How to choose a PhD research question that holds up to scrutiny

Most of the difficulty in a PhD comes from the question. Too broad, and you never finish. Too narrow, and nobody cares about the answer. Unanswerable with the data you can get, and you cannot defend it. Choosing well is the highest leverage decision of your doctorate.

A topic is not a research question

A topic is an area. A research question is a specific, answerable problem inside that area. The difference is easiest to see side by side.

The second version tells you what data you need, what you will measure, and what a positive or negative answer would look like. That is what makes it researchable. The example comes from my own field, and the background is in explainable AI and uncertainty in medical imaging.

Five tests a good research question passes

1. It is answerable with evidence you can actually obtain. Name the dataset, the experiment, or the fieldwork. If you cannot, the question is not ready.
2. It is new, and you can say how in one sentence. New data, a new method, a new setting, or a new combination. Vague novelty does not survive review.
3. It matters to someone beyond you. A clinician, an engineer, a policy maker, or other researchers who would change what they do based on the answer.
4. It fits your time and resources. A doctorate lasts a few years. Your question should be answerable within that window, with room for things to go wrong.
5. It can fail informatively. A good question teaches you something even when the hypothesis turns out to be false.

Clinical researchers use a similar checklist called FINER: Feasible, Interesting, Novel, Ethical, and Relevant. It transfers well to most fields, including AI.

From topic to question in four steps

Step 1. Collect open problems

Read the limitations and future work sections of recent papers in your area. Authors are telling you, in writing, what remains unsolved. Keep a list.

Step 2. Narrow along one axis at a time

Restrict by data (which population or dataset), by method (which family of approaches), or by setting (which real world constraint). Each restriction makes the question more answerable.

Step 3. Write it as one sentence and one hypothesis

If you need two sentences and three "and" clauses, you have several questions. Pick one. Then state what you expect to find, so that your experiments have something to confirm or refute.

Step 4. Stress test it

Present it to your supervisor and to at least one person outside your group. Ask them to attack it: has this been done, can it be measured, would the answer matter? Revise, then repeat.

Red flags

• You cannot name the data you will use.
• The answer already appears in a paper from the last few years, and you have no clear angle beyond it.
• The whole project depends on one resource you do not control, such as a single hospital partnership with no backup.
• Only one outcome would count as success, so a negative result would leave you with nothing to report.
• You are choosing the question mainly because the method is fashionable.

It is allowed to change

Your question will evolve as you learn. That is healthy. What matters is that at any moment you can state your current question precisely, and explain why you changed it when you do.