Re: End of semester MCQ and short answer question

From: "Leonid Gibiansky" leonidg@metrumrg.com Subject: Re: [NMusers] End of semester MCQ and short answer question Date: Sun, July 17, 2005 2:14 pm Nick, Thanks for the reference, results seems very reasonable: if you ignore zeros, predicted concentrations may decay not as fast as they should (leading to lower CL, higher V). However, this might be strongly related to the design and sampling. In my examples, fraction of BQLs was relatively small (less than 5%), most zeros were very suspicious (related either to the non-compliance or data errors) because LLOQ was 3-4 orders of magnitude less than Cmax while sampling points were not that far from the dose to warrant zeros. With the good design, fraction of BQLs in the data set is small, and efforts that are needed to include those are not warranted by the gain that you may get from inclusion of those points. As to the true values, you may be forced to use special segment of the error model to account for the BQL measurements. This can be very similar to LLOQ/2 imputation with BQL variance fixed at (LLOQ/2)^2. My justification of the idea to ignore zeros (or not to use "true" values) is that we are not interested in the very fine details of the PK behavior (the deaper you look, the more compartments you may descover) and restrict the model to the range of concentrations that are relevant to your problem. Sampling should also be consistent with the goal of the study: samples should be located at characteristic points of the profiles (while BQLs are definitely not in that range/position). One can imagine situations when zeros are important: for example, if there is a sub-population with much higher CL: ignoring zeros may hide the fact that concentration decay for a sub-population is much faster than was expected, especially if the sample timing was not designed for the sup-population with the high CL. But even in this case, BQL values may help you to discover the problem with your design, but will not help to build the correct model: it is difficult to build correct model based on the very noisy measurement. If you do not have sufficient non-BQL time points to define the terminal phase and impute some BQL values (or use the true value that will be a reflection of the random noise generated by the instrument), the model for the subpopulation will be defined by the timing of the sampling point rather than by the actual CL. If you have sufficient number of non-BQL time points to define the terminal phase, ignoring the BQLs should not adversely affect the model. Leonid