Measurement
Part of PR agencies measurement: reject AVE, keep evidence levels apart
PR attribution methods compared, and why allocation is not incrementality
Compare PR attribution methods for an England buyer, separating allocation, matched observation, modelling and causal estimates without naming a winner.
PR agencies attribution methods answer different questions. A rule can allocate recorded credit. A causal design can estimate what changed because of an intervention. Treating those as interchangeable creates certainty the evidence does not contain.
This comparison uses one bounded case: an organisation operating in England wants to understand a defined stakeholder action after several communications routes. No route, value, audience result or agency effect is supplied.
What to take away
- Allocation suits transparent bookkeeping, but it does not estimate what would have happened without the communication.
- Design the comparison before the campaign starts, because an experiment cannot be added afterwards.
- Linked records describe who was exposed. Association can remain unless the design varied exposure.
- Model output depends on its inputs, and sparse or shifting series can make a coefficient unusable.
- Give each published number a label and the method version that produced it.
Common comparison fields
Each method must state the decision, eligible people or events, intervention, comparison, unit, period, timezone, channel coverage, identity and deduplication rule, missingness, assumptions, output, uncertainty, owner and stop condition. The same outcome definition and analysis period apply across the table.
PR attribution methods compared
Rule allocation
- Question answered
- Which touchpoint gets credit
- Evidence produced
- Assigned count or value
- Main hold condition
- Rule or window unknown
Platform attribution
- Question answered
- Platform allocates observed events
- Evidence produced
- Platform-modelled credit
- Main hold condition
- Consent or model unclear
Matched observation
- Question answered
- Which outcomes link to exposure
- Evidence produced
- Linked association
- Main hold condition
- Matching quality hidden
Experiment
- Question answered
- Difference vs counterfactual
- Evidence produced
- Causal estimate
- Main hold condition
- Comparison or power fails
Question answered
- Rule allocation
- Which recorded touchpoint receives credit under a declared rule?
- Platform attribution
- How does one platform allocate events it can observe?
- Matched observation
- Which recorded outcomes can be linked to exposure or contact records?
- Marketing mix modelling
- How do aggregated inputs covary with an outcome over time?
- Experiment or quasi-experiment
- What difference is estimated against a credible counterfactual?
Evidence produced
- Rule allocation
- Assigned count or value
- Platform attribution
- Platform-modelled channel credit
- Matched observation
- Linked descriptive association
- Marketing mix modelling
- Modelled contribution under assumptions
- Experiment or quasi-experiment
- Causal estimate with uncertainty
Main hold condition
- Rule allocation
- Rule, window or eligible paths are unknown
- Platform attribution
- Consent, cross-channel loss or model version is unclear
- Matched observation
- Matching quality, unmatched share or timing is hidden
- Marketing mix modelling
- Too few periods, unstable inputs or confounding defeat the model
- Experiment or quasi-experiment
- Comparison assumptions, power or implementation fail
There is no universal winner. The right method follows the question and available design.
Allocation is not incrementality
A last-touch or weighted rule divides observed credit. It is useful for consistent reporting when the rule, window and unassigned share are visible.
Google Analytics provides a current first-party example of platform allocation in its attribution documentation. That page describes Google's own reporting methods and settings. It is not independent evidence of completeness, cross-platform performance or PR causation.
In the UK, the Information Commissioner's Office enforces the consent rules behind analytics and marketing cookies. A platform figure therefore reflects who agreed to be measured, plus the coverage and window that supplier applies.
The Association for the Measurement and Evaluation of Communication (AMEC) publishes the Integrated Evaluation Framework. UK agency teams use it to name objectives, inputs, outputs and outcomes rather than one headline score.
Matched observation links records using a declared key or probability model. Report eligible records, matched and unmatched shares, false-match risk, excluded people and the permitted purpose. Association may remain after matching because exposure was not randomly assigned.
Models require visible assumptions
Marketing mix modelling works with aggregated series. Specify time unit, geography, outcome, communications inputs, other drivers, transformations, lag structure, validation and uncertainty. A coefficient is model dependent. Sparse or changing data can make a result unsuitable for the decision, especially when PR activity overlaps advertising, product changes or external news.
Model assumptions to specify
- Time unit and geography
- Outcome and communications inputs
- Other drivers and transformations
- Lag structure
- Validation and uncertainty
HM Treasury's Magenta Book is government guidance, but its distinction between process, impact and value-for-money evaluation is useful here. Its quality guidance for impact evaluation says that monitoring a change does not show the intervention caused it and stresses a credible counterfactual.
Experiments intentionally vary exposure under an approved design. Quasi-experiments use a defensible comparison created by timing, allocation or observed characteristics. Both need qualified statistical, ethical, privacy and operational review. Stop causal language when groups differ materially, spillover is uncontrolled, implementation is weak or the outcome definition changes.
Publish a bounded verdict
If your question is which recorded touchpoint gets credit, use rule allocation. If it is how one platform splits the events it can see, use platform attribution. If it is whether contact records link to outcomes, use matched observation.
If it is how aggregate inputs covary over time, use marketing mix modelling. If it is what changed against a credible counterfactual, use an experiment or quasi-experiment.
Small budgets and thin records point to rule allocation or platform attribution. Matching needs linkable identifiers. Modelling needs many stable periods. Experiments need headroom for a clean comparison. More complex does not mean better.
The report must label every result attributed, modelled or causal, retain the method version and show uncertainty. AVE and media-space prices answer none of these questions and should not enter the comparison. Publication remains on HOLD until named reviewers inspect the actual data, permissions, design and reconciliation.
Before you act
- State the decision, eligible people, intervention and comparison.
- Declare the rule, window and unassigned share for allocation.
- Report matched and unmatched shares and false-match risk.
- Specify model inputs, lag structure, validation and uncertainty.
- Stop causal language when groups differ materially.
- Keep publication on hold until named reviewers inspect the data.
Common questions
What is the difference between allocation and incrementality?
A rule that divides recorded credit keeps reporting consistent when the rule, window and unassigned share are visible. Incrementality estimates effect against a counterfactual, so only it supports causal language.
When should marketing mix modelling be used?
MMM works with aggregated series and suits a sufficiently rich aggregate question. Specify time unit, geography, outcome, communications inputs, other drivers, transformations, lag structure, validation and uncertainty. Sparse or changing data can make a result unsuitable for the decision.
