There is no single approach that answers all marketing questions - Marketing Mix Modeling has its purpose too

When it comes to evaluating the efficiency of marketing activities, many questions quickly arise that are diverse and interwoven and all have a reason to be answered. This is because their answers can be used to inform data-driven marketing decisions and are of help to various stakeholders. Even though Marketing Mix Modeling is a flexible approach that makes it possible to use many influencing factors to explain sales or revenue, it is important to keep the focus and not exaggerate this approach. Three main questions are answered with MMMs:

1. What was the impact of the marketing channels on my sales or revenue?
2. What is the return on investment (ROI) per marketing channel?
3. How can I optimise my marketing budget allocation for the future?

To bring it even more to the point: MMM measures the causality of media effects on a target variable, it is not about the predictive accuracy of the target variable in general.

Keep focus when selecting the influencing factors to be considered

What influence does focusing on the above questions have when it comes to selecting the variables to be included in the model? The model should explain the impact of marketing channels on sales (or any other KPI that is relevant to the company) - that's why the main variables to be included in the model are the marketing channels themselves. Here it is recommended to go to the granularity level that is relevant for the budget decision making. For example, search ads can be further categorised into brand, non-brand, and pmax. Many other variables can be considered in the discussion of influencing factors, such as holiday periods, weather, seasonal factors, or special events. These can sometimes have a strong influence on the explanation of the target variable. However, the question is whether they also influence the decision regarding the advertising budget.

And here is the point where there is no right or wrong decision. Rather, depending on the individual case, a distinction needs to be made as to which other non-media factors are necessary. These non-media factors are also referred to as control variables, which are further defined by their role in the models.

Confounding control variables affect both media and sales. A good example of this is a demand-driven increase in search queries thus influencing paid search ads as well as sales. Usually, it is recommended to include confounding control variables in MMMs.

Mediator control variables are affected by media exposure in the first step and positively affect Sales in the second step. Usually, it is recommended to exclude those variables as they may bias the MMM.

Predictor control variables have an isolated effect on Sales, without interacting with media variables in the model. The price is regularly cited as an example here.

It is often the case that the effects and their size are only estimated in advance of the modelling, but are not known in detail. Before making decisions about their inclusion or exclusion, we recommend analysing the relevant data of possible control variables and looking at their correlation with the media and sales variables. Depending on the strength of the influence and the individual requirements, a solution can then be defined together.

This discussion shows that although it makes sense to follow a standardised procedure when creating MMMs, there can certainly be no one size fits all solution.

Happy analysing

Source: Google Developers - Control Variables