• Fujian Song, Khalid S Khan, Jacqueline Dinnes, and Alex J Sutton.
• Department of Public Health and Epidemiology, University of Birmingham, Edgbaston, UK. f.song@bham.ac.uk
• Int J Epidemiol. 2002 Feb 1;31(1):88-95.

BackgroundDespite the great possibility of publication bias in studies of diagnostic test research, empirical studies about publication bias have mainly focused on studies of treatment effect.MethodsA sample of 28 meta-analyses of diagnostic accuracy was selected from the Database of Abstracts of Reviews of Effectiveness (DARE). Methods used to deal with publication and related biases in these meta-analyses were examined. Asymmetry of funnel plot of estimated test accuracy against corresponding precision for each meta-analysis was assessed by three statistical methods: rank correlation method, regression analysis, and Trim and Fill method.ResultsIn reviews of diagnostic accuracy, there was a general lack of consideration of appropriate literature searching to minimize publication bias, and the impact of possible publication bias has not been systematically assessed. The results of the three different statistical methods consistently showed that in a large proportion of the 28 meta-analyses evaluated, the smaller studies were associated with a greater diagnostic accuracy. Exploratory analyses found that the fewer the literature databases searched, the greater the funnel plot asymmetry in meta-analyses. Funnel plot asymmetry tended to be greater in meta-analyses that included smaller number of primary studies. Our data revealed no consistent relationship between funnel plot asymmetry and language restriction in reviews.ConclusionsFurther research is required to explain why smaller studies tended to report greater test accuracy in a large proportion of meta-analyses of diagnostic tests. In systematic reviews of diagnostic studies, literature search should be sufficiently comprehensive and possible impact of publication bias should be assessed.

5 keywords...

hide…