Four attribution approaches for corporate communications, compared on one decision

Corporate communications attribution methods answer different questions. For Cedar Vale Engineering Ltd's fictional site-consolidation message, the common question is: what evidence, if any, connects the approved communication version with a later, buyer-defined outcome among the eligible England population? The outcome, time window, population and unit remain blank until an authorised buyer supplies them.

HM Treasury's May 2026 Magenta Book distinguishes process, impact and value-for-money evaluation. Its methods are government guidance, not evidence that any approach will work for Cedar Vale.

One comparison frame

Each option is assessed against the same fields: causal question, eligible population, exposure definition, outcome event, comparison, time window, deduplication, confounders, data access, unit, uncertainty, privacy risk and decision use. No method receives a score because the real feasibility and buyer tolerance are unknown.

Approach	What it can establish	Essential boundary
Descriptive chronology	Which controlled events and later outcomes were recorded	Sequence alone does not attribute or prove cause
Rule-based allocation	Which channel receives credit under a declared rule	The rule creates the allocation
Matched observation	Whether comparable observed groups differ after adjustment	Unmeasured differences can remain
Experimental or quasi-experimental evaluation	An estimate against a defensible counterfactual	Design assumptions, power and interference need specialist review

Descriptive chronology is the baseline

Start with the approved copy ID, channel event, timestamp and independently defined outcome record. This method is inexpensive in conceptual terms and useful for audit. It can reveal missing hand-offs or corrections. It cannot say the message produced the outcome. A third event may explain both, or the outcome may have happened anyway.

Rule-based allocation answers an accounting question

A last-recorded-channel or first-recorded-channel rule can allocate an event consistently. Define the identity key, eligibility window, treatment of repeated contacts and unmatched cases before running it. Report both the raw outcome and the allocated label. A different rule may produce a different answer without changing a single observed event.

Matched observation needs credible comparability

An analyst may compare exposed and unexposed records that appear similar on observed characteristics. The common unit might be a person, establishment or account, but never a mixture. Selection, missing data and unobserved differences limit interpretation. Privacy and access gates apply before linkage, independently of analytical usefulness.

Causal evaluation asks the strongest question

HM Treasury's quality guidance for policy impact evaluation says monitoring a change does not establish that an intervention caused it and emphasises a robust counterfactual. Randomisation may be inappropriate or impossible in a corporate decision; quasi-experimental designs also depend on demanding assumptions. A qualified evaluator must decide whether interference between groups, changing implementation or insufficient observations defeats the design.

Withhold the winner

Chronology is appropriate for controlled delivery audit. Allocation is useful when finance or workflow needs a declared convention. Observational analysis can explore bounded associations. A causal design is required for a causal claim, but it is not automatically feasible or proportionate.

Before opening any outcome data, register the chosen question, method, window, exclusions and disconfirming evidence. Keep later exploratory analyses under different labels. This prevents a convenient result from being presented as the original test. The method owner should also record interference risks, such as recipients discussing the message across comparison groups, and any concurrent operational change that could affect the outcome.

Use one reporting unit throughout. If delivery is counted per message but the outcome exists per establishment, document the linking rule and unresolved many-to-one matches. Do not average incompatible estimates to create an apparently precise conclusion.

Cedar Vale should therefore record an approach as admissible, unresolved or not suitable for the actual question. If the outcome was defined after results appeared, the comparison population changed, or the counterfactual cannot be defended, stop at description. A longer report cannot repair a missing design.