Background: A typical meta-analysis of diagnostic test accuracy studies uses a single two by two table per study. A bivariate random-effects meta-analysis can synthesise these tables, utilising their binomial distribution within-studies. For tests on a continuous scale, each studyás two by two table corresponds to a threshold value dichotomising the test into 'positiveá and 'negativeá groups. Sometimes studies report multiple two by two tables, relating to multiple threshold values, that collectively follow a multinomial within-study distribution. However researchers often ignore this multinomial structure, and meta-analyse thresholds separately assuming a binomial dsitribution. \HI{Aims & Methods:} Hamza 2009 meta-analysis studies with multiple thresholds results and account for their multinomial structure; they focus on complete data (all studies provide all thresholds) and estimating ROC curves. In this poster we evaluate their approach given missing data (some studies provide an incomplete set of thresholds) and focus on sensitivity and specificity estimates. We apply the method to 30 studies of endovaginal ultrasound for detecting endometrial disease, with up to 7 different thresholds per study and missing data.

Results: We show that, compared to a separate meta-analysis of each threshold assuming a binomial distribution, the multinomial approach produces different summary estimates for each threshold (with sensitivity or specificity changed by up to 10%) and smaller standard errors (reduced by up to 45%). For example, for a threshold of 8mm the summary sensitivity was 85% (binomial) and 79% (multinomial), with standard errors of logit-sensitivity of 0.17 (binomial) and 0.09 (multinomial).

Conclusions: When studies provide test accuracy results for multiple thresholds, these thresholds should be jointly meta-analysed accounting for their multinomial structure; otherwise statistical precision is lost and clinical conclusions may be misleading.

Reference
1. Hamza TH, Arends LR, van Houwelingen HC, Stijnen T. Multivariate random effects meta-analysis of diagnostic tests with multiple thresholds. BMC Medical Research Methodology 2009;9:73