We’re working on a hierarchical model to combine results from different TV series across a broadcast season. However, each show follows a different schedule some premiere in the Fall, others midseason, and they vary in episode counts and duration. That means the series don’t share a consistent timeline or presence throughout the season.
On top of that, marketing for each show tends to follow a “blitz” pattern, with intense activity during specific windows (e.g., premieres, finales). Has anyone come across examples of hierarchical models that handle:
- Entities (like shows) with non-overlapping or staggered timelines?
- Marketing or treatment effects that follow a concentrated “blitz” or pre-launch pattern?
Any papers, blog posts, or code examples would be greatly appreciated!
It’ll depend on how you’d like those hierarchies to be structured, and how many levels you’ll be dealing with. As always, it’s best to start with something that’s as simple as possible while still being useful. Make sure it works, and then build from there. It’s worth thinking carefully about what information should be pooled and what shouldn’t. Is there a way to simplify your problem and then work up your desired complexity in stages?
For coding hierarchical models in PyMC, the Radon example is a great starting point, if you haven’t seen it.
1 Like
What is the quantity of interest you’re trying to model and infer? Is it something like predicting popularity of a show in terms of fraction of viewership or overall viewership (working in terms of proportions/composition vs. absolute number is a big distinction in models, though sometimes they can be factored). Whatever you’re doing, you’re going to need to include effects for things like seasonality, day of week, holidays, etc.
I’d say that’s more the norm than the exception with real data. For example, Covid patients came in acyclically on a staggered timeline within an overall seasonal and ad hoc dynamics.
You can include effects for timing before/after launch and provide overall effects and duration of effect.