9 replies to this topic

[joybaker]

Hi all,

 

I would like to ask how to make the choice between covariate vs inter-occassional variations in a cross-over food study? 

This is a study where each subject underwent fasted/low fat/high fat (food_num=1/2/3) stages. I used covariate (on Ka/Cl) and IOV (on Ka/Cl), and IOV model is better (AIC much better, not a little). However IOV treats each run as an orderless run for each subject, and covariate model more likely to say that food states affect parameter with a direction. So though statistically I tend to use IOV, how could I make a conclusion? Graphically you can still tell that in fed state the AUC and Cmax are higher than fasted, so it is hard to draw a conclusion that the IOV masks any effect generated by food. (mixed model also performs worse than IOV alone)

 

Q1

How can I value the IOV? Say "from the model IOV can explain 20% of the variation" etc. As it is not reported in the Theta table like covariates. I can only see the results from posthoc, but I cannot have a general idea how big it is?

 

Q2

Another question comes from this example, how could I explain the significant effect towards Cl (or Cl/F)? We usually think that food effect should be put on absorption(Ka/Tlag etc), as when the API comes in the system, the clearance is still its clearance. 

 

I have attached the file with 5 subjects for an easy example.

Attached Thumbnails

• 

Attached Files

•  123FE.phxproj   2.41MB   62 downloads

[smouksassi1]

Unless ou model the F your parameters will be CL/F V/F . you can recode your model to have covariate and or OCC effect on your Frel.

 

IOV is a variance component and can be described by a variance or % CV of between occasion variation

[smouksassi1]

Q1

IOV is part of the OMEGA you get the info in it: you get a small variance for tlag and ka while for CL you have:

0.26846402  --> sqrt(0.27 )~ 50 % IOV

 

Q2 you can use IOV to diagnose time varying effects which in our case here is food status ( changes within subject after each reset)

so instead of using the food label we use a occasian lable 1,2,3 ideally the food status order is not always the same to be able to differentiate between sequence, carryover and food effect

IOV will show you that the effect is not random ( plot of random effect of occasion versus occasion) and then we cannot consider it as random variation rather it is systematic and we need a theta (fixed effect for it)

 

 

 

 

Attached Files

•  123FE_SM.phxproj   2.57MB   58 downloads

[joybaker]

Hi,

This case is slightly special, as the sequence is always fasted-high fat-low fat (as you can see from the figure in the first post). So I'm trying to understand once I assign the covariate affect to the parameter (eg Cl), are there anything left that can be explained by IOV (which can be orderless), or which one takes a dominate role? Thus the two food columns can be used for either food effect and sequence effect or both as they are colinear.

 

So I build a covariate model/IOV and a mixed model. The "XY Plot check Cl" showed that generally Cl is higher in fasted, tend to say covariate takes a dominate role that food decreases Cl. However if I compare the AIC of three models, the IOV model is always better (even better than mixed model). So which model should I choose? 

(No. 16/17 might be tricky example, but the rest of the subjects have some trend that seems to be roled mostly by covariate rather than IOV)

 

Q1

IOV is part of the OMEGA you get the info in it: you get a small variance for tlag and ka while for CL you have:

0.26846402  --> sqrt(0.27 )~ 50 % IOV

 

Q2 you can use IOV to diagnose time varying effects which in our case here is food status ( changes within subject after each reset)

so instead of using the food label we use a occasian lable 1,2,3 ideally the food status order is not always the same to be able to differentiate between sequence, carryover and food effect

IOV will show you that the effect is not random ( plot of random effect of occasion versus occasion) and then we cannot consider it as random variation rather it is systematic and we need a theta (fixed effect for it)

 

 

Attached Thumbnails

• 

Attached Files

•  123FE_SM_z.phxproj   2.61MB   64 downloads

[joybaker]

Hi, is it possible to make VPC stratified by IOV parameter?

I tried once but the three VPC looks the same, so IOV was not accounted.

Thanks

Jo

 

Q1

IOV is part of the OMEGA you get the info in it: you get a small variance for tlag and ka while for CL you have:

0.26846402  --> sqrt(0.27 )~ 50 % IOV

 

Q2 you can use IOV to diagnose time varying effects which in our case here is food status ( changes within subject after each reset)

so instead of using the food label we use a occasian lable 1,2,3 ideally the food status order is not always the same to be able to differentiate between sequence, carryover and food effect

IOV will show you that the effect is not random ( plot of random effect of occasion versus occasion) and then we cannot consider it as random variation rather it is systematic and we need a theta (fixed effect for it)

 

[smouksassi1]

it should work with latest phoenix

it used to be a bug where vpc only kept first value of time varying covariates (OCC is by definition a time varying one).

it might be good to have another copy of OCC in your dataset to use it for plotting or other splits that you might need

 

[joybaker]

That's very helpful to know. I use v8.1, maybe that's why. Is it possible to report typical value (Cl) in three occasions?

 

Since Cl = tvCl * exp (nCl + nClx0(FE==1) + nClx1(FE==2) +nClx2(FE==3) ), but nClx0 =nClx1 =nClx2, the typical values calculated will be the same for 3 occasions. But this is obviously not true.

 

it should work with latest phoenix

it used to be a bug where vpc only kept first value of time varying covariates (OCC is by definition a time varying one).

it might be good to have another copy of OCC in your dataset to use it for plotting or other splits that you might need

 

 

[smouksassi1]

you can ask for the table at different values if your time varying covariate and then you can filter do further computation to compute mean Cl by occasion 

 

when you use it as a fixed effect covariate you can compute it from your equation:

stparm(Ka = tvKa * exp(dKadFOOD2*(FOOD==2)) * exp(dKadFOOD3*(FOOD==3)) * exp(nKa + nKax0*(FOODNUM==1) + nKax1*(FOODNUM==2) + nKax2*(FOODNUM==3)))

 

 

ka food1 = tvka 

ka food2= tvKa * exp(dKadFOOD2)

ka food3= tvKa * exp(dKadFOOD3)

Attached Thumbnails

• 

[joybaker]

Thanks, I think a posthoc table can do this as well. I used to think it can report a population value similar to a fixed effect covariate （as typical value might not equal to the mean/median value）, but your clarification helps.

 

The second suggestion, I think its a choice between IOV and IIV? In this example IOV gives better OFV , I think it is a compromise the way I can find the typical values by mean.

 

 

 

you can ask for the table at different values if your time varying covariate and then you can filter do further computation to compute mean Cl by occasion 

 

when you use it as a fixed effect covariate you can compute it from your equation:

stparm(Ka = tvKa * exp(dKadFOOD2*(FOOD==2)) * exp(dKadFOOD3*(FOOD==3)) * exp(nKa + nKax0*(FOODNUM==1) + nKax1*(FOODNUM==2) + nKax2*(FOODNUM==3)))

 

 

ka food1 = tvka 

ka food2= tvKa * exp(dKadFOOD2)

ka food3= tvKa * exp(dKadFOOD3)

 

[Simon Davis]

Also next month Phoenix 8.4 will be released that has quite a few NLME enhancements.