Adopt non-redundant data structures, supported by identifiers and automated validation procedures (Control)
Why it matters
A core principle of robust data management is to minimize redundant storage of information to keep datasets internally consistent (Codd 1974). In literature reviews, this principle is often violated: the same entity (a paper, a decision, a status) is represented multiple times across files, tables, or derived artifacts. Once redundancy enters the workflow, contradictions become easy to create and hard to detect.
This is particularly problematic when status information or inclusion/exclusion decisions are stored redundantly (for example, in multiple tables or repeated across representations of the same paper). Contradictory values can accumulate silently—e.g., the same paper appears “included” in one place and “excluded” in another—because there is no single source of truth.
Practical implementation
Use non-redundant data structures: Represent each entity type once and link entities through identifiers:
- Paper-level entity: one canonical record per paper, keyed by stable identifiers (e.g., DOI, PMID, ISBN) and/or a repository-internal ID.
- Decision-level entity: screening/extraction/synthesis decisions stored once (as fields or separate linked tables), not duplicated across multiple exports.
- Derived artifacts: aggregated outputs (counts, profiles, PRISMA diagrams, dashboards) are generated from primary data rather than stored as editable, redundant files.
This makes “what the dataset says” unambiguous: there is one authoritative place to read and update each piece of information.
Rely on identifiers, not repeated values: Non-redundant structures work best when every record can be referenced reliably:
- Prefer globally meaningful identifiers (DOI/PMID) when available.
- Maintain stable internal IDs when external identifiers are missing.
- Track provenance through an origin field rather than duplicating records.