Hi Yuma, My experience is that some model modifications can greatly reduce objfn but make prediction actually worse. I like to use repeated 2-fold cross-validation since I am usually interested in accurate predictions for out-of-sample data. This may or may not be what you want your model to do. Once you have decided what you actually want your model to do then test for whatever that thing is along with objective function, accepting into your model what improves both measures.
Look closely as to why some individuals get higher residual error. You can put it into or omit it from your model but you should have a good reason why. Do you trust the doses? Are there outlier data? Are all the covariates correct? Did people simply write down incorrect things? Look at the individuals who get assigned high residual error. Are the data reasonable? Did somebody write down a wrong weight or age or height or BMI? One danger is that you mask model misspecification with IIV on residual error. If residual error correlates with say, obesity and your model works poorly in the obese then you get improved model fit to the data by effectively reducing the impact of obese on the model fit by assigning them higher residual error. You dont want to mathematically reduce the impact of those individuals that demonstrate real shortcomings of the structural model. Warm regards, Douglas Eleveld ________________________________________ From: [email protected] [[email protected]] on behalf of Y.A. Bijleveld [[email protected]] Sent: Friday, January 16, 2015 3:09 PM To: [email protected] Subject: [NMusers] IIV on res error Dear all, I am modeling multi-center log-transformed neonatal data and have constructed a 2-compartment model with ETA’s on Cl, V1 and V2. However, when introducing interindividual variability on the residual error the MOFV drops >150 points, while previously significant relationships between clearance and covariates disappear. I find it strange that the introduction of the IIV has such an impact and don't fully understand. I have already checked the data for (extreme) outliers. Can anyone shed some light? Thank you so much. Yuma Bijleveld. $PK F1=(BIO1**FDS12) * (BIO2**FDS34) TVV1=THETA(1)*(WT/70000) V1=TVV1*EXP(ETA(1)) TVCL=THETA(2)*(WT/70000)**0.75*(GA/281)**THETA(6) CL=TVCL*EXP(ETA(2)) TVQ=THETA(4)*(WT/70000)**0.75 Q=TVQ TVV2=THETA(5)*(WT/70000) V2=TVV2*EXP(ETA(3)) S1=V1 $ERROR IPRED=LOG(0.0001) IF(F.GT.0)IPRED=LOG(F) IRES = DV-IPRED W=1 IF(F.GT.0)W = SQRT(THETA(3)**2) IWRES = IRES/W Y = IPRED+W*EPS(1)*EXP(ETA(4)) $THETA (0, 75.7) ;1 V1 (0, 2.09) ;2 CL (0, 0.376) ;3 add (0, 3) ;4 Q (0, 31.8) ;5 V2 (0, 3.3) ;6 GA $OMEGA BLOCK(2) 0.167 ;1 V1 0.0824 0.12 ;2 Cl $OMEGA 0.1 ;3 V2 $OMEGA 0.1 ;4 RES $SIGMA 1 FIX ________________________________ AMC Disclaimer : http://www.amc.nl/disclaimer ________________________________ ________________________________ De inhoud van dit bericht is vertrouwelijk en alleen bestemd voor de geadresseerde(n). Anderen dan de geadresseerde(n) mogen geen gebruik maken van dit bericht, het niet openbaar maken of op enige wijze verspreiden of vermenigvuldigen. Het UMCG kan niet aansprakelijk gesteld worden voor een incomplete aankomst of vertraging van dit verzonden bericht. The contents of this message are confidential and only intended for the eyes of the addressee(s). Others than the addressee(s) are not allowed to use this message, to make it public or to distribute or multiply this message in any way. The UMCG cannot be held responsible for incomplete reception or delay of this transferred message.
