Subject: Transformations of data

Date: Wed, 27 Jun 2001 14:02:29 -0400

Could somebody clarify this query?

I am analysing a data set in NONMEM in which I was trying to arrive at better data fits and explanation by using nontransformed as well as transformed data. I then examined various plots using XPOSE (DV vs PRED, DV vs IPRED, WRES vs TIME, WRES vs PRED etc). Although DV vs IPRED is known to be better than DV vs IPRED I was observing that DV vs IPRED was the best when I was using a square root transformation of the data. The values of the parameter estimates were also good. ALso the spread of the residuals etc was much more symmetrical when I am using a square root transformation of the data in comparison to log and nontransformed data. In case of log transformation the IPRED showed more curvature as the DV was increasing (I have 4 to 5 fold differences in DV ie., comparing the lowest to highest value). This I did not observe while I was using a square root transformation. The best predictions (individual) I observed were for square root transformed data. I would be happy if somebody could explain me how the distributions etc get better when we use log transformation and square root transformation of data. Also how should the plot of absolute values of IRES vs IPRE look?

I did analysis with FO and FOCE with interaction. But model diagnostics wise FO was much better. Could somebody tell me what is happening statistically?

Date: Wed, 27 Jun 2001 14:47:37 -0400

Subject: Transformation of data

I think I need to add this information to my previous question:

I used proportional plus additive model for untransformed data while I used a simple additve model for transformed data (both log and square root) in the error structure.