DAGWOOD abstract submitted to SER! But I also wanted to share it with y'all, because it's the single most surprisingle brain-warpingly difficult thing I think I've ever worked on, and I am super proud of what we've got so far.
A THREAD!
#epitwitter
1/n
A THREAD!
#epitwitter
1/n
DAGs are a way to organize and display causal inference models and their assumptions.
But where in the DAG are those assumptions?
They are hidden in the negative space between nodes, taking the idea of being transparent about our assumptions very literally.
2/n
But where in the DAG are those assumptions?
They are hidden in the negative space between nodes, taking the idea of being transparent about our assumptions very literally.
2/n
For example, a causal model typically assumes a sharp (or negligible) causal null for any and all missing confounders. Those confounders are hidden somewhere in the ocean, but we can't see them in a typical DAG.
Hard to think critically about invisible things.
3/n
Hard to think critically about invisible things.
3/n
Then there is the problem that DAGs in practice are WILD. What exactly goes in a DAG? How does it relate to what you actually did?
There are some schools of thought for "best practices" in theory, but the reality is pretty crazy.
4/n
There are some schools of thought for "best practices" in theory, but the reality is pretty crazy.
4/n
And there are other problems with causal inference methods, in part due to language differences.
Like did you know that the no unmeasured confounders assumption in epi is nearly identical in structure to the instrumental variables exclusion restriction?
5/n
Like did you know that the no unmeasured confounders assumption in epi is nearly identical in structure to the instrumental variables exclusion restriction?
5/n
Enter: the DAGWOOD (DAG with omitted objects displayed)!
DAGWOODs are a combo graphical overlay and corresponding list of the most critical assumptions in any causal inference model, and a framework for putting it all together in a practical package.
6/n
DAGWOODs are a combo graphical overlay and corresponding list of the most critical assumptions in any causal inference model, and a framework for putting it all together in a practical package.
6/n
Everything is based on the model implied by what you actually ran or propose to run, or the "empirical DAG"
DAGWOODs add two layers on top of that: exclusion restriction pathways (ERPs) and misdirection or reverse pathways (MRPs).
7/n
DAGWOODs add two layers on top of that: exclusion restriction pathways (ERPs) and misdirection or reverse pathways (MRPs).
7/n
DAGWOODs add two layers on top of that: exclusion restriction pathways (ERPs) and misdirection or reverse pathways (MRPs).
Ya know all that stuff in the limitations sections that list confounders that might mess things up? Those are known unknown ERPs.
8/n
Ya know all that stuff in the limitations sections that list confounders that might mess things up? Those are known unknown ERPs.
8/n
But the DAGWOOD magic is in the unknown unknowns.
What if there was some kind of confounding or direct DAG-compatible pathway that the authors didn't think of, control for, write about, etc.?
That's an unknown unknown ERP! And we can generate them algorithmicly!
9/n
What if there was some kind of confounding or direct DAG-compatible pathway that the authors didn't think of, control for, write about, etc.?
That's an unknown unknown ERP! And we can generate them algorithmicly!
9/n
An unknown unknown ERP represents what happens if we assume that there is some kind of non-negligible causal pathway here. If that would mess up your estimate, it's an important assumption.
In many/most models, it's the MOST important (and least likely to hold) assumption.
10/n
In many/most models, it's the MOST important (and least likely to hold) assumption.
10/n
But we aren't done yet! What about all those assumptions about how you drew your DAG arrows?
What if your confounder is a collider?
What if you have reverse causality? Or hidden time nodes that represent multiple periods?
Enter: the MRPs!
11/n
What if your confounder is a collider?
What if you have reverse causality? Or hidden time nodes that represent multiple periods?
Enter: the MRPs!
11/n
These are a little weirder, since they are inherently not DAG pathways (they'd break the A part if they were).
Instead, they represent alternative ways your DAG could have been drawn. Again, totally generated by algorithm.
12/n
Instead, they represent alternative ways your DAG could have been drawn. Again, totally generated by algorithm.
12/n
And the extra magic? Just like DAG nodes and arrows correspond to causal math, DAGWOODs correspond to model assumptions.
And we can list them!
Every causal model already has these assumptions built in; DAGWOODs pull them out of the negative space.
13/n
And we can list them!
Every causal model already has these assumptions built in; DAGWOODs pull them out of the negative space.
13/n
Each item on this list must be individually justified as being negligible or non-existent in order to be considered reasonably unbiased.
So now the author has a structure on which they can (and must) justify those model assumptions.
14/n
So now the author has a structure on which they can (and must) justify those model assumptions.
14/n
More over, you can use a DAGWOOD to help BUILD a model! Make a DAG, run the DAGWOODs, get assumptions. If an unknown unknown ER isn't justifiable, maybe you need to control for something there, so add it into the DAG.
Then rerun the DAGWOOD algorithm and see what you get.
15/n
Then rerun the DAGWOOD algorithm and see what you get.
15/n
Look, new assumptions! So maybe you have a new area to concentrate on. Build another node. Run another DAGWOOD. Keep going until your assumptions are justifiable (or in some cases, realize they aren't, which is also great!).
16/n
16/n
You can do the same thing for critically evaluating an existing DAG too! You draw the DAG implied by what was run, get back list of assumptions to look at.
You can use DAGWOODs as the central component for some kind of causal inference review tool (hinty winky face).
17/n
You can use DAGWOODs as the central component for some kind of causal inference review tool (hinty winky face).
17/n
DAGWOODs don't change the bar for causal inference, but they tell us where the bar is.
They take a causal map, look at the oceans, and warn you "Here there be dragons."
All models rely on critical assumptions, DAGWOODs help us confront them directly.
18/n
They take a causal map, look at the oceans, and warn you "Here there be dragons."
All models rely on critical assumptions, DAGWOODs help us confront them directly.
18/n
And of course, it couldn't have been done without an mind-boggling number of rewrites, scraps, new ideas, and general brain melding from all-start co-authors and colleagues, @anecdatally @SarahWieten and @BreskinEpi, with a shout out to @Emma_C_Clarke
More to come!
19/19
More to come!
19/19
