Choosing variabilities

Hi all, I am trying to model a system and I am running into the issue of which parameters should I include variabilities. The model is complex with more than 6-7 parameters and its a dose driven effect model, fitting the effect observations only. Some parameters describe the dose dynamics and other contribute to effect. Leaving aside the variabilities on effect part if I concentrate on dose part variabilites, it leaves me in a puzzle. What dose parameters should I choose to add variability on. There is very little physiological or mechanistic information on the dose part, which may also be called a hypothesis. I chose different models with variabilities on different dose parameters. Almost all run with different end OBJ function. The final fit is more or less similar between most of the models. It's like parameter values change but they compensate each other in final effect. Now, my puzzle is, which model (with which variabilities) should I choose. The one with best OBJ func. The one that may make some physiological relevance (although this relevance may get refuted). Any insight into experience with such problems would be helpful. Thanks, Pratik Bhagunde

If model fit is similar for two (or several) competing models, I would go with the simplest model. Similarity can be assessed by plots of PRED vs PRED, IPRED vs IPRED (of one model versus the competing one) and also by comparing VPC plots (to make sure that the variability is captured correctly by both models). RSEs of parameter estimates is also a good guide: OMEGAs with high RSEs can be removed. Correlation of ETAs scatter plot is helpful to identify over-parametrization (that would show up as extremely strong correlations of ETAS). So I would start with model that includes many random effects, remove as many as possible based on the OF comparison, and then remove as many as possible based on the similarity of the model fit (see above). It also depends on the quality of your data. If data support (allow to estimate with good precision) many random effects, then it is fine to retain those in the model. Physiological relevance is always a good thing: if you can guess that the parameter is stable in the population, then it adds additional support to the removal of the random effect on that parameter. Regards, Leonid -------------------------------------- Leonid Gibiansky, Ph.D. President, QuantPharm LLC web: www.quantpharm.com e-mail: LGibiansky at quantpharm.com tel: (301) 767 5566