5. Consumer Analytics and Causal Insight

Introduction and Context

Global marketing decisions increasingly depend on the ability to convert dispersed consumer information into defensible managerial knowledge. Firms can observe transactions, digital interactions, search behaviour, media exposure, customer-service conversations, social commentary, and responses to experiments. Yet observation alone does not establish why consumers behave as they do. A rise in sales may result from advertising, seasonality, distribution expansion, a competitor’s stock-out, or an unmeasured economic change. The central challenge of consumer analytics is therefore not merely prediction. It is the disciplined development of insight that distinguishes association from causation and translates evidence into decisions.

This challenge is particularly demanding in global markets. Data-generating processes differ across countries because of language, regulation, payment systems, cultural meanings, media environments, economic conditions, and levels of digital adoption. A model developed in one market may not transport successfully to another. Even apparently universal constructs—such as trust, value, loyalty, or convenience—may have different meanings and measurement properties across cultures.

Consumer analytics combines statistical modelling, research design, experimentation, machine learning, and substantive marketing theory. Its purpose is to answer several related questions: Who are the customers? What influences their choices? Which interventions change behaviour? How long will customers remain valuable? How should resources be allocated across markets and segments? This chapter develops an integrated approach to these questions, with particular attention to multivariate analysis, structural equation modelling, experiments, causal inference, choice models, text analytics, social listening, forecasting, and customer-lifetime-value estimation.

Learning Objectives

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

1. Distinguish descriptive, predictive, and causal consumer analytics.

2. Select and interpret appropriate multivariate and structural models.

3. Design experiments and observational studies that support credible causal conclusions.

4. Apply choice models to product, pricing, and market-entry decisions.

5. Evaluate text analytics and social-listening evidence in multilingual environments.

6. Develop forecasts and customer-lifetime-value estimates while recognising uncertainty.

7. Integrate diverse analytical outputs into global marketing decisions.

8. Identify ethical, governance, validity, and implementation risks.

From Data Description to Causal Insight

Consumer analytics can be viewed as a hierarchy of questions. Descriptive analytics asks what happened: sales, conversion, retention, sentiment, and customer composition. Predictive analytics asks what is likely to happen, such as whether a customer will churn or respond to an offer. Prescriptive analytics asks what action should be taken. Causal analytics asks a more demanding question: what would happen if a particular intervention were changed, holding other relevant conditions constant?

The causal question is counterfactual. For a customer, one may wish to compare the outcome under exposure to an advertisement with the outcome that would have occurred for the same customer without exposure. Only one of these outcomes can normally be observed for an individual at a given time. Causal inference therefore depends on research design, assumptions, comparison groups, and sensitivity analysis rather than on statistical sophistication alone.

A useful distinction is between an outcome variable, treatment or intervention, covariates, mediators, moderators, and confounders. A confounder influences both the treatment and the outcome. For example, high-intent consumers may be more likely to click a search advertisement and more likely to purchase; a simple comparison of clickers and non-clickers may overstate advertising effectiveness. A mediator explains part of the pathway between an intervention and an outcome, such as perceived value mediating the effect of a price presentation on purchase intention. A moderator changes the strength or direction of an effect, such as cultural uncertainty avoidance moderating the relationship between trust cues and adoption.

The potential-outcomes framework formalises causal effects by defining outcomes under alternative treatment conditions. The average treatment effect is the average difference between these potential outcomes. In practice, identification requires assumptions such as random assignment, conditional independence, stable treatment conditions, and adequate overlap between treated and comparison units. These assumptions should be argued substantively, not simply declared after a model has been estimated.

Multivariate Analysis for Consumer Understanding

Marketing data contain multiple variables that are often correlated. Multivariate analysis allows the researcher to examine these relationships jointly, reducing the risk of drawing conclusions from isolated bivariate associations.

Multiple regression estimates the relationship between an outcome and several predictors. A model of customer spending might include income, tenure, service usage, promotion exposure, and market-level controls. Regression coefficients are conditional associations unless the design supports a causal interpretation. Important diagnostic issues include nonlinearity, interaction effects, multicollinearity, heteroscedasticity, influential observations, missing data, and model specification.

Logistic regression is appropriate when the outcome is binary, such as renewal versus non-renewal. Its coefficients are expressed in log-odds, although predicted probabilities or marginal effects are usually more interpretable for managers. Count models, including Poisson and negative-binomial regression, can be used for purchase frequency or service contacts; the latter is often preferable when variance exceeds the mean.

Segmentation methods seek to identify groups with meaningful differences in needs or behaviour. Cluster analysis groups observations according to a chosen distance measure, while latent-class analysis treats segment membership as probabilistic. Segments should not be accepted merely because an algorithm produces them. Their stability, interpretability, economic value, actionability, and cross-market comparability must be assessed. A segment that is statistically distinct but impossible to reach or serve may have little managerial value.

Factor analysis reduces correlated observed measures to a smaller number of latent dimensions. Exploratory factor analysis is useful when the structure is uncertain; confirmatory factor analysis tests a theoretically specified measurement model. Researchers should examine factor loadings, reliability, convergent validity, discriminant validity, and measurement invariance. Cronbach’s alpha alone is not sufficient evidence of construct quality because it can be inflated by redundant items.

Principal component analysis and factor analysis are related but conceptually different. Principal components summarise variance mathematically, whereas factor analysis models common variance as generated by latent constructs. The distinction matters when interpreting dimensions such as brand trust, perceived risk, or relationship quality.

Structural Equation Modelling: Linking Constructs and Outcomes

Structural equation modelling (SEM) is particularly valuable when marketing theory includes latent constructs and mediated relationships. SEM combines a measurement model, which links constructs to observed indicators, with a structural model, which specifies relationships among constructs.

A model might propose that communication quality increases perceived trust, trust increases perceived value, and perceived value increases loyalty intention. SEM enables the researcher to test the whole network while accounting for measurement error. It also permits comparisons of direct, indirect, and total effects. However, a statistically significant indirect effect does not by itself prove the proposed causal sequence. Temporal ordering, theory, design, and alternative explanations remain essential.

Covariance-based SEM is commonly used for theory testing and global model evaluation, whereas partial least squares SEM is often used for prediction-oriented work and complex models with formative indicators. Neither method eliminates the need for sound conceptualisation. Model-fit indices should be interpreted collectively rather than used as mechanical pass–fail criteria. Researchers should avoid repeatedly modifying a model in response to sample-specific fit statistics without theoretical justification.

In global research, measurement invariance is crucial. Configural invariance asks whether the same factor structure exists across markets. Metric invariance asks whether factor loadings are comparable, supporting comparisons of relationships. Scalar invariance concerns intercepts and is generally needed for comparing latent means. If full invariance is not achieved, partial invariance or alignment approaches may be considered, but conclusions must be qualified. Translation equivalence, response styles, acquiescence, and culturally specific interpretations can all threaten validity.

Experiments and the Logic of Causal Inference

Randomised experiments provide the clearest design for estimating intervention effects because treatment assignment, if successfully randomised, breaks the systematic connection between treatment and pre-existing characteristics. In digital marketing, experiments may compare creative executions, landing pages, recommendation systems, prices, messages, or promotional incentives. A/B tests are simple forms of randomised experiments, but doctoral-level practice requires attention to power, interference, treatment compliance, repeated testing, and business relevance.

A well-designed experiment specifies the unit of randomisation, treatment conditions, primary outcome, analysis population, duration, and stopping rule before examining results. Power calculations should reflect the minimum effect that would justify managerial action, not merely the smallest statistically detectable effect. Multiple outcomes and repeated experimentation increase the probability of false positives; pre-registration, correction procedures, or hierarchical testing can reduce this risk.

Field experiments can have high external validity but may suffer from contamination. Customers may see multiple treatments, share offers, or change behaviour because they know they are being observed. Cluster randomisation may be needed when treatment affects groups, stores, regions, or social networks. In global settings, treatment effects should be examined by market, but subgroup analysis must distinguish theoretically motivated heterogeneity from accidental noise.

When randomisation is impossible, observational causal methods may be useful. Propensity-score methods attempt to balance observed covariates between treated and untreated units. Difference-in-differences compares changes over time between an affected group and a comparison group, relying heavily on a parallel-trends assumption. Regression discontinuity exploits a threshold-based assignment rule. Instrumental-variable methods use a variable that affects treatment but influences the outcome only through treatment; credible instruments are difficult to find. Synthetic-control methods construct a weighted comparison unit for a treated market or region.

Each method replaces randomisation with assumptions. Researchers should report balance diagnostics, pre-treatment trends, placebo tests, robustness checks, sensitivity to unobserved confounding, and the practical scope of the estimated effect. A causal estimate is not automatically transportable: an intervention effective among existing customers in one country may not produce the same effect among prospective customers elsewhere.

Choice Models and Market Response

Choice models explain how consumers select among competing alternatives. In a multinomial logit model, the probability of choosing an option depends on its attributes and the consumer’s preferences. A utility function may include price, brand, delivery time, warranty, sustainability information, and country of origin. Choice probabilities can then support simulations of market share, price changes, product launches, or feature configurations.

The basic multinomial logit model relies on the independence of irrelevant alternatives assumption, which may be unrealistic when alternatives are close substitutes. Nested logit models relax some restrictions by grouping related options. Mixed logit models allow preferences to vary across consumers and can represent richer substitution patterns. Hierarchical Bayesian methods can estimate individual-level preferences, particularly when each respondent provides repeated choices.

Choice experiments require careful attribute selection, realistic levels, clear alternatives, and attention to respondent burden. Stated preferences may differ from revealed behaviour because hypothetical choices lack financial or practical consequences. Combining survey choices with transaction data can improve insight, but the integration process must address scale differences, sampling selection, and privacy.

In global marketing, choice utilities may vary because consumers attach different meanings to attributes. A fast delivery promise may be decisive in one market but less important where consumers prioritise price or local availability. Willingness-to-pay estimates should therefore be treated as model-based behavioural indicators rather than literal guarantees of purchase.

Text Analytics and Social Listening

Consumer voice data include reviews, complaints, call transcripts, open-ended survey responses, online forums, and social-media posts. Text analytics can identify topics, entities, emotions, sentiment, intentions, and emerging issues. A rigorous workflow begins with a research question and sampling frame, followed by data cleaning, language identification, coding or model selection, validation, and interpretation.

Dictionary-based sentiment analysis is transparent but sensitive to context, irony, negation, slang, and domain-specific language. Supervised classification can be more accurate when labelled training data are available, but performance depends on representativeness and drift. Topic models identify recurring word patterns, whereas modern language models can classify or summarise text. Human validation remains necessary, especially when decisions affect customers or employees.

Social-listening volume is not equivalent to consumer importance. Highly vocal users may not represent the customer base, and platform users vary across countries. Bots, coordinated campaigns, duplicate content, and algorithmic amplification can distort apparent attention. Sentiment scores should therefore be combined with reach, verified customer status, transaction evidence, and issue severity.

Multilingual analytics introduces additional risks. Direct translation may remove cultural nuance, and the same word may carry different emotional meanings across markets. Models should be evaluated separately by language and dialect. Data collection must comply with platform terms, privacy law, consent requirements, and organisational policies. De-identification and restricted access are particularly important for sensitive conversations.

Forecasting Demand and Behaviour

Forecasting supports inventory, media allocation, staffing, capacity planning, and financial decisions. Time-series methods model patterns in level, trend, seasonality, cycles, and shocks. Exponential smoothing is often effective for operational forecasts with evolving levels and seasonality. Autoregressive models use historical relationships, while dynamic regression incorporates predictors such as price, promotion, weather, exchange rates, or macroeconomic indicators.

Forecast evaluation should use rolling or expanding holdout periods rather than random cross-validation that destroys temporal order. Mean absolute error is easy to interpret, while root mean squared error penalises large errors more strongly. Percentage-based measures can be misleading when actual demand is near zero. Prediction intervals should accompany point forecasts because uncertainty is operationally significant.

Marketing interventions create a special forecasting problem. Historical data may reflect past prices and campaigns but provide weak evidence about a new strategy. Forecasts should distinguish baseline demand from incremental demand attributable to an intervention. Scenario analysis, structural breaks, and stress testing are particularly important in global markets affected by currency changes, regulation, supply disruption, or geopolitical events.

Customer Lifetime Value and Resource Allocation

Customer lifetime value (CLV) estimates the present value of expected future contribution from a customer or segment. A simplified formulation is:

CLV = Σ [(Expected revenue_t − Expected variable cost_t) × Retention probability_t] / (1 + discount rate)^t − Acquisition cost.

The equation is conceptually simple but operationally demanding. Retention, purchase frequency, order value, margin, service cost, discounting, and future marketing costs must be estimated. For contractual businesses, survival or churn models may be appropriate. For non-contractual settings, probabilistic models such as buy-till-you-die approaches estimate purchase frequency and dropout from transaction histories.

CLV should be calculated using contribution margin rather than revenue where possible. It should also account for uncertainty, referrals, returns, fraud, payment costs, and differences in service burden. A high-revenue customer is not necessarily a high-value customer if discounts and service costs are substantial. Similarly, a low-current-value customer may have high potential value if the firm can profitably improve retention or share of wallet.

The most useful managerial concept is often incremental CLV: the expected change in lifetime contribution resulting from a proposed action. Targeting customers solely on predicted CLV can produce weak decisions because the customers with the highest predicted value may purchase without intervention. Uplift modelling estimates who is persuadable, who is unaffected, and who may be harmed by an offer. This supports contact policies that consider both value and treatment responsiveness.

Integrating Analytics into Global Decisions

No single method answers every marketing question. A robust analytical programme often combines qualitative understanding, descriptive profiling, causal testing, forecasting, and economic valuation. For example, interviews may identify trust concerns; text analytics may reveal their frequency; SEM may test relationships among trust constructs; an experiment may evaluate reassurance messages; and CLV analysis may determine whether the intervention merits scale.

Integration should occur at the level of a decision, not merely through the accumulation of models. Managers should specify the decision, target population, intervention, outcome, time horizon, constraints, and threshold for action. Analysts should then identify the minimum evidence needed and make assumptions explicit. A dashboard of correlations is not a decision system unless it clarifies what action follows from each result.

A useful governance arrangement separates model development, validation, deployment, and monitoring. Models should be checked for performance drift, subgroup disparities, data leakage, unstable features, and unintended incentives. Human review is appropriate when automated decisions affect access, pricing, eligibility, or customer treatment. Documentation should record data provenance, variable definitions, exclusions, model versions, validation results, and approved uses.

Illustrative Case: Aurora Mobility’s Cross-Market Retention Decision

Background

Aurora Mobility, a fictional subscription-based urban mobility company, operates in three countries. The firm observes declining monthly retention in one market and proposes a personalised discount for customers predicted to churn. Executives initially believe that the highest-risk customers should receive the largest incentives.

Situation and Problem

The analytical team finds that customers who receive discounts are more likely to renew, but these customers also had unusually high prior engagement and were selected by local account managers. A simple comparison therefore cannot establish that discounts caused renewal. In addition, the proposed intervention could reduce margin among customers who would have renewed without assistance.

Analysis

The team first develops a churn model using tenure, usage frequency, service failures, payment history, and prior complaints. It evaluates performance separately by country and checks whether variables have comparable meanings across markets. Text analysis of service conversations identifies two distinct issues: price sensitivity in one market and reliability concerns in another. A latent-class analysis suggests that high churn risk contains at least two groups—price-responsive users and reliability-frustrated users.

Aurora then conducts a randomised field experiment. Eligible customers are randomly assigned to a control group, a modest discount, a reliability guarantee, or a service-recovery message. The primary outcome is renewal after three months, and the economic outcome is incremental contribution margin. The experiment is stratified by country and customer class to avoid imbalanced allocation.

Results show that the discount increases renewal among price-responsive customers but generates little incremental value among customers who would have renewed anyway. The reliability guarantee has a stronger effect among reliability-frustrated users, particularly in the market where service interruptions are more frequent. A CLV model indicates that the reliability intervention produces greater incremental lifetime contribution despite a smaller short-term renewal effect. Forecasting is used to estimate service-capacity requirements if the guarantee is scaled.

Approach and Solution

Aurora replaces its single churn-targeting rule with a treatment-policy framework. Customers are classified by predicted risk, expected response, incremental CLV, and potential harm from irrelevant contact. Discounts are reserved for customers with positive expected net value, while service-recovery interventions are directed to reliability-frustrated customers. The company monitors effects by market and recalibrates models as service conditions change.

Outcome

The company’s decision improves because the analysis distinguishes risk from persuadability and revenue from contribution. The case does not claim that a particular numerical uplift is universal; rather, it demonstrates how segmentation, text evidence, experimentation, causal reasoning, forecasting, and CLV can be connected to one decision.

Learning Points

First, predictive accuracy does not establish intervention effectiveness. Second, customer heterogeneity matters for both response and economics. Third, qualitative and text evidence can improve treatment design. Fourth, the right outcome is often incremental contribution rather than conversion alone. Finally, global implementation requires market-specific validation rather than automatic transfer of a central model.

Limitations, Risks, and Common Errors

Several recurring errors undermine consumer analytics. Treating correlation as causation is the most fundamental. Other problems include post-treatment adjustment, selection bias, survivorship bias, omitted variables, leakage from future information, unrepresentative samples, p-hacking, overfitting, and excessive reliance on statistical significance. Large datasets do not correct poor measurement or weak design.

Data quality problems are especially serious in global settings. Customer identifiers may not match across systems; currency and calendar definitions may differ; missingness may be systematic; and translations may alter constructs. Simpson’s paradox can occur when an aggregate relationship reverses within markets or segments. Researchers should inspect disaggregated patterns and test whether aggregation is substantively justified.

Models also create operational risks. A model may optimise clicks while reducing customer trust, maximise short-term sales while increasing returns, or target vulnerable consumers with aggressive persuasion. Forecasts can create self-fulfilling resource allocations, and automated sentiment systems can misclassify minority languages. Monitoring must therefore include business, customer, fairness, and reputational outcomes.

Ethics, Governance, and Professional Responsibility

Consumer analytics involves personal data, inferred attributes, and behavioural prediction. Responsible practice requires a lawful basis for processing, purpose limitation, data minimisation, appropriate retention, security, and meaningful transparency. Consent should not be treated as a universal solution when consumers lack real choice or cannot understand complex data uses.

Researchers should assess whether a variable is necessary, whether a proxy reproduces protected characteristics, and whether an intervention creates unequal treatment. Privacy-preserving approaches may include aggregation, pseudonymisation, access controls, secure computation, and differential privacy where appropriate. Ethical review is particularly important for sensitive data, vulnerable populations, location tracking, biometric information, and automated eligibility decisions.

Professional integrity also requires reporting null results, uncertainty, alternative specifications, and limitations. Managers should not present a model-based estimate as a fact independent of assumptions. In doctoral research, reproducibility, preregistration where suitable, transparent coding, and careful distinction between exploratory and confirmatory analysis strengthen credibility.

Emerging Directions

Consumer analytics is moving toward integrated, near-real-time decision systems that combine transaction, behavioural, text, and experimentation data. Generative AI can assist with coding, summarisation, synthetic data generation, and conversational analysis, but generated outputs require validation and may reproduce training-data biases. Causal machine learning methods increasingly seek to estimate heterogeneous treatment effects, while privacy-enhancing technologies aim to support measurement without centralising raw personal data.

The most important development is not the replacement of theory by automation. It is the combination of machine-scale pattern detection with explicit causal questions, high-quality measurement, experimental discipline, and human judgement. As regulatory expectations and consumer concerns increase, explainability, provenance, and accountable governance will become strategic capabilities rather than compliance afterthoughts.

Best Practices for Doctoral-Level Consumer Analytics

A rigorous programme should:

• begin with a clearly defined managerial and causal question;
• connect constructs and variables to established theory;
• establish measurement validity across relevant markets;
• distinguish prediction from causal estimation;
• pre-specify key outcomes and decision rules where feasible;
• validate models temporally, geographically, and across customer groups;
• report effect sizes, uncertainty, assumptions, and robustness checks;
• evaluate incremental profit or contribution, not only model accuracy;
• combine automated analysis with domain and cultural expertise; and
• monitor deployed interventions for drift, harm, and unintended consequences.

Chapter Summary

Consumer analytics extends beyond describing customers or predicting their behaviour. Its advanced purpose is to generate reliable insight for decisions about interventions, markets, products, communication, retention, and resource allocation. Multivariate methods reveal conditional relationships and latent structures; SEM connects measurement with theoretically specified pathways; experiments and observational designs support causal inference under explicit assumptions; choice models estimate responses to attributes and prices; text analytics and social listening capture consumer voice; forecasting anticipates demand; and CLV connects behaviour to economic value.

Global application requires attention to measurement invariance, cultural meaning, language, market heterogeneity, privacy, and transportability. No model can compensate for a weak research question or an invalid comparison. The strongest practice integrates theory, data quality, experimental or quasi-experimental design, uncertainty analysis, economic reasoning, and ethical governance. Most importantly, firms should target incremental value rather than merely high predicted value and should distinguish customers who are at risk from customers whose behaviour can profitably be changed.

Key Terms

Causal inference: Methods for estimating the effect of an intervention relative to a counterfactual outcome.

Confounder: A variable associated with both the treatment and the outcome that can distort their observed relationship.

Customer lifetime value: The discounted expected future contribution generated by a customer.

Difference-in-differences: A quasi-experimental method comparing changes over time between treated and comparison groups.

Factor analysis: A method for modelling relationships among observed variables through underlying latent factors.

Forecast interval: A range expressing uncertainty around a future prediction.

Incremental CLV: The change in lifetime contribution attributable to a marketing intervention.

Latent class: An unobserved probabilistic group used to represent heterogeneous consumer segments.

Measurement invariance: Evidence that a construct is measured comparably across groups or markets.

Mediator: A variable that explains part of the pathway between an intervention and an outcome.

Moderator: A variable that changes the strength or direction of a relationship.

Multinomial logit: A choice model estimating probabilities among multiple alternatives.

Potential outcomes: The outcomes that would occur for a unit under alternative treatment conditions.

Propensity score: The estimated probability that a unit receives a treatment given observed covariates.

Randomised experiment: A design assigning units to treatments by chance to support causal comparison.

Structural equation modelling: A framework combining latent-variable measurement models with hypothesised structural relationships.

Text analytics: Computational analysis of written or spoken language to identify patterns, meanings, or classifications.

Treatment effect heterogeneity: Variation in an intervention’s effect across people, segments, or markets.

Uplift modelling: Estimation of how an intervention changes behaviour for different types of customers.

Willingness to pay: The maximum amount a consumer is estimated to exchange for a product or attribute.

Review Questions

Short Answer Questions

1. Why does a correlation between advertising exposure and purchase not necessarily establish advertising effectiveness?

2. Distinguish a mediator from a moderator in consumer research.

3. What is measurement invariance, and why is it important in global marketing research?

4. Why may contribution margin be more appropriate than revenue in CLV calculations?

5. Identify two limitations of social-listening data.

Long Answer Questions

1. Explain how SEM combines measurement and structural models, and discuss the risks of interpreting mediation as proof of causality.

2. Compare randomised experiments with difference-in-differences and propensity-score methods.

3. Discuss the assumptions and managerial uses of multinomial logit, nested logit, and mixed logit models.

4. Develop an integrated analytics process for evaluating a new global customer-retention programme.

5. Critically evaluate the challenges of applying consumer analytics across languages and cultures.

Analytical / Application Questions

1. A company observes that customers receiving coupons renew at higher rates than customers who do not receive coupons. Design an analysis that would estimate the incremental effect of coupons and explain the assumptions required.

2. A global brand’s sentiment score falls sharply in one country, but sales remain stable. Identify possible explanations and propose a triangulated investigation using text analytics, transaction data, and qualitative research.

3. A firm’s churn model accurately identifies customers likely to leave but produces little profit when used for targeting. Explain why this may occur and design a treatment-policy approach using uplift modelling and incremental CLV.

22. Multiple Choice Questions with Answers

Question 1

A company finds that customers exposed to an advertisement purchase more frequently than unexposed customers. Which issue most directly prevents this comparison from proving a causal effect?

A. The outcome is measured as a purchase

B. Exposure may be associated with pre-existing purchase intent

C. The sample contains more than one customer

D. Advertising is a marketing activity

Correct Answer: B. Exposure may be associated with pre-existing purchase intent

Question 2

In a structural equation model, perceived trust explains how communication quality affects loyalty intention. Perceived trust is therefore a:

A. Moderator

B. Confounder

C. Mediator

D. Sampling unit

Correct Answer: C. Mediator

Question 3

Which condition is generally required to compare latent construct means across countries in a conventional SEM analysis?

A. Scalar measurement invariance

B. Perfect predictive accuracy

C. Identical sample sizes

D. A statistically significant treatment effect

Correct Answer: A. Scalar measurement invariance

Question 4

A firm wants to estimate the effect of a new landing page and randomly assigns visitors to the existing or new page. The principal advantage of this design is that it:

A. Eliminates all measurement error

B. Guarantees a large effect

C. Balances treatment assignment in expectation

D. Makes market differences irrelevant

Correct Answer: C. Balances treatment assignment in expectation

Question 5

Which customer should generally be prioritised under an incremental-CLV targeting policy?

A. A high-value customer who would buy without intervention

B. A low-value customer who cannot be reached

C. A customer whose intervention-related contribution is expected to exceed its cost

D. Every customer with a high predicted churn probability

Correct Answer: C. A customer whose intervention-related contribution is expected to exceed its cost

Question 6

The independence of irrelevant alternatives assumption is most closely associated with the basic:

A. Multinomial logit model

B. Difference-in-differences model

C. Factor analysis model

D. Survival model

Correct Answer: A. Multinomial logit model

Question 7

Why should a sudden increase in social-media mentions not automatically be interpreted as an increase in customer importance?

A. Social data never contain useful information

B. Mention volume may reflect bots, amplification, or vocal minorities

C. Transaction data cannot be analysed

D. Sentiment cannot be measured in any language

Correct Answer: B. Mention volume may reflect bots, amplification, or vocal minorities

Question 8

Which validation approach is most appropriate for a time-series demand forecast?

A. Randomly shuffling all observations before splitting the data

B. Using only the first observation as the test set

C. Rolling or expanding holdout evaluation

D. Evaluating fit only on the training period

Correct Answer: C. Rolling or expanding holdout evaluation

Question 9

In a difference-in-differences study, the parallel-trends assumption means that, absent the intervention:

A. Both groups would have had identical customer characteristics

B. The treatment group would have experienced the same outcome trend as the comparison group

C. Both groups must receive the same treatment

D. The intervention must be randomly assigned

Correct Answer: B. The treatment group would have experienced the same outcome trend as the comparison group

Question 10

Which practice best addresses cross-market risks in automated text analytics?

A. Use one translated sentiment dictionary without validation

B. Ignore dialect and platform differences

C. Evaluate performance separately by language and validate with human coding

D. Treat all online authors as representative customers

Correct Answer: C. Evaluate performance separately by language and validate with human coding