Literature searching: question to traceable sources

A literature search is not simply the collection of everything published on a topic. Searching is the work of narrowing the question being asked; where the question is not sharp enough, the number of results rises while their usefulness falls.

An answerable question first

Clinical questions are commonly structured into four components: the patient or population, the intervention under examination, the comparison, and the outcome measured. That structure reveals the search terms almost by itself. "Does this drug work" is not searchable; written as "in this population, compared with this alternative, on this outcome", it becomes so.

Choosing the sources

No single database covers the whole literature, and the differences in coverage have to be worked with knowingly:

• PubMed / MEDLINE — the principal index of biomedical literature, with structured searching through MeSH terms.
• Cochrane — systematic reviews and registers of controlled trials.
• Trial registries — which reveal studies that are ongoing or were never published.

The last of these matters for publication bias. Studies with unfavourable results tend to be published later or not at all, so a search confined to published literature can produce a systematically optimistic picture.

Recording the search

For a search to be reproducible, the search itself has to be documented. What should be recorded includes:

• The databases used and the date of access
• The full search string, the operators and any field restrictions
• Limits applied by language, year or publication type
• Inclusion and exclusion criteria
• The number of records removed at each stage, and why

Without this record, there is no way to explain why the same search returns a different set some months later. The PRISMA flow diagram is the structure most widely used to report that screening process visibly.

From screening to appraisal

Screening on titles and abstracts is fast but error-prone, so exclusion criteria are best kept narrow at that stage. At full-text appraisal the methodological quality of each study is assessed separately. In randomised studies, risk of bias is judged by examining the randomisation method, allocation concealment, blinding and the handling of missing data.

Certainty of evidence is a separate question

That a result is statistically significant does not mean confidence in it is high. The GRADE approach grades the certainty of evidence across study design, risk of bias, consistency, directness, precision and publication bias. The distinction makes visible that "was found significant" and "has been reliably demonstrated" are not the same statement.

A search as a living document

Literature does not stand still. Even after a search is complete, if the topic continues to be followed the search string should be retained and re-run at intervals. An update then becomes the continuation of a recorded piece of work rather than a new search built from scratch.