Best practices for Multi-Product modeling using "Stacked Geos" (Geo * Product hierarchy)

[PavanKumarDharmoju]

Hi Meridian Team,

I am exploring ways to model a multi-product portfolio (e.g., Product A and Product B) within Meridian. Since the library currently supports a single KPI target per model, I am evaluating a "Stacked Geo" workaround where I treat Geo-Product combinations as unique units in the geo dimension.

• Standard Geo: New York
• Stacked Geo: New York_ProductA, New York_ProductB

I have successfully transformed the data tensors to fit this structure, but I have three specific questions regarding the statistical validity and feature engineering required to make this robust:

1. Population Scaling & Target Audience:
Since Meridian normalizes media spend using the population array (), simply repeating the census population for both NY_ProductA and NY_ProductB seems incorrect, as it would dilute the media density.

• Question: In this stacked approach, is the best practice to replace census population with Product-Specific Target Audience sizes?
• Concern: If I do this, does the model handle the variance in population scale well (e.g., if Product A has a 1M audience and Product B has a 10M audience in the same physical geo), or will this variance destabilize the hierarchical priors?

2. Handling Shared "Brand" Media:
This stacked structure implies that NY_ProductA and NY_ProductB are independent territories.

• Question: How should we handle "Brand Level" media (e.g., a Brand TV spot) that theoretically impacts both rows?
• Concern: If I split the Brand Spend arbitrarily (e.g., 50/50) between the two stacked rows, I introduce bias. If I duplicate the full spend for both rows, I inflate the total budget. Is there a recommended constraint or prior setup to handle shared media across stacked geos without violating independence assumptions?

3. Hierarchical "Borrowing" of Strength:
One of my main motivations for stacking is to allow a new product (Product B) to "borrow" statistical strength from a mature product (Product A) via the hierarchical nature of the model.

• Question: Does the geo-level hierarchy actually facilitate information sharing between NY_ProductA and NY_ProductB? Or, because they have distinct geo IDs, does the model treat them as completely disjoint populations with no correlation other than sharing the global distribution parameters?

Any guidance on the feasibility of this architecture would be appreciated!

[dattatreyam23]

Hi @PavanKumarDharmoju ,

Thank you for contacting us!

While it is possible to use non-geo hierarchies in Meridian, we generally recommend against using a product-level hierarchy, especially when halo effects are expected (i.e., media spent on one product impacts the KPI of another). The halo effect seems to be present in the brand TV example you provided and “splitting” channel execution will not resolve the issues from a causal inference perspective. See non-geo hierarchy for more information.

If you are confident that no halo effects are present and would still like to proceed, it would be preferred to use geo-product-specific target populations on these channels. You may also consider setting unique_sigma_for_each_geo = True to account for differences in variability across geo-products. Keep in mind that under the model hierarchy, coefficients (betas) will benefit from partial pooling across geo-products and Adstock, Hill, and seasonality parameters will be completely pooled.

Feel free to reach out if you have any questions or suggestions regarding the same.

Thank you,

Google Meridian Support Team