Bayesian Media Mix Modeling - Overview

Summary

Media mix models (MMM) are regression models advertisers use to measure media effectiveness and guide budget allocation. Jin, Wang, Sun, Chan & Koehler (Google, 2017) propose an MMM with flexible functional forms for two phenomena linear regression cannot capture — carryover (advertising’s lagged effect) and shape (saturation / diminishing returns) — estimated in a Bayesian framework so prior knowledge can compensate for the low information content of a single MMM dataset. Key finding: the model recovers parameters well on large data, but for typical small samples (a couple of years of weekly data) the priors dominate and estimates can be biased; the optimal media mix derived from the model has large variance and must be trusted cautiously.

Overview

An MMM relates aggregated sales (weekly or monthly, national or geo-level) to media spend across channels plus control variables (price, distribution, seasonality, macro factors). It descends from the marketing “4Ps” tradition (Borden 1964; McCarthy 1978) and is fundamentally a regression that infers causation from observational correlation. Randomized experiments across media are expensive and rarely feasible, so observational regression remains the workhorse despite its causal fragility (see Activity Bias in Advertising for one such confound).

Two well-documented features of advertising response break the classic linear decision model (Guadagni & Little 1983):

1. Carryover / lag effect — a portion of an ad’s impact occurs in periods after the exposure (delayed consumer response, inventory effects, word-of-mouth). Modeled via the adstock transformation. See Carryover (Adstock) Functional Forms.
2. Shape / saturation effect — response is not linear in spend; high spend yields diminishing returns (the “shape effect”, Tellis 2006). Modeled via a curvature function. See Shape (Saturation) Effects.

Because these transformations make the model nonlinear in the parameters, ordinary least squares / MLE is awkward, and the paper turns to Bayesian estimation via MCMC (see Bayesian Estimation and Priors for MMM). The Bayesian framing is motivated less by philosophy than by data scarcity: as Chan & Perry (2017) note, the information content within a single MMM dataset is low relative to the number of parameters, so priors drawn from industry experience or prior/related media-mix models are essential.

Main Content

The MMM regression equation (combined model)

Data form (Sec. 2 intro, p. 2). Citing Chan & Perry (2017), the paper notes that MMM data come in several forms: daily or weekly, national or geo-level. Weekly national data is the most commonly used, and the paper’s model is written for it. Geo-level data reappear in Sec. 9 as a way to get more information (Sun et al. 2017). The paper does not try to choose the control variables, since they vary by advertiser and industry; it proposes a general model and estimation method for any MMM.

For weekly national data over weeks $t=1,\dots ,T$, with $M$ media channels and $C$ control variables, the response (sales, or log-sales) $y_t$ is modeled as

$$y_t=\tau +\sum _{m=1}^M{\beta }_m\text{Hill}\left(x_{t,m}^{\ast };{\mathcal{K}}_m,{\mathcal{S}}_m\right)+\sum _{c=1}^C{\gamma }_c{z}_{t,c}+{ϵ}_t,$$

where

• $x_{t,m}^{\ast }=\text{adstock}(x_{t-L+1,m},\dots ,x_{t,m};w_m,L)$ is the carryover-transformed spend of channel $m$ (Eq. 1, see Carryover (Adstock) Functional Forms);
• $\text{Hill}(\cdot )$ is the shape/saturation transform with shape (slope) ${\mathcal{S}}_m$ and half-saturation ${\mathcal{K}}_m$ (see Shape (Saturation) Effects);
• ${\beta }_m\ge 0$ is the regression coefficient (maximum effect) of channel $m$;
• $\tau$ is baseline sales (intercept), ${\gamma }_c$ the effect of control variable $z_{t,c}$;
• ${ϵ}_t$ is white noise, uncorrelated, constant variance.

Media effects are assumed additive (no synergy/interaction between channels — a simplification, cf. Zhang & Vaver 2017). Carryover is applied before shape (adstock then Hill), which is appropriate when per-period spend is small relative to cumulative spend.

Why Bayesian

Bayesian inference treats parameters $\Phi$ as random variables with a posterior $p(\Phi \mid \mathbf{y},\mathbf{X})\propto \mathcal{L}(\mathbf{y}\mid \mathbf{X},\mathbf{Z},\Phi )\pi (\Phi )$. The prior $\pi (\Phi )$ injects external knowledge to offset the weak signal in a single dataset, and the full posterior (not just a point estimate) supplies credible intervals and propagates parameter uncertainty into downstream attribution metrics (ROAS, mROAS, optimal mix).

Examples

Shampoo advertiser case study (preview)

The model is applied to 2.5 years of weekly volume-sales data for a shampoo advertiser (TV, magazines, display, YouTube, search), with price/distribution/promotion as controls. Four functional-form specifications are compared by BIC; the most parsimonious (geometric adstock + reach transformation, BIC 51.49 vs 98.12 for the most flexible) wins on the complexity penalty alone, since the negative log-likelihoods differ by at most 0.06. The posterior of the retention rate reproduces its prior, the Hill and reach models disagree about 3- to 4-fold on magazines mROAS, and the optimal TV/magazine split has spikes at both a 0% and a 100% TV share with the posterior mean (about 0.47) between them — the data cannot reliably guide allocation. Residual autocorrelation (about 0.6 at lag 1) remains to lag 15. Full treatment, with every table and figure reading, in Jin 2017 MMM - Shampoo Case Study Data, Diagnostics and Posteriors; the BIC concept summary is in MMM Model Selection and Application.

Connections

• Carryover generalizes the classic distributed-lag idea: the geometric-decay adstock weight is the continuous/marketing analog of the Koyck lag. See Carryover (Adstock) Functional Forms and Carryover Effects and Distributed Lags.
• Shape corresponds to the concave-vs-S-shaped response debate in Shape of the Marketing Response Function.
• Empirical advertising-elasticity generalizations (~0.10 short-run) provide priors and sanity checks for ${\beta }_m$; see Advertising and Promotion Effects.
• Bayesian/MCMC machinery: Bayesian Estimation and Priors for MMM, MCMC Basics, Bayesian Linear Regression.

Source Map

Added by the 2026-10-13 coverage audit. Source: Jin, Wang, Sun, Chan & Koehler, “Bayesian Methods for Media Mix Modeling with Carryover and Shape Effects”, Google Inc., 14 April 2017, 34 pp. Every section, equation, table and figure maps to a note:

Source location	Content	Note
Abstract, Sec. 1 (pp. 1-2)	MMM definition, 4Ps, causation-from-correlation caveat, carryover and shape motivation, paper roadmap	this note; scope limits in Jin 2017 MMM - Conclusions, Limitations and Extensions
Sec. 2 intro (p. 2)	data forms per Chan & Perry (daily/weekly, national/geo; weekly national most common), notation $x_{t,m}$, $z_{t,c}$, controls not selected	this note
Sec. 2.1, Eqs. 1-3, Fig. 1 (pp. 3-4)	adstock, geometric decay, delayed adstock, choosing $L$ (common vs per-medium, a properly chosen $L$ helps, large $L$ approximates infinity), $L=13$	Carryover (Adstock) Functional Forms (choosing $L$ ^choosing-L; Fig. 1 weights table ^figure1; delayed does not nest geometric ^not-nested)
Sec. 2.2, Eqs. 4-6, Fig. 2, footnotes 1-2 (pp. 3-5)	Hill, $\beta$Hill, reach transform, identifiability examples, reach/GRP definitions	Shape (Saturation) Effects
Sec. 2.3, Eq. 7 (pp. 5-6)	adstock-then-shape order, additive model	this note; Jin 2017 MMM - Conclusions, Limitations and Extensions
Sec. 3, Eqs. 8-9, footnote 3 (pp. 6-7)	MLE vs posterior, Gibbs/slice vs Stan/HMC, prior rationale, support, Sec. 3 citation map	Bayesian Estimation and Priors for MMM; Jin 2017 MMM - STAN Sampler Code and MCMC Implementation
Sec. 4.1, Eqs. 10-11, Fig. 3 (pp. 7-9)	ROAS, mROAS, change/pre/post periods, why restrict the window (^window-rationale), perturbation-spread caveat, MLE plug-in, posterior plug-in rule	ROAS, mROAS, and Optimal Media Mix; Fig. 3 in Jin 2017 MMM - Simulated ROAS, mROAS and Optimization Results
Sec. 4.2, Eqs. 12-16 (pp. 9-10)	constrained optimal mix, flight pattern ($M$ free parameters; multiplicative alternative), two Bayesian approaches, extension to >2 channels	ROAS, mROAS, and Optimal Media Mix; algorithm in Jin 2017 MMM - Simulated ROAS, mROAS and Optimization Results
Sec. 5.1, Eq. 17, Tables 1-3, Fig. 4 (pp. 10-13)	simulated DGP, generating parameters, signal strength, priors, sampler timing	Jin 2017 MMM - Simulation Design, Signal Strength and Priors
Sec. 5.2, Figs. 5-6 (pp. 13-15)	simulated ROAS/mROAS recovery, optimisation scenarios I/II, figure readings	Jin 2017 MMM - Simulated ROAS, mROAS and Optimization Results (^fig5, ^fig5-readings, ^fig6)
Sec. 6, Table 4, Figs. 7-8 (pp. 14-19)	500-replicate sample-size study, two vs sixty years, curve underestimation, per-panel Fig. 7 readings, Fig. 8 scales, multi-prior bias mechanism	Jin 2017 MMM - Sample Size vs Bias Study (Two vs Sixty Years) (arms in rows, about 104 vs 3,120 weeks, scenario-dependent required size and Stan run-time link ^sixty-year-rows; ^table4, ^fig7, ^fig7-panels, ^fig8, ^bias-mechanism)
Sec. 7.1, Table 5, Figs. 9-10 (pp. 19-20)	prior on $\beta$ sensitivity, Fig. 10 panels	Jin 2017 MMM - Prior Sensitivity Experiments on Beta and Half-Saturation; Jin 2017 MMM - Prior Sensitivity Experiments on Beta and Half-Saturation
Sec. 7.2, Table 6, Figs. 11-12 (pp. 20-22)	prior on $K$ sensitivity, $K$ inside vs outside data range, zero prior mass on $K=2$, Fig. 11 densities, Fig. 12 panels	Jin 2017 MMM - Prior Sensitivity Experiments on Beta and Half-Saturation; Jin 2017 MMM - Prior Sensitivity Experiments on Beta and Half-Saturation; Jin 2017 MMM - Prior Sensitivity Experiments on Beta and Half-Saturation
Sec. 8, Table 7, Eq. 18, Table 9 (pp. 22-24)	four specifications, BIC selection	MMM Model Selection and Application; footnotes 5-7 in Jin 2017 MMM - Shampoo Case Study Data, Diagnostics and Posteriors
Sec. 8, Table 8, Fig. 13, footnote 4 (pp. 22-23)	shampoo data and provenance, priors “similar” to Table 3, control/media correlations, orthogonalisation, run times	Jin 2017 MMM - Shampoo Case Study Data, Diagnostics and Posteriors
Sec. 8, Figs. 14-17 (pp. 24-27)	per-model ROAS/mROAS with approximate medians, TV/magazine optimisation with mode/line positions, posterior vs prior with $\alpha$/$\theta$ densities and $\beta$Hill medians, residual and sales ACF by lag	Jin 2017 MMM - Shampoo Case Study Data, Diagnostics and Posteriors (^fig14, ^fig15, ^fig16, ^fig17)
Sec. 9 (pp. 27-28)	conclusions, hierarchical extensions, richer simulators	Jin 2017 MMM - Conclusions, Limitations and Extensions
Appendix: Stan code (pp. 28-30)	full Stan program; code map incl. geometric-adstock change and ec upper bound	Jin 2017 MMM - STAN Sampler Code and MCMC Implementation; Jin 2017 MMM - STAN Sampler Code and MCMC Implementation
Acknowledgment (p. 28)	thanks to colleagues for support, comments and discussions	Jin 2017 MMM - Conclusions, Limitations and Extensions
Appendix: Fig. 18 (pp. 31-32)	single-dataset posteriors, panels (a)-(f) read against Table 1 truths, sampler unstated	Jin 2017 MMM - Simulation Design, Signal Strength and Priors
References (pp. 33-34)	all 31 references	full reference map below (^reference-map)

Reference Map (pp. 33-34)

Every reference in the paper, where it is used, and where the vault covers it.

Reference	Used in the paper for	Vault note
Bertsekas (2014) Constrained Optimization and Lagrange Multiplier Methods	solving the budget-constrained optimum, Eqs. 12-13 (Sec. 4.2)	ROAS, mROAS, and Optimal Media Mix
Borden (1964); McCarthy (1978)	marketing-mix / 4Ps origin (Sec. 1)	this note, Overview
Carpenter et al. (2016) Stan	HMC sampler (Sec. 3, Appendix)	Jin 2017 MMM - STAN Sampler Code and MCMC Implementation
Chan & Perry (2017)	low information content of MMM data; noisy real data (Secs. 1, 5.1)	Chan Perry 2017 - Challenges in Media Mix Modeling - Overview
Chen et al. (2017) bias correction for paid search	causal-reliability issues the paper does not address (Sec. 1, p. 2)	Jin 2017 MMM - Conclusions, Limitations and Extensions
Friedman, Hastie & Tibshirani (2009), p. 212	radial-kernel reading of delayed adstock (Sec. 2.1)	Carryover (Adstock) Functional Forms
Gelfand & Smith (1990); Geman & Geman (1984)	Gibbs sampling (Sec. 3)	Jin 2017 MMM - STAN Sampler Code and MCMC Implementation; MCMC Basics
Gelman (2006) “Prior distribution”, pp. 1634-1637	priors with the same support give similar posteriors when the data are informative; the prior dominates when they are not (Sec. 3)	Bayesian Estimation and Priors for MMM; Influence of Likelihood and Prior
Gelman & Rubin (1992)	$\hat{R}$ convergence check, reported as equal to or close to 1 for all shampoo models (Table 8 caption)	Jin 2017 MMM - Shampoo Case Study Data, Diagnostics and Posteriors; What to Do About Convergence Problems
Gesztelyi et al. (2012); Hill (1910)	pharmacological origin of the Hill function (Sec. 2.2)	Shape (Saturation) Effects
Guadagni & Little (1983)	linear decision models of marketing (Sec. 1)	this note, Overview
Hanssens, Parsons & Schultz (2003) Market Response Models	geometric decay as the standard carryover form; negative-binomial delayed alternative (Sec. 2.1)	Carryover (Adstock) Functional Forms; Market Response Models - Overview; Carryover Effects and Distributed Lags
Hartigan (2012) Bayes Theory; Hoff (2009) A First Course in Bayesian Statistical Methods	quantile-based credible intervals (Sec. 3)	Jin 2017 MMM - STAN Sampler Code and MCMC Implementation
Hoffman & Gelman (2014) NUTS	HMC (Sec. 3)	HMC and Stan in Practice
Jin, Shobowale, Koehler & Case (2012)	reach-vs-GRP curve, Eq. 6 (Sec. 2.2)	Shape (Saturation) Effects
Lehmann & Casella (1998), p. 449	variance and asymptotics of the MLE, Eq. 8 (Sec. 3)	Jin 2017 MMM - STAN Sampler Code and MCMC Implementation
Neal (2003) slice sampling; Scott (2016) BOOM	custom Gibbs sampler; Fig. 8 plotting (Secs. 3, 6)	Jin 2017 MMM - STAN Sampler Code and MCMC Implementation
Rosenbaum & Rubin (1983)	propensity-score methods, hard to use with the little data an MMM has (Sec. 1, p. 2)	Jin 2017 MMM - Conclusions, Limitations and Extensions
Schwarz (1978)	BIC, Eq. 18 (Sec. 8)	MMM Model Selection and Application; Overfitting and Information Criteria
Spiegelhalter, Best, Carlin & van der Linde (2002)	DIC as an alternative to BIC (Sec. 8, p. 23)	Overfitting and Information Criteria; Model Comparison
Vehtari, Gelman & Gabry (2015); Watanabe (2010)	WAIC / PSIS-LOO as alternatives to BIC (Sec. 8, p. 23). The paper adds that all these criteria pick the best regression model, “not the model with the most accurate causal inference”	Model Selection Using Predictive Performance; Overfitting and Information Criteria
Tellis (2006) “Modeling marketing mix”	names the carryover and shape problems that the paper sets out to address (Secs. 1-2)	this note, Overview; Shape of the Marketing Response Function
Wang et al. (2017); Sun et al. (2017)	hierarchical category-level and geo-level extensions (Secs. 3, 9)	Jin 2017 MMM - Conclusions, Limitations and Extensions; Hierarchical Models
Zhang & Vaver (2017) AMSS	synergy and cross-channel effects; simulation-based evaluation (Secs. 2.3, 4.1, 9)	Zhang Vaver 2017 - Aggregate Marketing System Simulator - Overview
^reference-map

Headline validation result (Sec. 6, Table 4)

With two years of weekly data and correctly located, weakly informative priors, the fitted $\beta$Hill response curves are underestimated by 18-40% for every channel. With sixty years the bias drops to between -0.2% and +10.4%. The authors note that sixty years is a simulation device and not a practical option. Details: ^table4.

See Also

• Jin 2017 MMM - Simulation Design, Signal Strength and Priors
• Jin 2017 MMM - Simulated ROAS, mROAS and Optimization Results
• Jin 2017 MMM - Sample Size vs Bias Study (Two vs Sixty Years)
• Jin 2017 MMM - Prior Sensitivity Experiments on Beta and Half-Saturation
• Jin 2017 MMM - Shampoo Case Study Data, Diagnostics and Posteriors
• Jin 2017 MMM - STAN Sampler Code and MCMC Implementation
• Jin 2017 MMM - Conclusions, Limitations and Extensions
• Carryover (Adstock) Functional Forms
• Shape (Saturation) Effects
• Bayesian Estimation and Priors for MMM
• ROAS, mROAS, and Optimal Media Mix
• MMM Model Selection and Application
• Index: Bayesian Media Mix Modeling
• Hierarchical Models — geo-level and brand-level MMMs pool channel parameters through hierarchical priors
• Model Validation and Interpretability - Overview — validation checklist for MMMs and driver rankings
• Abowd 2022 - Conclusion, Limitations and Marketing Relevance — DP-noised KPIs as measurements with known variance