8. Multivariate and Predictive Market Analytics

Introduction and Context

Advertising decisions rarely depend on a single variable. A campaign may succeed because the creative idea is distinctive, the audience is appropriately selected, the media schedule is well timed, the offer is attractive, and the brand already enjoys trust. Conversely, weak performance may result from several interacting causes. Multivariate market analytics provides the methods required to examine these simultaneous influences rather than treating each factor in isolation.

In earlier stages of marketing research, analysts often describe audiences through frequencies, averages, cross-tabulations, and simple comparisons. These techniques remain useful, but postgraduate-level analysis requires a more integrated perspective. The analyst must estimate relationships while controlling for other variables, identify meaningful customer groups, predict likely behaviour, understand latent attitudes, and translate statistical evidence into decisions about targeting, positioning, campaign design, and resource allocation.

This chapter examines regression, logistic regression, analysis of variance, factor analysis, clustering, segmentation, discriminant analysis, and conjoint analysis. These methods address different questions. Regression estimates the size and direction of relationships; logistic models predict categorical outcomes; ANOVA compares group means; factor analysis reduces observed measures to underlying dimensions; clustering discovers groups; discriminant methods classify cases into known groups; and conjoint analysis estimates the value consumers attach to product or communication attributes. Used together, they create an analytical chain from data to decision.

The central principle is that sophisticated modelling does not compensate for weak research design. A statistically impressive model can still mislead if the sample is biased, variables are poorly measured, assumptions are ignored, or correlation is interpreted as causation. Therefore, model evaluation, managerial judgement, ethics, and transparent communication are treated as integral parts of the analytical process.

Learning Objectives

After studying this chapter, students should be able to:

1. Explain the purpose, logic, and appropriate use of major multivariate methods in advertising and market research.

2. Select suitable dependent and independent variables for regression, logistic models, ANOVA, factor analysis, clustering, discriminant analysis, and conjoint analysis.

3. Interpret coefficients, significance tests, fit measures, effect sizes, classification results, segments, and attribute utilities.

4. Evaluate assumptions, model performance, validity, reliability, and the risk of overfitting.

5. Apply analytical findings to targeting, positioning, media allocation, creative strategy, and offer design.

6. Recognise ethical, technical, and managerial limitations in predictive market analytics.

The Analytical Problem: Many Variables, One Decision

Multivariate analysis studies situations in which several variables are examined simultaneously. A typical advertising dataset may include exposure frequency, reach, media channel, creative execution, brand familiarity, perceived relevance, purchase intention, income, age, geography, device usage, and actual conversion. The analyst must distinguish between variables that explain an outcome, variables that merely correlate with it, and variables that create noise.

A useful conceptual model begins with four elements:

• Outcome: what the organisation wants to explain or predict, such as sales, brand consideration, click-through, conversion, or churn.
• Predictors: variables believed to be associated with the outcome, such as exposure, price, message relevance, or audience characteristics.
• Controls: variables included to reduce confounding, such as seasonality, region, prior brand usage, or campaign period.
• Decision rule: the managerial action linked to the result, such as increasing investment, changing creative, or prioritising a segment.

Prediction and explanation are related but not identical. An explanatory model seeks interpretable relationships that support theory or managerial understanding. A predictive model seeks accurate performance on new observations. A variable may improve prediction without offering a convincing causal explanation. For example, a browsing pattern may predict conversion but may not be a practical lever for management.

The measurement level of each variable matters. Continuous outcomes such as revenue or attitude scores invite linear regression. Binary outcomes such as purchase versus no purchase require logistic regression. Nominal categories, ordered ratings, counts, and repeated observations may require other models or transformations. Incorrectly treating a categorical variable as continuous can produce misleading estimates and inappropriate conclusions.

Regression as a Foundation for Market Explanation

Multiple Linear Regression

Multiple linear regression estimates the relationship between a continuous dependent variable and several predictors. In simplified form:

Y = β₀ + β₁X₁ + β₂X₂ + ... + βₖXₖ + ε

Here, Y may represent brand preference, sales, or campaign response; the X variables may represent advertising spend, reach, perceived relevance, and price; β coefficients estimate the expected change in Y associated with a one-unit change in a predictor while other included predictors are held constant; and ε represents unexplained variation.

The phrase “holding other variables constant” is important. If a coefficient for perceived relevance is positive, the model estimates its association with the outcome after accounting for the other variables in the model. This is not automatically proof that increasing relevance will cause the outcome to rise. Causal interpretation requires appropriate design, temporal ordering, credible controls, and often experimentation.

Regression output commonly includes coefficient estimates, standard errors, confidence intervals, significance tests, R-squared, adjusted R-squared, and an overall F-test. R-squared indicates the proportion of observed variation explained by the model in the sample. Adjusted R-squared penalises the addition of unnecessary predictors. Neither measure alone establishes that the model is useful for prediction or that the relationships are causal.

Interaction terms allow the effect of one variable to depend on another. For example, the effect of social media exposure may differ between high- and low-involvement consumers. A model containing exposure, involvement, and their interaction can represent this conditional relationship. Analysts should interpret interactions graphically or through predicted values rather than relying only on the coefficient table.

Regression Assumptions and Diagnostics

Linear regression generally assumes linearity, independent observations, appropriately specified errors, constant error variance, and limited problematic multicollinearity. Normality of residuals is particularly relevant for small-sample inference, although prediction may remain acceptable when samples are large and other assumptions are reasonable.

Diagnostics should include residual plots, tests or visual checks for non-constant variance, leverage and influence measures, and variance inflation factors for multicollinearity. Multicollinearity occurs when predictors overlap strongly. It inflates standard errors and makes individual coefficients unstable, even when the model predicts reasonably well. The remedy may involve combining variables, removing redundant measures, using theory to choose among them, or applying regularisation.

Outliers deserve investigation rather than automatic deletion. A high-value customer, unusually large order, or exceptional campaign week may be a genuine observation. Removing it merely because it weakens a desired conclusion is poor practice. Analysts should document the rule used and assess results with and without influential cases.

Logistic Regression for Behavioural Prediction

Many market outcomes are categorical. A consumer either converts or does not convert; a lead may be qualified or unqualified; a respondent may choose one brand over another. Logistic regression models the probability of an event. It uses the logit transformation of probability so that predicted values remain between zero and one.

The logistic model can be expressed as:

log[p/(1-p)] = β₀ + β₁X₁ + ... + βₖXₖ

The exponentiated coefficient, exp(β), is an odds ratio. An odds ratio above one indicates higher odds of the event as the predictor increases, controlling for other variables; an odds ratio below one indicates lower odds. Odds are not the same as probability, so managers should also receive predicted probabilities for realistic customer profiles.

Performance should be evaluated using a confusion matrix, accuracy, precision, recall, specificity, F1 score, and—where appropriate—ROC-AUC or precision-recall curves. Accuracy can be deceptive when the event is rare. If only two per cent of prospects convert, a model that predicts “no conversion” for everyone achieves 98 per cent accuracy but has no targeting value. Threshold selection should reflect the cost of contacting a prospect, the cost of missing a likely customer, and available capacity.

Comparing Groups: ANOVA and Experimental Reasoning

Analysis of variance, or ANOVA, tests whether the means of a continuous outcome differ across two or more groups. In advertising, a researcher might compare brand attitude across three creative executions, average purchase intention across media conditions, or sales across regions.

ANOVA partitions total variation into variation between groups and variation within groups. The F-statistic compares these sources of variation. A significant omnibus test indicates that not all group means are equal, but it does not identify which groups differ. Post-hoc comparisons, such as Tukey-type procedures, are required when multiple groups are compared.

A two-way ANOVA examines two categorical factors simultaneously and can test interaction. For instance, creative format and audience involvement may jointly influence persuasion. A significant interaction means that the effect of creative format depends on involvement level. Main effects should not be interpreted in isolation when a substantial interaction is present.

Important assumptions include independence, approximate normality of residuals, and homogeneity of variance. When variances are unequal, robust alternatives or appropriate transformations may be preferable. Statistical significance should be accompanied by effect size and confidence intervals. A tiny difference can be statistically significant in a very large sample while having little managerial importance.

ANOVA is strongest when embedded in a sound experimental design. Random assignment improves causal inference. Observational comparisons of existing customer groups may reveal associations but should not be described as controlled effects without qualification.

Discovering Structure in Attitudes: Factor Analysis

Market research instruments often contain many correlated statements: “This brand is trustworthy,” “This brand keeps its promises,” and “This brand is dependable.” Factor analysis identifies a smaller number of latent dimensions that account for patterns of correlation among observed variables.

Exploratory factor analysis is used when the underlying structure is uncertain. The researcher examines factor loadings, communalities, eigenvalues, scree plots, and theoretically sensible solutions. Rotation improves interpretability. Oblique rotation permits factors to correlate and is often appropriate for psychological constructs, because perceived quality, trust, and relevance may be related.

Confirmatory factor analysis tests a hypothesised measurement structure. It evaluates whether observed indicators represent specified constructs and may report fit indices such as comparative fit measures, residual-based measures, and information criteria. Fit should be considered alongside theory, reliability, validity, and sample size rather than treated as a mechanical pass-or-fail exercise.

Factor analysis supports scale development, brand-image mapping, questionnaire reduction, and construction of composite variables for subsequent models. Analysts must distinguish statistical factors from substantive constructs. A factor is useful only when its indicators share a defensible conceptual meaning. Reliability measures can support internal consistency, but high reliability alone does not establish validity.

From Variables to Audiences: Clustering and Segmentation

Segmentation divides a market into groups that are relatively similar within themselves and meaningfully different from one another. Clustering is a data-driven approach to discovering such groups. Common approaches include hierarchical clustering and partitioning methods such as k-means. Hierarchical methods build a tree of solutions; k-means assigns observations to a pre-specified number of clusters by minimising within-cluster distance.

The number of clusters should not be selected solely because a software output suggests it. Analysts should combine statistical criteria, stability across samples, interpretability, segment size, reachability, and managerial usefulness. Standardising variables is often necessary when measures use different scales. Otherwise, a variable measured in large numerical units can dominate distance calculations.

A useful segmentation is not merely a demographic description. It should identify a difference in needs, motivations, media behaviour, profitability, responsiveness, or strategic opportunity. Segments should be profiled using variables not used to create them, when possible, to reduce circular interpretation. They should also be actionable: the organisation must be able to identify, reach, and serve them.

Clustering is exploratory and sensitive to variable selection, scaling, outliers, and algorithm choice. A cluster solution found in one sample may not reproduce in another. Validation through holdout samples, replication, or external behavioural outcomes is therefore essential.

Classification with Discriminant Methods

Discriminant analysis classifies cases into known groups using predictor variables. Unlike clustering, which discovers groups, discriminant analysis begins with an established group label—for example, loyal versus occasional customers—and estimates functions that distinguish them.

Linear discriminant analysis commonly assumes multivariate normality within groups and similar covariance matrices. It can provide classification functions, group centroids, posterior probabilities, and a confusion matrix. If covariance structures differ substantially, quadratic discriminant analysis or other classification methods may be more appropriate.

Classification accuracy should be compared with a baseline and assessed on data not used to estimate the model. Cross-validation helps identify overoptimistic performance. In marketing applications, misclassification costs are rarely equal. Misclassifying a high-value prospect as low priority may be more costly than contacting a marginal prospect, so the decision rule should reflect economic consequences.

Discriminant analysis is especially useful when an organisation has meaningful segment definitions and needs to assign new prospects to them. However, classification should not be confused with discovering psychologically natural categories. The quality of the original group labels determines the quality of the classifier.

Conjoint Analysis: Measuring Preference Trade-offs

Consumers evaluate offers as bundles of attributes. A mobile plan, for example, may combine price, data allowance, network quality, and contract flexibility. Conjoint analysis decomposes evaluations of profiles into part-worth utilities associated with attribute levels.

The basic process involves defining relevant attributes and realistic levels, constructing a manageable set of product or service profiles, asking respondents to rate, rank, or choose among alternatives, and estimating utilities. Choice-based conjoint designs are particularly useful when respondents select between competing alternatives. Attribute importance is derived from the range of utilities within an attribute, although importance values depend on the selected levels and should not be interpreted as universal psychological weights.

Conjoint analysis can estimate willingness to pay, simulate market shares under alternative configurations, identify preference heterogeneity, and reveal trade-offs. It is valuable for positioning and offer design because it asks not simply what respondents claim to like, but how they choose when desirable features compete.

Design quality is crucial. Too many attributes create cognitive burden; unrealistic combinations weaken validity; and omitted attributes can force respondents to overvalue included ones. Simulations are scenario estimates, not guaranteed forecasts. They should be calibrated against observed market behaviour where possible.

Building an Integrated Analytical Workflow

A strong multivariate project follows a reasoned sequence rather than beginning with whichever technique is available in software.

First, define the decision and the estimand. Is the organisation trying to predict conversion, estimate a campaign effect, identify segments, or optimise an offer? Second, clarify the data-generating process, including sampling, timing, measurement, missingness, and possible sources of bias. Third, inspect distributions, relationships, coding, and unusual observations. Fourth, select a method consistent with the outcome, research design, and theoretical purpose.

The data should then be divided appropriately into training, validation, and test sets when prediction is central. Cross-validation is useful when data are limited. Feature selection must occur within the resampling procedure; selecting variables using the full dataset before validation leaks information and creates overly optimistic results. Hyperparameter tuning, transformations, and imputation should also be performed without allowing test-set information to influence the model.

The fifth stage is estimation and diagnosis. Analysts assess assumptions, residuals, collinearity, class imbalance, stability, and sensitivity to alternative specifications. Sixth, performance is evaluated using measures appropriate to the task. Seventh, findings are translated into profiles, scenarios, financial implications, and recommended actions. Finally, results are monitored after implementation because market conditions, media costs, consumer behaviour, and competitive actions change.

A model should be judged by both statistical and managerial criteria. A slightly less accurate model may be preferable if its logic is transparent, its variables are available before the decision, and its recommendations can be acted upon. Conversely, a highly interpretable model may be insufficient when the cost of prediction errors is substantial.

Applied Case: Choosing Growth Targets for a Subscription Brand

Background

The following is an illustrative case. “StreamPlus,” a fictional subscription entertainment service, has a large prospect database and a limited promotional budget. The organisation wants to increase paid subscriptions without repeatedly targeting people who would have subscribed anyway.

Situation and Problem

The marketing team has campaign exposure data, website behaviour, survey measures of entertainment needs, household characteristics, prior trial status, and subscription outcomes. A simple demographic rule produces a large audience but a low conversion rate. Senior management asks whether analytics can identify priority audiences and guide message and offer design.

Analysis

The research team first defines conversion within thirty days as the primary outcome. A logistic regression includes prior trial, recent site visits, content-category interest, price sensitivity, advertising exposure, device usage, and region. The team reports predicted probabilities rather than treating coefficients as direct percentage-point changes. It discovers that prior trial and recent site visits are strong predictors, while exposure has a smaller association after controls are included.

An ANOVA of an experimental message test compares three creative themes: exclusive content, family value, and flexible cancellation. The overall test indicates differences in purchase intention, and post-hoc comparisons show that flexible cancellation performs particularly well among respondents with high price sensitivity. A two-way analysis reveals an interaction between message theme and sensitivity level.

Factor analysis of twelve attitude statements produces interpretable dimensions labelled content excitement, perceived risk, and service convenience. These factor scores are used to profile audiences, not as unquestioned psychological truths. A clustering analysis using needs, viewing behaviour, and price sensitivity identifies three candidate segments: content enthusiasts, convenience-oriented households, and cautious value seekers. The solution is checked for stability and external differences in conversion and retention.

Finally, a choice-based conjoint study compares monthly price, advertising presence, offline viewing, content access, and cancellation flexibility. Simulations suggest that cancellation flexibility has particular value among cautious value seekers, whereas content access dominates among enthusiasts. The results do not justify offering every feature to every customer; they support differentiated propositions.

Strategy and Solution

StreamPlus creates a propensity score for prioritising prospects but uses a separate uplift-oriented test to determine whom advertising changes. High-propensity prospects are not automatically treated as the only valuable audience, because some would convert without intervention. The company assigns content enthusiasts to exclusive-content creative, cautious value seekers to risk-reducing messages, and convenience-oriented households to practical-use communication.

The media plan uses predicted value, contact cost, expected incremental response, and capacity constraints. A holdout group remains unexposed so that incremental conversions can be estimated. The team monitors conversion, cancellation, frequency, and complaints rather than optimising only short-term sign-ups.

Outcome

The illustrative case does not claim a universal numerical result. Its outcome is a more disciplined decision system: audiences are selected using behavioural and attitudinal evidence, messages are matched to segment needs, offer simulations inform design, and campaign impact is assessed against a control or holdout. This reduces the risk of confusing correlation with incremental effectiveness.

Learning Points

The case demonstrates that no single technique answers every managerial question. Logistic regression supports prioritisation; ANOVA evaluates controlled message comparisons; factor analysis organises attitudes; clustering creates exploratory audience groups; conjoint analysis examines trade-offs; and experimental validation tests whether targeting produces incremental value.

Challenges, Ethics, and Governance

Multivariate analytics creates several risks. Sampling bias can make a model systematically inaccurate for underrepresented audiences. Historical campaign data may encode prior targeting decisions, meaning that the model learns who was contacted rather than who would respond under equal opportunity. Proxy variables may reproduce sensitive characteristics even when those characteristics are excluded explicitly. Excessive personalisation may feel intrusive and undermine trust.

Privacy principles require purpose limitation, data minimisation, appropriate consent or lawful basis, secure storage, controlled access, and defined retention periods. Sensitive data should receive heightened protection. Organisations should explain important automated decisions where required, provide human oversight, and establish procedures for contesting harmful outcomes.

Analysts should document data sources, variable definitions, missing-data treatment, exclusions, model versions, validation design, thresholds, and known limitations. Fairness should be assessed across relevant groups using measures suited to the decision context. There is no single fairness metric that resolves every trade-off; false-positive and false-negative costs must be considered explicitly.

Common professional mistakes include running many models until a desirable p-value appears, reporting accuracy without a baseline, treating statistical significance as business significance, using post-treatment variables as predictors, validating on the training data, assuming segments are permanent, and presenting a forecast as a certainty. Transparent uncertainty is a mark of analytical quality, not weakness.

Technology, Automation, and Emerging Practice

Contemporary analytics platforms automate data preparation, model comparison, audience scoring, and dashboard reporting. Machine-learning methods such as regularised regression, random forests, gradient boosting, and neural networks may improve predictive performance, particularly with complex and nonlinear data. However, advanced algorithms do not eliminate the need for valid labels, leakage control, calibration, interpretability, and governance.

Generative AI can assist with code drafting, variable documentation, scenario generation, and plain-language explanation, but its outputs require verification. Automated systems may also optimise toward short-term clicks while degrading brand outcomes, customer quality, or long-term profitability. Marketing organisations increasingly need model cards, audit trails, drift monitoring, and clear ownership of automated decisions.

Emerging practice is moving from descriptive targeting toward incrementality, customer lifetime value, causal inference, privacy-preserving measurement, and continuous experimentation. The important shift is not merely technological. It is conceptual: the question is becoming not only “Who is likely to respond?” but also “Who will respond because of this intervention, at what cost, and with what long-term consequence?”

Best Practices for Professional Use

A sound practice is to begin with a decision, not a method. Use theory and prior research to choose variables, preserve temporal ordering, and distinguish prediction from causal explanation. Predefine primary outcomes where feasible, inspect data before modelling, and validate using procedures that mirror future use.

Report effect sizes and uncertainty, not only significance. Translate coefficients into predicted outcomes, financial scenarios, or audience counts that managers can understand. Compare models against simple baselines. Test stability across time, markets, and demographic groups. Combine quantitative results with qualitative understanding when interpreting segments and motivations.

Most importantly, connect analytical output to an accountable action: a revised brief, a media allocation, a testable proposition, a service change, or a decision not to target. Analytics becomes strategically valuable when evidence changes behaviour and when the organisation learns from the results.

Chapter Summary

Multivariate and predictive market analytics enables researchers to study several influences simultaneously and connect evidence with advertising decisions. Multiple regression estimates continuous outcomes, while logistic regression predicts categorical events and requires careful attention to probabilities, thresholds, class imbalance, and validation. ANOVA compares group means and can reveal interactions in experimental designs. Factor analysis identifies underlying dimensions among correlated measures. Clustering discovers potential segments, whereas discriminant analysis classifies cases into established groups. Conjoint analysis estimates preference trade-offs and supports offer and positioning decisions.

These methods must be selected according to the decision, measurement structure, research design, and intended use. Assumption checking, effect sizes, holdout validation, cross-validation, sensitivity analysis, and transparent reporting are essential. A useful model is not necessarily the most complex or the one with the highest in-sample fit. It is a model that is sufficiently accurate, interpretable, stable, ethical, and connected to an actionable decision. Contemporary practice increasingly combines predictive modelling with experimentation, incrementality analysis, privacy governance, and continuous monitoring.

Key Terms

Multivariate analysis: Statistical analysis involving multiple variables examined together.

Dependent variable: The outcome a model seeks to explain or predict.

Predictor: An explanatory or input variable used in a model.

Multiple linear regression: A model estimating a continuous outcome from several predictors.

Interaction effect: A relationship in which the effect of one predictor depends on another variable.

Multicollinearity: Strong correlation among predictors that can destabilise individual coefficient estimates.

Logistic regression: A model estimating the probability of a categorical, commonly binary, outcome.

Odds ratio: The multiplicative change in event odds associated with a one-unit predictor change.

ANOVA: A method for testing mean differences across two or more groups.

Effect size: A measure of the practical magnitude of a relationship or group difference.

Factor analysis: A method for identifying latent dimensions underlying correlated observed variables.

Factor loading: The strength of association between an observed variable and a latent factor.

Cluster analysis: An exploratory method for grouping similar observations.

Segmentation: Division of a market into groups with strategically meaningful similarities and differences.

Discriminant analysis: A classification method assigning observations to known groups using predictors.

Conjoint analysis: A preference method estimating the value consumers attach to attribute levels.

Cross-validation: Repeated partitioning of data to evaluate model performance on unseen observations.

Overfitting: A condition in which a model captures sample-specific noise and performs poorly on new data.

Incrementality: The outcome caused by an intervention beyond what would have occurred without it.

Model drift: Decline in model performance as relationships or data patterns change over time.

Review Questions

Short Answer Questions

1. What is the difference between prediction and causal explanation in market analytics?

2. Why can a high R-squared value fail to establish that a model is useful or causal?

3. What does an odds ratio represent in logistic regression?

4. Why are post-hoc comparisons needed after a significant ANOVA?

5. Distinguish clustering from discriminant analysis.

Long Answer Questions

1. Explain the assumptions and diagnostic procedures relevant to multiple linear regression in advertising research.

2. Discuss how factor analysis can support questionnaire development and brand-positioning research.

3. Explain the process of designing and interpreting a segmentation study using clustering methods.

4. Describe how conjoint analysis can inform offer design, pricing, and positioning, including its major limitations.

5. Develop an integrated workflow for building and validating a predictive campaign-response model.

Analytical / Application Questions

1. A campaign-response model has 96 per cent accuracy, but only 4 per cent of prospects convert. Evaluate whether the model is useful and recommend additional performance measures.

2. A two-way ANOVA shows a significant interaction between creative format and audience involvement. Explain how the advertiser should interpret and apply this result.

3. A clustering solution produces five statistically distinct groups, but one group represents only 1 per cent of the market and cannot be reached through available media. Should it be retained? Justify your answer using analytical and managerial criteria.

22. Multiple Choice Questions with Answers

Question 1

Which statement best distinguishes an explanatory model from a predictive model?

A. An explanatory model always has higher accuracy than a predictive model.

B. An explanatory model focuses on interpretable relationships, whereas a predictive model focuses on performance for new observations.

C. A predictive model can only use categorical variables.

D. An explanatory model does not require data validation.

Correct Answer: B. An explanatory model focuses on interpretable relationships, whereas a predictive model focuses on performance for new observations.

Question 2

A logistic regression coefficient has an exponentiated value of 1.8. What does this most directly indicate?

A. The predicted probability is exactly 1.8.

B. The outcome increases by 1.8 percentage points.

C. The odds of the event are multiplied by 1.8 for a one-unit predictor increase, holding other variables constant.

D. The model explains #### 80 per cent of outcome variance.

Correct Answer: C. The odds of the event are multiplied by 1.8 for a one-unit predictor increase, holding other variables constant.

Question 3

Why can accuracy be misleading when evaluating a conversion model?

A. Accuracy cannot be calculated from a confusion matrix.

B. Accuracy may appear high when the model predicts the majority class for nearly everyone.

C. Accuracy is only used for linear regression.

D. Accuracy automatically measures causal impact.

Correct Answer: B. Accuracy may appear high when the model predicts the majority class for nearly everyone.

Question 4

A significant omnibus ANOVA result establishes that:

A. Every group mean differs from every other group mean.

B. At least one group mean differs from another, requiring further comparisons to locate the differences.

C. The independent variable causes the outcome in all research designs.

D. The within-group variance is zero.

Correct Answer: B. At least one group mean differs from another, requiring further comparisons to locate the differences.

Question 5

What is the primary purpose of rotation in exploratory factor analysis?

A. To increase the sample size.

B. To make the factors more interpretable by clarifying the loading pattern.

C. To remove all measurement error.

D. To convert categorical outcomes into probabilities.

Correct Answer: B. To make the factors more interpretable by clarifying the loading pattern.

Question 6

Which feature is most important for an actionable market segment?

A. It is the smallest group produced by the algorithm.

B. It is defined only by age and gender.

C. It can be identified, reached, and served through a meaningful proposition.

D. It always contains equal numbers of customers.

Correct Answer: C. It can be identified, reached, and served through a meaningful proposition.

Question 7

How does discriminant analysis differ from clustering?

A. Discriminant analysis begins with known group labels, while clustering discovers groups.

B. Discriminant analysis cannot classify new cases.

C. Clustering requires a dependent variable and discriminant analysis does not.

D. They are identical methods with different names.

Correct Answer: A. Discriminant analysis begins with known group labels, while clustering discovers groups.

Question 8

What does conjoint analysis primarily estimate?

A. The causal effect of advertising exposure on sales.

B. The relative utility consumers attach to attribute levels and the trade-offs among them.

C. The statistical significance of a regression coefficient.

D. The number of latent factors in an attitude scale.

Correct Answer: B. The relative utility consumers attach to attribute levels and the trade-offs among them.

Question 9

A model performs very well on training data but poorly on a test sample. This is most consistent with:

A. Overfitting.

B. Random assignment.

C. Homogeneous variance.

D. Incrementality.

Correct Answer: A. Overfitting.

Question 10

Why should a high-propensity audience not automatically be treated as the best advertising target?

A. High-propensity audiences cannot be measured.

B. They may have converted without advertising, so propensity does not establish incremental response.

C. Propensity models only apply to brand attitudes.

D. Advertising can never affect high-propensity consumers.

Correct Answer: B. They may have converted without advertising, so propensity does not establish incremental response.