A larger panel can still contain little new information
Geo-level data can provide useful variation, but duplicating national media values across every region does not create independent regional exposure. Before fitting Meridian, establish which inputs are genuinely measured by geography and which are allocated estimates.
- Define stable geographies
- Audit the geo-week grid
- Reconcile national totals
- Check usable variation
Record the source and allocation rule for each series. The Meridian guide explains the modelling workflow; the data requirements guide covers the business contract behind the arrays.
A reproducible workflow makes data checks, modelling assumptions and decision sensitivity visible. The mmm tools, code and platforms learning path connects this idea to the assumptions, practical checks and next steps.
To work through the connected decisions, read MMM in Python: validate the weekly data contract and Geo-lift market selection: review more than pre-test fit.
Build a complete geo-week grid
Define stable geographic identifiers and one week-start convention. Check for duplicate geo-week rows, missing periods and boundary changes. A missing region should not silently disappear from national totals, and a renamed region should not become a new observational unit.
Suppose three regions each contribute £100,000 in one week, but one is absent the next week. The observed national total falls from £300,000 to £200,000 even if the business is unchanged. Filling the missing region's outcome with zero would convert an ingestion problem into a false sales shock.
Check units and scaling inputs
Keep KPI units, currency and media definitions consistent across regions. If population is used for scaling, document the source, date and geographic boundary. The population of an administrative area may not match a media market with a similar name.
For reach-frequency channels, preserve the same geography and time granularity as the relevant model inputs. Check whether reported reach is a count or a percentage and whether frequency is calculated over the same window. Do not sum deduplicated reach across overlapping periods and call it a unique audience.
Reconcile before investigating model diagnostics
Aggregate the panel to national totals and compare it with finance and platform reports. Explain differences in coverage, currency, attribution or timing. Inspect channels that are constant, nearly identical across regions or active only during a single event. More geographies do not guarantee that channel effects can be separated.
Maintain a data-quality report with coverage by week, duplicates, revisions, source totals and allocation rules. Freeze the validated input version before sampling. If a diagnostic later looks strange, this report helps distinguish a model problem from an upstream data change.
Where regional variation is weak, consider whether a national model, a combined channel or a deliberately designed geo experiment better fits the decision. The geo-lift guide describes how interventions can create useful variation, but an observational panel is not itself a randomised experiment.
Sources and further practice
This article’s worked example is illustrative, not a measured client result. Use the linked methodology to check the assumptions before applying it.