In NLME, while doing population analysis, (single dose (5 mg) data, 10 subjects) the final model is copied, with all the fixed and random effects were accepted.
By checking the sim /pred.check box (changing from simple) in the copied model and using # of replicates as 10, with N iterations changed to 0, 10 replica were generated. Nothing was selected in Binning option. Adding a simulation table generated some more data (with time points mentioned in the sequence).
Is it the way pred check is done ?
What does the option " Pred. Corr" in the main tab mean ??
I would like to simulate for a new dose say for e.g., 10 mg instead of 5 mg. How would i do that by keeping the final model intact. Do i have to add another excel to the data object ?? Any example file would be helpful !!
In NLME, while doing population analysis, (single dose (5 mg) data, 10 subjects) the final model is copied, with all the fixed and random effects were accepted.
By checking the sim /pred.check box (changing from simple) in the copied model and using # of replicates as 10, with N iterations changed to 0, 10 replica were generated. Nothing was selected in Binning option. Adding a simulation table generated some more data (with time points mentioned in the sequence).
Is it the way pred check is done ?
Yes but visual predictive checks will show in output only the predictions at the observed times (90% confidence bounds superimposed with observed data) . The add/sim table is an “add on” and it results in a simulation of different patients (different random numbers) where the link with the data is only through the dosing and covariate information.
What does the option " Pred. Corr" in the main tab mean ??
It should be a correction for the binning option as observed times across the population can be different (one patient t=3.3, other 3.3, other 3,etc…, difference between nominal and actual times or just differences across patients by design)
I would like to simulate for a new dose say for e.g., 10 mg instead of 5 mg. How would i do that by keeping the final model intact. Do i have to add another excel to the data object ?? Any example file would be helpful !!
You just replace the observed data set by a template one where you put only the dosing information. each id patient will show a different dosage regimen and/or different covariate setting.
In add/sim table you select the times of observations.
The vis pred simulation has no meaning in that case as you just do pure simulation. Only rawsim…csv output is the one that you should look at.
I am out of town. If not clear enough let me know and I will send you an example this week-end.
what is the minimum number of observations (or # of replicates) that need to be generated in predictive check. Will it depend on number of subjects in the Original study ??
If you want just visual predictive checks, then you use the observed times as in your study. How many replicates is not easy to know as it depends on the sensitivity of your model parameters to the response as well as the variability.
I am confused about your input “If you want just visual predictive checks, then you use the observed times as in your study”. For example, I have 10 subjects and each subject has 20 observations. What is the observed times? 10? 20? or 200?
Another questions about the visual predictive checks. When I finished a pred. check (replicate # is 100) and got a AllPCData.csv. I am not sure whether my method to valuate my model is right or not. I summarize the csv file and got median DV at each time point for each project and then compared the mean DV with observed concentration. I saw a video in which the author compared the observed concentration to the median and 90th prediction interval of simulated concentration for each subject. Can we calculate the prediction interval in Phoenix NLME?
niter=1000 or anything else is irrelevant when using PRED/SIM.
You have number of replicates in run options and Pred/SIM selection that tells you how many duplicate of the observed design you want to simulate.
Predictive check will take the exact design as the one observed and will simulate at these time points only and nrep times where nrep is the number of replicates.
If you click on add/sim table, you have the option to select the observed times you want and perform a simulation at theses times. The link with the input data set is in the dosing and covariate information but not the times.
When using ad//sim table, you can also replace your data set by any template table where you control completely the design.
These the main differences between Predictive checks and pure simulation.
If now you define a covariate like dosing as a categorical covariate, you can stratify by that covariate.
This is a good idea to stratify if enough information is available to superimpose observed stratified data versus predicted stratified confidences bounds and see to what extent about 90% of the observed data are within the 90% confidence bounds you predict and the average data around the 50% predicted quantile. If you see a trend showing foe example the observed data starting to drift apart from the 90% confidence region, this can be a sign of non linear clearance. This would not be seen without stratification as all the data together may look good.
.
If you have very few data, at some point stratification will prevent you to see anything but just randomness.