Title: On summary models for meta-analysis of diagnostic studies
Authors: ShengLi Tzeng - National Sun Yat-sen University (Taiwan) [presenting]
Abstract: Summarizing performance metrics is essential to a systematic review of a diagnostic performance. When a gold standard is available, every individual study in a meta-analysis has merely four numbers from a dichotomized test, i.e., number of true positives, false negatives, true negatives, and false positives. The goal of such a meta-analysis is to produce the summarized sensitivity, specificity, and a summary line of the receiver operating characteristic (ROC) curve. There are various summary models for the performance metrics in the literature, and hence one inevitably faces the problem of which model(s) to use. However, the suitability of existing asymptotic model selection approaches is doubtful since the meta-analysis here typically has a rather small sample size. Directly applying information criteria may fails to select or to combine good models for meta-analyses. Novel simulation scenarios mimicking a typical data collection process are conducted. The simulation avoids generating data from a certain model that would be biased towards specific assumptions. Even though the data never follow any probabilistic mechanism of a candidate model, we do know the underlying ROC curve. Then several model determination methods were compared accordingly, including simply using the most popular one, model averaging, and criterion-based model selections. Some suggestion and discussion about the better model determination strategy will be given based on the simulation results.