NONMEM

From: "Ritu Karwal" Ritu.Karwal@ranbaxy.com Subject: [NMusers] NONMEM Date: Thu, 14 Sep 2006 09:48:21 +0530 Dear nmusers: Can anybody suggest me how to perform optimization of sampling time points in NONMEM or using any other software? I have read somewhere that there is a command $SUPER in NONMEM to perform this. Also, I have got a control stream (which I have got from nmusers previous messages only posted by Ludger Banken) for this but that is not working. Below is that control stream. Also can anybody tell me what RATD, COVA, EVID, SS stands for in this control stream. I have recently started working on population pharmacokinetics and NONMEM so I dont know these words? With regards, Ritu Karwal, M.Sc (Statistics) Research Associate, Metabolism and Pharmacokinetics, Ranbaxy Research Laboratories Limited. $PROB Popkin Dummy Simulation for first starting value of the seed ;This problem is only here to allow the use of the seed (-1) later on $THETA (.1,1.4 ,10) ;Ka 1 $THETA (.1,3.5 ,50) ;CL 2 $THETA (1,20 ,1000) ; V 3 $THETA (1 ,1 ,1 ) ;F 4 $THETA (-2,0.2 ,2) ; CL (Covar) 5 $OMEGA 0.223144 ;kA $OMEGA 0.039221 0.039221 ;CL V $OMEGA .0001 FIXED;F $SIGMA 0.039221 ;ERR Mult $SIGMA 0.005 ;Err Add $SIGMA 1 FIXED; Random time $DATA SLB.nm $INPUT ID TIME AMT DV MDV EVID SS II ADDL RATD COVA $SIMULATION (35436 ) ONLY SUB=1 $SUBROUTINE ADVAN2 TRANS=2 $PK ;This PK block has to be used in all problems below ;Therefore all model must be covered by this model KA = THETA(1) * EXP(ETA(1)) CL = THETA(2) * EXP(ETA(2))*COVA**THETA(5 ) V = THETA(3) * EXP(ETA(3)) F1 = THETA(4) * EXP(ETA(4)) S2=V $ERROR CP=F Y = F * EXP( EPS(1) ) + EPS(2) IF (AMT.GT.0) THEN ATIME=TIME RTIME=TIME LTIME=TIME ENDIF MDV2=MDV IF (TIME.EQ. 101) RTIME=ATIME+1.7320508076 *EXP(0.3339544233 *EPS(3)) IF (TIME.EQ. 102) RTIME=ATIME+23.916521486 *EXP(0.050780836 *EPS(3)) IF (TIME.EQ. 201) RTIME=ATIME+3.4641016151 *EXP(0.0874491407 *EPS(3)) IF (TIME.EQ. 202) RTIME=ATIME+5.4772255751 *EXP(0.055421818 *EPS(3)) IF (TIME.EQ. 203) RTIME=ATIME+7.4833147735 *EXP(0.0405906612 *EPS(3)) IF (RTIME.LE.LTIME) THEN RTIME=LTIME+.01 MDV2=1 ENDIF LTIME=RTIME $SUPER SCOPE=4 ITERATIONS=100 ;This is the big DO Loop around the following Models $PROB Popkin Simulation 1 to generate random Time $THETA (.1,1.4 ,10) ;Ka 1 $THETA (.1,3.5 ,50) ;CL 2 $THETA (1,20 ,1000) ; V 3 $THETA (1 ,1 ,1 ) ;F 4 $THETA (-2,0.2 ,2) ; CL (Covar) 5 $OMEGA 0.223144 ;kA $OMEGA 0.039221 0.039221 ;CL V $OMEGA .0001 FIXED;F $SIGMA 0.039221 ;ERR Mult $SIGMA 0.005 ;Err Add $SIGMA 1 FIXED; Random time $INPUT ID TIME AMT DV MDV EVID SS II ADDL RATD COVA $SIMULATION (-1) ONLY SUB=1 $TABLE ID RTIME AMT DV MDV2 EVID SS II ADDL RATD COVA FILE=SIMUL1 NOPRINT NOHEADER NOFORWARD $PROB Popkin Simulation 2 to generate random Concentrations $THETA (.1,1.4 ,10) ;Ka 1 $THETA (.1,3.5 ,50) ;CL 2 $THETA (1,20 ,1000) ; V 3 $THETA (1 ,1 ,1 ) ;F 4 $THETA (-2,0.2 ,2) ; CL (Covar) 5 $OMEGA 0.223144 ;kA $OMEGA 0.039221 0.039221 ;CL V $OMEGA .0001 FIXED;F $SIGMA 0.039221 ;ERR Mult $SIGMA 0.005 ;Err Add $SIGMA 1 FIXED; Random time $DATA SIMUL1 (12E12.0) NRECS=700 NOOPEN $INPUT ID TIME AMT DV MDV EVID SS II ADDL RATD COVA $SIMULATION (-1) ONLY SUB=1 $PROB Popkin Model $THETA (.1,1.4 ,10) ;Ka 1 $THETA (.1,3.5 ,50) ;CL 2 $THETA (1,20 ,1000) ; V 3 $THETA (1 ,1 ,1 ) ;F 4 $THETA (-2,0.2 ,2) ; CL (Covar) 5 $OMEGA 0.223144 ;kA $OMEGA 0.039221 0.039221 ;CL V $OMEGA .0001 FIXED;F $SIGMA 0.039221 ;ERR Mult $SIGMA 0.005 ;Err Add $SIGMA 0.01 FIXED; Random time ;Without data statement the data from the last step are used $INPUT ID TIME AMT DV MDV EVID SS II ADDL RATD COVA $ESTIM MAXEVALS=2000 SIGDIGITS=3 PRINT=0 METHOD=0 $PROB Popkin H0 $THETA (.1,1.4 ,10) ;Ka 1 $THETA (.1,3.5 ,50) ;CL 2 $THETA (1,20 ,1000) ; V 3 $THETA (1 ,1 ,1 ) ;F 4 $THETA 0.0001 FIXED; CL (Covar) 5 $OMEGA 0.223144 ;kA $OMEGA 0.039221 0.039221 ;CL V $OMEGA .0001 FIXED;F $SIGMA 0.039221 ;ERR Mult $SIGMA 0.005 ;Err Add $SIGMA 0.01 FIXED; Random time ;Without data statement the data from the last step are used $INPUT ID TIME AMT DV MDV EVID SS II ADDL RATD COVA $ESTIM MAXEVALS=2000 SIGDIGITS=3 PRINT=0 METHOD=0

From: Justin Wilkins justin.wilkins@farmbio.uu.se Subject: Re: [NMusers] NONMEM Date: Thu, 14 Sep 2006 13:40:08 +0200 Dear Ritu, The example you quote seems to be more suitable for producing a limited series of Monte Carlo simulations than for producing a truly optimal design. If you want to do it in the most appropriate way, you will need to use a tool designed specifically for the purpose. You will have to decide what your needs are (whether you need exact sampling times or sampling windows, which will inform your choice of approach - D- or ED-optimality, for example), and then select one of the several pieces of software available based on these requirements. Two that are commonly used for generating population designs are PopED - http://depts.washington.edu/rfpk/rd/software_popED.html PFIM - http://www.bichat.inserm.fr/equipes/Emi0357/download.html Each (there are others as well) has its own strengths and weaknesses but I have used PopED with some success in the past, and PFIM has a good reputation. Both packages require other software to run - PopED needs O-Matrix or Matlab, and PFIM needs R or S-PLUS. Both rely on minimization of the inverse of the Fisher information matrix with respect to some variable (time, in your case), but others on this group will probably be able to give you more detail about this than I can. $SUPER is a NONMEM control record used to program a series of problems for NONMEM to perform within a single file. As for the other NONMEM parameters you mention, EVID is the PREDPP event identification variable, is usually required, and has 5 possible values - 0 for observation events, 1 for dose events, 2 for 'other-type' events such as compartment resets, 3 for reset events (which re-initialize the PK kinetic system) and 4 for reset-and-dose events (a combination of 1 and 3). SS is the steady-state data item for PREDPP, and can have 4 possible values - 0, indicating that the dose is not a steady-state dose, 1, indicating a steady-state dose with reset, 2, indicating steady-state dose without reset, and 3, which is very similar to 1. The above is a (very) basic summary from the NONMEM online help, which is accessed by the commands >> nmhelp SUPER >> nmhelp EVID >> nmhelp SS from the command prompt. Have a look for more detail. COVA and RATD have no special meaning in NONMEM and appear to have been user-defined variables (covariates, perhaps) supplied by the author of the quoted code. Justin

From: Mark Sale - Next Level Solutions mark@nextlevelsolns.com Subject: Re: [NMusers] NONMEM Date: Thu, 14 Sep 2006 08:21:14 -0700 My 2 cents: First, you need to define what you mean by optimal. Traditional statistical optimal design - e.g. ANOVA is aimed at getting parameter estimates as precisely estimated as possible - a laudable goal in traditional statistics, nicely implemented in a number of applications, I'm most familiar with Steve Duffuls ( http://www.bichat.inserm.fr/equipes/Emi0357/download.html), but I'll let him tell you how wonderful it is. The three central problems, IHMO with existing methods are: 1. Limited to D (or ED, or perhaps C|S) optimal for model parameters. What if we want to get a model with the smallest mean absolute error? What if we want a study design that estimates some other statistic (survival time, AUC, difference between treatment A and placebo, etc) as precisely as possible? 2. No/little flexibility in sample number - they address sample times only, the sample number is (usually) fixed. What if our question is (the more realistic) - what is the optimal study design (which I think means getting and "adequate" answer for the lowest cost) to answer this question if samples cost $200 each and subjects cost $5000 to enroll and $1000/week to keep in the study. 3. They don't tell you if a study is adequate - only which design is best, for a given number of subjects/samples. We have done a small demonstration project optimizing a BE study. Essentially: Find the optimal number of subjects, number of samples, sample times to result in a "successful" BE study (e.g., 1-beta = 90%, 1-alpha = 90%), given a specified cost of samples and subjects. NCA pk was done. Model also included uncertainty about model parameters. The algorithm used for the optimization was (my favorite algorithm) Genetic algorithm. Interestingly, the resulting design was quite close to what we usually do - except the GA answer had fewer sample times than the traditionally designed study (and so was a little cheaper). It also was clear that D optimal was very different from NCA optimal sampling times. I haven't had time to continue persuing this - any grad student interested, I'm happy to share the code that I have. Mark Mark Sale MD Next Level Solutions, LLC www.NextLevelSolns.com

From: "Andrew Hooker" andrew.hooker@farmbio.uu.se Subject: Re: [NMusers] NONMEM Date: Thu, 14 Sep 2006 18:28:51 +0200 Hi Mark, I agree with most of your assessment about the current state of optimal design. The designs are generally based on getting the best possible parameter estimates of your model and it important to develop methods for looking at other test statistics. Two points: 1. In PopED (www.rfpk.washington.edu ) we have to ability to optimize over the number of samples per individual in a study, the number of individuals in a study, and other design variables other than sample times. See: M. Foracchia, A. Hooker, P. Vicini and A. Ruggeri. PopED, a software for optimal experimental design in population kinetics. Comput Methods Programs Biomed, 74: 29-46, 2004. 2. To me it is not so surprising that the optimal design that you attempted using NCA resulted in sample times similar to the designs people come up with normally (without any fancy optimal design) for these studies, because most people design their studies based on NCA type thought processes anyway. The question that comes to mind is: why shouldn't we use the information we gain from population mixed effect models in our design calculations? It would be interesting to compare the performance of your NCA based design and the D-optimal design you calculated. -Andy Andrew Hooker, Ph.D. Assistant Professor of Pharmacometrics Div. of Pharmacokinetics and Drug Therapy Dept. of Pharmaceutical Biosciences Uppsala University Box 591 751 24 Uppsala Sweden Tel: +46 18 471 4355 www.farmbio.uu.se/research.php?avd=5

From: Mark Sale - Next Level Solutions mark@nextlevelsolns.com Subject: Re: [NMusers] NONMEM Date: Thu, 14 Sep 2006 09:50:09 -0700 Andy, There are several practical obstacles to this. The first reason that no one uses a formal optimization (based on a pop pk model) to optimize NCA, or other study endpoints it that it is pretty hard. We estimated saving about 10% of sample assay costs (maybe 100 samples per study, ~$10,000 dollars in a $500,000 study), and the sample assay budget came from a different group than the people designing the study, so the study designers didn't lose a lot of sleep over assay costs, prefering the CYA approach. It took me several of weeks of work to do the optimization, another downside. The second study optimization would obviously go a lot faster, but it isn't clear that there is a business case for it, until someone writes a general application to do it. Hence my offer of any code I have to anyone who wants to pursue it. It also is very computationally intensive, running Monte Carlo simualtion on 1000's of designs, doing the pk and statistics for each design (ANOVA for NCA) etc. Probably the pay off for BE studies is marginal. The payoff for large, expensive, difficult to recruit studies may be significant, and they wouldn't be much harder to optimize. Another practical issue is that the stats groups were skeptical - because we basically would control the SE of the AUC - finding an "optimal" SE - not a minimal value- by controling sample number and times. They told us that stats was responsible for estimating the SE of the parameters, not clin pharm. They prefered to use historical values for SE of AUC (and worst case scenario at that), and so the formal power analysis, which was done by stats, didn't reflect the optimization, only the SE of the NCA quantities from an historical study. These are all reasons why I gave up on this a while ago. But, I think in theory it is a very practical way to formally optimize study designs - much more powerful than just doing some simulations in Trial Simulator and manually tweaking some study parameters. Mark Sale MD Next Level Solutions, LLC www.NextLevelSolns.com

From: "Stephen Duffull" stephen.duffull@stonebow.otago.ac.nz Subject: Re: [NMusers] NONMEM Date: Fri, 15 Sep 2006 08:00:34 +1200 Hi all >> The example you quote seems to be more suitable for producing >> a limited series of Monte Carlo simulations than for >> producing a truly optimal design. If you want to do it in the >> most appropriate way, you will need to use a tool designed >> specifically for the purpose. I think it is worth a note here that, in my experience of optimizing designs for both industry and academia, I have never been asked to find an optimal design. What I do get asked is to provide a sufficient design that requires minimum effort, for example: fewest patients, reduced number of doses, fewest samples per patient etc. The term sufficient means a design that meets the needs for which the model is intended to be used. It is also my experience that it is almost impossible to do this within any acceptable time-frame using NONMEM or some other package that has MC sim + estimation - and that an information theoretic approach is both practical and very quick at doing this. >> PopED - http://depts.washington.edu/rfpk/rd/software_popED.html >> PFIM - http://www.bichat.inserm.fr/equipes/Emi0357/download.html Of course, I have to plug: POPT - www.winpopt.com WinPOPT - www.winpopt.com POPT requires MATLAB and WinPOPT runs independently of any other software. Also I think that France Mentre has released PFIM_OPT for S (or R). Steve -- Professor Stephen Duffull Chair of Clinical Pharmacy School of Pharmacy University of Otago PO Box 913 Dunedin New Zealand E: stephen.duffull@otago.ac.nz P: +64 3 479 5044 F: +64 3 479 7034 Design software: www.winpopt.com

From: Nick Holford n.holford@auckland.ac.nz Subject: Re: [NMusers] NONMEM Date: Fri, 15 Sep 2006 08:31:05 +1200 Steve, If I may split the academic hair a little more ... I suspect that in fact your collaborators would have initially asked you for an 'optimal design' i.e. these are the words they would have used when asking you to help them. But you would have offered a 'sufficient design' as being good enough. I accept that methods based on the Fisher information matrix (FIM) are much faster than brute force Monte Carlo (MC) methods but the FIM methods are limited to minimizing parameter precision as the objective. There are other objectives for trial design e.g. power, which can be (tediously) explored using MC methods but which are only indirectly optimized using FIM methods. Nick -- Nick Holford, Dept Pharmacology & Clinical Pharmacology University of Auckland, 85 Park Rd, Private Bag 92019, Auckland, New Zealand email:n.holford@auckland.ac.nz tel:+64(9)373-7599x86730 fax:373-7556 http://www.health.auckland.ac.nz/pharmacology/staff/nholford/

From: "Stephen Duffull" stephen.duffull@stonebow.otago.ac.nz Subject: Re: [NMusers] NONMEM Nick >> If I may split the academic hair a little more ... Of course :-) >> I suspect that in fact your collaborators would have >> initially asked you for an 'optimal design' i.e. these are >> the words they would have used when asking you to help them. >> But you would have offered a 'sufficient design' as being good enough. Sometimes - but not always. Many 'sponsors' really do want the 'minimally effective' design - and don't ask for an 'optimal' design. And of course some sponsors know what they want but inadvertently use the term 'optimal' anyway. So, I don't agree with your assertion here. >> I accept that methods based on the Fisher information matrix >> (FIM) are much faster than brute force Monte Carlo (MC) >> methods but the FIM methods are limited to minimizing >> parameter precision as the objective. (Hair split - you mean maximize precision.) Not true - although I accept that this is their most common use in practice. In addition to maximizing precision, FIM based designs can be used to: 1) determine designs for model discrimination 2) determine designs with minimum bias 3) determine designs for power to reject the null hypothesis (for model building decisions only) 4) determine designs that carry the highest probability of success (GLMs) And of course any combination of the above (including with the standard maximizing parameter precision). I'm not of course advocating that FIM based methods should replace all MC methods - but I think both have a complementary role. Steve -- Professor Stephen Duffull Chair of Clinical Pharmacy School of Pharmacy University of Otago PO Box 913 Dunedin New Zealand E: stephen.duffull@otago.ac.nz P: +64 3 479 5044 F: +64 3 479 7034 Design software: www.winpopt.com

From: Mark Sale - Next Level Solutions mark@nextlevelsolns.com Subject: Re: [NMusers] NONMEM Date: Thu, 14 Sep 2006 13:49:20 -0700 I agree with Steve, optimal design for studies is not generally available. Instead, we use simulation, to plagarize Steves term, to find a sufficient design by tweaking parameters and doing MC simulation. I would however, suggest that given sufficient computational power, formal optimization is possible on a reasonable time line - a couple of days. For most optimizations (i.e., BE studies, NCA pk studies, dose-response, survival etc), you won't use NONMEM anyway. You use SAS or Splus in some non-iterative alorithm (like ANOVA) that is very fast. So, for most trial optimizations, you don't need the (NONMEM) estimation, only the simulation. But, we currently don't have the tools to do this. Mark Sale MD Next Level Solutions, LLC www.NextLevelSolns.com

From: "Stephen Duffull" stephen.duffull@stonebow.otago.ac.nz Subject: Re: [NMusers] NONMEM Date: Fri, 15 Sep 2006 09:25:45 +1200 Mark >> I agree with Steve, optimal design for studies is not >> generally available. Instead, we use simulation, to >> plagarize Steves term, to find a sufficient design by >> tweaking parameters and doing MC simulation. I would >> however, suggest that given sufficient computational power, >> formal optimization is possible on a reasonable >> time line - a couple of days. I would suggest that using a information theoretic technique you would get rid of the simulation component completely and do this more efficiently using an FIM approach. I do not advocate simulation where quicker methods are available. If you use simulation then: 1) how do you know what designs to choose? 2) how do you determine when you have a minimally effective design - could there be another design around the corner that you didn't think of which is more efficient? Why not just let an FIM search do its business, get the answer for the most efficient-sufficient design. In a sense the most efficient design is how you would determine 'optimal' (or 'best') - just that the design is cast from finding the minimum experimental effort rather than some other goal. Steve -- Professor Stephen Duffull Chair of Clinical Pharmacy School of Pharmacy University of Otago PO Box 913 Dunedin New Zealand E: stephen.duffull@otago.ac.nz P: +64 3 479 5044 F: +64 3 479 7034 Design software: www.winpopt.com

From: Mark Sale - Next Level Solutions mark@nextlevelsolns.com Subject: Re: [NMusers] NONMEM Date: Thu, 14 Sep 2006 14:55:07 -0700 >I would suggest that using a information theoretic technique you would get >> rid of the simulation component completely and do this more efficiently >> using an FIM approach. I do not advocate simulation where quicker methods >> are available. Absolutely, the only problem is the quicker method often don't answer the question you want answered (they tell you how to get the smallest SE of model parameters, not how to do the cheapest/most powerful study/fastest). >> If you use simulation then: >> 1) how do you know what designs to choose? You optimize, based on user defined criteria (1-alpha >=0.9, 1-beta >>=0.9, minimal cost, shortest duration) >> 2) how do you determine when you have a minimally effective design - could >> there be another design around the corner that you didn't think of which is >> more efficient? That what optimization does, I can reference you to a text. GA (unlike FIM) does NOT guarantee the "best" answer, the textbooks always say "near optimal solution". But, some analysis more sophisticated than I understand, done mostly at Illinois University has examined "GA deceptive searches", and you can be pretty sure that properly done GA is truly optimal - but not guaranteed. Also, as an aside I have combined search algorithms and greatly increased the "robustness" of the search, specifically addressing the problem you mention. >> Why not just let an FIM search do its business, get the answer for the most >> efficient-sufficient design. In a sense the most efficient design is how >> you would determine 'optimal' (or 'best') - just that the design is cast >> from finding the minimum experimental effort rather than some other goal. Absolutely (again), if your goal is minimal SE of model parameters. If your goal is cheapest/most power/fastest, I'm not sure FIM will work. Mark Sale MD Next Level Solutions, LLC www.NextLevelSolns.com

From: "Gastonguay, Marc" marcg@metrumrg.com Subject: Re: [NMusers] NONMEM Date: Thu, 14 Sep 2006 19:12:58 -0400 Steve, I'm curious to learn of examples where FIM-based optimization methods have been useful for achieving designs that minimize bias (#2 in your post below) of parameter estimates. It was my understanding that in order to assess bias of parameter estimates, it is necessary to run batches of MC simulation-estimation cycles and compare estimated values back to a reference "true" value (e.g. simulation model parameter values). Thanks in advance for your insight. Marc Marc R. Gastonguay, Ph.D., President & CEO, Metrum Research Group LLC 2 Tunxis Rd. Suite 112, Tariffville, CT 06081 Direct:860-670-0744, Main:860-735-7043, Fax:860-760-6014 Email:marcg@metrumrg.com, Web:www.metrumrg.com

From: "Stephen Duffull" stephen.duffull@stonebow.otago.ac.nz Subject: Re: [NMusers] NONMEM Date: Fri, 15 Sep 2006 11:58:49 +1200 Marc Thought that I might escape comment on this :-) >> I'm curious to learn of examples where FIM-based optimization >> methods have been useful for achieving designs that minimize >> bias (#2 in your post below) of parameter estimates. I know of no examples for NL MEM. That doesn't mean it can't be done of course... >> It was >> my understanding that in order to assess bias of parameter >> estimates, it is necessary to run batches of MC >> simulation-estimation cycles and compare estimated values >> back to a reference "true" value (e.g. simulation model >> parameter values). As a gross simplification, I might suggest that bias in parameter estimates can arise from 2 main sources (I am sure someone will correct me here). 1) due to not finding the true maximum of the likelihood 2) due to model misspecification The latter case is a special construct since bias doesn't truly exist it just appears to exist because you're thinking about one model and fitting another and then the parameter estimate you obtain you apply to the wrong model and hence it appears to be biased. I don't think there's much you can do about case 1 (other than use a better search algorithm). Case 2 is a real problem and can be implicit (e.g. due to linearisation) or explicit (e.g. your data does not support estimation of a slow distribution phase hence your estimate of half-life while perhaps accurate for your model is inaccurate to describe the actual time course of disposition). Case 2, you can optimize designs (to minimize the effects of this bias) by using information theoretic techniques without the need for MC. I hasten to add here that this may not necessarily be based solely on FIM - but may have other analytic computation. So I should correct my last statement: >>> > In addition to maximizing precision, FIM based designs can > >> be used to: To: "In addition to maximizing precision, information theoretic designs (not necessarily based solely on FIM but under the overall umbrella of "Optimum Design of Experiments") can be used to:" Regards Steve -- Professor Stephen Duffull Chair of Clinical Pharmacy School of Pharmacy University of Otago PO Box 913 Dunedin New Zealand E: stephen.duffull@otago.ac.nz P: +64 3 479 5044 F: +64 3 479 7034 Design software: www.winpopt.com

From: Timothy H Waterhouse WATERHOUSE_TIMOTHY_H@Lilly.com Subject: Re: [NMusers] NONMEM Date: Fri, 15 Sep 2006 12:04:57 -0400 Hi all, While it's true, as far as I know, that "existing methods" (meaning software such as POPT, PFIM etc) don't allow for optimal design for precise estimation of other statistics, I think this is a relatively simple problem to address. If the feature of interest (such as AUC or tmax) can be written as a function of your PK parameters, a c-optimal design will give you precise estimates of this feature, and can be obtained using a function of the FIM. The problem with c-optimal designs is that they often give you a singular FIM, meaning they have zero efficiency for parameter estimation. This was actually addressed in a talk by Anthony Atkinson at the DEMA conference in Southampton last weekend. He used a combination of the "c" and "D" criteria to find designs which are efficient for both, using a simple PK model as an example. His methods were for fixed effects models, but I think they will extend to mixed effects models using the usual approximate information matrix. If you want to design for maximum power, minimum bias, etc, the information matrix may not be quite as helpful... Tim

From: Mark Sale - Next Level Solutions mark@nextlevelsolns.com Subject: Re: [NMusers] NONMEM Date: Fri, 15 Sep 2006 09:19:59 -0700 Key phrase "If the feature of interest (such as AUC or tmax) can be written as a function of your PK parameters," I'm assuming closed form solution. But, we are still talking about precision in estimating parameters. What about the vast majority of phase II and III studies that we do using hypothesis testing? using ANOVA, Fisher exact test etc. Not infrequently we can develop models for the endpoint. Can we optimize the study (obviously, I beleive we can). We have not had a lot of success convincing people that we should move from a hypothesis testing mode to an estimation mode. BTW, the reason I'm obsessed with this - and I admit that I am, is that I think this is a real opportunity for Clin pharm/modeling to impact later phase development. Companies like Pharsight (and others) have really done an excellent job of raising awareness that modeling can contribute to phase III design. Part of the reason that this role hasn't been fully embraced everywhere, (IMHO) is that they are trying to wrest control of study design away from stats (and others). Methods to optimize phase III trials are new turf - we wouldn't have to fight anyone for it, it really is very complimentary to stats - since it requires their methods to implement. If we could actually demonstrate saving money (optimizing for lowest cost for an adequate study), I think we can seriously impact late phase development. Of course in academics, money is never an issue .... But, tools need to be developed, and we need more computers. Mark Sale MD Next Level Solutions, LLC www.NextLevelSolns.com

From: "A.J. Rossini" blindglobe@gmail.com Subject: Re: [NMusers] NONMEM Date: Fri, 15 Sep 2006 21:13:44 +0200 On 9/15/06, Mark Sale - Next Level Solutions mark@nextlevelsolns.com wrote: <-- lots of comments deleted --> > Part of the reason that this role > hasn't been fully embraced everywhere, (IMHO) is that they are trying > to wrest control of study design away from stats (and others). Methods > to optimize phase III trials are new turf - we wouldn't have to fight > anyone for it, it really is very complimentary to stats - since it > requires their methods to implement. Why need to wrestle anything from stats? Any decent M&S group ought to have a few statisticians around... (speaking from the Novartis M&S viewpoint, where I'm in the statistics subgroup...). More seriously, this IS a critical problem -- there are general design principles for PK sampling design on the programmatic level, i.e. akin to the "adaptive designs" work that more traditional stats groups have been working on. Part of that will be thinking through the range of "critical path" questions that PK sampling might answer, and being a bayesian (and PopKin scientist) about it, leveraging compound family and indication/pathway knowledge about the challenges that could be faced. Analysis is one thing, but design, now that is completely different. And especially working the patterns, i.e. "given models X through Z at Ph I with dense sampling, which resulted in sparse sampling design W for Ph II and PhIb trials which were fit to models X1-- Z1, what does the range of reasonable models suggest that we should use for sampling at Ph III?" (i.e. design using the previous work, or even more interesting, the trajectory of previous work). After all, as I used to say when I was a biostat/stat prof, "nothing like a bad design or sloppy operations to generate another PhD thesis...". But getting the design right, sigh, that's contextual and hard. best, -tony blindglobe@gmail.com Muttenz, Switzerland. "Commit early,commit often, and commit in a repository from which we can easily roll-back your mistakes" (AJR, 4Jan05). _______________________________________________________