You have a biomarker, or a questionnaire, or a set of criteria, or any other tool by which you want to be able to diagnose persons with a specific disease, or to predict a disease in the future. The idea is that such a tool should be able to detect the presence or absence of a particular process: it should classify, with none or little error, subjects into healthy or sick. As you can imagine, this is very complex by the single reasoning that health is not black and white, but more like a spectrum. In addition, you need to test your diagnostic tool by determining the degree of similarity between the results of your tool and an external criterion. This criterion, in principle, says the truth about the condition of the patients: disease yes/disease no. Although everything seems very easy in diagnostic tests reasoning, as we make everything to fall into a 2x2 table, it is not that simple, and biases appear at every corner. There is also a lot of circularity, and the literature is full – yes, even in our field, and even in our best journals – of examples of wrong designs and interpretations.

Disclosure of Interest None declared

DOI 10.1136/annrheumdis-2014-eular.6172