I have a few questions for all, open to any input.
When selecting from a 2 to 3 compartment , and determining best model fit, is it standard to use the overall model comparison within NLME to compare the -2LL, Akaike and Bayesian Criterion.
First question: At what point is a drop in the -2LL worthy of justifying the addition of adding an extra "3rd" compartment. For example, is a 50 point drop meaningful?
Second question: After looking at the covariate plots, and manually selecting body weight on the theta parameters, when the decision to add in BW as a covariate do you add it on all of the PK parameters , i.e. V, CL, V2, CL2, V3, CL3, or do you add in only a select few? To make the model more physiologically meaningful, is it not best practice to include on all parameters? What about Ka, I see most of the time this is not reported. Any opinions?
Third question: I do see that after adding in the PD as an indirect response model , when I do add an effect compartment the -2LL and model criteria do improve. Again the same question, at what level is it justifiable to include this effective compartment as an additional delay to the PD.
How do you feel about adding in Baseline as a covariate on Kin. Does this make any sense to do this since of course baseline=kin/kout and we are expecting this. I do notice significant improvement in the fits when I do add baseline on Kin as a PD covariate into the model.
We do have a lot of internal discussion and debates about the justifications for 2vs3cmpt, indirect without and with an effective cmpt added for our compounds "class effect".
Thanks for your help!
When selecting from a 2 to 3 compartment , and determining best model fit, is it standard to use the overall model comparison within NLME to compare the -2LL, Akaike and Bayesian Criterion.
2 and 3 compartments models are nested models which means that you can define a fixed value for at least one the parameters for the big model (here the 3 compts) and come back to the 2 compt model. Here the flow between central and second peripheral compt is shut down (Q3=0) and you are back to the 2 compts.
The consensus is to use a decrease larger than 10 in -2LL as the threshold value, not AIC or BIC. It corresponds to a p value of about 0.001 and the idea is to be conservative when selecting the bigger model to overcome all other factors that can affect the true p value (model misspecification, multiple comparisons, non linear model making the chisquare only an approximation etc..).
The other tests like AIC and BIC just tell you that if you do not have any statistical test, then use the model with smallest AIC or BIC but you cannot put any approximate p-value to be used as reference.
second question: The answer to me is in between.
You can use the covariate search option we have implemented in run options where you select a priori all PK parameters with weight. The program uses the forward/backward covariate selection method which in general gives you a good estimate of the best covariate model.
On the other hand once you have too many PK/PD parameters, it is better to combine this approach with only a prior selection of all "making sense" PK/PD parameters that could be link to your covariate. The reason is that you always take a risk to get by chance the selection of a specific relationship when it does not make any sense.
Most of the time because of its link to physiology, V and Cl are assumed to be correlated to covariate like weight but it is not a rule to be used al the time. Understanding the mechanism may help making the decision to select certain PK parameters only before starting the covariate search. This applies also to how to use more than one covariate.
Taking all the possible covariate can lead to no sense conclusions, waste of time and numerical difficulties.
Third question: Adding an effect compartment adds one parameter that (ke0) that when going to infinity degenerates into the smaller model(here basic indirect response model)
Therefore, again the -2LL decrease of 10 would be to me the rule to use to select a keo based model.
baseline question: E0=kin/kout applies only to steady state conditions before you give the drug.
If you gave a new drug at t=0 but the system was still under the effect of the previous drug, then it would be possible to have Kin depending on the baseline. I do not have a good example for this but it should be possible.
You have also to make the decision is the baseline is a random variable or a covariate(you read the value from your data and it is assumed to be fixed). You cannot use it as both. This makes me thinking that using Kina s a function of the baseline may be a bad idea even for non steady state models.
You have a problem, try it first and then send to the forum. We will help you.
best Regards;
Serge
