Measurement
Part of Measuring a franchise so that activity and return never get confused
Franchise attribution methods: choosing the right design for a change
Compare four attribution methods for franchise changes using the same causal question, implementation evidence, counterfactual test and reporting limits.
Attribution asks whether a specific action caused a measured change. It is different from reporting that sales rose after training or advertising. This comparison considers four methods against one question: did the intervention change a pre-defined outcome for comparable franchise units over a stated period?
The criteria are design strength, data need, practical burden and the claim the result can support. No method guarantees a commercially useful effect, and calculations should receive qualified analytical and financial review.
Franchise attribution methods
Design
- Before-and-after
- compare one unit with its own earlier period
- Matched comparison
- compare treated units with similar untreated units
- Randomised or staggered trial
- allocate or phase the intervention by a pre-set rule
- Contribution analysis
- test an explicit causal chain with mixed evidence
Main advantage
- Before-and-after
- quick and understandable
- Matched comparison
- supplies an observed counterfactual
- Randomised or staggered trial
- strongest protection from selection bias when implemented well
- Contribution analysis
- useful when controlled rollout is unavailable
Main limitation
- Before-and-after
- season, price, capacity and local events also change
- Matched comparison
- matching cannot remove every hidden difference
- Randomised or staggered trial
- may be impractical, contaminated or underpowered
- Contribution analysis
- supports a reasoned contribution claim, not a precise effect
Defensible wording
- Before-and-after
- outcome changed after the action
- Matched comparison
- effect estimate subject to matching assumptions
- Randomised or staggered trial
- estimated incremental effect for the tested population
- Contribution analysis
- evidence is consistent with contribution
What to take away
- Before-and-after shows an outcome changed but not that the action caused it.
- Matched comparison supplies an observed counterfactual but cannot remove every hidden difference.
- Randomised or staggered trials give the strongest protection from selection bias when implemented well.
- Contribution analysis supports a reasoned contribution claim rather than a precise effect estimate.
- Choose the smallest credible design and state clearly what it cannot prove.
Before-and-after
Fix the measure, window and comparison period before looking at the result. Adjust only under a documented rule. This design may be adequate for an operational check, such as whether a software defect disappeared after a fix, where other explanations are limited. It is weak evidence for marketing return or long-term sales because many influences move together.
Before-and-after design limits
Before change
- Measure
- Fixed in advance
- Window
- Pre-period set
- Comparison
- Baseline period
- Adjustment
- Documented rule
- Causal claim
- Not established
After change
- Measure
- Same definition
- Window
- Post-period set
- Comparison
- Same trading period
- Adjustment
- Documented rule only
- Causal claim
- Sequence is not causation
HM Treasury's impact-evaluation quality guide states that monitoring an outcome can establish change but generally not whether an intervention caused it. Report the observation without upgrading sequence into causation.
Matched comparison
Select comparison outlets using pre-intervention characteristics such as service mix, maturity, capacity and earlier trend. Record the matching rule before viewing post-change performance. Use identical definitions and trading periods, then show where the groups still differ.
Matched comparison vs network average
Matched comparison
- Selection
- Pre-intervention traits
- Service mix
- Matched
- Maturity
- Matched
- Capacity
- Matched
- Earlier trend
- Matched
- Remaining risk
- Local confounders
Network average
- Selection
- All outlets
- Service mix
- Uncontrolled
- Maturity
- Uncontrolled
- Capacity
- Uncontrolled
- Earlier trend
- Uncontrolled
- Remaining risk
- Definition gaps
This is more informative than comparing with the network average, but local competition, management and unrecorded implementation can remain confounders. ONS's framework for data quality highlights coherence and comparability; without those, the apparent contrast may be a definition gap.
Randomised or staggered trial
Where enough comparable units exist and the change can fairly be phased, allocate timing in advance. Define population, intervention, control, outcome and period. Check whether control locations adopt the new practice anyway, whether the rollout happened as assigned and whether the sample can detect a decision-relevant difference.
The 2026 Magenta Book explains how experimental and quasi-experimental methods use a counterfactual to estimate impact. Its government context does not remove the need for commercial judgement, consent, lawful data use or an affordability test.
Contribution analysis
Write the expected chain: activity, immediate change, customer response and business outcome. Gather implementation records, interviews, operational measures and evidence about alternative explanations. Government guidance characterises contribution analysis as an evidenced line of reasoning rather than definitive proof.
Keep observation separate from interpretation. GDS guidance on analysing research recommends distinguishing what was seen or heard from the team's inference. That boundary is equally valuable in commercial review notes.
Select the smallest credible design
Use before-and-after for narrow control questions, a planned comparison for meaningful operational or marketing changes, and specialist experimental advice when a high-value decision warrants it. Choose contribution analysis when causal mechanisms matter but allocation is unavailable. In every case, verify implementation first, publish the definition and state what the design cannot prove.
Before you act
- Fix the measure, window and comparison period in advance.
- Record the matching rule before viewing post-change performance.
- Define population, intervention, control, outcome and period.
- Write the expected causal chain before gathering evidence.
- Verify implementation before interpreting any result.
- Publish the definition and state what the design cannot prove.
Common questions
Which attribution design gives the strongest protection from selection bias?
A randomised or staggered trial, where the intervention is allocated or phased by a pre-set rule, gives the strongest protection from selection bias when implemented well. However, it may be impractical, contaminated or underpowered, so specialist experimental advice is worth taking for high-value decisions.
When is contribution analysis the right choice?
Use contribution analysis when causal mechanisms matter but controlled rollout is unavailable. It tests an explicit causal chain with mixed evidence, including implementation records, interviews, operational measures and evidence about alternative explanations. It supports a reasoned contribution claim, not a precise effect, so keep observation separate from interpretation.
What should be checked before trusting a matched comparison?
Select comparison outlets using pre-intervention characteristics such as service mix, maturity, capacity and earlier trend. Record the matching rule before viewing post-change performance, use identical definitions and trading periods, then show where the groups still differ. Local competition, management and unrecorded implementation can remain confounders.
