Re: SIMULATION QUESTIONS

Hi Live Below is the mixture of some answers and some questions concerning your model: 1. How can you use Erlang distribution in the PK model? Most of the times, normal or log-normal should be sufficient for the model description, and use of more exotic distributions should be justified somehow. Details of the model would help to understand whether this distribution is relevant for a PK model that you study. 2.If you use linear two-compartment model, why would you use ADVAN5? Advan3-4 should be sufficient. Is it because of some complicated absorption model (and Erlang distribution also relates to the absorption)? 3. If you use SUBPROBLEMS, NONMEM will supply seeds for each problem. If you run the same script multiple times and then combine the results, you need to supply different seed for each run. If you do predictive check simulations (is this what you are trying to do?) then you need few hundred sets, may be 500 if the data set is very large or 1000 otherwise. 4. If simulations over-predict the data that were used to fit the model, this indicates deficiency of the model rather then insufficient information given to NONMEM. You should look on model diagnostic plots and try to understand what is inconsistent. 5. Most often for predictive check simulation some overall statistics such as Cmax, AUC, Cmin, Ctrough are used. If you use me or rmse, it should be done for all points separately, one by one, and then summarized. Simulated IPRED rather than PRED should be used for comparison. 6. You data set lists observation compartment as CMT=8, is this correct(given the two-compartment model)? Leonid ------------------------------------------------------ Leonid Gibiansky President QuantPharm LLC www.quantpharm.com [EMAIL PROTECTED] wrote: > Dear nmusers, > > I'm working on a pharmacokinetic population analysis of a data set in > NONMEM for my master thesis. The data set I'm analysing consists of 12 > hours concentration-time profiles. A 2-compartment model with Erlang > distribution describes the data well, and I'm now validating my model. > However, I've run into some questions. > > I've performed a data splitting analysis, and have some questions about > predictive performance/simulation in NONMEM. > > 1. What is the difference between using different seeds and subproblems > when doing a simulation? And how many seeds/subproblems are generally > considered to be needed? > > 2. Simulating my subsets for the data splitting seems to give a > over-prediction of the concentrations observed. Is there any way I can put > in the C0/C2 concentration, in order to give NONMEM more information and > hopefully better PRED when doing a simulation? Only age was found to be a > significant covariate, and I suspect the reason for over-predicting the > concentrations is that NONMEM needs more information. > > 3. When having different time measurements, how do I calculate out me (mean > error) and rmse (root mean square error) for the subset; can I use the > average for the different time measurements in the subset? > > Part of the inputfil: > ID AMT TIME DV MDV SS II CMT AGE RATE > 8 225 0 0 1 2 12 1 59 0 > 8 0 0 0 0 0 0 8 59 0 > 8 0 0.23 0 0 0 0 8 59 0 > 8 0 0.48 0 0 0 0 8 59 0 > 8 0 0.98 0 0 0 0 8 59 0 > 8 0 1.43 0 0 0 0 8 59 0 > 8 0 1.98 0 0 0 0 8 59 0 > 8 0 3.03 0 0 0 0 8 59 0 > 8 0 4.07 0 0 0 0 8 59 0 > 8 0 5.98 0 0 0 0 8 59 0 > 8 0 8.03 0 0 0 0 8 59 0 > 8 0 9.97 0 0 0 0 8 59 0 > 8 0 1.90 0 0 0 0 8 59 0 > > Part of the controlfile: > $PROB > $DATA > $INPUT > $SUBROUTINE ADVAN5 SS5; > $MODEL > $PK > $ERROR > > $THETA > (FIX 7.98) ; K12 (B) > (FIX 27.2) ; Q/F > (FIX 57) ; V8 > (FIX 203) ; V9 > (FIX 23.4) ; CL/F > (FIX 0.107) ; age effect > > $OMEGA > (FIX 0.06) > (FIX 0.09) > (FIX 0.27) > (FIX 0.91) > (FIX 0.02) > > $SIGMA > (FIX 0.02) > (FIX 1370) > > $SIMULATION (49682607) ONLYSIMULATION SUBPROBLEMS=10 > > $TABLE > > Your help is highly appreciated, thank you in advance. > > Best regards, > Live > > Live Storehagen > Master student > Institute of Pharmacy > University of Oslo > Norway