Most research problems that show up late, such as a study that doesn't answer anything, a literature review that goes in every direction, or a supervisor who keeps asking "so what are you actually asking?", start with a question that was too vague at the beginning.
A good research question does a lot of work. It tells you what to read, whom or what to study, what to measure or explore, and which methods fit. It also tells your reader, in one sentence, what your paper or thesis is for. This guide covers how to test a question, which framework to use, and how to fix the most common problems, with worked examples from eight fields.
From topic to question
A topic is an area of interest: "social media and teenagers". A question is a specific thing you want to find out about it: "Is time on image-based platforms associated with depressive symptoms in 13- to 16-year-olds?" The work is in getting from one to the other, and it usually goes like this:
- Start with the topic and read enough recent work to know what's been done. Our guide to finding a research gap covers this step.
- Pick one angle: a population, a setting, a time period, a variable or an experience.
- Decide what kind of answer you want: a description, a comparison, a relationship, a cause or an account of experience. This decides your design (see the table below).
- Write the question, then test it against FINER.
- Rewrite it. Most good questions go through five or six drafts.
Test it with FINER
The FINER criteria come from Designing Clinical Research by Hulley and colleagues, a standard text in clinical research methods. They work in any field.
| Criterion | Ask yourself | What failing looks like |
|---|---|---|
| Feasible | Can I get enough participants or data, with the skills, time and money I have? | A national survey planned for a one-year master's project |
| Interesting | Will the answer interest me, my supervisor and people in my field? | A question chosen only because data happened to be available |
| Novel | Does it confirm, refute or extend earlier findings, or find something new? | A question a recent systematic review has already answered |
| Ethical | Would an ethics committee approve it? Is the burden on participants justified? | A design that withholds treatment people need |
| Relevant | Will the answer matter for knowledge, practice, policy or future research? | An answer that changes nothing whichever way it comes out |
Feasibility is the one most often underestimated. If you're not sure, ask your supervisor how long similar studies took, or run a small pilot.
The type of question decides the design
The wording of your question signals what sort of study you'll need. Getting this right early saves you from collecting data that can't answer what you asked.
| Type | Typical wording | Designs that fit |
|---|---|---|
| Descriptive | What is…? How common is…? | Surveys, audits, document analysis |
| Comparative | Do A and B differ in…? | Group comparisons, quasi-experiments |
| Relational | Is X associated with Y? | Correlational and cohort studies |
| Causal | Does X cause or reduce Y? | Experiments and randomized trials |
| Exploratory | How do people experience or make sense of…? | Interviews, focus groups, qualitative analysis |
A common mismatch is asking a causal question ("Does social media cause depression?") and then running a survey, which can only show association. Either change the design or change the question to match what your design can show.
Frameworks: PICO, PEO and SPIDER
Frameworks break a question into parts so you can check nothing is missing. They also turn directly into search terms for your literature review.
| Framework | Elements | Best for |
|---|---|---|
| PICO (or PICOT) | Population, Intervention, Comparison, Outcome (and Time) | Effects of treatments, programmes and interventions |
| PEO | Population, Exposure, Outcome | Associations, risk factors and experiences of an exposure |
| SPIDER | Sample, Phenomenon of Interest, Design, Evaluation, Research type | Qualitative and mixed-methods questions |
A 2014 comparison by Methley and colleagues ran the same qualitative search with PICO and SPIDER in three databases. PICO found more of the relevant studies; SPIDER returned fewer irrelevant ones. Their advice: use PICO when the search must be comprehensive, and treat SPIDER's precision as a time-saver rather than a replacement. For framing the question itself, use whichever framework fits your study.
Eight weak questions, rewritten
Each weak question below is a real pattern: a topic dressed up as a question. Each rewrite adds the limits that make it answerable.
Weak: Does social media affect mental health?
Better: Among 13- to 16-year-olds in secondary schools, is daily time spent on image-based social media associated with depressive symptoms over one school year?
What changed: Names the population, narrows "social media" to a measurable exposure, names the outcome and a timeframe, and asks about association, which a survey can answer, rather than effect, which it cannot.
Weak: Is online learning effective?
Better: For first-year undergraduate statistics students, do final exam scores differ between course sections that use weekly online quizzes and sections that do not?
What changed: Replaces "effective" with a specific outcome and turns an open judgment into a comparison between two groups.
Weak: What do nurses think about burnout?
Better: How do intensive care nurses describe coping with moral distress after a patient's death?
What changed: Focuses on one group and one experience, and uses "how do… describe", which signals a qualitative design such as interviews and thematic analysis.
Weak: How can battery life be improved?
Better: How does charging rate (0.5C, 1C and 2C) affect the capacity retention of lithium iron phosphate cells over 500 cycles at 25 °C?
What changed: Defines the variable you will change, the material, the outcome, the duration and the conditions, so the experiment almost designs itself.
Weak: Why do startups fail?
Better: What reasons do founders of software startups that closed within five years give for the closure, and do these reasons differ between bootstrapped and venture-funded firms?
What changed: Limits the sector and timeframe, says whose account you are collecting, and adds a comparison that makes the answer more useful.
Weak: What is the impact of climate change on agriculture?
Better: How have yields of rain-fed maize in one region changed with growing-season rainfall variability between 2000 and 2024?
What changed: Swaps two enormous concepts for one crop, one region, one measurable climate variable and a defined period with available data.
Weak: How was women's suffrage portrayed in the press?
Better: How did British national newspapers frame suffragette militancy between 1908 and 1914, and did framing differ by the newspaper's political leaning?
What changed: Defines the sources, the aspect of the movement, the period and the angle of analysis, so the archive search has edges.
Weak: Is exercise good for back pain?
Better: In adults with chronic non-specific low back pain, does a 12-week supervised exercise programme, compared with usual care, reduce pain intensity at 12 weeks?
What changed: Fills every PICO element: population, intervention, comparison and outcome, plus a timepoint.
Too broad, too narrow, or ready?
| Too broad | Too narrow | Ready |
|---|---|---|
| Thousands of relevant papers, no obvious place to start | Answerable with a quick search or a single number | A defined body of literature you can read in weeks |
| Contains words like "impact", "effective" or "role" without saying how they'll be measured | So specific that nobody else would find the answer useful | Every key term can be measured or clearly explored |
| Would take a research team years | Answerable with yes or no and nothing further | Fits your time, data access and skills |
Question, hypothesis or aim?
- Aim: what the study sets out to do. "To examine the relationship between sleep and exam performance in first-year students."
- Research question: what you want to find out. "Is sleep duration associated with exam scores in first-year students?"
- Hypothesis: a testable prediction of the answer, used mainly in quantitative work. "Students who average under six hours' sleep in exam weeks score lower than those who average seven or more."
Qualitative studies usually have questions but not hypotheses. Theses usually have one main question with two to four sub-questions, each answered by a chapter or study.
Common mistakes
- Two questions in one. "How do students use AI tools and do they improve grades?" Split them.
- A built-in assumption. "Why is remote work harming productivity?" assumes the answer. Ask whether, and how, it affects productivity.
- A value question. "Should schools ban phones?" is a policy debate. Research can answer what happens when they do.
- Unmeasurable terms. "Impact", "effectiveness", "wellbeing" and "success" need to be defined as something you can observe.
- A design mismatch. Causal wording with a correlational design, or "how many" with interviews.
Where AI tools help
AI is useful for stress-testing a question you've written: asking for alternative wordings, flagging vague terms, or checking a draft against FINER. It can't tell you whether the question is genuinely new, or whether it's the right question for you. That comes from the literature and from your supervisor.
In LitraMind, you can draft the question at the top of a document in Write New Paper and build the outline beneath it, so every section is tied to what you're asking. Research & Cite helps you find published work as you check whether the question has already been answered. Once you have a question, our guide to starting a paper from your sources picks up from there.
Checklist
- It names or clearly implies the population, setting or material
- Every key term can be measured or explored
- Its wording matches a design you can carry out
- It passes all five FINER criteria
- It asks one thing, without assuming the answer
- You've checked it hasn't already been answered
- Your supervisor or a colleague understands it on first reading
