Marketing Mix Modeling in Pharma: Optimising Spend Across Channels

What is marketing mix modelling, and what does it actually answer?

Marketing mix modelling — sometimes written as MMX modelling — is a statistical method that relates an outcome, usually sales or prescriptions, to the marketing activity that preceded it, using aggregate data over time and usually across geographies. It produces an estimated contribution and return for each channel, and from those a recommended budget split.

The most useful thing to establish first is which question it answers, because a great deal of disappointment with MMM comes from asking it a question it was never built for.

The architecture that actually works, and why it is rarely built

Experiments calibrate the model; the model allocates the budget; attribution runs the week. Each does what the others cannot.

The practice-oriented literature on modern MMM is explicit about the first half of that: results should be validated against experimental methods — field experiments, advertising experiments, or lower-level attribution results — rather than accepted on statistical fit alone. The three leading open-source tools all provide a mechanism for it, whether through calibration in the optimisation or through Bayesian priors.

It is rarely built because the three live in different teams on different clocks. Experiments belong to whoever will run a controlled test; the model belongs to analytics; attribution belongs to digital marketing — and the calibration step requires the first to finish before the second is refreshed.

The practical version for a commercial team: run one experiment a year on the largest channel, and use it to set the prior. That single connection converts a correlational model into a partially calibrated one, and it is achievable without reorganising anything.

How MMM works: adstock, saturation and the shape of the answer

Two mechanics carry most of the method, and both are commonly described loosely enough to be misused.

Adstock — the effect does not land on the day you spend

Advertising and promotional effect persists and decays rather than arriving and ending. The concept, coined by Simon Broadbent, is expressed in its simple form as a decay model in which current adstock equals a decay weight multiplied by the previous period's adstock, plus this period's activity. The parameter that matters is the half-life — the time for the effect to fall to half its level.

The divergence is the finding. A threefold disagreement between academic and practitioner estimates of the central parameter means the number cannot be borrowed with confidence — and in pharma the case for borrowing is weaker still, because a detail visit to a specialist is not a television impression on a shopper. Estimate adstock from your own data, present it as a range, and treat any vendor quoting a single confident half-life for your therapy area as having chosen a number rather than measured one.

Saturation — each additional unit does less

The second mechanic is diminishing returns: beyond a threshold, additional exposure produces progressively smaller increases in response, which is why adstock is usually transformed into a non-linear shape such as a logistic or negative exponential curve before it enters the model. This is the same phenomenon the call-planning response curve describes, arriving through a different door — and it carries the same caution, that the estimate is least reliable at the high-spend end where the data is thinnest.

Why MMM estimates are usually too generous

This is the section most published MMM content omits, and it is the difference between a model that informs a budget and one that flatters last year's decisions.

The variation problem, which is the most practical constraint on this page

A model cannot estimate the effect of something that never changed. If a channel received roughly the same budget in every territory in every period, its coefficient is not measured — it is assumed, borrowed from a prior, or absorbed into the baseline.

This is the reverse of how most teams think about MMM. The instinct is that more history produces a better model. In fact more of the same history produces a more confident version of the same non-answer, because the information in the data comes from variation, not from volume.

The operational implication is uncomfortable and cheap: to measure a channel, you have to be willing to change it. Different budget levels across matched territories, or a deliberate step change in one period, generate more usable information in two quarters than three years of stable spending. A commercial team that will not vary anything is asking for a model of a system it has kept deliberately still.

Who built the tools, and why that matters

Open-source MMM has made the method accessible to organisations that could never have commissioned it, and that is a genuine advance. It also carries a structural fact worth stating plainly, because almost no published guidance does.

Two of the three most-used tools for deciding how to divide an advertising budget were built by two of the largest sellers of advertising. The concerns raised about this are specific rather than insinuating, and they are worth knowing before defaults are accepted.

• Default settings carry influence. As one industry chief executive put it, there is a great deal of power in setting where everybody starts in a solution — and many advertisers run the default specification without customising it.
• Data access is asymmetric. Meridian is deeply integrated with its parent's ecosystem and can draw on platform data that competing channels cannot supply on equal terms.
• Adoption has been commercially incentivised. Reporting indicates sales targets were tied to Meridian adoption — meaning the entity being measured had a commercial interest in distributing the measurement tool.
• The defence is reasonable and partial. Open-sourcing does let practitioners inspect the method, and MMM is an improvement on click-based attribution regardless of who publishes it. The point is not that these tools are unusable — it is that defaults are decisions.

What a pharma team should do about this

The concern is mostly academic for a pharmaceutical brand whose largest channel is a field force rather than paid search, which is a genuine advantage of this category.

But the general lesson transfers directly, and it is the one worth carrying: whoever supplies the model has an interest in its priors. That applies with equal force to a commercial vendor whose platform is one of the channels being evaluated, and pharma has several of those.

Three practical safeguards. Ask what the priors are and who set them. Ask whether the channel with the most favourable estimated return happens to be the one the supplier sells. And calibrate at least one channel with an experiment you designed yourself — an independent anchor is worth more than an independent auditor.

Why MMM suits India better than the methods usually imported with it

This inverts the usual framing, and the inversion holds. Most measurement guidance reaching Indian commercial teams assumes prescriber-level prescription data — deciles, individual response models, patient-level attribution. India's dominant commercial data infrastructure operates at distributor, stockist and chemist level rather than at prescriber level, so those methods arrive already broken and teams conclude that rigorous measurement is not available to them.
Marketing mix modelling does not need prescriber-level data. It works on aggregate outcomes by geography and time period — which is exactly the shape of the data Indian pharmaceutical companies actually hold, in the form of secondary sales by territory. The method that survives the Indian data environment best is the one least often attempted here, because it arrived packaged with an American framing that treats measurement as an individual-level problem.

Two cautions keep this honest. Secondary sales are not prescriptions — they measure movement into the trade rather than out of it, and stocking behaviour introduces noise that a monthly model will read as demand. And territories are not randomly assigned: the best representatives and the largest budgets go to the best territories, which is confounding of exactly the kind described above. Both are manageable with design — longer periods to smooth stocking, and territory characteristics as controls — and neither is a reason to prefer a method whose data does not exist.

A worked example: reading an MMM output honestly

Model outputs are usually presented as a table of returns by channel and a recommended reallocation. The following works through an illustrative output for a mid-size brand and shows which numbers can carry a decision. The figures demonstrate the reading, not a real model.

What this output actually licenses, and what it does not

One channel supports a confident decision. One supports a directional one. Three support none.

The naive reading takes the ranking at face value, cuts the lowest-returning channel and moves its budget to the highest. In this example that decision is made almost entirely on the two weakest estimates in the table — a channel that could not be separated from another, and one whose coefficient was never measured because its funding never varied.

The disciplined reading acts on detailing, sets a band for digital, and treats the remaining three as untested rather than as poor performers. The correct next step for medical education is not a cut; it is to vary its funding across regions for two quarters so the next model can actually estimate it.

This is why a confidence interval belongs beside every number in an MMM output, and why an output presented without one should be sent back. A ranking without intervals invites exactly the decision the model cannot support.

Where marketing mix modelling fails

1. Asking it which doctor to call. MMM has no individual-level view by construction. A model asked for targeting guidance will produce something, and it will be an artefact of aggregation.
2. Reading returns without confidence intervals. A ranking of five channels where three coefficients are unmeasured looks identical to one where all five are solid. The interval is the part that says which is which.
3. Reallocating away from channels that never varied. An unmeasured coefficient is not a poor result. The correct response to no variation is to create some, not to cut the budget.
4. Ignoring reverse causality. Budgets follow demand as often as they create it, and a model built purely on observation credits the spend either way. Without at least one experimental anchor, every return in the table is an upper bound.
5. Treating collinear channels as separable. When two channels move together the split between them is arbitrary. More history does not fix it; different history does.
6. Accepting default priors from an interested party. Defaults are decisions, and whoever supplies the model has an interest in its assumptions. Ask what the priors are and who set them.
7. Borrowing an adstock half-life. Academic and practitioner estimates diverge roughly threefold, and the most-quoted single figure comes from FMCG. A detail visit to a specialist is not a television impression on a shopper.

A sequence for building an MMM that can be trusted

Ordered so that the model's credibility is established before its output is used to move money.

1. Audit variation before commissioning anything. For each channel, ask whether spend has meaningfully differed across territories or periods. Channels with no variation cannot be estimated, and knowing that in advance prevents a model being blamed for a silence in the data. This audit takes days and routinely changes the scope.
2. Fix the geography of the spend data. MMM needs activity by the same geographic unit as the outcome. Centrally-held channel budgets that cannot be allocated to territory are the most common blocker, and it is a reporting fix rather than an analytical one.
3. Run one experiment on the largest channel. A matched-territory design at two activity levels, over a full response window. This is the anchor that converts a correlational model into a partially calibrated one, and one channel is enough to start.
4. Build the model with intervals and out-of-sample validation, not fit statistics. Hold out a period that contains a genuine budget change and test whether the model predicts it. A high R-squared on the training period is not evidence of anything the decision needs.
5. Act only on the estimates the data supports, and create variation for the rest. Reallocate where the interval is narrow and the variation was real. For the channels the model could not estimate, deliberately vary their funding for two quarters — that is the input to the next refresh, and it is the step that compounds.

One expectation to set with a steering group. The first model will be able to answer fewer questions than were asked of it, because most organisations have not varied their spending enough to support the analysis they want. That is not a failure of the model or of the analyst. It is the accumulated cost of stability, and the only way to pay it down is to introduce variation on purpose.

What governs decisions between refreshes

MMM refreshes quarterly or half-yearly. Budgets, campaigns and field plans move continuously. The gap between the two is where most measurement programmes quietly stop being used, because a model that last spoke four months ago cannot answer the question in front of someone today.

The discipline that keeps a model useful between refreshes is unglamorous: maintain an event log. Competitor launches, supply interruptions, formulary changes, territory vacancies, campaign start and stop dates, recorded with dates and geographies as they happen. Reconstructed twelve months later it is guesswork; captured contemporaneously it is the difference between a model that explains an anomaly and one that absorbs it into a channel coefficient.

This is also the cheapest available improvement to the next model, and it requires no analytics capability at all. A shared log maintained by the brand team for two quarters will improve the following refresh more than any change to the model specification, because it converts unmeasured confounders into measured controls — which is precisely what the causation problem described earlier demands.

Key takeaways

The seven actions this article argues for, separated from the evidence that supports them.

• Audit variation before commissioning a model. A channel whose spend never varied cannot be estimated, and knowing that in advance stops a model being blamed for a silence in the data.
• Get spend onto the same geographic unit as the outcome. Centrally-held channel budgets that cannot be allocated to territory are the most common blocker, and the fix is a reporting change rather than an analytical one.
• Run one experiment on the largest channel. It is the anchor that converts a correlational model into a partially calibrated one, and a single channel is enough to start.
• Judge the model on out-of-sample validation, not on fit. Hold out a period containing a genuine budget change and test whether it predicts — a high R-squared on the training period is not evidence of anything the decision needs.
• Act only on estimates with narrow intervals and real variation behind them. A ranking without confidence intervals invites exactly the decision the model cannot support.
• Do not cut a channel the model could not estimate. The correct response to no variation is to create some, deliberately, for two quarters.
• Ask who set the priors. Whoever supplies the model has an interest in its assumptions, including any vendor whose own platform is one of the channels being evaluated.

Conclusion

Marketing mix modelling has a reputation problem in both directions. It is oversold as an objective arbiter of channel value, and dismissed as a black box that produces whatever the analyst expected. Neither is right. It is a reasonable method with well-understood limits, and almost all the disappointment attached to it comes from asking it questions it was never built to answer or accepting outputs the underlying data could not support.

The single most useful idea here inverts a common assumption. More history does not produce a better model — more variation does. A channel funded at a flat rate across every territory in every period is invisible to the analysis regardless of how many years of it exist, which turns an analytical exercise into an operational instruction: to measure a channel, be willing to change it. For teams in markets without prescriber-level data, that is a genuinely encouraging finding, because the aggregate method works on exactly the data those teams already hold.

Three questions decide whether an output can move money. Did this channel's spend actually vary? Is the interval narrow enough to act on? And is there an experimental anchor anywhere in the model establishing that spend caused the outcome rather than following it? A ranking that cannot answer all three is a description of last year's budget rather than a recommendation for next year's.

Frequently Asked Questions For Marketing Mix Modeling in Pharma