Marketing Mix Modeling and Incrementality

Marketing Mix Modeling & Incrementality

MMM is a top-down, regression-based method that uses aggregated time-series data (spend, impressions, sales) to estimate the incremental contribution and ROI of each marketing channel, controlling for baseline, trend, seasonality, price, and promotions. It is privacy-durable (no user-level tracking) and the dominant method in the post-cookie / post-ATT era. This skill is the marketing-measurement application layer; for the general causal-inference toolkit see da-12-ab-testing-causal-inference.

The modern measurement stack is a calibration triad: MMM frames strategy and which channels to test, incrementality/geo experiments produce ground-truth causal estimates, and those estimates calibrate the MMM (as priors or likelihood constraints). MTA fills the short-term tactical-optimization gap where consented user-level signal still exists.

Core Concepts

Adstock / carryover

Advertising effect persists and decays over subsequent periods. Geometric adstock: adstock_t = x_t + alpha * adstock_{t-1}, where alpha in [0,1) is the retention/decay rate (higher = longer carryover). Often truncated at a max lag l_max (e.g. 4-8 weeks). Delayed/Weibull adstock adds a peak-delay parameter (theta) so the effect peaks days after exposure (TV, brand) rather than immediately — used by Robyn and described in Google’s carryover paper.

Saturation / diminishing returns

Each channel’s response is concave: incremental spend buys less incremental outcome as the channel saturates. Common forms:

• Hill function (from pharmacology): response = x^s / (k^s + x^s), with shape s and half-saturation k. Used by Meridian and DeepCausalMMM.
• Logistic saturation: used by PyMC-Marketing; saturation_lam controls curvature.
• Michaelis-Menten / exponential variants in other tools. Adstock is applied before saturation: transform spend → carryover → saturated response → linear coefficient.

Response curves and budget allocation

The fitted saturation curves yield diminishing-return curves per channel. The optimizer reallocates budget so marginal ROAS is equalized across channels (move spend from saturated to under-invested channels until marginal returns match). This — not the historical ROAS point estimate — is the decision-grade MMM output.

Bayesian MMM

Treats all parameters (baseline, channel betas, adstock alpha, saturation shape) as distributions. Advantages: encodes priors from domain knowledge / past lift tests, produces full posterior uncertainty (credible intervals on ROI), and supports hierarchical / geo-level pooling. Sampling via NUTS/HMC (PyMC, TensorFlow Probability). Frequentist MMM (ridge regression, e.g. Robyn) instead penalizes coefficients to handle multicollinearity among correlated channels.

Tools / Frameworks

Choosing: largest spend on Google → Meridian integrates more easily; largest spend on Meta → Robyn. Prefer Bayesian (Meridian/PyMC-Marketing) when you have priors/lift tests to fold in or need uncertainty quantification; Robyn for fast, non-technical, decomposition-heavy workflows.

MTA vs MMM (and the unified view)

Treating these as either/or is outdated. Unified Marketing Measurement (UMM) combines MMM (strategy) + MTA (tactics) + experiments (ground truth). Neither MTA nor MMM is causal by itself — only experiments are; experiments calibrate both.

Incrementality & geo experiments

Incrementality = the causal lift attributable to advertising vs. a counterfactual where it never ran (not the same as last-click attributed conversions, which include organic/baseline demand).

Geo experiments (GeoLift / geo-lift) are the practical gold standard:

• Markets (DMAs, regions, ZIPs) are randomized or selected into treatment (campaign on / spend change) vs control (held out).
• A counterfactual for treated geos is built from control geos.
  • Synthetic control (e.g. Haus, Meta GeoLift): a weighted blend of control geos that best matches the treated geo’s pre-period trajectory.
  • TBR (time-based regression) / CausalImpact: Google’s Bayesian structural time-series builds a counterfactual from control-market series and returns lift with credible intervals.
• Key outputs: incremental conversions/revenue, iROAS (incremental ROAS), iCPA, and confidence/credible bounds.
• Power / MDE: run a pre-test power analysis to pick test length, number of geos, and the minimum detectable effect; underpowered tests produce inconclusive lift. Account for delayed conversions (e.g. long consideration windows) by extending the post-period.

Pitfalls: spillover/contamination between adjacent geos; too few or poorly matched control geos; post-hoc tweaking of the analysis window (p-hacking); ignoring effect size in favor of p-values; insufficient pre-period for the counterfactual fit.

Calibrating MMM with experiments

This is where the methods become one system. A geo/lift test gives a causal estimate for a channel; feed it back into the MMM so the model’s belief about that channel is anchored to reality.

Two mechanisms:

1. ROI/ROAS priors (Meridian): set each channel’s ROI prior from past lift tests, benchmarks, or experiments; the Bayesian model shrinks toward them when the time series is weak/collinear.
2. Likelihood-based saturation calibration (PyMC-Marketing add_lift_test_measurements()): each lift test contributes {channel, x (pre-test spend), delta_x (spend change), delta_y (measured sales change), sigma (uncertainty)} — effectively two points on the channel’s saturation curve. The framework adds these as constraints on the saturation function itself, so more lift tests at different spend levels keep improving the curve, not just one anchored ROAS point.

Operating cadence: at least one lift/geo test per major channel per quarter; re-calibrate the MMM and review diagnostics on that cadence.

Privacy-era / post-cookie measurement

• iOS ATT (App Tracking Transparency) and Chrome third-party cookie deprecation gutted user-level tracking → MTA degraded, MMM and experiments resurged because they need no individual identifiers.
• Data clean rooms (e.g. Google ADH, Amazon Marketing Cloud, retail-media clean rooms) allow privacy-safe joins of advertiser + platform data for aggregated lift/incrementality measurement without exposing PII.
• Durable stack: MMM (durable, no PII) + geo/lift experiments (causal truth) + clean-room aggregated measurement + consented first-party data, with MTA only where consent persists.

Practical Patterns

• Transform order: spend → adstock (carryover) → saturation → linear term. Getting the order wrong inverts the economics.
• 2-3 years of weekly data typical; ensure spend variation per channel (a channel with flat spend is unidentifiable).
• Decompose first: isolate baseline (trend, seasonality, holidays, price, promo, distribution) so media coefficients capture incremental media effect.
• Validate with holdout / time-series CV and NRMSE / R²; check residuals.
• Calibrate, then optimize: anchor channels to experiments before trusting the budget optimizer’s reallocation.
• Report uncertainty: present credible intervals on ROI, not point estimates, to set decision risk.
• Quarterly experiment cadence keeps the MMM honest as creative, audiences, and saturation shift.

Anti-Patterns

• Treating MTA last-click as incrementality — it credits demand that would have converted anyway; inflates ROAS on bottom-funnel/retargeting.
• Skipping calibration — uncalibrated MMM ROIs are easily confounded by collinear channels (TV and search rising together) and will misallocate budget.
• Over-trusting the optimizer beyond the observed spend range — saturation curves extrapolate poorly; cap reallocation to a sane band of historical spend.
• Ignoring carryover — fitting media to same-week sales only understates TV / brand and overstates fast-response channels.
• Reusing LightweightMMM for new builds — deprecated; migrate to Meridian.
• One-and-done MMM — a model not re-fit/re-calibrated drifts within a quarter.
• Underpowered geo tests — running a test too short or with too few geos yields wide intervals and an inconclusive read presented as “no lift.”

Troubleshooting

• Implausible / negative channel ROI → multicollinearity; add experiment priors, drop/aggregate correlated channels, or use ridge (Robyn) / stronger Bayesian priors.
• Saturation curve looks linear (no diminishing returns) → insufficient high-spend observations; add lift tests at higher spend, or constrain priors.
• MMM and lift test disagree → trust the experiment; recalibrate. Persistent gaps suggest omitted controls or wrong adstock length.
• Geo test inconclusive → re-run power analysis; lengthen post-period for delayed conversions; verify control geos match the pre-period; check for spillover.
• Posterior won’t converge (high R-hat, divergences) → tighten priors, reparameterize, increase samples; check for collinear predictors.

References

1. The Role of Adstock and Saturation Curves in MMM (ResearchGate, 2024) — https://www.researchgate.net/publication/388175908_The_Role_of_Adstock_and_Saturation_Curves_in_Marketing_Mix_Models_Implications_for_Accuracy_and_Decision-Making
2. Carryover and Shape Effects in Media Mix Modeling (Towards Data Science, 2024) — https://towardsdatascience.com/carryover-and-shape-effects-in-media-mix-modeling-paper-review-fd699b509e2d/
3. DeepCausalMMM: Deep Learning Framework for MMM with Causal Inference (arXiv, 2025) — https://arxiv.org/html/2510.13087v1
4. Diminishing Return Curves Turn MMM into Budget Decisions (Measured, 2024) — https://www.measured.com/faq/media-mix-modeling-diminishing-return-curves-mmm-budget-decision/
5. Google Meridian MMM: The 2025 Guide (Eliya, 2025) — https://www.eliya.io/blog/media-mix-modeling/google-meridian-mmm
6. Master Bayesian MMM with PyMC-Marketing (Eliya, 2025) — https://www.eliya.io/blog/media-mix-modeling/pymc-marketing-bayesian-mmm-guide
7. Bayesian Media Mix Modeling for Marketing Optimization (PyMC Labs, 2024) — https://www.pymc-labs.com/blog-posts/bayesian-media-mix-modeling-for-marketing-optimization
8. Meridian vs Robyn: Comprehensive Comparison for 2025 (Eliya, 2025) — https://www.eliya.io/blog/media-mix-modeling/Meridian-vs-Robyn
9. Google Meridian vs Meta Robyn — What’s Next for MMM (Double, 2025) — https://www.double.io/newsletter/google-meridian-vs-meta-robyn-whats-next-for-mmm
10. Exploring Meridian, Google’s new open-source MMM (Search Engine Land, 2024) — https://searchengineland.com/exploring-meridian-googles-new-open-source-marketing-mix-model-438754
11. Open Source Battle for MMM: Robyn vs LightweightMMM (Forvio, 2024) — https://www.forvio.com/resources/blog/open-source-battle-for-mmm-robyn-vs-lightweightmmm
12. Multi-touch attribution vs marketing mix modeling (Funnel.io, 2024) — https://funnel.io/blog/mta-vs-mmm
13. MTA vs MMM: marketing measurement in a privacy-first world (Usercentrics, 2025) — https://usercentrics.com/knowledge-hub/mta-vs-mmm/
14. MTA vs MMM (Haus, 2024) — https://www.haus.io/blog/mta-vs-mmm-choosing-between-multi-touch-attribution-and-marketing-mix-modeling
15. Incrementality testing vs MMM vs MTA pros/cons (Measured, 2024) — https://www.measured.com/faq/what-are-the-pros-and-cons-of-incrementality-testing-versus-mmm-or-mta/
16. GeoLift Framework: Incrementality Testing Guide (Andava, 2024) — https://www.andava.com/learn/geolift-framework-incrementality-testing-guide/
17. GeoLift 101: Geo-Based Incrementality Testing (Triple Whale, 2024) — https://www.triplewhale.com/blog/geolift-geo-based-incrementality-testing
18. Geo Experiments: The Fundamentals (Haus, 2024) — https://www.haus.io/blog/geo-experiments-the-fundamentals
19. Geo-Based Incrementality Testing Playbook for 2025 (Lifesight, 2025) — https://lifesight.io/blog/geo-based-incrementality-testing/
20. MMM Calibration with Lift Tests and Bayesian Methods (PyMC Labs, 2024) — https://www.pymc-labs.com/blog-posts/mmm_roas_lift
21. Data Clean Rooms: Privacy-Safe Marketing Attribution Guide (Hashmeta, 2025) — https://hashmeta.com/blog/data-clean-rooms-the-complete-guide-to-privacy-safe-marketing-attribution/
22. Post-Cookie Attribution Playbook for 2026 (GrowthMarketer, 2025) — https://growthmarketer.com/blog/post-cookie-attribution-playbook/
23. MMM vs MTA vs Lift Tests 2026: The Measurement Matrix (Digital Applied, 2025) — https://www.digitalapplied.com/blog/media-mix-vs-attribution-vs-mta-2026-decision-matrix
24. The 2025 State of Data Clean Rooms in Retail Media (Skai, 2025) — https://skai.io/blog/data-clean-rooms-in-retail-media/