What a Real Preregistration Looks Like in a Clinical Trial

When a clinical trial appears in a peer-reviewed journal, somewhere in the methods section there is usually a registration number. ClinicalTrials.gov, ISRCTN, the EU Clinical Trials Register. Readers are meant to find that reassuring. The trial was logged before it ran. The primary endpoint was declared in advance. The analysis plan was committed to paper before anyone unblinded the data.

In practice, the distance between a preregistration that functions as advertised and one that provides only the appearance of constraint is considerable. Understanding that distance matters if you are reading a trial, writing about one, or designing one.

The core purpose of preregistration is to close off what researchers call the garden of forking paths: the many legitimate-seeming analytic choices that, taken together, can produce a statistically significant result from data that contain no real signal. Which patients get excluded post hoc. Whether the outcome is the raw score or the change from baseline. Whether the analysis is per-protocol or intention-to-treat. Any one of these decisions is defensible. The problem is that there are dozens of them, and if the researcher can choose after seeing the data, the nominal p-value of 0.05 stops meaning what it is supposed to mean.

A preregistration that does real work specifies the primary endpoint with enough precision that there is no ambiguity when the analysis is run. Not just "depression score" but which instrument, which subscale, which time point, and what summary statistic. It names the statistical test, states the covariates that will enter the model, and distinguishes confirmatory from exploratory analyses. It describes the rules for handling missing data before missingness is known. It documents the sample size calculation and what assumptions that calculation rests on.

A preregistration that does not do real work often looks similar on the surface. It names an outcome. It gives a time frame. But it leaves enough interpretive slack that the registered plan and the published analysis can diverge without anyone being technically wrong. Researchers who have compared registered protocols to published papers, a genre of methods literature that has grown substantially since the mid-2010s, consistently find that a meaningful fraction of trials shift their primary outcome between registration and publication, drop secondary outcomes that did not pan out, or add outcomes that did. This is sometimes called outcome switching, and it is not always cynical. Priorities change. Unexpected safety signals emerge. But the record should show the change, and it often does not.

Timestamp hygiene is one underappreciated dimension. Registries record when entries are created and when they are modified. A registration submitted after data collection is complete but before publication does not serve the same function as one submitted before enrollment opens. Some registries flag late registration explicitly. Others require the submitter to self-report the timing. Readers who want to evaluate a trial properly should check the registration date against the dates of first enrollment.

Amendments are legitimate and sometimes necessary. A protocol that bans any modification is not realistic. The question is whether amendments are logged with a rationale, time-stamped, and distinguishable from the original plan. Versioned, dated amendments visible in the registry record are a feature. Wholesale rewrites with no audit trail are not.

None of this makes preregistration worthless. It remains one of the more structurally sound tools available for constraining post hoc flexibility, and the field has moved toward more rigorous standards over the past decade. Registered Reports, a journal format that offers conditional acceptance before data collection, push the logic further by separating the editorial judgment of the hypothesis from the judgment of the result.

But a registration number in a methods section is not the same thing as a binding pre-analysis plan. Reading the actual registry entry, checking the timestamp, comparing the registered outcomes to the published ones: that is the work that turns a reassuring footnote into a real constraint.