Unveiling Marketing Mix Modelling: The Future of Marketing Measurement

Executives have always dreamed of a "single source of truth" to measure marketing performance and avoid wasting advertising budgets. Then came the wake-up call with iOS14, when Apple allowed users to disable tracking. In the post-iOS14 era, marketing attribution is about combining methods to calibrate your business model so you know which levers drive performance and make growth targets achievable.

In this article I'll share insights and experiences from my Marketing Mix Modeling Service.

Table of Contents

What is Marketing Mix Modeling?

Why Marketing Mix Modeling?

And what about Google Analytics 4?

The Challenges of Marketing Mix Modeling

What are the implementation options?

What the new MMM-SaaS solutions can do

Disadvantages of Lightweight MMM and Robyn MMM

Hybrid Attribution and Triangulation as Paths to the Future

Frequently Asked Questions about Marketing Mix Modeling

What is Marketing Mix Modelling?

Marketing mix modelling (MMM) allows you to measure the results of your marketing and advertising campaigns and determine how different channels contribute to your goal (e.g. sales) based on econometric models of your historical data.

Marketing mix modelling is an established approach that was used and optimised by media companies long before digital marketing. However, as you can guess, this was manual, time-consuming and therefore cost-intensive manual work. Even today, it is usually still agencies that collect data over long periods of time and convert it into models.

This breaks down the business metrics to distinguish between the contributions of marketing and promotional activities (incremental factors) and other factors (basic factors). These factors that influence the marketing mix can be defined as follows:

• Incremental factors: Business results achieved through marketing activities such as TV and print ads, digital editions, price discounts, promotions, etc.
  
  
• Base driver: The base result is achieved without any advertising. It is due to the brand equity built up over years. Base results are usually fixed unless there are economic or environmental changes. Seasonality and other non-market factors such as GDP, growth rate, consumer sentiment can also be subsumed under this.
  
  
• Other drivers: These are a sub-component of the base factors and are measured as the brand equity accumulated over a period of time due to the long-term impact of marketing activities.

Statistical analysis in media mix modelling uses multi-linear regression to determine the relationship between the dependent variable, e.g. sales or engagement, and the independent variables, e.g. advertising spend in the different channels.

For companies using marketing mix modelling, it is important to be critical in selecting the data they want to measure and the data they can measure. Data quality cannot be neglected, so companies need to spend time aggregating and cleaning data from internal databases, third-party sources or both. Media mix models often use two to three years of data, allowing for factors such as seasonality.

WHY Marketing Mix modeling?

A picture speaks more than a thousand words:

(Source: Sellforte.)

Marketing Mix Modelling therefore enables the following:

MMM can be used to identify the most appropriate marketing channel (e.g. TV, online, print, radio, etc.) to achieve marketing objectives and maximise returns.

2. Better implementation of advertising campaigns

MMM can be used to forecast optimal spending levels for the marketing channels used to avoid saturation.

3. Testing business scenarios

Mit Marketing Mix Modellierung können Geschäftskennzahlen auf der Grundlage geplanter Marketingaktivitäten prognostiziert und dann verschiedene Geschäftsszenarien simuliert werden, z. B. die Erhöhung der Ausgaben um 10 % in Zeitraum y, die Höhe der Ausgaben, die erforderlich sind, um eine 10-prozentige Steigerung der Umsätze zu erreichen usw.

Dazu kommt: Attributionsmodelle überschätzen die relative Wirksamkeit von Direct-Response-Kanälen und unterschätzen die Wirkung von markenbildenden Medien wie TV. Dabei ist TV noch immer das Medium, das die meisten inkrementellen Verkäufe generiert. Dennoch haben Attributionsmodelle Schwierigkeiten, die Auswirkungen von Medien wie TV zu messen.

Ein weiterer Grund, warum die digitale Zurechnung den ROI verfälscht, ist, dass sie kurzfristig ist. Eine Studie von Meta bestätigt, dass Werbeeffekte Monate oder Jahre andauern können, und der größte Teil des Paybacks kommt von diesen langfristigen Effekten. Attributionsmodelle blicken selten über ein paar Tage hinaus.

Dabei darf man nicht vergessen, die tatsächliche Definition von Attribution ist "Ursache und Wirkung". Sehr oft ist es der Fall, dass Kampagnenplaner einer Social-Media-Kampagne Credits zuweisen und dies als "Attribution" bezeichnen:

But this is only a very small part of the marketing mix, and true attribution is the incremental effect of:

• Market research
• Product development
• customer research and segmentation 
• mental and physical availability 
• and brand distinctiveness and differentiation

So it's really about more than lower-funnel sales and Google search clicks. And Marketing Mix Modelling is a step in that direction. MMM uses aggregated data and can therefore evaluate a larger number of channels, both traditional and digital. In addition, MMM allows you to take into account external influencing factors such as seasonality, promotions, holidays, etc.

I would be happy to help you create an MM model. Please feel free to rech out.

And what about Google Analytics 4 (GA4)?

Just like Apple, Meta and other ad networks, Google Analytics takes a walled-garden approach to its advertising business, meaning it provides analytics and optimisation tools for Google channels and only for those channels. Google will never recommend investing part of your budget in meta or vice versa, even if it would be better for your business.

Due to the latest privacy laws and the data breaches caused by adblockers (up to 30% data loss from this alone) and Apple's ATT, GA(4) relies on machine learning to compensate for the data loss. So-called data sampling is nothing more than an estimate, even if Google goes to great lengths to package it in a cool way.

At the end of the day, it's a web analytics tool, nothing more. And there may be better alternatives for this, such as Amplitude or TripleWhale.

The intricacies of marketing mix modelling

MMM does not only have advantages. One of the problems with marketing mix models is that they require a lot of data. This can be a problem if you are working with a small marketing budget or if your company has just started advertising campaigns on a larger scale. This is because most traditional marketing mix models require at best 2 years of historical data to make tractable predictions.

And even if a company invests heavily in marketing, it is often not so easy to collect all the aggregated data needed for MMM analysis. Data is often collected in different data silos. Only data-savvy companies have the right solutions, such as a marketing data hub and a data warehouse. But often a solution can be found after all.

What are the options for implementing MMM?

(Source: Analysts guide to MMM)

MMM-3rd Parties

There are third-party providers such as Nielsen that specialise in MMMs and usually do the whole job including data collection, modelling and recommendations. Most offer software into which an MMM can be loaded to enable predictive simulations/optimisations and forecasts.

Advantages: Provides an end-to-end service, requires the least time and effort (compared to other options), can compare your MMM results with benchmarks (depending on the provider);

Disadvantages: The provider charges a fee and is usually the most expensive option compared to other options.

Semi-automated MMM tools or software-as-a-service platforms (Saas)

These new entrants allow MMMs to be created in GUIs. This requires less in-depth statistical knowledge. MMM SaaS solutions offer automated modelling techniques that can be run continuously and typically include optimisation planning and simulation modules for predictive scenario planning. One provider in this area is Cassandra, a self-serve platform that aims to make MMM as easy as possible.

I have summarised how to create a marketing mix model with Cassandra in this article: Marketing Mix Modelling Made Easy: An Introduction to Cassandra.

Advantages: Suitable for those with less specialised MMM or analytical skills, cheaper alternative to working with an MMM provider;

Disadvantages: Less guidance than working with an MMM provider.

Development of an internal MMM solution

For companies that want to run and manage everything completely in-house, there is the option of developing their own MMM solution.

Advantages: Apart from staff,time and server costs, there are no ongoing costs for running MMM;

Disadvantages: Requires in-house analytical / data science skills, requires the most internal investment of time and resources of all the options.

Marketing Mix Modeling-as-a-Service

To make Marketing Mix Modeling as accessible as possible for companies, I also offer Marketing-Mix-Modelling-as-a-Service.

What the new MMM SaaS solutions can do

After Apple's iOS14 update, which drastically reduced the amount of data towards Meta, TikTok and co. with ATT (read more in this article on Apple's SkAdNetwork), Meta started working on Robyn MMM, an open source media mix modelling tool. Google also picked up the ball and developed Lightweight MMM.

Both libraries are open source, which has created a community that is constantly improving the libraries and adding new features.

Lightweight MMM uses the Bayesian algorithm and Numpyro as a backend and allows users to integrate their own data into the model. This makes the analysis more accurate.

One of the main advantages of Bayesian modelling in marketing applications is that it provides a flexible framework for dealing with uncertainty and variability in the data. By using probability distributions to represent uncertainty, Bayesian models can provide more accurate predictions and insights than traditional linear models.

Robyn MMM relies on R and Meta's Nevergrad Python Library, as does the Prophet Library. Prophet allows the model to select holidays by country and automatically determines trend and seasonality.

Nevergrad, in turn, runs a user-selected number of iterations (the documentation recommends no less than 2,000) that create multiple models. It then minimises the selected metrics and selects the Pareto-optimal models from these.

In addition, Robyn MMM uses hyperparameter optimisation, running the model thousands of times with different parameters to find the best parameter combination that gives the most accurate results.

As a central result, Robyn, like Lightweight, provides an optimal budget distribution. For this, it compares the historical budget allocation with the algorithm's recommendation and also shows the predicted changes in media impact after their application:

(Grafik von Analysts guide to MMM)

Disadvantages of Lightweight MMM and Robyn MMM

Die Vorteile im Einsatz der Bibliotheken und Regressionsfilter sind gleichzeitig die Nachteile. So sind die Variablen von Prophet (Saison, Trend, Feiertage) zwar effektive Annahmen, um den gesamten Prozess zu automatisieren, aber ohne manuelle Änderung ist es schwierig, aussagekräftige Schlussfolgerungen zu ziehen. Beim klassischen Ansatz der Marketing-Mix-Modellierung können wir untersuchen, welche Elemente der "Saison" sich auf die abhängige Variable auswirken, und welche genauen Feiertage sie positiv oder negativ beeinflussen.

Eine weitere Herausforderung sind die zugrunde liegenden Datenquellen. Wenn diese Daten unvollständig, ungenau oder inkonsistent sind, können die Modelle falsche oder irreführende Ergebnisse liefern. Hier ist es hilfreich, Beratung an der Hand zu haben, die diesen Prozess begleiten kann, um verlässliche Ergebnisse sicherzustellen.

Hybrid attribution and triangulation as future solutions

Hybrid attribution is, in my opinion, the future of marketing: you have to triangulate from multiple angles instead of relying on just one method.

This point is usually passed over by those not in the field, but most data protection changes don't really cause problems. The GDPR required a lot of legwork, but things continued to work afterwards. iOS14 was different, as it really changed everything and we had to start from scratch again for the most part, but SKAN is now okay. But what no one can afford to do is sit idle and wait, because with Privacy Sandbox, the same change that happened with iOS14 is coming to Android.

The key takeaway is that different attribution methods have different use cases. For example, SKAN is for real-time optimisation/granularity, MMM is for strategic budget allocation. None of them can do everything on their own.

(Source: Sustainable Measurement for Mobile Gaming)

Therefore, the future consists of the following methods:

1. multi-touch attribution: for decisions on a daily basis.

2. experiments such as conversion lift tests and geo-lift tests: validate causal relationships. 3. modelling: holistic view of marketing budgets and business drivers.

3. modelling: holistic view of marketing budgets and business drivers.

4. user surveys: The best way to find out early on who you are marketing to and how they became aware of your business.

Frequently asked questions about Marketing Mix Modelling

What is marketing mix modelling and what is it used for?

It uses statistical models and analytical methods to examine how sales respond to change. The goal of marketing mix modelling is to quantify the impact of each marketing activity on a company's sales success and thus make informed decisions regarding marketing budget and strategy. It answers, on an aggregate basis, the question of which factors have the greatest influence on sales success.

Marketing mix modelling can help in evaluating the effectiveness of advertising campaigns, setting marketing budgets and identifying growth opportunities. It is also useful for analysing competitors' marketing activities and finding out which factors have the greatest influence on sales success.


What data sources are needed for marketing mix modelling?

Various data sources are needed for marketing mix modelling, such as sales data, advertising spend, competitor data, channel data, market shares, prices, general economic data and other relevant variables. These data can come from various sources, such as internal company systems, market studies or publicly available data sources.

What kind of variables are considered in a marketing mix modelling?

As far as possible, all those variables that can influence the success of marketing activities are taken into account. These include, for example, advertising expenditure, price levels, competitive activities, distribution channels and seasonal trends.
How is the effectiveness of promotions measured with marketing mix modelling?

What techniques are used in marketing mix modelling?

Statistical analysis, regression, time series analysis and other data analysis techniques. The choice of techniques depends on the data available, the specific objectives of the analysis and other factors.

How is marketing mix modelling carried out?

First, the relevant variables are identified, then the data are collected, analysed and a model is created. This model is refined over time based on current data.

Modern SaaS solutions often work with assumptions to make these steps as simple as possible and reduce the time required.

What are the challenges of doing marketing mix modelling?

The availability and quality of the data, the complexity of the analysis, the identification of relevant variables and the validation of the results. Again, SaaS solutions help by working with assumptions.

Are there limits to marketing mix modelling?

Yes, there are. It is based solely on the available data and these may not be all the relevant factors that influence the success of marketing activities. Also, marketing mix modelling cannot predict the future, but is based on historical data. Unexpected events, such as Covid-19 or Russia's attack on Ukraine, can affect the outcome. Therefore, I recommend a hybrid approach together with multi-touch attribution, experiments like Geo-Lift and user surveys.