Marketing Mix Modeling and Incrementality

Parent: data analysis · researched 2026-05-30T22:06:56.555Z· 24 sources · 9 concepts · skill da-22-marketing-mix-modeling

MMM is a top-down, regression-based method that uses aggregated time-series data

Marketing Mix Modeling & Incrementality

MMM is a top-down, regression-based method that uses aggregated time-series data [source]
(spend, impressions, sales) to estimate the incremental contribution and ROI of [source]
each marketing channel, controlling for baseline, trend, seasonality, price, and [source]
promotions. It is privacy-durable (no user-level tracking) and the dominant method [source]
in the post-cookie / post-ATT era. This skill is the marketing-measurement [source]
application layer; for the general causal-inference toolkit see [source]
da-12-ab-testing-causal-inference. [source]
The modern measurement stack is a calibration triad: MMM frames strategy and [source]
which channels to test, incrementality/geo experiments produce ground-truth [source]
causal estimates, and those estimates calibrate the MMM (as priors or [source]
likelihood constraints). MTA fills the short-term tactical-optimization gap where [source]
consented user-level signal still exists. [source]

Adstock / carryover

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

Saturation / diminishing returns

Each channel's response is concave: incremental spend buys less incremental [source]
outcome as the channel saturates. Common forms: [source]
Hill function (from pharmacology): response = x^s / (k^s + x^s), with [source]
shape s and half-saturation k. Used by Meridian and DeepCausalMMM. [source]
Logistic saturation: used by PyMC-Marketing; saturation_lam controls [source]
Michaelis-Menten / exponential variants in other tools. [source]
Adstock is applied before saturation: transform spend → carryover → saturated [source]
response → linear coefficient. [source]

Response curves and budget allocation

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

Bayesian MMM

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

Tools / Frameworks

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

MTA vs MMM (and the unified view)

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

Incrementality & geo experiments

Incrementality = the causal lift attributable to advertising vs. a [source]
counterfactual where it never ran (not the same as last-click attributed [source]
conversions, which include organic/baseline demand). [source]
Geo experiments (GeoLift / geo-lift) are the practical gold standard: [source]
Markets (DMAs, regions, ZIPs) are randomized or selected into treatment [source]
(campaign on / spend change) vs control (held out). [source]
A counterfactual for treated geos is built from control geos. [source]
Synthetic control (e.g. Haus, Meta GeoLift): a weighted blend of control [source]
geos that best matches the treated geo's pre-period trajectory. [source]
TBR (time-based regression) / CausalImpact: Google's Bayesian [source]
structural time-series builds a counterfactual from control-market series and [source]
returns lift with credible intervals. [source]
Key outputs: incremental conversions/revenue, iROAS (incremental ROAS), [source]
iCPA, and confidence/credible bounds. [source]
Power / MDE: run a pre-test power analysis to pick test length, number of [source]
geos, and the minimum detectable effect; underpowered tests produce [source]
inconclusive lift. Account for delayed conversions (e.g. long [source]
consideration windows) by extending the post-period. [source]
Pitfalls: spillover/contamination between adjacent geos; too few or poorly [source]
matched control geos; post-hoc tweaking of the analysis window (p-hacking); [source]
ignoring effect size in favor of p-values; insufficient pre-period for the [source]

Calibrating MMM with experiments

This is where the methods become one system. A geo/lift test gives a causal [source]
estimate for a channel; feed it back into the MMM so the model's belief about that [source]
channel is anchored to reality. [source]
ROI/ROAS priors (Meridian): set each channel's ROI prior from past lift [source]
tests, benchmarks, or experiments; the Bayesian model shrinks toward them when [source]
the time series is weak/collinear. [source]
Likelihood-based saturation calibration (PyMC-Marketing [source]
add_lift_test_measurements()): each lift test contributes [source]
`{channel, x (pre-test spend), delta_x (spend change), delta_y (measured sales [source]
change), sigma (uncertainty)}` - effectively two points on the channel's [source]
saturation curve. The framework adds these as constraints on the saturation [source]
function itself, so **more lift tests at different spend levels keep improving [source]
the curve**, not just one anchored ROAS point. [source]
Operating cadence: at least one lift/geo test per major channel per quarter; [source]
re-calibrate the MMM and review diagnostics on that cadence. [source]

Privacy-era / post-cookie measurement

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

Practical Patterns

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

Anti-Patterns

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

Troubleshooting

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

References

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 [source]
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/ [source]
DeepCausalMMM: Deep Learning Framework for MMM with Causal Inference (arXiv, 2025) - https://arxiv.org/html/2510.13087v1 [source]
Diminishing Return Curves Turn MMM into Budget Decisions (Measured, 2024) - https://www.measured.com/faq/media-mix-modeling-diminishing-return-curves-mmm-budget-decision/ [source]
Google Meridian MMM: The 2025 Guide (Eliya, 2025) - https://www.eliya.io/blog/media-mix-modeling/google-meridian-mmm [source]
Master Bayesian MMM with PyMC-Marketing (Eliya, 2025) - https://www.eliya.io/blog/media-mix-modeling/pymc-marketing-bayesian-mmm-guide [source]
Bayesian Media Mix Modeling for Marketing Optimization (PyMC Labs, 2024) - https://www.pymc-labs.com/blog-posts/bayesian-media-mix-modeling-for-marketing-optimization [source]
Meridian vs Robyn: Comprehensive Comparison for 2025 (Eliya, 2025) - https://www.eliya.io/blog/media-mix-modeling/Meridian-vs-Robyn [source]
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 [source]
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 [source]
Open Source Battle for MMM: Robyn vs LightweightMMM (Forvio, 2024) - https://www.forvio.com/resources/blog/open-source-battle-for-mmm-robyn-vs-lightweightmmm [source]
Multi-touch attribution vs marketing mix modeling (Funnel.io, 2024) - https://funnel.io/blog/mta-vs-mmm [source]
MTA vs MMM: marketing measurement in a privacy-first world (Usercentrics, 2025) - https://usercentrics.com/knowledge-hub/mta-vs-mmm/ [source]
MTA vs MMM (Haus, 2024) - https://www.haus.io/blog/mta-vs-mmm-choosing-between-multi-touch-attribution-and-marketing-mix-modeling [source]
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/ [source]
GeoLift Framework: Incrementality Testing Guide (Andava, 2024) - https://www.andava.com/learn/geolift-framework-incrementality-testing-guide/ [source]
GeoLift 101: Geo-Based Incrementality Testing (Triple Whale, 2024) - https://www.triplewhale.com/blog/geolift-geo-based-incrementality-testing [source]
Geo Experiments: The Fundamentals (Haus, 2024) - https://www.haus.io/blog/geo-experiments-the-fundamentals [source]
Geo-Based Incrementality Testing Playbook for 2025 (Lifesight, 2025) - https://lifesight.io/blog/geo-based-incrementality-testing/ [source]
MMM Calibration with Lift Tests and Bayesian Methods (PyMC Labs, 2024) - https://www.pymc-labs.com/blog-posts/mmm_roas_lift [source]
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/ [source]
Post-Cookie Attribution Playbook for 2026 (GrowthMarketer, 2025) - https://growthmarketer.com/blog/post-cookie-attribution-playbook/ [source]
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 [source]
The 2025 State of Data Clean Rooms in Retail Media (Skai, 2025) - https://skai.io/blog/data-clean-rooms-in-retail-media/ [source]

Children

Adstock and carryover transformations (frontier)
Saturation and Hill curves (frontier)
Bayesian MMM and priors (frontier)
Open-source MMM frameworks (Robyn, Meridian, PyMC-Marketing) (frontier)
Multi-touch attribution vs MMM (frontier)
Incrementality testing (frontier)
Geo-lift experiments (GeoLift, TBR, CausalImpact) (frontier)
Privacy-era post-cookie measurement (frontier)
MMM calibration with experiments (frontier)

Frontier under this node: Adstock and carryover transformations, Bayesian MMM and priors, Geo-lift experiments (GeoLift, TBR, CausalImpact), Incrementality testing, MMM calibration with experiments, Multi-touch attribution vs MMM, Open-source MMM frameworks (Robyn, Meridian, PyMC-Marketing), Privacy-era post-cookie measurement, Saturation and Hill curves

← the whole tree · 3D view· how to read this page