03 · What You Need to Know
What Actually Separates Exploration From Confirmation
Exploratory Research Asks What the Evidence Might Reveal
Exploratory research is oriented toward discovery.
Researchers may investigate patterns, characterize a phenomenon, identify possible relationships, develop constructs, compare alternative explanations, generate hypotheses, or determine which questions deserve more focused study.
The analytical path may evolve as researchers learn from the evidence. An unexpected relationship can lead to another analysis. A surprising case can redirect attention. A visualization can suggest a subgroup worth investigating.
This responsiveness is not a methodological defect. It is part of what exploration is for.
The important limitation is inferential. When a pattern helps generate a hypothesis and the same pattern is then presented as evidence confirming that hypothesis, discovery and testing are no longer independent.
Confirmatory Research Puts an Advance Expectation at Risk
Confirmatory research begins with a claim that is specified before the evidence used to evaluate it is known.
The researcher might predict that an intervention improves a particular outcome, that two variables are positively associated, or that one theoretical model will outperform another according to specified criteria.
The crucial feature is that the prediction is exposed to the possibility of being wrong.
Nosek and colleagues describe preregistration in terms of distinguishing prediction from postdiction. A prediction made before an outcome is known has not had the opportunity to adapt itself to that outcome. An explanation developed afterward can fit information already observed.
Exploration
Uses evidence to discover patterns, develop questions, generate explanations, or identify hypotheses worth pursuing.
Confirmation
Uses evidence to evaluate a hypothesis or prediction that was specified independently of that evidence, together with a sufficiently defined test.
The Distinction Is About the Relationship Between the Claim and the Evidence
This point is more useful than simply asking whether the data existed when the hypothesis was written.
Suppose a national survey dataset was collected five years ago, but you have never examined it. You develop a theoretically motivated hypothesis and preregister the variables and analysis before accessing the relevant data.
The data are old, but the evidence is new to you.
Now consider the opposite situation. You collect brand-new data, inspect the correlations, notice a surprising association, develop a hypothesis explaining it, and then write the hypothesis into the paper as though it had been predicted before analysis.
The data were newly collected, but the hypothesis was informed by the evidence it supposedly tests.
This is why “before data collection” is a useful default for many studies but not the conceptual definition of confirmation.
Why Data-Dependent Hypotheses Need Different Interpretation
Imagine drawing a target around an arrow after the arrow has already landed. The final picture can look impressive, but the process did not test whether the archer could hit a target specified beforehand.
Data-dependent hypotheses create a related inferential problem. The observed pattern influenced which hypothesis was selected for attention. The same evidence therefore contributes both to generating the claim and to making the claim appear supported.
This does not mean the explanation is wrong. It means the current evidence has a different role.
The appropriate next step may be to treat the finding as hypothesis-generating and evaluate the hypothesis prospectively using new or otherwise independent evidence.
Exploration and Confirmation Can Occur in the Same Study
A research project does not need to choose a single identity.
You might preregister three hypotheses and corresponding analyses while also exploring additional variables. The prespecified tests can serve a confirmatory role, while analyses inspired by unexpected patterns remain exploratory.
The Center for Open Science explicitly preserves this distinction in its Registered Reports guidance. Registered Reports permit unregistered exploratory analyses while requiring them to be distinguished from preregistered confirmatory analyses.
This arrangement recognizes that good science needs both.
Researchers should therefore avoid language suggesting that exploratory work is what remains after the “real” analyses are finished. Exploration generates many of the questions that later become worth testing.
You Do Not Have to Decide Every Exploratory Question in Advance
By definition, some questions arise because researchers learn something unexpected.
If you notice a surprising pattern during analysis, investigate it. If the result suggests an alternative mechanism, develop that explanation. If a subgroup behaves differently, examine the possibility carefully.
The requirement is not foresight. It is accurate reporting.
When an analysis was inspired by the observed data, describe it accordingly. Do not quietly move it into the confirmatory narrative simply because it produces an appealing result.
You Can Plan Exploration in Advance
Exploration itself can be prospective.
You might state before analysis that you intend to examine relationships among a defined set of variables without directional hypotheses, compare several possible models, or use visualization and clustering to identify patterns worth future study.
That work remains exploratory because its purpose is discovery rather than testing a specific predicted outcome.
This is why exploratory research can itself be preregistered. Preregistration and confirmation overlap, but they are not synonyms.
A Preregistered Analysis Is Not Automatically Confirmatory
Suppose you preregister: “We will examine all pairwise relationships among 40 variables and investigate any interesting associations.”
The plan was unquestionably recorded in advance. Yet the eventual hypothesis that emerges from one of those associations was not predicted before the evidence was known.
Calling the resulting discovery “preregistered” may be accurate with respect to the broad exploratory procedure, but calling the specific finding a preregistered confirmatory hypothesis would not be.
Researchers should therefore use these labels carefully. “Preregistered” describes the timing of the documented plan. “Confirmatory” describes the inferential function of a test.
Existing Data Can Support Confirmation Under Some Conditions
Secondary-data research makes the distinction particularly interesting.
A dataset can already exist while the focal hypothesis remains genuinely prospective from the researcher's perspective. If the researcher has not examined the relevant outcome relationships and the hypothesis comes from independent theory or evidence, preregistration before accessing those results can preserve an important form of prospective testing.
However, independence is rarely all-or-nothing. Researchers may know summary statistics, have worked with related variables, have seen prior publications using the dataset, or know patterns from previous analyses.
These exposures should be disclosed because they help readers judge how independent the eventual test really was.
The Open Science Framework currently provides a Secondary Data Preregistration template specifically for research using existing datasets, underscoring that prospective planning does not necessarily require prospective data collection.
Splitting Data Can Separate Discovery From Testing
One way to combine exploration and confirmation within a single dataset is to partition the data before analysis.
For example, researchers might use one subset to explore patterns and generate hypotheses. They then freeze the hypothesis and analysis plan before opening an untouched holdout subset used for confirmation.
This creates a more independent test because the confirmatory observations did not contribute to generating the hypothesis.
The strategy requires genuine separation. Repeatedly inspecting the holdout sample while refining the hypothesis erodes the independence it was intended to provide.
Data splitting also has costs. Each subset contains fewer observations, which can reduce precision or statistical power. Whether the strategy is worthwhile depends on sample size, analytical goals, and the availability of future independent data.
Replication Can Turn an Exploratory Finding Into a Prospective Test
Another common sequence is simpler.
A first study produces an unexpected result. Researchers develop a hypothesis to explain it. They then preregister a second study designed specifically to test that hypothesis.
Explore Observe an unexpected pattern and investigate plausible explanations.
Generate Formulate a specific hypothesis based on the exploratory evidence.
Preregister Specify the hypothesis, design, outcome, and analysis before collecting or examining the new evidence.
Test Use independent data to evaluate whether the prediction survives a prospective test.
This sequence does not diminish the exploratory study. It gives the first and second datasets different jobs.
The Boundary Can Sometimes Be Messy
Research rarely divides perfectly into untouched predictions and completely unconstrained discovery.
A hypothesis may be theoretically motivated but refined after a pilot study. Researchers may know broad properties of an existing dataset without knowing the focal relationship. A planned model may need revision because an assumption fails. Prior studies may already suggest likely outcomes.
It can therefore be misleading to treat “exploratory” and “confirmatory” as moral categories or perfectly binary states.
A more useful practice is to report the chronology and relevant prior knowledge. What generated the hypothesis? What evidence had researchers already seen? Which analytical decisions were specified before the focal result? What changed afterward?
Readers can then evaluate the strength of the confirmatory claim without pretending that all research fits a pristine binary classification.
Neither Type of Research Is Automatically Better
Confirmatory research is useful when a sufficiently developed claim is ready for a prospective test. Exploratory research is useful when researchers need to discover patterns, develop explanations, refine concepts, or generate hypotheses.
A field consisting only of confirmation would eventually run out of interesting claims to test. A field consisting only of exploration could accumulate intriguing patterns without subjecting them to sufficiently independent tests.
The two activities can therefore form a productive research cycle.
Watch Out
The problem is not conducting exploratory analyses after seeing the data. The problem is allowing the history of those analyses to disappear and presenting a data-generated hypothesis as though the same evidence had independently confirmed a prediction made beforehand.
