Method library WORKFLOW / 041

Research workflow/4 steps

Draw the DAG, find the adjustment set (ggdag & dagitty)

@ggdag ✓ P D · Jun 29, 2026 · 1k views

Source ggdag / dagitty — Barrett; Textor et al.

@misc{ggdagdagitty,
  title        = {ggdag / dagitty},
  author       = {Barrett; Textor and others},
  howpublished = {\url{https://r-causal.github.io/ggdag/}},
  note         = {Software / documentation}
}

In the margins · Peer notes@wei_zhang · Testable implied conditional independencies are underused — it's the closest thing to a falsification test for you… 2

The method workspace

Follow the analysis.

Select a node to explore the method

Method mapInput → 4 steps → output

START WITHInput data

ARRIVE ATResults

Read step 01

Inside the methodReasoning, equations & implementation

Data prep

Encode your assumptions as a DAG

Step notesField note & discussion1

Data preparation — shapes the raw inputs into what the estimator expects.

What happens here

Nodes and directed edges make the causal story explicit and falsifiable — the modelling happens here, not in the regression.

Formula

G=(V,E);\quad \tau=\mathbb E[Y\mid do(X{=}1)]-\mathbb E[Y\mid do(X{=}0)]

Reads fromNot specified Feeds intoNot specified

Key code

dag <- dagify(y ~ x + a + b, x ~ a + b,
              exposure='x', outcome='y')

Reference / docs ↗

Discussion on this step (0)

StatsOtterDiscussion starter · Editorial note

Before using this step, make its assumptions and its output explicit. A useful peer note explains what you would check, what could go wrong, or what would change your interpretation.

Questions to explore

Which assumption matters most here, and what evidence could challenge it?
What would you check before passing this output to the next step?

Peer discussion

Bring your experience, a question, or a different perspective.

Diagnostic / pre-tests

Enumerate paths; spot backdoors & colliders

Step notesField note & discussion1

A pre-flight check — run this before trusting any estimate downstream.

What happens here

Open non-causal paths bias the effect; a collider is the trap — conditioning on it OPENS a path rather than closing it.

Formula

\text{backdoor: } X\leftarrow\cdots\to Y;\quad \text{collider: } X\to C\leftarrow Y\ (\text{do not condition})

Reads fromNot specified Feeds intoNot specified

Key code

ggdag_paths(dag)

Reference / docs ↗

Discussion on this step (0)

StatsOtterDiscussion starter · Editorial note

Before using this step, make its assumptions and its output explicit. A useful peer note explains what you would check, what could go wrong, or what would change your interpretation.

Questions to explore

Which assumption matters most here, and what evidence could challenge it?
What would you check before passing this output to the next step?

Peer discussion

Bring your experience, a question, or a different perspective.

Estimation

Minimal sufficient adjustment set

Step notesField note & discussion1

The core estimate — where the causal quantity itself is computed.

What happens here

The smallest covariate set that blocks every backdoor path (and conditions on no descendant of treatment) identifies the effect.

Formula

Z\ \text{blocks all backdoors}\Rightarrow \mathbb E[Y\mid do(x)]=\textstyle\sum_z \mathbb E[Y\mid x,z]\,P(z)

Reads fromNot specified Feeds intoNot specified

Key code

ggdag_adjustment_set(dag)

Reference / docs ↗

Discussion on this step (0)

StatsOtterDiscussion starter · Editorial note

Before using this step, make its assumptions and its output explicit. A useful peer note explains what you would check, what could go wrong, or what would change your interpretation.

Questions to explore

Which assumption matters most here, and what evidence could challenge it?
What would you check before passing this output to the next step?

Peer discussion

Bring your experience, a question, or a different perspective.

Robustness check

Test the DAG's implications

Step notesField note & discussion1

A robustness check — does the headline result survive a different lens?

What happens here

A DAG implies conditional independencies you can actually check in the data — and warns when several DAGs are observationally equivalent.

Formula

d\text{-sep}(A,B\mid C)\ \Rightarrow\ A\perp\!\!\!\perp B\mid C\ \text{(testable)}

Reads fromNot specified Feeds intoNot specified

Key code

library(dagitty); impliedConditionalIndependencies(dag)

Reference / docs ↗

Discussion on this step (0)

StatsOtterDiscussion starter · Editorial note

Before using this step, make its assumptions and its output explicit. A useful peer note explains what you would check, what could go wrong, or what would change your interpretation.

Questions to explore

Which assumption matters most here, and what evidence could challenge it?
What would you check before passing this output to the next step?

Peer discussion

Bring your experience, a question, or a different perspective.

Data contract

Inputs & data

Subject-matter assumptions about what causes what: nodes (treatment, outcome, confounders, mediators, colliders) and the directed edges between them.

Show data format & exampleHide example

Format — declare the edges; mark the exposure and outcome.

library(ggdag)
dag <- dagify(y ~ x + a + b,
              x ~ a + b,
              exposure = 'x', outcome = 'y')
ggdag_adjustment_set(dag)   # minimal sets that block the backdoors

Read the evidence

Results & interpretation

2 figures Workflow discussion 2

01FigureThe minimal adjustment set {a, b}: conditioning on those nodes blocks the backdoor paths (greyed) and leaves the direct x → y effect identified.

02FigureEvery open path from treatment x to outcome y: the direct causal path (red) plus the backdoor paths (teal) that confound it until they're blocked.

Figures reproduced from ggdag / dagitty — Barrett; Textor et al. — unofficial community showcase; all credit to the original authors.

Context & interpretation

⚠️ Unofficial community showcase of ggdag. Not affiliated with the authors; all credit to them.

Before any estimation: encode your assumptions as a causal graph, enumerate the backdoor paths from treatment to outcome, and let the graph hand you the minimal set of covariates to adjust for.

Peer notes

Discuss the workflow (2)

2

@rahul_nair · Jun 29, 2026

Doing this BEFORE you touch the data is the discipline that prevents collider bias. The 'do not condition on a collider' warning can't be repeated enough.
0

@wei_zhang · Jun 29, 2026

Testable implied conditional independencies are underused — it's the closest thing to a falsification test for your DAG.

1

@diego_morales · Jun 29, 2026

This pairs perfectly with a DoWhy/propensity pipeline: ggdag picks the adjustment set, then you estimate and refute.