Choose your study type, and design backwards from its reporting guideline, Djandan Tadum Arthur Vithran

In Part 1 you turned a topic into a question with named, measurable parts. If you did that honestly, you have already made most of your next decision for you. The shape of a question quietly dictates the shape of the study that can answer it. The trainees who struggle here are usually the ones who pick a design they admire, almost always a randomised trial, and then try to bend their question to fit it. That is backwards. The design serves the question, never the other way round.

So this part does two things. First, it matches the kind of question you are asking to the kind of study that can answer it. Second, and this is the part almost nobody teaches, it shows you how to read the reporting guideline for that study before you collect a single data point, so that the paper writes itself instead of fighting you at the end.

The question already names the design

Start by asking what kind of thing your question wants to know.

If you are asking “does this cause that?” or “does this treatment work?”, a question about effect, you need a comparison, and ideally a randomised one. That is a randomised controlled trial (RCT).

If you cannot randomise, because the exposure is a disease, a habit, an injury, or anything you would never assign to a person, you are in observational territory. A cohort study follows groups forward in time and asks what happens to the exposed versus the unexposed. A case-control study starts from the outcome and looks backward for the exposure, efficient when the outcome is rare. A cross-sectional study takes a single snapshot and is the right tool for “how common is this?”

If you are asking “what does all the evidence, taken together, say?” you need a systematic review, and where the data allow, a meta-analysis that pools the numbers. When the field is so large that even the reviews need reviewing, you write an umbrella review of those meta-analyses.

And if you have seen something genuinely unusual, one striking patient, or a short series of them, that is a case report or case series. Modest in the hierarchy, but the form in which much of medicine first notices something new.

There is a rough ladder of strength here, case report at the bottom, then cross-sectional, case-control, cohort, RCT, and systematic review of RCTs near the top. It is worth knowing, but do not worship it. A clean, well-run cohort study answers many questions better than a small, badly-run trial ever could. The best design is not the one highest on the ladder; it is the strongest one your question, your data, and your resources can actually support. That feasibility test from Part 1, can I really do this?, decides this as much as ambition does.

The move almost nobody makes: read the guideline first

Here is the single most useful habit in this entire series, and it costs you one afternoon.

For every major study type, the research community has published a reporting guideline, a checklist of everything a complete, trustworthy paper of that type must contain. These are not house styles or journal quirks. They are the field’s hard-won consensus on what a reader needs in order to believe and reproduce your work. Most good journals now require you to submit the completed checklist alongside the manuscript.

The mistake is to discover these checklists at the end, when the study is done and you are writing up, and to find that the item the checklist demands is the measurement you never took, the consent detail you never recorded, the flow of participants you can no longer reconstruct. By then it is a hole in the paper, and reviewers see holes.

The fix is almost embarrassingly simple: read the checklist before you begin. Print it. Treat it as the blueprint for your data collection. If the guideline says you must report how many patients were eligible, excluded, and why, then build that count into your spreadsheet from day one. If it says you must state your primary outcome in advance, then state it, in writing, before you look. Design the study so that the reporting is already complete when the data stop coming in. The write-up stops being an act of reconstruction and becomes an act of transcription.

The checklist you need depends on your design:

CONSORT, for randomised controlled trials.
STROBE, for observational studies (cohort, case-control, cross-sectional).
PRISMA, for systematic reviews and meta-analyses.
CARE, for case reports.
ARRIVE, for animal research.
STARD for diagnostic-accuracy studies, TRIPOD for prediction models, and others for the less common designs.

You do not need to memorise these. You need to know they exist, and to find the right one before you start. They are all kept in one place, the EQUATOR Network (equator-network.org), which is the first website I send every trainee to.

Register before you start, it is not optional any more

Two designs carry a further obligation that, if you skip it, can make your study unpublishable no matter how good it is.

For a randomised trial, you must register the protocol in a public registry (such as ClinicalTrials.gov) before you enrol your first patient. Most reputable journals will not publish an unregistered trial, this is a direct legacy of the era when inconvenient trials simply vanished. Registration fixes your primary outcome in public, before the results can tempt you to change it.

For a systematic review, register your protocol in PROSPERO before you begin screening. It is quick, it is free, and it protects you: if another group publishes first, your timestamped registration shows your work was independent and underway.

Both are forms of the same discipline, committing to your question and your primary outcome before you see the answer. That commitment is what separates a test of a hypothesis from a hunt for a flattering result.

A worked example, from my own work

When my colleagues and I set out to settle how strongly physical activity protects against death, the question was about effect at the population scale, but no single trial could ever answer it, and dozens of meta-analyses already existed, each with its own slice of the evidence. The question, what does the totality of the meta-analytic evidence say about physical activity, sedentary behaviour, and mortality?, pointed unambiguously to one design: an umbrella review, a systematic review sitting on top of the existing meta-analyses.

That decision made the rest follow. Because the design was a systematic review, the guideline was PRISMA, and the obligation was a PROSPERO registration before screening. We wrote the eligibility criteria, the search strategy, and the quality-appraisal method in advance, against the PRISMA checklist, which meant that when the screening was done, the Methods section was effectively already written, and the participant-flow diagram was just a matter of filling in numbers we had been counting all along. The paper was published in the Journal of Cachexia, Sarcopenia and Muscle. It was not a harder paper to write than a weak one, it was an easier one, because the design and its checklist had done the organising before the writing ever began.

The lesson generalises. The afternoon spent choosing the right design and reading its guideline is not overhead. It is the cheapest time you will ever spend on the paper, and it buys back weeks at the end.

The traps at this stage

Design envy. Reaching for an RCT when your question and resources can only support an observational study. A strong cohort beats a doomed trial every time.
The retrofitted checklist. Finding STROBE or PRISMA after the data are collected, and discovering the one thing it requires is the one thing you never recorded.
Skipping registration. Running the trial or the review first and registering after, which, for many journals, is the same as not registering at all.
Outcome drift. Letting the interesting result you happened to find quietly become the “primary outcome” you claim you set out to test. Registration exists precisely to stop this, including stopping you from fooling yourself.

Your turn, one afternoon

Before you collect anything:

Write your question from Part 1 at the top of a page. Name what kind of question it is, effect, frequency, association, or synthesis.
Choose the strongest design your data and resources can actually support. Write one sentence on why a stronger design is not feasible, and be honest.
Go to the EQUATOR Network, find the reporting guideline for that design, and print the checklist.
Go through it line by line and mark every item as already planned, need to add to data collection, or not applicable. The middle category is your to-do list, handle it now, while you still can.
If you are running a trial or a systematic review, register it. Today, before you start.

Do this, and you will have done something most published authors never quite manage: you will know, before the first data point, exactly what the finished paper must contain.

Next in this series, Part 3: The skeleton before the prose, outline, and build your figures first. Why the fastest way to a finished manuscript is to design the figures and the section-by-section outline before you write a single paragraph, and how the “one message” from Part 1 decides what every figure is allowed to show.