From: Toufigh GordiSubject:[NMusers] Covariate effect Date: Fri, 07 Mar 2003 14:47:15 -0500 Dear all, I have a general question with regard to incorporation of covariate(s) into the model. If I understand it correctly, a covariate can partly explain the variability of a particular model parameter, meaning that upon insertion of the covariate effect the inter-individual variability normally decreases. A significant decrease in objective function value is also another indication that the presence of the covariate in the model has improved the general fit. What if the incorporation of a covariate results in a large decrease of the OFV but very small or no change in the ETA of the particular parameter? Am I correct to conclude that the decrease in OFV is probably the result of having an extra parameter rather than having the correct reason for the observed variability? Would a simpler model be as good as the more complete (with covariate effect built into it)? Put another way, which model is preffered: one without a covariate or one with a covariate with significantly lower OFV but no change in ETA? Toufigh Gordi _______________________________________________________ From: "Serge Guzy" Subject:RE: [NMusers] Covariate effect Date: Fri, 7 Mar 2003 13:58:23 -0800 Without entering into mathematical details, my impression is that an influential covariate should lead to a better objective function (ignoring the degree of freedom effect) and at the same time decrease the ETA. The use of a covariate separates to some extent the whole population into subset of populations, each of them characterized by different population means. Suppose the 1 compt model, bolus injection and V lognormal V=Vfixed*exp(Vrandom): Suppose Vfixed=exp(a+b*weight(i)) and the covariate weight is really significant(i for individual i). If I am coming back to each individual and plug their corresponding weight into the Vfixed term, I see that Vfixed will be different for every individual. If the covariate is influential, it will decrease the residual part for every individual(V random).Consequently, the eta will decrease. Serge Guzy President POP-PHARM(510 453 7443) Head of Pharmacometrics and Preclinical Statistics Xoma: tel(510) 2047476 _______________________________________________________ From: "Gobburu, Jogarao V" Subject:RE: [NMusers] Covariate effect Date:Fri, 7 Mar 2003 18:17:50 -0500 The case you describe is not unusual. I am assuming there is not much prior information about this covariate. If there is no drop in the unexplained variability, inspite of having a significant drop in OFV, the full model might not be of practical use (for example dose adjustment). But certainly it might be a signal to probe more into influential individuals/data points (Mats Karlsson performed some work on these lines). Another source of this happening could be that it is false-positive finding. Regards, Joga Gobburu Pharmacometrics CDER, FDA _______________________________________________________ From:"Kowalski, Ken" Subject: RE: [NMusers] Covariate effect Date:Mon, 10 Mar 2003 10:04:01 -0500 Toufigh, It is likely that you have substantial power to detect relatively small differences in the covariate effect. This occurs quite frequently with continuous covariates like body weight when you have a lot of subjects and a wide range of values. Since the change in OFV is the yardstick we use to determine statistical significance of the covariate effects, I wouldn't discard it and exclude the covariate effect from the model simply because it doesn't appear to be clinically relevant. I would use clinical relevance only to intrepet the parameters for the final model and their impact on dosing. I would not use a clinical relevance assessment to guide model building as this can be very subjective (based on one's assessment of what is clinically relevant for each covariate effect). Ken _______________________________________________________