Which Analyses Were Secondary or Exploratory?

Which results extend the main question rather than answer it directly?

A paper may begin with one primary question and finish with dozens of results. Some examine secondary outcomes. Others test subgroups, alternative models, interactions, additional time points, potential mechanisms, or relationships that became interesting during analysis.

These analyses can be scientifically valuable. The problem arises when their role becomes blurred. A striking exploratory result can be presented as though it carried the same confirmatory weight as a prespecified primary analysis. Conversely, calling every non-primary analysis “mere exploration” can dismiss useful evidence. The task is to identify what each analysis was intended to do and when that decision was made.

Classify analyses by their purpose, hierarchy, and timing

How to distinguish additional analyses from the study's main analysis

Start by establishing the primary analysis

You cannot reliably identify secondary or exploratory analyses until you know what the primary analysis was.

Once the primary question, outcome, time point, comparison, analysis population, and model are clear, ask what role each additional analysis plays relative to that primary analysis.

Some analyses address different prespecified outcomes. Some test whether the primary conclusion is robust. Others examine whether effects vary between subgroups. Still others investigate patterns that were not central to the original question.

Secondary and exploratory are not simply synonyms

Terminology varies among disciplines, so the distinction is not perfectly universal. Still, a useful conceptual separation is possible.

An exploratory analysis can be prespecified. A secondary analysis can sometimes be developed after data collection. This is why the analytical label and the timing of the decision should be recorded separately rather than treated as interchangeable.

Prespecified does not automatically mean primary

A common mistake is to divide analyses into “prespecified primary” and “post hoc exploratory.” Real analytical plans are more complicated.

A trial might prespecify one primary analysis, five secondary outcome analyses, three subgroup analyses, and several sensitivity analyses. All were planned, but they do not all carry the same role.

CONSORT 2025 explicitly asks reports to describe methods for additional analyses such as subgroup and sensitivity analyses and to distinguish those that were prespecified from those conducted post hoc.

This gives you two separate questions:

• What role did the analysis have: primary, secondary, sensitivity, subgroup, exploratory, or another role?
• When was the analysis decided: before the relevant results were known or afterward?

Post hoc and exploratory are related but not identical

Post hoc describes timing: the analysis was formulated after a relevant point in the study process, commonly after data or results were available. Exploratory describes inferential purpose more than chronology.

A study can prespecify exploratory analyses because researchers know in advance that they want to investigate potential patterns without treating those analyses as confirmatory. Conversely, an unplanned post hoc analysis may investigate a scientifically plausible question but should still be identified as having been developed later.

When timing matters, examine what was prespecified and what appears to have been decided later.

Secondary outcome analyses answer additional outcome questions

A trial might designate symptom severity at 12 weeks as its primary outcome while also measuring quality of life, treatment satisfaction, adverse effects, and functional status as secondary outcomes.

Those secondary outcomes may be important, sometimes very important. Their secondary status means they occupy a different position in the study's inferential hierarchy, not that they are scientifically trivial.

When several secondary outcomes are tested, multiplicity can become relevant because the opportunity for chance findings increases as more hypotheses are examined. Whether and how multiplicity should be controlled depends on the study's confirmatory claims and analytical framework.

Subgroup analyses ask whether results differ across groups

Researchers may examine whether an intervention effect differs by age, sex, baseline severity, institution, disease subtype, socioeconomic status, or another characteristic.

The critical statistical question is usually not whether the intervention is statistically significant in one subgroup and non-significant in another. It is whether the effect itself differs between subgroups, often assessed through an interaction or other formal test of heterogeneity.

Subgroup findings deserve additional caution when many groups were examined, sample sizes are small, subgroup definitions were chosen after looking at results, or no plausible rationale existed before analysis.

Sensitivity analyses have a different job

A sensitivity analysis is generally intended to test the robustness of conclusions to assumptions or analytical choices rather than answer a new substantive research question.

For example, researchers might repeat the primary analysis using alternative assumptions about missing data, a different but defensible model specification, or another approach to a consequential analytical uncertainty.

ICH E9(R1) emphasizes that sensitivity analyses should be aligned with the same estimand as the main analysis and examine how assumptions affect the reliability of the primary estimate. It distinguishes these from supplementary analyses, which can provide additional insights and may address different estimands.

Exploratory analyses can be useful precisely because they explore

Exploration is a legitimate part of science. Unexpected patterns can suggest new mechanisms, identify questions worth testing, reveal heterogeneity, or motivate subsequent studies.

The problem is not exploration. The problem is presenting an exploratory finding as though it were a clean confirmatory test of a hypothesis selected independently of the data.

A transparent paper can say, in effect: “We observed this pattern in an exploratory analysis; it warrants further investigation.” That interpretation preserves the finding without pretending that the study provided stronger confirmation than it did.

The number of analyses matters

If researchers test one hypothesis at a conventional significance threshold, chance findings remain possible. If they test dozens or hundreds of outcomes, subgroups, transformations, models, and time points, opportunities for apparently interesting results increase.

This is the broader multiple-testing problem. The appropriate response depends on the inferential framework. Some confirmatory analyses use formal multiplicity adjustments or hierarchical testing procedures. Exploratory research may instead emphasize effect estimates, uncertainty, consistency, plausibility, and replication rather than treating every nominal P value as a definitive discovery.

When reading a paper, therefore, ask not only “Is this result statistically significant?” but also “How many opportunities were there to find something like this?”

Secondary analyses may use different analytic populations

A secondary outcome might be measured only in a subset. A mechanistic analysis may require laboratory data unavailable for many participants. A subgroup analysis deliberately restricts the population. A long-term analysis may include only those retained at later follow-up.

Thus, additional analyses may involve different participants and data from the primary analysis. Check what data were actually analyzed rather than assuming the primary-analysis denominator applies throughout the paper.

Alternative models are not all sensitivity analyses

Researchers sometimes present many model specifications and call them robustness checks. That label does not automatically establish that they function as genuine sensitivity analyses.

Ask what assumption each alternative analysis changes and why. ICH E9(R1) recommends a structured approach in which the assumptions being varied are made explicit. Simply trying numerous models without explaining what uncertainty each addresses makes interpretation difficult.

Exploratory findings should not quietly become the paper's conclusion

One of the most important reading checks occurs in the discussion and abstract. Compare the prominence of the conclusion with the status of the analysis supporting it.

If the paper's strongest claim depends on a small, unplanned subgroup analysis while the primary result was inconclusive, that hierarchy should remain visible in your interpretation. The exploratory result may still be interesting. It should not silently inherit the evidential status of the primary analysis.

Secondary does not mean disposable

It is equally important not to overcorrect. Safety outcomes, quality of life, adverse events, implementation measures, and other secondary outcomes can be central to practical decision-making even when they are not designated primary.

The analytical hierarchy tells you how the study was structured. It does not provide a universal ranking of scientific importance. A secondary outcome can matter enormously while still requiring interpretation according to how it was planned, measured, analyzed, and tested.

Sorting one study's results into their actual analytical roles

Common mistakes when interpreting secondary and exploratory analyses

Give every analysis a role before giving its result a meaning

When a paper contains many analyses, annotate them by function rather than reading them as one undifferentiated stream of results. This is particularly useful when the discussion emphasizes findings that were not central to the original analytical plan.

This classification does not tell you mechanically which findings to believe. It tells you what inferential job each analysis was designed to perform. That is a much better starting point than sorting results into “significant” and “not significant” and hoping the methods section forgives us.

Before interpreting additional results, identify what kind of analyses produced them

Questions about secondary and exploratory analyses

Additional analyses can inform the evidence without becoming the primary evidence

How to Cite This Guide

This guide is intended to be read, shared, and used in research, teaching, and academic work. If you draw on its ideas, explanations, or other content, please acknowledge the source by citing the guide. Doing so gives appropriate credit and helps your readers locate the original resource.