Marketing mix modeling interview questions fall into four groups: conceptual questions about what MMM is and how adstock, saturation and incrementality work; technical questions about specification, multicollinearity, priors and validation; case questions where a model gives a strange or uncertain result; and stakeholder questions about explaining results and handling disagreement. Below are 43 of them with model answers, plus flashcards so you can practise and track which ones you’re still shaky on.
I co-founded and ran a 50-person growth marketing agency that worked with hundreds of clients, and I’ve built marketing mix models for companies like Monzo, Booking.com and Time Out and advised Google, Appsflyer and Recast on MMM. The pattern I’d point out to anyone preparing is that interviewers rarely care whether you can recite a formula. They care whether you treat a model as evidence to be questioned rather than an answer, so the best responses to almost every question here include a check, a caveat or a test. The model answers are written to show that habit, and they’re deliberately concise: in a real interview you’d add an example from your own work.
How to use these marketing mix modeling interview questions
Try each question out loud before you look at the answer, because explaining adstock fluently to another person is a different skill from recognizing a correct explanation. Score yourself honestly: “got it” means you’d have said the substance of the model answer without prompting, “shaky” means you knew the idea but fumbled the explanation, and “missed” means you need to go and read about it. Your scores are saved in this browser, so you can come back and filter to the cards you haven’t got yet. Each section below links to the library guide that covers the topic in depth.
No cards left in this set. Untick the filter or pick another category.
The seniority of the role changes the mix. Junior analyst interviews lean on the conceptual and technical sections and often include a short coding exercise, such as implementing geometric adstock or computing ROI from a fitted regression. Senior and lead roles spend more time on the case and stakeholder sections, because the hard part of the job at that level is deciding what to do when the model and the business disagree.
Conceptual questions about marketing mix modeling
These test whether you understand what an MMM is for and how its main pieces work. If any feel shaky, the pillar guide to what marketing mix modeling is, the guides to saturation curves and marginal ROAS, and the comparison of MMM vs multi-touch attribution cover the ground.
1. What is marketing mix modeling, and what question does it answer?
Marketing mix modeling is a regression-based method that explains an outcome such as sales or sign-ups over time as a baseline plus the contributions of marketing channels and other drivers like price, promotions and seasonality. It answers how much each channel incrementally contributed and what would likely happen if budgets changed. It works on aggregated data, usually weekly, so it needs no user-level tracking. A strong answer also says what it can't do: it estimates effects from historical variation, so it's only as causal as that variation and the controls allow.
2. How is MMM different from multi-touch attribution?
Multi-touch attribution assigns credit for conversions to the touchpoints a tracked user saw, so it depends on user-level data and only sees trackable digital channels. MMM models aggregate outcomes against aggregate spend over time, so it covers offline media, price and seasonality and doesn't need cookies or device IDs. Attribution describes paths but doesn't measure incrementality: a touchpoint can get credit for a conversion that would have happened anyway. MMM aims at incrementality but at a coarser level and with wider uncertainty. Many teams use both, plus experiments, and reconcile them.
3. What is adstock, and why do we need it?
Adstock is a transformation that carries part of a period's advertising effect into later periods, because people remember ads and act on them after a delay. The simplest version is geometric: this week's adstock equals this week's spend plus a retention rate times last week's adstock. Without it the model assumes all effect happens in the week of spend, which undercounts channels like TV and can attribute delayed effects to whatever ran the following week. You can describe it by its half-life, the number of weeks until half the effect has gone.
4. What is a saturation curve, and how is it different from adstock?
A saturation curve describes diminishing returns within a period: each additional unit of spend buys less extra outcome than the last, because you reach the most responsive people first and then pay for frequency. Common forms are the Hill function, the logistic, and the exponential or log transformation. Adstock is about time (when the effect lands), saturation is about scale (how much effect a given level of spend produces). Most MMMs apply adstock first and then saturation, and the order is a modelling choice worth stating.
5. What is the baseline in an MMM?
The baseline is the part of the outcome the model attributes to things other than the marketing it measures: brand equity built up over years, distribution, trend, seasonality and organic demand. It's what the model predicts you'd sell if measured media spend were zero, which is itself an extrapolation, often far outside the data. The baseline share is sensitive to specification: adding a trend or seasonality term moves credit from media to baseline, and dropping one does the opposite. A very high or very low baseline share is a reason to look at the specification, not a finding in itself.
6. What's the difference between average ROI and marginal ROI, and which should drive budget decisions?
Average ROI is total incremental outcome divided by total spend for a channel. Marginal ROI is the incremental outcome from the next unit of spend, read off the slope of the response curve at current spend. Because of diminishing returns, marginal ROI is lower than average ROI for any saturating channel. Budget decisions should use marginal ROI: a channel with a great average return can still be the worst place for the next dollar if it's already saturated, and an optimal allocation equalises marginal returns across channels subject to constraints.
7. What does incrementality mean, and why does MMM care about it?
Incrementality is the difference between what happened and what would have happened without the marketing: the causal effect. A conversion that would have happened anyway isn't incremental even if an ad was clicked first. MMM is meant to estimate incremental contributions, which is what makes it useful for budgeting, but it does so from observational data, so its incrementality claims rest on assumptions about controls and variation. That's why good practice calibrates or validates MMM against experiments such as geo tests or conversion lift studies.
8. What are the main limitations of MMM?
It needs enough history and enough independent variation in each channel, so new or always-on channels with flat spend are hard to measure. Channels that move together can't be separated well. It's observational, so omitted confounders bias the estimates. It works at an aggregate level, so it's poor at creative, audience or campaign-level questions. Results are uncertain and sensitive to specification choices such as adstock form and priors. And it predicts best near observed spend levels; recommendations far outside them are extrapolations.
9. Why has MMM become more popular again in recent years?
Privacy changes such as Apple's App Tracking Transparency and the decline of third-party cookies weakened user-level attribution, so marketers needed a method that works on aggregate data. At the same time open-source libraries from Meta (Robyn), Google (LightweightMMM and then Meridian) and PyMC Labs (PyMC-Marketing) made MMM cheaper to build in-house, and cloud compute made Bayesian estimation practical. Finance teams also pushed for measurement that covers all channels, including offline, on a consistent basis.
10. What is a measurement framework, and where does MMM fit in it?
A measurement framework assigns each question to the method best suited to answer it and defines how the methods inform each other. MMM usually answers strategic questions: total budget, cross-channel mix and long-run planning. Experiments answer specific causal questions with high confidence, such as whether a channel is incremental. Attribution and platform reporting support fast in-channel optimisation. A good framework calibrates MMM with experiments, uses MMM to choose which experiments to run next, and is explicit about which number wins when methods disagree.
Technical MMM interview questions
These are where analyst and data scientist interviews spend most of their time. Interviewers often follow up by asking you to write the code, so it’s worth being able to sketch a regression with transformed media in Python or R; the from-scratch Python MMM is good practice. The guides to multicollinearity, MMM validation, model diagnostics and Bayesian MMM go deeper.
11. Write down a basic MMM specification and explain each term.
A simple form is: sales at time t equals an intercept, plus the sum over channels of a coefficient times the saturated adstock of that channel's spend, plus coefficients on control variables such as price, promotions, seasonality and trend, plus an error term. The intercept and trend capture the baseline, the transformed media terms capture marketing effects with carryover and diminishing returns, and the controls absorb other drivers so they aren't credited to media. In a log-log version, coefficients become elasticities. A good answer mentions that the transformation parameters are either searched over, as in Robyn, or estimated with priors, as in Bayesian models.
12. How do you handle multicollinearity between channels?
First diagnose it, with pairwise correlations and variance inflation factors, and by checking whether coefficients swing when you add or drop a channel. Options include combining channels that always move together, regularisation such as ridge regression (which Robyn uses), informative priors in a Bayesian model, calibration with experiment results, and, best of all, creating independent variation by deliberately varying one channel's spend or running a geo test. More data from the same plan doesn't fix it, because the problem is that the channels never vary separately.
13. How do you choose adstock and saturation parameters?
There are three broad approaches. You can grid-search or optimise them against a fit criterion, as Robyn does with Nevergrad, using multi-objective selection to avoid implausible decompositions. You can estimate them jointly with the coefficients in a Bayesian model, with priors reflecting what's plausible for each channel type. Or you can fix them from external evidence, such as experiments or published benchmarks. Whichever you use, check that the results are plausible (TV carryover longer than search, for instance), and test how sensitive the ROIs are to the choice.
14. How do you validate a marketing mix model?
Fit on training data alone isn't enough, because a flexible model can fit history well and still be wrong about cause. Use a time-based holdout or rolling cross-validation and report out-of-sample error such as MAPE or NRMSE. Check residuals for autocorrelation, structure and outliers. Check stability: do coefficients hold up when you change the window or drop a variable? Check plausibility against benchmarks and business knowledge. Most importantly, compare the model's estimates with experiment results where you have them, and calibrate if they disagree.
15. What is DECOMP.RSSD in Robyn and why is it used?
DECOMP.RSSD is the decomposition root sum of squared distance: it measures how far each channel's share of effect is from its share of spend. Robyn uses it alongside NRMSE as a second objective, so model selection prefers models that fit well without assigning wildly different effect shares from spend shares. The idea is to avoid technically good-fitting models with implausible decompositions. The trade-off is that it nudges results toward spend shares, which can hide genuinely very efficient or inefficient channels, so it's a regulariser on plausibility, not evidence.
16. What is the role of priors in a Bayesian MMM, and how do you set them?
Priors encode what you believe about parameters before seeing this dataset, such as that media effects are positive, carryover is shorter for search than for TV, and ROIs are probably in a plausible range. With limited, collinear data, priors materially affect results, so they should be explicit and justified, ideally from experiments or previous models. Meridian, for example, lets you put priors directly on ROI. Good practice is to run prior predictive checks, compare priors with posteriors to see where the data was informative, and test sensitivity to looser priors.
17. How would you check whether a Bayesian MMM has converged?
Check R-hat for every parameter (values close to 1, commonly below 1.01, indicate chains agree), effective sample size (enough independent draws for stable estimates), and divergent transitions (any divergences suggest the sampler struggled with the geometry, often from poorly scaled data or flat priors). Look at trace plots for mixing. Then check the posterior predictive fit against the data. Convergence only says the sampler explored the posterior well; it says nothing about whether the model is right.
18. What does a residual plot tell you in an MMM?
Residuals should look like noise. A pattern over time suggests missing trend or seasonality; runs of positive or negative residuals, or a low Durbin-Watson statistic, indicate autocorrelation, which makes standard errors too small; isolated spikes suggest a missed promotion, outage or event; a level shift suggests a structural break such as a tracking change. Residuals that correlate with lagged spend suggest the carryover is wrong. Fixing these usually changes channel estimates, which is why reading residuals is part of review.
19. What data would you ask for to build an MMM?
A consistent outcome series, usually weekly, ideally two to three years long; spend (and where possible impressions or GRPs) by channel for the same weeks; price and promotion history; distribution or availability; seasonality and holiday markers; known events such as launches, outages, stockouts and PR; and relevant external factors such as competitor activity or macro data. A data dictionary is as important as the data: definitions, currencies, net versus gross spend, and whether the outcome includes returns. I'd also ask for any past experiments or lift tests, which are the most valuable inputs for calibration.
20. How much data do you need for an MMM?
A common rule of thumb is at least two years of weekly data, enough to see seasonality repeat, and something like ten or more observations per estimated parameter. But the binding constraint is usually variation, not row count: a channel with flat spend or one that always moves with another can't be estimated well from any amount of history. Geo-level data helps by multiplying the observations. If data is short, informative priors and experiment calibration become more important.
21. What's the difference between a linear, log-linear and log-log MMM?
In a linear model, coefficients are absolute effects: one more unit of spend adds a fixed amount of sales, so diminishing returns must come from a saturation transform. In a log-linear model (log of sales on raw inputs), coefficients are approximately percentage changes in sales per unit of input, so effects scale with the baseline. In a log-log model, coefficients are elasticities: a 1 percent change in spend gives a coefficient-percent change in sales, which builds in diminishing returns and is natural for price. The choice changes how effects interact and how you compute ROI.
22. How do you compute ROI and response curves from a fitted model?
Compute each channel's contribution by taking the model's prediction with the channel at actual spend minus the prediction with that channel set to zero, summing over the period, in original units. ROI is contribution divided by spend (or margin-adjusted contribution divided by spend for profit ROI). A response curve repeats the prediction at a range of spend multipliers to show total contribution against spend, and its slope gives marginal ROI. In a Bayesian model, do this for each posterior draw to get intervals, not just a point.
Questions on experiments and incrementality
Most marketing science teams now pair models with experiments, so expect at least a couple of these. The guides to incrementality testing, geo-lift testing, statistical power and calibrating MMM with experiments cover each in depth.
23. How would you use an experiment to calibrate an MMM?
Run a test that measures one channel's incremental effect, such as a geo holdout or conversion lift study, and convert the result to the model's units, such as incremental outcome per dollar over the same window. Then use it to constrain the model: as a prior on that channel's ROI or coefficient in a Bayesian model, as a calibration penalty as Robyn does, or as a check that rejects models inconsistent with it. Be careful that the experiment's window, geography and spend level match what the model estimates, and document how much the calibration moved the result.
24. Explain how a geo-lift test works.
You split geographic regions into treatment and control groups, change spend in the treatment regions (increase it, decrease it, or switch it off) and compare outcomes against what the control regions imply would have happened. Methods include difference-in-differences, matched markets and synthetic control, where the counterfactual is a weighted combination of control regions fitted on the pre-period. The design needs regions whose outcomes track each other before the test, enough regions and duration for statistical power, and a plan for spillover between regions.
25. What is statistical power, and why do many marketing tests fail to reach it?
Power is the probability that a test detects an effect of a given size if it exists. It depends on the effect size, the noise in the outcome, the number of units and the duration. Marketing effects are often small relative to the natural variation in sales, so tests need large samples or long durations to detect them, which is the point Lewis and Rao made about measuring advertising returns. Running an underpowered test and calling a null result 'no effect' is a common mistake; you should compute the minimum detectable effect before launching.
26. What is difference-in-differences and its key assumption?
Difference-in-differences estimates an effect as the change in the treated group minus the change in the control group over the same period, which removes fixed differences between the groups and shared shocks over time. Its key assumption is parallel trends: without treatment, both groups would have moved in parallel. You check it by looking at pre-period trends, and it fails if, for instance, treatment regions were chosen because they were growing faster.
27. Why might platform conversion lift results differ from MMM results for the same channel?
They measure different things over different windows. Platform lift studies typically measure short-window conversions among users the platform can track, often an intent-to-treat effect, and may count conversions the MMM attributes to baseline or other channels. MMM measures aggregate outcomes with carryover over longer windows, and may miss short-lived effects or absorb them into other terms. Differences in outcome definition, attribution window, spend level and geography all matter, so you reconcile them by aligning definitions before deciding which is wrong.
28. When would you run an experiment instead of relying on the model?
When the decision is large and the model is unsure about it: a channel whose interval crosses zero or spans a decision threshold, a channel the model can't separate from another, a new channel with little history, or a recommendation that moves spend far outside observed levels. I'd also test when the model's result contradicts strong prior evidence. Experiments are expensive, so the model is useful for deciding which test has the most value of information.
29. What is synthetic control?
Synthetic control builds a counterfactual for a treated unit, such as a region where you launched TV, as a weighted average of untreated units chosen so that the weighted combination closely matches the treated unit's outcome before the intervention. After the intervention, the gap between the treated unit and its synthetic version estimates the effect. It works well with few treated units and long pre-periods, and tools like GeoLift and CausalImpact use related ideas. Its weak points are poor pre-period fit and shocks that hit only the treated unit.
Case questions: what would you do if the model said this?
Case questions don’t have a single right answer, and that’s the point. A good structure is to check the data first, then the specification and possible confounders, then the uncertainty, and only then decide what to recommend, which is usually a test before a large budget move. The guide to confounding variables in MMM is the most useful background for this section.
30. The model says one channel has an ROI of 12 while others are around 1 to 2. What do you do?
Treat it as a reason to investigate, not a finding. Check whether the channel's spend coincides with something else that drives sales, such as promotions, seasonality or another channel, and whether a control is missing. Check how much spend variation the estimate comes from, and how wide the interval is. Compare with benchmarks and any experiments. Check the decomposition for errors in units or scaling. If it survives all that, propose a test before moving large budget, because the cost of being wrong is high and the model is extrapolating if you scale spend a lot.
31. A channel shows a negative coefficient. How do you handle it?
Negative media effects are rarely real, so first look for causes: collinearity with another channel, spend that's reactive to falling sales (budget increased when sales dropped), a missing control, or a data problem such as misaligned dates. Constraining coefficients to be non-negative, as most MMM tools do by default, hides the symptom but not the cause, so investigate before constraining. Report what you found and how the constraint affected other channels' estimates.
32. The CMO wants to cut a channel because its MMM interval includes zero. What do you advise?
An interval that includes zero means the data can't tell whether the channel works, not that it doesn't. Cutting it on that basis is acting on absence of evidence. I'd look at why the estimate is uncertain (little spend variation, collinearity), what the upper end of the interval would mean for the business, and propose a test, such as a geo holdout or spend pause in some regions, to settle it. If budget must be cut now, a partial cut in a test design gives you the evidence and the savings.
33. You're asked to build an MMM for a brand with only 12 months of data. What do you say?
Twelve months won't show seasonality repeating, so the model can't easily separate seasonal demand from seasonal spend. I'd explain the risk, then suggest ways to make the most of it: use geo-level or daily data to increase observations, use external seasonality data or priors, use informative priors from benchmarks or experiments, limit the number of channels estimated separately, and plan experiments to calibrate. I'd present results as provisional with wide intervals and set a date to refit with more data.
34. The model's holdout error is much worse than its training error. What's happening and what do you do?
That's a sign of overfitting or of a change between periods. Too many terms for the data, flexible transformations tuned to the training window, or dummy variables for every outlier can all fit noise. Alternatively, something changed in the holdout period: tracking, pricing, a new competitor or a structural break. I'd simplify the model, use regularisation or priors, check for breaks in the holdout period, and compare against a simple benchmark such as a seasonal naive forecast. A model that can't beat a simple baseline out of sample shouldn't drive budget decisions.
35. Branded search shows the highest ROI in the model. Should you increase its budget?
Probably not on that evidence. Branded search mostly captures people already looking for the brand, often because of other marketing, so it tends to steal credit from upstream channels like TV and social and to look highly efficient. I'd check whether it's being treated as a cause when it's partly an outcome, consider modelling it as a mediator or leaving it out of the media set, and recommend a test such as a branded search holdout to measure its true incrementality.
36. Two models with similar fit give very different channel ROIs. Which do you choose?
Similar fit with different ROIs means the data can't distinguish them, which is itself the key finding. I'd look at which is more consistent with experiments and benchmarks, which has more plausible adstock and saturation, which is more stable across time windows, and whether the difference would change the decision. If it would, present both with the uncertainty and propose the test that would tell them apart. Choosing one silently hides the uncertainty from the people making the decision.
Stakeholder and communication questions
Every MMM eventually has to persuade someone to change a budget, so these questions come up even in technical interviews. The guide to communicating MMM results has a decision memo template that turns most of these answers into a document.
37. How would you explain MMM results to a CFO?
Lead with the decision and the expected financial effect as a range: what we recommend changing, by how much, and what it's expected to do to revenue or profit. Then say why we believe it, distinguishing what experiments confirmed from what only the model says, and what could make us wrong. Avoid jargon, convert effects into money and margin, and show how we'll check the result, such as a test or a metric to watch with a date to report back. CFOs trust ranges and stated risks more than precise-looking point estimates.
38. A channel manager disputes the model because their platform reports a much higher ROAS. How do you respond?
Acknowledge that both numbers measure something real but different: platform ROAS counts attributed conversions, many of which may have happened anyway, while the model estimates incremental effect. Explain the difference without dismissing their data, and look for specific reasons the model might be wrong for their channel. Then propose a test that both sides agree in advance will settle it. Agreeing the test design before seeing results turns a political argument into an empirical one.
39. How do you communicate uncertainty without losing credibility?
Present ranges as the normal state of measurement and tie them to decisions: say whether the decision holds across the range, and if it doesn't, what would resolve it. Use plain language such as 'we're confident it's between X and Y' rather than statistics jargon. Separate what's well established from what's uncertain, and have a plan for reducing the uncertainty that matters. In my experience, stakeholders lose confidence when a precise number turns out wrong, not when an honest range turns out right.
40. Leadership wants the MMM to tell them the ROI of every campaign. What do you say?
MMM works on aggregate time series, so it estimates channel-level effects well but generally can't separate individual campaigns, creatives or audiences that run at the same time, because there isn't enough independent variation. I'd explain what level of detail the data supports, suggest which questions belong to experiments or in-platform tests, and agree on a measurement framework where each question goes to the right method rather than overloading the model.
41. How would you set up an MMM project with a new client or team?
Start with the decision the model is meant to inform and the outcome that matters, such as profit or new customers, and agree who will act on the results and when. Then audit the data, document definitions and gaps, and agree the specification and validation approach before results are seen, including how experiments will be used. Set expectations on timing, uncertainty and what the model can't answer. Plan for refreshes and a feedback loop where results drive tests that improve the next model.
42. What would you do if a senior stakeholder asked you to change the model to give a different answer?
I'd ask what they believe is wrong and why, because sometimes they know something the model doesn't, such as an event or a data issue, and that's a legitimate reason to change the specification. If there's a real reason, I'd make the change transparently and document it. If the request is simply for a different number, I'd explain that changing assumptions to get a desired result makes the model worthless for decisions, show how sensitive the result is to the assumption they want changed, and propose a test to settle the disagreement.
43. Tell me about a time a model result was wrong or surprising. What did you learn?
This is a behavioural question, so answer it with a real example in a situation, action, result structure. Good answers show that you checked the data and specification before trusting or dismissing a surprising result, that you used an independent check such as an experiment or a holdout, and that you communicated the correction clearly. Interviewers want to see that you treat models as fallible and that you changed your process afterwards, for instance by adding a data check or a validation step.
If you want to go further than reading, building a model end to end is the best preparation, because it gives you real choices to talk about. The Vexpower MMM course runs one invented brand from measurement plan to decision memo, the MMM courses roundup lists the free library notebooks, and the glossary is useful for a last-minute check on terms like DECOMP.RSSD and half-life.
