Measurement
Four attribution approaches for corporate communications, compared on one decision
Compare four communication attribution approaches on one England decision, using consistent questions, evidence fields and explicit causal limits.
What to take away
- Meltwater Media Intelligence is admissible for a delivery audit: it orders controlled events and later records, and cannot attribute an outcome.
- Google Analytics 4 last-click is admissible for an accounting convention only; the declared rule creates the allocation.
- Propensity score matching with MatchIt stays unresolved when key differences between records are unmeasured.
- Geo-based incrementality testing with CausalImpact or Meta Conversion Lift is the only option here that can support a causal claim, and only if the counterfactual, power and interference checks hold.
- Register the question, outcome, window, unit and exclusions before opening any outcome data.
One frame for four named approaches
Judge each candidate on the same fields: causal question, eligible population, exposure definition, outcome event, comparison, time window, deduplication, confounders, data access, unit, uncertainty, privacy and decision use.
Starting from the population rather than a channel list is the discipline set out in population-based audience research. HM Treasury's Magenta Book separates process, impact and value-for-money evaluation.
One frame for four approaches
What it can establish
- Meltwater Media Intelligence timeline
- Which controlled events and later outcomes were recorded
- Google Analytics 4 last-click
- Which channel receives credit under a declared rule
- Propensity score matching with MatchIt
- Whether comparable observed groups differ after adjustment
- Geo-based incrementality testing with CausalImpact
- An estimate against a defended counterfactual
Causal claim?
- Meltwater Media Intelligence timeline
- No
- Google Analytics 4 last-click
- No
- Propensity score matching with MatchIt
- Only with a defensible design
- Geo-based incrementality testing with CausalImpact
- Yes, if the design holds
Cost and requirement
- Meltwater Media Intelligence timeline
- Quote-based annual subscription, no public list price; needs delivery logs
- Google Analytics 4 last-click
- Free standard GA4; Analytics 360 is paid enterprise, quoted annually; needs events and a property
- Propensity score matching with MatchIt
- MatchIt is a free open-source R package on CRAN; needs linked exposure and outcome records
- Geo-based incrementality testing with CausalImpact
- CausalImpact is a free open-source R package from Google; Meta Conversion Lift is an advertiser option; needs comparable geographies
Linking exposure records to outcome records is a data protection question as much as an analytical one. The Data Protection Act 2018 governs that processing in the UK.
Example: registering the Cedar Vale decision first
For Cedar Vale Engineering Ltd's hypothetical site-consolidation message, a registered record might read as follows.
The question is whether the approved message changes the rate of one defined action among eligible accounts in named regions. The population is accounts matched on sector and size; exposure is confirmed delivery of the approved copy ID; the outcome is the buyer-defined action recorded in the CRM.
The unit is the account, the window is twelve weeks before and six weeks after the send, and the exclusions cover accounts contacted by another campaign. Name the disconfirming evidence that would count against the message, which is the reason the framework is fixed before results arrive.
That record yields the decision rule:
Registering the Cedar Vale decision
- Delivery logs complete, no counterfactualMeltwater timeline admissible, claim stops at description.
- Finance needs one conventionGA4 last-click admissible as accounting.
- Comparable records exist, key differences unmeasuredMatchIt unresolved.
- Matched regions and pre-period data defensiblegeo test admissible for a causal claim.
Which approach fits which decision type
Which approach fits which decision
- Delivery auditMeltwater Media Intelligence timeline, admissible.
- Allocated budget conventionGoogle Analytics 4 last-click, admissible as accounting.
- Pattern exploration among similar recordspropensity score matching with MatchIt, unresolved.
- Repeat, scale or stop a messagegeo-based incrementality testing, admissible for a causal claim.
Label each approach as admissible, unresolved or not suitable before acting on any number. If the outcome was defined after results appeared, the population changed, or the counterfactual cannot be defended, record it as not suitable and stop at description.
Is the approach admissible?
Was the outcome defined after results appeared?
Stop at description
Check comparison population
When a geo test is worth its cost
Proportionality turns on decision stakes, sample size and budget. A geo test earns its cost when the decision is expensive to reverse and the audience is large enough to detect a meaningful effect.
Where the decision is cheap to reverse, run something smaller first, as the note on turning evidence gaps into small, reversible tests argues. uk/government/publications/the-magenta-book/quality-in-policy-impact-evaluation-qpie-html) states that monitoring a change does not establish that the intervention caused it and points to a strong counterfactual. Interference between regions and concurrent operational change are the usual ways a promising design fails.
Common questions
Which of the four can support a causal claim?
Only geo-based incrementality testing, and only when matched control geographies, pre-period data and a qualified review of interference and power all hold. The other three order, allocate or compare rather than identify a cause.
Why is Google Analytics 4 last-click not causal?
It gives credit under a declared rule. Another rule can produce a different answer without any observed event changing, which makes it an accounting convention.
What makes propensity score matching unresolved?
Unmeasured differences, selection and missing data between matched records. The comparison is only as credible as the characteristics the analyst could observe.
When should I stop at description?
When the outcome was defined after results appeared, the comparison population changed, or no counterfactual can be defended. A longer report cannot repair a missing design.
