Choosing a Statistics Research Question states what the data can realistically describe, compare or explore. It emerges after the researcher has identified a meaningful problem, the unit or population of interest, measurable variables, an obtainable data route and the type of claim the design can support. The question should come before the test or software. For introductory quantitative work, the most defensible questions are usually descriptive, comparative or association-focused; causal wording is appropriate only where the design has the control, timing and evidence to justify it. A clear question guides the variables, sample, data collection, analysis and interpretation, while an unclear one forces those decisions to be guessed later (Barroga and Matanguihan, 2022; Texas A&M University, 2026).
A topic is not yet a question
A topic identifies an area of interest. It does not specify what will be measured or how evidence will be used. “Student engagement”, “housing affordability” and “online shopping” are topics. They become researchable only when the investigator decides what is meant by the key concept, whose experience or records matter, what data are available, and what comparison or pattern would be useful.
This distinction protects research quality. A question like “Does social media cause anxiety?” contains a large causal claim, an ambiguous outcome and potentially sensitive data. A smaller question such as “Is reported daily social-media use associated with a validated wellbeing score among consenting adult survey respondents in the available dataset?” is still demanding, but it identifies the type of relation, population and data boundary. Its wording does not promise more than a non-experimental design can usually establish.
| Starting point | Why it is insufficient | Researchable revision | What the revision clarifies |
| Student wellbeing | No population, measure, timeframe or analytical purpose. | How do self-reported wellbeing scores vary by commuting-time category among consenting college survey respondents? | Outcome, grouping variable and restricted sample. |
| Crime in cities | “Crime” and “cities” are undefined; cause is not stated. | How did recorded bicycle-theft rates vary across selected London boroughs from 2022 to 2025? | Indicator, geography and period. |
| Advertising and sales | “Effect” overclaims unless the design supports causality. | What association is observed between monthly promotional spending and sales revenue in an anonymised retailer dataset? | Association rather than cause; data unit is a month. |
| Climate change | Too broad for a bounded quantitative study. | How did daily maximum temperature vary during July at one documented monitoring station, 2015–2025? | Single measurable outcome, location and timeframe. |
The Question–Evidence–Claim framework
Use the Question–Evidence–Claim framework before deciding on methodology or a statistical procedure. It exposes whether the desired conclusion is stronger than the available data can bear.
1. Question
Write the question in plain language. It should identify a phenomenon or outcome, a population or unit of analysis, and the intended purpose: description, comparison, association or carefully justified prediction. A question is not improved by technical words that do not change what will be measured.
2. Evidence
Specify the data source, variables, coverage, access conditions and likely limitations. A question about a national population needs evidence with an appropriate sampling or coverage route; a local class dataset cannot support that same population claim. The UK Data Service recommends identifying required data type, geography and time period before searching, then reading collection and documentation details before use (UK Data Service, 2026).
3. Claim
Choose language that matches the design. A cross-sectional survey can describe respondents, compare observed groups or investigate an association. It normally cannot demonstrate that one factor produced an outcome. An experimental or quasi-experimental design may support a stronger claim, but only if its allocation, measurement, timing and potential alternative explanations are addressed.
| Question purpose | Useful question pattern | Evidence requirement | Language suitable for a conclusion |
| Descriptive | What pattern is visible in [outcome] for [unit] over [place/time]? | Clearly defined records or measurements. | “The data show…” or “The observed pattern was…” |
| Comparative | How does [outcome] differ between [defined groups/periods]? | Comparable measures and groups. | “The groups differed in the available data…” |
| Association | Is [X] associated with [Y] in [population/dataset]? | Two measured variables and suitable data structure. | “An association was observed…” |
| Predictive exploration | How well do [inputs] predict [outcome] in [dataset]? | Adequate documented data and validation plan. | “The model performed within this dataset…” |
| Causal evaluation | What changed after [intervention], compared with an appropriate counterfactual? | Design with control/comparison, temporal order and confounding strategy. | “The design supports an estimated effect, subject to assumptions…” |
A six-screen feasibility check
Before adopting a question, test it through six screens. A failure at any screen is a prompt to revise the idea; it is not a reason to force unsuitable data or a more elaborate test.
| Screen | Question to ask | A positive sign | A reason to redesign |
| Purpose | What answer is the study seeking? | The question can be classified as descriptive, comparative, association or prediction. | It tries to describe, explain and evaluate everything at once. |
| Unit | What does one row of data represent? | The answer is unambiguous: person, household, area, firm, month or event. | The plan shifts between individual and area-level conclusions. |
| Variables | Can each central concept be measured or identified in documentation? | Definitions, coding and units are known. | A key concept is only a vague label. |
| Data | Is an appropriate and permitted source available? | Source, coverage, access route and metadata are recorded. | The work depends on inaccessible, unapproved or undocumented data. |
| Scope | Can the study be completed at the available level and timeframe? | Variables, geography and period are deliberately bounded. | The study needs many datasets, groups and outcomes without a rationale. |
| Ethics | Can the evidence be obtained and reported responsibly? | Public aggregates or approved non-sensitive data are used appropriately. | It requires informal collection of sensitive or identifiable personal information. |
The six screens follow the broader design principle that the problem, question, methods and sample should align. Texas A&M’s four-step sequence—define the problem, develop the question, design methods and determine the sample—makes the ordering explicit: statistical analysis is a consequence of design, not its starting point (Texas A&M University, 2026).
From broad interest to quantitative question
A reliable route has four stages. First, identify a practical or scholarly issue. Second, decide what particular feature of it could be measured. Third, check data availability. Fourth, write a question whose wording matches the resulting design.
Suppose the broad interest is “working while studying”. A researcher might initially consider academic attainment, weekly paid-work hours and study time. A data search may reveal that an approved survey records work hours and study time but not verified grades. The question should then be adapted to the variables actually available: “What association is observed between reported weekly paid-work hours and reported weekly study hours among respondents in the available survey?” This is not weaker than an unsupported attainment claim; it is more honest and replicable.
Researchers should keep the research question separate from the title. The title introduces the study; the question names what will be answered. The Statistics Research Titles resource explains how to create a concise title after the question, evidence and claim are coherent.
Operationalising concepts without hiding assumptions
Operationalisation means translating a concept into a measurable variable. It is an interpretive decision, not merely a technical one. “Financial pressure” could mean arrears, self-rated difficulty meeting costs, debt level or a composite scale. Each measure captures something different. The question, title and conclusion should refer to the measure actually used rather than a broader concept it cannot fully represent.
| Concept | Possible measure | What it can support | What it should not be treated as |
| Academic engagement | Attendance percentage in one term | A record of attendance patterns. | A complete measure of learning or motivation. |
| Local deprivation | Published area-level index score | Comparison of geographic areas. | An individual resident’s income or circumstances. |
| Customer satisfaction | A documented survey scale | Reported satisfaction among respondents. | Objective product quality. |
| Physical activity | Reported weekly activity minutes | A self-reported behaviour measure. | A clinical health assessment. |
| Financial wellbeing | A validated survey item or scale | Respondents’ reported assessment. | Verified household financial position. |
Fictional planning application
The following scenario is fictional. Yusuf, an undergraduate logistics student, wants to investigate whether remote working makes people more productive. He begins with a single broad question and expects to compare people who work remotely with those who do not.
His tutor asks him to apply the six-screen feasibility check. Yusuf finds an open employee survey dataset that includes self-reported remote-working days, a perceived productivity measure, job role category and survey date. It does not include objective output, an experimental allocation or repeated follow-up. He revises his question to: “How is the number of reported remote-working days associated with perceived productivity among survey respondents in the available employee dataset?”
The revision changes the project’s logic. The unit is a survey respondent, not an organisation. The outcome is perceived productivity, not verified output. The claim is association, not a causal effect. Yusuf plans descriptive summaries, a graph and a cautious association analysis, then discusses role type and self-selection as potential explanations. His key learning point is that precision was achieved by limiting the question, not by adding an advanced model.
Practical planning sequence
Start with two or three possible questions, then complete a short question specification for each: intended purpose, unit of analysis, outcome, explanatory variable or comparison group, source, scope and claim boundary. Put the design choice in writing before selecting software or a test.
Next, inspect data documentation. Confirm how variables were collected, whether values are comparable, what population the source covers, whether missingness is recorded and what access conditions apply. If primary data collection is proposed, seek appropriate institutional advice before recruitment, consent or collection; this is particularly important for health, disability, finances, protected characteristics and other sensitive matters.
Finally, make a one-page design map. It should align the question, variable definitions, sample or source, planned descriptive displays and limits. That map becomes the basis for the Statistics Project Methodology stage, where sampling, measurement and data-management choices are set out in more detail.
Critical limitations and safeguards
A precise question cannot compensate for invalid measurement, biased sampling or inappropriate data access. A question about “people in the UK” should not be answered with the unexamined views of a small convenience sample. An area-level association should not be interpreted as a statement about every individual in the area. A large dataset can still be poorly matched to a question.
Avoid collecting identifiable or sensitive data without an approved ethical and safeguarding pathway. Do not redesign a question after looking at results solely to make an output appear more striking. Keep an audit trail of question changes, source versions and analytical decisions. Transparent revision is legitimate when data feasibility requires it; concealed revision after seeing results is not sound research practice.
FAQs
Q1. Must every statistics project have a hypothesis?
No. Descriptive and exploratory questions may not require a formal hypothesis. When a hypothesis is used, it should be a testable prediction grounded in prior reasoning, not a conclusion selected after reviewing results.
Q2. How many variables should a research question include?
Use the smallest set that can answer the question. One outcome and one or two clearly defined comparison or explanatory variables are often more manageable than an extensive list of weakly understood fields.
Q3. Can I ask a causal question using survey data?
A survey can explore associations, but causal claims require stronger design and assumptions. State what the data can show and seek relevant methods supervision where causal evaluation is genuinely required.
Q4. What if the available data do not match my first idea?
Revise the question to match documented variables and coverage, or choose another source. It is better to narrow the question than to force a dataset to answer something it cannot support.
References
American Statistical Association (2016) Guidelines for Assessment and Instruction in Statistics Education: College Report 2016 (Accessed: 25 August 2026).
Barroga, E. and Matanguihan, G.J. (2022) ‘A Practical Guide to Writing Quantitative and Qualitative Research Questions and Hypotheses in Scholarly Articles’, Journal of Korean Medical Science, 37(16), e121 (Accessed: 25 August 2026).
Texas A&M University (2026) Choosing the Right Research Design (Accessed: 25 August 2026).
UK Data Service (2026) Finding and Accessing Data for Your Project (Accessed: 25 August 2026).
USC Libraries (2026) Quantitative Methods (Accessed: 25 August 2026).