Research Design and Project Proposal

Introduction

A business analytics capstone project is not simply an exercise in applying software to a dataset. It is a structured investigation in which a student defines a problem, establishes what evidence is needed, selects a defensible research design, analyses data appropriately, and communicates findings to an academic and professional audience. The research design connects the research question to the evidence, methods, analysis and conclusions. The project proposal makes that design explicit before substantial resources are committed.

In business analytics, project topics commonly include customer churn, demand forecasting, fraud detection, workforce planning, pricing, marketing effectiveness, operational efficiency and sustainability reporting. Each topic can be investigated in several ways. A question about the causes of employee turnover may require qualitative interviews, while a question about whether a revised email campaign increases conversion may require an experiment. A question about predicting next month’s sales may require an observational predictive study using historical data. Selecting a method because it is familiar, convenient or available is weaker than selecting it because it fits the question and the decision context.

A sound proposal therefore demonstrates alignment among five elements:

1. the business problem and research purpose;
2. the research questions and, where appropriate, hypotheses;
3. the research design and data requirements;
4. the analysis plan and intended deliverables; and
5. feasibility, ethics, governance and risk controls.

Learning Objectives

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

• distinguish research paradigms, purposes and designs used in business analytics;
• formulate a focused problem statement, research aim, questions and hypotheses;
• select quantitative, qualitative, mixed-method, experimental, observational or case-study designs;
• define scope, assumptions, population, sample, variables and unit of analysis;
• assess feasibility, resources, milestones, risks and deliverables;
• identify data protection, consent, fairness, confidentiality and ethical approval requirements; and
• prepare a coherent, defensible project proposal.

From Business Problem to Researchable Question

A business problem is a practical condition requiring attention; a research problem is a gap in knowledge or evidence that can be investigated systematically. For example, “customer churn is too high” is a business problem. A researchable formulation might be: “Which observable customer behaviours are associated with voluntary churn among monthly subscribers during the 2024 financial year, and how accurately can churn be predicted 30 days in advance?”

A strong problem statement normally identifies:

• the organisational or industry context;
• the undesirable condition or opportunity;
• the affected stakeholders;
• the evidence currently missing;
• the consequences of inaction; and
• the boundary of the proposed investigation.

The aim expresses the overall purpose in one sentence. Objectives translate the aim into achievable actions, usually beginning with verbs such as assess, compare, identify, estimate, evaluate, develop or interpret. Research questions should be answerable with the proposed evidence. They should not combine several unrelated investigations.

A quantitative question may ask, “To what extent does delivery time predict customer satisfaction?” A qualitative question may ask, “How do customers describe the experience of delayed delivery?” A mixed-method question may ask both how frequently delays occur and how customers interpret them. A causal question must be especially precise because it concerns whether an intervention produces an outcome rather than merely whether two variables are associated.

Hypotheses and Conceptual Models

A hypothesis is a testable proposition about a relationship or difference. For example, “Customers experiencing delivery delays of more than two days have lower satisfaction scores than customers receiving on-time deliveries.” A null hypothesis generally states that there is no specified difference or relationship, while an alternative hypothesis states that one exists. Hypotheses are most appropriate when theory or prior evidence supports a quantitative test; exploratory qualitative studies do not need to force hypotheses into the design.

A conceptual model shows proposed relationships among constructs. In a churn study, perceived service quality may influence satisfaction, satisfaction may influence loyalty, and loyalty may influence churn intention. The model helps identify variables and prevents data collection from becoming an unfocused search for correlations. It does not prove causation.

Research Paradigms and Methodological Logic

A research paradigm contains assumptions about reality, knowledge and how knowledge can be justified. Business analytics projects often draw on three broad positions.

Post-positivist or quantitative approaches treat measurement, systematic observation and statistical inference as important tools. They recognise that measurements are imperfect and that conclusions are probabilistic rather than absolutely certain. These approaches are suitable for estimating prevalence, testing relationships, comparing groups and building predictive models.

Interpretivist or constructivist approaches emphasise how people create meaning within social and organisational contexts. Interviews, focus groups, observations and document analysis are often appropriate. The researcher seeks depth, context and multiple perspectives rather than statistical generalisation.

Pragmatic approaches select methods according to the research question and decision need. A pragmatic project may combine transaction data with interviews because numerical patterns alone do not explain operational causes. Pragmatism does not mean that any method is acceptable; the design must still be coherent and transparent.

Methodology is the rationale connecting philosophical assumptions, design, data collection and analysis. Methods are the specific techniques, such as a survey, regression model, thematic coding or A/B test. Confusing methodology with methods can produce a technically polished but logically weak project.

Selecting a Research Design

Quantitative Design

Quantitative research uses numerical measurement and structured analysis. It is appropriate when constructs can be operationalised and the project seeks estimates, comparisons, associations, predictions or tests of hypotheses. Typical data include sales, prices, click-through rates, customer ratings and processing times.

A quantitative design requires clear definitions of variables. An outcome variable is the phenomenon to be explained or predicted; explanatory variables are proposed predictors or inputs. A categorical variable records groups, a discrete variable records countable values, and a continuous variable can take values on a scale. The unit of analysis might be a customer, transaction, store, employee, product or day. Mixing units—for example, predicting customer churn using store-level averages without justification—can create invalid conclusions.

Quantitative designs may be descriptive, correlational, predictive or explanatory. Descriptive analysis summarises what happened. Correlational analysis examines association. Predictive analysis estimates unknown or future outcomes. Explanatory analysis evaluates theoretically motivated relationships. Prediction and explanation are not identical: a model can predict well without identifying a causal mechanism.

Qualitative Design

Qualitative research investigates meanings, experiences, practices and processes. Common methods include semi-structured interviews, focus groups, observation and document analysis. Sampling is usually purposive: participants are selected because they possess relevant experience or knowledge. The researcher develops codes and themes through a systematic process, maintains an audit trail and supports interpretations with evidence from the data.

Quality is assessed through credibility, dependability, confirmability and transferability. Useful practices include triangulation across data sources, participant validation where appropriate, reflexive notes, transparent coding decisions and searching for disconfirming evidence. A small sample is not automatically weak; its adequacy depends on the research purpose, information richness and diversity of relevant perspectives.

Mixed-Method Design

Mixed-method research integrates quantitative and qualitative evidence within one study. It is valuable when the project needs both the scale of numerical data and the contextual explanation provided by participants.

• In an explanatory sequential design, quantitative results are collected first and qualitative work explains surprising or important patterns.
• In an exploratory sequential design, qualitative findings inform the development of measures, hypotheses or a subsequent survey.
• In a convergent design, quantitative and qualitative data are collected during a similar period and then compared or integrated.

Integration must be planned, not added as an afterthought. A joint display, for example, can place a customer segment’s churn rate beside interview themes explaining its behaviour. Contradictory findings should be reported and examined rather than concealed.

Experimental Design

An experiment deliberately changes an independent variable and observes its effect on an outcome. Random assignment, when feasible, helps balance other influences between treatment and control groups. A digital A/B test might randomly assign website visitors to two page designs and compare completed purchases.

A simple treatment effect can be expressed as:

Here, Ȳ_T is the mean outcome for the treatment group, Ȳ_C is the mean outcome for the control group, and Δ is the observed difference. This difference is most plausibly interpreted as a causal effect when assignment is random, treatment conditions are implemented as intended, outcomes are measured consistently, and interference and major attrition are controlled.

A before-and-after comparison without a control group is weaker because changes may reflect seasonality, market conditions or regression to the mean. Where randomisation is impossible, a quasi-experimental design may use matched groups, interrupted time series or difference-in-differences. These designs require explicit assumptions and cautious language.

Observational Design

In observational research, the researcher measures naturally occurring conditions without assigning treatment. Cross-sectional studies collect data at one point or period; longitudinal studies follow units over time. Observational designs are often realistic for business data, but confounding can distort relationships. For example, high advertising expenditure and high sales may occur together because managers spend more during peak demand, not because advertising alone caused the sales increase.

Association should therefore not be described as causation unless the design and assumptions justify it. Predictive modelling can still be valuable, but the proposal must distinguish a forecasting objective from an intervention claim.

Case Study Design

A case study examines a bounded contemporary phenomenon in its real-world context. The case may be a firm, department, implementation project, product launch or supply chain disruption. Evidence can include interviews, performance data, reports, system logs and observations. A case study is particularly useful when context and process are central and when the boundary between the phenomenon and its environment is difficult to separate.

A single case can provide analytical insight but usually cannot support statistical generalisation to all organisations. Multiple-case designs permit comparison, although they require more time and access. The proposal should define the case boundary, justify its selection and explain how evidence will be triangulated.

Scope, Population, Sampling and Measurement

Scope is the deliberate boundary of the project. It should specify the organisation or market, period, geography, population, outcome, technologies and exclusions. “Analyse customer behaviour” is too broad; “predict voluntary monthly subscription churn among UK customers active between January and December 2024” is more manageable.

The population is the full set of units to which the study refers. A sample is the subset actually studied. Probability sampling—such as simple random, stratified or systematic sampling—supports statistical generalisation when a valid sampling frame exists. Non-probability approaches—such as convenience, purposive and snowball sampling—may be appropriate for exploratory or qualitative research but limit generalisation.

Measurement requires operational definitions. “Customer engagement” might be measured using monthly sessions, active days, support contacts or a validated survey scale. A proposal should explain why the measure represents the construct and acknowledge measurement error, missingness and possible bias. For survey instruments, reliability and validity deserve attention; a measure can be consistent without measuring the intended construct.

Analysis Planning and Assumptions

The analysis plan should be written before examining results in depth. It should identify data preparation, descriptive analysis, modelling or coding procedures, evaluation metrics and interpretation rules.

For a binary classification problem, common metrics include accuracy, precision, recall and F1 score. If TP, TN, FP and FN denote true positives, true negatives, false positives and false negatives:

These measures assume a clearly defined positive class. Accuracy may be misleading when classes are imbalanced, while precision and recall reflect different error costs. A proposal should state which error is more consequential for the decision.

For supervised predictive work, data should be separated into training and test information, with validation procedures used where appropriate. Information from the future must not leak into training predictors. Time-dependent business data often require chronological rather than random splitting. The proposal should identify missing-data treatment, outlier policy, feature construction, model comparison and reproducibility procedures.

Qualitative analysis should similarly specify transcription, familiarisation, coding, theme development, review and reporting. Quantitative and qualitative procedures should not be presented as automatic recipes; their appropriateness depends on the research question and data quality.

Feasibility, Deliverables and Project Governance

A feasible project balances value, rigour, time, access and capability. A useful feasibility assessment asks:

• Can the required data be accessed legally and ethically?
• Is the sample or case available within the academic calendar?
• Are the variables measured at adequate quality and granularity?
• Can the student perform or learn the required methods?
• Are software, computing, storage and supervision available?
• Can the project be completed without depending on an uncertain external event?

Typical deliverables include an approved proposal, data dictionary, cleaned and documented dataset or interview corpus, analysis notebook, methodology chapter, results, limitations, recommendations, final report and presentation. Code and documentation should be version-controlled where possible. A dashboard may be a useful communication product, but it does not replace methodological explanation.

A milestone plan should contain dependencies. For example, ethical approval precedes recruitment; data access precedes cleaning; cleaning precedes modelling; and validated findings precede recommendations. A simple schedule can use weekly milestones, with contingency time for delayed access, recruitment problems or re-analysis.

Risk Register

A risk register records the risk, likelihood, impact, response and owner. Common risks include unavailable data, small samples, changing business requirements, data leakage, participant withdrawal, software failure and unrealistic scope. Risk responses include avoidance, reduction, transfer or acceptance. A project should define decision points at which the scope or method will be revised rather than waiting until the final week.

Ethics, Data Governance and Responsible Analytics

Ethical approval may be required when research involves human participants, personal data, sensitive organisational information, vulnerable groups, deception, intervention or access to restricted records. Requirements vary by institution and jurisdiction, so students should consult the relevant ethics committee before collecting data. Retrospective secondary data are not automatically exempt.

Important principles include:

• Informed consent: participants should understand the purpose, procedures, risks, benefits, voluntary nature of participation and withdrawal arrangements.
• Privacy and confidentiality: collect only necessary data, restrict access, remove direct identifiers where possible and avoid reporting combinations that enable re-identification.
• Data security: use approved storage, encryption where required, access controls, secure transfer and a defined retention and deletion period.
• Minimisation and purpose limitation: do not collect data merely because they might become useful later.
• Fairness: assess whether data or models produce systematically worse outcomes for relevant groups.
• Transparency: document data provenance, transformations, model limitations and intended use.
• Responsible communication: avoid overstating causality, certainty or business benefit.

Generative AI can support brainstorming, code explanation, documentation or language editing, subject to institutional policy. It must not be treated as an authoritative source, allowed to expose confidential data, or used to fabricate references, participants, results or analysis. Students remain responsible for verifying generated code and reporting meaningful use of AI.

Structure of a High-Quality Project Proposal

A proposal should be concise enough to be read but detailed enough to permit critical assessment. A robust structure is:

1. Working title: specific and informative.
2. Background and rationale: context, importance and evidence gap.
3. Problem statement: the precise problem being investigated.
4. Aim and objectives: one overall aim and measurable objectives.
5. Research questions or hypotheses: aligned with the design.
6. Literature and conceptual framework: relevant theory, evidence and constructs.
7. Methodology: paradigm, design, population or case, sampling, data sources and procedures.
8. Analysis plan: preparation, analytical techniques, evaluation and assumptions.
9. Ethics and governance: consent, privacy, security, approvals and risks.
10. Feasibility and resources: skills, software, access, budget and support.
11. Timeline and deliverables: milestones, dependencies and outputs.
12. Limitations: anticipated threats to validity and mitigation.
13. References and appendices: instruments, draft data schema, risk register or schedule.

Alignment can be tested with a traceability table.

Illustrative Case: Retention Analytics at Northstar Telecom

This is a fictional illustrative case.

Background and problem. Northstar Telecom reports increasing monthly churn among prepaid customers. Managers want a model to identify customers at risk and a recommendation for a retention campaign. The initial request—“build a churn model”—is too broad because it does not specify the prediction horizon, action, data permissions or success criteria.

Research design. The student reframes the project into three linked questions: (1) How accurately can churn within the next 30 days be predicted using information available at the prediction date? (2) Which behavioural patterns are associated with elevated risk? (3) Does a low-cost reminder offer improve renewal among eligible high-risk customers? The first two questions use an observational, quantitative design. The third requires an experiment and may need separate approval.

Scope and assumptions. The population is active prepaid customers in one country during 2024. Churn is operationally defined as no renewal or usage for 30 consecutive days, subject to confirmation with the business owner. Predictors must be available before the prediction date. The student excludes customers already in a complaint resolution process because their treatment pathway differs. The model is intended to prioritise human review, not automatically deny service or impose penalties.

Feasibility and ethics. The data steward confirms access to anonymised customer identifiers, recharge history, usage summaries and service contacts. Exact location and message content are excluded as unnecessary. The student obtains approval for the intervention, prepares a retention schedule, documents data lineage and establishes role-based access. Fairness checks compare performance across relevant groups only where lawful, ethically justified and sufficiently represented.

Analysis and outcome. The student uses chronological training and test periods, compares a simple baseline with candidate models, and reports recall and precision because missing a genuinely at-risk customer is costly. Results show that recent recharge decline and repeated service contacts are useful predictors, but performance falls for a small customer segment with sparse history. Interviews with service agents explain that some apparent “churn” reflects temporary financial hardship rather than dissatisfaction. A controlled pilot then estimates whether the message changes renewal behaviour.

Learning points. The project succeeds because it separates prediction from explanation and explanation from causal evaluation. It defines the outcome, prevents future information from entering predictors, includes a realistic fallback for sparse data, and treats ethical approval and operational use as design requirements rather than administrative details.

Challenges and Limitations

Research proposals necessarily contain uncertainty. Access may change, measures may be imperfect and early assumptions may prove incorrect. The appropriate response is not to promise certainty but to specify how uncertainty will be managed.

Common threats include selection bias, non-response, survivorship bias, confounding, data leakage, inconsistent definitions, small samples and researcher interpretation. A model may perform well in historical data yet fail after a policy change. A case study may provide deep understanding but limited generalisation. Interviews may reveal articulate participants’ perspectives while excluding less visible groups. Mixed-method projects may become unmanageable if integration is not planned.

Limitations should be specific. “The study has limitations” is unhelpful; “the single-company sample limits transferability to other industries” is informative. Students should distinguish limitations that can be mitigated—such as inconsistent coding—from those that must simply be acknowledged—such as the inability to randomise a historical intervention.

Best-Practice Checklist

Before submitting a proposal, confirm that:

• the problem is important and expressed in researchable terms;
• every objective is feasible within the available time;
• the research questions match the chosen design;
• the unit of analysis and outcome definitions are explicit;
• data access, sampling and missing-data issues are addressed;
• the analysis plan identifies assumptions and evaluation criteria;
• causal language is used only when justified;
• ethics, privacy, security and AI-use requirements are documented;
• deliverables, milestones, dependencies and risks are realistic; and
• limitations and contingency plans are stated honestly.

Chapter Summary

Research design is the logic linking a business problem to credible evidence and defensible conclusions. Quantitative designs support measurement, comparison and prediction; qualitative designs explain experience, meaning and process; mixed methods integrate breadth and depth; experiments support causal inference under appropriate conditions; observational studies analyse naturally occurring data but require caution about confounding; and case studies provide contextual understanding of bounded real-world phenomena.

A strong proposal defines the problem, aim, objectives, questions, population, unit of analysis, measures, data sources and analysis plan. It also establishes scope, feasibility, resources, deliverables, milestones and risks. Ethical approval, privacy, security, fairness and responsible communication are integral to analytics research. The best proposal is not the most ambitious one, but the one whose claims, evidence, methods and resources are consistently aligned.

Key Terms

Research design: The overall plan connecting research questions, evidence, methods and conclusions.

Business problem: A practical organisational condition requiring investigation or action.

Research question: A focused, answerable question guiding an investigation.

Hypothesis: A testable proposition about a relationship or difference.

Conceptual model: A representation of proposed relationships among constructs or variables.

Quantitative research: Research using numerical measurement and structured analysis.

Qualitative research: Research examining meanings, experiences, practices or processes.

Mixed-method research: An integrated study using quantitative and qualitative evidence.

Experiment: A design in which an intervention or treatment is deliberately assigned or introduced.

Observational study: A study measuring naturally occurring conditions without researcher assignment.

Case study: An in-depth investigation of a bounded phenomenon in context.

Population: The full set of units to which a study refers.

Sample: The units actually included in a study.

Unit of analysis: The entity about which data and conclusions are organised.

Operationalisation: Defining an abstract construct through observable measures.

Confounding: Distortion of an apparent relationship by another variable related to both factors.

Data leakage: Use of information unavailable at the intended prediction time.

Triangulation: Comparing evidence from multiple sources, methods or perspectives.

Feasibility: The practical possibility of completing a project with available resources and constraints.

Informed consent: Voluntary agreement based on adequate information about participation.

Review Questions

Short Answer Questions

1. What is the difference between a business problem and a research problem?
2. Define the unit of analysis and explain why it matters.
3. When is a qualitative design more suitable than a quantitative design?
4. What is data leakage in predictive analytics?
5. Identify three issues that may require ethical approval.

Long Answer Questions

1. Compare experimental and observational designs, including the strength of causal claims each can support.
2. Explain how an explanatory sequential mixed-method design could investigate customer churn.
3. Describe the main sections that should appear in a business analytics project proposal.
4. Discuss how sampling, measurement and scope affect the credibility of a capstone project.
5. Explain the role of feasibility, deliverables, milestones and risk management in project planning.

Analytical / Application Questions

1. A retailer wants to know whether a discount notification increases repeat purchases. Recommend a design, identify its assumptions and explain what evidence would support a causal conclusion.
2. A predictive model has high accuracy but performs poorly for a small customer segment. Analyse possible reasons and propose governance and methodological responses.
3. Design a mixed-method study for a hospital whose appointment no-show rate has increased. Specify the research questions, data sources, sampling, analysis and ethical safeguards.

Multiple Choice Questions with Answers

Question 1

Which design is most appropriate for testing whether a new website layout causes an increase in purchases?

A. Randomised experiment

B. Cross-sectional case study

C. Unstructured observation

D. Historical descriptive report

Correct Answer: A. Randomised experiment

Question 2

What is the primary purpose of a conceptual model in a research proposal?

A. To guarantee statistical significance

B. To show proposed relationships among constructs

C. To replace data collection

D. To eliminate sampling decisions

Correct Answer: B. To show proposed relationships among constructs

Question 3

Which statement best distinguishes prediction from explanation?

A. Prediction always proves causation

B. Explanation never uses numerical data

C. A model may predict accurately without identifying a causal mechanism

D. Explanation is possible only in qualitative research

Correct Answer: C. A model may predict accurately without identifying a causal mechanism

Question 4

In an observational study, confounding occurs when:

A. The sample is larger than the population

B. Another variable distorts the apparent relationship between two variables

C. All participants provide informed consent

D. The outcome is measured more than once

Correct Answer: B. Another variable distorts the apparent relationship between two variables

Question 5

Which sampling approach is generally most appropriate when participants are selected for their specialised experience?

A. Purposive sampling

B. Simple random sampling

C. Systematic sampling

D. Cluster sampling

Correct Answer: A. Purposive sampling

Question 6

If TP = 80 and FP = 20, what is precision?

A. 20%

B. 50%

C. 80%

D. 100%

Correct Answer: C. 80%

Question 7

What is the main ethical reason for data minimisation?

A. It guarantees a representative sample

B. It reduces collection of unnecessary personal information

C. It increases model complexity

D. It removes the need for consent

Correct Answer: B. It reduces collection of unnecessary personal information

Question 8

Which feature is essential to a well-defined project scope?

A. An unlimited list of possible analyses

B. Explicit boundaries concerning population, period or setting

C. A promise that findings will generalise universally

D. Exclusion of all limitations

Correct Answer: B. Explicit boundaries concerning population, period or setting

Question 9

What is the main purpose of chronological splitting in a time-dependent prediction problem?

A. To ensure future information does not enter model training

B. To make every class equal in size

C. To remove all missing values

D. To guarantee causal inference

Correct Answer: A. To ensure future information does not enter model training

Question 10

Which statement about a single case study is most accurate?

A. It always supports statistical generalisation

B. It cannot use quantitative evidence

C. It can provide contextual insight but may have limited transferability

D. It requires random assignment

Correct Answer: C. It can provide contextual insight but may have limited transferability