RE: Indirect response model

This is another recent thread that I would like to comment on. On the issue of LOCB vs LOCF: The question is discussed at great length in NONMEM Users Guide VI, PREDPP, in chapter III, section B.2 Time- Varying Concomitant Variables. The idea is that, prior to advancing the state vector from time t1 to time t2, the PK routine is called (i.e., $PK is evaluated) with the next record (time t2), so that the values computed by PK for the record with TIME=t2 are used during the advance from t1 to t2. This is easily justified. If the values on the record with t1 were used, the values (other than TIME) recorded on the last record in the data set would never be seen by PK, and could not enter into the model. Linear interpolation such as you are discussing could not be carried out properly. As it is, the first data record is always seen by PK (because there is always a call to PK with the first data record of the individual record), and subsequent data records are seen prior to the advance to that record. The linear interpolation suggested by Jakob could also be carried out in $PK without any change to the data set. I think that I recall some of Lewis' fellows doing similar interpolations in the past. The following code is a suggestion. I have not tested it!! $PK KFOR = THETA(1) KCL = THETA(2)*EXP(ETA(1)) IC50 = THETA(3) IMAX = THETA(4) F1 = KFOR/KCL IF (TIME.EQ.0) THEN PCOP=0 PTME=0 SLOPE=0 OCOP=0 ELSE SLOPE=(COP-PCOP)/(TIME-PTME) OCOP=PCOP ENDIF PCOP=COP PTME=TIME $DES ICOP=OCOP+SLOPE*(T-PTME) COEF = IMAX*ICOP/(IC50+ICOP) DADT(1) = KFOR-KCL*(1-COEF)*A(1) On Thu, 17 May 2007 11:52:54 +0200, "Jakob Ribbing" <[EMAIL PROTECTED]> said: > Dear all, > > I agree, implementing the actual PK model (i.e. IPP) or using linear > interpolation to describe the individual PK profiles as suggested by > Juergen > and David is necessary in this example. > > However, for understanding when input concentrations can be used directly > as > a reasonable approximation I have the following question: If the input > concentrations were to be used, as originally presented by Nishit, > wouldn't > nonmem use the input concentrations according to last observation carried > BACKWARD, rather than forward? In that case, for ID 1101 (the third > individual in the dataset): > 0 < TIME <= 840 -> COP=0.13884 > 840 < TIME <= 842 -> COP=0.12987 > 842 < TIME <= 844 -> COP=0.12147 > 844 < TIME <= 1848 -> COP=227.79 > ... > I haven't checked myself that nonmem uses LOCB rather than LOCF, but have > been told so by Mats Karlsson which usually makes checking superfluous:>) > I > think this could be important in other situations as well: The LOCB-rule > could induce false covariate relations if the drug affects a potential > covariate. For example, disease level/score/stage may falsely appear to > affect the drug clearance if investigated within nonmem. To properly > quantify such a covariate relation a simultaneous fit of the PK-PD model > may > be necessary, treating the covariate (biomarker) as an integrated part of > the model. Getting back to this thread: > > Nishit, if using nonmem version 6, the dummy dose into compartment 1 is > not > needed for initializing the baseline. Dummy-dosing records can be removed > from the datafile and this line in the model file: > F1 = KFOR/KCL > can be replace by > A_0(1) = KFOR/KCL > > Lastly, either fit log-transformed DV:s using an additive error model OR > add > "INTERACTION" on the $ESTIMATION, to account for the interaction between > ETA1 and EPS1 in the current error model. I would prefer the first > alternative for several reasons. > > Kind regards, > > Jakob > > $INPUT ID TIME DV AMT=DOSE COP MDV > ID 1101 in the dataset: > > 1101 0 . 1 0 1 > 1101 0 68.1 . 0 0 > 1101 840 88.1 . 0.13884 0 > 1101 842 105.5 . 0.12987 0 > 1101 844 108.8 . 0.12147 0 > 1101 1848 113.3 . 227.79 0 > 1101 1850 62.6 . 379.54 0 > 1101 1852 138.7 . 412.18 0 >