Re: Covariate Models Using Weight

Date: Mon, 22 Nov 1999 23:55:10 +0100 From: Pierre Maitre <pmaitre@freesurf.ch> Subject: Re: Covariate Models Using Weight >Nick Holford wrote: >> >> James Gallo wrote: >>> >>> James Wright wrote: >>> >>>> "Piotrovskij, Vladimir [JanBe]" wrote: .... ....May I too? Dear respected scientists May a modest clinician add a word of philosophy to your religious debate? You have been discussing the way body weight influences the elimination clearance and I liked this hot discussion very much. It is indeed important to understand the relationship between weight and elimination. I am just asking myself weither it is appropriate to plug into a pharmacokinetic model a formula like Males CLcr = (143.5 - (1.095*Age) * Wt * 0.07 Females CLCr = (119.5 - (0.915*Age) * Wt * 0.07 or any other equation you want, without some evidence in your data to support your choice. My NONMEM mentor, Sam Vozeh, used to tell me, back in 1984, that nonlinear regression is an art, not a science. So please allow me tu use unscientific words like " I believe". I believe that clinicians need that we tell them the average pharmacokinetics of a drug, the size of the inter-individual variability, and the influence of important covariates (and among them, possibly, body weight). "Important" is the keyword here and is the hardest to define. The size of "Important" has first to be compared with the size of the interpatient variability for the parameter in question. Who cares about a small 8% change of clearance due to a particular covariate, if the remaining inter-individual variability is 40% for that parameter? ... forget this covariate, it won't help the clinician. For a clinician, "important" would mean changing the dosing by at least 20% (this is not science, again, but my experience as an anesthesiologist administering drugs every day). How do we sort out the "important" covariates from those who are no important? The P value doesn't help in this matter: as you all know, statistically significant does not mean important. My answer to this question might be: what you see in your data is worth modeling, what you don't see in your data is not worth modeling. Graphical methods based on the plots of etas vs covariates are a good start (Xpose can be used for this purpose) and allow one to pick up the "Important" covariates and to find a simple model that would fit the data. Dosing recommendations for the clinician must be kept simple in order to be safe. The above equations are certainly very exact, but they are just too complex to be used at bedside. And their complexity give to the clinician a false sense of scientific truth (that the clinician translates into "precise prediction") whereas the prediction of the concentration is very vague and unprecise, due to the cloud of interpatient variability. To conclude, my point would be: for your model to be useful, keep is simple. Pierre Maitre Genolier / Geneva