Re: Storing simulations

From:Nick Holford Subject:Re: [NMusers] Storing simulations Date:Fri, 12 Jul 2002 08:14:54 +1200 I am not sure what you are trying to do but let me guess. If you want to simulate a data set using NONMEM then estimate the parameters of that data set using NONMEM you have 2 choices: 1. One Stage Stage 1: Add the $SIMULATION (1234567) SUBPROBLEMS=100 to an existing control stream. This will use the dataset as a template, simulate a new DV which replaces the original DV value in a temporary (hidden from the user) dataset, then estimate the parameters using the simulated data. With 100 subproblems this will be repeateed 100 times. The parameter estimates will appear in the usual NONMEM output listing but instead of one problem you will find the results of 100 problems listed one after another. Sounds too easy? 'Fraid so. NONMEM has a bug/feature that means that if one of the subproblems encounters certain errors (which arise commonly when dealing with simulated data and using the same initial estimates for all runs) it will not go onto the next problem but just stops. This unreliability makes it worthless for serious work. 2. Two Stage Stage 1: Add the $SIMULATION (1234567) ONLSIMULATION to an existing control stream. This will simulate a new data set as in the one stage method but you can then save the simulated results (and ID and other covariates) in a TABLE file in the usual way. The generated "population data" will be saved in the TABLE file. Stage 2: Change the dataset in your control stream to the name of the table file generated in Stage 1. (Remove any $SIMULATION record!). This will then run NONMEM in the usual way. the But if you are doing Monte Carlo simulation and need to do this 100 times or more then you need to find a way to automate Stages 1 and 2. With both methods you will need to find an automated way of extracting things such as objective function values and parameter estimates in order to perform a meta-analysis. Wings for NONMEM ( http://wfn.sourceforge.net) will handle this for you if you use the One Stage method. I have also made programs for doing the Two Stage method but they are a bit more clumsy and not currently supported directly in the distributed version of WFN. The basic idea is to run Stage 1 with ONLYSIMULATION and SUBPROBLEMS=100. Then extract, one at time, the 100 simulated data sets saved in the TABLE file output and run each of these extracted data sets with a standard NONMEM control stream. -- Nick Holford, Divn Pharmacology & Clinical Pharmacology University of Auckland, 85 Park Rd, Private Bag 92019, Auckland, New Zealand email:n.holford@auckland.ac.nz tel:+64(9)373-7599x6730 fax:373-7556 http://www.health.auckland.ac.nz/pharmacology/staff/nholford/