“Research data” is not one kind of thing, and treating it as one is a common source of trouble in data management plans (DMPs) and repository selection. A genomic sequence file, a set of interview transcripts, a spreadsheet of survey responses, and the output of a climate model all count as research data, but they need different metadata, different storage formats, different levels of access control, and often different repositories entirely. This guide sets out the taxonomies research administrators and researchers actually use to classify data, and why the distinctions matter operationally rather than just terminologically.
No single classification is exhaustive or mutually exclusive on its own; most real datasets sit at the intersection of several of these axes at once (a raw, quantitative, experimental dataset in a structured file format, for example). The value of the taxonomy is in forcing explicit answers to a small set of questions—how is this data organized, how was it produced, and how sensitive is it—each of which drives a different part of a data management plan.
Why the distinction matters before you write a data management plan
Funder DMP requirements (see CASRAI’s guide to NIH vs. NSF data management plans and worked DMP examples) generally ask for the same handful of commitments: what data will be generated, how it will be described, where it will be stored, and how it will be shared. Every one of those answers depends on data type:
- Metadata schema choice depends on form and generation method—tabular quantitative data and unstructured qualitative interview data are typically described using different schemas (see CASRAI’s guide to choosing a metadata schema).
- Repository selection depends heavily on data type—a disciplinary repository built for structured tabular data (e.g., a genomics or social-science data archive) may not accept unstructured qualitative data or large simulation outputs, and vice versa (see how to choose an open data repository).
- Sensitivity classification determines whether data can be deposited openly at all, or requires a controlled-access repository and a data sharing agreement.
- Volume and format (raw instrument output vs. a cleaned analysis-ready file) determine storage cost, whether version control is needed, and what "the dataset" even means for citation and reuse purposes.
Getting the classification wrong at the planning stage is one of the more common reasons a DMP commitment turns out to be unworkable at the point of actual deposit—see CASRAI’s guide to the research data lifecycle for how planning-stage decisions propagate through later stages.
Classifying research data by form
Structured data
Structured data is organized into a predefined schema—rows and columns, fixed fields, a relational database—so that individual data points are directly machine-addressable without additional parsing. Examples: a spreadsheet of experimental measurements, a relational database of clinical trial case report forms, a CSV of survey responses with a fixed column layout.
Unstructured data
Unstructured data has no predefined schema and typically requires additional processing (transcription, coding, natural-language processing, manual annotation) before it can be analyzed systematically. Examples: interview audio recordings and their transcripts, free-text field notes, images, video, PDF documents. Qualitative research produces predominantly unstructured data, though the coded output of qualitative analysis (a codebook applied to transcripts) is itself a structured derivative of unstructured source material—see CASRAI’s dictionary entry on qualitative data collection techniques.
Semi-structured data
Semi-structured data sits between the two: it has organizational markers (tags, key-value pairs, a consistent hierarchy) but not a rigid fixed schema. Examples: JSON or XML data exports, electronic health record fields that mix coded values with free-text clinical notes, genomic file formats like FASTQ or VCF that combine structured headers with sequence data.
This distinction matters practically because repository ingest systems and metadata schemas are frequently built around one form or the other; a generalist repository designed for structured tabular deposits may handle an unstructured video corpus poorly, or require it to be packaged and described differently.
Classifying research data by generation method
A second, independent axis classifies data by how it came into being. This is the distinction most directly relevant to documenting data provenance, since it describes the process that produced the dataset, not just its file format.
Observational data
Collected by observing a phenomenon as it naturally occurs, without the researcher manipulating variables. Examples: field ecology surveys, astronomical observations, epidemiological cohort data, social-media data collected for content analysis. Observational data is typically not reproducible in the strict sense—the exact conditions of observation cannot be recreated—which raises the stakes on documenting exactly how, when, and under what conditions it was collected.
Experimental data
Generated under controlled conditions where the researcher deliberately manipulates one or more variables. Examples: laboratory assay results, randomized controlled trial outcome measures, controlled-condition psychology experiments. Because the researcher controls the conditions, experimental data is generally more reproducible than observational data, provided the experimental protocol itself is documented and preserved alongside the data.
Simulation data
Generated by a computational model rather than measured from the physical or social world. Examples: climate model outputs, molecular dynamics simulations, agent-based economic models, computational fluid dynamics runs. Simulation data raises a distinct documentation requirement: the data alone is not reproducible or interpretable without the code, model parameters, and software environment (container image, dependency versions) that produced it—a point the National Academies’ work on computational reproducibility addresses directly, and which CASRAI’s guide to the research data lifecycle touches on for preservation planning. Some simulation output is large enough (climate and physics runs in particular) that repository selection and storage cost, not just documentation, become the binding constraint.
Derived (or secondary) data
Produced by transforming, aggregating, or analyzing other data—raw or otherwise—rather than by direct collection. Examples: a statistical summary table computed from raw survey responses, a normalized gene expression matrix derived from raw sequencing reads, a machine-learning model’s training set assembled from multiple upstream sources. Derived data needs its lineage documented back to its source data (see data provenance) so that a reuser can understand what transformations were applied and whether the derivation itself is reproducible. Increasingly, synthetic data—data generated algorithmically to resemble real data without being drawn directly from real observations—is treated as a distinct sub-case of derived data with its own disclosure expectations; see CASRAI’s guide to synthetic data in research.
Raw vs. processed data: a stage-of-processing distinction, not a data type
A related but separate distinction is raw versus processed (or "clean"/"analysis-ready") data. Raw data is the direct, unmodified output of a collection instrument or process—the sequencer’s base-call file, the survey platform’s unedited response export, the sensor’s logged readings. Processed data has had cleaning, filtering, normalization, or transformation applied.
This is a stage-of-processing distinction that cuts across the form and generation-method taxonomies above—raw data can be structured or unstructured, observational or experimental. It matters for data management planning because funders and repositories increasingly ask which version(s) of a dataset will be retained and shared. Retaining only processed data can make a result harder to independently verify or reanalyze; retaining only raw data without documented processing steps can make published results difficult to reconstruct. Good practice, reflected in most funder DMP guidance, is to document the processing pipeline (ideally as versioned, executable code) that connects the two, rather than treating "the dataset" as a single undifferentiated artifact.
Classifying research data by qualitative vs. quantitative content
Independent of both form and generation method, data is often described as qualitative or quantitative based on the nature of the content itself:
- Quantitative data consists of numerical measurements or counts that support statistical analysis—lab measurements, survey Likert-scale responses, instrument readings.
- Qualitative data consists of non-numerical content—text, images, audio, video—that is typically analyzed through coding, thematic analysis, or interpretation rather than statistics.
- Mixed-methods data combines both within a single study design; see CASRAI’s dictionary entry on mixed-methods research for how this affects study design and, by extension, what a DMP for a mixed-methods project needs to account for on both sides.
This axis overlaps heavily but not perfectly with the structured/unstructured distinction above—quantitative data is usually, but not always, structured, and qualitative data is usually, but not always, unstructured (a structured codebook applied to interview transcripts is qualitative in content but structured in form).
Classifying research data by sensitivity and openness
A final axis, orthogonal to all of the above, is how openly the data can be shared. This classification governs access control, repository choice, and whether a data sharing agreement is required, independent of the data’s form or how it was generated.
- Open data—can be shared without restriction, typically under a Creative Commons or similar open license, once any necessary anonymization or aggregation is complete.
- Restricted or controlled-access data—can be shared, but only under conditions (a data use agreement, an approved-researcher process, a secure enclave) because it contains identifiable, commercially sensitive, or otherwise restricted information.
- Confidential or closed data—cannot be shared at all, whether for legal, ethical, contractual, or national-security reasons, though its existence and high-level characteristics may still need to be documented in a DMP.
Personal and health data occupies a large share of the restricted-data category in practice; CASRAI’s dictionary entry on data ownership and its guide to describing reuse rights and permissions for a shared dataset cover how sensitivity classification interacts with licensing and permissions once a repository decision has been made.
How this taxonomy drives repository and metadata decisions in practice
Putting the axes together, a research administrator reviewing a DMP or advising a researcher on repository choice is really asking a small number of compounding questions for each dataset the project will produce:
- What form is it in—structured, unstructured, or semi-structured—and does that match what candidate repositories are built to ingest?
- How was it generated—observed, measured experimentally, simulated, or derived—and what does that imply for what needs to be preserved alongside the data itself (protocol, code, model parameters) for it to be usable by someone else?
- Is what will be shared the raw form, the processed form, or both, and is the pipeline between them documented?
- How sensitive is it, and does that rule out open deposit in favor of a controlled-access repository or a data sharing agreement?
Each answer narrows the field of appropriate repositories and metadata schemas. A repository evaluated for CoreTrustSeal certification publishes exactly this kind of scope information—what data types, formats, and sensitivity levels it is built to accept—which is why matching data type to repository scope early in DMP drafting avoids having to re-plan storage and access arrangements later in the data lifecycle, closer to the point of actual deposit.
Frequently asked questions
Is “raw data” the same as “primary data”?
Not exactly. Primary data usually refers to data a researcher collected first-hand for the specific study (as opposed to secondary data reused from another source), while raw data refers to the unprocessed state of a dataset regardless of who collected it. A dataset can be primary and processed, or secondary and effectively raw from the reusing researcher’s perspective.
Can one dataset belong to more than one category in this taxonomy?
Yes, and in practice almost every real dataset does. A single dataset is typically describable on all four independent axes at once—for example, a raw, quantitative, experimental dataset stored in a structured tabular format with open-access licensing. The axes are complementary lenses, not a single either/or classification.
Does the data type affect which metadata schema I should use?
Yes. Structured tabular data, geospatial data, software/code, and unstructured text or media are commonly described using different metadata schemas because each has different descriptive requirements. See CASRAI’s guide to choosing a metadata schema for a dataset for how to match schema to data type.
Where does synthetic data fit in this taxonomy?
Synthetic data is generated data, not observed or directly measured, which places it closest to the derived/simulation end of the generation-method axis—but it carries its own disclosure expectations distinct from ordinary derived data, since a reuser needs to know the data does not represent real observations. See CASRAI’s guide to synthetic data in research.
Do funder DMP templates ask researchers to classify their data using this taxonomy explicitly?
Funder templates rarely use this exact terminology, but they consistently ask questions this taxonomy is built to answer—what data will be generated, in what format, and how sensitive it is—which is why working through the classification before drafting a DMP tends to produce a more specific, defensible plan than describing "the data" as a single undifferentiated category. See CASRAI’s NIH vs. NSF DMP comparison and worked DMP examples for how this plays out in real plan language.
