Cause is read by a declared design
Answers Epistemology
A reading whose value rests on a claim of cause names the evidence design it is read by: a holdout experiment, a geo experiment, a matched-market test, a marketing mix model, or attribution. The attribution rules are one design among these and type into it, and the kernel names the record road credit is divided over, an event attributed to a tactic, so attribution has a declared object.
What goes wrong without it
Two readings of the same outcome, one from a randomised holdout and one from a last-touch rule, are not two estimates of one number; they are answers to different questions, and a dashboard that shows them side by side without saying which is which invites the wrong decision. The measurement literature is explicit that observed credit and assigned effect diverge, that the returns to advertising are small against the noise, and that the mix model is a third design with limits of its own. Naming the design on the reading, and typing the attribution rules into one design, is what lets a plan say what kind of evidence it is asking for.
What proves it — 2
Each is a gate the build runs: a check inside a ring, a whole ring, or an eval beside them. A build stops at the first that refuses.
- registry integritykernel integrity
Every kind and relationship the kernel names is one the registries declare: a spine's verbs are verbs its kind states or is reached by, the metric derivation names specifications that exist, the contract-status escape names kinds that exist, and the grounding reaches kinds that exist, and the machine's kinds and verbs and the instrumenting kind are kinds and verbs the registries declare. - registry integrityspine integrity
Every row of a kind the model leans on states the relationships that kind exists to carry, so a relationship declared across the registry and asserted nowhere is a failure rather than a thin line on a map.
Derived from — 6
- Gordon, Zettelmeyer, Bhargava and Chapsky, Marketing Science 38(2), 2019A Comparison of Approaches to Advertising Measurement: Evidence from Big Field Experiments at Facebook
- Dalessandro, Perlich, Stitelman and Provost, ADKDD 2012Causally motivated attribution for online advertising
- Kohavi, Tang and Xu, Cambridge University Press, 2020Trustworthy Online Controlled Experiments: A Practical Guide to A/B Testing
- Vaver and Koehler, Google, 2011Measuring Ad Effectiveness Using Geo Experiments
- Lewis and Rao, Quarterly Journal of Economics, 2015The Unfavorable Economics of Measuring the Returns to Advertising
- Chan and Perry, Google, 2017Challenges and Opportunities in Media Mix Modeling
Kinds it licenses
Relationships it licenses
read by · reads as design
Where this model stands
Evidence Designs is a vocabulary kind of five in Measurement and evidence; Metrics read by Evidence Designs, 0..1, names the design a reading is established by, asserted today on the three attributed readings; every attribution model reads as the attribution design, on a typing road bound 1..1.
registry/kernel.json `evidence` names the kind, the road, the record road credit divides over and the rules kind; FND-026 holds each to the registries, and the question by what design a reading is established walks the road.
A metric may also count the occupants of a consumer state, a fourth way the Metrics spine lets it derive; Retention Rate counts occupants of Account Holder, which is the reading it always was.
What it does not claim
An incremental reading is a reading read by an assigning design, and the model does not yet carry the reading itself as a record; the record shapes declare its join.
RUL-050