KPI definition with regards to the types of variables

[SteliosSake]

Hello everyone!

I have a question regarding the definition of the KPI, y variable in the model specification.

Lets have a scenario were we have:

• 8 marketing channels
• 1 non-media treatment variable (discounts)
• 4 control variables like an economy index, competitor sales, last month's marketing spend (sum of all media-treatments) and organic demand (deriving from unknown sources or direct purchases to the site without an ad intervention).

My question is the following:

• Regarding modeling based on causality rules, should the KPI be gross revenue (meaning that the KPI = Revenue + Discounts) because we want the non-media treatment variable of discounts to be part of the KPI? or should the KPI be net revenue (meaning that the KPI = Revenue - Discounts), because it is okay for the KPI to not include a variable that is considered a non-media treatment in terms of causality modeling?

Additionally:

• Can the same can be asked regarding the rest types of variables (e.g., regarding control variables, if hypothetically I HAD to consider discounts as a control variable, meaning that it is not intervenable, then the KPI should include discounts?)

This question came up due to the fact that many people have told me to consider the KPI being the net revenue but I don't know if this is correct, because if part of the KPI is a variable to the model (either this variable being control, media or non-media treatment), then by subtracting it from the KPI we are in some manner "taking it already into account", so that steals the chance of the meridian model to understand its complete relationship.

I would really like to hear your thoughts on this!! Thanks in advance for your time and effort!

Kind regards,
SteliosSake

[dattatreyam23]

Hi @SteliosSake,

Thank you for using Meridian!

I understand that there are two questions here, (1) about the definition of revenue and KPI and (2) inclusion of discounts as one of the variables in the model.

For (2), if discount is an important variable that impacts KPI, we should include it in the model. Otherwise, we risk introducing bias or even causing a negative baseline. Since discount is an intervenable measure, and discount is not media-related but product-related, we should treat it as a non-media treatment. If discount is non-intervenable but an important confounder (i.e., causally affects both KPI and other media channels), then we should treat it as control.

For (1), based on your question, maybe allow me to first clarify the word regarding ‘revenue’? Can we understand as follows:
Gross revenue = sales revenue,
Net revenue = sales - discounts ?

In this case, from a causality standpoint, setting either gross revenue or net revenue as the KPI is fine, as long as we include discount in our model (if it is important), and we are clear on our model interpretation (e.g., what are the estimated causal effects of media on gross revenue or net revenue). We recommend you use whichever KPI definition that is most useful to the business.

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

Thank you,

Google Meridian Support Team

[SteliosSake]

Hello @dattatreyam23 ,

First of all thank you for taking the time to answer to my questions. Allow me to answer to some really important points you raised.

(2)
Regarding the 2nd question and the inclusion of the "discounts" variable, I know that it is important so I have made two analyses, one treating it as control and one treating is as a non-media treatment. As I understand, the only thing that changes between them is the interpretation of the baseline. For example in the analysis with the KPI being net revenue:

• For discounts as control, the discounts are treated as not intervenable leading to the assumption that it is something that we can't control. Hence the baseline is just the net revenue generated if no media spend was utilized.
• For discounts as non-media treatment, discounts are intervenable, leading to the baseline being treated as the net revenue generated when no media spend was utilized and discounts were at their minimum value.

Correct me if I am wrong at any of the points I raised.

(1)
Regarding the 1st question, i am glad that it is okay to choose whichever KPI we want. To answer your question regarding definitions , we define revenue as follows: Revenue = Sales + Discounts. So Gross Revenue for me is Sales + Discounts while Net Revenue is Revenue - Discounts (or just Sales). The only true difference between my definitions of Gross and Net revenue is that gross revenue contains discounts inside it, while net revenue is after discounts (they are not included).

My follow up question on this is:

• Is it correct to define the KPI as Gross Revenue (Sales + Discounts), when I have already included Discounts as an input variable in the model (either as a control or a non-media treatment)?

The question stems from my wondering if it is okay to have a formula Y = f(X1, X2, X3, ...), with Y being the KPI and X1, X2, X3,... being the input variables when we know that Y includes one of the input variables (lets say X3, so Y = Z+X3). This changes the formula to: Z+X3 = f(X1, X2, X3, ...). Discounts is X3 in this example and Y is gross revenue because Gross Revenue (Y) = Sales (Z) + Discounts (X3).

Thanks again for all your help!

[cpulavarthi]

Hello @SteliosSake,

Thank you for sharing detailed additional information!

For (2), Your interpretation about baselines is sound. To be more specific, by default, Meridian specifies the baseline or counterfactual value for a non-media treatment to be the minimum value, but it could also be a user-supplied value.

For (1), thank you again for further clarification. It appears that you are making a choice between two models:

• Model 1: Z+X3 = f(X1, X2, X3),
• Model 2: Z = f(X1, X2, X3),

where Z = Sales, X3 = discount, and X1 and X2 are two other input variables.

In Meridian, non-media treatments and controls have a linear relationship with KPI, so you could rewrite Models 1 and 2 as

• Model 1: Z + X3 = f(X1, X2) + gamma * X3
• Model 2: Z = f(X1, X2) + gamma * X3

Then Model 1 could be re-expressed as Z = f(X1, X2) + (gamma-1) * X3.
So, again, we recommend you use whichever KPI definition that is most useful to the business.

Hope this helps.

Do reach out if you have any further questions or suggestions.

Google Meridian Support Team