03 · What You Need to Know
A Dataset Is a Research Resource, Not a Single Research Question
Research datasets are often richer than the projects that originally generated them. A dataset might contain demographic information, behavioral measures, outcomes, exposures, repeated observations, contextual variables, qualitative material, or linked administrative records. No single paper could or should necessarily answer every worthwhile question those data permit.
This creates an important distinction between the data source and the research project.
Dataset
An organized collection of observations or information available for analysis.
Research project
An investigation organized around a defined research question or coherent set of questions, with an appropriate analytical and interpretive plan.
Because those are not the same thing, one dataset may support several projects. Conversely, one project may use several datasets.
A New Question Can Create a Distinct Research Project
Consider a large survey of university students containing measures of digital competence, academic engagement, generative AI use, financial stress, social support, and academic performance.
One project might examine factors associated with students' adoption of generative AI. Another might investigate the relationship between financial stress and academic engagement. Both analyses use the same rows of data and may even use some of the same control variables, yet their scientific purposes are substantially different.
The strongest reason for treating them as separate projects is not that different variables happen to be selected. It is that each analysis begins from a different research problem and requires its own rationale, literature, analytical choices, interpretation, and contribution.
Secondary Analysis Is a Normal Form of Research
Secondary data analysis uses existing data to address a research question beyond the analysis for which those data were originally collected or assembled. The precise definition varies somewhat across fields, particularly because public databases, administrative records, archived qualitative material, and trial datasets originate under different conditions.
Secondary analysis can offer substantial advantages. It may reduce the need to collect new data, make fuller use of research investments, allow analysis of large or difficult-to-recruit populations, and permit questions to be examined that were not central to the original study.
Those efficiencies do not make secondary analysis methodologically effortless. The researcher inherits the characteristics and limitations of the existing data.
The Dataset Must Fit the New Question
Perhaps the most consequential mistake in secondary research is beginning with available variables rather than a defensible question. A dataset can contain something that resembles the construct you want to study without measuring that construct adequately for your purpose.
Before using existing data, examine how the data were generated. Relevant considerations may include:
- the population and sampling procedure;
- inclusion and exclusion criteria;
- how constructs and variables were operationalized;
- measurement reliability and validity where applicable;
- when and under what conditions observations were collected;
- missing data and attrition;
- changes in measurement across waves;
- the original purpose of data collection; and
- whether important confounders or contextual variables are absent.
The logic should remain question first, data second. If you repeatedly alter the research question until it fits whatever variables happen to be available, you may end up answering a convenient question rather than an important one.
This is why the decision to add a research question because data are available deserves separate scrutiny.
Different Analyses of the Same Data Are Not Automatically Different Projects
Running another statistical model does not necessarily create another study. Neither does changing the dependent variable, analyzing a subgroup, adding a moderator, or replacing one analytical technique with another.
The International Committee of Medical Journal Editors recognizes that manuscripts based on the same dataset may legitimately differ in analytical methods and conclusions. Its guidance nevertheless states that manuscripts using the same dataset should add substantially to one another to warrant consideration as separate papers, with appropriate citation of previous publications from the dataset.
The important phrase is add substantially. A genuinely new question may justify a distinct project. A minimally altered analysis designed primarily to produce another manuscript may not.
| Situation |
Distinct project may be defensible when... |
Separation is questionable when... |
| Different outcome |
The outcome represents a distinct scientific question requiring its own rationale. |
The outcome is switched mainly to generate another analysis. |
| Different subgroup |
The subgroup question has a substantive theoretical or practical justification. |
Many subgroups are searched after seeing the data without a compelling rationale. |
| Different method |
The analytical approach answers a meaningfully different question or tests an important alternative explanation. |
The same substantive conclusion is repackaged using another technique. |
| Different variables |
The variables operationalize a separate research problem. |
Variables are rearranged into minimally different models without a new contribution. |
| Different publication |
The manuscript adds substantially to existing work using the dataset. |
Findings substantially duplicate an existing publication without adequate justification or disclosure. |
Data Reuse Does Not Mean the Evidence Is Independent
Two papers can represent distinct research projects while still drawing evidence from the same observations. That distinction should remain visible.
Suppose two articles use the same national student survey. One investigates AI literacy, while another examines academic engagement. They may answer different questions, but they do not constitute two independently recruited samples.
This becomes particularly important when evidence is later synthesized. If overlapping datasets are treated as independent studies in a meta-analysis or systematic review, the same participants may effectively be counted more than once. Researchers conducting evidence synthesis therefore sometimes need to identify overlapping populations or data sources before combining estimates.
Related Publications Should Be Transparent
When several publications arise from the same dataset, readers should be able to understand their relationship.
ICMJE recommends appropriate citation of previous publications from the same dataset. For secondary analyses of clinical trial data, it specifically recommends citing the primary publication, identifying the work as secondary analysis, and retaining the relevant trial registration and persistent dataset identifiers.
These recommendations arise from biomedical publishing, so their exact procedural requirements should not be generalized mechanically to every discipline. The underlying principle is much broader: researchers should not obscure material overlap among studies using the same data.
Watch Out
Do not present analyses from one dataset in ways that imply they come from independent samples or entirely unrelated research when meaningful overlap exists. Transparent citation and description allow readers, reviewers, and evidence synthesizers to understand how the projects are connected.
Dataset Reuse Has Ethical and Governance Conditions
The ability to access a dataset does not necessarily establish permission to use it for any research purpose. Secondary research may be governed by participant consent, ethics approval, institutional policies, data-use agreements, licensing terms, repository conditions, privacy requirements, or applicable law.
These conditions vary considerably. Some de-identified datasets are made available specifically for broad secondary research. Others impose restrictions on research topics, populations, linkage, redistribution, commercial use, or attempts at re-identification. Certain secondary analyses may require ethics review even when researchers do not interact directly with participants.
Researchers should therefore verify the conditions attached to the specific dataset rather than relying on a general assumption that existing data are unrestricted data.
Qualitative Data Can Also Be Reanalyzed
Dataset reuse is not limited to quantitative research. Interview transcripts, field notes, diaries, open-ended responses, audiovisual records, and other qualitative materials may sometimes be analyzed to address questions beyond the original study.
Qualitative secondary analysis introduces particular questions about context. Researchers may not have been present during the original data collection, and the material may have been generated for purposes that differ from the new analytical interest. Assessing the fit between the existing material and the new question is therefore especially important.
Consent, confidentiality, contextual integrity, and the possibility that participants could be identifiable from rich qualitative material also require careful attention.
A Dataset Should Not Become a Question-Generating Machine Without Boundaries
Large datasets make it technically possible to test enormous numbers of relationships. Scientific justification should constrain that possibility.
Researchers should distinguish planned confirmatory analyses from exploratory or post hoc analyses where that distinction is methodologically relevant. Searching extensively for associations and then presenting the most interesting findings as though they were prespecified can distort the evidentiary meaning of the results.
Exploration itself is not the problem. Exploratory analysis can generate valuable hypotheses. The problem arises when the history of the analysis is concealed or when the availability of data becomes the sole rationale for adding questions.
One Dataset Can Support a Research Program
Large longitudinal cohorts, registries, institutional datasets, and major surveys may support dozens or even hundreds of projects over time. In those contexts, it makes little sense to equate the dataset with one study.
Instead, the dataset becomes research infrastructure from which different projects emerge. Those projects may be coordinated within a larger research program, with each project contributing a defined piece of a broader scientific agenda.
