Data Extraction Errors in Meta-analyses of Diagnostic Tests
Mei Chung , Center for Clinical Evidence Synthesis, Tufts-NEMC
Joseph Lau, Tufts-New England Medical Center
Gowri Raman, Tufts-New England Medical Center
Christopher H Schmid, Tufts-New England Medical Center
*Thomas Trikalinos, Center for Clinical Evidence Synthesis, Tufts-NEMC
Keywords: meta-analysis, diagnostic test, error, misclassification, data quality
We scrutinized 378 papers included in 23 diagnostic test meta-analyses. The latter were randomly chosen from all identified in Medline through 2003. Numbers in the 2by2 tables of each study were extracted in duplicate and independently, and contrasted with those reported in the published meta-analyses. Leaving aside discrepancies due to rounding errors, we could not replicate the result for at least one of the included studies in 18 meta-analyses. Discrepancies were due to obvious errors (i.e., the meta-analysis transposed the 2 by 2 tables on all studies), inconsistent application of eligibility criteria (i.e., exclusion of a subpopulation in only some of the studies), pooling together different units of analysis (i.e., patients and body parts), or secondary to unclear reporting in the original papers. We discuss findings and the challenges we encountered.