I am curious as to why NLME sets a diagonal omega matrix instead of a full block omega matrix as the default when first attempting to converge a population PK model. Wouldn’t it be best to first attempt converging a model that has a full block omega matrix (aka off diagonal covariances estimated as well as variances) and then check the estimated covariance values to confirm that there is negligible (say less than 0.0025) instead of setting a diagonal omega matrix from the start?
Dan
Hi Dan,
I guess there are different approaches that people take. I think the general design of the Phoenix model object was to guide newer users try to follow the general Rule of Parsimony, i.e. Choose the simplest model that fits the data and addresses questions of interest.
So yes it’s an assumption that by defaulting a diagonal omega matrix we are suggesting all random effects independent but if the model fits the data well enough then that means less parameters to be adjusted so a simpler model. As you note it’s easy enough to change and test subsequently.
Simon.
I would speculate also that this is the default for historical reasons and because the default engine is FOCE-ELS which might have issues fitting full omegas and or run times can be very long.
Depending on your situation I would indeed go into full omega and a more modern EM based engine to start with.
Indeed we always strive for a parsimonious model but keep in mind the intended use of your model.
Reduce your model but not beyond the intended use of it
If your data does not support a full omega and you want to simulate from your model you might end up simulating un-realistic combination of the independent etas. Better make a strong assumption about a value correlating Vc/CL for example.