Clarification required for running Non-Revenue Model

[annie-gulati]

Issue1: We are consistently getting low baseline
Issue2: We scaled the Target variable by a factor of 1,10, 20, 100 etc. We are getting very different result for media contribution. In general, we should not be seeing this. Change in scale should be adjusted in the Coefficient and variable rank order should maintain.
Please guide us on these concerns.
You can see below the result:

Factor 1

Factor 10

[cpulavarthi]

Hello @annie-gulati,

Thank you for contacting us!

For Issue 1 you mentioned, please note that the negative baseline estimate means that the revenue attributed to the treatment variables (paid media, organic media, and non-media treatments) exceeds the historical revenue. In other words, ROI estimates are too high. This can happen if you haven’t accounted for all of the confounding variables, or when the model isn’t sufficiently explaining the time effects, or when there is low information in your data and informative ROI priors are required. Please check our documentation on Low or Negative Baseline Attribution for guidance on how to debug this issue if you haven’t done so already.

Regarding Issue 2, scaling the target variable (by changing either the kpi directly or the revenue_per_kpi factor) is expected to affect the results when you don’t correspondingly scale the chosen ROI prior distribution. Please note that the internal data transformations for the KPI are scale invariant (Ref - Transformation: KPI Units, however If you change the total revenue (or total KPI if revenue data is not being used) but do not change the ROI prior distribution, then you are effectively changing the prior distribution on the percentage of revenue attributed to each channel. This will have a significant impact on the model fit.

Feel free to reach out if you have any further queries regarding this.

Thank you

Google Meridian Support Team