WinNonlin Error in NCA AUC(0-infinity) Calculation
As a PhD biostatistician, I developed a keen interest in PK/PD and completed a course in Principles of Pharmacokinetics at UNC School of Pharmacy.
I tried to understand how WinNonlin calculates NCA PK parameters.?
While AUC(0-last) was straightforward, AUC(0-infinity) was mysterious, because I didn’t know how many data points WinNonlin uses to estimate lambda_z, which is vital for AUC(last-infinity) calculation.
Eventually, I could figure it out through reverse engineering using ADPC and ADPP, which includes NCA PK parameters from WinNonlin.
While doing so, I could find an error WinNonlin is making.?WinNonlin is using adjusted r^2 to choose the optimal n for lambda_z.?This is the Certara’s justification why they use the adjusted r^2, instead of r^2:
The author seems confused. The adjusted r^2 penalizes not the added data points, but the added explanatory variables.?
Basically, WinNonlin uses the adjusted r^2 over r^2 to use LESS data points when estimating lambda_z.?
However, my observations are just the opposite. r^2 criterion chooses the least data points in every case that I’ve ever encountered.
The following is a simulated dataset.
Using the adjusted r^2 criterion, lambda.z.n.points is 6,
but using r^2, it is 3, which is obviously less than 6.
????time? ? ? conc
1 ? 0.00 0.0000000
2 ? 0.25 3.3330979
3 ? 0.50 4.8341330
4 ? 1.00 7.5964924
5 ? 2.00 7.8812186
6 ? 3.50 7.3032869
7 ? 5.00 6.5859969
8 ? 7.00 5.9507540
9 ? 9.00 5.2168118
10 12.00 3.7971380
11 17.00 2.7493894
12 24.00 1.5989055
13 32.00 0.7595756
Biostatistician and Pharmaceutical Scientist
4 个月An interesting discussion for sure that I think really highlights the role of biostatisticians and pharmacometricians. I think making appropriate use of the other lambda_z acceptance criteria provided by WinNonlin or using the Best Fit method and specifying an appropriate Start Time Not Before rule negates the concern you raise about including concentrations impacted by ongoing absorption. As with any analysis, it needs to be specified appropriately.
Pharmaceutical Executive
4 个月Ok but aren’t there more important things to solve?
Executive Consultant and Vice-President, Strategic Consulting
4 个月In general, I think the use of r-squared (adjusted or not) is more for historical reasons than actually being a good method. The best method would be to compute the mean squared prediction error. This could be implemented in WinNonlin easily. An old paper by Sheiner and Beal (J. PK Biopharm 1982;9:503-512) explains why r-squared is not a good way to measure fit.
Biopharmaceutics Scientist at Certara UK (Simcyp Division)
4 个月I agree with the comments above.. the idea is to not select more or less number of points, but the correct number of points which are relevant to the terminal phase. Which is a complex task as the each and every profile is different and the ideal way is to visually look at the profile and manually select the points, but to do it manually for each PK profile is very time consuming, thats where winnonlin comes in does it properly, there by removing human error aswell.. once the NCA is done user have the option to look at the points the software has used, perhaps change it if user isn't happy.. in all no, it is not an error in winnonln.
Director Clinical Pharmacology at BioNTech
4 个月I think the statement that there is an "error" in winnonlin is misleading. The algorithm that is used is clearly stated in the user manual and correctly implemented in the software: (https://onlinehelp.certara.com/phoenix/8.3/responsive_html5_!MasterPage!/WinNonlin%20User%27s%20Guide.pdf) - go to page 134. The particular algorithm implemented in winnonlin is one of undoubtedly many that could be used. Consistent, repeatable selection criteria which penalizes inclusion of data points far from the regression line (ex. distribution phase data) while maximizing the selection of terminal phase measurements are ideal. The adjusted R2 algorithm that is used is adequate for this purpose as it rewards use of additional measures while penalizing points that are far from the regressed line. I think improving on this algorithm is an interesting topic (Nathan Teuscher has commented within this thread). Perhaps there may be some small advantages with certain types of data to use different methods? But overall, the method that is used is generally acceptable, explicitly stated in the manual, and correctly implemented in the software.