Confirmatory versus exploratory findings: how to label them so reviewers trust both

Why the label changes the meaning

A p-value only means what it is supposed to mean when the analysis was fixed before the data were seen. Once you choose an outcome, a subgroup, or a model because it looked promising in the data, the reported uncertainty is understated, sometimes badly. Nothing about the arithmetic changes. What changes is how many chances the data had to produce something that looks like a result.

That is why the same odds ratio with the same confidence interval deserves different language depending on its origin. A pre-specified comparison supports a claim. A comparison found while exploring supports a question.

What makes a finding confirmatory

Three things, all written down before analysis: the outcome and how it is scored, the comparison or model including the covariates, and the decision rule, meaning the test and threshold. A registered protocol or a dated analysis plan is the cleanest evidence. An internal document shared with the team before the data were unblinded also works, and journals increasingly ask to see it.

Studies that reuse existing data can still have confirmatory analyses. The plan is written before the analyst opens the dataset for that question, and the analysis follows it. What cannot be made confirmatory after the fact is a question that emerged from the data themselves.

How to report the exploratory set

State how many comparisons were examined, not only how many are reported. Apply a correction across the exploratory family, such as the Benjamini-Hochberg false discovery rate procedure, and report the adjusted values. Present the results as associations that suggest follow-up rather than as effects. And resist the temptation to move an exploratory finding into the confirmatory section because it turned out to be interesting. Reviewers recognize the pattern, and the guidance against it is explicit.

In practice this means two clearly separated blocks in the results: the pre-specified analyses first, then a section titled exploratory analyses with its own table, its own correction, and its own more cautious wording.

• Confirmatory: pre-specified, reported first, plain claims with confidence intervals.
• Exploratory: labelled, with the number of comparisons examined and a false-discovery-rate correction.
• Language: confirmatory findings show, exploratory findings suggest.
• Next step: exploratory results become the pre-specified questions of the following study.

A worked example from a client study

In the RRIPG injury questionnaire study on this site, the deliverables were reported as two confirmatory findings and three exploratory findings. The confirmatory pair answered the questions the research group brought to the project. The exploratory three were labelled as such and reported separately, so the research group could read each set with the right level of certainty and decide which exploratory questions to carry into future data collection. Keeping the two sets apart is what lets a reader trust both.