Logit transformation and covariate model

Dear NM users, I am running an extravascular pop PK model with dual absorption and a logit scale to define F and its IIV. I would like to test some covariates on F. Has anyone tested covariate models (categorical and continuous) through logit transformation? Kind regards Marie

Dear Marie, If you run the SCM via PsN it is easy (but still quite error prone, so I include some instructions, below). In the scm configuration file you list the parameter to be on logit scale, and then define the code for testing a continuous covariate as linear, additive (same for categorical, but then it would be different code for additive, and not -5, since that is for continuous exponential). For testing covariates on other parmameters (not on logit scale) you do not need to define the code yourself, in the configuration file, since SCM will handle per default. First of all, SCM tests covariates on the typical values and expects a variable name starting with TV for the logit-transformed variable (so not TVF1, because that is a fraction between 0 and 1). In the example below this is the variable TVPHI defined in the nonmem control stream, just below. If you also have IIV on the fraction for each absorption route, you can of course rename in the below code, from TVPHI to e.g. TVFLGT, or whatever you want, as long as you use the consistent parameter name in the scm config file. Likewise, my example has F1 as the fraction, but you may have a variable which is the total F for the two routes of absorption. So you may be using an entirely different variable name (than TVF1 and F1), which is fine. TVF1 = THETA(6) ; Typical F (0-1) TVPHI = LOG(TVF1/(1-TVF1)) ; Typical F on logit scale ;TVPHI = PHICOV+TVPHI ; Delete this line (not for you to include in control stream, will be added by SCM), I only included for illustration of what additive looks like F1 = EXP(TVPHI + ETA(5))/(1 + EXP(TVPHI + ETA(5))) ;Individual F1 (0-1) Here is an example of what the configuration file may look like, for a continuous covariate, where I assume zero or negative values (e.g. -99) codes for a missing covariate value. Adapt if not suitable in your case. The below code assumes that (multiplicative) exponential relation is the default for a continuous covariate when tested on other parameters (adapt code if that is not what you are using, the -5 needs to be changed), which for this parameter (TVPHI) is changed to (additive) linear. Upper and lower bounds, I have just arbitrarily set to one. Should be adapted to the range of covariate values. Finally, add similar code for categorical covariates (PHI:*-2). SCM does this all automatically for testing covariates on other parameters, but not when you define your own code, as below. logit=PHI …. [code] PHI:*-5=IF(COV.GT.0) THEN \ PARCOV= THETA(1)*(COV - median) \ ELSE \ PARCOV= 0 \ ENDIF \ [lower_bounds] PHI:*-5=-1 [upper_bounds] PHI:*-5=1 [inits] PHI:*-5=0.001 You can find more support on using the PsN scm here: https://uupharmacometrics.github.io/PsN/docs.html https://github.com/UUPharmacometrics/PsN/releases/download/v5.6.0/scm_userguide.pdf Best wishes Jakob Jakob Ribbing, Ph.D. Principal Director, MIDD Consultant & Senior Expert www.pharmetheus.com 

It can be done the same way as for any other parameters if you add covariates to the logit rather than to F. Interpretation is slightly more difficult as the effect needs to be converted back to the F scale to be useful, but technically, there is no difference with untransformed version Thank you Leonid