The Influence of Normalization Weight in Population Pharmacokinetic Covariate Models

Transcription

1 Cln Pharmacoknet SHORT COMMUNICATION The Influence of Normalzaton Weght n Populaton Pharmacoknetc Covarate Models Sebastaan C. Goulooze 1 Swantje Völler 1 Pyry A. J. Vältalo 1 Elsa A. M. Calver 1 Leon Aarons 2 Elke H. J. Krekels 1 Catherjne A. J. Knbbe 1,3 Ó The Author(s) 2018 Abstract In covarate (sub)models of populaton pharmacoknetc models, most covarates are normalzed to the medan value; however, for body weght, normalzaton to 70 kg or 1 kg s often appled. In ths artcle, we llustrate the mpact of normalzaton weght on the precson of populaton clearance (CL pop ) parameter estmates. The nfluence of normalzaton weght (70, 1 kg or medan weght) on the precson of the CL pop estmate, expressed as relatve standard error (RSE), was llustrated usng data from a pharmacoknetc study n neonates wth a medan weght of 2.7 kg. In addton, a smulaton study was performed to show the mpact of normalzaton to 70 kg n pharmacoknetc studes wth paedatrc or obese patents. The RSE of the CL pop parameter estmate n the neonatal dataset was lowest wth normalzaton to medan weght (8.1%), compared wth normalzaton to 1 kg (10.5%) or 70 kg (48.8%). Typcal clearance (CL) predctons were ndependent of the normalzaton weght used. Smulatons showed that the ncrease n RSE of the CL pop estmate wth Sebastaan Goulooze and Swantje Völler contrbuted equally to ths work. Electronc supplementary materal The onlne verson of ths artcle ( contans supplementary materal, whch s avalable to authorzed users. & Catherjne A. J. Knbbe c.knbbe@antonuszekenhus.nl 70 kg normalzaton was hghest n studes wth a narrow weght range and a geometrc mean weght away from 70 kg. When, nstead of normalzng wth medan weght, a weght outsde the observed range s used, the RSE of the CL pop estmate wll be nflated, and should therefore not be used for model selecton. Instead, establshed mathematcal prncples can be used to calculate the RSE of the typcal CL (CL TV ) at a relevant weght to evaluate the precson of CL predctons. Key Ponts Normalzaton to a weght outsde the observed weght range (e.g. 70 kg normalzaton n a paedatrc study) can ncrease the uncertanty of parameter estmates n pharmacoknetc covarate models. The predctve performance of pharmacoknetc models and ther covarate submodels s unaffected by weght normalzaton. When normalzng outsde the observed covarate range, the RSEs of the correspondng populaton estmates should generally not be used for model evaluaton. The RSE of the typcal parameter at a relevant covarate value can be used nstead Dvson of Systems Bomedcne and Pharmacology, Leden Academc Centre for Drug Research, Leden Unversty, Leden, The Netherlands Dvson of Pharmacy and Optometry, Unversty of Manchester, Manchester, UK Department of Clncal Pharmacy, St Antonus Hosptal, Neuwegen, The Netherlands

2 S. C. Goulooze et al. 1 Introducton In populaton pharmacoknetc modellng, covarate models are bult to descrbe between-subject varablty n pharmacoknetc parameters (e.g. clearance [CL]) based on patent nformaton (e.g. weght) [1, 2]. These covarate models can then be used to support personalzed pharmacotherapy [3]. One of the most commonly dentfed covarates n populaton pharmacoknetc models s the body weght of the patent [4 8]. The relatonshp between drug CL and weght s often descrbed usng a power functon: WT EXPWT CL TV ¼ CL pop ð1þ CL ¼ CL TV e ETA ; WT norm ð2þ where CL TV represents the predcted CL for a typcal ndvdual wth weght equal to WT,CL pop represents the populaton estmate of CL for an ndvdual wth a weght equal to the normalzaton weght (WT norm ), WT represents the ndvdual s weght, EXP WT represents the exponent that characterzes the nfluence of weght on CL, CL represents the ndvdual post hoc predctons of CL for ndvdual, and e ETA represents the post hoc estmate of the devaton of ndvdual from the predcted CL of a typcal ndvdual. EXP WT n ths weght-based Eq. 1 can be fxed a pror (for nstance to 1, 0.75 or 0.67) or estmated [9 12]. For contnuous covarates other than weght, t s common to normalze to the mean or the medan covarate value of the covarate n the dataset, whereas for weght, normalzaton to 70 kg s often chosen [2, 13 15]. The ratonale for ths approach s that the estmate of CL pop wll then represent the value of a typcal 70 kg adult, whch can easly be compared wth other (adult) studes [2, 5, 10]. For smlar reasons, a WT norm of 1 kg can be chosen or, when no explct normalzaton s performed, ths normalzaton s mplctly chosen [9]. Recently, Mahmood and Tegenge nvestgated the mpact of 70 and 1 kg weght normalzaton on CL TV predctons and concluded that weght normalzaton has no mpact on CL TV predctons [9]. The concept of normalzaton has been extensvely studed n lnear regresson [16, 17]. When normalzng the data to the mean, the relatve standard error (RSE) of the ntercept term s mnmzed, whle normalzng outsde of the data range can result n estmates wth poor precson [16]. In ths context, the power functon n Eq. 1 can be consdered a lnear model n the log doman, wth an ntercept of log CL pop [13]. If the same concepts apply, we mght expect the RSE of the estmate of CL pop to be mnmal when normalzng weght to the geometrc mean. Alternatvely, normalzng to 1 kg or 70 kg mght result n hgh RSE of CL pop, especally n populatons wth hgh or low weghts. As the mpact of 70 kg normalzaton on the RSE of CL pop could be consderable, especally when analysng data from neonatal or morbdly obese patents, nsght nto the statstcal consequences of the selected normalzaton weght seems mportant among those nvolved n populaton pharmacoknetc modellng. In ths artcle, we provde mathematcal dervatons of the phenomena and llustrate the mpact of weght normalzaton on the precson of the parameter estmate of CL pop. For ths, we used an exstng neonatal pharmacoknetc dataset, as well as smulated data of varous paedatrc and obese populatons. 2 Methods 2.1 Case Study A dataset from a prevously publshed populaton pharmacoknetc analyss of phenobarbtal n term and preterm neonates was used [18]. Ths dataset contaned phenobarbtal plasma concentratons collected durng therapeutc drug montorng from 53 neonates up to 80 h after the last phenobarbtal dose. The weght of these neonates ranged from 0.45 to 4.5 kg and had a medan value of 2.7 kg [18]. These data were modelled usng a one-compartment model wth nterndvdual varablty on CL and volume of dstrbuton (V) and a proportonal error model. The covarate model conssted of a lnear model (exponent fxed to a value of 1) for weght on V and a power model wth an estmated exponent for weght on CL (Eq. 1). For three dfferent values of WT norm (1, 2.7, and 70 kg), parameter estmates were obtaned usng NONMEM 7.3 [19]. The collnearty of the parameter estmates was assessed usng the condton number, whch s defned as the square root of the rato between the largest and smallest egenvalue of the correlaton matrx. RSEs of the parameter estmates (.e., CL pop and EXP WT ) n NONMEM were obtaned usng two dfferent methods: calculaton from the estmated varance covarance matrx n NONMEM, and calculaton calculated from 1000 bootstrap runs, whch were performed usng PsN for each of the three models [20]. Furthermore, from these 1000 bootstrap runs, the estmates of CL pop and EXP WT (see Eq. 1) were used to calculate the predcted functon of CL TV over weght for each of the bootstrap runs over a weght range of kg. For each of the three weght-normalzed models, we used these functons of CL TV over weght to obtan the 95% confdence nterval and RSE of CL TV predctons over the weght range of kg for all three values of WT norm.

3 Influence of Normalzaton Weght n Populaton PK Covarate Models RSE of CL TV predctons over weght were also calculated usng the varance covarance matrx obtaned from NONMEM [13]. For ths, we log-transformed the power functon of Eq. 1, whch resulted n the followng lnear model: WT Log CL TV ¼ LogCL pop þ EXP WT log ; ð3þ WT norm where LogCL pop represents the estmate of the natural logarthm of CL of an ndvdual whose weght s equal to the normalzaton weght. Wth a lnear model, we can use prncples from lnear regresson to calculate the RSE of the predctons of CL TV at an arbtrary weght WT : VARðLog CL TV Þ ¼ VAR LogCLpop þ 2 COVAR LogCLpop ;EXP wt WT log þ VAR EXPwt WT norm WT 2 log ð4þ WT norm pffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff RSEðCL TV Þ ¼ e VARðLogCL TV Þ 1; ð5þ where VAR LogCLpop represents the varance of the estmate of LogCL pop, COVAR LogCLpop ;EXP wt represents the covarance between the estmates for LogCL pop and EXP WT, and VAR EXPwt represents the varance of the estmate of EXP WT. Mathematcal dervaton of Eqs. 4 and 5 are suppled n the electronc supplementary materal. 2.2 Impact of WT norm on the Relatve Standard Error of CL pop for Dfferent Weght Dstrbutons To further study the mpact of WT norm n a covarate functon on the RSE of the estmate of CL pop wth dfferent weght dstrbutons, we generated pharmacoknetc datasets of patent populatons wth sx dfferent weght dstrbutons n R verson We then ftted a onecompartmental pharmacoknetc model descrbed n the prevous subsecton, re-estmatng all parameters. For ths purpose, three paedatrc weght dstrbutons (PEDIAT1, PEDIAT2 and PEDIAT3) and three weght dstrbutons of adult populatons ncludng obese patents (OBESE1, OBESE2 and OBESE3) were used to generate smulated datasets (Table 2). For each weght dstrbuton, 250 datasets consstng of 50 patents were randomly sampled from the dstrbutons usng R software. For each patent, concentratons were smulated for 24, 72 and 120 h after a sngle 10 mg/kg dose. The smulated datasets were ftted wth the models descrbed n Sect. 2.1 and weght was normalzed to ether 70 kg or the expected geometrc mean of the weght dstrbuton. Geometrc mean was chosen as we hypotheszed that ths would result n the mnmum RSE of CL pop, as dscussed n the Introducton. The proporton of successful covarance steps was compared for a statstcally sgnfcant dfference between the dfferent normalzaton strateges, usng a two-sample test for equalty of proportons wth contnuty correcton (prop.test functon n R). For datasets that yelded successful covarance steps n both model fts, we calculated the rato of the RSEs for the estmate of CL pop of both model fts: RSE rato ¼ 3 Results 3.1 Case Study RSE CL pop 70 kg normalzaton RSE CL pop Geometrc mean normalzaton ð6þ Table 1 shows that only the RSE of the estmate of CL pop vared wth normalzaton weght. Normalzng to medan weght (2.7 kg) resulted n a lower RSE of CL pop, compared wth 1 kg and 70 kg normalzaton. RSEs were 10.6, 8.0, and 48.2% for 1, 2.7 and 70 kg, respectvely (Table 1). These RSE values were obtaned from NONMEM s covarance step, but smlar results were obtaned usng a bootstrap (1 kg: 10.5%; 2.7 kg: 8.1%; 70 kg: 48.8%). Addtonally, there was a stronger correlaton between the uncertanty of the parameter estmates of CL pop and EXP WT when normalzng to 1 kg or 70 kg (Table 1). Fnally, ncreased collnearty between the parameters was observed for the 1 kg and 70 kg normalzatons, as dentfed by a hgher condton number (Table 1). The parameter estmates of CL pop vared wth normalzaton weght (Table 1), whch results from the fact that ths parameter represents the typcal CL of a subject whose weght s equal to the normalzaton weght. However, the same predcted CL TV (Eq. 1) s obtaned for the three model fts wth dfferent normalzaton weghts snce CL TV ¼ 0:00615 WT 0:665 1kg 0:0119 WT 0:665 0:104 WT 0:665 2:7kg 70kg Ths s further llustrated by Fg. 1 where the results on predcted CL TV for the model wth a normalzaton weght of 2.7 kg s shown, whle the models wth dfferent normalzaton weghts produced equvalent results. Addtonally, Fg. 1a shows that the 95% bootstrap confdence nterval of predcted CL TV broadens the further the weght moves away from the centre of the weght dstrbuton of the patent populaton n the dataset.

4 S. C. Goulooze et al. Table 1 Parameter estmates and relatve standard errors (%) from the NONMEM covarance step for the neonatal dataset usng dfferent normalzaton weghts WT norm 1 kg 2.7 kg (medan) 70 kg OFV CL pop (L/h) (10.6%) (8.0%) (48.2%) V (L) 2.37 (4.4%) 2.37 (4.4%) 2.37 (4.4%) EXP WT (20.3%) (20.3%) (20.3%) Proportonal error [%] 2.89 (23.5%) 2.89 (23.5%) 2.89 (23.5%) Condton number Correlaton CLpop, EXPwt a a Correlaton of the uncertanty of the parameter estmates of CL pop and EXP WT OFV objectve functon value, CL pop typcal clearance of subject whose weght s equal to normalzaton weght, V volume of dstrbuton, EXP WT exponent n Eq. 1, WT norm normalzaton weght n Eq. 1 Fg. 1 Clearance predctons versus weght ( kg) n an example neonatal dataset. (a) Medan (sold black lne) and 95% confdence nterval (dotted lne) of 1000 functons of CL TV versus weght obtaned from 1000 bootstrap runs; green dots represent the ndvdual post hoc CL estmates of the studed patents. (b) Estmated functon of CL TV versus weght from the orgnal dataset (sold black Fgure 1b llustrates that the CL TV functons from dfferent bootstrap samples are very smlar wthn the range of the data, and dverge outsde the weght range of the orgnal dataset, explanng the broader confdence ntervals seen n Fg. 1a. Fgure 2 shows the RSE of the predcted functon of CL TV over weght, as well as the RSE of the estmate of the parameter CL pop. The RSE of the CL TV functon was calculated usng ether bootstrap or the varance covarance matrx (Eq. 5), wth both methods resultng n smlar results (Fg. 2). Usng Eq. 5, the mnmum RSE of CL TV (5.8%) was calculated at 1.9 kg, whch s close to, but not equal to, measures of the central tendency of the weght dstrbuton, such as the mean (2.4 kg, RSE = 6.6%), medan (2.7 kg, RSE = 7.5%) or geometrc mean (2.1 kg, RSE = 6.0%). The RSE of the estmate of the CL pop parameter reported by NONMEM for a model wth a gven normalzaton weght matches the RSE of the CL TV functon at the normalzaton weght. The former results n lne) and llustratve set of functons of CL TV versus weght (grey sold lnes) obtaned n sx (of 1000) separate bootstrap runs; green dots represent the ndvdual post hoc CL estmates of patents n the orgnal dataset. Depcted results were obtaned usng a normalzaton weght of 2.7 kg. CL TV clearance for a typcal ndvdual, CL clearance for ndvdual 6.4% RSE wth 1.9 kg normalzaton, and 8.0% RSE wth normalzaton to the medan. 3.2 Impact of 70 kg Normalzaton n Dfferent Paedatrc and Obese Populatons Evaluaton of the mpact of 70 kg normalzaton on the RSE of CL pop estmates for smulated paedatrc and obese datasets wth varous weght dstrbutons showed that, generally, 70 kg normalzaton resulted n a hgher RSE of the CL pop estmate compared wth normalzaton to geometrc mean weght, resultng n an RSE rato above 1 (Table 2, Fg. 3). The results show that the degree of mpact of 70 kg normalzaton depends on the weght dstrbuton n the dataset. The three paedatrc weght dstrbutons had a geometrc mean weght of 20 kg, but dfferent dsperson around the geometrc mean. The RSE rato was hghest (medan RSE rato = 4.3) for the lognormal dstrbuton wth a standard devaton on the logarthmc scale of 0.25 (PEDIAT1).

5 Influence of Normalzaton Weght n Populaton PK Covarate Models Fg. 2 Relaton between weght and the RSE of both CL TV and CL pop n an llustratve neonatal dataset. The sold lne represents the RSE of CL TV predctons from 1000 bootstrap runs, the dotted lne represents the RSE of CL TV predctons obtaned from the varance covarance matrx (Eq. 5), and the red dots represent the RSE of the estmated CL pop parameter usng the correspondng normalzaton weght, obtaned from the covarance step of a sngle NONMEM run. The vertcal tck marks on the bottom of the graph depct the body weghts of subjects n the dataset. RSE relatve standard error, CL TV clearance for a typcal ndvdual, CL pop populaton clearance The weght dstrbutons of OBESE1 and OBESE2 have a smlar standard devaton on a logarthmc scale, but dfferent geometrc means of the weght (162 and 118 kg, respectvely). The results n Table 2 and Fg. 3 show that ths results n a lower RSE rato n OBESE2, compared wth OBESE1 (medan RSE rato of 2.3 and 3.5, respectvely). Normalzng to 70 kg has only margnal mpact on RSE n OBESE3 (medan RSE rato = 1.1), wth a geometrc mean that s closer to 70 kg than OBESE2 (97 kg vs. 118 kg), as well as a hgher standard devaton on a logarthmc scale (0.35 vs. 0.2). Table 2 shows that for both paedatrc and adult populatons, the mpact of 70 kg normalzaton on the RSE of the estmate of CL pop was hghest for studes wth a narrow weght range (low standard devaton on a logarthmc scale) and a medan weght away from 70 kg. The percentage of datasets for whch a successful covarance step could be obtaned n both model fts ranged from 28 to 84% n the dfferent scenaros. The proporton of successful covarance steps was sgnfcantly hgher usng geometrc mean normalzaton compared wth 70 kg normalzaton n scenaros PEDIAT1 (p \ 0.001) and PEDIAT2 (p \ 0.001), but not n any of the other scenaros. The most common cause of the mssng covarance steps was boundary ssues due to the dffculty n estmatng the varance for the nterndvdual varablty of CL. 4 Dscusson Ths report llustrates the statstcal prncple that, when estmatng the exponent n a body weght-based covarate submodel of a populaton pharmacoknetc model (Eq. 1), Table 2 Characterstcs of the dfferent weght dstrbutons and summary of the smulaton results Dstrbuton Geometrc mean (kg) SD on logscale Dstance between geometrc mean and 70 kg (n SD on logscale) Medan RSE rato CL pop (Eq. 6) for 70 kg normalzaton Covarance step successful wth 70 kg normalzaton (%) Covarance step successful wth geometrc mean normalzaton (%) Covarate step successful n both normalzatons (%) PEDIAT1 Log-normal PEDIAT kg PEDIAT kg OBESE OBESE OBESE RSE relatve standard error, SD standard devaton, CL pop populaton clearance

6 S. C. Goulooze et al. Fg. 3 RSE rato of CL pop (Eq. 6) when usng 70 kg normalzaton compared wth geometrc mean weght normalzaton. For each weght dstrbuton, 250 datasets were generated and reftted. Only results of datasets for whch the covarance step was successful for both the 70 kg and geometrc mean weght normalzaton were ncluded n ths graph (Table 2). RSE relatve standard error, CL pop populaton clearance the use of a normalzaton weght outsde the observed weght range can result n an nflated RSE of the estmate of CL pop. Ths holds true for 70 kg normalzaton, but also for 1 kg normalzaton (whch s sometmes referred to as no normalzaton ) [9]. The RSE of the CL parameter CL pop represents the RSE of the predcted typcal CL (CL TV ) at a partcular normalzaton weght, and s therefore not a unversal measure for the precson of the estmate of CL (Table 2). As we show, the RSE of the predcted CL TV s not constant, but dependent on the weght of the subject for whom CL TV s predcted (Fgs. 1 and 2). Ths means that the RSE of CL pop represents how precsely CL can be estmated for a subject at the appled normalzaton weght. When estmatng CL pop at a normalzaton weght outsde the observed weght range, the RSE of CL pop wll be nflated and cannot be used as a crteron for model selecton. Our example wth 70 kg normalzaton n a neonatal pharmacoknetc model showed, for nstance, that the RSE of the CL pop estmate ncreased sxfold compared wth the estmate obtaned wth normalzaton to the medan. In lnear regresson, the mnmum RSE s obtaned by normalzng to the mean value. Because the power functon becomes a lnear model n the log doman, one mght expect the mnmum RSE at the geometrc mean weght as ths s equvalent to normalzng to the mean value of log of weght. The results of the case study show that ths s not necessarly the case for covarate models of non-lnear mxed-effects models. In our case study, we found that the mnmum RSE of CL pop was obtaned by normalzng to 1.9 kg, rather than the mean, medan or geometrc mean weght. The normalzaton weght wth mnmum RSE of CL pop can be predcted by mnmzng Eq. 5, although ths does requre that an ntal model wth a test normalzaton weght s run to obtan an estmate of the varance covarance matrx. In our case study, we normalzed to the medan weght as ths s the most commonly used normalzaton weght. Ths ncreased the RSE of CL pop to 8.0%, from the mnmum RSE of 6.4% at a normalzaton weght of 1.9 kg. However, normalzng to the medan weght wll lkely be ft-for-purpose n most cases. Regardless of the normalzaton weght that s used, Eqs. 4 and 5 can be used to calculate RSE values for any gven body weght based on a varance covarance matrx. Ths can be useful as normalzaton to 70 kg as a standardzed ndvdual s sometmes advocated to mprove comparsons of results from dfferent studes [5]. In these cases, normalzaton to 70 kg can appled and Eqs. 4 and 5 can be used to calculate the RSE values for a relevant body weght to enable ts use as a model selecton crteron. Alternatvely, model estmaton can be performed wth medan weght normalzaton, and both the estmated CL pop parameter for the medan weght and the derved CL TV for a 70 kg ndvdual are reported, n whch case the calculated RSE of the latter s relevant when comparng results from dfferent studes. The expected ncrease n RSE of centrng to 70 kg s dependent on both the varance of the weght dstrbuton and the dstance of the mean of the dstrbuton from the chosen normalzaton weght (Fg. 3). The effect seems to be largest n cases of narrow dstrbutons wth a mean covarate value far away from the centre of the data. Ths

7 Influence of Normalzaton Weght n Populaton PK Covarate Models especally holds true for (pre)term neonates, nfants and young chldren and (morbdly) obese patents. In cases where the populaton mean s away from 70 kg but the range of weghts ncludes 70 kg (such as scenaro OBESE3), normalzng to 70 kg wll lkely result n an RSE ncrease of CL pop that wll not affect ts applcaton n model selecton (Fg. 3). It s mportant to realze that weght normalzaton only mpacts the precson of the CL pop parameter estmate, n case a covarate model accordng to Eq. 1 s chosen. If the exponent s fxed, the RSE of the estmate of CL pop wll be unaffected by normalzaton weght. Smlarly, n ths stuaton, the RSE of the predcted CL TV wll be ndependent of the subject s weght. Whether or not the exponent should be estmated or fxed s an ongong dscusson that s outsde the scope of ths paper [5, 11, 12]. In the smulaton study, the calculaton of the RSE rato requred that both model fts resulted n a successful covarance step. Ths requrement ntroduces the potental for selecton bas as the results from the excluded datasets mght have had a dfferent mpact of normalzaton on the RSE of CL pop. However, performng a bootstrap on each of the 3000 model fts to obtan the RSE rato ndependent of covarance step success was not feasble due to ts computatonal demands. Interestngly, the smulaton study showed that n two of the sx smulated scenaros, a hgher percentage of runs wth successful covarance steps was obtaned when usng geometrc mean normalzaton nstead of 70 kg normalzaton. Addtonally, the paedatrc case study showed that wth normalzaton outsde the observed weght range, collnearty between the CL pop and EXP WT parameter estmates s ncreased (Table 1). Ths suggests that normalzaton affects the stablty of the parameter estmaton process, whch has been descrbed for both lnear models, as well as populaton pharmacoknetc models [15]. Although the ncreased stablty was not observed n all smulated scenaros, t mght be a reason to advocate medan weght normalzaton over normalzaton outsde the observed weght range. 5 Conclusons Normalzng body weght-based covarate relatonshps n populaton pharmacoknetc models to 1 kg or 70 kg can nflate the RSE of the parameter estmate of CL pop n populaton pharmacoknetc models. The predctve performance of the models are unaffected by normalzaton. However, when normalzng wth a weght outsde the observed weght range, the CL pop RSE represents the precson of the CL TV at ths extrapolated weght, and ths value should therefore not be used for model selecton. Instead, the precson of CL TV at a relevant weght value can be calculated from the covarance matrx. Acknowledgements The authors would lke to thank the researchers who provded the data used for the case study n ths paper, most notably Dr. Leo M. Stolk and Dr. Snno H. P. Smons. Complance wth Ethcal Standards Ths artcle does not contan any studes wth human partcpants performed by any of the authors. Detals on the ethcal approval and nformed consent from the prevously publshed clncal studes of whch the anonymzed data was used as a case example are reported elsewhere [18]. Fundng Catherjne A. J. Knbbe was supported by the Innovatonal Research Incentves Scheme of the Dutch Organzaton for Scentfc Research (NWO, Vd grant, May 2013). Conflcts of nterest Sebastaan C. Goulooze, Swantje Völler, Pyry A. J. Vältalo, Elsa A. M. Calver, Leon Aarons, Elke H.J. Krekels, and Catherjne A.J. Knbbe have no conflcts of nterest that are relevant to the content of ths paper. Open Access Ths artcle s dstrbuted under the terms of the Creatve Commons Attrbuton-NonCommercal 4.0 Internatonal Lcense ( whch permts any noncommercal use, dstrbuton, and reproducton n any medum, provded you gve approprate credt to the orgnal author(s) and the source, provde a lnk to the Creatve Commons lcense, and ndcate f changes were made. References 1. Mould DR, Upton RN. Basc concepts n populaton modelng, smulaton, and model-based drug development. CPT Pharmacometrcs Syst Pharmacol. 2012;1:e6. 2. Mould DR, Upton RN. Basc concepts n populaton modelng, smulaton, and model-based drug development part 2: ntroducton to pharmacoknetc modelng methods. CPT Pharmacometrcs Syst Pharmacol. 2013;2:e Krekels EHJ, van Hasselt JGC, van den Anker JN, Allegaert K, Tbboel D, Knbbe CAJ. Evdence-based drug treatment for specal patent populatons through model-based approaches. Eur J Pharm Sc. 2017;109S:S Anderson BJ, Allegaert K, Holford NH. Populaton clncal pharmacology of chldren: modellng covarate effects. Eur J Pedatr. 2006;165(12): Anderson BJ, Holford NH. Mechansm-based concepts of sze and maturty n pharmacoknetcs. Annu Rev Pharmacol Toxcol. 2008;48: Sh R, Derendorf H. Pedatrc dosng and body sze n botherapeutcs. Pharmaceutcs. 2010;2(4): Knbbe CA, Brll MJ, van Rongen A, Depstraten J, van der Graaf PH, Danhof M. Drug dsposton n obesty: toward evdencebased dosng. Annu Rev Pharmacol Toxcol. 2015;55: Eleveld DJ, Proost JH, Absalom AR, Struys MM. Obesty and allometrc scalng of pharmacoknetcs. Cln Pharmacoknet. 2011;50(11):751 3 (dscusson 5 6). 9. Mahmood I, Tegenge MA. Populaton pharmacoknetcs: some observatons n pedatrc modelng for drug clearance. Cln Pharmacoknet. 2017;56(12):

8 S. C. Goulooze et al. 10. Holford NH. A sze standard for pharmacoknetcs. Cln Pharmacoknet. 1996;30(5): Fsher DM, Shafer SL. Allometry, Shallometry! Anesth Analg. 2016;122(5): Calver EA, Krekels EH, Valtalo PA, Rostam-Hodjegan A, Tbboel D, Danhof M, et al. Allometrc scalng of clearance n paedatrc patents: when does the magc of 0.75 fade? Cln Pharmacoknet. 2017;56(3): Wang Y, Jadhav PR, Lala M, Gobburu JV. Clarfcaton on precson crtera to derve sample sze when desgnng pedatrc pharmacoknetc studes. J Cln Pharmacol. 2012;52(10): Joerger M. Covarate pharmacoknetc model buldng n oncology and ts potental clncal relevance. AAPS J. 2012;14(1): Bonate PL. Pharmacoknetc-pharmacodynamc modelng and smulaton. 2nd ed. New York: Sprnger; Snee RD, Marquardt DW. Comment: collnearty dagnostcs depend on the doman of predcton, the model, and the data. Am Stat. 1984;38(2): Belsley DA. Demeanng condtonng dagnostcs through centerng. Am Stat. 1984;38(2): Voller S, Flnt RB, Stolk LM, Degraeuwe PLJ, Smons SHP, Pokorna P, et al. Model-based clncal dose optmzaton for phenobarbtal n neonates: an llustraton of the mportance of data sharng and external valdaton. Eur J Pharm Sc. 2017;109S:S Boeckmann AJ, Shener LB, Beal SL. NONMEM users gude part VIII. Ellcott Cty: ICON Development Solutons; p Kezer RJ, Karlsson MO, Hooker A. Modelng and smulaton workbench for NONMEM: tutoral on Prana, PsN, and Xpose. CPT Pharmacometrcs Syst Pharmacol. 2013;2:e50.