Re: Simultaneous PK and Cell Life Span PD modeling

Toufigh, Mark Sale said: > When I first started this some very smart people told me that you should > put an ETA on everything, because, biologically, there is variability on > everything. This is still correct. There is variability in all biological parameters. Many of them are biologically connected and thus apparently random differences are statistically correlated. So it makes sense to estimate the variance-covariance matrix for the parameter random effects (OMEGA with full BLOCK() in NONMEM terms). The problem with trying to estimate the variability is caused by 1) sub-optimal designs 2) approximate estimation methods. The first can be improved by using an optimal design program (see France Mentre's excellent review at PAGE 2007 http://www.page-meeting.org/?abstract=1179 ). The second can be improved by using a better estimation method (see see France Mentre's excellent review at PAGE 2005 http://www.page-meeting.org/default.asp?id=26&keuze=abstract-view&goto=abstracts&orderby=author&abstract_id=833 } So if you have been given data from a typical sub-optimal design and are using NONMEM (none of the better algorithm implementations are as flexible (except possibly S-Adapt MCPEM http://www.page-meeting.org/?abstract=1111 ) ) then you have to be pragmatic and let the data tell you which parameters have distinguishable between subject variability. NONMEM gives some indirect help with this by issuing a warning about parameters being at a constraint boundary. If you haven't turned off this semi-helpful feature and your THETAs are well clear of their boundaries then you should check which OMEGAs are close to zero and consider fixing them to zero. Nick Mark Sale - Next Level Solutions wrote: > > Toufigh, > When I first started this some very smart people told me that you should > put an ETA on everything, because, > biologically, there is variability on everything. Since, then, I've decided > that this isn't a feasible approach. The > same argument would be used to justify physiologically based pk models - even > though they have two many > parameters to be feasible. At the other end, if you talk to some very > traditional statistician, they are entirely > "data driven", don't put anything in that the data don't tell you to, i.e., > ignore your knowledge of biology entirely. > So, we all decide where we are on this empirical/theoretical spectrum. > Personally, I like to think I'm somewhere > in the middle - biologically based, but really need the data to tell me what > should go into a model. The good news, > I think, is that Bayesian statistics lets us do this. Could be done > formally, but you'd have to specify, up front, > how confident yo! > u were that each feature is in the model. If you have lots of reason to > believe that volume is a > linear function of weight, then (IMHO) you are justified putting it in even > if the current data in hand don't really > tell you that. On the other hand, I'd be hesitant to put clearance as a > function of astrological sign, even if the > current data support that (unless the support is very, very strong) - because > my prior is so low. People more > knowledgeable about Bayesian methods could put this in a more formal > structure. > But, to your question, my answer is NO, you might start with a prior that > there is an ETA on any given > parameter, but if the current data don't support it (by which I mean either > in terms of log-likelihood test [I know, > you can't formally test whether and ETA is significant] or failure to > converge) I don't hesitate to remove it. I don't > see a reason to treat variance parameters any different from structural model > parameters or co! > variate parameters > in this regard. > Mark > > Mark Sale MD > Next Level Solutions, LLC > www.NextLevelSolns.com >