If you are using a composite profile to generate PK parameter estimates for NCA, do you use data where one or more animals is missing data? For example, if you have the following samples available (missing samples BLQ)
Time (h) 0 1 2 3 4 5 6 7 8
N 3 3 3 3 3 3 3 2 1
Should you exclude the 7 or 8 hour timepoints as there is data missing, or should you include it? The Sparse sampling option for NCA includes N and SEM estimates in the output, but is it still appropriate to use the later timepoints in lambda z ranges?
Hi John, I don’t know of any regulatory guidance specific to this, however my personal thought is that with such small samples I think it’s hard to apply a rule to remove them from Lz automatically. Do you think they are affecting the Lz calculation a lot?
If so you might want to consider presenting a second analysis with e.g. the 8h sample excluded from Lz calcs and cover this in the discussion etc.
They could be causing the AUC to be overestimated if they are excluded as the remaining point could be an outlier. it’s a bit of a tricky subject. The consensus from PharmPK was to include the observed concentration data, but there are a number of methods to account for missing samples, all of which have positives and negatives (setting as 0, setting as 0.5*BLQ etc).
As you might imagine I disagree with the PharmPK consensus as I think the predicted value Clast is a better value to extrapolate from (and remember if the observed point is below the regression line it will cause an underestimation).
I thought the case for predicted Clast expounded by Helmut much more convincing than relying on the guidance that is only that and not IMO up todate with what most software can easily give now. I’ve also used Predicted since I was back at Covance in 2000 and since we stated up front in our SOP/protocol this is the value we would use we never got any problems to my knowledge.
Yeah, I have a preference for predicted values as well. The guidance is a bit ambiguous to be honest. I’ve used the predicted values since I started doing NCA. The results are likely to be very similar too unless the terminal phase is poorly characterised.
data sets like John’s are not uncommon in TK / serial sampling – though I would prefer the magick n=6.
Some thoughts:
IMHO, setting to zero doesn’t make sense. After a dose zero is exactly the only value we know to be false. Would give an AUCt similar to PHX/WNL’s AUCall, which I don’t like at all. Might lead in an extreme situation to AUC∞ < AUCt!
½LLOQ is a relic – old beliefs die hard. It would be only justified if the time interval after Clast ≈ t ½. How likely is that? In PopPK it was shown that ½LLOQ gives biased estimates. There is hope: I know of a group working on a NCA-method for unbiased estimating AUC∞ in serial sampling (taking missing samples and BQLs into account).
My procedure (arbitrary, of course): Drop data points if Nmiss ≥⅓N. In your case I would keep the 7h and drop the 8h. Still positive bias; I know, I know.
@admins: Would you please change the tag “EMEA” (globally in the data base?) to “EMA”? The European Agency was renamed back in 2005.
