Skip to main content

Attribution and Its Causal Limits

Attribution assigns observed outcome credit according to a rule. Causal inference asks what outcome would have occurred under another intervention. Last touch, first touch, linear credit, and fitted attribution models can reorganize observed conversions; none creates the missing counterfactual by itself.

Prerequisites​

Complete path and journey analysis. For intervention claims, review prediction, description, and causation and randomized versus observational evidence.

Learning goals​

  • Distinguish deterministic credit allocation from incremental causal effect.
  • Declare conversion, eligible touch, lookback, identity, and missing-source handling.
  • Recognize selection, exposure, consent, and cross-device gaps.
  • Route product-impact questions to an experiment or explicit identification strategy.

Credit is not incrementality​

The fixture has no advertising impression. It does contain a declared source on signup_started, so the worked rule assigns each restated account creation to that actor's earlier signup source. This is transparent bookkeeping. It does not estimate what would happen if the homepage or docs source were removed, nor whether either source caused the signup.

For causal attribution, specify the intervention and estimand. A randomized eligible-unit holdout can identify an intention-to-treat effect when assignment, measurement, interference, and attrition assumptions are credible. Observational approaches require a defensible assignment mechanism, pre-treatment confounder control, overlap, consistency, and sensitivity analysis. Model fit is not that evidence.

Metric contract​

  • Population: actors with a valid, consented account_created by the restated close and a prior eligible signup source in the fixture.
  • Unit: resolved actor × first eligible conversion.
  • Observation window: signup touches and account creation occurring on 2026-10-01 UTC, with arrivals through 2026-10-04.
  • Denominator: two eligible conversions; missing-source conversions remain in the denominator and are reported separately.
  • Missingness: unobserved touches, denied measurement, other devices, offline exposure, and malformed events limit descriptive credit.
  • Identity: fixed 2026-10-02 mapping; no probabilistic cross-device stitching and no household inference.

Worked example​

import json
from collections import Counter
from datetime import datetime
from pathlib import Path

root = Path("tests/fixtures/product-analytics")
rows = [json.loads(line) for line in (root / "events-v1.jsonl").read_text().splitlines()]
plan = json.loads((root / "tracking-plan-v1.json").read_text())
parse = lambda value: datetime.fromisoformat(value.replace("Z", "+00:00"))
deduped = {}
for row in sorted(rows, key=lambda item: parse(item["received_at"])):
deduped.setdefault(row["event_id"], row)
def valid(row):
required = plan["events"][row["event_name"]]["required_properties"]
kinds = {"string": str, "boolean": bool}
return all(key in row["properties"] and isinstance(row["properties"][key], kinds[kind]) for key, kind in required.items())
def actor(row):
return row["user_id"] or plan["identity_snapshot"]["links"].get(row["anonymous_id"], row["anonymous_id"])
eligible = [row for row in deduped.values() if row["analytics_consent"] and row["purpose"] == "product_measurement" and valid(row)]
sources = {actor(row): row["properties"]["source"] for row in eligible if row["event_name"] == "signup_started"}
conversions = {actor(row) for row in eligible if row["event_name"] == "account_created"}
credits = Counter(sources[key] for key in conversions if key in sources)
print(f"rule=signup_source_credit credits={dict(sorted(credits.items()))}")
print(f"conversions={len(conversions)} missing_source={sum(key not in sources for key in conversions)} causal_effect=False experiment_required=True")

Expected output:

rule=signup_source_credit credits={'docs': 1, 'homepage': 1}
conversions=2 missing_source=0 causal_effect=False experiment_required=True

Limitations and common mistakes​

  • Calling credited conversions “caused conversions.”
  • Omitting unattributed conversions from the denominator.
  • Letting post-treatment behavior enter a model intended to estimate an intervention effect.
  • Comparing channels without common eligibility, exposure opportunity, or identity coverage.
  • Treating modeled recovery of unobserved conversions as individual truth or causal identification.
  • Ignoring interference when one actor's exposure changes another actor's outcome.

Practice​

Write two separate questions: “How does a declared last-touch rule distribute observed conversions?” and “What is the incremental effect of assigning a homepage promotion?” For the second, name the unit, assignment, outcome window, estimand, guardrails, and interference risk.

Next steps​

Ticket 63 continues with dashboards as decision surfaces, carrying these explicit metric and causal boundaries into monitoring and decision workflows.

Primary sources​

Checked 2026-10-19: Google Analytics attribution definition, Google Analytics modeled key-event limits, and Causal Inference: What If.