Marketing Mix, Attribution and Budget Allocation

Introduction

Marketing investment decisions increasingly depend on an answer to a difficult question: which activities caused additional business outcomes, and how much should be invested in each activity? The question is difficult because customers encounter several channels, outcomes occur with delay, campaigns interact, and many important exposures are controlled by platforms that do not share user-level data.

Marketing attribution and marketing-mix modelling address related but different problems. Attribution generally studies journeys at the individual or event level, whereas marketing-mix models (MMMs) usually estimate aggregate effects over time and across markets. Neither method is universally superior. The appropriate method depends on the decision, data-generating process, level of aggregation, experimental capability, and acceptable assumptions.

This chapter develops an integrated approach to attribution, incrementality, marketing-mix modelling and budget allocation. Particular attention is given to carryover, saturation, interaction effects, identification, walled gardens and the managerial risks of treating measured performance as causal performance.

Learning objectives

By the end of this chapter, the reader should be able to:

1. Distinguish attribution, incrementality analysis and marketing-mix modelling.
2. Explain carryover, saturation and interaction effects in response models.
3. Compare rule-based, algorithmic and causal attribution methods.
4. Specify and interpret an aggregate marketing response model.
5. Recognize identification threats, platform measurement limitations and cross-channel bias.
6. Allocate a constrained budget using marginal returns rather than historical averages alone.
7. Design a measurement system that combines experiments, MMM and operational diagnostics.

1. The managerial problem: observed credit versus causal contribution

Suppose a customer purchases after receiving a paid-search impression. A naïve report assigns the sale to paid search. Yet the customer may already have intended to buy, may have seen television advertising earlier, or may have visited through organic search. The observed sequence is not the same as the causal process.

A useful distinction is:

• Observed outcome: what happened after an exposure.
• Attributed outcome: the amount of credit assigned by a measurement rule.
• Incremental outcome: the additional outcome caused by the activity relative to a credible counterfactual.

Incrementality is fundamentally counterfactual. For person or unit i, the incremental effect of treatment is:

Here, Y_i(1) is the outcome if the unit receives the marketing activity and Y_i(0) is the outcome if it does not. Only one potential outcome is observed for each unit, so individual incrementality is not directly observable. Experiments or credible quasi-experimental designs estimate an average treatment effect:

The expectation is over the relevant target population. The result is causal only under assumptions such as valid randomization, no material interference between units, consistent treatment definition and reliable outcome measurement.

Attribution can be useful for operational optimization, such as allocating creative or managing bids, but attribution credit should not automatically be interpreted as incremental sales.

2. Three measurement perspectives

2.1 Attribution

Attribution assigns conversion credit to touchpoints. It is often granular, fast and connected to activation platforms. Its limitations include selection bias, identity loss, cookie restrictions, incomplete exposure data and the fact that a conversion path is not a randomized experiment.

2.2 Incrementality measurement

Incrementality compares outcomes under treatment and a counterfactual. Common designs include randomized holdout tests, geo experiments, switchback tests and carefully constructed observational studies. Incrementality is usually the strongest evidence for causal contribution, but tests may be expensive, short-lived or difficult when channels overlap.

2.3 Marketing-mix modelling

MMM estimates the relationship between aggregate outcomes and marketing and non-marketing drivers over time and across geographies. It is particularly useful for television, radio, out-of-home, sponsorships and other channels where person-level exposure is unavailable. MMM supports strategic budget planning, but it requires sufficient variation, careful specification and disciplined validation.

3. Attribution methods and their assumptions

3.1 Rule-based attribution

Common rules include first-touch, last-touch, linear, time-decay and position-based attribution. Last-touch assigns all credit to the final recorded touchpoint; linear attribution divides credit equally; time-decay assigns more credit to recent events. Position-based rules assign larger shares to selected positions, such as the first and last touch.

These methods are transparent but encode arbitrary assumptions. Last-touch tends to reward channels near conversion, including branded search and retargeting. First-touch tends to reward discovery channels. Time-decay reflects a belief that recent exposures matter more, but it does not estimate a causal decay process.

3.2 Algorithmic attribution

Algorithmic methods may use Markov chains, Shapley-value decompositions or predictive models. In a Markov approach, the analyst estimates how conversion probabilities change when a channel is removed from observed paths. In a Shapley approach, credit is distributed according to the average marginal contribution of a channel across possible channel coalitions.

For a value function v(S), where S is a set of channels, the Shapley value for channel j is:

N is the set of all channels and S is a subset not containing j. The formula fairly averages channel j’s incremental contribution across orderings, provided that v(S) is meaningfully defined. In observational attribution, however, v(S) may be estimated from non-random exposure patterns. The resulting decomposition is not automatically causal.

Algorithmic attribution also faces identity resolution problems. If a user switches devices, signs in late, deletes cookies or is exposed in a walled garden, the path is incomplete. Missingness may be systematic rather than random.

3.3 Causal attribution and experiments

A randomized holdout assigns eligible units to treatment or control. If treatment assignment is independent of potential outcomes, the difference in average outcomes identifies the treatment effect. For a binary outcome, an estimated incremental conversion rate is:

where p̂_T and p̂_C are conversion rates in treatment and control groups. If the treatment group contains n_T units and x_T conversions, then p̂_T = x_T/n_T; analogous definitions apply to control.

The incremental number of conversions in a treated population of size N_T is approximately N_T × Δp̂. Incremental return on ad spend can be expressed as:

Units must be consistent: revenue and spend should use the same currency and attribution window, and revenue should be defined consistently, for example as net revenue rather than gross bookings.

Tests require attention to power, treatment contamination, seasonality, multiple testing, delayed conversion and interference. A geo test may be preferable when user-level randomization is impossible, but geographic balance and spillovers must be assessed.

4. Marketing-mix models: structure and interpretation

A basic MMM models an outcome such as sales in market m and period t as a function of baseline demand, marketing variables, price, distribution, competition, seasonality and macroeconomic conditions:

Y_{m,t} is the outcome, α_m is a market effect, δ_t is a time effect, A variables are marketing inputs, X is a vector of controls, β and γ are parameters, and ε is the unexplained error. The model assumes that the included variables adequately represent systematic demand drivers and that the functional form is suitable for the decision.

A coefficient is not automatically a causal effect. For example, marketing may increase when sales are expected to fall, creating reverse causality. Marketing may also be concentrated in high-potential markets, creating selection bias. Fixed effects remove stable market differences but do not remove time-varying confounding.

4.1 Carryover and adstock

Advertising may influence demand after the period in which it is delivered. A simple geometric adstock transforms current spend X_t into an effective stock A_t:

X_t is current-period spend or exposure, λ is the retention parameter, and A_0 is an initial stock, often set to zero or estimated. The units of A_t are the units of X_t if λ is dimensionless. A larger λ implies longer carryover. This model assumes a constant proportional decay and linear accumulation before any saturation transformation.

If the response is approximately negligible after h periods, a rough half-life is:

This is applicable for 0 < λ < 1 and measures the number of periods for a remaining effect to fall to one-half. It is a model parameter, not a universal property of a medium.

Alternative specifications include finite distributed lags and flexible decay functions. Analysts should avoid estimating highly flexible carryover curves without sufficient time variation, because carryover, trend and seasonality can be difficult to distinguish.

4.2 Saturation and diminishing returns

Incremental response commonly declines as effective exposure increases. One useful transformation is the Hill function:

A ≥ 0 is adstocked activity, s > 0 controls curve shape, and θ > 0 is the half-saturation point because f(θ) = 0.5. The function is bounded between 0 and 1. It should be multiplied by a scale parameter if the result is to represent sales rather than a normalized response.

A concave alternative is:

for A ≥ 0 and θ > 0. It is increasing and diminishing-return, with f(0) = 0. Analysts must choose a response function based on evidence and interpretability rather than convenience.

The marginal response of a channel is the derivative of its expected outcome with respect to spend. If expected incremental profit from channel j is Π_j(x_j), the local marginal return is:

When a channel saturates, MR_j falls as x_j increases. Budget allocation should therefore compare marginal returns at the proposed allocation, not merely historical average ROAS.

4.3 Interaction effects

Channels can reinforce or weaken one another. An interaction model may include:

The marginal effect of A is β_1 + β_3B_t. Thus, if β_3 > 0, the effect of A increases with B; if β_3 < 0, the channels substitute for one another. Interaction estimates require sufficient independent variation in both channels. If channels are always purchased together, the interaction is weakly identified.

5. Budget allocation under response curves

Let x_j be spend on channel j, with total budget B. A stylized optimization problem is:

Π_j may represent incremental contribution profit, not merely revenue. If functions are differentiable and the solution is interior, the optimum satisfies equal marginal returns:

μ is the shadow value of relaxing the budget constraint. In practice, fixed minimum spends, inventory limits, brand requirements, uncertainty and operational constraints may prevent equality.

Worked example

Assume two channels have estimated incremental contribution functions:

and

where x is measured in thousands of currency units and contribution is measured in thousands of currency units. The total budget is 100, so x_A + x_B = 100. The derivatives are:

and

At an interior optimum, 4/√x_A = 3/√x_B. Squaring gives 16/x_A = 9/x_B, so 16x_B = 9x_A. Combining this with x_A + x_B = 100 gives x_A = 64 and x_B = 36. The marginal returns are both 0.5 contribution units per budget unit. Estimated contribution is 8√64 + 6√36 = 64 + 36 = 100 thousand currency units.

This result is conditional on the response functions, budget range and absence of interaction. It is not evidence that the channels have identical historical ROAS.

6. Identification and validation

Identification asks whether the available data contain enough independent information to distinguish competing explanations. Key threats include:

• Confounding: a third factor affects both spend and outcome.
• Reverse causality: managers increase or decrease spend in response to expected demand.
• Multicollinearity: channels move together, making individual effects unstable.
• Trend and seasonality: sales and media rise together for unrelated reasons.
• Selection bias: high-value customers receive more marketing.
• Measurement error: spend, impressions, conversions or revenue are recorded inaccurately.
• Interference: one person’s treatment changes another person’s outcome.
• Post-treatment controls: controlling for a variable affected by marketing can remove part of the true effect.

Validation should be both statistical and managerial. Holdout periods, geo tests and pre-registered experiments can calibrate model estimates. Out-of-sample prediction checks whether a model generalizes, but predictive accuracy alone does not establish causal validity. Sensitivity analysis should vary plausible carryover, saturation, exclusion and prior assumptions.

A strong practice is triangulation: compare MMM estimates with randomized tests, platform reports, brand-lift studies and business knowledge. Disagreement is not necessarily a failure; it may reveal different estimands, populations or time horizons.

7. Walled gardens and cross-channel measurement

A walled garden is a platform that controls data, identity, exposure measurement and often the reporting methodology within its environment. It may provide highly detailed internal attribution while limiting impression-level data, user-level export or independent verification. Platform-reported conversions may therefore be accurate within a platform’s rules but incomplete or non-comparable across platforms.

Cross-channel measurement problems include:

1. Non-comparable denominators: one platform reports clicks, another reports impressions, and a third reports modeled conversions.
2. Overlapping credit: several platforms claim the same conversion.
3. Identity fragmentation: users cannot be consistently linked across devices and environments.
4. Different windows: platforms use different click and view-through lookback periods.
5. Modeled conversions: privacy restrictions require statistical estimation rather than direct matching.
6. Conversion latency: early reports understate outcomes that occur later.

Governance should establish a common conversion definition, reporting calendar, currency, attribution window and treatment of cancellations or returns. Independent experiments should be used where possible. Privacy-preserving methods—such as aggregated conversion measurement, clean rooms and secure multi-party computation—can support analysis, but they do not remove identification problems. A clean room can improve data collaboration while still containing biased or non-random exposure data.

8. An integrated measurement architecture

A practical measurement system assigns different jobs to different tools:

• Experiments: estimate causal lift for selected channels, audiences and conditions.
• MMM: estimate portfolio-level response over longer periods and include offline media.
• Attribution diagnostics: monitor journeys, creative, frequency and delivery anomalies.
• Customer-level models: estimate heterogeneous response and lifetime value, subject to valid treatment definitions.
• Finance and operations data: convert outcomes into contribution after costs, returns and service effects.

The sequence matters. First define the decision and estimand: short-run sales, long-run customer value, contribution profit or brand demand. Next audit data quality and exposure definitions. Then estimate models with explicit assumptions, validate against experiments and stress-test uncertainty. Finally translate estimates into constrained decisions with guardrails.

AI and machine learning can improve forecasting, creative classification and heterogeneous-effect estimation. However, a highly accurate response predictor can still be non-causal if it learns that high-intent customers are more likely to receive advertising. Causal forests, uplift models and Bayesian MMMs are useful only when their treatment, outcome and identification assumptions are carefully specified.

9. Illustrative case: Northstar Home, a fictional retailer

Background

Northstar Home, a fictional omnichannel retailer, sells furniture through stores and its website. It invests in paid search, social video, connected television, email and local out-of-home advertising. The finance team observes that platform-reported ROAS is highest for branded search, while the brand team believes connected television creates demand that appears later through search.

Problem

The company must allocate an additional 2 million currency units for the next quarter. Its existing last-touch report recommends increasing branded search and retargeting while reducing television. However, senior management suspects that this would harvest existing demand rather than create incremental demand.

Analysis

The analytics team defines the primary estimand as incremental contribution profit over 12 weeks, including delayed conversions. It conducts three activities:

1. A randomized paid-search holdout in eligible regions.
2. A geo-level television test with matched treatment and control markets.
3. A two-year weekly MMM including media, price, distribution, promotions, holidays, competitor activity and economic controls.

The search experiment finds a modest lift: many reported search conversions would have occurred without the ads. The television test shows a delayed lift, with the strongest effect in weeks two through five. The MMM estimates positive television carryover and diminishing returns, but the television coefficient is sensitive to assumptions about promotions. Email shows strong attributed performance but limited incremental lift among customers already receiving frequent brand communications.

The team also discovers that the social platform uses a seven-day click and one-day view window, whereas the internal report uses a thirty-day window. Direct comparison is therefore invalid.

Approach and solution

The company rejects a single “winner” metric. It uses experimental lift to calibrate channel response curves, MMM to represent offline and delayed effects, and journey reports to manage execution. The optimization model includes minimum television investment for reach, a cap on branded-search expansion, and uncertainty ranges around all response parameters.

The recommended allocation increases television moderately, maintains non-brand search for high-intent demand, reduces low-incrementality retargeting, and reserves part of the budget for a follow-up test. The plan is approved conditionally: investment is released in stages, and decisions are revisited when test evidence arrives.

Outcome and learning points

After the quarter, total contribution improves relative to the prior planning baseline, although the realized result is below the model’s central estimate. The shortfall is traced to a promotion change and an overestimated television response in one market. Because the plan included uncertainty and staged investment, the error is contained.

The case illustrates five principles: attributed credit is not incrementality; delayed effects matter; platform windows must be normalized; model uncertainty should affect allocation; and measurement is a learning system rather than a one-time report.

10. Best practices for managers and analysts

1. Define the causal estimand before selecting a tool.
2. Use contribution profit or customer value when revenue is an incomplete objective.
3. Separate measurement windows and conversion definitions across platforms.
4. Include carryover, saturation and plausible interactions, but avoid unnecessary flexibility.
5. Design experiments before spending heavily when a decision is material.
6. Report confidence or credible intervals, not only point estimates.
7. Treat platform-reported attribution as directional unless independently validated.
8. Conduct sensitivity analysis for missing channels, identity loss and delayed outcomes.
9. Record allocation constraints, business rules and unmeasured interventions.
10. Re-estimate models as consumer behavior, privacy rules and media technology change.

Ethical, regulatory and governance considerations

Marketing measurement uses personal and behavioral data, so lawful basis, purpose limitation, data minimization, security and retention controls are essential. Consent and opt-out choices must be respected, and modeled measurement should not be used to recreate prohibited individual profiles. Fairness audits are important when optimization systematically excludes protected groups or directs costly offers toward vulnerable consumers.

Governance should document data provenance, model versions, assumptions, experiment protocols, platform contracts and decision overrides. Managers should be able to explain why a budget changed and what evidence supported the change. Transparent uncertainty is preferable to false precision.

Chapter Summary

Attribution assigns credit; incrementality estimates causal contribution; and marketing-mix modelling explains aggregate variation across time and markets. Rule-based and algorithmic attribution can support operational decisions but are not automatically causal. Experiments provide stronger identification when treatment and counterfactual conditions can be controlled. MMM extends measurement to offline channels, delayed effects and privacy-constrained environments, but it is vulnerable to confounding, aggregation and weak variation.

Carryover can be represented with adstock, saturation with bounded or concave response functions, and interaction with product terms. Budget allocation should maximize incremental contribution subject to constraints, generally moving investment toward channels with higher current marginal returns rather than higher historical average ROAS. Walled gardens, inconsistent windows, identity fragmentation and modeled conversions make cross-channel comparison difficult. The strongest managerial practice is triangulation: combine experiments, MMM, operational attribution and financial analysis, while documenting assumptions and uncertainty.

Key Terms

Incrementality: The additional outcome caused by a marketing activity relative to a counterfactual.

Attribution: Assignment of conversion or revenue credit to marketing touchpoints.

Counterfactual: The outcome that would have occurred under an alternative treatment condition.

Average treatment effect: The average difference between treated and untreated potential outcomes.

Marketing-mix model: An aggregate statistical model linking outcomes to marketing and contextual drivers.

Adstock: A transformation representing the persistence or carryover of marketing effects.

Carryover: Influence of prior-period marketing on current outcomes.

Saturation: Declining incremental response as marketing activity increases.

Marginal return: Additional objective value generated by one additional unit of investment.

Interaction effect: A condition in which the effect of one channel depends on another variable or channel.

Identification: The ability to distinguish a causal effect from alternative explanations using available data.

Confounding: Distortion caused by a variable related to both marketing exposure and outcome.

Multicollinearity: Strong correlation among predictors that makes separate effects difficult to estimate.

Walled garden: A platform ecosystem that controls access to data, identity and measurement.

Geo experiment: A randomized or quasi-randomized test using geographic units.

Holdout test: A design in which a control group is intentionally withheld from treatment.

Shapley value: A cooperative-game allocation of average marginal contribution across channel coalitions.

ROAS: Revenue divided by advertising spend.

iROAS: Incremental revenue divided by marketing spend.

Response curve: A function describing expected outcome as investment changes.

Review Questions

5 Short Answer Questions

1. Distinguish attributed conversions from incremental conversions.
2. What managerial assumption is embedded in last-touch attribution?
3. Explain the purpose of an adstock transformation.
4. Why can a platform’s reported ROAS be non-comparable with another platform’s ROAS?
5. What does an interior optimum in budget allocation imply about marginal returns?

5 Long Answer Questions

1. Compare rule-based attribution, algorithmic attribution, experimentation and MMM, including the principal assumptions and limitations of each.
2. Explain how carryover and saturation affect the interpretation of channel coefficients and budget recommendations.
3. Discuss the major identification threats in an MMM and propose practical validation methods.
4. Explain why walled gardens create cross-channel measurement problems and design governance responses.
5. Develop an integrated measurement architecture for an omnichannel retailer and justify the role of each component.

3 Analytical / Application Questions

1. A company observes high last-touch ROAS for branded search but finds that a randomized holdout produces only a small lift. Explain at least three reasons for the discrepancy and recommend a budget decision process.
2. Two channels have response curves with diminishing returns, but one channel also increases the effectiveness of the other. Specify how an interaction term could change allocation and identify the evidence needed before acting on it.
3. Design a geo experiment for a television campaign when user-level exposure data are unavailable. State the treatment, control, outcome, timing, balance checks and principal threats to validity.

Multiple Choice Questions with Answers

Question 1

Which measure most directly represents causal contribution?

A. Last-touch attributed revenue

B. Incremental outcome estimated against a credible counterfactual

C. Total platform-reported conversions

D. Number of recorded impressions

Correct Answer: B. Incremental outcome estimated against a credible counterfactual

Question 2

In the adstock equation A_t = X_t + λA_t−1, what does a larger λ generally imply?

A. Shorter carryover

B. Longer carryover

C. Lower current-period spend

D. Automatic channel interaction

Correct Answer: B. Longer carryover

Question 3

In a Hill saturation function, θ is commonly interpreted as:

A. The advertising budget constraint

B. The error variance

C. The half-saturation level

D. The number of channels

Correct Answer: C. The half-saturation level

Question 4

Why can last-touch attribution overvalue branded search?

A. It ignores all conversion events

B. It assigns credit only to offline media

C. It tends to reward touchpoints close to conversion

D. It always uses randomized assignment

Correct Answer: C. It tends to reward touchpoints close to conversion

Question 5

If two channels are always purchased together, which estimation problem is especially likely?

A. Perfect randomization

B. Weak identification due to limited independent variation

C. Guaranteed negative interaction

D. Elimination of seasonality

Correct Answer: B. Weak identification due to limited independent variation

Question 6

At an interior optimum with separable response functions and a binding budget, the usual condition is that:

A. Historical ROAS is identical across channels

B. Average spend is identical across channels

C. Marginal returns are equal across active channels

D. The largest channel receives the entire budget

Correct Answer: C. Marginal returns are equal across active channels

Question 7

A randomized holdout test primarily helps estimate:

A. The causal lift of treatment relative to control

B. The platform’s preferred attribution window

C. The number of impressions purchased

D. The historical average ROAS only

Correct Answer: A. The causal lift of treatment relative to control

Question 8

Which is a characteristic problem of walled gardens?

A. They always provide complete cross-platform identity

B. They eliminate modeled conversions

C. They may restrict independent access to exposure and identity data

D. They guarantee comparable attribution windows

Correct Answer: C. They may restrict independent access to exposure and identity data

Question 9

In the interaction model Y_t = β₀ + β₁A_t + β₂B_t + β₃(A_t × B_t) + ε_t, what does β_3 describe?

A. The intercept

B. The effect of A when B is zero only

C. How the effect of A changes with B

D. The residual outcome

Correct Answer: C. How the effect of A changes with B

Question 10

Which statement about predictive accuracy is correct?

A. High predictive accuracy automatically proves causality

B. Prediction and causal identification are identical

C. A model can predict well while learning non-causal exposure patterns

D. Predictive models cannot be used in marketing

Correct Answer: C. A model can predict well while learning non-causal exposure patterns