How to handle too many channels and channels with limited historical data

[aadityachouhan-ah]

Hello, I have couple of questions around channel inclusion in Meridian.

1. The Meridian documentation advises training only on channels we want to evaluate. However, since we don’t know the KPI contribution of each channel upfront, we can’t simply exclude others. Is the recommended approach to group all unselected channels into a single aggregate channel?
2. What’s the recommended strategy for introducing new channels with limited history? If modeled separately, their posterior may closely mirror the prior due to insufficient data. Is there guidance on how many weeks of data should be collected per channel before including a channel for performance evaluation?

Thanks,
Aaditya

[dattatreyam23]

Hi Aaditya,

Thank you for contacting us.

1. You are correct we do recommend only including the channels that you are interested in estimating causal effects for (see model-fit). Is the issue that you have too many channels and are having issues with convergence? As a guideline we do recommend maintaining the number of media channels below 20 to help ensure sufficient variation and volume for each media channel to yield a robust estimation. This is a guide though and if you have sufficient data it may not be an issue. If you are having convergence issues and there are channels with low spend it may be be reasonable to include them into a single channel which may help. Please see media granularity for more information.

2. We don't have specific guidelines around channel length at the moment. amount-data-needed are the general guidelines for the whole model. I would suggest including the channel in the full model with the other channels you do have a full history for but encode the new channel with 0's in the period before it was active. If it ends up where the posterior = prior for that channel's media effects after fitting the full model then would suggest not enough signal.

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

Thank you,

Google Meridian Support Team