5 Questions on Marketing Mix Modeling

What can marketing mix modeling do? What can’t it do, and what should you pay attention to when choosing methods and providers? We’ll go through the five most important questions that you should ask.

Marketing mix modeling arose from the realization that the simultaneous effects of individual marketing measures can’t be determined through descriptive data evaluation. Correlation coefficients or key indicator comparisons, “with vs. without measure X,” are inherently biased – this is the classic spurious correlation effect.

The regression analysis approach can help to determine which measures have which effect. In principle, they can do that. The effectiveness of the results is based on what the data set looks like and which methods of analysis you use. You should pose the following 5 questions before starting.

1. Can your Marketing Mix Modeling algorithm detect non-linear relationships and interactions between marketing channels?

Every marketing channel becomes saturated at some point. Further investment will not bear the same dividends as the initial investment. It is even often the case that after a certain point, additional expenditures can even have negative effects. This phenomenon is called non-linearities.

Furthermore, among media experts, it is almost general knowledge that marketing channels together have more of an effect than the sum of the individual effects. This phenomenon, which is used in management synergy, is called, “interactions” in statistical language. So, it’s surprising that most marketing mix models manage to get along completely without modeling these interactions.

2. Can your Marketing Mix Modeling algorithm find what is unknown up to this point?

In principle, with the help of econometric models, it has been possible to model non-linearity and interactions for a long time. The requirement is, however, that you have to know beforehand which non-linearities and interactions appear where. Then, the methods can determine its parameters.

Marketing mix models can have 100 variables. This means 10,000 possible interactions. On top of that, there are different kinds of interactions (AND, OR, XOR, etc.). That makes this especially time-consuming.

Furthermore, practical experience shows that the existing knowledge available is virtually never enough to stipulate the exact parameters for the econometric models.

That is precisely the large, practical advantage and the purpose of Universal Structural Modelling and NEUSREL. It doesn’t need any prior knowledge; it finds every arbitrary non-linearity and interaction independently. The result is a significantly higher explanation-power.

3. Do the Marketing Mix Modeling methods take indirect effects into account?

Regression and econometric models depict the relationship between driver variables and result variables. They measure direct effects exclusively, under the assumption that the drivers don’t mutually influence each other. An assumption which, in a digital age with many interactive touch points, can no longer be maintained.

In a marketing mix project for a mobile telecommunications provider, the Universal Structural Model and NEUSREL made it clear that after the television advertisement, people began to google the offer. This, in turn, led to significantly more Google AdWord clicks which, in turn, led to product sales.

TV is the driving cause. However, TV only had a very small direct effect on product sales. And this is only proven by regression analyses. Without taking the indirect effects into consideration, a very unrealistic view on TV’s effectiveness would arise.

4. Does your data set make the evaluation of long-term effects possible?

Marketing mix models count as methods that only measure short-term effects. The situation isn’t a question of analysis methods, but rather a question of the data used. Also, two steps also lead to making long-term effects measurable.

First, longer-term effects can be considered by using multiple target variables that differ on the time horizon.  So, not only the sales numbers of the next weeks are used, but rather also those in 2 weeks, one month or even in 6 months.

Secondly, market research-based variables such as brand preference can be included as intermediate sizes in the model. These steps allow you to measure which influence marketing channels have on brand preference. Also, they allow you to measure which influence the changes in brand preference have on sales numbers. The key feature is that the increase of the brand strength is a short-term, intermediate effect. But it also has a long-term impact because it is proven that brand attitudes are long-term values. With this background knowledge, the long-term effects are realized over a horizon of 1 to 2 years.

5. Does the Marketing Mix Modeling method make the simultaneous visualization of all relevant channels possible in a given number of cases?

Marketing mix modeling was simpler a few years ago. There were TV, radio, print, and billboards. Today there is AdWords, thousands of types of online banners, online affiliate ads, pre-rolls, and so on. (More examples: game commercials, press releases, inbound telephone contacts, shop visits, video-based outdoor formats, website traffic, social media engagement, mobile ads, company events, sponsoring, product launch specials, competitions spendings, etc. etc. etc.)

When we assume that these dozens of channels influence each other in their effects and build synergies, it will make sense to integrate them all into one model.

The problem:

The more variables are incorporated into a model; the more data points are necessary to validly specify the model. At best, however, data are mostly used every week, which leads to data sets that have merely 50 to 100 data points. In this number of cases, regression approaches can only tabulate 5-10 variables sensibly – if you want to receive robust results.

We have implemented a technology in NEUSREL that has already been used very successfully in genetics, among other fields. Here, data sets are processed that have significantly more variables (=chromosomes) than data sets (=test subjects). This is possible thanks to a methodology trick. The method doesn’t try to explain an outcome variable anymore, but rather many target variables simultaneously.

Sales numbers in various sales channels are an example. All these outcome variables have common drivers and causes. The method compresses the driver variables into a few proxies (so-called components).  The known Principle Components Analysis or Factor Analysis work differently. They create proxies that are in the position to account for the driver variables as far as possible. The new method creates the proxies in such a way that these can then optimally forecast the target variables. This is a fundamental and decisive difference.

The importance of the method is enormous. In the usual case numbers, it finally incorporates all of the relevant, influential factors. And this is precisely what delivers significantly more representative models. Additionally, there is a need to incorporate ever shorter timeframes into the analysis, especially when campaigns only run for a limited period.

Summarized …

the answer to these five questions gives the answer to which method should be used:

1. Can the method of analysis depict non-linear relationships and interactions between marketing channels?
2. Can the analysis methods of such non-linearities and interactions find what is unknown up to this point?
3. Do the analysis methods take indirect effects into account?
4. Does your data set make the evaluation of long-term effects possible?
5. Does the analysis method make possible the simultaneous visualization of all relevant channels in a few cases?

At the moment, there is only one methodology and one software that lives up to all the demands. It is the Universal Structure Modelling that is implemented in NEUSREL software.

p.s. More on NEUSREL and Universal Structural Modeling here.