How to Formulate a Research Question

A topic tells a reader what area you care about. A research question tells them what you will find out, and what evidence would count as an answer. “Deep learning for medical images” is a topic. “On chest X-rays from a second hospital, does a model trained at the first hospital keep its sensitivity for a named finding, compared with a model fine-tuned on 200 local cases?” is a question a reviewer can score.

Start from a decision, not from a technique

Write one sentence about the decision your result would change. A clinician choosing whether to trust a model off the shelf. A programme chair deciding whether a new training trick is worth the compute. A student deciding whether a method is stable enough to build a thesis on. If you cannot name the decision, you have a technique in search of a problem, and the question will wander every time a new paper comes out.

Four parts, in one sentence

Clinical reviews often use PICO: population, intervention, comparison, outcome. The same four slots work in computing, education, and lab work if you rename them. Fill all four before you open a search box.

Qualitative work uses the same discipline with different words. The “comparison” might be two groups of interviewees. The “result” might be a set of themes you will look for, written down before the interviews so you cannot quietly move the target. A question with no recorded outcome is a prompt for a conversation, not a study.

One question, then the aims

Keep a single main question. Two or three aims can sit under it if each aim is a piece of the same answer. Three unrelated questions in one project is three projects, and a thesis or a paper that tries to finish all of them will finish none in a way a reader trusts.

• Too wide. “How can machine learning improve healthcare?” No dataset, no decision, no end.
• Still a topic. “A survey of transformers.” A survey is a method. The question is what the survey is for: which design choices have actually been compared, on which tasks.
• Answerable. “On the named public benchmark, which of these three training choices moves accuracy by more than the run-to-run variation, against the published baseline?”

Check that you can finish it

Read the question back and mark anything you do not already have: data access, ethics approval, a supervisor who knows the method, compute, or the months the design needs. Cut or narrow the question until the missing piece is something you can get inside the time you actually have. A precise question you cannot run is not more rigorous than a smaller one you can.

Write it so a stranger can tell if you succeeded

Hand the sentence to someone outside your group. Ask them what result would make the answer “yes” and what result would make it “no”. If they cannot say, the question is still a topic. When they can, that sentence is the one to put at the end of the introduction and, later, at the top of a proposal. The companion guide How to write a research proposal is the outline those aims drop into.