Following PROBAST (signaling question 1.1, about “appropriate data sources”), a non-nested case-control design would pose a high risk of bias for the development and/or validation of diagnostic models (in short, due to spectrum bias). The Circulating Cell-free Genome Atlas Study (CCGA) follows exactly that design and developed/validated an omics-based multi-cancer screening tool. Without addressing the merit of multi-cancer screening, I wonder if the fact that their study included over 40 centers in US and Canada impacts this risk of bias due to the data source. My understanding is that it makes things worse: big data of inadequate quality. Though it seems some people would feel that its multicentric nature, along with careful matching, would make it more likely to capture the target distribution that would be expected in a prospective cross-sectional design (except for baseline prevalence).
In short, my question is: can non-nested case-control designs ever be considered low risk of bias for developing and validating diagnostic models?
Thank you very much!
Best,
Giuliano