Julia vs Phoenix WinNonlin (examples) [Software]
Hi Helmut!
Thank you for this great explanation and comparison! I didn't test Metida vs Phoenix, only vs SPSS and SAS. And as I know CSH and FA(2) are completely identical for data with 2 levels in random model (in this case G have 2x2 dim). And I think the difference in results with Phoenix for CSH and FA(2) is the issue (unfortunately it can't be submitted on GitHub). Also if we look at SPSS/SAS documentation we can find that rho optimized in '-1<rho<1' range. And it can't be more 1.0, but we have 'cshCorr_11 1.0005616', rly? It is difficult to make good optimization with Newton methods if parameters are limited, in Metida sigmoid function is used to change the optimization region for rho from -1:1 to -inf:+inf. In SAS and SPSS this feature works fine and I never get rho more than 1 and less -1. But maybe in Phoenix, this limiting work not so good if we see 1.0005616. From one side we can just look at REML values and get a better fitting, from the other - only the fact that we see covariance coefficient = 1.0005616 in the variance-covariance matrix leads that other computations and REML estimation have no sense.
In this situation maybe it is a good choice - never use Heterogenous Compound Symmetry in Phoenix.
If we see at point estimate and variance values for Metida and Phoenix with FA(0) - they are identical. But the minimal difference may be in DF estimation. I observe differences for DF in all software - between SPSS, SAS and Phoenix, sometimes it depends on data (especially if the covariance matrix is ill-conditioned). Maybe the cause is different approaches for derivating REML function, but it is only an assumption.
So also I can say that spatial covariance was added to Metida, and interface to any custom structure. Now I try to work on bootstrapping and multiple imputations for Metida. Maybe some other packages can be interesting:
ODMXMLTools.jl - experimental package for ODM-XML
MetidaNCA.jl - Julia NCA package
Thank you for this great explanation and comparison! I didn't test Metida vs Phoenix, only vs SPSS and SAS. And as I know CSH and FA(2) are completely identical for data with 2 levels in random model (in this case G have 2x2 dim). And I think the difference in results with Phoenix for CSH and FA(2) is the issue (unfortunately it can't be submitted on GitHub). Also if we look at SPSS/SAS documentation we can find that rho optimized in '-1<rho<1' range. And it can't be more 1.0, but we have 'cshCorr_11 1.0005616', rly? It is difficult to make good optimization with Newton methods if parameters are limited, in Metida sigmoid function is used to change the optimization region for rho from -1:1 to -inf:+inf. In SAS and SPSS this feature works fine and I never get rho more than 1 and less -1. But maybe in Phoenix, this limiting work not so good if we see 1.0005616. From one side we can just look at REML values and get a better fitting, from the other - only the fact that we see covariance coefficient = 1.0005616 in the variance-covariance matrix leads that other computations and REML estimation have no sense.
In this situation maybe it is a good choice - never use Heterogenous Compound Symmetry in Phoenix.
If we see at point estimate and variance values for Metida and Phoenix with FA(0) - they are identical. But the minimal difference may be in DF estimation. I observe differences for DF in all software - between SPSS, SAS and Phoenix, sometimes it depends on data (especially if the covariance matrix is ill-conditioned). Maybe the cause is different approaches for derivating REML function, but it is only an assumption.
So also I can say that spatial covariance was added to Metida, and interface to any custom structure. Now I try to work on bootstrapping and multiple imputations for Metida. Maybe some other packages can be interesting:
ODMXMLTools.jl - experimental package for ODM-XML
MetidaNCA.jl - Julia NCA package
Complete thread:
- Metida.jl: experimental Julia LinearMixedModel software PharmCat 2020-10-03 19:59 [Software]
- Metida.jl: validation in progress… PharmCat 2021-02-08 21:33
- user structure PharmCat 2021-02-18 18:44
- Validation results PharmCat 2021-03-05 14:17
- Validation results Weidson 2022-01-11 23:57
- Julia ≠ 🇷 Helmut 2022-01-12 01:02
- Julia ≠ 🇷 Weidson 2022-04-04 14:50
- Julia vs Phoenix WinNonlin (examples) Helmut 2022-01-13 15:18
- Julia vs Phoenix WinNonlin (examples)PharmCat 2022-05-20 17:54
- Julia ≠ 🇷 Helmut 2022-01-12 01:02
- Validation results Weidson 2022-01-11 23:57
- Validation results PharmCat 2021-03-05 14:17
- user structure PharmCat 2021-02-18 18:44
- Metida.jl: validation in progress… PharmCat 2021-02-08 21:33