Marketing decisions are made under uncertainty. Managers may need to estimate demand, understand why customers switch brands, evaluate a new advertisement, or determine whether a product change improves satisfaction. Marketing research provides evidence for these decisions, but evidence is useful only when the research design and data sources are appropriate to the decision.

Research design is the overall plan for obtaining and analysing evidence. It connects the management problem to research questions, data sources, measurement procedures, sampling, analysis, and conclusions. A poorly designed study may produce precise-looking results that answer the wrong question. A well-designed study makes the limits of the evidence explicit and supports defensible action.

This chapter examines three broad purposes of research—exploration, description, and causal explanation—and distinguishes cross-sectional from longitudinal designs. It then evaluates primary, secondary, syndicated, internal, and open data. Finally, it presents a framework for judging evidence according to relevance, credibility, validity, timeliness, and cost.

Learning Objectives

1. From Management Problem to Research Design

A management problem concerns an action or decision: for example, declining sales, low app usage, or uncertainty about a proposed price. A research problem translates that concern into a question that evidence can address. The distinction is important. “How can we increase sales?” is a broad management problem; “What factors are associated with repeat purchase among customers who bought in the last six months?” is a researchable question.

Research designs are not rigid stages that must always occur separately. A project may begin with interviews, proceed to a survey, and end with an experiment. The design should reflect the decision, not a preference for a particular method.

2. Three Purposes of Marketing Research

2.1 Exploratory design

Exploratory research investigates a problem that is poorly understood. Its purpose is to clarify concepts, discover possible explanations, identify relevant variables, and generate hypotheses. It is especially useful when managers do not yet know which questions to ask.

Common methods include literature and database reviews, analysis of internal records, expert interviews, focus groups, social listening, ethnographic observation, and qualitative analysis of customer complaints. Samples are usually purposive rather than statistically representative. The researcher seeks information-rich cases, not estimates of population percentages.

Exploration is flexible and iterative. Initial findings may change the research question or reveal that a supposedly simple problem has several dimensions. For example, interviews with lapsed subscribers may show that “cancellation” reflects price concerns, confusing billing, poor service recovery, or a change in household circumstances.

Exploratory evidence is valuable for insight but should not automatically be treated as proof of prevalence or causation. A few participants may identify an important issue, but they cannot establish how common it is. Similarly, an observed association in online comments may suggest a hypothesis without demonstrating that the issue causes behaviour.

2.2 Descriptive design

Descriptive research measures characteristics, behaviours, attitudes, or market conditions. Typical questions include:

Surveys, structured observation, panels, transaction databases, and market audits are common descriptive methods. Descriptive studies require clear operational definitions, an appropriate sampling frame, reliable measurement, and suitable analysis. If a study reports that 42% of customers are satisfied, the researcher must explain who was surveyed, how satisfaction was measured, when the data were collected, and how nonresponse was handled.

Description may be univariate, such as a frequency distribution of purchase occasions, or comparative, such as differences in awareness between two customer segments. It may also examine associations, such as whether satisfaction differs by delivery speed. An association, however, does not by itself show that delivery speed causes satisfaction.

2.3 Causal design

Causal research evaluates whether changing one factor produces a change in another. It is used for questions such as:

The central feature is deliberate variation in an independent variable and comparison of outcomes. Strong causal designs commonly include random assignment, a control or comparison condition, manipulation checks, and procedures that reduce alternative explanations.

A causal claim requires more than correlation. Three conditions are generally relevant: the presumed cause precedes the outcome; the cause and outcome covary; and plausible alternative explanations are controlled or made less likely. Randomised experiments are particularly powerful because random assignment tends to balance both observed and unobserved factors across groups, subject to chance and adequate implementation.

Experiments may be conducted in laboratories, stores, websites, mobile applications, or geographic markets. A/B testing is a familiar field experiment in which users are randomly assigned to version A or version B. Quasi-experiments may be necessary when random assignment is impractical or unethical. Examples include comparing regions before and after a campaign while using another region as a comparison, or exploiting a policy threshold. Such designs can be informative but require stronger assumptions than a properly randomised experiment.

Comparing the three designs

Feature	Exploratory	Descriptive	Causal
Main purpose	Discover and clarify	Measure and profile	Test effects
Typical question	What might be happening?	What, who, when, and how much?	What happens if X changes?
Common methods	Interviews, focus groups, observation	Surveys, panels, audits, records	Experiments, A/B tests, quasi-experiments
Sampling emphasis	Information richness	Representativeness	Valid comparison and assignment
Typical output	Concepts and hypotheses	Estimates and patterns	Effect estimates and decision evidence
Main limitation	Limited generalisability	Weak causal inference	Cost, complexity, and implementation risk

The designs are complementary. Exploration can improve the wording and content of a descriptive survey. Description can identify segments for an experiment. An experiment can test whether a relationship suggested by earlier studies is causal.

3. Cross-Sectional and Longitudinal Designs

Cross-sectional studies

A cross-sectional study collects information at one point in time or during a relatively short field period. It provides a snapshot of market conditions. A brand-tracking survey conducted in May, for example, can estimate awareness and preference in May.

Cross-sectional designs are often efficient and comparatively inexpensive. They are suitable for estimating current attitudes, comparing groups, and informing immediate decisions. Their limitations include difficulty establishing temporal order, inability to observe individual change, and vulnerability to seasonality. A cross-sectional association between social-media use and brand preference cannot establish whether social-media use changed preference, preference caused social-media use, or another factor influenced both.

Longitudinal studies

A longitudinal study collects data on multiple occasions. It can reveal trends, persistence, transitions, and individual change. Common forms include:

Longitudinal data are useful for studying retention, customer journeys, advertising effects over time, and changes in attitudes. They provide stronger evidence about temporal ordering than a single cross-section. However, they cost more, require participant retention, and may suffer from panel conditioning: respondents can change their behaviour because they know they are being observed. Attrition can also bias results if those who leave differ systematically from those who remain.

Decision consideration	Cross-sectional	Longitudinal
Time perspective	Snapshot	Change and sequence
Cost and speed	Usually lower and faster	Usually higher and slower
Individual change	Not directly observed	Can be observed in panels
Main risks	Seasonality and weak temporal order	Attrition, conditioning, and tracking burden
Best use	Current market estimate	Trends, retention, and behavioural paths

A repeated cross-sectional design may be preferable when tracking population-level change is more important than following the same individuals. For example, a brand may survey a new representative sample each quarter. A panel is preferable when the research question concerns individual switching or purchase histories.

4. Sources of Evidence

4.1 Primary data

Primary data are collected specifically for the current research purpose. Examples include a newly administered survey, interviews with target customers, observations in stores, or an experiment designed by the research team.

The main advantage is fit: the researcher controls questions, definitions, sampling, timing, and procedures. Primary research can address information that does not exist elsewhere. Its disadvantages include cost, time, respondent burden, and risks arising from poorly designed instruments or fieldwork.

Primary data require a documented protocol. Researchers should specify the target population, recruitment method, consent process, questionnaire or interview guide, quality checks, data treatment, and analysis plan before collection where feasible.

4.2 Secondary data

Secondary data were collected previously for another purpose, whether by the organisation or by an external body. Examples include government statistics, prior customer surveys, published industry reports, transaction records, and academic studies.

Secondary data can reduce cost and accelerate analysis. They are often useful for market sizing, trend analysis, benchmarking, and constructing a sampling frame. Their limitations include mismatched definitions, unknown collection procedures, missing variables, restricted access, and uncertain quality. “Available” does not mean “appropriate.” A government measure of household income may not match the organisation’s definition of disposable income, and an old customer database may underrepresent recent buyers.

Before using secondary data, the researcher should inspect metadata: who collected the data, why, from whom, when, how variables were defined, how missing values were treated, and whether revisions or weighting were applied.

4.3 Internal data

Internal data originate within the organisation. They may include point-of-sale transactions, customer relationship management records, website analytics, call-centre logs, loyalty activity, returns, complaints, campaign records, and financial data.

Internal data are often highly relevant to operational decisions and can support detailed segmentation and behavioural analysis. Their weaknesses include incomplete coverage, inconsistent coding across systems, duplicate records, changes in data definitions, and selection effects. For example, website analytics describe tracked visitors, not necessarily all potential customers. A CRM record may be accurate for active customers but omit prospects who never completed registration.

Data integration creates additional risks. Customer identifiers may not match across systems, timestamps may use different time zones, and a “customer” may mean a household in one database and an individual in another. A data dictionary and documented data lineage are essential.

4.4 Syndicated data

Syndicated data are collected by a research or analytics provider and sold or licensed to multiple clients. Examples include retail panels, media-audience measurement, consumer panels, brand trackers, and industry benchmarks.

The shared cost can make sophisticated measurement affordable. Syndicated providers may offer large samples, repeated measurement, specialist sampling, and benchmarks across competitors or markets. However, the organisation may have limited control over definitions, questionnaire design, sampling, timing, and access conditions. Commercial confidentiality and provider methodology must be considered carefully.

A buyer should examine the provider’s sample design, weighting, response rates, panel recruitment, quality controls, variable definitions, reporting frequency, and licensing restrictions. Apparent precision should not obscure uncertainty or methodological limitations.

4.5 Open data

Open data are publicly accessible data that can be used under stated legal and licensing conditions. Sources may include national statistics offices, regulatory bodies, transport authorities, public health agencies, international organisations, and open geographic datasets.

Open data support market mapping, demographic analysis, location planning, policy analysis, and contextual interpretation. They may be authoritative, but openness is not a guarantee of accuracy. Researchers must evaluate coverage, update schedules, geographic boundaries, revisions, missing values, and licensing. Data scraped from a public website may be technically accessible but unstable, incomplete, or prohibited for certain uses by the site’s terms.

5. Evaluating Evidence

A practical evidence assessment should consider five criteria. No source will maximise every criterion, so evaluation is a reasoned trade-off rather than a simple score.

Relevance

Relevance asks whether the evidence addresses the decision and the target population. Relevant evidence uses suitable constructs, units, geography, market category, and time period. A national awareness estimate may be irrelevant to a decision about urban first-time buyers. Relevance also includes granularity: data aggregated at national level may not answer a store-location question.

Credibility

Credibility concerns the trustworthiness of the source and the transparency of its methods. Consider the institution’s expertise, independence, incentives, documentation, auditability, and willingness to disclose limitations. A credible source can still provide irrelevant evidence, so credibility is necessary but not sufficient.

Validity

Validity concerns whether the evidence supports the intended interpretation. Important forms include:

Reliability—consistency of measurement—is related to but different from validity. A consistently wrong scale is reliable but not valid.

Timeliness

Timeliness concerns whether the evidence reflects the market conditions relevant to the decision. Rapidly changing prices, technology, regulations, and consumer preferences can make old data misleading. Historical data may nevertheless be appropriate for a trend study or seasonality analysis. The key is alignment between the data period and the decision horizon.

Cost

Cost includes more than purchase price. Researchers should consider licensing, fieldwork, cleaning, integration, respondent incentives, analyst time, delays, and the cost of making a wrong decision. A cheap dataset with serious bias may be more expensive in consequence than a carefully designed study.

Criterion	Questions to ask
Relevance	Does it measure the right population, construct, geography, and decision period?
Credibility	Who produced it, with what expertise, incentives, and documentation?
Validity	Are measures and inferences defensible? What alternative explanations remain?
Timeliness	When were the data collected, updated, and revised?
Cost	What are direct, indirect, integration, and error costs?

Researchers should record evidence provenance: source, extraction date, version, transformations, exclusions, weighting, and responsible analyst. This makes later review and replication possible.

6. Integrating Multiple Sources

No single source usually answers a complex marketing question. Triangulation combines different methods or sources to improve understanding. For example, declining sales might be examined using internal transaction data, customer interviews, competitor pricing information, and a survey of category buyers. Convergence increases confidence; disagreement may reveal segmentation, measurement differences, or a genuine market tension.

Integration must be analytical rather than merely additive. Researchers should identify which source answers which question and recognise that sources may represent different populations. Transaction data show behaviour among observed customers, while a survey may include stated intentions among recruited respondents. Combining them without reconciling definitions can create false precision.

Information need	Preferred evidence	Supplementary evidence	Main limitation
Current purchase pattern	Internal transactions	Retail panel	Coverage and coding
Reasons for switching	Interviews	Survey open responses	Recall and social desirability
Size of target market	Official statistics and panel data	Primary survey	Definition mismatch
Effect of campaign	Randomised test	Brand tracker	Implementation and spillover

7. Ethics, Governance, and Responsible Use

Research design must protect participants and the integrity of decisions. Informed consent should explain participation, intended use, risks, incentives, and withdrawal rights where applicable. Researchers should collect only data necessary for the stated purpose, restrict access, secure identifiers, and establish retention and deletion rules.

Privacy law and sector regulation may impose requirements concerning personal data, profiling, cookies, children, sensitive information, cross-border transfers, and automated decision-making. Legal compliance is a minimum standard; ethical practice also asks whether collection is proportionate, understandable, and fair.

Researchers must avoid deceptive recruitment, undisclosed manipulation that creates unreasonable risk, discriminatory sampling, and misleading presentation of uncertainty. Incentives should not unduly pressure participation. Reporting should disclose missing data, exclusions, weighting, nonresponse, and limitations.

Artificial intelligence can assist with coding open-ended responses, identifying patterns, synthetic data generation, and exploratory analysis. It can also reproduce sampling bias, hallucinate interpretations, expose confidential information, or obscure how classifications were made. AI outputs require human review, documented prompts and procedures where relevant, privacy safeguards, and validation against known data. Synthetic data are not a substitute for evidence about real consumers unless their generation and limitations are clearly established.

8. Illustrative Case: Choosing Evidence for a Subscription Service

Background

Imagine a fictional streaming service, Northstar, whose monthly cancellation rate has increased. Managers initially propose a discount, but they do not know whether cancellations are driven by price, content gaps, technical problems, or changing household needs.

Problem

The decision is whether to introduce a discount, improve service reliability, expand content, or redesign retention communications. The management problem is broad; the research problem is to identify important drivers of cancellation and estimate which intervention changes retention profitably.

Analysis and approach

Northstar first conducts exploratory interviews with recent cancellers and reviews complaint logs. Interviews reveal three possible explanations: price sensitivity, repeated playback failures, and insufficient local-language content. The team treats these as hypotheses rather than population estimates.

Next, analysts use internal data to describe cancellation rates by tenure, plan, device, region, payment method, and reported service incidents. They discover that cancellations are concentrated among newer customers who experienced repeated failures. A cross-sectional survey of current and former customers measures perceived value, reliability, content satisfaction, and cancellation intentions. The survey provides prevalence estimates, but because it is cross-sectional it cannot establish that reliability caused cancellation.

Finally, Northstar conducts a controlled field test among eligible at-risk customers. Participants are randomly assigned to a reliability-support message, a price discount, or a standard communication. The organisation pre-specifies the primary outcome as continued paid subscription after a defined follow-up period and monitors adverse effects, such as discounting customers who would have stayed anyway. A longer panel or linked transaction analysis can examine whether effects persist beyond the initial period.

Outcome

Suppose the test shows that the support message produces a modest retention improvement among customers with documented technical incidents, while the discount produces a larger short-term retention effect but reduces contribution margin. The appropriate decision is not simply to select the largest retention percentage. Managers must compare incremental retention, margin, implementation cost, and customer experience, while recognising that results may not generalise to customers without technical incidents.

Learning points

The case demonstrates a staged design: exploration identifies plausible explanations, description measures their distribution, and experimentation evaluates interventions. Internal, primary, and longitudinal evidence answer different questions. It also shows why segmentation and decision economics matter: an average treatment effect may conceal meaningful differences between customer groups.

9. Best Practices and Common Errors

Best practices include defining the decision before choosing a method; using multiple sources when the decision is consequential; documenting operational definitions; separating association from causation; checking representativeness and missingness; pretesting instruments; planning for nonresponse and attrition; and reporting uncertainty and limitations.

Common errors include selecting data because they are convenient, treating online comments as representative, confusing a large sample with a valid sample, using outdated market estimates without adjustment, assuming syndicated metrics are automatically comparable, and presenting correlations as causal effects. Another error is beginning with an available method—such as a survey or dashboard—rather than identifying the information needed.

Chapter Summary

Research design is the plan linking a marketing decision to evidence, measurement, sampling, analysis, and conclusions. Exploratory research clarifies poorly understood problems and generates hypotheses. Descriptive research measures market characteristics and patterns. Causal research tests whether deliberate changes produce outcomes and requires attention to temporal order, comparison, and alternative explanations.

Cross-sectional studies provide efficient snapshots but offer limited evidence about change and causality. Longitudinal studies reveal trends and individual transitions but require greater time, cost, and management of attrition and panel conditioning. Primary data offer purpose-built control; secondary data offer speed and economy. Internal data are operationally rich but may contain coverage and integration problems. Syndicated data provide shared specialist measurement, while open data support accessible contextual and geographic analysis.

Evidence should be judged for relevance, credibility, validity, timeliness, and cost. Triangulation can strengthen conclusions, but only when differences in populations, definitions, and measurement are understood. Ethical research protects participants, documents provenance, governs personal data, and treats AI as an assistive tool requiring validation and human accountability.

Key Terms

Exploratory research: Research used to clarify a problem and generate concepts or hypotheses.

Descriptive research: Research that measures characteristics, behaviours, attitudes, or conditions.

Causal research: Research that evaluates whether a change in one factor produces a change in another.

Cross-sectional study: A study measuring units at one point or short period in time.

Panel study: A longitudinal study involving repeated measurement of the same units.

Trend study: A study using different samples from the same population over time.

Cohort study: A study examining a defined group sharing a characteristic across time.

Construct validity: The extent to which a measure represents the intended concept.

Internal validity: The extent to which an observed effect is attributable to the presumed cause.

Data provenance: A documented record of a dataset’s origin, processing, and transformations.

Panel conditioning: Changes in participant responses or behaviour caused by repeated observation.

Review Questions

Short Answer Questions

Long Answer Questions

Analytical / Application Questions

Multiple Choice Questions with Answers

Question 1

Which research design is primarily intended to generate hypotheses and clarify an poorly understood problem?

Question 2

Which conclusion is most directly supported by a well-designed randomised experiment?

Correct Answer: B. A change in the manipulated factor caused a change in the outcome

Question 3

A survey conducted once in June to estimate current brand awareness is an example of which design?

Question 4

Question 5

Question 6

Which source is most likely to contain point-of-sale transactions and CRM records?

Question 7

Which criterion asks whether evidence measures the correct population, construct, geography, and period?

Question 8

Question 9

Correct Answer: C. To combine evidence while examining convergence and differences

Question 10

Why should AI-generated coding of open-ended responses be reviewed by researchers?

Chapter 3 Research Design and Sources of Evidence

Research Design and Sources of Evidence

Introduction