How to model Retail shops in a geo hierarchy model

[georgemarsci]

Hi there,

I tried to search but couldn't find relevant information.

How can you structure retails shops in a model with geo hierarchy.

Thanks

[dattatreyam23]

Hi @georgemarsci ,

Thank you for contacting us!

If you are referring to modeling retail shops as geos, it is generally not recommended.

Why modeling retail shops as geos is not recommended

1. To be able to use geo-level modeling, data input such as media, media spend, KPI, must be broken down into mutually exclusive geographic regions (in your case, shops). While sales (KPI) can often be tracked at the shop level, it is typically difficult to accurately attribute media impressions or spend to a specific individual shop.

2. You must provide a "population" variable for each geo (shop), in order to scale media metrics. This helps put different units on a comparable scale and assists in model convergence. For non-geographic categories like individual shops, defining a meaningful scaling factor (such as a shop's specific trade area population) is difficult.

3. You lose the ability to leverage the advantage of actual geo-level models, because shop is used to replace geos.

4. The Stable Unit Treatment Value Assumption (SUTVA) assumes that media exposure in one unit (shop) does not affect the outcomes in any other unit (shop). Using individual shops as the hierarchical variable often leads to significant "cross-category contamination", where advertising for one shop drives traffic to another nearby location, violating this core assumption.

For more information, please see the official Meridian FAQ: Can Meridian's hierarchical model structure be used for a categorical variable besides geo?

​Hope this helps!

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

Thank you,

Google Meridian Support Team

[georgemarsci]

Thank you so much @dattatreyam23

Great. Now I get it. How should I then model a geo hierarchical model (one country with multiple regions) while I have shops (physical locations) as well. Should I include let's say Sqm as control variable? How's the best practice?

[dattatreyam23]

Hi,

You should consider including total square meters (Sqm) per region as a control, if you think it causally affects both your target KPI (sales) and your media execution. For example, if regions with a larger physical retail footprint naturally generate more sales and historically receive higher media budgets, Sqm acts as a confounding variable that must be controlled for to get unbiased causal estimates of your media.

When preparing your data, ensure the aggregated Sqm control variable matches the dimensions of your KPI data (geo, time). Because the total square meters of retail space in a region will likely scale with the region's population size, you should enable population scaling (the control_population_scaling_id argument in ModelSpec) for this control variable so Meridian can accurately put all geos on a comparable scale.

Note, control variables in Meridian must exhibit variation across time and, depending on your model configuration, across geos. You can leverage Meridian's EDA module for these checks. See more details in the official documentation Perform an exploratory data analysis.

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

Thank you,

Google Meridian Support Team