Re: Covariate Models Using Weight

Date: Fri, 19 Nov 1999 12:23:36 +0000 From: James <J.G.Wright@ncl.ac.uk> Subject: Re: Covariate Models Using Weight Dear Nick, The point where you and I differ, I think, is in what we consider relevant evidence. I don't consider between species scaling to have any relevance to the use of weight in a human population. Nor am I impressed by elaborate fractal models. As you have provided only two references that I consider relevant, (and you are an author on both of them), I remain totally underwhelmed. Steve raises several interesting issues, and the truth is that unless you have a truly massive sample there is always a certain amount of subjectivity in how you construct a covariate model (A Miller, Subset Selection in Regression, is a very scary book). If you require predictive accuracy, then you should be very wary of overfitting, so I guess I am not that keen on including covariates that are not justified by improvements in fit. Steve's example of creatinine clearance and the Cockcroft & Gault formula is an interesting one - the Cockcroft & Gault formula contains someone else's prior knowledge, but if your population isn't the same as theirs (or you are using a modern creatinine assay) maybe this will do more harm than good. Personally, I would use the raw covariate (and weight etc separately (if justified), although this will use more parameters) in this particular case. My view is we should use prior information, but evaluate its relevance critically. James