Michael KaminskyMarketing Science | Causal Inference | Incrementality

How Hierarchical Models Fix One of the Biggest Flaws in Media Mix Modeling

Name: How Hierarchical Models Fix One of the Biggest Flaws in Media Mix Modeling
Uploaded: 2026-06-08T13:00:23+00:00
Duration: 1 min 43 s
Channel: Michael Kaminsky
Description: In this segment Michael Kaminsky explains how Recast uses hierarchical models to address a key flaw in traditional media mix modeling: the inability to…

8 Jun 2026 · 2 min

Watch on YouTube

1:43

In this segment Michael Kaminsky explains how Recast uses hierarchical models to address a key flaw in traditional media mix modeling: the inability to capture the two‑stage conversion funnel involving branded search. Branded search—when users type a brand name into Google—can be driven both by existing brand awareness and by other marketing activities (e.g., TV ads) that prompt the search. Simple linear regression treats each channel independently and misattributes the effect of ads that indirectly generate branded search. Recast’s hierarchical approach first builds a model that predicts branded search volume from all other marketing channels, then incorporates that prediction into the overall media mix model. This structure lets the model understand the causal relationship between upper‑funnel activities and downstream branded search, leading to more accurate attribution of spend across channels.

Chapters

So hierarchical models often mean different things. It's a very general concept. At Recast, again, we use the idea of a hierarchical model to model a couple of different things. But the most important is that we allow the model to capture the relationship between certain types of special channels. And in specific, those special channels are generally things like branded search. So branded search, of course, is when someone searches for your brand on Google. Go to Google, search for Coca-Cola. The theory is that branded search basically comes from two different sources. So one is ongoing brand awareness, which is probably a lot of what Coca-Cola searches come from, but also from other marketing that you're doing. Maybe someone sees an ad for Coca-Cola on TV and they say, "Oh, Coke is launching some new interesting product. I'm gonna Google Coca-Cola and then try to learn more about that product." What happened was that the user saw the ad on television that caused the branded search and then maybe eventually caused a purchase down the line. The problem is that if you use a simple media mix model or like a simple linear regression to try to model this marketing mix, the model will get it wrong, right? Because the model won't understand this multi-stage part of the conversion funnel where someone is seeing one ad and then seeing another type of ad because they saw the first ad, right? It's really important to build into the model some understanding of the two-stage nature of the interaction of these marketing channels. And that's what we do with hierarchical models in Recast is we build in a separate model that estimates which channels are driving branded search activity most, how much of our branded search spend can be predicted by all of the other activity in the marketing channels. And this allows the model to better understand those true relationships.