population PK modelling of very sparse data

Dear all, I would like to seek some advice from you regarding population PK modelling of very sparse data. I'm trying to fit a population PK model to a set of very sparse data. There are about 700 subjects in the dataset. The intention was for each of these subjects to self-administer daily doses for 7 days (loading phase) followed by weekly doses for 10 weeks (maintenance phase). For each subject, I've zero, one or two concentration measurements of the parent drug and its major metabolite taken at least one week after the final dose. In addition, I've information regarding which doses, if any, were missed by the subjects (i.e. I know each subject's adherence to the dosage regimen). BQL values are present in the data set, and comprise about 15% of all data. In the literature, a two-compartment model for the parent and a two-compartment model for the metabolite (including one compartment for the depot compartment) has been suggested. However, because of my overall data sparseness, NONMEM was not able to produce a successful two-compartment model. This is so even after I've fixed Ka, intercompartmental clearances for both the parent and the metabolite, as well as the parent drug's metabolic clearance to the metabolite (fixed at 15.2% of the total clearance of the parent drug). After repeated model iterations, the best performing model to date is a one-compartment model for the parent drug and a one-compartment model for the metabolite. Ka and the parent drug's metabolic clearance to the metabolite were fixed. CL, V(parent drug comp), CL(metabolite) and V(metabolite comp) were estimated. IIV was estimated for CL and CL(metabolite). I log-transformed the data and used the M3 method to account for BQL values. RUV is exponential error (additive in the log scale). In addition, the model was more stable after I've incorporated allometric scaling (by weight) to CL, V(parent drug comp), CL(metabolite) and V(metabolite comp). Although this is the best performing model, it is still not optimal because of its poor prediction of high concentration values for the parent drug and metabolite. Could you request for assistance on how to improve this model? Thank you and best wishes, Kok-Yong Seng, PhD DSO National Laboratories Singapore