Attribution, Marketing Mix and Budget Decisions

Introduction

Digital marketing creates many observable interactions: an impression, a search click, a product-page visit, an email open, a remarketing exposure, and finally a purchase. The central managerial problem is not simply to count these interactions, but to decide which activities contributed incrementally to business outcomes and how the next unit of budget should be allocated.

Attribution, marketing mix modelling (MMM), and incrementality analysis address this problem from different perspectives. Attribution assigns credit across customer journeys. MMM estimates the relationship between aggregate marketing investment and business outcomes over time. Incrementality asks the strongest causal question: what additional outcome occurred because of a particular intervention?

No single method is universally correct. Attribution is detailed but vulnerable to tracking gaps and selection bias. MMM is useful for strategic budget allocation and offline channels but may be too aggregated for individual journeys. Experiments can provide strong causal evidence but may be expensive, operationally difficult, or limited to particular contexts. Professional decision-making therefore combines methods rather than treating one model as a perfect account of marketing effectiveness.

Learning Objectives

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

1. Explain the purpose and assumptions of marketing attribution.
2. Compare first-touch, last-touch, linear, position-based, and data-driven attribution.
3. Identify common attribution biases and measurement failures.
4. Calculate and interpret marginal return on advertising spend.
5. Explain how MMM and incrementality studies support budget decisions.
6. Recommend a practical measurement and allocation approach for a digital marketing organisation.

1. Attribution as a Credit-Assignment Problem

Attribution is the process of assigning some share of a conversion, revenue amount, or customer value to marketing touchpoints. A conversion may be a purchase, lead, subscription, app installation, or other defined outcome. A touchpoint is an observable interaction between a customer and a marketing activity.

Suppose a customer sees a social advertisement, clicks a search advertisement, receives an email, and then purchases. A single-conversion attribution model may assign the entire conversion to one touchpoint or divide credit among several touchpoints. The assigned credit is not automatically the same as causal impact. A touchpoint may receive credit because it occurred before a purchase, even though the purchase would have happened without it.

For a set of conversions, an attribution system should ideally satisfy several properties:

• Coverage: relevant interactions and conversions are recorded.
• Consistency: the same rules are applied across comparable journeys.
• Non-duplication: total assigned credit does not exceed the conversion value.
• Actionability: results can inform channel, campaign, and budget decisions.
• Validity: assigned credit reflects, as far as possible, incremental contribution rather than mere correlation.

If conversion value is V and the model assigns weights w₁, w₂, …, wₙ to n touchpoints, a conservation rule is:

The credit assigned to touchpoint i is:

Here, Cᵢ is credit in currency units and wᵢ is dimensionless. This formulation assumes that the conversion value is fully distributed among the observed touchpoints. It does not prove that the touchpoints caused the conversion.

2. Rule-Based Attribution Models

2.1 First-touch attribution

First-touch attribution assigns 100% of the credit to the first recorded marketing interaction in the conversion path. It is useful when the strategic question concerns demand creation or customer acquisition.

For a path whose first touchpoint is j:

Strengths: simple, easy to explain, and useful for identifying channels that introduce customers to a brand.

Limitations: it ignores nurturing, evaluation, conversion assistance, and retention activity. It can overvalue broad-reach channels and understate lower-funnel activity.

2.2 Last-touch attribution

Last-touch attribution assigns all credit to the final recorded touchpoint before conversion. It is common because it is operationally simple and aligns with many platform reporting systems.

Strengths: straightforward reporting; useful for assessing immediate conversion capture.

Limitations: it often rewards branded search, direct traffic, remarketing, or email reminders that occur after demand has already been created. It can encourage excessive investment in channels that harvest existing intent.

2.3 Linear attribution

Linear attribution distributes credit equally among all recorded touchpoints. If a journey has n touchpoints, then:

For a $120 conversion with four recorded touchpoints, each receives $30 of attributed credit. This is transparent and avoids choosing a single supposedly decisive interaction.

Its weakness is that equal credit is an assumption, not an empirical finding. The model treats an awareness impression, a product demonstration, and a final checkout reminder as equally influential, even when their roles differ substantially.

2.4 Position-based attribution

Position-based attribution, often called a U-shaped model, gives relatively high weights to the first and last touchpoints and distributes the remainder across intermediate interactions. A common configuration assigns 40% to the first touch, 40% to the last touch, and 20% equally among middle touches.

For n > 2 touchpoints:

For a two-touch journey, the first and last touchpoints may each receive 50%, because there are no middle interactions. The chosen percentages are conventions, not universal truths.

The model recognises acquisition and conversion roles, but it remains rule-based. It cannot establish whether the first or last interaction caused additional demand.

2.5 Time-decay models

Although not required in the objective, time-decay attribution is a useful related model. It assigns greater credit to touchpoints closer to conversion. A typical formulation uses an exponential decay function:

where Δtᵢ is the time between touchpoint i and conversion, λ > 0 is a decay parameter, and e is the base of natural logarithms. The weights are normalised so that they sum to one. The model assumes recency is a reasonable proxy for influence, which may not hold in long-consideration purchases.

3. Data-Driven Attribution

Data-driven attribution estimates credit from observed journey data rather than applying fixed rules. Depending on the implementation, it may use logistic regression, survival analysis, Markov chains, Shapley-value methods, causal models, or machine-learning algorithms.

A simple removal-effect approach asks how predicted conversion outcomes change when a channel is removed from observed paths. If a channel's removal causes a substantial reduction in predicted conversions, the channel may receive greater credit. However, removal effects depend on the data, model specification, path definition, and assumptions about interactions among channels.

A Shapley-value approach originates in cooperative game theory. It evaluates each channel's average marginal contribution across possible combinations of channels. In simplified form, the Shapley value for channel i is:

Here, N is the set of channels, S is a subset not containing i, v(S) is the value associated with that subset, and |S| is its size. The formula assumes that a suitable value function v has been defined and estimated. In practice, exact calculation becomes expensive as the number of channels grows, so approximations are often used.

Data-driven methods can model interaction effects and heterogeneous journeys, but they require sufficient, representative, and privacy-compliant data. They are not automatically causal. If a high-intent customer is more likely to click search advertising, a predictive model may assign search substantial credit even when the click added little incremental demand.

Comparison of attribution models

4. Attribution Bias and Measurement Failure

Attribution bias occurs when the measured credit systematically differs from the true incremental contribution of an activity. Important sources include the following.

Selection bias: customers who are already likely to buy may be more likely to interact with advertising. A search click may be associated with conversion because of existing intent, not because the click caused the purchase.

Survivorship bias: analysis may focus only on customers who reached conversion, excluding people exposed to marketing who did not buy. This makes it difficult to estimate what prevented or caused conversion.

Last-click bias: final-touch models favour channels that appear late in the journey, particularly branded search, retargeting, and CRM reminders.

Cookie and identity bias: deleted cookies, browser restrictions, multiple devices, logged-out activity, and consent choices fragment journeys. The observed path may be incomplete.

Platform bias: an advertising platform has an incentive to report credit for its own impressions and clicks. Comparing platform-reported conversions without deduplication can double-count outcomes.

Exposure bias: impressions are not equivalent to meaningful attention. A served advertisement may be unseen, technically viewable but ignored, or seen by a customer already planning to purchase.

Window bias: a seven-day click window and a one-day view window produce different reported outcomes. Attribution windows should reflect the buying cycle and be applied consistently.

Aggregation bias: customer-level results may not generalise to regions, products, or periods with different demand conditions.

Brand-demand bias: branded search and direct traffic often capture demand generated by other channels, public relations, word of mouth, or prior customer experience.

A useful diagnostic is to compare attributed performance with controlled lift. If a channel reports a 10% conversion share but a holdout test shows only a 2% incremental lift, the difference signals possible attribution inflation or a mismatch between measured and causal outcomes.

5. Marginal Returns and Budget Allocation

Budget decisions should focus on the next unit of spend, not merely on average historical performance. Average return on advertising spend is:

where R is attributed or incremental revenue and S is advertising spend. ROAS is dimensionless, commonly expressed as a ratio such as 4.0:1.

Marginal ROAS measures the additional revenue generated by an additional amount of spend:

where ∆R is the change in revenue and ∆S is the change in spend. For profit-based decisions, contribution margin should be considered. If m is the contribution margin rate and variable costs are excluded from R, marginal contribution is m∆R. A spend increase is financially attractive when:

where ∆F represents any additional fixed or operational cost associated with the activity. If ∆F = 0, the break-even marginal ROAS is:

For a 40% contribution margin, the break-even marginal ROAS is 1 ÷ 0.40 = 2.5. A marginal ROAS of 3.0 generates $3 of revenue per $1 spent and $1.20 of contribution before advertising cost; after $1 advertising cost, incremental contribution is $0.20 per dollar.

Diminishing returns

Marketing channels commonly exhibit diminishing returns: early spending reaches highly responsive audiences, while later spending reaches less responsive audiences or generates frequency without equivalent action. A simple response function is:

where R(S) is expected incremental revenue, α > 0 is a saturation ceiling, β > 0 controls responsiveness, and S ≥ 0 is spend. Its marginal return is:

This derivative declines as S increases. The function is illustrative rather than universal; real channels may show thresholds, carryover, interaction, or saturation at different rates.

A practical allocation rule is to move budget toward the channel with the highest credible marginal contribution, subject to constraints such as inventory, reach, cash flow, brand objectives, legal requirements, and minimum viable spend. The rule should use incremental contribution rather than platform-attributed revenue whenever possible.

6. Marketing Mix Modelling

MMM is a statistical approach that relates aggregate business outcomes to marketing investment and other explanatory variables over time. It is especially useful where user-level tracking is incomplete or unavailable, and where offline media, pricing, distribution, seasonality, and macroeconomic factors matter.

A simplified model may be written as:

where Yₜ is the outcome in period t, Xⱼₜ is marketing activity for channel j, βⱼ is its estimated association with the outcome, Zₜ is a vector of control variables such as price, promotions, seasonality, or economic conditions, γ is a vector of control coefficients, and εₜ is unexplained error.

Marketing effects often persist beyond the period of exposure. Adstock models represent carryover. One simple recursive form is:

where Aₜ is the transformed advertising stock, Xₜ is current activity, and θ is the retention parameter. This assumes geometric decay and a stable carryover process. If θ = 0, there is no carryover; as θ approaches 1, effects persist longer.

Saturation transformations may then be applied to Aₜ. A common Hill-type function is:

where s > 0 controls curve shape and k > 0 is the half-saturation point: f(k) = 0.5. Such transformations should be estimated and validated rather than chosen merely because they fit a desired narrative.

MMM can estimate contribution, response curves, and scenario outcomes. Its weaknesses include confounding, limited sample size, unstable coefficients when channels move together, and sensitivity to model specification. It usually needs sufficient time periods and variation in spend. A model cannot reliably distinguish channels whose investments always rise and fall together without additional identifying information.

7. Incrementality and Causal Measurement

Incrementality is the difference between observed outcomes under marketing and the outcomes that would have occurred without the marketing intervention. The latter is the counterfactual and cannot be observed directly for the same unit at the same time.

For a treatment group and a comparable control group, an experimental lift estimate is:

where Yₜ and Y꜀ are outcome totals for treatment and control groups, and Nₜ and N꜀ are their respective populations. This assumes random assignment or sufficiently credible comparability, consistent measurement, and no substantial interference between groups.

If the treatment conversion rate is 6% and the control rate is 5%, absolute lift is 1 percentage point. Relative lift is:

The distinction matters: absolute lift is measured in percentage points, while relative lift is a proportion of the control rate.

Common designs include:

• Randomised holdout tests: eligible users, geographic areas, or accounts are randomly assigned to treatment or control.
• Geo experiments: regions receive different marketing treatments, with pre-period checks and adjustments for differences.
• Switchback tests: treatments alternate over time, useful where geographic randomisation is difficult.
• Difference-in-differences: compares treatment and control changes before and after an intervention; it requires a credible parallel-trends assumption.
• Synthetic controls: constructs a weighted comparison from untreated regions or entities.

Incrementality tests can be underpowered, contaminated by spillover, or distorted by poor randomisation. A control group may still encounter the campaign through another channel, reducing measured lift. Tests should therefore specify the unit of randomisation, primary outcome, test duration, minimum detectable effect, decision rule, and treatment cost before launch.

8. Integrating Evidence for Budget Decisions

A mature measurement system uses different methods for different decisions.

A practical governance process is:

1. Define the business outcome, such as contribution profit rather than clicks.
2. Establish data quality, consent, identity, and conversion definitions.
3. Use descriptive attribution for journey understanding, not as unquestioned causal truth.
4. Run controlled tests on important channels and calibrate attribution or MMM estimates with the results.
5. Estimate response curves and marginal returns.
6. Allocate budget subject to profitability, capacity, brand, and risk constraints.
7. Monitor outcomes and refresh assumptions as market conditions change.

9. Illustrative Case: Northstar Homewares

Background

Northstar Homewares, a fictional online retailer, spends $400,000 per quarter across paid search, paid social, display, email, and affiliates. Its analytics dashboard reports channel ROAS using last-touch attribution. Management notices that branded search and email have the highest ROAS and proposes moving most budget away from social and display.

Problem

The decision is questionable because social and display frequently appear earlier in customer journeys, while branded search and email occur close to purchase. In addition, the company has incomplete cross-device tracking and does not separate new customers from returning customers.

Analysis

The team first compares last-touch, linear, and position-based results. Last-touch gives 48% of revenue credit to branded search and 22% to email. Position-based attribution reduces these shares because social and display receive more first-touch credit. The analysis is informative but does not establish incrementality.

Northstar then conducts a four-week geographic holdout test for prospecting social campaigns. Matched regions are randomly assigned to continued advertising or a reduced-exposure control condition. The treatment regions generate 1,080 purchases among 30,000 exposed eligible households; control regions generate 960 purchases among 30,000 comparable households.

Treatment rate:

Control rate:

Absolute lift is 0.4 percentage points, and relative lift is:

The test therefore estimates 120 incremental purchases in the treatment population, not all 1,080 purchases. If contribution per incremental purchase is $70 and campaign spend is $6,000, incremental contribution is 120 × $70 = $8,400. Incremental contribution after advertising cost is $2,400, giving an incremental contribution-to-spend ratio of 1.4.

Northstar also builds a quarterly MMM using weekly sales, media activity, price, promotions, seasonality, and stock availability. The model suggests that paid social has lower average ROAS than email but stronger remaining marginal returns at current spend. Email has high average ROAS but little scalable volume because its audience is limited.

Approach and solution

Management retains email as an efficient retention channel but does not transfer unlimited budget to it. It expands social cautiously, using the experimental lift estimate to calibrate the model. Branded search budgets are protected for profitable demand capture, while non-branded search is tested separately. The finance team requires each channel to report incremental contribution, not only attributed revenue.

Outcome

The revised allocation increases total contribution without maximising any single platform's reported ROAS. The company also adopts quarterly holdout tests, a common conversion taxonomy, separate new-versus-returning customer reporting, and a rule that channels with high attributed performance but no credible incremental evidence receive capped budgets pending testing.

Learning points

1. High last-touch ROAS does not prove high incremental value.
2. Average performance and marginal performance can point to different allocation decisions.
3. Small, scalable channels should not be judged only by current average ROAS.
4. Attribution, experiments, and MMM answer related but distinct questions.
5. Budget governance must include profit, capacity, privacy, and measurement uncertainty.

10. Ethics, Privacy and Governance

Attribution systems process behavioural data and may involve identifiers, consent signals, device information, and sensitive inferences. Organisations should collect only necessary data, document lawful purposes, honour consent and opt-out choices, apply retention limits, and secure data access. Identity resolution should not be used to create deceptive or discriminatory customer profiles.

Models should be audited for unequal measurement quality across devices, regions, languages, or customer groups. Missing consent may be systematically higher among some populations, creating unequal visibility rather than unequal marketing response. Automated budget systems should include human review, clear objectives, spending limits, anomaly detection, and an audit trail.

AI can assist with forecasting, segmentation, experiment design, and model diagnostics. It does not eliminate causal uncertainty. A complex model may produce precise-looking recommendations from biased or incomplete data. Practitioners should document training data, features, assumptions, validation results, model drift, and the consequences of incorrect allocation.

Best-Practice Checklist

• Define conversion and contribution profit before selecting a model.
• Distinguish attributed outcomes from incremental outcomes.
• Report new, returning, and high-value customers separately.
• Reconcile platform data with independent sales and finance records.
• Use consistent attribution windows and disclose them.
• Test major channels using randomised or quasi-experimental designs.
• Estimate marginal rather than only average returns.
• Include saturation, carryover, seasonality, price, and availability where relevant.
• Treat branded search, direct traffic, and retargeting as possible demand capture.
• Record uncertainty and avoid false precision.
• Govern personal data, consent, model access, and automated decisions.
• Re-test because customer behaviour, privacy rules, competition, and media costs change.

Chapter Summary

Attribution assigns conversion credit to marketing touchpoints, but assigned credit is not equivalent to causal impact. First-touch highlights customer acquisition, last-touch highlights conversion capture, linear distributes credit equally, position-based emphasises journey endpoints, and data-driven models estimate credit from observed patterns. Each approach depends on assumptions and can be distorted by selection, tracking, platform, window, and demand-harvesting biases.

Budget decisions should consider marginal return, contribution margin, diminishing returns, and operational constraints. MMM uses aggregate time-series evidence to estimate channel contributions, carryover, saturation, and scenario outcomes. Incrementality studies estimate causal lift by comparing treatment with a credible counterfactual. The strongest professional practice combines descriptive attribution, MMM, and controlled experimentation, while maintaining privacy, transparency, and financial governance.

Key Terms

Attribution: Assigning credit for a conversion or value to marketing touchpoints.

First-touch attribution: Giving all conversion credit to the first recorded touchpoint.

Last-touch attribution: Giving all conversion credit to the final recorded touchpoint.

Linear attribution: Dividing credit equally among recorded touchpoints.

Position-based attribution: Giving larger shares to the first and last touchpoints.

Data-driven attribution: Estimating credit from observed data and a statistical or machine-learning model.

Incrementality: The additional outcome caused by a marketing intervention.

Counterfactual: The outcome that would have occurred without the intervention.

Holdout test: A controlled design in which a comparable group does not receive the intervention.

Lift: The difference in outcome between treatment and control groups.

Marketing mix modelling: Aggregate statistical modelling of outcomes and marketing inputs over time.

Adstock: A representation of advertising carryover across periods.

Saturation: The decline in incremental response as spend increases.

ROAS: Revenue divided by advertising spend.

Marginal ROAS: Additional revenue divided by additional advertising spend.

Contribution margin: Revenue remaining after variable costs, expressed as an amount or rate.

Selection bias: Systematic distortion caused by differences in who receives or engages with marketing.

Attribution window: The period during which an interaction may receive credit for a conversion.

Demand harvesting: Capturing demand that was substantially created before the final marketing interaction.

Geo experiment: A controlled marketing test conducted across geographic areas.

Review Questions

Short Answer Questions

1. What is the main difference between attribution and incrementality?
2. State one strength and one limitation of last-touch attribution.
3. What does marginal ROAS measure?
4. Why can branded search receive excessive credit in a last-touch model?
5. What is the purpose of a control group in an incrementality test?

Long Answer Questions

1. Compare first-touch, last-touch, linear, position-based, and data-driven attribution. Explain the decision context for which each may be useful.
2. Explain four sources of attribution bias and describe how each can affect budget decisions.
3. Describe how MMM accounts for carryover, saturation, and non-marketing influences.
4. Explain why average ROAS may be a poor guide to the next dollar of spending.
5. Design a measurement governance process that combines attribution, MMM, and experimentation.

Analytical / Application Questions

1. A channel produces $50,000 in additional revenue from an additional $10,000 of spend. Its contribution margin is 35%. Calculate marginal ROAS and determine whether incremental contribution after advertising cost is positive.
2. In a holdout test, treatment conversion is 4.5% and control conversion is 3.5%. Calculate absolute lift and relative lift, and explain which measure is more appropriate for a finance report and why.
3. A company’s email channel has the highest attributed ROAS but a small addressable audience, while paid social has lower average ROAS and evidence of positive experimental lift. Recommend how the company should approach the next budget cycle, including at least three constraints or additional analyses.

Multiple Choice Questions with Answers

Question 1

Which attribution model assigns all conversion credit to the first recorded marketing interaction?

A. First-touch attribution

B. Last-touch attribution

C. Linear attribution

D. Position-based attribution

Correct Answer: A. First-touch attribution

Question 2

A journey contains four recorded touchpoints and produces $120 in conversion value. Under linear attribution, how much credit does each touchpoint receive?

A. $20

B. $30

C. $40

D. $120

Correct Answer: B. $30

Question 3

Which statement best describes incrementality?

A. The number of impressions served by a campaign

B. The revenue reported by an advertising platform

C. The additional outcome caused by a marketing intervention

D. The percentage of conversions assigned to the final touchpoint

Correct Answer: C. The additional outcome caused by a marketing intervention

Question 4

A campaign produces $24,000 of additional revenue from $6,000 of additional spend. What is its marginal ROAS?

A. 0.25

B. 2.0

C. 4.0

D. 18.0

Correct Answer: C. 4.0

Question 5

If a business has a 40% contribution margin, what marginal ROAS is approximately required to break even before additional fixed costs?

A. 0.40

B. 1.00

C. 2.50

D. 4.00

Correct Answer: C. 2.50

Question 6

Which method is generally most directly suited to estimating the causal lift of a particular campaign?

A. Last-touch attribution

B. Randomised holdout test

C. Linear attribution

D. Platform-reported view-through attribution

Correct Answer: B. Randomised holdout test

Question 7

In the adstock equation Aₜ = Xₜ + θAₜ₋₁, what does θ represent?

A. The conversion rate

B. The contribution margin

C. The carryover or retention parameter

D. The number of channels

Correct Answer: C. The carryover or retention parameter

Question 8

Why may last-touch attribution overvalue branded search?

A. Branded search always has the highest media cost

B. It often captures demand created by earlier activities

C. It cannot record clicks

D. It assigns equal credit to every touchpoint

Correct Answer: B. It often captures demand created by earlier activities

Question 9

Treatment conversion is 6% and control conversion is 5%. What is the relative lift?

A. 1%

B. 5%

C. 20%

D. 120%

Correct Answer: C. 20%

Question 10

Which statement about marketing mix modelling is most accurate?

A. It requires complete user-level tracking for every customer

B. It can help estimate aggregate channel contributions and scenario outcomes

C. It automatically proves that every coefficient is causal

D. It is useful only for email campaigns

Correct Answer: B. It can help estimate aggregate channel contributions and scenario outcomes