What is preregistration in research?

The problem preregistration solves is that a hypothesis written after seeing the data is indistinguishable, on the page, from one written before. Both appear in the introduction, both are followed by a test, and a reader has no way to tell which came first. Where that distinction matters, and it matters whenever a p value is being interpreted, a timestamp is the only evidence available.

Registering is less onerous than it sounds. You state the question, the hypotheses, the design, the sample size and how you decided it, the variables, the analysis you will run, and the criteria for excluding data. Most of that you have to decide anyway before collecting, and writing it down usually improves the design because vagueness becomes visible.

The level of detail that makes a registration useful is higher than most first drafts reach. Henderson and Chambers set the bar at enough detail for an independent researcher to repeat the study without asking you for clarification, which includes the order in which you will apply exclusion criteria, the exclusions themselves, and what you will not do, such as stopping after two unsuccessful attempts to contact a study author (Henderson and Chambers, 2022, p. 6). The target is removing undisclosed procedural and analytical flexibility, not removing all changes.

The benefit people notice first is not credibility but clarity. Committing to an analysis plan forces you to work out what you will do with the data before the data can influence the decision, which prevents the situation where three plausible analyses give three different answers and you have no principled way to choose.

The main cost is that a preregistration can be wrong. Your planned analysis may turn out to be inappropriate once you see the data's actual properties. That is allowed. You report the deviation, explain it, and report both the planned and the revised analysis where possible.

A registration also does not make a result trustworthy by itself. Yamada argues that registration can be cracked and strategically misused, so a registered finding should not get automatic credit, and recommends several preregistered replications instead. Yamada ties the concern to the 2012 report by John and colleagues that more than 30% of psychological researchers admitted involvement in questionable research practices (Yamada, 2018, pp. 1-2). Ioannidis qualifies the scope differently. Upfront registration applies most cleanly to randomized trials, while hypothesis-generating research may need more flexible registration or networks that link data collections and investigators, and the worked examples put the positive predictive value of an adequately powered randomized controlled trial with little bias at 0.85, against 0.20 for an adequately powered exploratory epidemiological study and 0.0010 for discovery-oriented research with massive testing under substantial bias (Ioannidis, 2005, pp. 698-701). Registration tells a reader which analyses were planned. It does not repair low power, biased measurement, or selective outcome reporting.

Does preregistration stop me exploring the data?

No. It separates confirmatory from exploratory analysis rather than banning the second. Report the preregistered analyses as planned, then report exploratory findings clearly labelled as exploratory. That labelling is the whole point, and it makes exploratory work more credible rather than less.

Exploratory analysis is how most interesting findings appear, and nobody serious argues against it. What preregistration prevents is exploratory findings being presented as though they had been predicted, which is where the inferential damage happens.

A useful habit even without a formal registration: date and save your analysis plan before you look at the data, and keep it. It is weaker evidence than a public registry entry and much better than nothing.

What is a registered report?

A publication format where the introduction and methods are peer reviewed before data collection, and acceptance in principle is granted regardless of how the results turn out. It removes the incentive to find a significant result, and it means reviewer feedback arrives while the design can still change.

The two stages divide the work in a specific way. Stage 1 carries the introduction, the hypotheses where relevant, detailed methods, and analysis plans, and acceptance in principle commits the journal to publishing the completed study whether the results are positive or negative, provided you follow the approved plan. Stage 2 adds the results and discussion, and review at that point asks whether you complied with Stage 1 and whether the conclusions fit the results (Henderson and Chambers, 2022, p. 1). That results-agnostic decision is aimed at publication and reporting bias, and it is not a promise that every preregistered study gets published.

Most of the feasibility planning has to happen before you submit Stage 1. Pilot data are not required, but Henderson and Chambers point out that piloting often reveals exclusion criteria you had not anticipated and lets you write the analysis code in advance, while warning that pilot data are usually a poor basis for a power calculation because they introduce bias. They also ask you to disclose any work completed before Stage 1 submission, and after acceptance to preregister the approved report with its materials, analyses, code, and any pilot data or simulations, either publicly or under embargo (Henderson and Chambers, 2022, pp. 3, 9).

For a doctoral student this format has a specific advantage: acceptance in principle before the study runs removes the risk of spending a year on work that turns out to be unpublishable because the effect was null.

The cost is time at the front. Review of the stage one submission takes months, which has to be planned into the schedule. Where the timeline allows it, it is one of the better deals available in publishing.