Evaluating the proportion of retrieved citations and relevant studies when database selection is restricted to MEDLINE, CENTRAL and Embase: Part 4 of a meta-research study of rapid review methodological trade-offs.

OBJECTIVE: Rapid reviews streamline systematic review methods to provide evidence for decision-makers more efficiently. A common methodological trade-off is the restriction of database selection. We aimed to determine the proportion of citations retrieved and recall when restricting database selection in three systematic reviews of interventions for the management of chronic primary low back pain (LBP). We also describe where missed studies were identified, duplication across databases and incremental recall of each database. DESIGN AND SETTING: We used original search files from three systematic reviews of randomized controlled trials for interventions (acupuncture, education, and TENS) conducted to inform the WHO guideline on non-surgical management of chronic primary LBP. Databases searched were CENTRAL(C), MEDLINE(M), Embase(E), CINAHL, and ICTRP. The acupuncture review also included CNKI and WanFang, and the education and TENS reviews included PEDro. Following Cochrane Rapid Review Methods Group recommendations, the restricted search included CENTRAL, MEDLINE and Embase. We report the relative yield, proportion of citations retrieved compared to the full search, and relative recall, proportion of relevant studies in the restricted search compared to the full search. We calculated incremental recall of each database (unique studies from each database) and cumulative number needed to read (NNR) (number of citations/number of relevant studies). RESULTS: The relative yield ranged from 54.7% to 86.0. Relative recall ranged from 65.4% to 100.0%, missing up to 34.6% of relevant studies. All studies missed in the acupuncture review were identified in Chinese databases and the study missed in the TENS review was identified in PEDro. Incremental recall ranged from 0.0% to 20.0% for CENTRAL, 7.7% to 20.0% for MEDLINE and 0.0% for Embase. NNR ranged from 11.0 to 158.7 for CENTRAL, 31.3 to 172.8 for MEDLINE and 14.2 to 68.5 for Embase. CONCLUSIONS: Restricting database selection reduces the workload while retrieving most studies. In reviews of chronic LBP, the selection of two biomedical databases is sufficient to identify all relevant studies from a restricted search. However, there is a risk of missing relevant studies when topic-specific databases are key contributors. Excluding such databases may introduce selection bias, therefore inclusion of topic-specific databases should be determined on a case-by-case basis.