How to handle paused campaigns data?

[jayjoshi33]

I have 100 weeks of data for 3 campaigns from a single marketing channel. I'm using this data to run a model to analyze the performance of each campaign. However, due to certain reasons, the campaigns were paused for about 4–5 weeks in between.

During these paused periods, all metrics (spend, impressions, conversions, etc.) have values 0. I’m not sure how to handle this, especially because removing these rows might disrupt the modeling particularly the adstock calculations, which depend on time continuity and lagged values.

This seems like a common issue in real-world datasets. What is the best way to address it?

[gousesb]

Hello @jayjoshi33 ,

Thank you for contacting us!

To begin, we suggest gathering additional data as the current amount might not be sufficient for running Meridian. For geo-level models, a minimum of two years' worth of weekly historical data is recommended, while national-level models need three years. If only monthly data is available, we recommend a minimum of three years' worth. Adequate data points are crucial for accurate model calculations. However, the exact amount of data needed can be complex and depends on your specific data characteristics. For more detailed guidance, please refer to Amount of data needed.

Secondly ,I would suggest keeping all rows with zero values. These entries are not data errors but are crucial observations of business reality, and removing them would invalidate your model. Retaining this data is holistically essential for several reasons. It preserves the temporal integrity of your time series, as deleting rows would break the weekly sequence and corrupt the model's understanding of trends and lagged effects. It also ensures proper causal inference, as the pause acts as a natural experiment where the model can observe how performance behaves when the marketing stimulus is removed. Most critically for your specific concern, keeping the zeroes is required for accurate adstock parameter estimation.Deleting these rows would make the model blind to this effect, causing it to severely underestimate decay.

Feel free to reach out for any further queries or feedback regarding Meridian.

Google Meridian Support Team