Is this common for everyone else? Little variation in adstock curves.

[Jon77Ruler]

I'm wondering if this is common for others too?

Very rarely I'll get one channel that sticks out from the pack, but across multiple Meridian models using different business' data, I pretty much get the default decay curves with most channels following it quite closely (i.e. after week1, all channels have about 50% left).

I'd REALLY like to be able to use a weibull cdf, as it feels like Meridian doesn't do much to calibrate the geometric curves unless you set different priors for each channel. Even then, it's almost like a minor consideration in the model fitting process.

[AdimDrewnik]

Maybe your data is just not providing any information for a posterior to move away from the prior.

[cpulavarthi]

Hello @Jon77Ruler,

Thank you for contacting us!

It's quite common for the adstock decay curves to appear similar across channels in Meridian models, especially when the underlying data doesn't contain a strong signal to differentiate them.

You might try adjusting the max_lag argument in ModelSpec to change the maximum number of future time periods affected. You could also try using custom alpha_m priors.

Our team is constantly working on improving Meridian's functionality and you may follow the CHANGELOG page in GitHub.

Feel free to reach out for any questions or suggestions relating to Meridian!

Google Meridian Support Team

[Jon77Ruler]

" It's quite common for the adstock decay curves to appear similar across channels in Meridian models, especially when the underlying data doesn't contain a strong signal to differentiate them. "

I fully appreciate that Robyn works very differently to solve model fit, but I know that when I put the same data through Robyn, even sticking to geometric curves, I will get far stronger differentiation between channels. I'm fortunate that I have access to multiple different sets of data, so I would say in my experience this is common. I am just wondering how "hard" Meridian works the adstock section part of the model to achieve fit, and whether a flexible baseline with high knot count is detrimental in these cases?

[cpulavarthi]

Hello @Jon77Ruler,

Here are some suggestions. Please also check Model debugging.

• Customize alpha priors. If you have domain knowledge of your media channel’s adstock decay rates, you can incorporate it through custom alpha priors.
  For example, you have three channels and your prior belief of their decay rates are described by Beta(1, 9), Beta(10, 90), Beta(1, 1), respectively. The first two channels share a mean decay rate of 10% with varying levels of confidence; the last channel is essentially uniform, indicating an uninformative prior. The corresponding code is as follows.

prior = prior_distribution.PriorDistribution(
  alpha_m=tfp.distributions.Beta([1, 10, 1], [9, 90, 1])
)
model_spec = spec.ModelSpec(prior=prior)

• Signals in data. If your data doesn't contain a strong enough signal to differentiate the true, underlying carryover effects across channels, Meridian's Bayesian framework will lean more on the default uninformative priors. This can result in posterior distributions for alpha that appear similar across channels. This often happens for channels with low spend, very low variation in their media execution, or high correlation with other channels. Here are some suggestions that might boost signal in data.
  • Lower the scope of the MMM. You can estimate fewer media channels (either by dropping a channel with low-spend or combining channels), use fewer knots to estimate time effects, and remove any extraneous controls. However, don't remove important confounders.
  • Get much more data. For example, use three years' of weekly data instead of two. Adding more data will reduce the variance in inference, but might make the inference less relevant.
  • Alternatively, consider adding geo granularity to your data and using a geo model instead of lowering the scope or adding more data.

Do contact us if you have any further questions or suggestions regarding Meridian!

Google Meridian Support Team

[Jon77Ruler]

Thank you; I knew about alpha_m priors but the clarification on Beta coefficients for decay confidence was extremely helpful.

I think my overall concern was that maybe with so many other levers that can be adjusted to fit the data (particularly the knot setting for baseline), whether adstock was taking a bit of a back seat given that across multiple models I don't see much variance there [from the default alpha_m settings].

Intuitively with many offline channels you'd expect to see longer adstocks, but that doesn't seem to get picked up at all in Meridian (whereas it does in Robyn). That's hard to explain to the business.

Apologies - I know that doesn't make it right. I'm just speculating because I've not a detailed understanding of the model workings! For reference; we're always using >2 years of weekly, National data, and this is across multiple sets of different business' data.

[cpulavarthi]

Hello @Jon77Ruler,

Thank you for reaching out!

I would first like to clarify that knots are specified (by the number of knots and/ or their locations) to capture time effects. They aren't specifically set only for the baseline.

I wouldn't say that capturing the carryover effects isn't prioritized because of the customizability options in Meridian. Hill-Adstock transformation capturing the saturation and carryover effects are the first ones to be applied on media spend (after the initial data scaling of course). Meridian uses geometric adstock to capture the carryover effect so a max_lag in the range of 2-10 (default value is 8) provides a balance between capturing the carryover effects, improving model convergence, managing model runtimes and maximizing data usage (reducing variance). Setting max-lag and customizing alpha priors to incorporate your business knowledge regarding the media behaviour will definitely help the model better capture carryover effects.

Thank you again for your suggestions and insights. We shall take them into consideration to improve capturing the carryover effects. Feel free to contact us if you have any other questions or suggestions regarding Meridian.

Google Meridian Support Team