I suppose it really comes down to what you are going to do with the model. Many times I have checked the SAME assumption when modeling inter-occasional variability, and found that sometimes, removing it does indeed improve the fit significantly. In almost every case I've retained it (despite the better fit) for the exact reasons Leonid cites: it makes your model completely data-dependent. I suppose if the model was meant as a description or summary of the data, then it would not matter, but I make all of my models work for a living...
There is a related topic which I'd be interested in hearing from the group about. Many times, we take several Phase 1 studies and put them together in order to develop a population model early in development. I've learned through experience to be careful when doing this, because often, one or more studies will appear to have a different mean response for some parameter, e.g., CL or V2. Of course, you can introduce study as a covariate, but this intrduces the same problem as above; in a simulation context, which CL value is correct? There is a work-around for this (use both values) but this doubles the number of simulations you have to do, and from a scientific stand-point it is not very satisfying. What we need is another level of random effects at the STUDY level, similar to what is routinely done when performing hierarchical modeling in something like WinBUGS. I'd love to see this feature in a future version of NONMEM. "Leonid Gibiansky" <[EMAIL PROTECTED]> Sent by: [EMAIL PROTECTED] 17-Oct-2008 09:30 To "Nick Holford" <[EMAIL PROTECTED]> cc "nmusers" <nmusers@globomaxnm.com> Subject Re: [NMusers] More Levels of Random Effects Nick, This is exactly what I meant. If you have a model for English, Irish and Welsh, you may at least extrapolate it to Australians and New Zealanders (of British descent :) ). With occasion treated as non-ordered categorical covariate, you cannot extrapolate the model at all because time cannot be repeated, so your covariate (occasion) will have different value (level) at any future trial. Leonid -------------------------------------- Leonid Gibiansky, Ph.D. President, QuantPharm LLC web: www.quantpharm.com e-mail: LGibiansky at quantpharm.com tel: (301) 767 5566 Nick Holford wrote: > Leonid, > > I dont understand what you mean by "we lose predictive power of the > model: we do not know what will be > the variability on the next occasion.". > > Or are you concerned about the situation where you have say 3 occasions > and the IOV seems to be different on each occasion but you now want to > predict the IOV for a future study on the 4th occasion? > > I agree it is hard to extrapolate to future occasions but this seems to > be just like any other non-ordered categorical covariate - e.g. if we > see differences between English, Irish and Welsh what difference would > you expect for Russians? :-) > > Nick > > > Leonid Gibiansky wrote: >> Hi Xia, Nick >> Technically, one can use different variances on different occasions but >> then we loose predictive power of the model: we do not know what will be >> the variability on the next occasion. One can use occasion-dependent IOV >> variance to check for trends (for example, to investigate the time >> dependence of the IOV variability, or to check whether the first >> occasion (e.g., after the first dose of a long-term study) is for some >> reasons different from the others) but the final model should have some >> condition that specifies the relations of IOV variances at different >> occasion (SAME being the simplest, most reasonable and the most-often >> used option). >> >> Thanks >> Leonid >> >> -------------------------------------- >> Leonid Gibiansky, Ph.D. >> President, QuantPharm LLC >> web: www.quantpharm.com >> e-mail: LGibiansky at quantpharm.com >> tel: (301) 767 5566 >> >> >> >> >> Nick Holford wrote: >>> Xia, >>> >>> There is no requirement to use the SAME option. However, it is a >>> reasonable model for IOV that it has the same variability on each >>> occasion. >>> >>> If you dont use the SAME option then you just need to estimate an >>> extra OMEGA parameter for each occasion you dont use SAME. You can >>> test if the SAME assumption is supported by your data or not by >>> comparing models with and without SAME. >>> >>> Nick >>> >>> PS Your computer clock seems to be more than 2 years out of date. >>> Your email claimed it was sent in 17 Jan 2006. >>> >>> Xia Li wrote: >>>> Dear All, >>>> Do we have to assume the variability between all occasions are the >>>> same when >>>> we estimate IOV? What will happen if I don't use the 'same' >>>> constrain in the >>>> $OMEGA BLOCK statement? Any input will be appreciated. >>>> >>>> Best, >>>> >>>> Xia Li >>>> >>>> -----Original Message----- >>>> From: [EMAIL PROTECTED] >>>> [mailto:[EMAIL PROTECTED] On >>>> Behalf Of Johan Wallin >>>> Sent: Wednesday, October 15, 2008 9:17 AM >>>> To: nmusers@globomaxnm.com >>>> Subject: RE: [NMusers] More Levels of Random Effects >>>> >>>> Bill, >>>> Is it really an eta you want, or is this rather solved by different >>>> error >>>> models for the different machines? >>>> >>>> If still want etas, one way would be to model in the same way as >>>> IOV. In the >>>> case of intermachine-variability you would have to assume the >>>> variability >>>> between all machines are the same... Or would you rather assume >>>> interindividual variability is different with >>>> different machine, and you then would want one eta for TH(X) for every >>>> machine...? It depends on what you mean by different slope every day, >>>> regarding on what your experiments like, but calibration differences >>>> should >>>> perhaps be taken care of by looking into your error model, eta on >>>> epsilon >>>> for starters... >>>> >>>> Without knowing your structure of data, a short example of IOV-like >>>> variability would be: >>>> >>>> MA1=0 >>>> MA2=0 >>>> IF(MACH=1)MA1=1 >>>> IF(MACH=2)MA2=1 >>>> ;Intermachine variability: >>>> ETAM = MA1*ETA(Y)+MA2*ETA(Z) >>>> >>>> PAR= TH(X) *EXP(ETA(X)+ETAM) >>>> >>>> $OMEGA value1 >>>> $OMEGA BLOCK(1) value2 >>>> $OMEGA BLOCK(1) same >>>> >>>> /Johan >>>> >>>> >>>> _________________________________________ >>>> Johan Wallin, M.Sci./Ph.D.-student >>>> Pharmacometrics Group >>>> Div. of Pharmacokinetics and Drug therapy >>>> Uppsala University >>>> _________________________________________ >>>> >>>> >>>> -----Original Message----- >>>> From: [EMAIL PROTECTED] >>>> [mailto:[EMAIL PROTECTED] On >>>> Behalf Of Denney, William S. >>>> Sent: den 15 oktober 2008 14:39 >>>> To: nmusers@globomaxnm.com >>>> Subject: [NMusers] More Levels of Random Effects >>>> >>>> Hello, >>>> >>>> I'm trying to build a model where I need to have ETAs generated on >>>> separately for the ID and another variable (MACH). What I have is a PD >>>> experiment that was run on several different machines (MACH). Each >>>> machine appears to have a different slope per day and a different >>>> calibration. I still need to keep some ETAs on the ID column, so I >>>> can't just assign MACH=ID. >>>> >>>> I've heard that there are ways to do similar to this, but I have been >>>> unable to find examples of how to set etas to key off of different >>>> columns. >>>> >>>> Thanks, >>>> >>>> Bill >>>> Notice: This e-mail message, together with any attachments, contains >>>> information of Merck & Co., Inc. (One Merck Drive, Whitehouse Station, >>>> New Jersey, USA 08889), and/or its affiliates (which may be known >>>> outside the United States as Merck Frosst, Merck Sharp & Dohme or >>>> MSD and in Japan, as Banyu - direct contact information for >>>> affiliates is >>>> available at http://www.merck.com/contact/contacts.html) that may be >>>> confidential, proprietary copyrighted and/or legally privileged. It is >>>> intended solely for the use of the individual or entity named on this >>>> message. If you are not the intended recipient, and have received this >>>> message in error, please notify us immediately by reply e-mail and >>>> then delete it from your system. >>>> >>>> >>>> >>> >> >
