Causal Diagramming¶
Causal model (a directed diagram) — instantiates Confounder Control
Draws the assumed causal structure — exposure, outcome, confounders, mediators, colliders — as a diagram, so the decision of what to control is made from the assumptions before the data, not by the data after the fact.
Every other mechanism in this archetype does something to the data; Causal Diagramming decides what should be done in the first place. It draws the hypothesized causal structure — the focal exposure→outcome arrow plus every variable that plausibly feeds into them — as a diagram whose arrows encode who causes whom. The single idea that makes it this mechanism and not its siblings: it produces no estimate at all. Its entire output is a plan — which variables are common causes worth controlling, which are consequences that would poison the estimate if controlled — derived from stated assumptions rather than from correlations in the data. It is the map the design and analysis mechanisms then execute.
Example¶
An analyst wants to know whether coffee drinking raises heart-disease risk. Before running a single model, they draw the structure they believe generates the data. Smoking gets an arrow into coffee (smokers drink more coffee) and an arrow into heart disease; coffee gets the focal arrow into heart disease. Reading the diagram, a back-door path jumps out — coffee ← smoking → heart disease — a route by which the two are correlated without coffee causing anything. That single path tells them smoking must be controlled.
The diagram earns its keep just as much by what it forbids. "Diagnosed arrhythmia" sits downstream of coffee, on the causal pathway to heart disease; conditioning on it would block part of the very effect being measured. Without the diagram, a well-meaning analyst might have thrown arrhythmia into the model as "just another control" and quietly biased the answer. The output is not a number but an adjustment set — {smoking}, and explicitly not {arrhythmia} — that the analysis is now obligated to honour.
How it works¶
The distinguishing move is that arrows carry direction, so the diagram encodes causal and temporal order rather than mere association. From that structure you read off the paths mechanically: control the back-door paths (common-cause routes), and leave the causal pathway alone. A variable's role is not intrinsic — the same covariate is a confounder for one claim, a mediator for another — so the diagram is drawn per focal claim. It is deliberately qualitative and assumption-driven: it commits you, on the record, to what you believe causes what, and every arrow you omit is a strong claim of no effect.
Tuning parameters¶
- Resolution / granularity — how many nodes and paths you include. More detail can surface a hidden common cause, but an over-full diagram becomes unfalsifiable and hard to read.
- Formalism — an informal sketch versus a formal acyclic diagram supporting algorithmic identification. Formality buys a defensible adjustment set but demands acyclicity and completeness.
- Breadth of candidate search — how widely you cast for possible confounders before pruning. Wider reduces the chance of a missed common cause; narrower keeps the map legible.
- Assumed-absent edges — how many arrows you are willing to declare absent. Each omission strengthens identifiability and raises the stakes if it is wrong.
When it helps, and when it misleads¶
Its strength is that it forces control decisions into the open and makes them auditable, replacing the "throw every available covariate into the regression" reflex with an explicit rationale. It is also the only mechanism here that tells you what not to adjust for — the guard against conditioning on mediators and colliders that no covariate-hungry model provides on its own.
Its central weakness is that a diagram is only as good as its assumptions: a confounder you never thought to draw is a confounder you will never control, and the tidy graph can lend false rigor to a shaky causal story ("we drew a diagram, so we're causal"). The classic misuse is to draw the diagram after the analysis to rationalize the covariate set already chosen. The discipline that guards against this is to derive the adjustment set from the diagram before touching outcome data — the back-door criterion makes that derivation explicit[n1] — and to submit the diagram to domain-expert challenge for the arrows it is missing.
How it implements the components¶
focal_causal_claim— the diagram is built around one named exposure→outcome arrow; that claim is what fixes every other variable's role.causal_map— it is the map: the structured representation of exposure, outcome, confounders, mediators, and colliders.confounder_candidate— each common-cause node it draws is a candidate confounder, justified by an assumed mechanism rather than by correlation.temporal_order_check— arrow directions encode which variables precede the exposure, the sanity check that separates a pre-exposure confounder from a downstream consequence.collider_and_mediator_guard— reading the graph is precisely what distinguishes a confounder to control from a collider or mediator to leave alone.
It does not implement the acting-on-data components — adjustment_strategy and design_control belong to the mechanisms that execute the plan (Statistical Adjustment, Restriction or Eligibility Control, Random Assignment), and sensitivity_bound belongs to Sensitivity Analysis for Unmeasured Confounding.
Related¶
- Instantiates: Confounder Control — Causal Diagramming supplies the adjustment set that the rest of the archetype's machinery depends on.
- Sibling mechanisms: Statistical Adjustment · Restriction or Eligibility Control · Random Assignment · Control Group Design · Matched Comparison · Stratified Analysis · Instrumental Variable Strategy · Sensitivity Analysis for Unmeasured Confounding · Negative Control Check
Editorial Notes¶
Form Classification¶
Form family: Analysis, Modeling & Optimization
Rationale: Draws the assumed causal structure — exposure, outcome, confounders, mediators, colliders — as a diagram, so the decision of what to control is made from the assumptions before the data, not by the data after the fact, making its operative form a computation, comparison, model, or analytic representation used to infer, estimate, or choose.
Independent corroboration: The frozen evidence defines Causal Diagramming as 'Draws the assumed causal structure — exposure, outcome, confounders, mediators, colliders — as a diagram, so the decision of what to control is made from the assumptions before the data, not by the data after the fact', so its operative form is Analysis, Modeling & Optimization.
Review outcome: Independent reviewer agreement; high confidence.
Origin Attribution¶
Primary origin: Statistics & Experimental Design
Origin pattern: Single lineage
Present-day reach: Multi-domain
Rationale: Causal inference developed pre-analysis DAGs and the backdoor criterion to derive adjustment sets from explicit assumptions rather than post hoc correlations.
Related originating lineages:
- Philosophy — Philosophy of science supplies the insistence that causal assumptions be stated before data can be interpreted as identifying an effect.
Review resolution: Statistics and experimental design is the agreed primary lineage because pre-analysis causal diagrams determine confounder adjustment from assumed structure rather than data-mined association. Philosophy is formative to the semantics of causal direction, while the established method remains a single lineage used across domains.
Review outcome: Reconciled after independent review; high confidence.
Notes¶
Causal Diagramming is upstream of every other mechanism here: the adjustment set it produces is the input a design or analysis mechanism consumes. Keeping the map separate from the model that uses it is what lets a team revise its causal assumptions — add a missing arrow, demote a suspected confounder to a mediator — without re-running the whole analysis from scratch.
[n1] A back-door path is a non-causal route connecting exposure and outcome through a common cause; the back-door criterion identifies which variables must be conditioned on to block all such paths without opening new ones. It is the formal basis for reading an adjustment set off a diagram, which is why the derivation should precede, not follow, the analysis. ↩