Causal Attribution: Four Questions
The author serves as CEO of C3 Metrics, a marketing attribution company. The four questions in this piece are applied to C3 first.
The marketing measurement industry has increasingly used a term: causal attribution. Attribution is the assignment of credit for an outcome to inputs. Causal is a claim about counterfactual reality. Causal attribution, as a phrase, joins them.
The industry has a long history of borrowing vocabulary from adjacent disciplines: scientific, statistical, financial. Some borrowings are merited. The field has extended real methodological work from causal inference, econometrics, and financial attribution into marketing measurement, in places. Other borrowings are unearned: terms adopted for their credibility without the underlying warrant. The four questions below are one way to tell them apart.
C3 Metrics does not use the term causal attribution. What follows is the four questions, C3’s answers to them, and a note on what the answers show.
The four questions.
The first three are attribution questions. The fourth is the causal question, and it is only meaningful if the first three have been answered.
At what unit of analysis is credit distributed?
What does the word attribution mean in the methodology’s use of it?
With what confidence interval, or with what disclosed assumptions in lieu of one?
If a causal claim is being made: under what identification strategy is that claim established, at what unit of analysis?
None of these is a gotcha. Any competent practitioner asks them of themselves before adopting a technical term. A methodology that answers all four consistently is doing what the marketing says it does. A methodology that answers three and stops at the fourth has answered the causal question in the same breath, by declining to make the claim.
The standard question three asks is disclosure. Assumptions are owed to the party relying on the number, in a form that party or its auditor can review. That is the bar. It is not that the framework must be public. It is not that the framework must be free. It is that the reliant party can review what the number rests on. Every provider is free to specify the framework. The standard is the disclosure to the reliant party.
C3 Metrics, held to the test.
C3’s four answers.
Unit of analysis: the individual conversion. Credit is distributed across touchpoints.
What attribution means, in C3’s use: the assignment of credit under a stated convention. Not a counterfactual claim.
Confidence interval and disclosed assumptions: the question offers two forks, and C3 takes the second, and says so plainly. No confidence interval is offered at the individual conversion, because no sampling frame exists at that unit. Disclosed assumptions stand in lieu. Those disclosures are documented in a proprietary framework, available to clients. The framework specifies the data sources used, the inferences drawn from them, and the assumptions behind those inferences. A C3 client, or an auditor authorized by the client, can review it. That is the form in which C3 meets the standard. Other providers can and do have other frameworks; the standard is not the framework but the disclosure to the reliant party in reviewable form.
If a causal claim were being made, at what unit? None. C3 does not claim causal identification at the individual conversion, and no counterfactual is available at that unit that would let us. We therefore do not use the word causal.
C3 answers all four. The answer to the fourth is what tells you why C3 does not use the term causal attribution.
The self-executing test.
The fourth question can only be checked against the first. If the unit in the fourth answer is coarser than the unit in the first — identification established at the geo, credit reported at the ad; identification at the channel, credit at the tactic; identification at the segment, credit at the conversion — the methodology owes a fifth disclosure: the rule by which the estimate is transported downward, and the name of that rule, which is allocation. The four questions do not force this comparison. But they enable it. That is what makes them a test.
The fifth disclosure is not exempt at this address, either. C3 distributes credit across touchpoints under a stated convention. That convention is an allocation rule, and we name it one.
This is a choice.
If C3 could use the term consistent with our own methodology, we would. It would be a stronger claim, a more impressive one, and the market might value it, at this moment. Choosing to use the term is a valid positioning choice. So is choosing not to. We have chosen not to, because we cannot use the term without knowing what it means, and without being able to test ourselves against that meaning. Until we can, we do not.
That is the same choice, in a different domain, as the distinction between measuring and confirming: to hold our language to what the underlying method can support, and to not adopt a modifier we cannot warrant.
The technical constraint.
Identification lives where the counterfactual was constructed. A randomized experiment across geos identifies at the geo. A media mix model calibrated by that experiment identifies at the channel. A conditional average treatment effect estimated for a stratum identifies at the stratum. Propagating the estimate downward, from geo to channel to ad, from stratum to individual, from population average to conversion, is an allocation rule, not an identification.
A weight can travel. A causal estimate cannot travel with it. Identification does not follow the weight down. Nothing transports it.
The frontier of causal inference over the last fifteen years has extended these tools to heterogeneous effects: conditional average treatment effects, causal forests, uplift models, and Pearl’s third rung of counterfactual reasoning. Each stratifies the population and yields an expected effect for a class. None, absent heroic assumptions the practitioner is obliged to disclose, delivers a causal effect for a specific input on a specific individual. The finest available estimate stops at the stratum. The stratum is not the person, and the person is where attribution asks its question.
Consider the smallest possible case. Two ads sit side by side, for exactly the same product, from the same company. Four out of five users choose the one on top. We can see the split. We cannot know why the fifth chose the one on the right, and no stratum, however finely cut, closes that gap. The field named this constraint before the industry existed: Holland called it the fundamental problem of causal inference, in 1986. The four questions are the test; this is the reason the test matters.
Shapley-value attribution, one common approach to principled attribution in marketing, decomposes a modeled outcome additively across features. This is a mathematically coherent operation. It is not, in its standard form, a causal claim about the underlying reality. Causal effects do not decompose additively in the presence of interaction, and interaction is the ordinary state of marketing exposures.
Causal Shapley variants exist. They respect an assumed causal graph and require the identifying assumptions to be stated explicitly. That specification does not work around the fourth question; it returns to it in different clothes: what is the identification strategy, and does the graph, treated as an assumption, survive review by someone who did not build it? Same test, same disclosure requirement.
The generous reading, granted in full.
A media mix model calibrated by properly designed geo experiments makes an explicit, defensible identification argument at its unit of analysis. Advisory committees composed of serious empiricists in causal inference review models, propose designs, and hold their principal investigators to a standard. This is real work, done at real institutions, and nothing in this piece diminishes it.
Advisory committees review models. They do not review headlines. Where the phrase causal attribution parts company with the discipline the four questions demand is at the transport: from the unit at which the model was identified, to the unit at which its outputs are marketed to end users. The failure is not in the mathematics. It is in the marketing.
The invitation.
The four questions are not proprietary. They are the questions any practitioner should ask themselves before adopting a technical term. C3 has answered them, publicly, above. Any vendor claiming to sell causal attribution can do the same. Point to the disclosed unit of analysis. Point to what attribution means in the methodology’s use. Point to the confidence interval or disclosed assumptions, in a form the reliant party can review. And, if a causal claim is being made, point to the identification strategy that establishes it at the reporting unit. If the reporting unit is finer than the unit of identification, disclose the transport rule and name it allocation.
If those answers are consistent, the phrase is doing what the marketing says. If they are not, the phrase is doing work the methodology does not authorize.
The difference is simple. Is this confirmation, or is it measurement? Only a full, consistent set of answers to the questions above can tell us. Absent that consistency, the phrase causal attribution is neither.
C3 does not use the term. The four questions are available to anyone who does.
Companion piece: Confirmation Is Not Measurement.