Standard Error in ka with Decreasing Data

It's probably not NONMEM, but the model misspecification associated with use of a first-order absorption model. If, say, everyone had their tmax at the same time, and everyone had the same elimination half-time, then ka would be perfectly determined from data only at tmax and later. Now if you add some weird concentrations earlier than tmax, they don't fit, and cause the estimate of ka to be less secure .. that's an idea, anyway ... of course one wants to look at the whole fit. There should be graphical evidence that the absorption model does not fit the early points for my explanationto be valid ...

I have a situation where I think the more experienced users of this list may be able to help. I am working with a drug that has variable absorption with a Tmax at about 4.5 hours. I am trying to decide whether it is useful to collect any samples at time points prior to the Tmax or rather to maximize the amount of useful data by collecting samples only at times past Tmax. Using some multiple dose data and a one compartment model (Advan2, Trans2), I analyzed steady state data following the last dose on day 18, times 0 to 72 hours. The standard error on Ka is 0.0746. Now if I remove the data from time points collected between 0 and 6 hours (all data prior to Tmax), the standard error on Ka is reduced to 0.0245. This is not what I expected. I reasoned that if the data is missing over the period of absorption, the error on the parameter Ka should increase. What does NONMEM do that makes my intuition incorrect?