Scientific Journal Article Databases: Choose by Evidence Job

Answer six questions before opening a search box.

In our editorial work, database choice fails when one interface silently becomes the discovery tool, eligibility rule, citation counter, and quality screen. We created this card to separate those jobs and preserve the handoff into claim verification. It does not rank providers universally; it tests whether the collection and controls can answer the declared research question.

1. Which disciplines must be covered?
2. Which document types count?
3. Is controlled vocabulary required?
4. Is citation-network analysis part of the method?
5. Must the complete search be reproducible?
6. Do you need records, lawful full text, or both?

Do not select a database because it returns the largest number. Recall without relevance creates screening cost; precision without coverage creates missed evidence.

Define concept blocks in plain language before translating syntax. For each database, record:

• platform and database name;
• coverage dates used;
• exact query;
• controlled terms and free-text synonyms;
• fields, limits, and document types;
• search date and result count;
• deduplication key and software;
• updates or alerts planned.

One master query pasted everywhere is not reproducible. Quotation, proximity, wildcards, subject headings, and field tags differ. Preserve the concepts while translating the syntax.

Test sensitivity with known relevant articles. If a database cannot retrieve them, inspect spelling, indexing, field restrictions, date limits, and vocabulary mapping before adding more terms.

Citation networks help find earlier foundations, later replications, critiques, and adjacent methods. Counts are database-dependent and time-dependent. A highly cited article can still be methodologically weak or irrelevant to the precise claim.

For every citation-led discovery, return to the article:

1. Verify title, DOI, journal, year, and version.
2. Read the methods and result supporting the proposed manuscript sentence.
3. Record population, design, estimate, uncertainty, and limitation.
4. Check corrections, expressions of concern, and retractions.
5. Classify support as direct, qualifying, conflicting, or background only.

This converts a citation graph into an evidence trail.

Product pages often describe millions of records, sources, or citations. Those numbers are not directly comparable. One service may count works, another source titles, another linked references, and another versions.

When coverage affects a method, record the provider's unit, snapshot date, included databases, language or document constraints, and source-list link. Avoid “comprehensive” unless the protocol defines what comprehensive means for the question.

For systematic work, disclose inaccessible databases and search limitations. Convenience is not a coverage argument.

A defensible choice records why each source is present before the search begins. Create one row per database and complete the fields below. The worksheet is useful for a single manuscript search, and it becomes essential when another researcher must reproduce or update the work.

Use the worksheet to assign complementary roles. PubMed may own biomedical subject retrieval, Crossref may reconcile DOI metadata, and OpenAlex may extend citation-network discovery. Overlap is not waste when each source has an explicit purpose. Unexplained overlap, by contrast, produces screening cost without a clear coverage benefit.

After a pilot search, add two observations: which known relevant records were missed, and which irrelevant record type dominated the results. Those observations supply a concrete reason to revise vocabulary, filters, or database coverage before the final run.

Deduplication should merge identities, not erase search provenance. Retain the database names, record identifiers, queries, and retrieval dates attached to every merged work. A DOI is the strongest common key when present, but DOI-only matching misses older records and can merge corrections or related objects incorrectly.

Use a staged match:

1. match exact DOI after normalizing case and URL prefixes;
2. match stable database identifiers such as PMID where appropriate;
3. compare normalized title, first author, year, journal, and pages or article number;
4. inspect corrections, errata, protocols, conference abstracts, and ahead-of-print versions manually;
5. preserve all source-database tags on the retained record.

Report the counts at each step: records retrieved per database, records before deduplication, duplicates removed, and unique records screened. That record lets a reader distinguish broad retrieval from duplicate inflation and reproduce the selection flow later.

The largest-number winner. A record total is treated as universal coverage. Repair it with discipline, document-type, date, and source-list checks.

The platform-database blur. A platform name is recorded without the specific databases searched. Repair the methods record with exact collections.

The unchanged-query pattern. Syntax is moved between platforms without translation. Repair it with a concept-to-syntax ledger for each database.

The indexed-equals-valid assumption. Inclusion is treated as quality evidence. Repair it with article-level appraisal and correction checks.

The citation-count verdict. A high count is treated as truth or relevance. Repair it by reading the claim-level evidence.

The metadata-full-text collapse. A record is assumed to contain the paper. Repair it by documenting the lawful version actually read.

In Manusights reviews, database work becomes useful when another reviewer can reconstruct why a source was retained and what manuscript sentence it supports.

Use a claim handoff card with database, query, search date, DOI, article version, study design, relevant result, important limitation, manuscript claim, and evidence role. If any of these is unknown, mark it unknown rather than filling the gap from the title or abstract.

The card also makes disagreement visible. Two papers may support different populations or outcomes; a synthesis should preserve that difference rather than vote by citation count.

Review source-to-claim continuity once the evidence table is attached to the draft.