Hello @aadityachouhan-ah,

Thank you for contacting us!

Meridian currently does not offer time-varying effects. As such, an implicit assumption of the model is that media effectiveness is constant across the modeling time period. However, there are still some ways to emphasize a particular period in Meridian’s current offering.

If a single channel’s behavior is assumed to be very different in one time period compared to another, a simple solution would be to separate the channel into two: one media channel for each time period. This allows the model to learn these differences without requiring separate models.

In optimization, you can also specify the start_date and end_date so that the optimization can be done using only a subset of time periods. This subsetting does not affect parameter estimation but does allow you to take into account differences in historical media execution.

Additionally, if you are using incrementality-informed ROI priors, you can specify the calibration period in the model as well. See our documentation on Setting ROI Calibration Period for more information.

The sequential approach described can lead to meaningfully different results than using the full time period, especially when there are significant differences in media effectiveness and/or execution patterns. While we don’t have specific guidance for using a two-stage model, we would suggest that if the two time periods are very different, training a model on only the most recent period may be preferred if the above approaches do not work. In that case, setting informative priors (for all parameters but especially ROI) based on intuition, incrementality experiments, and/or past MMM fits seems reasonable.

Feel free to reach out if you have any further questions.

Google Meridian Support Team