How does MMM handle correlation?

A rooster crows every morning just before the sun comes up. If you kept a log for a year, the pattern would be nearly perfect: crow, then sunrise, day after day. Nobody concludes the rooster raises the sun. The rooster is reacting to the same early light that comes before sunrise, so the two line up without one causing the other.

Marketing budgets behave a lot like that rooster. Brands raise spend right before and during the moments customers were already going to buy, so spend and sales rise together almost every time, bringing up big questions about correlation in marketing data. How does an MMM handle that correlation? An MMM starts from those patterns, then tries to separate how much of the sales increase came from marketing and how much came from everything else moving at the same time: seasonality, promotions, pricing, and other channels. How well it does that depends on how the model is built. Traditional MMMs struggle most when spend climbs alongside natural demand or when several channels scale together, and those are exactly the moments brands care about most.

That matters because the biggest budget calls of the year happen in those windows. If a model can't tell holiday shoppers apart from ad performance, its advice for Black Friday will be built on credit the ads never earned.

Key takeaways

• An MMM starts with correlation, the patterns between marketing spend and sales, and then tries to work out how much of the sales change marketing actually produced.
• Correlation is especially misleading in marketing because brands deliberately raise spend when demand is already rising, so spend and sales climb together even when ads aren't the reason.
• Traditional MMMs treat sales as a baseline plus separate channel contributions, and when spend tracks demand, many different splits fit the data equally well.
• Channels that always scale together, and channels that feed each other, like YouTube driving branded searches, are two more places traditional models misassign credit.
• Statistical penalties and extra history make a traditional model's answer steadier, but they don't make it more correct.
• Prescient's MMM builds how marketing works into the model, including funnel roles, timing, and shifting efficiency, which gives it more to go on than the correlations alone.
• No MMM proves cause and effect from observational data by itself, so the strongest confidence comes from combining the model with validation, stability checks, and test results.

Why correlation misleads marketers so easily

Correlation misleads marketers because almost nothing about a marketing plan is random. Budgets go up around launches, sales, and holidays because that's when they pay off, so high spend and high sales arrive together by design.

Take a brand that raises its Meta spend by half in November and sees revenue climb 35%. It's tempting to credit Meta for a large share of that jump. But a lot else changed that month, too:

• Holiday shoppers started buying more on their own.
• A Black Friday sale cut prices sitewide.
• Google search spend went up.
• Email sends became more frequent.
• Stores carrying the product ran their own deals.

Every one of those could explain part of the revenue lift, and they all happened in the same weeks. The Meta correlation is real, but it doesn't tell you how much revenue Meta actually added.

There's a second wrinkle: some of the relationship runs backward. The brand raised Meta spend because it expected November demand, so expected sales drove the spend at least as much as the spend drove sales. Statisticians call this endogeneity, and it's one of the main reasons marketing data is harder to read than it looks.

How MMMs try to separate marketing from everything else

MMMs try to separate marketing's contribution by looking at many factors at once, rather than comparing one channel's spend with revenue in isolation. Most use four main techniques.

Accounting for other things that move sales

An MMM includes factors that can change revenue on their own, so it doesn't hand their effect to advertising. The most common ones are:

This helps only to the extent the model includes the right factors and represents them correctly. If something drives both spend and sales and the model leaves it out, its effect can still end up credited to marketing.

Looking at how sales respond when spend changes

Instead of relying on the fact that big spend and big sales happen together, an MMM looks at whether changes in one channel's spend line up with changes in sales after the other factors are accounted for. That only works if channels move somewhat independently. When every budget rises and falls in lockstep, there's very little information about any one of them.

Allowing for delayed effects

Ads often pay off later. Someone sees a YouTube ad today, looks the brand up tomorrow, and buys two weeks later. MMMs account for that carryover so the model doesn't miss real effects just because they didn't land on the same day as the spend.

Allowing for diminishing returns

MMMs also allow for the possibility that each added dollar in a channel does less than the one before. That helps the model describe spend and sales more realistically, though neither carryover nor diminishing returns on its own settles which channel drove what.

Where traditional MMMs break down

Traditional MMMs break down under correlation in three situations: spend that rises with demand, channels that move together, and channels that feed one another. In each one, the data can't distinguish between explanations that the model has no structure to choose between.

When spend rises with natural demand

In most traditional MMMs, revenue is written as a sum: baseline demand, then one separate piece per channel. When marketing spend climbs exactly when baseline demand climbs, the data supports many different splits between the two. More credit to the holiday and less to ads fits about as well as the reverse. The data alone can't pick, so the model's built-in assumptions do.

When those assumptions push holiday demand into marketing's column, the result is called baseline leakage. Ads look far more efficient during peaks than they are, and recommendations follow. In Prescient's simulation research, two common open-source MMM approaches often overstated holiday effects by 20% to 30% relative to the known true values, while our model’s baseline estimates stayed close to them.

When channels scale together

Suppose a brand always sets its TikTok budget at half its Meta budget. Some weeks Meta gets $10,000 and TikTok $5,000, other weeks Meta gets $24,000 and TikTok $12,000, but the ratio never changes. Revenue rises in the bigger weeks. How much of that came from TikTok?

From the spend data alone, there's no way to know. The two channels carry exactly the same pattern, so any split of credit between them fits equally well. Real budgets are rarely that perfectly locked together, but they're often close, especially during promotions when everything scales at once. This is called multicollinearity, and it's why traditional models can swing big chunks of credit from channel to channel between runs.

It also explains why measuring more granularly is a double-edged sword. Breaking channels into campaigns gives the model more detail to work with, but it also gives it more pieces to tell apart. Without enough independent movement in the data, or enough structure in the model, extra granularity adds uncertainty instead of clarity.

When one channel creates the demand another channel captures

Branded search is the clearest example. Someone watches a YouTube ad, searches your brand name a few days later, and buys. Branded search and sales are strongly linked, but the search happened because of the YouTube ad.

That creates a trap with no good answer inside a traditional model. If branded search is treated as just another independent channel, it absorbs credit that belongs to YouTube. If the volume of brand-name searches gets treated as a background factor, the model subtracts away part of YouTube's effect along with it. In statistical terms, branded search is a mediator for YouTube: a path its effect travels through. A model that can't represent that path will misjudge one channel or the other.

Why penalties and more data don't fix it

When a traditional model can't decide between equally good explanations, a common fix is to add statistical penalties or starting assumptions that steady the answer. That does make results more stable, but stability isn't accuracy. The penalty simply picks one of the many explanations that fit equally well, and a different penalty would pick a different one.

\

More history doesn't rescue it either. If spend keeps rising with demand in the new data, the same ambiguity is still there. The model gets more confident, but more confident in whichever split its assumptions favor, not in the right one.

Starting assumptions aren't bad in themselves. Prescient's MMM uses them too, drawing on benchmarks from other brands to give low-data channels a realistic starting point. The problem is relying on them to resolve ambiguity that the model's structure can't.

How Prescient handles correlation

Prescient's MMM handles correlation by giving the model a realistic picture of how marketing works, so it has more to go on than which numbers moved together. Rather than slicing revenue into independent pieces, it treats marketing and demand as one system that changes over time.

Demand and marketing are modeled together

Under the hood, our model follows a few things you can't observe directly—how aware people are of your brand, how ready they are to buy, and how hard each ad dollar is working this week—and treats sales as what those conditions produce. Seasonal swings and marketing both act on the same underlying conditions, so there's no forced, after-the-fact division between "baseline" and "media."

Funnel roles give the model a direction to follow

Prescient’s MMM represents the difference between channels that create demand, such as YouTube, CTV, and prospecting, and channels that capture it, such as branded search and retargeting. That's what handles the branded search trap. Branded search is understood as a conversion point, and the demand it converts can be traced back to the channels that built it. Prescient measures that downstream credit as halo effects, sales a channel drives through another channel.

Timing and efficiency add more clues

Channels that launch together often fade at different speeds. A promotion's effect disappears in days while a brand campaign keeps working for weeks, and our model learns those patterns for each tactic. When ads perform better in a holiday week, Prescient’s MMM reads that as ads working harder in that context instead of as more baseline demand, which keeps peak-season gains attributed to the right source.

Models are chosen for stability, not just fit

Matching last year's revenue closely says little about whether credit was divided correctly, since many splits produce the same total. So Prescient also scores models on how they predict periods held back from training, whether results hold when trend and seasonality settings change slightly, and whether one part of the model is soaking up credit that belongs to another. Results that swing under small changes are treated as a warning sign, not an answer.

Test results are added as evidence

When a brand runs an incrementality test, Prescient can add the result to its MMM as a prior and show model accuracy with and without it. That lets the test inform the model without forcing every other estimate to bend around one region and one moment in time. The brand decides whether the test stays in.

What the peak-season evidence shows

Peak periods are where correlation is strongest, so they're the hardest test of how a model handles it. Prescient tested this with synthetic data where the true answers were known in advance, putting our model up against a regression-based and a Bayesian open-source approach, including a Black Friday and Cyber Monday-style peak.

• Budget advice at peak: The regression-based approach recommended about 81% too much peak spend after crediting holiday demand to ads. The Bayesian approach recommended about 11.5% too little after underestimating how far marketing could scale. Our model's recommendation was off by roughly 1%.
• Value given up: Following the two open-source approaches' full recommendations surrendered roughly 32% and 45% of the achievable result, compared with about 6% for Prescient’s MMM.
• Predicting peak lift: Our model's error in predicting each channel's extra peak-season sales was about half of one open-source approach's and roughly a quarter of the other's.

The common thread is that both open-source approaches fit past sales well. Their mistakes came from how they divided credit when spend and demand moved together, not from a poor fit.

Does MMM prove cause and effect?

Not on its own. MMM is designed to estimate how much marketing contributes to sales, but it works from observational data, so its estimates depend on assumptions that can't be fully verified from that data. A model can predict revenue very well and still be wrong about which channels produced it.

That doesn't make MMM unreliable. It means confidence should come from several kinds of evidence working together, each with its own limits.

Questions to ask an MMM provider about correlation

Asking about correlation is one of the fastest ways to tell how a provider's model will behave when it matters. These questions, and what a strong answer covers, are a good place to start.

Where Prescient comes in

Peak seasons, launches, and big multichannel pushes are the moments Prescient was designed around, because that's when coincident spend does the most damage to measurement. One model covers an omnichannel brand's website, Amazon, and retail sales, and it reports Modeled ROAS for each campaign along with halo effects and retail halo effects, refreshed every day.

Media Forecaster carries that same read into planning, so your Q4 budget is built on what your ads add above the holiday lift instead of on the lift itself. If you'd like to see how your own channels separate out, book a demo.

FAQs

Does marketing mix modeling prove causation?

No model proves causation from observational data alone. A marketing mix model estimates how much each channel contributes to sales, and those estimates depend on the model's structure and assumptions. Confidence comes from pairing the MMM with testing on unseen data, stability checks, and, where available, incrementality test results.

What is multicollinearity in MMM?

Multicollinearity is when two or more marketing inputs move together so closely that a model can't tell their effects apart. It's common when brands scale several channels at once for promotions. Traditional MMMs often respond by shifting credit unpredictably between those channels.

Can an MMM separate channels that always run at the same time?

Only with extra information. If two channels always move in exactly the same pattern, spend data alone can't separate them. A model can still tell them apart when they differ in other ways, such as their role in the funnel or how quickly their effect fades. Otherwise, it's more honest to report them together.

What is baseline leakage?

Baseline leakage is when an MMM credits marketing with sales that would have happened anyway, usually during seasonal peaks. It happens because budgets climb in the same weeks shoppers were already buying more, and traditional models can't reliably split the two. The result is inflated efficiency and recommendations to overspend at peak.

Do incrementality tests solve the correlation problem?

They help, but they don't solve it. A test measures one change in one place at one time, which can strengthen a model's estimate for that channel. It can't fix how the model handles correlation everywhere else, and its result may not hold in a different season or at a different budget.