Synthetic data collections
Creating, documenting, and sharing synthetic data
Synthetic data are mostly artificially generated datasets created to replicate some features of real-world data without disclosing actual individual records. The aim is often to enable preliminary analysis, method development, software testing, training, or exploration of data structures when access to the real underlying data is restricted. In some cases, synthetic datasets may also support wider discovery and documentation of a data collection that would otherwise remain inaccessible.
Although synthetic data can help address practical challenges, such as facilitating code testing without exposing sensitive records, they are not a substitute for real data. They do not inherit all analytical properties of the sources from which they are derived, and they do not automatically eliminate legal, ethical or governance obligations associated with those sources.
Synthetic data vary in fidelity. Low-fidelity synthetic data may preserve only broad distributions or structural features. Higher-fidelity synthetic data aim to reproduce more nuanced relationships between variables.
Several considerations apply when synthetic data are created, documented and shared.
Synthetic data are most useful when they support specific analytical or operational goals. For example:
- Enabling researchers to familiarise themselves with data structure and coding without needing secure access to real data.
- Permitting method development or debugging of code.
- Supporting training and teaching in contexts where real data are sensitive.
- Improving discoverability of a restricted dataset by revealing structure and content in a non-sensitive form.
However, synthetic datasets should always be accompanied by documentation that explains their purpose and limitations. Users should understand what the data represent and what they do not represent. This is essential because synthetic data can distort relationships or omit features that matter for particular analyses.
High-quality documentation is essential for synthetic data to be interpretable and usable. Users should be able to assess:
- why the dataset was generated
- how the synthetic data were produced
- what assumptions or models underlie the generation process
- what limitations and caveats apply to the data.
Synthetic datasets should be clearly and unambiguously labelled as synthetic. Documentation should avoid implying that synthetic data are equivalent to real data.
The Minimal Documentation Standard for Synthetic Data Collections provides a robust foundation for documenting synthetic datasets. It outlines the essential elements required to support transparency and reuse, including explanations of generation methods, intended use, and limitations.
High-quality documentation is essential for synthetic data to be interpretable and usable. Users should be able to assess:
- why the dataset was generated
- how the synthetic data were produced
- what assumptions or models underlie the generation process
- what limitations and caveats apply to the data.
Synthetic datasets should be clearly and unambiguously labelled as synthetic. Documentation should avoid implying that synthetic data are equivalent to real data.
The Minimal Documentation Standard for Synthetic Data Collections provides a robust foundation for documenting synthetic datasets. It outlines the essential elements required to support transparency and reuse, including explanations of generation methods, intended use, and limitations.
Synthetic data are not inherently “safe” simply because they are not real. Their generation, documentation and use must be governed by the same principles of accountability, transparency and clarity that apply to any research data. This includes:
- Being explicit about the intended analytical uses and limitations.
- Avoiding over-interpretation of patterns not supported by the generation model.
- Ensuring users understand that synthetic data may not replicate complex relationships present in real data.
Synthetic data can play a valuable role when used appropriately and documented transparently. They can improve access, reduce barriers and support methodological work. But they do not replace the responsibilities that come with generating, managing and sharing real research data.