Learn · Methods, in the open

Registered Reports: reviewed before you run the study

The most expensive peer review comment is simple: this study cannot answer its research question. After data collection, that problem cannot be fixed. Registered Reports review the design before the first participant is enrolled.

By Heidi Andersén, MD, PhD. Updated 1 August 2026.9 minute read

The Registered Report is one of the biggest changes in scientific publishing of the past decade, and it works in two stages.

Stage 1, before data collection

You submit the study before it exists:

• the research question and the rationale for it,
• the hypotheses,
• the methods,
• the sample size justification,
• the statistical analysis plan.

Peer reviewers evaluate whether the question is important and whether the methods are rigorous. There are no results to read, so there is nothing else to judge on.

If the study passes Stage 1 review, the journal grants an in-principle acceptance. At that point, the journal commits to publishing the paper provided that:

• the study is conducted in accordance with the approved protocol,
• any deviations from the protocol are transparently disclosed and justified, and
• the final manuscript satisfies the journal's pre-specified quality checks.

Stage 2, after the study

Reviewers check whether you followed the protocol and whether the conclusions match the evidence. They do not judge the paper on whether the results are positive, statistically significant, or exciting.

What the format reduces

Moving the review earlier takes the incentive out of four well-known problems at once:

• Publication bias. The decision to publish is made before the results exist, so the results cannot influence it. Negative and null findings become as publishable as positive ones.
• P-hacking. The analysis plan is fixed before anyone can see which analysis flatters the result. You can still run exploratory analyses. You have to label them as what they are.
• HARKing, hypothesising after the results are known. The hypotheses are on the record before there is any data for them to be fitted to.
• Selective reporting. The outcomes were named at Stage 1, so an outcome that goes missing at Stage 2 is visible as missing.

A fifth effect gets less attention, and for most authors it is the one that pays: design criticism arrives while the design can still change. The reviewer who would have written "underpowered for this comparison" in a rejection letter writes it while you can still fix the sampling.

Is it the right format for your study?

It fits confirmatory work: a stated hypothesis, a design chosen to test it, and data not yet collected. It fits badly where the work is genuinely exploratory, and forcing exploratory work into a confirmatory frame is its own distortion. Some journals also offer the format for meta-analysis or for secondary analysis of existing data, where the constraint becomes committing to the analysis before you look.

Which journals offer it

Many respected journals do, including titles published by Nature Portfolio (selected Nature journals), Springer Nature, PLOS, BMC, the Royal Society, BMJ, Elsevier, Wiley, SAGE and APA.

Read that as journals rather than as publishers. The format is adopted title by title, so one journal from a publisher offering it tells you nothing about the journal next to it. The Center for Open Science keeps the registry, more than 300 journals at the time of writing, some as a standing submission option and some for a single special issue: cos.io/initiatives/registered-reports. Check your target journal there, then read its own Stage 1 instructions, because what Stage 1 must contain varies between them.

Two checkpoints, two different questions

A Registered Report and a pre-submission check are often confused because both involve a checklist. They sit at opposite ends of the study and answer different questions.

They are not alternatives. A study can go through both, and a Stage 2 submission still benefits from the second one.

Which checklist belongs to which stage

A Stage 1 submission is a protocol. The completed-study checklists describe things that have not happened yet, so they are the wrong instrument for it:

For a randomised trial, a Stage 1 submission should be prepared as a protocol, using the journal's Registered Report template and the relevant protocol guideline, such as SPIRIT. CONSORT is intended for the completed trial report and includes results-stage information that does not yet exist.

How to use a checklist

Most advice stops at "complete the checklist for your study type". Five rules matter more than which checklist you picked:

1. Fill it while you write. Completed at the end it is a compliance form, and it changes nothing, because the paper is already written. Used while writing it is a design instrument: an item you cannot answer is usually a section you have not thought through, found at the point where thinking is still cheap.
2. A mark is your attestation, not a finding. When you tick an item you are telling an editor that your paper reports that item adequately. That is a claim you are making about your own work, and an editor relies on it. It is not the same kind of object as a tool telling you a sentence exists.
3. Presence is not adequacy. Finding a funding sentence in your manuscript tells you a sentence is there. Whether it says what this journal requires is a reading you have to do.
4. Do not let anything fill it in for you. A tool, or an AI agent, can tell you what it can see in the text. It cannot attest on your behalf. Transcribing its findings into your checklist means attesting to something you have not checked.
5. Give the location. Journals ask where each item is reported, by page or section. If you cannot point at a place, the item is not reported yet, whatever the tick says.

Where tools help, and where they stop

Stage 1 review asks a harder question than whether the proposed sample size reaches 80% power. It asks whether the study, as designed, can support the claim it intends to make.

A conventional power calculation tests one part of that problem. It does not show whether the estimand survives attrition, whether the planned estimator remains calibrated under plausible missingness, whether the measurement process distorts the target, or whether apparently successful estimates become systematically exaggerated.

RecoverLite tests the design as a system.

You specify the estimand, data-generating assumptions, measurement model, missingness process, recruitment and attrition pattern, and planned analysis. RecoverLite then simulates the full study under the target scenario, the null, and declared perturbations of the assumptions. The planned analysis is applied to every simulated dataset.

The output is not a prediction that the study will succeed. It is evidence about whether the proposed design can recover its intended target under the conditions tested. Results are evaluated against versioned decision criteria and reported as PASS, RISK, or FAIL. Sensitivity to stricter and more permissive thresholds is shown explicitly. A verdict that changes across profiles identifies a design close to a decision boundary.

RecoverLite is available as R and Python packages and runs locally.

My services at forskai.com extend this beyond a software run. The reviewed design service examines the alignment between the research question, estimand, sampling process, measurement, missingness, analysis, decision thresholds, and intended claim. The purpose is to identify where the study may fail before recruitment or data collection makes those failures expensive.

After the study, the Pre-Submission Check runs the second checkpoint in the browser: the statement scan, the reporting-guideline item walk, and links out to the reference check, which resolves identifiers against Crossref, DataCite and OpenLibrary, and to the claim-wording and claim-to-source steps. It flags what is open for you to judge. It does not decide whether the manuscript is ready, because that is not a thing a machine can settle.

If you work with an AI agent

An agent can run parts of the second checkpoint and none of the first. Ask it to say which is which:

Use the Vahtian presubmission-check skill:
https://raw.githubusercontent.com/heidihelena/vahtian/main/skill/presubmission-check/SKILL.md

My manuscript is finished. Run the steps you can run, list every step
you cannot as "not run" with its link, and give me the list of what to
fix. Do not mark any reporting-guideline item for me: those marks are
my attestation to the editor. Do not give me a readiness verdict.

The skill is written so that an agent reports the steps it cannot run rather than approximating them, which is the failure mode worth guarding against: an agent that quietly re-implements a check and reports the guess as a tool result. More on agent use at Vahtian for agents.

Related guides

Downstream: how to write a scientific paper and how to respond to reviewers. To make the analysis behind either stage citable, see make your analysis code citable. More guides on the Learn page.