, 17 tweets, 7 min read Read on Twitter
2/ In this chapter, Guido responds to the main criticism that is brought forward towards the potential outcomes framework by proponents of DAGs – the choice of covariates or how to justify ignorability.
3/ We're all familiar with the ignorability / unconfoundedness assumption that underlies classical matching estimators, for example. The corresponding DAG is depcited in Figure 8a of the paper.
4/ The problem is that ignorability in PO is a total black-box. What's this magical vector of controls W, that leads to an unconfounded treatment? There's nothing in the PO framework that could shed light on this question.
5/ This ambivalence is bad for science, because it leads to endless discussions in seminars (and in journal submissions) about what controls are / aren't necessary, based on the gut feelings of participants, and which are susceptible to arguments by authority.
6/ To alleviate the problem a little bit, you will usually see rough guidelines about which controls to include along the following lines. This advice is not sufficient, however, for the following four reasons.
7/ First, it is not necessary to control for all pre-treatment variables in your data set. This graph (taken from my teaching notes: p-hunermund.com/teaching/) has five pre-treatment variables, but controlling for only two of them at a time would be sufficient.
8/ Second, controlling for certain pre-treatment variables can actually introduce rather than remove bias. Guido discusses the collider bias example from the #BookofWhy, where adjusting for seat belt use, which is pre-treatment, opens up a backdoor path from smoking to cancer.
9/ Third, adjusting for post-treatment variables doesn't have to be problematic at all, as long as they don't introduce collider bias. Psychologists and management scholar do that all the time in mediation analyses, for example.
10/ And fourth, the very notion of pre-treatment and post-treatment isn't even defined without having a DAG in mind that tells you the causal ordering of variables. So funnily enough, PO actually can't even tell you what a pre-treatment variable is.
11/ To find suitable controls that guarantee ignorability, you need a model of the data generating process, either in form of a DAG or equations (although the latter don't capture the asymmetric nature of causal relationships).
12/ In my view, Guido provides a great argument for why carefully modeling the DGP is so important himself when he discusses violations of the collider / m-bias structure in the smoking and lung cancer example.
13/ While the causal effect in Figure 13a was identified without controlling for anything, it becomes unidentified as soon as we introduce an arrow from "attitudes towards safety" to "smoking" in the graph.
14/ That's pretty subtle! A small change to the DAG, only one additional edge, and the picture changes completely. The thing that Guido seems to ignore though is that such problems do not just arise with *these* methods (i.e., DAGs) but in all applied empirical work.
15/ Should we trust that we're able to solve such cases purely based on introspection and gut feeling? I think not. Instead, we should make our causal assumptions explicit, such that others can understand them and build on them in future research.
16/ Holding this explicitness as an argument against DAGs is pretty funny to me, to be honest. In this last paragraph of the chapter, Guido is basically saying that (1) specific model assumptions in a DAG are hard to defend, therefore (2) one should better restrict attention
17/ only to a specific DAG, which corresponds to the ignorability case discussed above. So DAGs are hard to come up with for a given context at hand, and that's why we should assume that a specific template DAG will be correct? I don't get it. 🤷‍♂️
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!