Draw Me a Reason
Asystole

I drew an arrow.
Therefore I proved the cause.

Your pencil's asking for a lawyer.

Start with the question: whose outcome, which choice?
Compared with what, by when? Give the target a voice.
Potential outcomes: what would happen under each plan;
For the same decision, one observed path is all we can
See for each person. The other stays unknown;
Causal inference compares more than what's simply shown.
Directed acyclic graph: a D-A-G;
Arrows state causal assumptions for the world we think we see.
Directed means the arrows have a way to go;
Acyclic: no directed loop around the show.
Time-indexed nodes let later feedback take its place;
Keep earlier and later measurements in their proper space.
Draw relevant shared causes and the routes they can create;
An absent arrow also makes a claim to contemplate.
You made a graph so pretty that the crowd began to clap;
It still needs defensible assumptions underneath the map.

Draw me a reason, state what you assume.
What would happen if we changed the choice in this room?
A diagram makes the causal story clear;
It doesn't make the hidden problems disappear.
Draw me a reason, state what you assume.
What would happen if we changed the choice in this room?
A diagram makes the causal story clear;
It doesn't make the hidden problems disappear.

A backdoor path starts with an arrow into treatment;
Shared causes can make comparisons need refinement.
Block the noncausal routes with a suitable measured set;
Keep the causal question in the method that you vet.
For a total effect, don't casually block the chain
Through a mediator carrying the change you want to explain.
Conditioning on a collider can open up a route;
Your adjustment list needs reasons for who stays in or out.
Observational data can support a causal aim;
The assumptions and the design must carry that claim.
Matching, weighting, regression: useful tools if they fit;
Hidden common causes won't just vanish when you hit submit.
Define eligibility and time zero with care;
Align assignment and follow-up for a comparison that's fair.
Immortal time can creep in if future survival picks a side;
Emulate the target trial. Keep your starting clocks aligned.

Draw me a reason, state what you assume.
What would happen if we changed the choice in this room?
A diagram makes the causal story clear;
It doesn't make the hidden problems disappear.
Draw me a reason, state what you assume.
What would happen if we changed the choice in this room?
A diagram makes the causal story clear;
It doesn't make the hidden problems disappear.

Consistency: observed outcomes match the treatment received,
With well-defined interventions in the claim you've conceived.
Exchangeability, conditional on the set:
Potential outcomes independent of the treatment people get.
Positivity: each option must be possible within
The covariate groups your comparison is in.
Measurement, missing causes, sensitivity checks:
Tell me where the claim could bend before you flex.

Draw me a reason, state what you assume.
What would happen if we changed the choice in this room?
A diagram makes the causal story clear;
It doesn't make the hidden problems disappear.
Draw me a reason, state what you assume.
What would happen if we changed the choice in this room?
A diagram makes the causal story clear;
It doesn't make the hidden problems disappear.

My graph has no unmeasured causes.

Because you measured them?

Because I didn't draw them.
...Pass the pencil.