A publishable question is specific enough to name its variables, answerable with evidence you can reach, and open enough that a result could surprise you. This is the six-step method for building one, plus the three tests that kill a bad question in ten minutes rather than ten weeks.
A good research project is mostly decided in the first week, before any reading, by the question you commit to. This is the method for building one.
If what you want is a list of worked questions, we have them for the EPQ and the Extended Essay, and a map of which subjects are workable. This page is the process you run when you want a question nobody else has.
The three tests
Run these before anything else. Each takes a minute and together they eliminate most of what students start with.
1. The variables test. Does your question name the specific things being compared? "The impact of social media on wellbeing" names a field. "Whether self-reported daily screen time predicts self-reported sleep duration" names two variables. If yours does not, it is still a topic.
2. The falsification test. Can you say in one sentence what evidence would change your answer? If no obtainable evidence would settle it, you will produce a summary of other people's arguments rather than a finding.
3. The surprise test. Is there a result that would genuinely surprise you? If every plausible outcome confirms what you already think, the question is closed and four months will confirm it slowly.
Most student questions fail test one. Nearly all of the rest fail test three.
The six steps
Step 1 — Pick territory, not a topic
Decide which discipline's methods you are using before you decide what you are curious about.
This sounds backwards and it prevents the commonest dead end: a question that is genuinely interesting and belongs to no field, so there is no method for answering it and no venue for publishing it. "Why are people lonely" is interesting. It is also sociology, psychology, economics and philosophy at once, and a project that cannot say which is being graded on nothing.
Step 2 — Find the dataset before the question
Open something you can actually reach and spend an hour inside it.
This is the step that distinguishes projects that finish from projects that do not, and it is the one students skip. A question invented in the abstract usually requires evidence that does not exist, is paywalled, or is too small. A question that comes out of a dataset you have already opened cannot have that problem.
Practically: pick a source from our list of free datasets, download it, open the codebook, and read variable definitions until you find one you did not expect to exist. Surprising variables are where unclaimed questions live, because nobody thought to use them.
Step 3 — Write three pairings
Two variables and a plausible relationship. Three of them, not one.
Writing one means you are already attached to it, and attachment is what stops you abandoning a question that does not work. Three means you are selecting rather than defending.
| Weak pairing | Why it fails | Stronger version |
|---|---|---|
| Social media and mental health | Neither variable is measurable as stated | Reported screen time and reported sleep duration, one survey year |
| Diet and health outcomes | Both are categories, not variables | Fast-food outlet density per 1,000 and recorded childhood obesity rate, by local authority |
| Education and income | Settled, and enormous | Class size and attainment in PISA, after controlling for national income |
Step 4 — Name the confounder
For each pairing, say out loud what else could explain the relationship.
If you cannot think of anything, you have not understood the pairing well enough to research it. If the confounder obviously explains the whole thing, drop that pairing and move to the next.
If the confounder is real but you can control for it, you have found your project — because "does X still predict Y once we account for Z" is a better question than "does X predict Y," and it is the shape most published papers actually take.
Step 5 — Check who has already done it
Search Google Scholar for your two variables plus the dataset name.
Three outcomes, all useful:
- Nothing comes back. Widen your search terms before celebrating. A complete absence usually means wrong terminology, not an untouched field.
- Close work exists. Good. It confirms the question matters and gives you something to position against. Narrow yours — a subgroup, a different country, a later period, a control they did not apply.
- Your exact study exists. Narrow further, or move to pairing two. This is why you wrote three.
Step 6 — Write the question in one sentence, with its limits
The finished form names the variables, the data, the population and the control.
"Does self-reported daily screen time predict self-reported sleep duration among US 12th graders in the 2023 Monitoring the Future survey, controlling for sex and region?"
Unlovely, and that is the point. Every clause is doing work, and anyone reading it knows exactly what you did and what you did not do.
Underneath it, write the sentence that says what result would change your mind. If you cannot write that second sentence, go back to step three.
What makes it publishable rather than just finishable
Three further things, and they are the gap between a good school project and a submission.
Original findings, not a summary. Most student research journals want new data or a genuinely novel analysis. A clean analysis of public data counts — the Journal of Emerging Investigators wants a finding, and an unreported relationship in an existing dataset is one. A literature review is not, at JEI; it is accepted at NHSJS and the Columbia Junior Science Journal, which is worth knowing before you start rather than after.
A method section somebody could repeat. Not "I analysed the data" but which extract, which years, which exclusions, which test. This is the section student submissions are weakest on and the section reviewers read first.
Honest limits. Reviewers trust a paper that names its own weaknesses more than one that claims more than it shows. Every capable student's first draft over-asserts; saying "this is correlational and here is what would be needed to establish causation" costs you nothing and buys credibility.
Think.Check.Submit. is worth reading before you choose where to send it, and the COPE, DOAJ, OASPA and WAME principles explain why a venue that sells expedited review is a venue to avoid.
The questions that do not work
Anything needing a lab. Not through lack of skill — most universities bar under-eighteens from BSL-2, radioactive and animal work. Work in the computational corner of your field instead.
A survey of your own school. Ninety respondents, no sampling frame, and your classmates answering a questionnaire they know you wrote. It will not carry the conclusion, and it raises consent questions your school has a policy about.
A question with a known answer. Smoking is bad for you. Deprived areas do worse on most outcomes. Executing these well still leaves you with nothing to report.
A tool rather than a finding. Building a working app is an achievement. It is not a research result unless you asked a question of it.
Where we fit
The Oxford Centre for Advanced Research pairs students with PhD candidates and early-career researchers at Oxford, Cambridge, other leading UK universities and Ivy League institutions. The Oxford Scholar Programme is £2,000 — about $2,600 — for ten contact hours over ten to fourteen weeks; the Oxford Publication Fellowship is £3,400, roughly $4,400, and is built around submission.
Steps one to six above are the whole of what happens in a mentor's first three sessions, and we have written them out here because they work whether or not anyone is paid. The thing a mentor genuinely adds is being the person who says your question fails test three — which the person who chose the question cannot see. A teacher with a doctorate, or a graduate student asked one bounded question, will do that for nothing.
The ten-minute version
Write your question down. Underneath it, write what evidence would change your answer. Underneath that, write the result that would surprise you.
Three sentences. If you can write all three, start. If you cannot write the second or third, you have saved yourself a term.
Frequently asked questions
What makes a research question publishable?
Three things. It names the specific variables being compared, it can be answered with evidence you can actually obtain, and the answer is not already known. Journals reject student work for failing the third far more often than the first — a well-executed study of something already settled has nowhere to go.
How do I know if my question is original?
Search Google Scholar for your two variables plus your dataset name. If nothing comes back, widen the search before celebrating — a total absence usually means you have the wrong search terms, not an untouched field. If you find close work, narrow: a subgroup, a different period, a control nobody applied.
Should I pick the question or the data first?
The data, almost always. This inverts the usual advice and it works better, because a question invented in the abstract tends to need evidence that does not exist. Starting from a dataset you can open guarantees the question is answerable before you commit to it.
How long should this take?
An afternoon to a week. If you have spent a month without a question, you are reading rather than deciding, and more reading will not resolve it. The move that breaks the deadlock is opening a dataset and looking for a pairing, not finding one more review article.
What if my result is boring?
A null result is a real result and some journals will publish it if the question was worth asking and the method was sound. What is genuinely unpublishable is a question whose answer was never in doubt. Choose for openness at the start and you avoid the problem.
Do I need a mentor to do this?
No, but you need someone to tell you your question fails. The failure mode is that the person who chose a question cannot see why it does not work. A teacher with a doctorate or a graduate student asked one bounded question will do this for free.