I am working on a population PK model. If I know that there are two parameters (CL and V) have a strong correlation between each other. How can I code in the PML model structure to fit the correlated coefficient? How to add co-variances between this two random effects?
Thanks,
Linda
Linda, note that you seem to be confusing 2 issues (correlations for fixed vs random effects). Yes, generally the structural values of Cl and V will be highly correlated. The model accounts for that. That is, the omega matrix automatically includes correlations for the fixed effects. The bigger issue is, are the etas for Cl and V correlated? PHX outputs help you assess this. If they are indeed correlated, you can specify a non-diagonal matrix for the random effects on the random effects tab. In my experience, in most cases. The Cl and V etas are not correlated enough to make a “significant” difference in the fit of the model.
Dan
Thank you for your response, Dan.
It is just use CL and V as an example. How should I assess whether etas for CL and V are correlated or not from PHX outputs? And how should I specify it bying coding in PML model structure?
Linda
Linda,
The model’s “Output data” includes a file of the etas, and is appropriately named “Eta”. Just do an x-y plot of the etas for V and Cl to visually assess if they appear correlated. You can also select “Generate Regression” - “LInear” to add Rsq to the plot. The correlation coefficient is the square root of the Rsq value. In you do decide to include a correlation for the etas when fitting the model, you can do so by clicking on the “Parameters” tab, and then click on the “Random Effects” tab and then unselect the “Diag” option. If you do fit the model with and without “Diag” option, you can then compare the outputs of the 2 models (-2LL, AIC, shrinkage, goodness of fit plots, etc.) to determine if adding the correlation improved the fit.
Hope that helps,
Dan