Control variables sugestion needed

[jayjoshi33]

Getting competitor sales data is quite challenging in real-world datasets. I'm looking for suggestions on potential control variables that can be used in Marketing Mix Modeling (MMM), especially when access to advanced MMM platforms or tools is limited.

So far, I've included holiday weeks flag using the holidays library. While Google search query volume is commonly recommended, it's not always accessible so, I'm looking for alternative signals.

My Setup:
I currently compare performance across two channels: Google Ads and DV360

I also test the model at campaign level for Google Ads, where I have access to GA4 data

What I’m Looking For:
Control variable suggestions (beyond query volume and holidays)

Reliable data sources/APIs to get these variables

Recommendations for any additional variables from Google Ads, DV360, or GA4 APIs that could be helpful in the Google Meridian MMM model, apart from standard ones like spend, clicks and impressions to improve the baseline

Thanks in advance! Would love to hear if anyone working on a similar problem

[Jon77Ruler]

You could try weather data; though seasonality may well capture elements of that anyway (e.g. using a time series decomposition, or Prophet or something like that). You can simplify the knots element doing that. Days of rainfall might be pertinent depending on business vertical. There are plenty of open source weather APIs and paid-for ones; depends on nationality needed.

Competitor interest rates (finance). In the UK savings ones tend to change only on Fridays, so it follows an easy cadence for weekly modelling.

I'm also interested to see if there are others.

[dattatreyam23]

Hi @jayjoshi33 ,

Thank you for contacting us.

To strengthen your Marketing Mix Model without direct competitor sales data, begin by incorporating crucial control variables from accessible sources. Start with your own internal data, creating flags for promotional periods, price changes, and new product launches. These account for self-inflicted shifts in demand. Next, integrate key macroeconomic indicators like the Consumer Price Index (CPI) or Consumer Confidence from public sources to control for the overall economic environment.

You can also create valuable competitor proxies, when competitor sales data is not available, like competitor pricing, observed promotional activity (if available), competitor's major product launch etc. since they can act as controls by influencing overall market demand. Social mentions or sentiments, such as positive or negative social media mentions, social reviews, etc. might also help as a control.

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

Thank you,

Google Meridian Support Team

[SteliosSake]

Hello again @cpulavarthi ,

One more important question (sorry for the rain of questions):

• In case I have a good estimate of the total organic revenue the company makes (meaning that I have a good estimate of the part of the KPI that is not due to the marketing activities), is it a good idea to increase the knots up until the point where the baseline outcome matches the estimate of the total organic revenue?

In the above analysis I will be utilizing as control variables either just the economic index or both the economic index and discounts (since I dont have any other controls available and the discounts can be considered not intervenable due to strict strategies in place by management).

Kind regards,
SteliosSake

[dattatreyam23]

Hi @SteliosSake ,

Thank you for contacting us.

To answer your initial queries:

• Can you explain further the part regarding the KPI having too much explaining power?

It's highlighting a potential issue where your model might be capturing a correlation rather than a true causal relationship. If a control variable is strongly driven by the KPI, the model might incorrectly attribute changes in the KPI to that control, or fail to properly disentangle the effect of media from the KPI's inherent dynamics.

Consider an extreme example where you include KPI as control while having the same KPI as model's response. The model will explain all KPI to be driven by the control (KPI) and hides true causal effects.

• ... Which of the two solutions to treat the negative or low baseline is more of a priority in your opinion? Or are they equally important?

  You can try 40-50% knots but should be cautious that a national model is more sensitive to n_knots than a geo-level model, it tends to assign more credits to time-varying intercept as n_knots becomes larger. At extreme when n_knots = n_times, you'll find KPI is fully explained by the intercepts.

  You should definitely investigate high quality controls. The primary purpose of control variables in Meridian is to improve causal inference by debiasing treatment effects. A negative baseline often signals that media is receiving too much incremental credit, which is fundamentally a bias problem. This bias stems from a failure to adequately account for unobserved factors (confounders) that drive both media spend and KPI.

  Therefore, while increasing knots can be a helpful step to better model time-based confounding, the most impactful solution will be to diligently identify and integrate more high-quality controls. This will get you closer to the "true" causal effects and a more realistic baseline.

Here are some additional suggestions:

• Check out Mitigate negative or low baseline.
• Consider geo-level model if possible as it greatly improves quality of estimates. Also, geo-level data can support the use of more knots to model time-varying intercepts.

Thank you,

Google Meridian Support Team

[SteliosSake]

Hello @dattatreyam23 ,

First of all, thank you for your quick reply to my queries! It really aids in accelerating my work.

I will get right on the task of identifying those high quality controls.

After your amazing feedback I have these follow-up questions:

• Regarding the knot selection (national model): I know that it is a good strategy to have dense knots around holidays to capture seasonality /time-varying behavior. In these dense areas of knots, it is better to have them really tight with each other (e.g., having a knot before Valentine's Day week, one knot during the week and one knot after the holiday) or is it better to have at least one or two time periods of space between them?

• Hypothetically, after identifying some high quality controls (lets say 4 control variable and a national model setting), then I should play around with the knot selection until I get a baseline outcome (revenue) that is close to the total organic revenue that the company estimates? I am asking this because you can never be sure of the contribution of time to your KPI.

• Geo level modeling: Right now I am using the national level model because the company does not gather data based on geographic location, but they do gather data based on different brands of products (e.g., the company sells items from 6 different vendors/brands). So instead of using the geo-level model by utilizing geo data, can I utilize the geo-modeling but with each geo representing a vendor/brand? If yes, what are the things to consider?

Thanks again for all your help! I hope my rain of questions actually helps the community of Meridian :P

Kind regards,
SteliosSake

[dattatreyam23]

• In case I have a good estimate of the total organic revenue the company makes (meaning that I have a good estimate of the part of the KPI that is not due to the marketing activities), is it a good idea to increase the knots up until the point where the baseline outcome matches the estimate of the total organic revenue?

Hi @SteliosSake ,

In regards to the above query that you raised yesterday, I would like to post a response. t is not recommended to increase knots just to make the model's baseline match an external organic revenue estimate. This can lead to overfitting and reduce the reliability of your causal estimates. Selecting the number of knots involves a trade-off between bias and variance, so we recommend monitoring the model fit closely (see Assess the model fit and the results and Choosing the number of knots for time effects in the model).

If you have a good estimate of your total organic revenue (which is essentially your baseline), you can use this to inform the model. You can do this by setting a total paid media contribution prior. This helps regularize the model by establishing a prior for the share of the KPI that is incremental from paid media, based on the difference between the total KPI and your organic revenue estimate.

We will get back to you with the latest query that you have raised.

Thank you,

Google Meridian Support Team

[dattatreyam23]

Hi @Steliosake,

Thank you for contacting us.

To clarify your latest queries:

• Knot Selection: Meridian does not explicitly discourage the use of consecutive knots. Instead, it recommends increasing knot density around periods of high seasonality to capture time-varying effects effectively. For isolated holidays, binary dummies may be preferable.

However, be careful not to select an excessive number of knots, as this can lead to overfitting. The tradeoffs between knot density and model stability are described in detail here.

• Geo-level Modeling Using Brands: Modeling brands as geographic units is generally not ideal. Geo-level Bayesian hierarchical media mix modeling (GBHMMM) relies on key assumptions, such as similar media impact on the target KPI across geos and well-defined, mutually exclusive population boundaries. These assumptions often do not hold when using brands as “geos.”

Brands may exhibit very different media sensitivities, which can confuse the model rather than improve estimation via pooling. Additionally, brands are not mutually exclusive in the way geos are, complicating population definitions and potentially violating model assumptions.

Please consider these points carefully when designing your model for the best balance of accuracy and interpretability.

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

Thank you,

Google Meridian Support Team