Bivariate meta-analysis of sensitivity and specificity of radiographers' plain radiograph reporting in clinical practice.

Abstract

Studies of diagnostic accuracy often report paired tests for sensitivity and specificity that can be pooled separately to produce summary estimates in a meta-analysis. This was done recently for a systematic review of radiographers' reporting accuracy of plain radiographs. The problem with pooling sensitivities and specificities separately is that it does not acknowledge any possible (negative) correlation between these two measures. A possible cause of this negative correlation is that different thresholds are used in studies to define abnormal and normal radiographs because of implicit variations in thresholds that occur when radiographers' report plain radiographs. A method that allows for the correlation that can exist between pairs of sensitivity and specificity within a study using a random effects approach is the bivariate model. When estimates of accuracy as a fixed-effects model were pooled separately, radiographers' reported plain radiographs in clinical practice at 93% (95% confidence interval (CI) 92-93%) sensitivity and 98% (95% CI 98-98%) specificity. The bivariate model produced the same summary estimates of sensitivity and specificity but with wider confidence intervals (93% (95% CI 91-95%) and 98% (95% CI 96-98%), respectively) that take into account the heterogeneity beyond chance between studies. This method also allowed us to calculate a 95% confidence ellipse around the mean values of sensitivity and specificity and a 95% prediction ellipse for individual values of sensitivity and specificity. The bivariate model is an improvement on pooling sensitivity and specificity separately when there is a threshold effect, and it is the preferred method of choice.