Bayesian MMM: a prior is not an identification strategy

In a July 2025 EMARKETER and TransUnion survey of 196 US marketing professionals, 46.9% planned to invest in marketing-mix modeling within the year, and 27.6% ranked MMM the most reliable measurement methodology available to them. The category’s vocabulary has shifted with the money: Lifesight, Sellforte, Measured and Haus now sell products literally named “Causal MMM,” and Google’s Meridian documentation calls MMM “a causal inference tool.”

Much of that confidence leans on one word: Bayesian. In that vocabulary, a prior is the ROI the modeler believes before seeing the data, and the posterior is that belief updated by the data. The reasoning goes that priors plus posteriors equal principled inference, therefore the ROI numbers are causal. I have built a marketing-mix model inside a commercial program and carried its outputs into budget meetings, so I have an interest in MMM being taken seriously. That is why the conflation bothers me. Bayesian machinery buys regularization, a way of keeping estimates within reason when the data are noisy, and honest uncertainty. Identification — the license to read a coefficient as “what would happen if we changed spend” — has to come from somewhere else, and a prior cannot supply it.

Why is “causal MMM” suddenly everywhere?

Because the alternative collapsed, and a reassuring label was available. Multi-touch attribution depended on user-level tracking that privacy regulation, Apple’s ATT and walled gardens have been dismantling for years. Marketers needed a measurement method that runs on aggregate data, and MMM was waiting with four decades of history and a new open-source toolchain.

The irony is that the third-party cookie itself survived. Google cancelled third-party-cookie deprecation in July 2024 and retired most Privacy Sandbox APIs in October 2025, citing low adoption. The renaissance kept going anyway, because the real driver was broader signal loss. The IAB’s State of Data 2026 report describes measurement systems “under strain” from “privacy regulation, signal loss, platform-embedded optimization, and fragmented data environments.”

46.9%

Into that demand walked the word “causal.” Measured’s FAQ compresses the sales pitch into one line: traditional MMM is correlation, “Causal MMM = Causation (proof of what truly drives growth).” Haus launched a product named Causal MMM in October 2025. And the toolmaker itself sets the tone: Meridian’s docs state that “all of the quantities that MMM estimates imply causality.” The label is doing commercial work. The question is what earns it.

What does Bayesian machinery actually buy you?

Two real things: regularization and honest uncertainty. An MMM asks a lot from a little data, and informative priors keep noisy, collinear channels (channels whose spend moves together) from producing nonsense. Recast, a Bayesian MMM vendor, is candid about the mechanics: “Without constraints, a model can produce negative ROIs or absurdly high ROIs just because the data are noisy or channels are collinear.” Posterior distributions, in turn, replace a false point estimate with a stated range. Both are genuine engineering advantages.

Notice, though, how Google’s own rationale page describes the role of priors: they “offer an intuitive way to regularize the fit of each parameter according to prior knowledge and the selected regularization strength.” Regularize. The causal weight in Meridian’s methodology sits elsewhere: on an assumed causal graph “constructed based on expert domain knowledge.” A causal graph is the modeler’s map of what drives what; the model inherits whatever that map gets wrong. The prior shapes the estimate. Whether that estimate means anything causal is decided upstream, by whether the assumed graph holds.

One more inconvenient detail for the “Bayesian, therefore causal” story: Meta’s Robyn, the other default open-source MMM, is not Bayesian at all. Under the hood it runs ridge regression with evolutionary hyperparameter search. The causal glow around these tools is a habit of the category, and nothing in the math produces it.

What is identification, and why can’t a prior supply it?

Identification is the question of whether your data and design could pin down the causal quantity at all — even with infinite data. Estimation asks how precisely you have measured a parameter; identification asks whether the thing you measured is the thing you claim. More data narrows the posterior without repairing a design that cannot distinguish “spend drove sales” from “sales forecasts drove spend.” A prior enters at estimation, downstream of that question; nothing downstream repairs a design.

The statistician David Freedman put the point plainly: “Data analysis on X and Y cannot decide whether the causal interpretation is viable. Instead, to make causal inferences from a system of regression equations, causation is assumed from the beginning. As Cartwright (1989) says, ‘No causes in, no causes out.’”

MMM is the textbook case. Google’s own researchers wrote in 2017 that typical mix models “produce correlational, not causal results. It is only under certain narrow conditions that these estimates can be considered causal.” The same paper counts what the models lean on: about 156 weekly data points, asked to support 20-plus channels with several parameters each. Google’s 2024 calibration paper concedes the deeper point: “the causal interpretation in MMMs relies on the uncheckable assumption of conditional exchangeability.” In plain terms: the model assumes that, given the controls it includes, spend varied as if assigned at random. Nothing in the data can test that assumption; it has to hold by design. A prior is one more assumption stacked on that uncheckable one — and it moves the posterior most where the data are too weak to overrule it. That weak-data region is where the causal question lives. In budget terms: where the data cannot answer, the modeler’s starting opinion becomes the ROI number.

Can the failure hide behind a good fit?

Yes, and there is a published demonstration. In a simulation study published by PyMC Labs, Juan Camilo Orduz built a Bayesian MMM with sensible priors on data containing one unobserved confounder: a variable that moves both spend and sales but is missing from the model. The model converged, fit the data and confidently reversed the true ranking of the two channels, estimating the weaker channel at roughly twice its true ROAS (return on ad spend).

A lift test is a controlled experiment that measures one channel’s incremental effect directly. Adding two per channel to the likelihood recovered the true values. Nothing inside the model signaled the problem; the experiment did.

Tool design can make this worse while looking like prudence. Robyn scores candidate models partly on DECOMP.RSSD, a criterion that favors models where each channel’s estimated share of the effect stays close to its share of the spend. The companion paper is frank that “business plausibility is a criterion used in model evaluation as much as or more than statistical fit.” That criterion rewards the model for agreeing with last year’s budget — the very allocation it was hired to interrogate.

Each piece of evidence an MMM can produce answers a narrower question than the one the budget meeting is asking:

The evidence	What it shows	What it cannot show
In-sample fit (R², NRMSE)	The model reproduces the past	That the decomposition is causal
Out-of-sample forecast accuracy	Prediction holds while spend and market stay in the same regime	What happens when you change spend
Posterior credible intervals	Uncertainty inside the assumed model	That the assumed model is right
Calibration against a lift test	The channel’s effect where it was tested	Channels and spend ranges no experiment touched

Where does causality actually enter an MMM?

Through a designed experiment, at calibration: the step where an experiment’s result is written into the model as a prior. Geo experiments, which randomize ad spend across geographic regions, create the exogenous variation the observational data lacks: spend that changes for a reason unrelated to demand. Google researchers formalized the method in 2011. Modern tooling then carries that evidence into the model. Meridian’s docs offer it as “one common approach”: to use “an experiment’s point estimate as the prior mean and its standard error as the prior standard deviation.” The experiment’s result becomes the model’s starting belief for that channel.

Robyn adds experiment error to its optimization objectives, and PyMC-Marketing ships lift-test calibration as a named feature. The Robyn paper goes furthest, claiming that “calibration serves as a method of identification.”

The Robyn claim is right in one direction only. A prior that carries an experiment imports evidence; a prior that carries opinion imports opinion, with the same mathematical dignity. This is the triangulation hierarchy in its proper order: the experiment is the ground truth, and its result becomes the prior that calibrates the mix model.

Calibration has honest limits, and the primary sources admit them. Meridian’s docs note that “the ROI measured by an experiment rarely aligns perfectly with the ROI measured by MMM,” because a weeks-long test and a model crediting months of carryover measure two different quantities. One experiment anchors one channel at one point on its response curve. It certifies the return at that spend level and says nothing about the return at double or half of it. Aryma Labs, the sharpest critics on this point, go further. Geo tests, they argue, should not set priors at all: “Your experiments may be univariable but MMM is not.” An experiment reads one channel in isolation; the model estimates every channel jointly, and those are not the same quantity. You do not need to settle that dispute to accept its premise: the causal content of a calibrated MMM is exactly the causal content of the experiments behind it, degraded by every mismatch between the two.

What did this split look like in a real program?

I ran both instruments side by side, the mix model and a control-cohort read, and they earned different kinds of trust. On a program I led at a top-5 pharma in the CIS, a marketing-mix “Sales-Impact Prediction” model decomposed sales by channel, forecast the impact of a candidate mix and computed ROI. Its forecast tracked fact, which earned it a place in planning. That is a prediction claim, and we kept it labeled as one.

The causal headline came from a different design. The program’s outcome was read against a control cohort: +7% incremental Rx, visited-versus-not-visited physicians, indexed 100 → 107; my role: AI lead and builder. A visited-vs-not-visited comparison is a designed read, not a randomized experiment, so the number holds for that cohort and that design, nothing wider. The discipline was refusing to let the MMM’s decomposition borrow the control read’s credibility, or the reverse.

A forecast that tracks fact earns planning trust. A causal claim has to be earned separately, by a design — and it stays exactly as wide as that design.

What does an auditable MMM show the CFO?

Everything a bank’s model-risk function would ask of a credit model. The Federal Reserve’s SR 11-7 guidance defines a model as any quantitative approach that “applies statistical, economic, financial, or mathematical theories, techniques, and assumptions to process input data into quantitative estimates.” An MMM fits that definition without a squeeze.

The same guidance asks for “effective challenge”: “critical analysis by objective, informed parties who can identify model limitations and assumptions and produce appropriate changes.” The test is independence plus authority: reviewers with no stake in the answer and the standing to force changes. Marketing reallocates real budgets on MMM outputs, and the IAB’s December 2025 best-practices guide already talks to finance in those terms, asking modelers for “CFO-readable, P&L-linked recommendations with confidence bands” built on auditable inputs and testable forecasts.

Concretely, an auditable MMM is a package the reviewer can attack:

1. Priors with provenance. Every informative prior traces to an experiment, a documented expert elicitation or a named judgment call, dated so the next refresh can be challenged against it.
2. Assumptions on the record. The causal structure the model assumes: which confounders are controlled and why, plus sensitivity of headline ROIs to adstock and saturation choices (how long an ad’s effect carries over, and where extra spend stops paying back).
3. Calibration evidence per channel. Which channels have experimental anchors matched on KPI, geography and time window, plus an explicit no-experiment flag on the channels that have none.
4. Outcomes analysis. Out-of-sample forecast tracking as ongoing monitoring, logged, with fit treated as necessary and never sufficient.
5. Effective challenge. Review by someone not paid for the model’s conclusions; SR 11-7 is blunt that a model’s developer “cannot be relied on as an objective or sole source” on model quality.

This is the same instinct as governance that clears procurement: the defensible position is an inventory you can put in front of a reviewer. The audit trail behind a decomposition chart is what turns it from a claim into evidence.

Isn’t Bayesian uncertainty already honesty?

The credible interval, the range a Bayesian model puts around its estimate, is honest about the wrong thing. It quantifies uncertainty inside the assumed model and stays silent on whether the assumptions hold. In the Orduz simulation, the reversed channel ranking came with perfectly well-behaved posteriors. An interval on an unidentified parameter is a precise statement about a quantity you cannot name.

The budget objection deserves a straight answer too. Lift tests cost money. Analytics practitioner Tim Hiebenthal states the hard line: “If you cannot run experiments to validate your MMM, you should not try to build one in the first place.” My version is softer but not by much: run the model, flag every channel that has no experimental anchor, and route the biggest reallocation decisions through the next test before you act on them. What you may not do is let unanchored posteriors wear the word causal into a budget meeting.

Michael Kaminsky of Recast puts the reviewer’s posture well: “The default assumption should be that MMMs in general are wrong.” As a starting point for effective challenge, that is model risk management doing its job.

Deciding what a measurement model is allowed to claim is operating-model work — the same discipline that kept the forecast and the causal read separate on the program measured against control. If your MMM’s next number is headed for a budget meeting, let’s pressure-test it against this checklist.