A common mistake in media mix modeling (MMM) is the belief that “more control variables are always better.”
This mindset comes from treating statistical models as causality machines. The thinking goes: If we just input everything, we’ll get perfect causal inferences on the other side.
That’s just not true.
In reality, overloading your model with control variables can create more problems than it solves. It can lead to overfitting, introduce instability, and make your model so complex that it’s impossible to interpret.
The goal of an MMM is not to estimate every possible factor that affects your business—that’s impossible. Instead, the focus should be on getting correct causal estimates for what you actually control – your marketing levers.
In this article, we’ll cover why adding too many control variables can hurt your model, the most common mistakes we’ve seen here, and how we approach seasonality, inflation, and macroeconomic factors.
Overloading Your MMM with Control Variables
More Variables ≠ More Accuracy
Many analysts were taught about omitted variable bias early in their careers but never learned about included variable bias –which is just as problematic.
Including excessive variables in your regression model creates several critical problems:
- Overfitting
Overfitting happens when a model is too complex relative to the amount of data available.
Instead of capturing meaningful patterns, the model memorizes the data –including noise, random fluctuations, and short-term anomalies that have no real predictive power.
We’re exaggerating but if you want to include “everything” in your MMM — from GDP growth to social media sentiment to the number of sunny days per month — the model might seem to have a high R-squared (which looks great), but it won’t identify true causal relationships.
It will give you a false sense of confidence and, when it looks at future data, it will break down.
- Multicollinearity
Multicollinearity happens when two or more independent variables in a model are highly correlated. This creates unstable parameter estimates – small changes in data can cause huge swings in predicted marketing effectiveness.
For example, let’s say you wanted to control for multiple economic indicators: consumer confidence, unemployment rate, GDP growth, and inflation.
Great… but these variables move together.
If consumer confidence drops, unemployment may rise. And then, because they are so correlated, the model will struggle to separate their individual impacts on sales.
The coefficient for one variable (e.g., GDP growth) might flip from positive to negative just by adding or removing another control variable—which is a clear sign of instability.
- Less Actionable
Beyond accuracy, your MMM must be actionable to deliver real business value. So when a model has too many variables, it becomes increasingly more difficult to understand what’s driving the results.
If your MMM has 30+ control variables, each interacting in complex ways, and your CMO asks, “why did our predicted TV ROI drop this quarter?”, your data science team won’t be able to explain it.
The model has now become a black box where small changes in multiple control variables are affecting the output in unpredictable ways. In the vast majority of cases, marketers don’t trust the results, and the MMM just gets ignored.
- More Analyst Bias
When analysts manually select control variables, they introduce bias into the model –often without realizing it (but sometimes they do realize, and still do it).
We’ve seen this before:
- Modelers test dozens of control variable combinations and pick the one that “looks best”.
- CMOs ask why the MMM doesn’t show what they expect (“I think TV is critical”) so analysts tweak the controls until the model aligns with leadership’s beliefs (and TV magically becomes incremental.)
Why Perfect Control Is Impossible (And How to Handle It)
The reality is there are infinite factors that impact business performance – most of which no one has reliable data for. If your ability to generate causal inferences depends on controlling for everything, you’ve already failed – because it’s impossible.
But there are a couple of variables that we want to put special emphasis on because they’re very common:
Seasonality (What Not to Do)
Many MMMs attempt to control for seasonality by estimating it separately and adding it as an independent variable. This approach makes intuitive sense because seasonality does affect demand.
But here’s the problem: seasonality doesn’t just impact baseline sales – it fundamentally alters marketing effectiveness across channels.
This misunderstanding leads to a dangerous scenario where marketers reduce spend during peak seasons (believing they’re over-investing) or increase spend during slow periods (thinking they’re under-investing). In reality, certain channels simply perform better at different times of the year.
The disconnect is clear: While marketers instinctively know to invest more when marketing is most effective, a poorly constructed model will paradoxically recommend the opposite strategy.
Inflation (What Not to Do)
Similarly, treating inflation as a straightforward control variable oversimplifies its complex impact on consumer behavior.
During inflationary periods, consumers become more price-sensitive, making them more responsive to promotions but less receptive to brand-building campaigns. Standard regression models miss these nuanced interactions and may incorrectly suggest reducing ad spend when performance appears to decline – when the real culprit is macroeconomic pressure on consumer spending.
How Recast’s Approach to MMM Avoids These Issues
1. A Time-Series Approach That Adapts Over Time
Rather than artificially removing seasonality, our time-series modeling approach allows the model to learn and adapt to historical patterns naturally. This helps distinguish between genuine shifts in marketing performance and predictable seasonal fluctuations.
The key advantage of time-series modeling is its flexibility – it doesn’t rely on static assumptions about seasonal trends. Instead, it continuously evaluates and refines its understanding as patterns evolve, eliminating the need for subjective manual adjustments that introduce bias.
For instance, if consumer behavior shifts and search ads begin performing better in Q3 rather than their traditional Q4 peak, our model recognizes and adapts to this change. Similarly, if summer TV campaigns that historically performed well begin to weaken, the model won’t cling to outdated assumptions.
2. Handling Macroeconomic Factors (Like Inflation)
We treat inflation as an interactive factor rather than a simple control variable, allowing our models to capture how inflation dynamically alters marketing effectiveness across channels.
This approach recognizes that during inflationary periods, tactics like price promotions often outperform brand campaigns. By modeling these channel-specific effects, we provide more accurate and actionable insights than models that apply inflation as a uniform adjustment.
3. Reducing Analyst Bias and Increasing Model Reliability
As we mentioned, any time you have an analyst interact manually with the model to choose (or to drop) a control variable, you risk injecting bias or tweaking the model to fit expectations.
That’s why our models have a fully automated variable selection – no cherry-picking, no individual decisions, and no need for ongoing model adjustments.
TL;DR – Are Too Many Control Variables Damaging Your Media Mix Model?
- More control variables don’t mean better results – they often introduce bias, overfitting, and make models harder to interpret.
- Seasonality and macroeconomic factors shouldn’t be manually controlled.
- Recast models inflation and external forces as interactive effects to prevent misattribution errors, and automates control variable selection to stop analyst bias.
