MMM data: testing identifiability before assembling more rows.
Evidence status
What was reproduced, and what it does not prove.
The CSV and Python verify the three triage indices, their authorised uses and the illustrative priority of TV. They validate neither an MMM, nor endogeneity, nor causality, nor market performance.
- Technical specification of the MMM identifiability matrix, specification in French (opens in a new tab)
v1, 31 August 2026 · CC BY 4.0 · specification in French · Method document written by marketing-mix.net. The figures are illustrative: they are not market measurements.
Key distinctions
Three conditions before calculating.
Usable variation
Useful variation is the part not entirely explained by season, price, promotions, distribution or other media. A high coefficient of variation can remain unusable if it follows another channel exactly.
Coherent grain
Spend, exposures, results and controls must share a geography, a period and a stable definition. Imputing a national channel into each region creates rows without adding independent information.
Proportionate decision
Feasibility depends on the use. Data may suffice for an overall media effect but not to separate two platforms. The authorised decision level must be coarser than the level of uncertainty.
Method
Assessing what the data can genuinely separate
- 01Write the decision model
List result, channels, controls, grain, window and expected decisions; discard the variables that cannot be defined stably over the whole period.
- 02Measure the variation
Compute non-zero weeks, coefficient of variation, level changes, geographic variation and the share of the support close to current spend.
- 03Test the separation
Examine correlations, variance inflation, conditioning and alignment with promotions or season; group the variables whose separation depends on a fragile assumption.
- 04Prescribe the missing evidence
Attach to each weakness an action — taxonomy, geographic collection, planned variation, experiment, prior or grouping; publish the authorised level of use.
The variation × separation × geography matrix
Three channels with the same history receive different feasibility levels according to their variation, their correlation and their geographic heterogeneity.
| Channel | Active weeks | Variation | Max. correlation | Index | Decision |
|---|---|---|---|---|---|
| Search | 100% | 0.28 | 0.82 | 0.030 | Group |
| Social | 95% | 0.32 | 0.77 | 0.049 | Calibrate |
| TV | 45% | 0.65 | 0.45 | 0.129 | Model |
Teaching scenario: 156 weeks are not enough to separate three channels
In this fictional teaching matrix, Search has 100% active weeks, a coefficient of variation of 0.28 and a maximum correlation of 0.82. Social reaches 0.32 and 0.77. TV is active only 45% of the time but its variation is 0.65 and its maximum correlation 0.45.
Triage index = active share × variation × (1 − correlation) × geographic factor. Search: 1 × 0.28 × 0.18 × 0.6 = 0.030. Social: 0.95 × 0.32 × 0.23 × 0.7 = 0.049. TV: 0.45 × 0.65 × 0.55 × 0.8 = 0.129.
The index is not a formal statistical test, but it shows where to investigate. TV offers the best independent variation; Search and Social must be grouped, calibrated or separated by an experiment before any fine allocation.
Acceptance conditions
What must be true to act.
- 01
The grain of spend, results and controls is coherent.
- 02
Each channel has a variation distinct from other commercial decisions.
- 03
The allocation level requested matches the level genuinely identifiable.
- 04
Every weak variable receives a grouping or an evidence plan.
Limits
What this analysis does not prove.
- The triage index proposed does not replace the model's diagnostics.
- A weak correlation does not eliminate the endogeneity of spend.
- Geographic granularity does not help if it rests on national imputations.
References
Works cited.
- Chan, D. & Perry, M. (2017), “Challenges and Opportunities in Media Mix Modeling”, Google Research.
- Sun, Y. et al. (2017), “Geo-Level Bayesian Hierarchical Media Mix Modeling”, Google Research.
- Zhang, Y. et al. (2024), “Media Mix Model Calibration With Bayesian Priors”, Google Research.