Share
73 Posts.
lightbulb Created with Sketch. 18
clock Created with Sketch.
23/06/17
10:18
Share
Originally posted by Southoz
↑
I would say you are one inch away from understanding my point Djwally.
Professor A is a legitimate scientist. The Leave-one out is a cross validation method is widely used in the development of prediction models. The aim of it is to try and control for the confirmation biases of model developers. But in and of itself it is not sufficient.
Even Wikipedia will tell you that people are pretty leery about exactly how much cross-validation techniques improve external validity. Hence newer methods - swap sampling - which I have never heard of – are attracting attention.
My argument is not which method is better – but rather there is a risk point that people aren’t seeing.
If prediction models were perfect with legitimate scientists and accepted cross validation methods the FDA wouldn’t require prospective studies. But they do ... for very good reasons.
And yes numbers in a sense you are exactly right. Matching is required. In other words a significant correlation to clinical diagnosis. This will get you approval. Further studies demonstrating sensitivity and specificity back to a reference standard often follow after this initial approval. Because it is only this that can demonstrate clinical utility / cost effectiveness which is necessary for commenrcial success.
As interesting as these discussions are unfortunately for me its back to the real world salt mine of research. Be interesting to see how what transpires here with RAP.
Expand
A few nervous comments surfacing from the commentariat.
I used to hate the presence of bots capping.
Now as I see them at work since 10:10 am, they give me re-assurance that the big boys are still in there.
Regards
Bluey