will focus on chapter 3 of Guido Imbens' new paper, in which he compares the potential outcome framework with graphical approaches to #causation. arxiv.org/abs/1907.07271 #BookofWhy #DAG #Causality #AI #Econometrics
2/ After having introduced his readers to the fundamentals of DAGs, Guido now constrasts them with the potential outcomes framework, to which he contributed massively in his work within econometrics.
3/ Again, this chapter of the paper is mostly definitional, so I only have a couple minor comments. But I would like to pick up some of the threads that the paperleaves open in certain places, in my view.
4/ In the beginning of chapter 3, Guido develops in more detail an argument that he already briefly touched upon in the intro. Namely, that some of the success of PO is due to the fact that
5/ it connects so well to the early econometrics literature from the 1930s and 1940s. This argument is quite surprising to me, I must say.
6/ In my view, DAGs are actually much closer in spirit to this foundational work in econometrics than PO, because the early econometrics papers were built on the basis of (linear) structural equation models (SEM).
7/ Here's an example from Haavelmo (1943). This system of equations is non-recursive and can therefore not be represented by a DAG. But I would like to focus on another aspect here.
8/ This SEM is a full-blown structural model of the data generating process (DGP; with all the strong assumptions that comes with it). In that sense, the model defines potential outcomes/counterfactuals of the sort u(r). That's true.
9/ But within PO, this model of the DGP is kept entirely implicit. DAGs, by contrast, represent an underlying structural causal (SCM) model that correspons exactly to the type of SEM that Haavelmo lays out in his paper (and in which counterfactuals are equally defined).
10/ Moving from the SCM to its graphical representation – the DAG – entails a deliberate choice by the analyst to only focus on the qualitative causal relationships of a specific domain. Distributional and functional-form assumptions, such as linearity, are discarded
11/ and identification is only based on the fully non-parametric d-separation relationships of the recursive / acyclic model. By the way, if you think that PO has an advantage when it comes to non-recursive models, this is not the case. See Heckman and Vytlacil (2007):
12/ So DAGs make the (qualitative) assumptions that are imposed on the DGP fully transparent, whereas they're kept largely implicit within PO. That's why I think DAGs actually connect much better to the seminal work in econometrics by Tinbergen, Haavelmo and others.
13/ PO, by contrast, has its origins largely in biometrics, and this shows when it comes to the way in which it aims at emulatiung randomized control trials (RCT).
14/ Identification in PO is essentially based on an "as good as random" metaphor, which prescribes to find some vector of pre-treatment control variables that lead to unconfoundedness of the treatment (or instrument). If unconfoundedness holds, an RCT is essentially mimicked.
15/ The problem is of course, that PO offers no guidance whatsoever in finding this vector of pre-treatment characteristics W, which establishes unconfoundedness, because the DGP is not modeled explicitly.
16/ DAGs do this. They tell you exactly what you need to control for, after you've made your causal assumptions explicit in them. They also tell you what not control for and which pre-treatment variables might be redundant as controls, by the way.
17/ Once you move away from the "as good as random" or "experiments are the gold standard" metaphors, you realize that there's indeed nothing special about RCTs.
18/ Causal inference involves predicting the effect of interventions. And experiments (if they're feasible) allow researchers to carry out these interventions themselves. That's a big plus.
19/ But they also have the drawback that they're often run in quite artificial settings (lab experiments) and they can be expensive to carry out (so your N will be small). In such cases, it might actually be better to estimate causal effects by
20/ large-scale observational studies that are licensed by our domain knowledge captured in the DAG. It's also not the case – as some might believe – that RCTs would allow "model-free" causal inference because if you want to apply their results in different domains
21/ or for populations that differ just slightly (e.g., outside of the lab) – which is almost always the case – then you'll need again a model that specifies how the DGP differs across populations.
22/ So there's really nothing special about RCTs. They don't circumvent the need for untestable assumptions about the DGP (i.e., a model) and, in addition, they can be expensive and might be only feasible in very artificial settings.
23/ That doesn't mean, of course, that RCTs aren't great for internal validity. But their drawbacks can definitely tip the scale in favor of other methods sometimes. Again, it all depends on the specific context at hand. There is no hierarchy of methods in causal inference!
Missing some Tweet in this thread?
You can try to force a refresh.

Like this thread? Get email updates or save it to PDF!

Subscribe to Paul Hünermund
Profile picture

Get real-time email alerts when new unrolls are available from this author!

This content may be removed anytime!

Twitter may remove this content at anytime, convert it as a PDF, save and print for later use!

Try unrolling a thread yourself!

how to unroll video

1) Follow Thread Reader App on Twitter so you can easily mention us!

2) Go to a Twitter thread (series of Tweets by the same owner) and mention us with a keyword "unroll" @threadreaderapp unroll

You can practice here first or read more on our help page!

Follow Us on Twitter!

Did Thread Reader help you today?

Support us! We are indie developers!

This site is made by just three indie developers on a laptop doing marketing, support and development! Read more about the story.

Become a Premium Member ($3.00/month or $30.00/year) and get exclusive features!

Become Premium

Too expensive? Make a small donation by buying us coffee ($5) or help with server cost ($10)

Donate via Paypal Become our Patreon

Thank you for your support!