What is Causal Marketing Mix Modeling (MMM)?

Causal MMM, or Causal Marketing Mix Modeling, is a concept that has emerged in the most recent innovation wave around marketing measurement. People familiar with econometrics might raise an eyebrow for this term: aren’t all Marketing Mix Models trying to be “causal”, meaning that they attempt to reveal the causal effect between marketing activities and revenue of a business? Are Causal MMMs are more causal than other MMMs?

In this article, we bring clarity to what Causal MMMs are and how they differ from Traditional MMMs. We’ll cover:

• What is Marketing Mix Modeling?
• What is Causality?
• What is Causal Marketing Mix Modeling?
• What are the benefits of Causal MMMs to Marketers?

What is Marketing Mix Modeling?

To understand Causal MMMs, we first need to discuss Marketing Mix Modeling and its key characteristics.

Introduction to Marketing Mix Modeling

Marketing Mix Modeling (MMM) is a time-series method for estimating how each marketing activity contributes to driving a business outcome, typically revenue of a business. After understanding the revenue contribution of each marketing channel and campaign, MMM can calculate their ROI (Return on Investment), which allows marketers to track the effectiveness of their media spend. MMM also provides marketers with response curves , which explain how media ROI behaves at different spend levels, enabling marketers to optimize media budget allocation across different channels and campaigns.

Data used by MMMs: In estimating marketing’s contribution to revenue, Marketing Mix Models use time-series data on sales, marketing activities, promotions, and other factors influencing sales.

How does Marketing Mix Modeling work? Example

To understand how MMM works on a very basic level, let’s use the illustration below as an example.

Learning Meta effectiveness from the first sales peak: The model first notices an increase in revenue coinciding with Meta advertising. If there’s nothing else to explain the sales increase, the model assumes it was driven by Meta advertising and assigns the incremental revenue during the sales peak to Meta.

Learning Google effectiveness from the second sales peak: When there’s a second peak in sales, a higher one, the model notices that the business has also activated Google advertising simultaneously with Meta. When it already knows Meta effectiveness from the earlier data points, it can estimate the incremental sales impact of Google from the second peak.

The model observes numerous situations like this where there’s variance in sales and spend levels for different channels, and thus is able to provide sales contribution of all channels. For a more in-depth review into MMM, take a look our Marketing Mix Modeling Guide .

What is Causality?

Before we can discuss Causal MMMs, we also need to understand what Causality is and how it relates to MMM.

Introductionto Causality

Causality is the principle that one event (the cause) directly influences another event (the effect). In other words, if a change in one variable consistently leads to a change in another, we say there is a causal relationship. This concept is fundamental to science, philosophy, and everyday decision-making, as it allows us to understand the mechanisms behind various phenomena. For instance, when we observe that heating water causes it to boil, we establish a causal link between temperature increase and the phase change of water.

Causality vs. Correlation: Where things can go wrong

Causality is sometimes confused with correlation. Distinguishing causality from mere correlation is a critical challenge. Just because two events occur together does not mean one causes the other. A classic example is the correlation between ice cream sales and drowning incidents—both tend to rise in the summer, but one does not cause the other; instead, a third factor (hot weather) influences both. To establish true causation, researchers rely on methods like controlled experiments, natural experiments, and statistical techniques such as instrumental variables or causal inference models.

Identifying causal relationships is challenging because real-world data is often influenced by multiple factors. For example, a rapid increase in sales might coincide with an increase in digital ad spend, but this does not always necessarily mean that the ad spend caused all of the increase—there can also be seasonality, competitor activity, or broader economic conditions at play.

In the context of marketing, understanding causality is crucial because it helps businesses determine which marketing activities genuinely drive sales.

Why do most Marketing Mix Models Fail at Causality?

Now let’s get back to Marketing Mix Modeling. Are all Marketing Mix Models causal, because they aim explaining how media activities drive sales? Sadly, no.

I’ll be blunt: Large majority of MMMs are Traditional Marketing Mix Models, which generate unreliable ROI estimates. Traditional MMMs typically perform poorly at estimating the incremental sales contribution of each channel and campaign. There two reason this.

1) MMM datasets are challenging

Already in 2017, Google Research explored the challenges in MMM datasets in their nominal paper “Challenges and Opportunities in Media Mix Modeling”. You can read the full paper here , but I’ll raise a few challenges in marketing datasets for discussion here to illustrate the point.

Noisy data. MMM datasets are inherently noisy, increasing the risk of the model attributing random variation to marketing activities.

Multicollinearity. Many marketing activities are done simultaneously and some channels are always on. This leads to high levels of correlation between marketing variables, making it hard for Traditional MMMs to isolate causal effects of individual channels.

Indirect causal effects. Marketing variables have indirect causal effects in driving revenue, which are challenging for Traditional MMMs to account for. One example of this is branded search. Branded search affects revenue directly, but search volume on branded keywords can also be significantly affected by other media, especially awareness media like Facebook Awareness campaigns or TV. As a demand capture channel, branded search is also affected by seasonal demand fluctuations. All of this makes branded search highly correlated with sales, which is illustrated in the example below showing one year of daily data for sales vs. branded search spend (indexed). Correlation in this data is 62%. Traditional MMMs struggle to distinguish between the pure impact of branded search spend, the indirect effects of other media working through branded search, and other effects driving search volumes.

Omitted variable bias and lack of proper feature engineering. Including key factors, even basic ones like promotions, can be challenging to properly. Sometimes analysts building Traditional MMMs might use poor quality data or skip key variables completely if they are not available, making estimation of causal effects challenging.

Some of the the signs of low quality MMMs include for example instability in modelling results (ROIs wildly shift from one daily update to another, or are hard to reproduce even for the same data), and poor performance in output validation (e.g., validatio against experiments).

2) “Let’s just plug in the data and see what comes out”

The promise of MMMs is very attractive – media optimization decisions can drive massive value for a business. As an example, our research shows that eCommerce brands can drive +6.5% more revenue by optimizing spend allocation across campaigns, channels, and weeks with MMM. At the same time, building MMMs is very easy if you don’t hold a high bar for quality. Just take Google’s Meridian, Meta’s Robyn MMM, or PyMC, and plug in your data.

The attractive value promise and simplicity of getting started has drawn a lot of attention to MMM: Consultants providing MMM services in short projects, entrepreneurs building MMM solutions, and businesses testing MMM in-house. Concerningly, many of these operate with “let’s just plug in the data and see what comes out”-mentality. While this has created awareness for Marketing Mix Modeling, it has also led to a market with a many Traditional MMMs, that are unable to capture causal effects of marketing reliably. Using these models, it can be a coin flip whether MMM’s recommendation leads to a revenue increase or decrease.

What is Causal Marketing Mix Modeling?

Now that we have covered the context properly, let’s finally discuss Causal MMMs! Causal Marketing Mix Models, like Sellforte’s MMM, differ from Traditional MMMs by employing advanced methods that enable them to measure causal effects between marketing variables and sales more accurately. Let’s discuss how Causal MMMs do this.

Starting with Beliefs: Mapping Causal Relationships with Directed Acyclic Graphs (DAGs)

Directed Acyclic Graph (DAG) is a visual tool for causal inference that helps modelers understand causal relationships even before any modeling is done. Instead of giving the model some data and seeing what sticks, Causal MMMs use DAGs to map hypothesized causal relationships between different variables.

Causal DAG is constructed by identifying key variables of the model, such as spend for each channel (e.g., “branded search spend”), seasonality, and sales, and then establishing directional arrows between the variables to indicate causal relationships. DAGs are based on research, but also on deep marketing & modeling experience of the team building models.

Causal MMMs use DAGs as part of the modeling workflow. A classic example of a DAG-inspired insight is how Branded Search works, which we already discussed briefly in our previous chapter. Branded Search spend has a direct effect on driving sales. However, the amount of people using your brand keywords is influenced by awareness channels, such as TV or Facebook Awareness campaigns, which also have a direct influence on driving sales. The model needs to take these complex interactions into account. A modeler might start building a DAG around this, as shown below.

Incorporating Insights from Causal Experiments with Model Calibration

Experimentation is a key tool for scientists to estimate causal effects between variables. As an example, Randomized Control Trials (RCTs) are often touted as the golden standard in clinical research. While pure RCTs are not available in marketing, there are Experiment design options that marketers can use for causal inference in marketing, such as Geo Lift Test and Conversion Lift Tests. These Experiments help distinguish between the true incremental impact and correlations that may arise due to other factors.

Causal MMMs incorporate causal experiments into the model using a process called model calibration, which is described in detail in this post . Model calibration leverages the ability of Bayesian models to ingest external information in the form of prior distributions, or just priors. When priors are formed based on evidence outside the model, they become informative priors. As an example, using an experiment, one could estimate ROI of a channel to be between 4.0 and 5.0, instead of the model searching for the ROI from anywhere between 0 and 100. With narrow priors, Causal MMMs are able to provide radically more robust ROI estimates, compared to traditional MMMs, which operate purely on time-series data ignoring experimetns. Below is an illustration of informative prior distribution, in comparison to a non-informative prior distribution.

Combining Statistical Modeling Techniques with Output Validation

There’s an abundance of literature around statistical modelling techniques for causal inference (see e.g., Causal inference in economics and marketing by Varian) in marketing, so we won’t cover those in-depth here. In addition to having high quality modelling techniques, Causal MMMs have strong output validation practices. In a nutshell, these methods try to understand whether the modeling outputs make sense when compared to other data, even if the model statistically might seem ok. High marks on output validation tests indicate that the model succeeds in estimating the incremental impact of channels. Here’s a few examples:

Model stability checks. Level of stability in the model is a key indicator for model’s quality. When daily model updates are conducted, do the model’s estimates fluctuate wildly?

Validation against Incrementality Factor benchmarks. Next Gen MMM platforms show MMM-estimated ROI for each channel, and ROAS reported by Google Analytics 4 last-click and Ad platform attribution. This enables Incrementality Factor validation, i.e. are the ratios of MMM ROI vs GA4 last-click ROAS in the range of what you typically expect?

Validation against Experiments. While Experiments are useful in model calibration, they can also be used for model validation - do the results align with experiment results? As an example, Sellforte’s Geo Lift Experiment analysis module that compares MMM results to a Geo Lift Experiment you conducted.

Forecast accuracy. MMM provides a forecast for the total sales of the company (split into base sales forecast, promotion-driven sales forecast, and media-driven sales forecast). This forecast can be compared to actualized sales to estimate model’s performance in predicting the future. Just remember that individual datapoints with high or low forecast accuracy does not always imply high or low model quality. In predicting the future, there’s lots of events outside the model that can influence the actual sales, ranging from world politics to inventory shortages.

What Are the benefits of Causal MMMs to Marketers?

Causal MMMs have three benefits to marketers.

1. Causal MMMs help marketers understand the true impact of each marketing channel by isolating incremental effects rather than just correlations. Causal MMMs provide accurate measurement, compared to Traditional MMMs which often overestimate or underestimate the true contribution of each channel.

2. Better Budget Allocation and ROI Optimization. By quantifying the causal impact of each marketing activity, marketers can allocate budgets more effectively across channels. Instead of relying on historical correlations, causal MMMs allow for data-driven media mix decisions, ensuring that spend is directed toward channels that truly drive revenue growth. This is particularly useful in multi-channel and cross-channel environments, where interactions between different media need to be properly accounted for.

3. Campaign-level optimization. The largest value creation lever in Marketing Mix Modeling, Campaign budget optimization, is only possible with Next Gen Causal MMMs, which provide budget recommendation for each campaign. Check the full research how campaign optimization drives more sales in our research article .

Conclusions

Causal MMMs are radically more capable in estimating the true incremental sales impact of each channel, by employing methods that enable them to better account for causal effects between marketing variables and sales.

The benefits for the marketer are obvious: More accurate modelling results, which enable better optimization of media budget allocation.

Ready to try Sellforte’s Causal MMM? Apply for a Free Trial.