How Bayesian Hierarchical Marketing Mix Modeling helps optimize regional media planning

Classic Marketing Mix Modeling (MMM) is an established tool for deriving recommendations for action in media planning. In our previous articles, we have already shown how the results of an MMM can be used and what advantages the methodological extension by a Bayesian approach offers.

For some time now, there has been a renewed interest in MMM for media planning, as they are very well suited to map the effectiveness of all relevant media channels and are not limited to the digital domain. However, they are usually only used on a national level for one brand, although many brands would also be interested in planning media for different regions or products. Using a traditional method, however, would involve an enormous amount of work to create individual models for each region or product. Other modern methods, on the other hand, are associated with poor interpretability.

Therefore, Bayesian Hierarchical Marketing Mix Modeling (BHMMM) is the appropriate approach with some methodological and practical advantages to optimize media planning for different subgroups of a brand such as regions or products.

In this article, we will answer the following questions:

• What is the difference between a BHMMM and traditional methods?

• What methodological and practical advantages does it bring with It?

• How to use it to optimize regional media planning?

What is Bayesian Marketing Mix Modeling?

The concept of marketing mix modeling is further developed in a Bayesian approach by using Bayesian statistics, which is based on the use of probabilities. This methodology makes it possible to incorporate prior knowledge and combine it with the data in the modeling process, which leads to more robust results and allows statements to be made regarding their certainty when interpreting them. An extension is Bayesian Hierarchical Marketing Mix Modeling, which allows the data for a large number of subgroups to be used in one model.

One of the methodological advantages of a BHMMM is that in one modeling run, the analysis of all subgroups happens together. For example, if the model is run for the different regions in which a brand is on the market, the regions form a larger system in which they learn from each other. A general media effect is thus assumed and general trends are also mapped, but individual characteristics of the regions are also taken into account so that a media channel can have a differently strong effect depending on the region. In addition, this also makes it possible to analyze regions for which data is only available over a shorter period of time because the procedure uses information from other regions to validate the influences of various factors. Furthermore, a hierarchical approach automatically uses a larger data set compared to a national model, i.e. a larger sample from which to learn. This leads to robust models and, by definition, information on their uncertainty is possible for all model parameters resulting from the Bayesian approach. Another advantage of the BMMM is that the parameters that reflect the time-delayed media effect (adstock effect) and the saturation of a media channel are also determined directly in the modeling process. The determination of these factors is associated with a time-consuming iterative process when using classic MMM methods, which is why a Bayesian approach contributes to a significant simplification in this point. Moreover, in a BMMM an extension is possible so that a changing media effect over time is taken into account. In contrast, classical methods provide a fixed point estimate for the influence of the respective variables. This makes the BHMMM a very comprehensive and flexible approach.

For media planning, Bayesian Media Mix Modeling has some practical advantages, as all model results are available for the subgroups, e.g. for the different regions. Thus, it can be seen, for example, if in region A the influence of out-of-home campaigns is comparatively high and in another region perhaps digital campaigns work particularly well. Based on the model results, media planning is then possible to adapt to the different regions. Implications for campaign planning can therefore be derived and the effects compared by means of forecasts for all the regions included. From this, it becomes obvious which of the plans promises the greatest effect on the target key figure (e.g. sales, website traffic, advertising perception, etc.). Furthermore, an optimization of the budget allocation to the different media channels is also possible for the regions, so that the different effectiveness depending on the region is also taken into account in the distribution of the budget.

What recommendations for action can be derived from Bayesian Hierarchical Media Mix Modeling?

Using a client example, we will now show how a BHMMM concretely supports regional media planning. For one of our clients at MMT, we implemented a project for 50 cities in which the brand is active. The aim was to provide the media agency in charge with recommendations for more efficient planning for each region and to compare different campaign plans based on the resulting model. In addition to media use, various factors were included to explain sales in the different cities. On the one hand, these were general influences such as seasonality, weather, and holidays. Furthermore, customer-specific influencing factors such as fees, new product launches, and an indicator of brand interest were taken into account.

The descriptive preliminary analysis already revealed differences in sales growth and, for example, in the reaction of sales to external factors such as the introduction of fees. These findings already provide a basis for the later result interpretation of the MMM. From a methodological point of view, a Bayesian approach was considered useful for this project, because the multitude of regions can be mapped well, and finally results from the MMM can be interpreted well. At the end of the project, it was confirmed that valuable strategic insights for regional media planning could be derived from the model results.

For all regions, it was shown how strongly the various influencing factors contributed to sales in the period under consideration. The consideration of the media contribution per region gives first indications of which channel in which region contributes how strongly to sales. From this, the return on investment or the cost per order can be calculated, which allows a deeper comparison of how efficient the different channels are in the respective regions. This is the basis for later decisions regarding the budget allocation per region and channel.

Conclusion

With hierarchical Bayesian media mix modeling, the general suitability of a media channel for different regions can be compared. This allows recommendations for the budget allocation per media channel and region, as it shows in which regions the use is particularly favorable or costly. Furthermore, concrete implications for the weekly media deployment per channel and region can be derived and compared by means of forecasts for different campaign scenarios in order to make a well-founded decision for optimized media planning.