1 / 23

Nonparametric (NP) methods: When using them? Which method to choose? Julie ANTIC

Nonparametric (NP) methods: When using them? Which method to choose? Julie ANTIC and advisors: D. Concordet, M. Chenel, C.M. Laffont, D. Chafa ï. A too restrictive normality assumption. • Usual population PK/PD studies assume normality of ETA.

laith-bates
Download Presentation

Nonparametric (NP) methods: When using them? Which method to choose? Julie ANTIC

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Nonparametric (NP) methods: When using them? Which method to choose? Julie ANTIC and advisors: D. Concordet, M. Chenel, C.M. Laffont, D. Chafaï

  2. A too restrictive normality assumption • Usual population PK/PD studies assume normality of ETA • But the true distribution of ETA may be more complex! Parametric estimation (normal) True distribution bimodal asymmetric heavy-tailed ETA

  3. How to detect departures from normality? • If ETA-shrinkage is low Parametric estimation (normal) Empirical Bayes Estimates (EBEs) True distribution ETA

  4. How to detect departures from normality? • But if ETA-shrinkage is high, EBEs can be misleading [Karlsson & Savic, 2007] Parametric estimation (normal) Empirical Bayes Estimates (EBEs) True distribution ETA

  5. A possible solution: NP methods NP method = estimates an increasing number of parameters with N (N= number of individuals in the sample) → for large samples, a lot of distributions are available! → no restrictive assumption on ETA distribution

  6. Several NP methods • Some discrete NP: - NP-NONMEM [Boeckmann & al., 2006] - NPML [Mallet, 1986] - NPEM [Schumitzky, 1991] - others: NP adaptative grid, extended grid… • Some continuous NP: - SNP [Davidian & al., 1993] - others: splines, kernels… frequencies support points

  7. Discrete NP Without assumption on ETA distribution, the MLE is (MLE = the maximum likelihood estimator) • discrete with at most N support points [Lindsay, 1983] → the likelihood is explicit ! • consistent[Pfanzagl, 1990] frequencies support points

  8. How to compute the discrete NP-MLE? frequencies NP-NONMEM [Boeckmann & al., 2006] • support points = EBEs • frequencies maximize the likelihood NPML [Mallet, 1986] and NPEM algorithm [Schumitzky, 1991] • increase the likelihood at each iteration • by modification of support points + frequencies • here implemented - using NP-NONMEM as starting point - in C++ - more details in [Antic, 2009] support points

  9. Smooth NP (SNP) SNP [Davidian & al., 1993] • = the MLE over a set of smooth distribution with density = polynomial² × normal density • examples • the degree of the polynomial increases with N • consistent [Gallant & al., 1987] density(ETA) = (1)²×exp(-0.5×ETA²)/√(2×PI) density(ETA) = (0.2+ETA)²×exp(-0.5×ETA²)/√(2×PI) density(ETA) = (0.3-0.4×ETA-0.6×ETA²)²×exp(-0.5×ETA²)/√(2×PI) density(ETA) = (0.9+0.06×ETA+0.06×ETA²+0.06×ETA3)²×exp(-0.5×ETA²)/√(2×PI) Normal distribution Asymmetric distribution Bimodal distribution Multimodal distribution

  10. Comparison of NP methods • several simulation studies:

  11. Details on the PK scenari Slow-metabolisers sub-population volume volume clearance clearance

  12. Details on the PK/PD scenario Non-responder sub-population baseline + disease progression(linear with time) baseline baseline + disease progression – effect (Emax model with effect compartment) 1 year time Effect at 100 days for a median AUC

  13. Simulation studies strategy • Strategy: for each scenari, repeat 100 times Dataset simulation with non-normal ETA Parametric estimation assuming normal ETA → estimation of residual variance , EBEs SNP nlmix code [Davidian & al., 1993] NP-NONMEM fixed NONMEM VI [Boeckmann & al., 2006] NPML (after NP-NONMEM) fixed implemented in C++[Antic & al., 2009] NPEM (after NP-NONMEM) fixed implemented in C++[Antic & al., 2009]

  14. Comparison of NP methods T1-distance True distribution Estimated distribution • T1 distance Estimated cumulative distribution function True cumulative distribution function ETA • Graphical inspection of marginal distributions Mean of estimated distributions

  15. ETA-shrinkage ~ 9%; PK IV bolus EBEs NP-NONMEM NPML (after NP-NONMEM) NPEM (after NP-NONMEM) SNP T1-distance Parametric EBEs and NP methods are roughly equivalent All methods seem consistent 0 N 50 100 200 300 400

  16. ETA-shrinkage ~ 9%; PK IV bolus clearance clearance clearance clearance clearance clearance N=200 TRUE EBEs NP-NONMEM NPML (after NP-NONMEM) All methods generally allow suspecting a departure from normality NPEM (after NP-NONMEM) SNP

  17. ETA-shrinkage ~ 34%; PK IV bolus EBEs NP-NONMEM NPML (after NP-NONMEM) NPEM (after NP-NONMEM) SNP T1-distance Parametric EBEs consistency is very slow! Only slight differences between NP methods N 50 100 200 300 400

  18. ETA-shrinkage ~ 34%; PK IV bolus; EBEs seem misleading clearance clearance clearance clearance clearance clearance N=200 TRUE EBEs No clear difference between NP methods NPML (after NP-NONMEM) NP-NONMEM NPEM (after NP-NONMEM) SNP

  19. ETA-shrinkage ~ 31%; PK oral EBEs NP-NONMEM NPML (after NP-NONMEM) NPEM (after NP-NONMEM) SNP T1-distance EBEs seem not consistent! NP-NONMEM is not as good as the other NP methods N 50 100 200 300 400

  20. ETA-shrinkage ~ 31%; PK oral; EBEs seem misleading clearance clearance NP-NONMEM seems biased clearance clearance clearance clearance N=300 EBEs TRUE NPML (after NP-NONMEM) NP-NONMEM SNP NPEM (after NP-NONMEM)

  21. ETA-shrinkage > 40%; PK/PD NP-NONMEM and NPML poorly detected the subpopulation Only NPEM and SNP appear to detect the non-responder sub-population EBEs NEVER detect the non-responder subpopulation TRUE EBEs 25% 25% Drug effect Drug effect NPML (after NP-NONMEM) 25% NP-NONMEM 25% Drug effect Drug effect 25% 25% NPEM (after NP-NONMEM) SNP Drug effect Drug effect

  22. Conclusion • EBEs are misleading when ETA-shrinkage is high (>30%) • NP methods appeared to be a good solution (with reasonable computation times) • Our recommendations: - use NP-NONMEM - easy to implement in NONMEM - quite fast to compute + a more advanced NP method (especially if ETA-shrinkage > 40%): ex. NPEM, SNP…

  23. To learn more on NP, go and see: • poster 107 [Comets, Antic & Savic] • poster 105 [Baverel, Savic & Karlsson] • poster 133 [Goutelle, Bourguignon, Bleyzac & al.] • poster 29 [Jelliffe, Schumitzky, Bayard & al.] • MM USC-PACK software demonstration [Jelliffe, Schumitzky, Bayard, & al.] Thanks for your attention.

More Related