ST3188 Statistical Methods for Market

Research
Strategic Exam Study Guide — Question Matrix, Coverage Strategy & Model Answers

1. The Question Matrix Across 6 Sittings

The exam is 2 hours: Section A (1 compulsory question = 40 marks) + Section B (answer 2 of 3, each worth 30
marks). Total 100 marks. The structure is highly stable: Q1 always tests sampling techniques (1a) plus sample-
size derivation (1b), and Q2(a) in Section B is always a multiple linear regression SPSS interpretation. Only the
rotating components (the (b) sub-questions and the multivariate technique tested in Q3(a)/Q4(a)) actually
vary.

1.1 The (b) sub-questions across all 6 sittings

Sitting 1(b)iii — sample-size 1(b)iv — sample-size Q2(b) [10] Q3(b) [10] Q4(b) [10]
context context
2023 May Effect of population Incidence rate & Post-test-only control Comparative scales Uses of factor analysis
variance σ² completion rate group design (steps + (purpose + 3 in market research
(effect on n) symbolic) examples)
2023 Oct Effect of confidence 95% CI for proportion Semantic differential Marketing-system Response rate, non-
level (1−α) (construction & vs Likert + verbal model (4 mix vars, response, strategies
interpretation) description + situational, across data-collection
construct example behavioural vs methods
performance)
2024 May Qualitative factors in Choice of confidence Stepwise regression Conjoint-analysis Projective techniques
choosing n for level for CI of a (purpose + model U(X) (define (definition + word
proportion proportion forward/backward/st terms + relative association + story
epwise) importance) completion)
2024 Oct Finite population Difference between One-way ANOVA Three research Semantic differential
correction (purpose + incidence rate and decomposition of designs (exploratory, and Stapel scale
formula) completion rate total variation descriptive, causal) + (purpose + examples)
examples
2025 May Define incidence rate Adjust statistically- Semantic differential Non-response bias Cluster vs
& completion rate determined n for vs Likert + verbal (definition + 3 ways discriminant analysis
incidence/completion description + improve, 3 adjust (similar/different +
construct example techniques) uses + interpretation)
2025 Oct Choice of confidence Finite population Post-test-only control One-way ANOVA Three research
level for CI on mean correction (purpose + group design (steps + decomposition of designs (exploratory,
willingness to pay formula) symbolic) total variation descriptive, causal) +
examples

1.2 The (a) sub-questions in Section B (from the 3 paper PDFs)

Although the matrix only catalogues the (b) parts, the corresponding (a) parts (worth 20 marks each) follow an
even more predictable pattern:

Paper (PDF in pack) Q2(a) — 20 marks Q3(a) — 20 marks Q4(a) — 20 marks

2022/23 academic year Multiple regression Factor analysis (lifestyle V1– Cluster analysis, Ward (Likert
(supermarket sales) V7) items)
2023/24 academic year Multiple regression (hotel Discriminant analysis (credit- Factor analysis (IT satisfaction)
 margins) card default)
2024/25 academic year Multiple regression (credit- Discriminant analysis (3 job Cluster analysis, Ward
card charges) classifications) (countries)

Take-aways from this pattern:

 Q2(a) is ALWAYS a multiple linear regression SPSS-output interpretation. This is a guaranteed 20 marks if
mastered.
 Q3(a) and Q4(a) jointly draw from {factor analysis, cluster analysis (Ward), discriminant analysis}. Across 3
papers, every paper contained 2 of these 3 — never the same pair.
 Q1(a) is ALWAYS the same 4 sampling techniques (convenience, quota, stratified, cluster), only the business
context (Starbucks, smartphones, cinema) changes.
 Q1(b) is ALWAYS sample-size derivation: parts (i) sampling distribution, (ii) estimator + its distribution, (iii)
the n ≥ z²σ²/e² formula derivation, plus (iv) a context part on FPC, incidence/completion rates, or CI level.
2. How to Score 100/100 in Every Paper
The principle: don't learn everything — learn enough that the union of your topics covers Q1 (compulsory, 40
marks) and any 2 of {Q2, Q3, Q4} (60 marks). Because Q2 is always regression, the cleanest strategy is "always
do Q2 + one alternate". Below are the topics you MUST master and a small set of optional add-ons for safety.

2.1 Section A — must master (40 guaranteed marks)

Q1(a) — 4 sampling techniques. Each must be explained as: (1) probability vs non-probability, (2) mechanics
of selection, (3) merits, (4) limitations + selection bias considerations, (5) how it would be applied in the
question's business scenario:

 Convenience sampling (non-probability) — easiest, cheapest; severe selection bias; useful only for
exploratory/pilot work.
 Quota sampling (non-probability) — fixes proportions on key strata (age, region) but selection within quota
is non-random; more representative than convenience but still biased.
 Stratified sampling (probability) — population divided into mutually exclusive strata, SRS within each; needs
a sampling frame and a meaningful stratifier; reduces sampling error if strata are homogeneous within and
heterogeneous between.
 Cluster sampling (probability) — population divided into clusters; whole clusters chosen at random; cheaper
for geographically dispersed populations but higher sampling error than stratified for the same n.

Q1(b) — sample-size derivation. Memorise these four moves cold:

 Sampling distribution = "the probability distribution of an estimator over infinitely many repeated samples".
 Estimator: X̄ ~ N(μ, σ²/n) for the mean (exactly if X is normal; approximately by CLT). For a proportion: P̂ = X̄
~ N(π, π(1−π)/n) approximately by CLT.
 Min n derivation: from z_{α/2} · SE ≤ e you get n ≥ z²_{α/2} σ² / e² (mean) or n ≥ z²_{α/2} π(1−π) / e²
(proportion). For the proportion case, set π = 0.5 if unknown — this is the conservative (worst-case) value
because π(1−π) is maximised at 0.5.
 Finite population correction: when n/N ≥ 0.10 use n_c = nN/(N + n − 1). The FPC reduces the required
sample size when n is a non-trivial fraction of N because each additional draw provides more information
from a finite pool.

Don't forget: incidence rate (proportion of population eligible) and completion rate (proportion of contacted
eligibles who finish the survey). Adjusted final n = (statistical n) / (incidence × completion). They lower IR or CR
→ larger required initial sample.

2.2 Section B — strategy: always Q2 + one alternate

Because Q2(a) is always multiple linear regression, you should always plan to do Q2. Then master enough to be
able to do AT LEAST ONE of Q3 or Q4 in any sitting. Concretely:

Tier 1 — non-negotiable (these alone secure ≥80/100 in every past paper):

 Multiple linear regression: theoretical model β specification + assumptions (ε ~ N(0, σ²)), F-test for joint
significance, individual t-tests, R² interpretation, standardised coefficients for relative importance, dummy-
 variable coefficient interpretation, residual normality plot, recommending model changes (drop insignificant
variables, add omitted ones, address multicollinearity).
 Multicollinearity, stepwise regression, relative importance in MLR, two-way ANOVA interaction — these are
the (b) topics that pair naturally with Q2(a). Of the last 3 papers, 2 had a regression-related (b) (multicoll.
2023; rel. importance 2025).
 Cluster analysis (Ward): minimise within-cluster variance, squared Euclidean distance δᵢ ⱼ = Σ ₖ(xᵢ ₖ − x ⱼₖ)²,
dendrogram + agglomeration coefficient jump → number of clusters, profile clusters via centroids/means.
 Discriminant analysis: theoretical D = β₀ + ΣβᵢXᵢ, standardised coefficients vs structure correlations for
relative importance, group centroids → cut-off rule, hit ratio (use cross-validated, not original, to avoid
overfitting bias).

Tier 2 — completes the 4-method coverage (recommended for safety):

 Factor analysis (PCA + varimax): KMO/Bartlett, eigenvalue >1 retention rule, % cumulative variance, rotated
component matrix interpretation + factor naming, residual matrix for fit, factor scores used in subsequent
regression/discriminant analysis.

Tier 3 — most-likely (b) topics to round out coverage:

 Scales: Likert vs semantic differential vs Stapel (definitions, examples, when used) — appears in 3 of 6
sittings.
 Comparative scales: paired comparison, rank order, constant sum.
 Three research designs: exploratory, descriptive, causal — appears in 2 of 6 sittings.
 Cross-sectional vs longitudinal designs.
 One-way ANOVA decomposition (SSᵧ = SSₓ + SS_error) — appears in 2 of 6 sittings.
 Non-response: response/incidence/completion rates, methods to improve, methods to adjust (subsampling,
replacement, substitution, subjective estimates, trend analysis, weighting, imputation).
 Projective techniques (word association, story completion, etc.).
 Conjoint analysis basic model U(X) = ΣᵢΣⱼ αᵢⱼxᵢⱼ.
 Post-test-only control group design.
 Operational data and the marketing-system model (controllable mix vs uncontrollable situational factors).

2.3 The minimum syllabus for 100/100

Combining the analysis above:

 Section A (compulsory): 4 sampling methods + the sample-size derivation toolkit. ~6 hours of revision.
 Section B (a): all 4 multivariate methods — regression, factor analysis, cluster analysis (Ward), discriminant
analysis. ~12 hours.
 Section B (b): the 12 short-answer topics in Tier 1+2+3 above. ~10 hours, ~50 minutes per topic to nail a 10-
mark answer.
3. Model Answers — Q2(b), Q3(b), Q4(b) (10 marks each)
Each answer below is calibrated to ~10 marks: enough definitions, formulae and worked examples to bank full
marks, but no padding. They are ordered by sitting and sub-question to mirror the matrix in Section 1.

3.1 Sitting: 2023 May

Q2(b) — Post-test-only control group design (steps + symbolic)

The post-test-only control group design is a true experimental design (i.e. it uses randomisation), in which only
post-treatment measurements are taken — there is no pre-test. Symbolic representation:

EG: R X O₁

CG: R O₂

Definitions: EG = experimental group; CG = control group; R = random assignment of participants to groups; X =

exposure to the treatment (the marketing stimulus, e.g. a new advertisement); O = a single post-treatment
measurement of the dependent variable (e.g. brand recall, purchase intent).

Steps to implement:

 1. Recruit a pool of test units (consumers, stores, etc.).

 2. Randomly assign units to either EG or CG — this controls selection bias and ensures the two groups are
equivalent in expectation on all pre-treatment characteristics, including the dependent variable.
 3. Expose only the EG to the treatment X.
 4. After a fixed interval, take a single measurement on each group: O₁ on EG, O₂ on CG.
 5. Estimate the treatment effect as TE = O₁ − O₂.

Because there is no pre-measurement, the design eliminates main and interactive testing effects. Its principal
weaknesses are that it cannot detect within-individual change and is sensitive to mortality (differential drop-
out across groups). Despite this, its simplicity, low cost and need for only two groups and one measurement
each make it the most popular design in commercial marketing research.

Q3(b) — Comparative scales: purpose + 3 examples

A comparative scale is a measurement scale in which participants directly compare stimulus objects against
one another, rather than rating each in isolation. The data produced are ordinal (rank-order) in nature. The
purpose is to detect small differences between stimuli, force participants to make a choice from a common
reference point, and reduce halo or carry-over effects between judgements. The trade-off is that conclusions
cannot be generalised beyond the specific stimuli scaled.

Three comparative scales:

 Paired comparison scaling. The participant is shown two stimuli at a time and chooses one according to a
stated criterion. With n brands there are n(n−1)/2 pairs. Best when the number of stimuli is small. Coca-Cola
reportedly conducted ≈190,000 paired comparisons before launching New Coke.
 Rank-order scaling. The participant sees all n stimuli simultaneously and ranks them 1 (most preferred) to n
(least preferred). Requires only n−1 decisions, more closely resembles the shopping environment,
eliminates intransitive responses, and is widely used in conjoint analysis.
 Constant-sum scaling. The participant allocates a fixed total (e.g. 100 points) across the stimuli to reflect
their relative importance or preference. Allows fine discrimination and is sometimes treated as metric, but
participants may allocate more or fewer than the requested total and rounding error can distort small
allocations.

A fourth example — the Q-sort — is also acceptable: it is a forced ranking technique used to discriminate
among a large number of objects rapidly.

Q4(b) — Uses of factor analysis in market research

Factor analysis is a class of interdependence techniques (the variables are not split into
dependent/independent) whose purpose is to reduce a large set of correlated observed variables into a smaller
set of underlying latent factors that account for most of the variance. In a market-research setting it is used in
several ways:

 Market segmentation. Consumers can be clustered into segments based on their factor scores — for
example, factors representing "price-conscious", "quality-conscious", "convenience-seeking" attitudes —
rather than on the raw correlated attitude items.
 Product research. Brand attribute ratings (durability, price, design, etc.) can be reduced to a smaller set of
brand image factors, used to compare competing brands or to position new products.
 Advertising studies. Identify underlying media-consumption habits of target customers (e.g. "digital-
engaged", "traditional-media") to design more effective media plans.
 Pricing studies. Identify factors associated with price sensitivity (e.g. coupon-using, deal-prone, brand-loyal)
and target promotional activity at price-sensitive segments.
 Data reduction prior to other multivariate techniques. Factor scores are uncorrelated (after orthogonal
rotation), which makes them ideal predictors in subsequent multiple regression or discriminant analyses
where multicollinearity in the original variables would otherwise be a problem.
 Construct measurement / multi-item scale development. Cronbach's alpha and confirmatory factor analysis
are used to verify that a battery of items measures a single underlying construct (e.g. brand loyalty,
satisfaction).

3.2 Sitting: 2023 October

Q2(b) — Semantic differential vs Likert + verbal description + example

(i) The semantic differential scale is a 7-point itemised rating scale anchored at each end by a pair of bipolar
adjectives (e.g. "Boring — Exciting", "Special — Routine"). The participant marks the position that best
describes their attitude towards the stimulus object. It is used principally to develop and compare brand,
product or company images, and the data are typically treated as interval and analysed via profile (mean) plots.

Example (a visit to an Odeon cinema):

Boring □ □ □ □ □ □ □ Exciting

Cosy □ □ □ □ □ □ □ Uncomfortable
The Likert scale is an itemised rating scale on which the participant indicates their degree of agreement
(typically 5 points: Strongly disagree → Strongly agree) with a series of statements about the object. Used
principally to measure attitudes; data are summed across items and treated as interval.

Example: "I enjoy going to the cinema for special occasions." — Strongly disagree / Disagree / Neither / Agree /
Strongly agree.

(ii) Nature and degree of verbal description. Researchers can label every category, only some, or only the
extremes. Strong anchors ("Completely disagree" / "Completely agree") concentrate responses away from the
extremes, producing peaked, less variable distributions. Weak anchors ("Generally disagree" / "Generally
agree") spread responses more uniformly. Labelling every category reduces ambiguity but providing labels at
all categories does not necessarily improve reliability.

(iii) Example of a question with an itemised rating scale:

"How satisfied are you with the seating quality at Odeon cinemas?"

Very dissatisfied □ Dissatisfied □ Neither □ Satisfied □ Very satisfied □

Q3(b) — Marketing-system model

(i) The four "marketing mix" variables — Product, Price, Promotion and Place (distribution) — are called
controllable because the marketing manager can directly decide their level. The firm decides what product
features to offer, what price to charge, what promotional message and channel to use, and through which
distribution outlets to sell. They are levers the firm operates to influence customer behaviour.

(ii) Situational factors are environmental forces outside the firm's control. Two examples are: (a)
macroeconomic conditions — recessions and inflation suppress consumer spending regardless of marketing
effort; (b) the legal and political environment — new regulations (tax changes, advertising restrictions, GDPR)
constrain what the firm can do. They are uncontrollable because the firm must adapt to them rather than
dictate them.

(iii) Behavioural responses (e.g. brand awareness, attitude, purchase intent, repeat purchase, brand switching)
are intermediate dependent variables that capture how customers think and act in response to the marketing
mix. Performance measures (e.g. sales volume, market share, profit, ROI) are the ultimate outcomes for the
firm. The distinction matters because behavioural responses are more sensitive — and more quickly observable
— than performance measures, and they help diagnose why a campaign succeeded or failed even when sales
numbers are slow to move or noisy.

Q4(b) — Response rate, non-response, strategies across data-collection methods

Response rate = number of completed interviews ÷ number of eligible units contacted. It is a critical quality
indicator because low response rates increase the probability of non-response bias (i.e. systematic differences
between participants and non-participants on the study variables). Non-response is the failure to obtain
information from sampled units, either through refusal, not-at-home, ineligibility, or breaking off.

Strategies to MINIMISE non-response (vary by method):

 Mail/postal: prior notification, personalisation, monetary incentives, reminder mailings, stamped reply
envelopes, attractive cover letters.
 Telephone: call-back protocols at varied times, trained interviewers to overcome refusals at the doorstep,
short questionnaire to minimise drop-off.
 Online: short questionnaires, mobile-friendly design, reminders, prize draws, progress bars to reduce mid-
survey abandonment.
 Face-to-face: appointments, return visits, well-trained interviewers who can build rapport.

Strategies to ADJUST for non-response after the fact:

 Subsampling of non-participants — re-contact a small sample of non-respondents (often by a different

mode) and project their responses to all non-respondents.
 Replacement — re-contact non-respondents from an earlier comparable survey.
 Substitution — substitute another sampling-frame unit expected to respond.
 Subjective estimates / trend analysis — extrapolate from the change observed between successive waves of
respondents (e.g. each wave's answer is X% of the previous wave's, project to non-respondents).
 Weighting — assign weights inversely proportional to sub-group response rates so that under-responding
groups are upweighted in the analysis.
 Imputation — assign a missing value based on similar respondents' values (e.g. mean of similar demographic
group, regression-based prediction).

3.3 Sitting: 2024 May

Q2(b) — Stepwise regression: purpose + three approaches

(i) Purpose. Stepwise regression is an automated variable-selection procedure used to identify, from a large
pool of candidate predictors, a parsimonious subset that explains most of the variation in the dependent
variable. It is useful when there are many candidate explanatory variables, when multicollinearity makes
simultaneous estimation unstable, and when the analyst wants a model that is both interpretable and
predictively adequate.

(ii) Three approaches:

 Forward selection. Start with an empty model (intercept only). At each step, compute an F-statistic (or
partial t-statistic) for each candidate variable not yet in the model. Add the most significant variable,
provided it meets a pre-specified entry significance level. Once a variable enters, it is never removed. Stop
when no remaining variable meets the entry criterion.
 Backward elimination. Start with the full model containing all candidate predictors. At each step, examine
the F-tests for each variable already in the model and remove the least significant variable, provided its p-
value exceeds a pre-specified removal level. Once removed, variables are not reconsidered. Stop when all
remaining variables meet the retention criterion.
 Stepwise selection. A hybrid: it begins like forward selection, but at each step after adding a new variable it
then performs a backward-elimination check — variables already in the model that no longer meet the
retention criterion are dropped. The procedure terminates either when no further variable meets entry, or
when the only candidate just entered is the only one removed in the subsequent backward step. Stepwise
 selection is generally preferred to pure forward or backward because the importance of a variable can
change as other variables enter or leave the model.

Q3(b) — Conjoint analysis basic model

(i) The basic conjoint analysis model is

U(X) = Σᵢ₌₁ᵐ Σⱼ₌₁^kᵢ αᵢⱼ xᵢⱼ.

Definitions of the five terms:

 U(X) — the overall (total) utility a respondent attaches to a product profile X (a particular combination of
attribute levels).
 m — the number of attributes describing the product (e.g. price, brand, screen size, battery life).
 kᵢ — the number of levels of the i-th attribute (e.g. price might have 3 levels: $499, $799, $1,099, so kᵢ = 3).
 αᵢⱼ — the part-worth utility associated with the j-th level (j = 1, …, kᵢ) of the i-th attribute (i = 1, …, m). It
quantifies how much that specific level contributes to overall utility.
 xᵢⱼ — a dummy indicator: xᵢⱼ = 1 if the j-th level of the i-th attribute is present in profile X, 0 otherwise.

(ii) Relative importance of attributes. The importance of attribute i, denoted Iᵢ, is defined as the range of its
part-worths across its levels:

Iᵢ = max_j(αᵢⱼ) − min_j(αᵢⱼ).

To express importance relative to all attributes, normalise:

Wᵢ = Iᵢ / Σ_{i=1}^m Iᵢ, so that Σᵢ Wᵢ = 1.

A larger range in part-worths means moving from the worst to the best level of that attribute moves utility
further, hence the attribute matters more to consumer preference. The Wᵢ values give managers a clear rank-
ordering — e.g. "price accounts for 45% of preference variation, brand 30%, battery life 25%" — and inform
product-design priorities.

Q4(b) — Projective techniques

(i) Projective techniques are unstructured, indirect data-collection techniques used in qualitative research.
They present participants with an ambiguous stimulus and ask them to interpret or complete it; the
assumption is that respondents will project their own underlying motivations, beliefs, attitudes and feelings
onto the stimulus. They should be used when the required information cannot be reliably elicited by direct
questioning — typically because respondents are unaware of, unable to articulate, or unwilling to admit, the
feelings or motivations of interest. Direct questioning in these circumstances tends to produce shallow,
rationalised or socially desirable responses.

(ii) Word association. Respondents are presented with a list of words one at a time and asked to give the first
word that comes to mind. The free associations are assumed to reveal underlying feelings about the topic.
Word association is widely used to test brand names and to uncover attitudes about brands, packages or
advertisements. Example: KLM presents service-related words ("check-in", "seating", "cabin crew") and the
spontaneous associations reveal customer perceptions that direct questioning would miss because
respondents would rationalise their answers.
(iii) Story completion. Respondents receive the beginning of a story directing attention to a topic but stopping
before the resolution; they are asked to complete it in their own words. The completion reveals the
respondent's expectations, anxieties or desired outcomes regarding the topic. Type of respondent: anyone
willing to use their imagination; particularly effective with people currently part-way through the experience
under study. Context: a lone business traveller waiting for a long-haul KLM flight could be given the start of a
story about an airline service experience; their completion would reveal their concerns and expectations about
the rest of the flight in a way a direct survey would not.

3.4 Sitting: 2024 October

Q2(b) / Q3(b) (25 Oct) — One-way ANOVA: decomposition of total variation

In a one-way ANOVA the dependent variable Y is metric and the independent variable X is categorical with c
levels (categories), with sample size n_j in category j (j = 1, 2, …, c) and total sample size N = Σ n_j. The total
variation in Y, measured by the corrected sum of squares, decomposes into a between-group component and a
within-group component:

SS_y = SS_between + SS_within (equivalently, SS_y = SS_x + SS_error)

Defining each term:

 SS_y = Σ_{i=1}^{N} (Yᵢ − Ȳ)² is the total sum of squares — total variation of all observations around the grand
mean Ȳ.
 SS_between = SS_x = Σ_{j=1}^{c} n_j (Ȳ_j − Ȳ)² is the between-group sum of squares — variation in Y
attributable to differences in category means Ȳ_j (the mean of Y within category j) from the grand mean.
This is the part of SS_y "explained" by X.
 SS_within = SS_error = Σ_{j=1}^{c} Σ_{i=1}^{n_j} (Y_{ij} − Ȳ_j)² is the within-group (residual) sum of squares
— variation around each category mean, NOT explained by X.

where Y_{ij} = the i-th observation in category j; Ȳ_j = mean of category j; Ȳ = grand mean across the whole
sample.

The strength of the effect of X on Y is η² = SS_x / SS_y, which lies in [0, 1]. The null hypothesis H₀: μ₁ = μ₂ = … =
μ_c is tested with the F-statistic

F = (SS_x / (c−1)) / (SS_error / (N−c)) = MS_x / MS_error,

which has an F distribution with (c−1, N−c) degrees of freedom under H₀. A large F (small p-value) leads to
rejection of equality of means.

Q3(b) — Three types of research design + examples

Research designs are classified into three broad types depending on the objective:

 Exploratory research. Purpose: provide insight into and understanding of a problem when little is known.
Information needs are loosely defined; the process is flexible and unstructured; samples are small and non-
representative; analysis is often qualitative. Methods: in-depth interviews, focus groups, expert surveys,
secondary data, pilot surveys. Example: A new entrant to the European craft-beer market conducts a series
 of focus groups with millennial consumers to understand which product attributes (taste, story, packaging)
drive their choice — the research question is too unstructured to test with a survey at this stage.
 Descriptive research. Purpose: describe market characteristics and the relationships between variables —
who, what, when, where, how often. Information needs are clearly defined; process is formal and
structured; large representative samples; analysis is quantitative. Methods: surveys, panels, observation,
secondary data analysis. Example: A retailer surveys 5,000 shoppers nationally to estimate the share of
customers who use each loyalty-card benefit and to profile their demographics.
 Causal research. Purpose: establish cause-and-effect relationships by manipulating an independent variable
in a controlled setting. Methods: experimentation (laboratory or field). Example: A consumer-goods firm
runs a controlled in-store experiment in which 50 stores are randomly assigned to receive a new shelf
display while another 50 act as controls; sales over the next 4 weeks are compared to estimate the causal
lift attributable to the display.

Q4(b) — Semantic differential and Stapel scale + examples

Both the semantic differential and the Stapel scale are non-comparative, itemised rating scales used in market
research to measure attitudes and brand image; they differ chiefly in whether the anchors are bipolar
(semantic differential) or unipolar (Stapel).

Semantic differential. A 7-point scale anchored at each end by a pair of bipolar adjectives. Used to develop and
compare brand, product and company images, design promotional strategies, and in new-product
development. Data are commonly analysed via profile (mean) plots.

Example — visit to an Odeon cinema:

Boring □ □ □ □ □ □ □ Exciting

Special □ □ □ □ □ □ □ Routine

Stapel scale. A unipolar 10-point scale numbered −5 to +5 (no zero), with a single descriptor placed in the
middle. The participant indicates how accurately the descriptor describes the object — higher positive numbers
= more accurate; larger negative numbers = more inaccurate. Often presented vertically; can be administered
over the phone (no need for a visual anchor pair). Pre-tests for true bipolarity are not required.

Example — visit to an Odeon cinema:

+5 +4 +3 +2 +1 "Full of energy" −1 −2 −3 −4 −5

+5 +4 +3 +2 +1 "A special event" −1 −2 −3 −4 −5

Both scales produce data treated as interval and analysed similarly; the Stapel is used less frequently but has
clear advantages for telephone administration and when finding genuine bipolar adjectives is difficult.

3.5 Sitting: 2025 May

Q2(b) for 2025 May = same as 2023 Oct Q2(b) (semantic differential vs Likert + verbal description + example) —
see §3.2 above.
Q3(b) — Non-response bias + improving response rates + adjusting
(i) Non-response bias arises when units sampled in a survey fail to respond and the non-respondents differ
systematically from respondents on the variables of interest. Even high overall sample sizes do not eliminate it:
if 30% of the sample doesn't answer and that 30% has a different mean income, brand preference, attitude,
etc., the estimates from the responding 70% are biased. It is problematic because (a) it cannot be corrected by
simply increasing the sample size, (b) it inflates standard errors and undermines representativeness, and (c) it
can lead to incorrect managerial decisions because the picture of the market is skewed.

(ii) Three ways to improve response rates:

 Prior notification — alerting the participant by letter/email/SMS that they will be contacted. This pre-
commits attention and signals legitimacy, increasing willingness to respond.
 Personalisation — using the participant's name, an explicit cover letter from a senior official, hand-signed
signatures. Increases perceived importance and the social cost of non-compliance.
 Incentives — monetary or non-monetary (cash, gift card, charitable donation, prize draw). Reciprocity-based
incentives included with the survey ("$2 enclosed") work better than promised post-completion rewards.

(Other valid answers: questionnaire-design improvements, follow-ups/reminders, callback protocols, mode

mix.)

(iii) Three techniques to ADJUST for non-response after data collection:

 Subsampling of non-participants — re-contact a small sub-sample of non-respondents (often by a different,

more intensive mode such as a phone call) to collect their answers, then project these to all non-
respondents and re-weight the survey total.
 Weighting — assign each respondent a weight inversely proportional to the response rate of their sub-
group, so that under-responding groups are upweighted in the final tabulations.
 Imputation — assign missing values for non-respondents based on the values of similar respondents (e.g.
mean of demographically matched respondents, or a regression-based prediction). Useful when only some
variables are missing for a given respondent.

Q4(b) — Cluster vs discriminant analysis: similarity, difference, uses, interpretation

(i) In what sense are they similar? Both are multivariate techniques concerned with classification: each
produces a grouping of cases into a small number of categories based on a vector of predictor variables. Both
are widely used for market segmentation purposes — assigning consumers to homogeneous groups.

In what sense are they different? Cluster analysis is an INTERDEPENDENCE technique: there is no a-priori
dependent variable; the algorithm DISCOVERS groupings from similarity in the data themselves. Discriminant
analysis is a DEPENDENCE technique: the group memberships are KNOWN in advance (the categorical
dependent variable), and the goal is to explain which predictor variables differentiate the known groups and to
classify new cases. In short: cluster = unsupervised; discriminant = supervised.

(ii) Uses of cluster analysis in marketing:

 Market segmentation — group consumers into segments based on benefits sought, attitudes, or behaviour,
so the firm can tailor a marketing mix to each.
 Understanding buyer behaviour — identify homogeneous groups of buyers and study how each group
responds to specific marketing variables.
 Identifying new product opportunities — clustering brands or products on perceived attribute similarity
reveals competitive sets and gaps in the market.
 Selecting test markets — group cities/regions into similar clusters and choose representative cities for test-
marketing.
 Reducing data — when the analyst has a large number of cases, clustering condenses them into a small
number of homogeneous groups for further analysis.

(iii) Basis for interpreting clusters. Clusters are interpreted by examining the cluster centroids — the mean
values of each clustering variable within each cluster — to characterise what makes each cluster distinct. The
analyst then assigns descriptive labels (e.g. "fun-loving shoppers", "economical shoppers", "apathetic
shoppers") based on which variables have the highest/lowest values in each cluster. It is also common to
PROFILE the clusters using variables NOT used for clustering (demographics, media usage, product usage), so
that downstream marketing actions (targeting, channel selection, messaging) can be designed for each cluster.
Statistical tools that support profiling include one-way ANOVA and discriminant analysis on the cluster-
membership variable.

3.6 Sitting: 2025 October

Q2(b) for 2025 Oct = post-test-only control group design — see §3.1 above.

Q3(b) for 2025 Oct = one-way ANOVA decomposition — see §3.4 above.

Q4(b) for 2025 Oct = three research designs — see §3.4 above.
4. Appendix — Additional 10-mark Model Answers from the 3 PDF Papers
The 3 examiner-commentary PDFs in your project pack contain three additional (b) topics not in the matrix
table. Master these too: they are recent, recurring, and high-leverage.

4.1 Multicollinearity (PDF 2023 paper, Q3b)

(i) Definition. Multicollinearity arises in a multiple linear regression when the predictor variables are highly
correlated with one another (intercorrelations close to ±1).

(ii) Problems caused:

 Partial regression coefficients are not estimated precisely — the standard errors of the β̂ ʼs are inflated, so
individual t-tests lose power and "true" predictors may appear insignificant.
 The magnitudes — and even the signs — of the partial regression coefficients can change dramatically
across samples or when a single observation is added/removed; estimates become unstable.
 It becomes difficult to assess the relative importance of the independent variables in explaining variation in
the dependent variable, because their effects are entangled.
 In stepwise regression, predictor variables may be incorrectly included or excluded, since the test of an
individual variable depends on which other (correlated) variables are already in the model.

(iii) How to detect:

 Inspect the Pearson correlation matrix of the predictors — pairwise correlations above |0.7|–|0.8| are
warning signs.
 Examine the Variance Inflation Factor (VIF) for each predictor; rule of thumb VIF > 10 (or tolerance = 1/VIF <
0.1) indicates a multicollinearity problem.
 Watch for the symptom of a high overall R² combined with individually insignificant t-statistics — a classic
multicollinearity signature.

4.2 Operational data (PDF 2024 paper, Q2b)

Operational data are the data generated by an organisation's daily activities and transactions — sales receipts,
invoices, accounts, HR records, customer-service logs, website clickstream, loyalty-card scans. They are a form
of internal secondary data: already collected, no incremental data-collection cost, no field-access problems,
and the quality is easier to verify than externally-generated data because the firm controls the source.

How they help build an understanding of customer behaviour, with examples:

 What products do customers buy and which are most profitable? — invoice and POS data answer this
directly. A grocery chain analyses basket compositions to identify cross-selling opportunities (e.g. wine +
cheese).
 Which customer segments repeat purchases vs which appear only on promotion? — loyalty-card data linked
to transaction logs tracks individual purchasing histories over time.
 Where are profitable customers located? — postal-code data on invoices supports geographic targeting and
store-location decisions.
 How do customers pay (cash, credit, instalment) and what does that imply for cash-flow planning and
credit-risk exposure? — accounts-receivable data answers this.
 What seasonal and weekly purchasing patterns exist? — time-stamped transaction data reveals patterns
that drive inventory, staffing and promotional calendars.

In short, operational data turn ordinary transactional records into a continuous, low-cost stream of behavioural
intelligence — the foundation of CRM, churn modelling, customer-lifetime-value analysis and modern data-
driven marketing.

4.3 Cross-sectional vs longitudinal designs (PDF 2024 paper, Q4b)

Definitions. A cross-sectional design measures a sample of population units at a single point in time. A
longitudinal design measures the SAME sample units repeatedly at multiple points in time (also called a panel).

Examples. Cross-sectional: a one-off February consumer-confidence survey of 2,000 adults. Longitudinal: a

panel of 1,000 households whose grocery purchases are recorded continuously every week for two years (e.g.
Kantar/IRI panels).

Relative advantages and disadvantages:

 Detecting change. Longitudinal is far better — repeated measures on the same units reveal individual-level
change, brand switching, and dynamic effects that cross-sectional designs hide. (Two cross-sections can
show a stable 30% market share for Brand B even though each year it is a completely different 30% of
consumers.)
 Volume of data. Longitudinal collects more data per unit because panel members agree to ongoing
participation and are usually compensated.
 Accuracy. Longitudinal data tend to be more accurate because purchases are recorded contemporaneously
(e.g. in a diary) rather than relying on memory.
 Representative sampling. Cross-sectional usually wins — easier to recruit a one-off representative sample
than to recruit and retain a panel that resembles the population.
 Response bias. Cross-sectional usually wins — panel members may become "professional respondents"
whose behaviour is conditioned by repeated participation.

Choice of design therefore depends on whether the research question concerns CHANGE or LEVEL: longitudinal
for the former, cross-sectional for the latter.

4.4 Two-way ANOVA: interaction between factors (PDF 2025 paper, Q2b)
In two-way ANOVA the dependent variable is metric and there are two categorical factors X₁ and X₂. The total
variation decomposes as

SS_y = SS_{x1} + SS_{x2} + SS_{x1x2} + SS_error,

where SS_{x1x2} captures the INTERACTION between the two factors.

Definition. An interaction occurs when the effect of one factor on the dependent variable depends on the level
of the other factor — i.e. the joint influence is not simply additive. If X₁ × X₂ is significant, you cannot
meaningfully interpret the main effect of X₁ in isolation from X₂ (and vice versa).
Patterns of interaction:

 Ordinal interaction. The rank order of the levels of X₁ across levels of X₂ is preserved — the lines on an
interaction plot do not cross — but the magnitude of the X₁ effect differs at each level of X₂. Example: a price
discount lifts sales of both premium and budget brands, but lifts the budget brand more.
 Disordinal, non-crossover interaction. The lines diverge but still do not cross within the observed range. The
X₁ effect changes substantially in size, possibly approaching zero at one level of X₂.
 Disordinal, crossover interaction. The lines cross — the rank order of X₁ levels reverses across levels of X₂.
Example: brand A outsells brand B in supermarkets but B outsells A in convenience stores. This is the
strongest form of interaction and the one with the most obvious managerial implication: the optimal level
of X₁ depends entirely on which level of X₂ applies.

Sketches of these three interaction plots (X₂ on the x-axis, dependent on the y-axis, separate lines for the levels
of X₁) are an excellent way to bank visual marks.

4.5 Adjusting for non-response (PDF 2025 paper, Q3b)

The standard list of seven techniques for adjusting for non-response after the survey, each briefly:

 Subsampling of non-participants. Take a small sample of those who did not respond, contact them by a
more intensive method (often phone), and project the resulting answers to all non-respondents.
 Replacement. Replace non-respondents with non-respondents from a comparable earlier survey,
attempting now to elicit a response (perhaps with a stronger incentive). The two surveys must have similar
respondents and short time gap.
 Substitution. Replace each non-respondent with another sampling-frame element judged likely to respond.
Useful but introduces selection bias.
 Subjective estimates. The researcher uses judgement, supported by external information (census, panel
data), to estimate likely responses of non-participants.
 Trend analysis. Examine how each successive wave of respondents (1st mailing → 2nd mailing → 3rd
mailing) differs from the previous, and extrapolate the trend to non-respondents. For example, if each
wave's mean spend is ≈85% of the previous wave's, project that pattern to estimate non-respondents'
spend.
 Weighting. Assign each respondent a weight inversely proportional to the response rate of their stratum
(e.g. high/medium/low income). Weights destroy the self-weighting nature of the design but correct sub-
group imbalance.
 Imputation. For each missing value, impute the value of a similar respondent — by group mean, hot-deck,
or regression-based prediction — exploiting the correlation between observed predictors and the missing
variable.

Best practice combines several of these (e.g. weighting + imputation) and reports the response rate alongside
any results.

4.6 Relative importance of independent variables in MLR (PDF 2025 paper, Q4b)
Because the predictors in a multiple regression are typically correlated, no SINGLE measure of "relative
importance" is unambiguously correct — different measures can give different rankings. The standard set of
approaches:
 Statistical significance. A predictor whose partial regression coefficient is not significant (incremental F-test,
or equivalently t-test) is normally judged unimportant — unless theory strongly predicts otherwise.
 Square of the simple correlation r². The r² between Y and Xᵢ measures the variance of Y explained by Xᵢ in a
bivariate setting; useful but ignores other predictors.
 Square of the partial correlation R²_{y x_i · x_j x_k}. The coefficient of determination between Y and Xᵢ
controlling for the other predictors — closer to "the share of variance explained by Xᵢ alone".
 Square of the part (semi-partial) correlation. The increase in R² when Xᵢ is added to a model that already
contains the other predictors.
 Standardised regression coefficients (beta weights). |Bᵢ| or B²ᵢ measure the change in Y (in standard-
deviation units) per one-standard-deviation change in Xᵢ, holding others constant. The most common quick
measure used in exam answers, but unstable under multicollinearity.
 Stepwise regression order of entry. Variables that enter first are typically more important; variables that
enter last (or never) are least important.

Limitations. All measures fail when predictors are highly correlated, because the variance explained "jointly" by
Xᵢ and Xⱼ cannot be uniquely allocated. Best practice: report several measures (e.g. standardised coefficients
PLUS partial correlations PLUS structure-correlation-style metrics from discriminant analysis) and, if rankings
disagree, note the multicollinearity caveat in the discussion.