causal-dag-identification

Causal DAG and Identification

Every causal study in task_type=causal_soo (selection-on-observables) begins with an explicit DAG. The DAG is the contract between the research question and the estimator — it names which variables play which roles (treatment, outcome, confounder, mediator, collider, instrument, pre-treatment correlate) and which back-door paths the adjustment set must close. Any method skill (M1–M5) composes this skill so the DAG check happens before any estimator runs.

Selection-on-observables identification

The selection-on-observables strategy assumes:

No-unmeasured-confounding (NUC): conditional on the adjustment set C, treatment assignment T is independent of the potential outcomes (Y(0), Y(1)). That is, every back-door path between T and Y is blocked by C, and C contains no descendants of T.

NUC is a checkable claim about the DAG, not a boilerplate assumption. Every causal study must state the claim, name the covariates that make it plausible, and acknowledge — by name — any confounder it cannot observe.

DAG drawing instructions

For every causal research question, construct a DAG with:

• Nodes = treatment T, outcome Y, and every covariate C in the candidate adjustment set.
• Edges = directed; an edge A → B asserts that A causally influences B.
• Serialization = the DAG MUST be saved in research_spec.dag as a DOT-string (round-trippable with NetworkX nx.DiGraph). DOT is preferred over pickled NetworkX so the DAG survives JSON round-trip with the rest of research_spec.json.

Adjustment-set selection — the back-door criterion (operational)

For an adjustment set C to identify the causal effect of T on Y:

1. Close every back-door path from T to Y. A back-door path is any path beginning with an edge into T (i.e., T ← ...). Conditioning on C blocks the path iff C contains a non-collider node on the path.
2. Include no descendants of T — conditioning on a post-treatment variable bars identification by absorbing part of the causal effect (post-treatment / mediator-as-confounder bias).
3. Include no colliders on the back-door paths unless the path requires conditioning on one of the collider's ancestors (the ancestor-of-collider rule).

If the DAG implies an un-block-able back-door path — i.e., the adjustment set requires a confounder that is not observed in the data — identification fails under selection-on-observables.

Mandatory pre-estimation checklist

Before any estimator runs, the Analyst must produce a covariate temporal-and-role table. For each candidate covariate C in research_spec.adjustment_set, declare:

Field	Allowed values
temporal_status	pre / post / contemporaneous (relative to T)
dag_role	confounder / mediator / collider / instrument / pre-treatment correlate

Save the table to research_spec.covariate_temporal_table and data_report.causal_identification.covariate_role_table. Any covariate with temporal_status: post is forbidden in the adjustment set per rule 2 of the back-door criterion above.

Identification-failure escalation

If dowhy.CausalModel.identify_effect returns None, OR if the DAG contains an un-block-able back-door path because a required confounder is unobserved:

• Analyst MUST stop. Do NOT silently proceed with "best available" adjustment.
• Critic MUST issue REVISE.
• The Writer must NOT report a causal estimate.

Write the failure into data_report.causal_identification:

{
  "adjustment_set": ["..."],
  "identified": false,
  "identification_method": "backdoor.linear_regression",
  "unmeasured_confounders_named": ["prior teacher quality", "..."]
}

validation_passed: false follows automatically.

Composition note

Every method skill (M1–M5) lists this skill in its references_skills so the DAG check + identification step always runs before any estimator. G2 (estimand definition) also composes this skill — the estimand is meaningful only against a DAG that identifies it.

Python implementation guidance

Primary library: dowhy (PyWhy). dowhy.CausalModel constructs the DAG from a DOT string (or NetworkX DiGraph); identify_effect derives the back-door / front-door / IV identification recipe; refute_estimate runs placebo / random-common-cause / unobserved-common-cause sensitivity checks.

Pin: dowhy>=0.11,<0.13. The dowhy API has historically had churn between minor versions — pinning the upper bound prevents silent breakage when an upstream rename lands.

Fallback: if dowhy becomes unmaintained, the CausalModel skeleton can be replaced with a thin custom wrapper around NetworkX

• a back-door criterion implementation (~80 LOC: enumerate paths between T and Y, classify each edge, check whether the proposed adjustment set blocks every back-door path).

Key functions / classes:

• dowhy.CausalModel(data, treatment, outcome, graph=DOT_string)
• model.identify_effect(method_name="default") → returns dowhy.causal_identifier.IdentifiedEstimand or None
• model.refute_estimate(estimate, method_name="random_common_cause" | "placebo_treatment_refuter" | ...)
• helpers in dowhy.causal_identifier

Function signatures the Analyst should produce:

def build_causal_model(
    df: pd.DataFrame,
    treatment: str,
    outcome: str,
    dag_dot: str,
) -> dowhy.CausalModel: ...

def identify_estimand(model: dowhy.CausalModel) -> dowhy.causal_identifier.IdentifiedEstimand: ...

def assert_no_post_treatment_in_adjustment(
    adjustment_set: list[str],
    treatment_var: str,
    var_temporal_table: dict[str, str],  # var -> 'pre' | 'post' | 'contemporaneous'
) -> None: ...

Library pitfalls:

• dowhy has had API churn; pin dowhy>=0.11,<0.13.
• The graph argument accepts DOT or NetworkX nx.DiGraph; prefer DOT for round-trip with research_spec.json.
• identify_effect will silently return None for unidentifiable estimands — the Analyst code must check and fail loudly.
• refute_estimate is best-effort, not proof of identification.

Validation criteria

The SKILL contract requires that:

1. The back-door criterion is operationally stated (the three rules above).
2. The pre/post/contemporaneous temporal-status mandate is enforced for every covariate.
3. The un-block-able-back-door escalation rule is honored (validation_passed: false).
4. The dowhy.CausalModel code skeleton is followed (no ad-hoc adjustment-set picking).
5. references_skills is empty (this skill is the root) and every method skill (M1–M5) composes this one.

A Writer using this skill must be able to produce a §Methods / Identification subsection that names the no-unmeasured-confounding assumption and lists every conditioning covariate with its DAG role.

An Analyst code artifact using this skill must produce:

• research_spec.dag as a DOT string,
• data_report.causal_identification with {adjustment_set, identified, identification_method, unmeasured_confounders_named},
• explicit validation_passed: false when identified == False.

Source provenance

Canonical source: the v3.0 causal-methods specification (internal) §3.1 (G1 per-skill specification, including back-door criterion, DoWhy implementation skeleton, dowhy version pin, and NetworkX fallback).