14 replies to this topic

[Richa Dua]

I am a new user and for one of my projects I want to figure out a limited sampling strategy to predict the AUC of a drug. I have data from about 12 subjects with extensive sampling and the AUC. Can you pick and choose the different time points and perform simulations? Any direction on how to approach this problems would be helpful.

Thank you!

[Simon Davis]

Hi Richa,
Have you attended one of the training courses or otherwise been introduced to VIF? Variance Inflation Factor is the multiplier of the residual mean square that is used in deriving the asymptotic variance of the parameter estimates.

The VIF is useful when performing simulations, to determine optimal experimental designs. A smaller value is better.
Note: VIF depends only on the study design. We increase the precision of the parameter estimates by making VIFs as small as possible.

Since we don’t have a method to determine the values of times that minimize the VIFs, our strategy is to compute values of the VIF for various choices of the sampling times. The collection of times that minimizes the values of VIFs of our parameters of interest relative to the other sets of times is preferable.

This can be accomplished using the simulation option in WinNonlin below I compared 3 sampling schedules for AUC, to do this I made three different sample schedules corresponding to my proposed protocols and simulated them.

You can also use Partial derivative plots to review visually which time regions are most of interest for your parameters but VIF is more 'objective'.

You probably want to "Split” the population curve into regions of time corresponding to absorption, distribution and elimination. Ideally for each subject, try to get at least one sample collected from each region. These could be fixed times for each region, or could be times randomly selected from a set of times for each region.

Simon

Attached Thumbnails

• 

[Richa Dua]

Hello Simon,

Thank you for your reply. So does that mean I have to build a pop pk model first? As opposed to a 1 or 2 compartment phoenix model?

Thanks.

[Simon Davis]

Not at all you can get VIF from the old WNL CLASSIC engines in simulation mode too, so you could get some indication from that. A population i.e. NLME model would probably give more confidence with sparse sampling but iI wouldn't say it was mandatory

 

 Simon

[Richa Dua]

Hello Simon,

I am trying to come up with a model that fits my data. It will probably be a 2 compartment model. However, I am not sure how to go about running simulations after that and obtaining the VIF values.

I guess my next question would be to know how to proceed with this problem once I have my 2 compartment (non pop Pk) model.

Is there a pdf I can refer to?

Thanks a lot for your help.

[Richa Dua]

Hello Simon,

Would it be okay if I were to send you my project file to see where I am getting stuck?

[Simon Davis]

Sure - you should be able to attach it to your existing post using the EDIT function. Try to use a name of 8 characters or less because we are having a few issues with attachments at the moment.

 

Then others can comment on directions to take with your model/simulations.

 

Even if it 'disappears' I can read it from the admin side and try to take a look at it.

 

   Simon

[Richa Dua]

Hello Simon

Here is my Project. I have struggled quite a bit and could not figure out how you came up with a simulated table to compare values of AUC, Cmax, and Clearance. Any direction on that would be very very helpful.

This project has IV data and oral data (both have a 2 compartment model in the file as well). IV 2 comp. model fits the data perfectly where as oral data - not so much. Not sure how to improve the fit.

Another question would be about estimates for 2 compartment. I tried doing curve stripping on each one in excel but still SDerror % for each subject when sorted by subject were horribly high.

 

Any directions on simulation would be really really helpful [file name=New_Project_V3.phxproj size=1017773]/extranet/media/kunena/attachments/legacy/files/New_Project_V3.phxproj[/file]

Attached Files

•  New_Project_V3.phxproj   993.92KB   0 downloads

[Richa Dua]

Simon,

I just sent you a message with my project file. However, it disappeared. Please let me know if you got something. If not, I can resend it.

[Richa Dua]

Hello Simon

Here is my Project. I have struggled quite a bit and could not figure out how you came up with a simulated table to compare values of AUC, Cmax, and Clearance. Any direction on that would be very very helpful.

This project has IV data and oral data (both have a 2 compartment model in the file as well). IV 2 comp. model fits the data perfectly where as oral data - not so much. Not sure how to improve the fit.

Another question would be about estimates for 2 compartment. I tried doing curve stripping on each one in excel but still SDerror % for each subject when sorted by subject were horribly high.

 

Any directions on simulation would be really really helpful [file name=New_Project_V3-20120406.phxproj size=1017773]/extranet/media/kunena/attachments/legacy/files/New_Project_V3-20120406.phxproj[/file]

Attached Files

•  New_Project_V3-20120406.phxproj   993.92KB   0 downloads

[Richa Dua]

Hello Simon,

Could you please give me an estimate as to when I can expect to hear back from you? Our project deadline is coming up very soon. If you are busy, is there anyone else who could possibly help me out with my questions.

Thank you,

Model world

[Simon Davis]

Hi Richa,
Somewhat unfortunately for you, you asked your question over Easter - I and I imagine many other users of the forum were on holiday.

If you have a formal support request then support@pharsight.com could be a better route for you in the future. Once you have your data fitted and have selected a model you can then perform a simulation trying different sampling sequences. I've done this quickly with a couple of WNL classic models to show how VIF might vary for parameters, the base11 uses the same sampling schedule as the original Oral data of 11 timepoints, the others are variations on your 13 points for simulation.

here is the attached project with WNL classic models - I'm rather stuck this week with some other projects but this illustrates the concept if you want to try this with the Phoenix models to get the VIF for secondary parameters, to get VIFs in 6.2.x you have to have some number in the conc column - no good reason why, just the way it is currently. /extranet/media/kunena/attachments/legacy/images/sim2-20120411.jpg

Simon [file name=Dua_62.phxproj size=1733338]/extranet/media/kunena/attachments/legacy/files/Dua_62.phxproj[/file]

Attached Thumbnails

• 
• 

Attached Files

•  Dua_62.phxproj   1.65MB   0 downloads

[Richa Dua]

Hello Simon,

Thanks soooo much for your input. I wanted to know if I had access to the file you attached. It says that it is password protected. Please let me know whenever you get a chance.

Thank you..

[Richa Dua]

Hello Simon,

Thank you so much for helping with the project questions. I had one question about VIF. Is there any value that a VIF can have?

For example, I am interested in looking at AUC and VIF change in AUC estimates. The base model which has extensive sampling provides me with a VIF of ~ 2000. Is that too big of a number? Or do I compare the sparse sampling AUC VIFs against this number?

Second question would be about interpreting partial derivative plots. What do you look for in those plots that is indicative of a good or bad design.

Thanks.

[Simon Davis]

Hi,

  Like AIC in model fitting there is no specific value that is a 'good' VIF, you can only assess i by comparing different study designs on the same model. A lower VIF is better, the ratio between them could be an approximate estimate of how you would expect CV% to change for those parameters if you ran those real experiments.

 

For interpreting partial derivative plots I was taught to look at where the parameter line bore away from the central axis when fitting data. When the line was further away that parameter benefited most from samples in that time region.

  Simon.