答复: A question about handling large-scale data

Dear Nick, THANKS for your quick suggestion. Because I want to construct a mechanism-based model, H+ concentration has to be considered, which allows us to receive circadian rhythm and dilution effect by food. The log-transformation could be helpful. My current error model is Y = LOG(IPRED)+EPS(1) for log-transformed DV value. Now, just “rounding error” or “termination” warnings show up. My question is if it is a problem when IPRED is large-scaled. If yes, how to handle it? And which kind of weighting factor should be preferred for Log error model? By the way, the model incorporated IDR, circadian rhythm, and food effects. Although it simulated data well, but fitting is in problem. In order to remove overparameter problem, I just opened 4 THETAs, 3 ETAs, and 1 EPS to be estimated. Even so, fitting can not be converged. Also, I tried different ADVAN13 and SAEM, both didn’t work. Hope to get more suggestions from you and other users. THANKS IN ADVANCE! Cheers, Dongyang 发件人: [email protected] [mailto:[email protected]] 代表 Nick Holford 发送时间: 2014年11月20日 15:40 收件人: Liudongyang_hotmail; [email protected] 主题: Re: [NMusers] A question about handling large-scale data Dear Dongyang, Transforming H+ concentration to the pH scale seems a reasonable idea. However I would not give much importance to warnings such as "rounding error" or "termination". You should evaluate your model based on appropriate changes in objective function value, plausible parameters, bootstrap confidence intervals of parameters and visual predictive checks of the model predictions. If you get errors associated with "numerical difficulty" then try a different DE solver e.g. ADVAN13 or switch from FOCE to SAEM. Best wishes, Nick