Prespecified and exploratory analysis

In one sentence

A prespecified analysis is planned before relevant results are examined, while an exploratory analysis investigates additional questions and must retain the context of that search.

The intuition

Before opening an envelope, you write down what would answer your question. After opening it, you notice something unexpected and ask a new question. Both can be useful. The difference is whether seeing the result shaped the question or rule you now present.

How it works

A trial's protocol and statistical analysis plan identify outcomes, comparisons, populations, methods and how multiple claims will be handled. Their dates and amendments help readers establish what was planned before relevant information became available. Merely writing a plan before publication is insufficient.

An analysis can be prespecified without being a confirmatory primary test. Exploratory endpoints can also be planned in advance. Check the analysis's assigned role and multiplicity, rather than treating “planned” as an automatic certificate of proof. FDA: endpoint hierarchy and exploration.

Post hoc means developed after the relevant results were available. Such analyses can generate hypotheses, investigate unexpected patterns or clarify limitations. Report what was searched and why. A new independent comparison can test the resulting hypothesis.

An amendment made for a reason unrelated to unblinded results is different from choosing the most favorable analysis after examining outcomes. Retain the amendment, timing, access to data and rationale. CONSORT 2025: transparent trial reporting.

Why it matters in cancer

Trials measure many markers and subgroups. An attractive response pattern in a small selected subgroup may suggest a next study. It does not establish that the marker selects a treatment benefit, especially when the subgroup was chosen after the outcome was known.

Two significance labels cannot establish that effects differ. A significant result in one subgroup and a non-significant result in another need a direct comparison of treatment effects, often an interaction test. Sun and colleagues: subgroup credibility.

Worked example

A fictional protocol names event-free survival as the primary outcome and a baseline marker-positive subgroup as exploratory. After seeing results, investigators test 12 further marker cutoffs and highlight the one with the smallest p-value.

Keep three statements separate: the planned primary result; the planned exploratory subgroup result; and the new cutoff-generated hypothesis. The last search should include all tested cutoffs and its selection method. A new study can lock the cutoff and evaluate whether it predicts a treatment difference.

Common confusions

Prespecified does not automatically mean adequately powered, unbiased or relevant.
Exploratory does not mean worthless; its job is different.
Calling a post hoc analysis “validation” does not create independent data.
A plan revised after results were seen cannot be treated as the original plan.
Response-defined groups formed after treatment are different from baseline categories.

Try it

The paper says an analysis was planned, but its uploaded plan is dated after the results. What would you ask for?

Answer: The relevant protocol and earlier plan versions, amendments, their dates, and who had access to the outcomes. Do not infer either fraud or confirmation from the label alone.

Sources and scope

Source check: October 9, 2026. Fictional protocol and searches. Expert and learner review remain pending.