An MMM is not reliable just because it has many weeks.
An MMM is not reliable just because it has many weeks. It must observe spend variation sufficiently distinct from trends, other channels and commercial decisions. 156 weeks of channels moving together can provide less information than 52 well-differentiated geo-test weeks. Audit identifiability, grain, variation, collinearity, demand-factor coverage and traceability before promising channel returns.
Decision and method.
Authorise an MMM only at the aggregation level that observed variations can actually identify; aggregate or experiment on channels that are insufficiently distinct. List outcome, channels, controls, grain, window and expected decisions, excluding variables that lack stable definition over the full period. Measure non-zero weeks, coefficient of variation, level changes, geographic variation and support near current spend. Examine correlations, variance inflation, conditioning and alignment with promotions or season; aggregate variables whose separation relies on fragile assumptions. Link every weakness to taxonomy, geo collection, planned variation, experiment, prior or aggregation, and publish authorised use level.
Worked example.
In this fictional teaching matrix, Search has 100% active weeks, variation 0.28, maximum correlation 0.82; Social 0.32, 0.77; TV active only 45%, variation 0.65, 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. It is not a formal statistical test: TV has best independent variation; group, calibrate or experimentally separate Search and Social before fine allocation.
Checks and limits.
Spend, outcome and controls share geography, period and stable definition; each channel has variation distinct from other commercial decisions; requested allocation matches identifiable level; weak variables receive aggregation or evidence plan. The index does not replace model diagnostics. Low correlation does not remove spend endogeneity. Geographic detail does not help when based on national imputation.
Resources and sources.
Download the MMM identifiability matrix. Related: check quality before modelling, choose the right grain, create experimental variation. Sources: Chan and Perry (2017); Sun et al. (2017); Zhang et al. (2024).