Using the example of a diagnostic prediction model using D-Dimer data of patients suspected of having deep venous thrombosis, they emphasize that differences in the distribution of predictor values across different studies need to consider both: a) true variation in the characteristics of patients (severity, clinical settings, geographical locations, or time); and b) also the ways that the predictors themselves (e.g., the D-Dimer) were measured. Non randomised studies are beginning to be accepted for demonstrating comparative clinical efficacy by approval agencies, perhaps in part in response to the real world evidence (RWE) movement (and the large recent investments in this by pharma) defined by FDA as the “the clinical evidence regarding the usage and potential benefits or risks of a medical product derived from analysis of real world data from observational or longitudinal patient studies”. Turning to cohort studies, Binder et al. provide evidence of the widespread occurrence in top journals of likely bias in cohort studies due to the bias that Binder described in this journal in 2014 3 - For example this occurs when the patient dies before the end of the cohort time period due to the risk factor of interest but the patient is then censored in the analysis. 58 of 125 eligible studies were at risk of this bias, but only 6 addressed this with appropriate sensitivity analyses. The authors reviewed the literature and then worked with a patient panel to develop questionnaire items based on literature review, tested in the target group using a think-aloud procedure, and validated by a cross-sectional study among patients receiving health care at the interface between general practice and hospital care in two regions in the Netherlands.