covariate models

From ltc_Ralf_Brueckner@wrsmtp-ccmail.army.mil Wed Jun 11 09:16:28 1997 Subject: covariate models Dear fellow NONMEM users, I have seen covariate models expressed as: a) V = THETA(1)+THETA(2)*WT and as b) V = THETA(1)+THETA(2)*(WT-70) so called "centering the covariates at their median". Can someone explain the rationale for the second method. Method a seems intuitively more useful as a clinician. Thank you Ralf Brueckner, MD Dept. of Clinical Pharmacology Div. of Experimental Therapeutics Walter Reed Army Institute of Research

From Rene_Braeckman@cc.chiron.com Wed Jun 11 11:05:34 1997 Subject: Re: covariate models The two methods differ in the interpretation of THETA(1) and THETA(2), as follows: a) THETA(1): part of V independent of WT THETA(2): change of V per unit WT (WT-dependent part) b) THETA(1): part of V for 70-kg WT THETA(2): change of V per unit of WT different from the standard 70-kg WT Method a) implies that there is a value for V equal to THETA(1) at a WT of zero, which is theoretically not possible. Method b) gives a direct estimate for V (THETA(1)) for a standard 70-kg person. Rene Braeckman, PhD Senior Director of Pharmacokinetics & Pharmacodynamics Chiron Corp., Emeryville, CA

From harrisof@capaccess.org Wed Jun 11 23:29:46 1997 Subject: Re: covariate models Thank you for the explanation. For weight and height, mightn't it make even more sense to split the model by gender?

From PRADHANR@cder.fda.gov Thu Jun 12 05:14:19 1997 Subject: Re: covariate models Ralf, In my experience, I have compared the following situation: 1.TVV1 = THETA(1) * (WT)**THETA(2) VS. 2.TVV1 = THETA(1) * (WT/70)**THETA(2) In a covariate rich model, #2 computationally performed better. In other words OFV minimization reached much sooner. Any comments would be appreciated. Thank you Raj Pradhan, Ph.D. OCPB, FDA

From FERDINAND.ROMBOUT.FR@bayer-ag.de Thu Jun 12 07:01:33 1997 Subject: Antwort: RE: COVARIATE MODELS It is all a matter of sensitivity. In example A one tries to estimate a slope and intercept. But since one, as has been mentioned, has no weight lets say smaller than 30, a small variation in slope can cause a big variation in intercept, therefore making the model more difficult to predict and sometimes instable. The result will be often bigger errors in the estimated parameters compared with method B were the intercept is more or less in the middle of the weight data, making a stable and more precise prediction of slope and intercept possible. This means we should not center around 70, but should center about the average of our covariate data in the dataset, ie average weight, height, albumin etc. F. Rombout Clinical Pharmacokinetics Department of clinical pharmacology Bayer AG Wuppertal, Germany

From ltc_Ralf_Brueckner@wrsmtp-ccmail.army.mil Thu Jun 12 08:23:23 1997 Subject: Covariate Models Thanks to all who have replied to my inquiry re: different ways of modeling covariates. In this inquiry, I used V as an exqample, but the question is equally applicable to any other parameter one wishes to evaluate. With respect to method 1: V = THETA(1)+THETA(2)*WT If THETA(1) drops out, the models yields a simple L/Kg relationship, something I was ingrained with as a pediatrician. The argument that when WT approaches 0, a Volume (THETA(1)) exists, which is theoretically impossible, is true. However, I have also seen NONMEM results presented using the second method: V = THETA(1)+THETA(2)*(WT-70) where THETA(1) and THETA(2) have values such that a 40 kg individual is predicted to have a negative volume (e.g. when THETA(1) = 10 L, THETA(2)=0.5 L/kg, WT=40 kg, V = -5 L. These are both really just examples of extrapolating the model to individuals not used in the model determination. It seems so far, that this is a question of interpretation, and one's frame of reference (i.e. a "typical" adult and deviation from this, versus a more straightforward - to my way of thinking - relationship between V and WT). I was actually expecting some statistical answer, or something more along the lines of what Raj Pradhan alluded to in terms of computational efficiency: Further comments on this phenomenon would also be appreciated. Ralf Brueckner, MD Dept. of Clinical Pharmacology Div. of Experimental Therapeutics Walter Reed Army Institute of Research

From Rene_Braeckman@cc.chiron.com Thu Jun 12 09:00:44 1997 Subject: Re[2]: covariate models Height, weight, body surface area, gender, etc are all demographic covariates that might be tested whether they influence a particular PK parameter. Some of these covariates, if not all, can be interrelated, i.e., are not independent from each other or are correlated among each other. If weight and gender are both statistically significant covariates on V, you can indeed calculate typical values of V for a standard 70-kg male and a standard 70-kg female. Rene Braeckman Dept. of PK/PD, Chiron Corporation

From VPIOTROV@janbelc1.ssw.jnj.com Fri Jun 13 00:44:01 1997 Subject: Gender as a covariate Any body size parameter is gender-dependent, therefore, if a typical value of CL is dependent on WT, it will be automatically gender-dependent. However, to my experience, you will hardly find any `pure' gender effect on CL: if you plot Bayesian estimates of CL vs. WT and use different symbols for females and males, all points will lie along the same line (points corresponding to females will only be shifted towards smaller weigth). Moreover, including in the model gender as a separate covariate affecting CL most probably will not improve the fit. The situation with V is different. You may easily find V weight- AND gender-dependent. I.e., on the V vs. WT plot you may see two clusters: points corresponding to females will lie above those corresponding to males. Accordingly, having separate THETAs for males and females in the model will substantially improve the fit. An explanation is quite simple: if the drug is lypophilic (most drugs are) it tends to distribute in adipose tissues, and fat contents in females is known to be higher (on average) than in males. Vladimir Piotrovsky, Ph.D. Janssen Research Foundation Clinical Pharmacokinetics vpiotrov@janbelc1.ssw.jnj.com B-2340 Beerse Belgium Fax: +32-14-605834 Email: vpiotrov@janbe.jnj.com

From rs@chdr.leidenuniv.nl Fri Jun 13 00:46:37 1997 Subject: Re: Covariate Models Ralf, In your case, methods 1 and 2 will make no difference from a statistical perspective, > > method 1: V = THETA(1)+THETA(2)*WT > > method 2: V = THETA(1)+THETA(2)*(WT-70) > because subtraction of 70 merely results in a linear transformation. The only ('statistical') reason for doing this is computational, i.e. avoiding numbers that become too large. This is comparable to the reason behing the rescaling of parameters that NONMEM does prior to estimation. In the case of weights there is probably no reason to go to the trouble. In Raj's example, numbers becoming too large may certainly be a problem in his first method, which his second method solves. > > method1: TVV1 = THETA(1) * (WT)**THETA(2) > > method2: TVV1 = THETA(1) * (WT/70)**THETA(2) > Naturally, the values of theta(2) change, resulting in a different interpretation (or requiring rescaling back to the original situation). Regards, Rik Schoemaker

From rs@chdr.leidenuniv.nl Fri Jun 13 10:33:12 1997 Subject: Re: Re[2]: Covariate Models Ken, You're right about theta(1) and theta(2) although I must admit I hadn't realized it! Thanks, Rik Schoemaker

From harrisof@capaccess.org Fri Jun 13 21:09:01 1997 Subject: interpretable models I really didn't mean adding another fitted parameter to the model. In typical clinical trials, especially the Phase III and IV variety I see more of myself, the gender ratio is not particularly related to the population ratio. Hence V=THETA(1)+THETA(2)*(WT-70) might be rather strange if 80% of the patients were of one gender, which is quite possible. My suggestion is that, where FLAG=1 for males and 0 for females, V=THETA(1)+THETA(2)*(WT-60-20*FLAG) where the constants "60" and "20" are chosen so 60+20*FLAG will yield the population means for each gender. In practice, clearance and volume of distribution can depend on gender in more complex and drug dependent ways, and whatever we know ought to be put to good use. I'm a statistician myself, and agree that estimation based on the data in hand is often best, but I'm always interested to learn what my medical colleagues regard as more readily interpretable. Elegant statistical results aren't worth much in my workplace if only statisticians can understand them.

From ETTEE@cder.fda.gov Sat Jun 14 12:43:42 1997 Subject: Re: Gender as a covariate I think that there is an over generalization on the issue of gender as covariate. It is worth noting that there may still be a gender effect on clearance (CL) after adjustment has been made for weight. This has been observed with some drugs. Omitting a gender effect (when present) because adjustment for weight has been made may have serious consequences. With data visualization, is a useful tool for informative population modeling, this error can be avoided. Ene Ette, Ph.D. Pharmacometrics Staff Office of Clinical Pharmacology & Biopharmaceutics FDA

From: KGKOWA@monsanto.com Fri, 13 Jun 1997 08:29:45 -0500 Subject: Re[2]: Covariate Models Rik, I agree. I was trying to say the same thing but you've said it more elegantly. One may prefer one parameterization over another for 'computational' (numerical) not 'statistical' reasons as they are essentially fitting the same model. In Raj's example, I believe you meant that the value of THETA(1) is different between the two parameterizations (method 1 vs method 2) not THETA(2). Specifically, THETA(1) for the first parameterization corresponds to THETA(1) * (1/70)**THETA(2) for the second parameterization. The value (and interpretation) of THETA(2) does not change between these two parameterizations. Ken Kowalski G.D. Searle