# Marketing Analytics Elective Course Guide: [Programme, course or term] > A course guide for the second-year MBA elective in marketing analytics: what the course is for, the competencies it > builds, a 12-session roadmap, the key methods with links to the concept cards (segmentation by clustering, conjoint > analysis, marketing experiments, marketing-mix models, digital attribution, customer lifetime value, diffusion and > brand equity), an original mini-case on Northfield Outdoor Co. with exhibits and a teaching note, typical assignments > and how they are graded, an exam preparation plan and a self-test. It is for MBA students choosing or taking the > elective and for managers who commission or must act on marketing analytics. It builds on the core course > (`mba/mba-marketing-guide`) and the data and decisions core (`mba/mba-decision-analytics-guide`) and links to them > instead of repeating them. Source basis: the primary works cited in each section, among them Bass (1969), Little > (1970), Green and Srinivasan (1978), McFadden (1974), Keller (1993), Gupta, Lehmann and Stuart (2004) and Shapley > (1953), and the structure common to marketing analytics electives at leading international programmes, written in > original words. The case, the company and all data are fictional; every number was computed in Python and is shown > with its Excel formula. Dated facts about privacy rules and tools were checked in October 2026 and carry a link. > Replace every [bracketed] value and every row marked *Example* with your own course details. ## Document control | Field | Value | |---|---| | Document ID | [MA-GUIDE-001] | | Programme and course | [MBA, marketing analytics elective] | | Term | [Term and year] | | Learner | [Your name] | | Instructor or coach | [Name] | | Reviewed by | [Study group / mentor] | | Software | Excel (Solver, SUMPRODUCT, T.TEST, LINEST); Python 3 with numpy, scipy, pandas and statsmodels | | Last updated | YYYY-MM-DD | | Version | [0.1] | | Classification | [Course work] | ## Contents 1. What the course is for 2. Prerequisites 3. Session roadmap 4. Key concepts and concept cards 5. Mini-case: Northfield Outdoor Co. and the budget that measured itself 6. Typical assignments and how they are graded 7. Exam preparation plan 8. Common struggles 9. Data, privacy and tools: dated facts 10. Self-test 11. Reading list ## 1. What the course is for The core marketing course teaches what to decide: whom to serve, how to position, how to set the mix. The marketing analytics elective teaches how to decide with evidence: how to find segments in data, measure what customers will trade off and pay, establish whether marketing caused a result, and allocate a budget across channels when every channel's report claims credit for the same sale. Its central idea is the difference between **correlation and incrementality**: a channel that is present before many purchases is not necessarily causing them. Little (1970) set the standard the course still follows: a model is useful to a manager when it is simple, robust, easy to control and complete on the important issues, and when the manager understands it well enough to argue with it. The course builds the competencies that accreditation bodies and programmes expect (see `mba/mba-core-curriculum-map`): | Competency | How this course builds it | Where you show it | |---|---|---| | Analytical and quantitative reasoning | Clustering, logit choice models, experiments and difference-in-differences, response curves and budget optimisation | Problem sets, mini-case, final | | Ethical decision making | Privacy, consent, fairness in targeting, honest reporting of uncertainty | Case discussion, project memo | | Communication | Turning model output into a one-page recommendation with the uncertainty stated | Memos, project presentation | | Leadership and teamwork | Team analytics project with a real or realistic data set | Team project, peer feedback | | Global perspective | Privacy regimes and platform rules that differ by country; cross-market experiments | Case discussion | | Technology, data and AI | Python or R workflows, open-source marketing-mix models, AI-assisted analysis with human review | Assignments, project | ## 2. Prerequisites | You should be able to | Check yourself | If not, use | |---|---|---| | Explain STP, the marketing mix and customer lifetime value | Write a positioning statement and compute a retention-based CLV | `mba/mba-marketing-guide`, `mba/mba-customer-lifetime-value-card` | | Read a regression output and spot omitted variable bias | Why can an ad coefficient be biased upwards in store data? | `mba/mba-regression-for-managers-card` | | Run and interpret a two-sample t-test and a confidence interval | What does a 95% interval that excludes zero tell you? | `academic/stats-t-tests`, `academic/stats-confidence-intervals` | | Explain logistic regression and maximum likelihood | What does a logit coefficient do to the odds? | `academic/stats-logistic-regression`, `academic/stats-estimation-mle` | | Allocate a budget with response curves and Solver | Why should marginal, not average, returns be equal across channels? | `mba/mba-marketing-mix-card`, `mba/pre-mba-excel-modeling-drills` | | Use Python or R for a small analysis | Load a CSV, compute group means, fit a regression | `mba/mba-decision-analytics-guide`, `mba/pre-mba-quant-refresher` | ## 3. Session roadmap | Session | Topic | Card or template to read | Case or exercise | Deliverable | |---|---|---|---|---| | 1 *(Example)* | Marketing data, metrics and the incrementality problem | This guide, sections 1 and 4 | Last-click report of a retailer | One-page list of questions the report cannot answer | | 2 *(Example)* | Segmentation by cluster analysis | `mba/mba-segmentation-clustering-card` | Cluster a survey of 400 buyers | Segment profiles and a typing tool | | 3 *(Example)* | Conjoint analysis I: ratings and choice models | `mba/mba-conjoint-analysis-card` | Design a choice study for a new product | Attribute and level table, design | | 4 *(Example)* | Conjoint analysis II: willingness to pay, simulators, pricing | `mba/mba-conjoint-analysis-card`, `mba/mba-pricing-card` | Price a product with a simulator | Profit curve and price memo | | 5 *(Example)* | Marketing experiments: A/B tests, power, pitfalls | `mba/mba-marketing-experiments-card`, `academic/stats-power-sample-size` | Plan an email or website test | Pre-registered test plan | | 6 *(Example)* | Field and geo experiments, difference-in-differences | `mba/mba-marketing-experiments-card`, `academic/stats-experiment-design` | Mini-case Exhibit 2 | DiD estimate with confidence interval | | 7 *(Example)* | Marketing-mix models I: adstock, saturation, estimation | `mba/mba-marketing-mix-modeling-card`, `academic/stats-time-series-arima` | Fit a small MMM | Model with response curves | | 8 *(Example)* | Marketing-mix models II: calibration and budget allocation | `mba/mba-marketing-mix-modeling-card`, `mba/mba-marketing-mix-card` | Mini-case Exhibit 3 | Solver allocation | | 9 *(Example)* | Digital attribution: rules, data-driven and Shapley | `mba/mba-digital-attribution-card` | Compare last-click and Shapley credit | Attribution comparison table | | 10 *(Example)* | Customer analytics: CLV, churn, targeting the right customers | `mba/mba-customer-lifetime-value-card` | Retention campaign targeting | Targeting rule with expected profit | | 11 *(Example)* | New products and brands: diffusion forecasts and brand equity | `mba/mba-new-product-diffusion-card`, `mba/mba-brand-equity-card` | Forecast a launch, value a brand | Bass forecast, brand premium | | 12 *(Example)* | Integration: the measurement plan, privacy and AI in analytics | This guide, sections 5 and 9 | Northfield mini-case discussion | Case write-up and measurement plan | ```mermaid flowchart LR A["Customers: segment and value them"] --> B["Design: conjoint, pricing, diffusion"] B --> C["Spend: channels and messages"] C --> D["Measure: experiments"] C --> E["Model: marketing-mix model, attribution"] D --> F["Calibrate the model with experiments"] E --> F F --> G["Decide: allocation, targeting, price"] G --> A ``` ## 4. Key concepts and concept cards | Concept | What it is | Originator | Where to study it | |---|---|---|---| | Post hoc segmentation | Groups found in data by clustering needs or behaviour | Smith (1956); MacQueen (1967); Ward (1963) | `mba/mba-segmentation-clustering-card` | | Conjoint analysis | Part-worths of attribute levels from evaluations of whole products | Luce and Tukey (1964); Green and Rao (1971); Green and Srinivasan (1978) | `mba/mba-conjoint-analysis-card` | | Multinomial logit choice model | Choice probabilities from random utility | Luce (1959); McFadden (1974) | `mba/mba-conjoint-analysis-card` | | Randomised experiments | Causal effects from random assignment | Fisher (1935) | `mba/mba-marketing-experiments-card`, `academic/stats-experiment-design` | | Difference-in-differences | Effect as the change in a treated group minus the change in a comparison group | Econometric practice | `mba/mba-marketing-experiments-card`, section 5 of this guide | | Adstock and saturation | Carry-over of advertising and diminishing returns to spend | Broadbent (1979) and response-modelling practice | `mba/mba-marketing-mix-modeling-card`, `mba/mba-marketing-mix-card` | | Marketing-mix model | Regression of sales on transformed marketing inputs and controls | Response-modelling practice; Little (1970) on decision models | `mba/mba-marketing-mix-modeling-card` | | Shapley attribution | Credit equal to a channel's average marginal contribution over all orders | Shapley (1953) | `mba/mba-digital-attribution-card` | | Customer lifetime value | Present value of a customer's expected margins | Gupta, Lehmann and Stuart (2004) | `mba/mba-customer-lifetime-value-card` | | Bass diffusion model | First purchases from innovation and imitation | Bass (1969) | `mba/mba-new-product-diffusion-card` | | Customer-based brand equity | Differential effect of brand knowledge on response to marketing | Keller (1993); Aaker (1991) | `mba/mba-brand-equity-card` | **Three ideas to carry through every session.** 1. **Average versus marginal.** The average return of a channel says whether the money spent so far paid back; the marginal return of the last dollar says where the next dollar should go. Budget decisions use marginal returns. 2. **Attributed versus incremental.** A report that credits a channel with a sale (attribution) is not evidence that the sale would not have happened without the channel (incrementality). Only experiments, or models calibrated with experiments, measure incrementality. 3. **Point estimate versus decision.** Every estimate in the course comes with uncertainty. A recommendation states the estimate, its interval, and what the manager should do across the plausible range. ## 5. Mini-case: Northfield Outdoor Co. and the budget that measured itself ### 5.1 Situation Two years after the Fastpack launch, Northfield Outdoor Co. was spending USD 6.0 million a year on paid marketing across five channels, and its new head of performance marketing, Leila Haddad, had a problem she described as "five reports that add up to more than the business". The digital analytics platform credited paid search with almost USD 10 million of revenue on a last-click basis. The affiliate network reported a return of four dollars per dollar. The video agency, whose campaigns rarely received last-click credit at all, argued that its work was what made people search for Northfield in the first place. Add up every channel's own claim and the total exceeded the revenue growth the finance team could see. The chief marketing officer, Mira Okonjo, wanted a budget for the coming year that she could defend line by line. The finance director, Elena Brandt, had a sharper question: "If we cut the whole budget by a quarter, how much revenue do we actually lose? Nobody here can tell me." Northfield's analytics lead, Tomás Reyes, had prepared for the question. Over the spring he had run a geographic holdout test: Northfield's 20 sales regions were split at random into 10 holdout regions, where all paid search was switched off for eight weeks, and 10 control regions, where search ran as normal. Store and online revenue was recorded for each region for the eight weeks before the test and the eight weeks of the test. The test had been uncomfortable: the paid search agency warned that Northfield would "disappear from the internet" in the holdout regions, and two regional managers asked to be excluded. Tomás had kept the randomisation intact. In parallel, a consultancy had built a marketing-mix model on three years of weekly data, with adstock and saturation for each channel and controls for seasonality, price and weather. The model's response curves (Exhibit 3) told a very different story from the last-click report: they gave paid social and online video much more credit, and affiliates much less. Leila suspected that many affiliate sales came from discount-code sites that customers visited at checkout, after they had already decided to buy. The consultancy's model had been finished before the geo test ended, so the test was a genuine check on it. Mira asked Tomás and Leila for four things before the budget meeting in two weeks: an honest estimate of what paid search really adds, a view on which of the two other sources (last-click or the model) to trust for the other channels, a reallocation of the USD 6.0 million that would raise incremental gross profit, and a measurement plan so that next year's budget meeting would not be an argument between reports. Northfield's gross margin on incremental revenue was 45%. Mira added one constraint: whatever they proposed had to respect the privacy commitments Northfield had made to its customers, which ruled out buying third-party data to track individuals across sites. ### 5.2 Exhibits **Exhibit 1. Paid marketing spend and last-click attributed revenue, last 12 months (USD thousands, fictional).** | Channel | Spend | Last-click revenue | |---|---|---| | Paid search | 1,800 | 9,900 | | Paid social | 1,500 | 3,000 | | Online video | 900 | 450 | | Catalogue and direct mail | 1,000 | 1,600 | | Affiliates | 800 | 3,200 | | Total | 6,000 | 18,150 | **Exhibit 2. Geo holdout test of paid search: average weekly revenue per region (USD thousands, fictional).** Regions were assigned at random. "Pre" is the 8 weeks before the test, "Test" the 8 weeks of the test; paid search was off in holdout regions during the test. Paid search spend before the test was USD 1.8 million a year, spread across the 20 regions in proportion to their revenue. | Region | Group | Pre | Test | Change | Region | Group | Pre | Test | Change | |---|---|---|---|---|---|---|---|---|---| | G01 | Holdout | 130.2 | 129.8 | -0.4 | G11 | Control | 168.0 | 172.4 | 4.4 | | G02 | Holdout | 156.0 | 157.9 | 1.9 | G12 | Control | 162.9 | 172.1 | 9.2 | | G03 | Holdout | 102.6 | 102.5 | -0.1 | G13 | Control | 148.4 | 156.0 | 7.6 | | G04 | Holdout | 184.9 | 185.3 | 0.4 | G14 | Control | 147.9 | 156.1 | 8.2 | | G05 | Holdout | 166.0 | 168.0 | 2.0 | G15 | Control | 154.0 | 158.4 | 4.4 | | G06 | Holdout | 142.7 | 146.9 | 4.2 | G16 | Control | 134.6 | 141.4 | 6.8 | | G07 | Holdout | 142.2 | 141.3 | -0.9 | G17 | Control | 139.9 | 144.5 | 4.6 | | G08 | Holdout | 157.6 | 160.9 | 3.3 | G18 | Control | 163.7 | 169.3 | 5.6 | | G09 | Holdout | 143.3 | 141.4 | -1.9 | G19 | Control | 146.7 | 153.6 | 6.9 | | G10 | Holdout | 144.4 | 145.3 | 0.9 | G20 | Control | 115.6 | 122.0 | 6.4 | | Mean | Holdout | 146.99 | 147.93 | 0.94 | Mean | Control | 148.17 | 154.58 | 6.41 | **Exhibit 3. Marketing-mix model response curves, annual incremental revenue $R_c(x) = a_c(1 - e^{-x/k_c})$ (USD thousands, fictional).** $a_c$ is the saturation level of channel $c$ and $k_c$ its spend scale; the form is the one used in `mba/mba-marketing-mix-card`. | Channel | $a_c$ | $k_c$ | Current spend $x_c$ | |---|---|---|---| | Paid search | 7,990 | 1,500 | 1,800 | | Paid social | 13,300 | 2,500 | 1,500 | | Online video | 6,380 | 1,500 | 900 | | Catalogue and direct mail | 2,100 | 800 | 1,000 | | Affiliates | 1,300 | 600 | 800 | ### 5.3 Discussion questions 1. Compute the last-click return on ad spend (ROAS) for each channel from Exhibit 1. Which channels is last-click likely to over-credit and under-credit, and why? 2. Using Exhibit 2, estimate the incremental revenue of paid search per region-week with difference-in-differences, its 95% confidence interval, and the incremental ROAS. What makes the estimate credible, and what could still bias it? 3. Using Exhibit 3, compute each channel's incremental revenue, average ROAS and marginal ROAS at current spend. Compare the model with Exhibit 1 and with your answer to question 2. 4. Reallocate the USD 6.0 million to maximise incremental gross profit. How much does the reallocation gain? Answer Elena: should the total budget be cut, and what does the model leave out? 5. Propose a measurement plan for next year that Mira can present to the board: which experiments, how the model is calibrated and refreshed, which decision each piece supports, and how privacy is respected. ### 5.4 Teaching note **Q1. Last-click.** ROAS = last-click revenue / spend (`=C2/B2`): paid search 5.50, paid social 2.00, online video 0.50, catalogue and direct mail 1.60, affiliates 4.00; total 3.03. Last-click gives all credit to the final touch before purchase. It over-credits channels that sit at the end of the journey and capture demand that already exists: brand search (customers who already know Northfield type its name) and coupon or cashback affiliates (visited at checkout). It under-credits channels that create demand earlier but are rarely the last click: video, social and the catalogue, whose effect often shows up as a search or a direct visit days later. It also ignores store sales driven by online media. **Q2. The geo test.** Difference-in-differences compares the change in holdout regions with the change in control regions, which removes both the fixed differences between regions and the seasonal rise that hit all regions: $$\widehat{\text{DiD}} = (\bar{y}^{H}_{\text{test}} - \bar{y}^{H}_{\text{pre}}) - (\bar{y}^{C}_{\text{test}} - \bar{y}^{C}_{\text{pre}}) = (147.93 - 146.99) - (154.58 - 148.17) = 0.94 - 6.41 = -5.47$$ Switching search off cost USD 5.47k of revenue per region-week. In Excel, with the changes in `E2:E11` (holdout) and `J2:J11` (control): `=AVERAGE(E2:E11)-AVERAGE(J2:J11)` and `=T.TEST(E2:E11,J2:J11,2,3)` for the two-sided Welch test. ```python import numpy as np from scipy import stats holdout = np.array([-0.4, 1.9, -0.1, 0.4, 2.0, 4.2, -0.9, 3.3, -1.9, 0.9]) # change in weekly revenue, USD k control = np.array([4.4, 9.2, 7.6, 8.2, 4.4, 6.8, 4.6, 5.6, 6.9, 6.4]) did = holdout.mean() - control.mean() se = np.sqrt(holdout.var(ddof=1) / 10 + control.var(ddof=1) / 10) res = stats.ttest_ind(holdout, control, equal_var=False) ci = res.confidence_interval(0.95) saved = 1800 / (20 * 52) # search spend saved per region-week, USD k print(f"DiD {did:.2f} se {se:.3f} t {res.statistic:.2f} p {res.pvalue:.1e}") print(f"95% CI {ci.low:.2f} to {ci.high:.2f}") print(f"iROAS {-did / saved:.2f} (CI {-ci.high / saved:.2f} to {-ci.low / saved:.2f})") ``` Output from the run: ```text DiD -5.47 se 0.801 t -6.83 p 2.4e-06 95% CI -7.15 to -3.79 iROAS 3.16 (CI 2.19 to 4.13) ``` The Welch test uses 17.66 degrees of freedom. If you build the interval in Excel with `=T.INV.2T(0.05,17.66)`, Excel truncates the degrees of freedom to 17 and returns 2.1098 instead of 2.1039, giving -7.16 to -3.78: the same answer for the decision. Search spend saved per region-week was $1{,}800/(20 \times 52) = 1.7308$ (USD k), so the incremental ROAS is $5.47/1.7308 = 3.16$ (95% interval 2.19 to 4.13). Over a year, search adds about $5.47 \times 20 \times 52 =$ USD 5.69 million of revenue, not the USD 9.9 million last-click credits: last-click overstates search by about 74%. Even the upper end of the interval (4.13) is well below 5.50. Credibility comes from random assignment of regions, which makes the groups comparable (pre-period means 146.99 and 148.17), and from the pre-post design, which removes fixed differences. Remaining risks: only 10 regions per group (a few unusual regions could matter; the interval reflects this), spillovers (holdout customers who saw search ads while travelling or shopping online across regions bias the effect towards zero), a short test that misses any long-term effect of search on brand, and novelty effects. The result applies to search as a whole; brand and non-brand search probably differ, which is the next test. **Q3. The model.** For each channel, $R_c = a_c(1 - e^{-x_c/k_c})$ (`=a*(1-EXP(-x/k))`), average ROAS $R_c/x_c$ and marginal ROAS $R_c'(x_c) = (a_c/k_c)e^{-x_c/k_c}$ (`=a/k*EXP(-x/k)`): | Channel | Incremental revenue (USD k) | Average ROAS | Marginal ROAS | Last-click ROAS | Last-click / model revenue | |---|---|---|---|---|---| | Paid search | 5,583.5 | 3.10 | 1.604 | 5.50 | 1.77 | | Paid social | 6,000.8 | 4.00 | 2.920 | 2.00 | 0.50 | | Online video | 2,878.6 | 3.20 | 2.334 | 0.50 | 0.16 | | Catalogue and direct mail | 1,498.3 | 1.50 | 0.752 | 1.60 | 1.07 | | Affiliates | 957.3 | 1.20 | 0.571 | 4.00 | 3.34 | | Total | 16,918.5 | 2.82 | | 3.03 | | The model's search ROAS (3.10, USD 5.58 million) sits inside the geo test's interval and close to its point estimate (3.16, USD 5.69 million), which is the strongest reason to trust the model's other channels more than last-click. The model reverses the last-click ranking: social and video, not search and affiliates, have the highest marginal returns. Gross profit after spend from paid marketing is $0.45 \times 16{,}918.5 - 6{,}000 =$ USD 1,613.3k. With a gross margin of 45%, a channel's last dollar pays for itself only if its marginal ROAS exceeds $1/0.45 = 2.22$: social (2.92) and video (2.33) clear the bar, search (1.60), catalogue (0.75) and affiliates (0.57) do not. ```vega-lite { "$schema": "https://vega.github.io/schema/vega-lite/v5.json", "description": "Northfield: last-click ROAS against model average and marginal ROAS by channel (fictional)", "width": 420, "height": 240, "data": {"values": [ {"channel": "Paid search", "measure": "Last-click ROAS", "value": 5.50}, {"channel": "Paid search", "measure": "Model average ROAS", "value": 3.10}, {"channel": "Paid search", "measure": "Model marginal ROAS", "value": 1.604}, {"channel": "Paid social", "measure": "Last-click ROAS", "value": 2.00}, {"channel": "Paid social", "measure": "Model average ROAS", "value": 4.00}, {"channel": "Paid social", "measure": "Model marginal ROAS", "value": 2.920}, {"channel": "Online video", "measure": "Last-click ROAS", "value": 0.50}, {"channel": "Online video", "measure": "Model average ROAS", "value": 3.20}, {"channel": "Online video", "measure": "Model marginal ROAS", "value": 2.334}, {"channel": "Catalogue and mail", "measure": "Last-click ROAS", "value": 1.60}, {"channel": "Catalogue and mail", "measure": "Model average ROAS", "value": 1.50}, {"channel": "Catalogue and mail", "measure": "Model marginal ROAS", "value": 0.752}, {"channel": "Affiliates", "measure": "Last-click ROAS", "value": 4.00}, {"channel": "Affiliates", "measure": "Model average ROAS", "value": 1.20}, {"channel": "Affiliates", "measure": "Model marginal ROAS", "value": 0.571} ]}, "mark": "bar", "encoding": { "x": {"field": "channel", "type": "nominal", "title": null, "sort": null}, "xOffset": {"field": "measure"}, "y": {"field": "value", "type": "quantitative", "title": "Revenue per USD of spend"}, "color": {"field": "measure", "type": "nominal", "title": null} } } ``` **Q4. Reallocation.** Revenue is maximised for a fixed budget when every funded channel has the same marginal ROAS $\lambda$. Setting $(a_c/k_c)e^{-x_c/k_c} = \lambda$ gives $x_c = k_c \ln(a_c/(k_c \lambda))$; choose $\lambda$ so that the $x_c$ add up to 6,000 (Solver: maximise `=SUMPRODUCT(a,1-EXP(-x/k))` subject to `=SUM(x)=6000` and $x \ge 0$). The solution is $\lambda = 1.8104$: | Channel | Current spend | Optimal spend | Change | Incremental revenue at optimum | |---|---|---|---|---| | Paid search | 1,800 | 1,618.8 | -181.2 | 5,274.5 | | Paid social | 1,500 | 2,694.9 | +1,194.9 | 8,774.1 | | Online video | 900 | 1,281.3 | +381.3 | 3,664.5 | | Catalogue and direct mail | 1,000 | 297.2 | -702.8 | 651.7 | | Affiliates | 800 | 107.8 | -692.2 | 213.8 | | Total | 6,000 | 6,000.0 | 0 | 18,578.6 | The same budget earns USD 1,660.0k more incremental revenue (18,578.6 against 16,918.5) and USD 747.0k more gross profit; gross profit after spend rises from USD 1,613.3k to USD 2,360.4k. Because the curves are estimates, move half way in the first half-year (search 1,709.4, social 2,097.4, video 1,090.6, catalogue 648.6, affiliates 453.9, which already lifts gross profit after spend to USD 2,162.5k) and confirm with a test before moving the rest. To Elena's question: at the optimum the marginal ROAS (1.81) is below the 2.22 break-even, so on these curves the short-run profit-maximising budget is smaller: set $\lambda = 2.22$ and the total falls to USD 4,600.8k (search 1,311.3, social 2,182.4, video 973.8, catalogue 133.3, affiliates 0), with gross profit after spend of USD 2,495.7k, only USD 135.4k more than reallocating the full 6.0 million. Most of the gain comes from reallocation, not from cutting. And the model measures revenue within its time window: it leaves out the lifetime value of new customers acquired (`mba/mba-customer-lifetime-value-card`), long-term brand effects (`mba/mba-brand-equity-card`), and any effect beyond its adstock decay. A deep cut is the decision most exposed to those omissions. Recommendation: hold the total at about USD 6.0 million this year, reallocate in two steps, and test the cut directly with a budget-level geo test before deciding. **Q5. Measurement plan.** (1) A rolling calendar of geo experiments: brand versus non-brand search first, then affiliates (switch off coupon-site partners in holdout regions), then a heavy-up test of social and video, each with a power calculation (`academic/stats-power-sample-size`). (2) Refresh the marketing-mix model quarterly and calibrate it with every experiment result, using an open-source framework if the team has the skills (section 9). (3) Use attribution (`mba/mba-digital-attribution-card`) only for within-channel optimisation, such as keywords and creatives, never for the budget split. (4) Report every number with its interval and the decision it supports. (5) Privacy: geo tests and aggregate models need no individual tracking across sites, which fits Northfield's commitments; customer-level analysis uses only first-party data with consent, reviewed by the privacy team. ## 6. Typical assignments and how they are graded | Assignment | What it asks | Typical weight | |---|---|---| | Class participation | Case discussions, cold calls, contributions on method choices | 15% to 25% | | Individual problem sets (3 to 5) | Clustering, conjoint estimation, experiment analysis, MMM allocation | 25% to 35% | | Team analytics project | A real or realistic data set, a question, an analysis and a recommendation | 25% to 35% | | Final exam or final case | Method choice, computation and interpretation under time pressure | 20% to 30% | **Rubric for a problem set or project memo (self-assessment).** | Criterion | Excellent | Adequate | Weak | |---|---|---|---| | Question and method | The business question is explicit and the method fits it (causal question, causal design) | Method fits but the question is vague | Method chosen because it is available | | Data and assumptions | Data sources, cleaning and assumptions stated; limits named | Some assumptions stated | Data used without checks | | Correct analysis | Calculations reproducible, uncertainty quantified, sensitivity shown | Correct point estimates, little on uncertainty | Errors in calculation or interpretation | | Interpretation | Average versus marginal and attributed versus incremental kept apart | Mostly correct with some slips | Correlation read as cause | | Recommendation | Clear decision, the range over which it holds, next test | Decision without range | No decision, or a decision the analysis does not support | | Communication | One-page summary, exhibits a manager can read, code and data appended | Clear but long | Output dumped without interpretation | ## 7. Exam preparation plan | Week | Focus | Activities | |---|---|---| | Final minus 3 *(Example)* | Methods map | For each card, write the question it answers, the formula, and one number from the worked example from memory | | Final minus 2 *(Example)* | Computation drills | Redo Worked examples A of the clustering, conjoint, experiment and MMM cards in Excel without notes; time each one | | Final minus 1 *(Example)* | Cases and interpretation | Redo this mini-case in 90 minutes; write a one-page memo; answer the self-test in section 10 | | Exam week *(Example)* | Review | Error log review (`mba/mba-error-log`), formula sheet (`mba/mba-formula-sheet`), sleep | Expect four types of question: choose the method for a business question and justify it; compute (logit shares, WTP, DiD with an interval, marginal ROAS, allocation); interpret output (cluster profiles, regression or choice-model estimates); and critique a flawed analysis (last-click used for budgets, an A/B test stopped early, a segmentation on unstandardised data). ## 8. Common struggles | Struggle | Why it happens | Remedy | |---|---|---| | Confusing attribution with incrementality | Platform reports look causal and are precise to the cent | For every number ask "compared with what would have happened otherwise?" | | Using average instead of marginal returns | Average ROAS is what reports show | Always compute the derivative of the response curve at current spend | | Over-trusting a model's fit | High $R^2$ in an MMM looks reassuring | Ask whether the model was checked against an experiment | | Getting lost in code | Python or R syntax absorbs the time | Reuse the short blocks in the cards; write the question first, then code | | Stopping A/B tests early | A significant result appears after a few days | Pre-register the sample size and the analysis (`mba/mba-marketing-experiments-card`) | | Choosing the number of segments by statistics alone | Fit indices look objective | Combine fit, stability and actionability (`mba/mba-segmentation-clustering-card`) | | Treating conjoint shares as market forecasts | The simulator produces precise percentages | Adjust for awareness and distribution; calibrate to sales | ## 9. Data, privacy and tools: dated facts | Topic | Fact as checked | Check before you rely on it | |---|---|---| | EU and EEA data protection | The General Data Protection Regulation, Regulation (EU) 2016/679, has applied since 25 May 2018 and covers profiling and the use of personal data for marketing | [EUR-Lex, Regulation (EU) 2016/679](https://eur-lex.europa.eu/eli/reg/2016/679/oj) and your national regulator | | Browser tracking | As of October 2026, Google's Privacy Sandbox site states that some Privacy Sandbox technologies are being phased out (page last updated 17 October 2025). Chrome's plans for third-party cookies changed several times between 2020 and 2025 | [privacysandbox.google.com](https://privacysandbox.google.com/) | | Open-source marketing-mix models | As of October 2026, Google's Meridian is published on GitHub under the Apache-2.0 licence, and Meta's Robyn under the MIT licence, described by its authors as experimental, written in R with a Python interface in beta | [github.com/google/meridian](https://github.com/google/meridian), [github.com/facebookexperimental/Robyn](https://github.com/facebookexperimental/Robyn) | Two practical consequences for the course. First, measurement that does not depend on tracking individuals across sites (geo experiments, aggregate marketing-mix models, first-party customer data with consent) is more robust to changes in browsers, platforms and law than user-level multi-touch attribution. Second, any customer-level analysis in a project (segmentation, CLV, targeting) needs a check of the legal basis and of the data owner's policy; anonymise or aggregate wherever the question allows. Using AI assistants to write analysis code is common practice; review every line, never paste personal data into a tool your organisation has not approved, and follow your course's AI policy for graded work. ## 10. Self-test 1. Paid search has a last-click ROAS of 5.50 and an incremental ROAS of 3.16 in the geo test. Explain the gap in two sentences. 2. A response curve has $a$ = 13,300 and $k$ = 2,500 (USD thousands). What is the marginal ROAS at a spend of 1,500? Give the Excel formula. 3. With a gross margin of 45%, what marginal ROAS must the last dollar of spend reach to pay for itself? 4. Two products have logit utilities 0 and -0.7. What are their shares of preference? 5. A Bass model has $p$ = 0.04 and $q$ = 0.60. When does adoption peak? 6. Write the difference-in-differences estimator with the numbers of Exhibit 2. 7. What does a silhouette near zero tell you about a customer in a segmentation? 8. Two channels: search alone produces 60 conversions, social alone 40, both together 120. What is each channel's Shapley credit? 9. With a geometric adstock decay of 0.5, a spend of 100 in week 1 and nothing afterwards, what is the adstock in weeks 1, 2 and 3? 10. Why should a marketing-mix model be calibrated with experiments, and why should attribution not be used to split the budget across channels? ## 11. Reading list **Primary works.** - Bass, F. M. (1969). A new product growth for model consumer durables. *Management Science*, 15(5), 215-227. - Little, J. D. C. (1970). Models and managers: the concept of a decision calculus. *Management Science*, 16(8), B466-B485. - Green, P. E. and Srinivasan, V. (1978). Conjoint analysis in consumer research: issues and outlook. *Journal of Consumer Research*, 5(2), 103-123. - McFadden, D. (1974). Conditional logit analysis of qualitative choice behavior. In P. Zarembka (ed.), *Frontiers in Econometrics*, 105-142. Academic Press. - Keller, K. L. (1993). Conceptualizing, measuring, and managing customer-based brand equity. *Journal of Marketing*, 57(1), 1-22. - Gupta, S., Lehmann, D. R. and Stuart, J. A. (2004). Valuing customers. *Journal of Marketing Research*, 41(1), 7-18. - Shapley, L. S. (1953). A value for n-person games. In H. W. Kuhn and A. W. Tucker (eds.), *Contributions to the Theory of Games, Volume II* (Annals of Mathematics Studies 28), 307-317. Princeton University Press. - Punj, G. and Stewart, D. W. (1983). Cluster analysis in marketing research: review and suggestions for application. *Journal of Marketing Research*, 20(2), 134-148. - Fisher, R. A. (1935). *The Design of Experiments*. Oliver and Boyd. - Broadbent, S. (1979). One way TV advertisements work. *Journal of the Market Research Society*, 21. **Textbooks and books (by name only).** Lilien, Rangaswamy and De Bruyn, *Principles of Marketing Engineering and Analytics*; Kohavi, Tang and Xu, *Trustworthy Online Controlled Experiments*; Hanssens, Parsons and Schultz, *Market Response Models*; Farris, Bendle, Pfeifer and Reibstein, *Marketing Metrics*; Wedel and Kamakura, *Market Segmentation*; Train, *Discrete Choice Methods with Simulation*. ## Answer key 1. Last-click credits search with every sale whose final click was a search ad, including customers who would have bought anyway (they searched for Northfield by name, often after seeing other media). The geo test measures only the revenue lost when search is switched off, which is the incremental part. 2. $(13{,}300/2{,}500)e^{-1{,}500/2{,}500} = 5.32 \times 0.5488 = 2.92$; `=13300/2500*EXP(-1500/2500)`. 3. $1/0.45 = 2.22$: each dollar of revenue brings 0.45 of gross profit, so 2.22 dollars of revenue are needed per dollar of spend. 4. $e^{0} = 1$ and $e^{-0.7} = 0.4966$: shares 66.8% and 33.2% (`=EXP(0)/(EXP(0)+EXP(-0.7))`). 5. $t^* = \ln(q/p)/(p+q) = \ln(15)/0.64 = 4.23$ years. 6. $(147.93 - 146.99) - (154.58 - 148.17) = 0.94 - 6.41 = -5.47$ (USD k per region-week). 7. The customer is about as close to another segment as to its own: a borderline case whose assignment is uncertain. 8. Search: $\frac{1}{2}(60) + \frac{1}{2}(120 - 40) = 70$; social: $\frac{1}{2}(40) + \frac{1}{2}(120 - 60) = 50$; the credits add up to 120. 9. 100, 50 and 25 (each week keeps half of the previous week's adstock). 10. A model fitted to observational data can confuse channels that move together with seasonality or with each other; experiments give unbiased anchors that the model must reproduce. Attribution assigns credit for sales that happened, not the sales that would be lost without a channel, so it rewards channels that capture existing demand and penalises those that create it. ## Revision history | Version | Date | Author | Change summary | |---|---|---|---| | 0.1 | YYYY-MM-DD | [Your name] | First draft from the template | | 0.2 | YYYY-MM-DD | [Your name] | *Example:* Replaced the roadmap with the term syllabus; added project data set | ## Standard and licence note Structure aligned with the marketing analytics electives common at leading international programmes and with the competency areas used by business school accreditation bodies; methods based on the primary works listed in section 11. The text, mini-case, data, code and exercises are original and fictional; no textbook, published case or course material was copied or paraphrased. This template is not affiliated with or endorsed by any business school, accreditation body, publisher, author or software vendor. Dated facts about law, browsers and tools were checked in October 2026; check the current official pages before you rely on them. Template licence: MIT.