advertisement

Human and Ecological Risk Assessment: An International Journal ISSN: 1080-7039 (Print) 1549-7860 (Online) Journal homepage: http://www.tandfonline.com/loi/bher20 Robust benchmark dose estimation using an infinite family of response functions Ishapathik Das To cite this article: Ishapathik Das (2018): Robust benchmark dose estimation using an infinite family of response functions, Human and Ecological Risk Assessment: An International Journal, DOI: 10.1080/10807039.2018.1438172 To link to this article: https://doi.org/10.1080/10807039.2018.1438172 Published online: 01 Mar 2018. Submit your article to this journal Article views: 18 View related articles View Crossmark data Full Terms & Conditions of access and use can be found at http://www.tandfonline.com/action/journalInformation?journalCode=bher20 HUMAN AND ECOLOGICAL RISK ASSESSMENT https://doi.org/10.1080/10807039.2018.1438172 Robust benchmark dose estimation using an inﬁnite family of response functions Ishapathik Das Department of Mathematics, Indian Institute of Technology Tirupati, Andhra Pradesh, India; Department of Statistical Science, Duke University, Durham, North Carolina, USA ABSTRACT ARTICLE HISTORY Researchers usually estimate benchmark dose (BMD) for dichotomous experimental data using a binomial model with a single response function. Several forms of response function have been proposed to ﬁt dose–response models to estimate the BMD and the corresponding benchmark dose lower bound (BMDL). However, if the assumed response function is not correct, then the estimated BMD and BMDL from the ﬁtted model may not be accurate. To account for model uncertainty, model averaging (MA) methods are proposed to estimate BMD averaging over a model space containing a ﬁnite number of standard models. Usual model averaging focuses on a pre-speciﬁed list of parametric models leading to pitfalls when none of the models in the list is the correct model. Here, an alternative which augments an initial list of parametric models with an inﬁnite number of additional models having varying response functions has been proposed to estimate BMD for dichotomous response data. In addition, different methods for estimating BMDL based on the family of response functions are derived. The proposed approach is compared with MA in a simulation study and applied to a real dataset. Simulation studies are also conducted to compare the four methods of estimating BMDL. Received 19 December 2017 Revised manuscript accepted 5 February 2018 KEYWORDS benchmark dose; binomial models; model misspeciﬁcation; family of response functions; interval estimation Introduction One of the main goals in quantitative risk assessment is to estimate the risk function R(d), which is the probability of adverse events, such as death, birth defect, weight loss, cancer, or mutation exhibited in a subject exposed at dose level d. Suppose n number of subjects are exposed to a dose level d and y number of adverse events are observed. Then, the response y is distributed according to a binomial distribution with parameter [n, R(d)], where R(d) is the probability of adverse events at dose level d. It is usually observed that R(d) increases with doses d, and hence a binomial model with a non-decreasing response function may be better to ﬁt such dichotomous data than any mathematical model. After estimating the risk function R(d), the extra risk function RE(d), deﬁned as RE ðd Þ D Rð1d¡Þ ¡p0p0 is computed, where p0 is the risk at minimum dose level usually called as background risk. The benchmark dose (BMD) is deﬁned by the dose CONTACT Ishapathik Das ishapathik@iittp.ac.in Department of Mathematics, Indian Institute of Technology Tirupati, Tirupati – Renigunta Road, Settipalli Post, Tirupati, Andhra Pradesh 517506, India. © 2018 Taylor & Francis Group, LLC 2 I. DAS level having the extra risk RE ðBMDÞ D BMR, where BMR is called the benchmark response, usually pre-speciﬁed in the range 0:01BMR0:1 (US EPA 2012). The benchmark dose lower bound (BMDL) is also determined using the risk function R(d). The accuracy of the estimation of BMD and BMDL is dependent upon the estimation of the risk function R(d). The benchmark analysis by estimating BMDs and corresponding BMDLs have been accepted as one of the important tools for quantitative risk assessment by a variety of agencies such as the U.S. Environmental Protection Agency (US EPA), the European Food Safety Authority (EFSA), the Organisation for Economic Cooperation and Development (OECD). Several agencies including the US EPA, the EFSA, the OECD provide guidance on BMDs (OECD 2008; US EPA 2012; EFSA 2017) and the estimation of BMDs is dramatically increased over the past few decades for quantitative risk assessment. Methods of estimating BMD and BMDL are discussed by several researchers such as Crump (1984); Bailer et al. (2005); Morales et al. (2006); Wheeler and Bailer (2007, 2009); West et al. (2012) to name just a few. Crump (1984) introduced methods of estimating BMD and BMDL by proposing four models for discrete responses and three models for continuous responses. There are eight models (Wheeler and Bailer 2007, 2009; West et al. 2012) that have been identiﬁed as standard models for estimating BMD and BMDL. One of the models from the set of standard models may be chosen for ﬁtting the data sets. However, the responses may be generated from the model other than the chosen model. Researchers (Wheeler and Bailer 2007, 2009; West et al. 2012) have shown that the estimation of BMD and BMDL are signiﬁcantly affected if the assumed model is incorrect. So, there is a recent rise in developing methods of accounting for model uncertainty in BMD estimation. To deal with model uncertainty in BMD and BMDL estimation, various model averaging (MA) methods have been proposed by Kang et al. (2000); Bailer et al. (2005); Wheeler and Bailer (2007); Shao and Small (2011); West et al. (2012); Piegorsch et al. (2013). The estimates of BMD and BMDL in model averaging methods are given by the weighted average of the estimates of BMD and BMDL when individually modeled. Bayesian methods and Bayesian model averaging methods for estimating BMD and BMDL are also proposed by Morales et al. (2006); Shao and Small (2012); Simmons et al. (2015). The model averaging approach may solve the problems of model uncertainty when the true model generating responses can be well approximated by some members of the model space containing the assumed models. The model space used in MA is always ﬁnite. However, there is an inﬁnite number of other models which cannot be approximated by the members of the model space used in MA leading to the problem of model uncertainty. So, model averaging techniques only provide a partial solution to the problem of model uncertainty. In this article, an inﬁnite family of models containing some of the standard models is used to ﬁnd the risk function. The family of response models is parameterized by two unknown parameters. There is an inﬁnite number of response models which can be represented by different values of parameters. Some standard response models correspond to some speciﬁc values of parameters. So, we may get better results for accounting model uncertainty in BMD and BMDL estimation using the proposed family of models. HUMAN AND ECOLOGICAL RISK ASSESSMENT 3 Methods Binomial models The binomial models are members of the generalized linear models (GLMs) described by three components given below. 1. Distributional components: let y1, y2, …, yn be n random samples of adverse events at dose levels d1, d2, …, dn, where for each i 2 {1, 2, …, n}, yi has binomial distribution with parameter (ni, ri), ri 2 [0, 1], and yi D nyii has scaled binomial distribution belongs to the exponential family having the form of probability mass function (pmf) given by (Fahrmeir and Tutz 2001) yi ui ¡ bðui Þ wi C cðyi ; wi ; ’Þ ; s yi jui ; wi ; ’ D exp ’ where ri D Rðdi Þ D E yi jdi , ui D log 1 ¡ri ri is the so-called natural parameters, b(ui) D h i log [1 C exp (ui)], wi D ni, f D 1, and cðyi ; wi ; ’Þ D log yi !ðnni i¡! yi Þ! . 0 2. Linear predictor: hðdi Þ D f ðdi Þb, where f(di) is a vector function of di, and b is the regression parameter vector. 3. Parametric response function: Rðdi Þ D h½a; hðdi Þ or g ½a; Rðdi Þ D hðdi Þ, where h is the parametric response function and g (the inverse of h) is the parametric link function. We usually assume that the inverse of h exists. For dose–response studies, the linear predictor is usually assumed to be h(d) D b0 C b1d, or h(d) D b0 C b1d C b2d2, and a single response function such as logistic, probit, log-log, complementary log-log or some other response functions are used to ﬁt the models. Here, instead of a single response function, we use a family of response functions (parametric response function) parameterized by a response parameter vector a to ﬁt the models. So, we denote the response function as hða; ¢Þ in place of h( ¢ ). Several researchers (Stukel 1988; Czado 1997) proposed various families of response functions to ﬁt binomial models. One such family of response functions for binomial models is given by Rðd Þ D Eðyjd Þ D h½a; hðd Þ D exp½Gða; hÞ ; 1 C exp½Gða; hÞ (1) where h h(d), and Gða; ¢Þ is called a generating family. There are several Generating families proposed in the literature (Stukel 1988; Czado 1989). Stukel (Stukel 1988) provides the following generating family: if h 0 (i.e., r 12), 8 expða1 hÞ ¡ 1 > > ; a1 > 0 > > < a1 a1 D 0 Gða; hÞ D h; > > > logð1 ¡ a1 hÞ > :¡ ; a1 < 0; a1 4 I. DAS and for h < 0 (i.e., r < 12), 8 1 ¡ exp ð ¡ a2 hÞ > > ; > > < a2 Gða; hÞ D h; > > log ð1 C a2 hÞ > > : ; a2 a2 > 0 a2 D 0 a2 < 0; 0 where h h(d), and r R(d). Note that, for a D ½0; 0 , we get the logistic response function. So, the logistic response function is a member of this family. Also, several important response functions such as the Probit, the complementary log–log response functions can be approximated by the members of this family (Stukel 1988). For estimating the risk function using the above models, we need to0 estimate the 0 0 for the comunknown parameters using an available data set. Let us denote d D b ; a bined parameter vectors including the unknown regression parameter vector b and the response parameter vector a. The unknown parameter vector d can be estimated using the maximum likelihood estimation (MLE) methods given in Stukel (1988). Due to the estimation of the response parameters along with regression parameters, the variances of the estimated regression parameters are increased (Taylor 1988). The variance inﬂations of the regression parameters are asymptotically zero if the response parameters are orthogonal to the regression parameters (Cox and Reid 1987). Czado (Czado 1997) proposed some conditions on the family of response functions providing local orthogonality between response and regression parameter vectors. A family of response functions L D fhða; ¢Þ : a 2 Vg provides local orthogonality between response and regression parameter vectors around a point h0 asymptotically, if the following conditions are satisﬁed: 1. there exists h0 and r0 such that hða; h0 Þ D r0 ; 8 a 2 V; (2) @hða; hÞ jðh D h0 Þ D s0 ; 8 a 2 V; @h (3) and 2. there exists s0 such that where V is denoted for the parameter space of a. Such a family L D fhða; ¢Þ : a 2 Vg satisfying conditions (2) and (3) is called (r0, s0) ¡ standardized at h0 (Czado 1997). Now, for estimating risk function R(d) using (r0, s0) ¡ standardized family at h0, we need to estimate extra three parameters r0, s0, and h0. For avoiding estimating extra three parameters, Czado (1997) proposed to choose r0 D b0, s0 D 1, and h0 D b0. By choosing the values such a way, the variance inﬂations of b are reduced P as the values of h vary around the point h0 D b0, when centered covariates (i.e., d D n1 niD 1 di D 0) are used (Czado 1997). For this, if dose levels are not centered, we need to transfer the available dose levels as xi D di ¡ d, and after estimating BMD/BMDL from the model, we make the inverse transformation to get the estimates of BMD/BMDL in the true range of dose levels. For constructing (r0 D b0, s0 D HUMAN AND ECOLOGICAL RISK ASSESSMENT 5 1) ¡ standardized family at h0 D b0, we adopt the methodologies given by Czado (1997). Here, we use Stukel’s (Stukel 1988) generating family to construct the family of response functions, and the (r0 D b0, s0 D 1) ¡ standardized at h0 D b0 generating family is given by if hc 0 [logitðr Þb0 ], 8 expða1 hc Þ ¡ 1 > > ; a1 > 0 > > < a1 a1 D 0 Gc ða; hÞ D b0 C hc ; > > log ð 1 ¡ a h Þ > 1 c >¡ : ; a1 < 0; a1 (4) and for hc < 0 [logitðr Þ < b0 ], 8 1 ¡ expð ¡ a2 hc Þ > > ; > > < a2 Gc ða; hÞ D b0 C hc ; > > logð1 C a2 hc Þ > > : ; a2 a2 > 0 a2 D 0 (5) a2 < 0; where h h(d), r R(d), hc D h ¡ b0, and logitðr Þ D log½r=ð1 ¡ r Þ. Hence, the risk function R(d) using the binomial model with (r0 D b0, s0 D 1) ¡ standardized at h0 D b0 generating family is given by Rðd Þ D Eðyjd Þ D h½a; hðd Þ D exp½Gc ða; hÞ ; 1 C exp½Gc ða; hÞ (6) where h h(d), and Gc ða; ¢Þ is given by Eqs. (4) and (5). In the next section, we provide an expression for the benchmark dose (BMD) using the above model. Benchmark dose estimation The benchmark dose (BMD) is deﬁned by the dose level having extra risk RE ðBMDÞ D BMR, where BMR is the benchmark risk usually pre-speciﬁed as 0.01, 0.05, and 0.1. So, BMD is the solution of the equation RE ðBMDÞ D ) RðBMDÞ ) BMD D D RðBMDÞ ¡ p0 D BMR 1 ¡ p0 p0 C ð1 ¡ p0 ÞBMR D BMRE; say R ¡ 1 ðBMREÞ; (7) where p0 is the background risk, i.e., the risk at the minimum dose level d1. We denote 0 0 0 d D ½b ; a for the joint parameter vector including the regression parameter vector b, and the response parameter vector a. For a ﬁxed value of BMR 2 ½0; 1, the BMD can be expressed as a function of d, SðdÞ, say. From Eqs. (6) and (7), we get an expression for 6 I. DAS BMD as BMD D SðdÞ D S1 ðdÞIfLBMRb0 g C S2 ðdÞIfLBMR < b0 g ; (8) where LBMR D log 1 ¡BMRE BMRE , and IfLBMRb0 g is the indicator function taking value 1 if LBMRb0 , and 0 otherwise. The functions S1 ðdÞ and S2 ðdÞ are given by 8 log½a1 ðLBMR ¡ b0 Þ C 1 > > ; > > > a1 b1 > > < LBMR ¡ b0 ; S1 ðdÞ D > b1 > > > > 1 ¡ exp½ ¡ a1 ðLBMR ¡ b0 Þ > > ; : a1 b1 a1 > 0 a1 D 0 a1 < 0; and, 8 log½1 ¡ a2 ðLBMR ¡ b0 Þ > > ; a2 > 0 ¡ > > > a2 b1 > > < LBMR ¡ b0 ; a2 D 0 S2 ðdÞ D > b1 > > > > exp½a2 ðLBMR ¡ b0 Þ ¡ 1 > > ; a2 < 0: : a2 b1 P Note that we require centered dose levels (i.e., d D n1 niD 1 di D 0) for using the model (6). If the dose levels are not centered, we make the transformation xi D di ¡ d to have the centered dose levels. After estimating BMD from the model we make the inverse transformation to get the estimated value of BMD within the true range of dose levels. Asymptotic results The asymptotic distributions of unknown parameters for q-dimensional multinomial response models with a family of response functions are discussed in Das and Mukhopadhyay (2014). For q D 1, we get the binomial models using a family of response functions. So, the similar results can be applicable for binomial models using a family of response functions. However, for making this article self-contained, we provide the required asymptotic 0 0 0 0 0 0 0 b ;a b for the MLE of d D ½b ; a , lðdÞ for the logresults here. We denote b d D ½b @l for the score function for the observed responses. Also, Jn is likelihood function, and @d denoted for the Fisher’s information matrix. The asymptotic results are given by the following Lemmas. Lemma 1. The score function mean 0 and variance Jn. @l @d has an asymptotic multivariate normal distribution with HUMAN AND ECOLOGICAL RISK ASSESSMENT 7 Proof. From the section “Methods”, the risk function is given by, Rðd Þ D h½a; hðd Þ; (9) 0 where hðd Þ D f ðd Þb, b is an unknown regression parameter vector and a is a vector of unknown response parameters. Also, 0 hðd Þ D f ðd Þb D g ½a; Rðd Þ; (10) where g is the inverse of h. & Now, from the section “Methods”, the log-likelihood function for the sample y1, …, yn is given by lðdÞ D n X li ðdÞ iD1 D n X iD1 (11) yi ui ¡ bðui Þ ni C constant: Thus, the score function is (Fahrmeir and Tutz 2001), @lðdÞ @d n @X y ui ¡ bðui Þ ni @d i D 1 i n X @ri ¡ 1 Var yi yi ¡ ri ; @d iD1 D D (12) and (Fahrmeir and Tutz 2001) ¡ @2 lðdÞ 0 @d@d D D n X @ri iD1 @d Var yi Hn ; ðsayÞ: ¡1 n @ri X @2 ui ¡ yi ¡ ri ni 0 0 @d i D 1 @d@d (13) From Eq. (13), we get the Fisher information matrix that is X n @2 lðdÞ @ri Var yi D Jn D ¡ E 0 @d @d@d iD1 ¡1 @ri 0 : @d (14) & From Eq. (12), using the central limit theorem we have @l@dðdÞ has asymptotic normal distribu& tion with mean 0 and variance Jn. Lemma 2. The MLE of d, b d has an asymptotic multivariate normal distribution with mean d and variance Jn¡1. 8 I. DAS Proof. By Taylor series expansion and approximating up to ﬁrst-order term, we have 0D @l bd @d 2 @lðdÞ @ lðdÞ b C D d¡d ; 0 @d @d@d which gives (Fahrmeir and Tutz 2001, p. 439), pﬃﬃﬃﬃ pﬃﬃﬃﬃ @lðdÞ pﬃﬃﬃﬃ ¡ 1 @lðdÞ D N Jn C Op N ¡ 1=2 : N bd ¡ d D N Hn¡ 1 @d @d Thus, the MLE of d, bd has an asymptotic normal distribution with mean d and variance & Jn¡1. d DS b Lemma 3. The estimate BMD d is a consistent estimator for BMD. Proof. The proof is trivial from the result that the MLE of d, bd is a consistent estimator of d, b and SðdÞ is a continuous function of d. Hence, S d is a consistent estimator for SðdÞ, i.e., d is a consistent estimator for BMD. BMD & In the next section, we provide conﬁdence intervals for BMD to ﬁnd BMDL from the proposed model. Conﬁdence intervals Here, we provide four methods of constructing conﬁdence intervals for BMD for a particular value of BMR D BMR0 . The methods are discussed as follows: 1. Conﬁdence interval using ML estimates: here, we use the asymptotic result of the distribution of bd for constructing the conﬁdence interval for BMD. From Lemma 2, we have bd has an asymptotic multivariate normal distribution with mean d and variance 0 d ¡ d has an asymptotic x2-distribution with p S D Jn¡1. Hence, bd ¡ d S ¡ 1 b degrees of freedom, where p is the order of the vector d. Hence, the 100(1 ¡ t)% conﬁdence region for d is given by 0 C D fd 2 Rp : bd ¡ d S ¡ 1 bd ¡ d x2p;ð1 ¡ tÞ g; (15) where x2p, (1 ¡ t) is the (1 ¡ t)th quantile of the x2 distribution with p degrees of freedom. For BMR D BMR0 , we compute BMD D SðdÞ, for d 2 C using Eq. (8).Let us denote SL SU D D MinfSðdÞ : d 2 Cg; and MaxfSðdÞ : d 2 Cg: (16) Now, from (16), we have d 2 C ) SðdÞ 2 ½SL ; SU , which implies PðSðdÞ 2 ½SL ; SU Þ Pðd 2 CÞ D 1 ¡ t. Hence, the 100(1 ¡ t)% conservative conﬁdence interval for BMD is given by [SL, SU]. HUMAN AND ECOLOGICAL RISK ASSESSMENT 9 2. Conﬁdence interval using LR test: here, we test the null hypothesis H0 : RE ðd Þ D BMR0 vs: H1 : RE ðd Þ 6¼ BMR0 ; (17) where RE ðd Þ D Rð1d¡Þ ¡p0p0 , with p0 is the background risk. Let D(d) be the deviance (Fahrmeir and Tutz 2001) under nullhypothesis and D b d be the deviance of the ﬁt ted model. Then, Lðd Þ D Dðd Þ ¡ D b d has an asymptotic x2-distribution with 1 degree of freedom. Let us denote Lmin D Minfd 2 R : Lðd Þx2p;ð1 ¡ tÞ g; and Lmax D Maxfd 2 R : Lðd Þx2p;ð1 ¡ tÞ g (18) Then, the 100(1 ¡ t)% conﬁdence interval for BMD is given h by i ½Lmin ; Lmax . 3. Conﬁdence interval using score test: let us denote u0 D @b@l b , where bd 0 is the MLE 0 d b 20 be the estimated variance0 of u0 at d D bd 0 . Then, of d under H0 given in Eq. (17). Let s s 20 has an asymptotic x2 distribution with 1 degree of freedom (Fahrmeir T ðd Þ D u20 =b and Tutz 2001). Let us denote Tmin D minfd 2 R : T ðd Þx21;ð1 ¡ tÞ g, and Tmax D maxfd 2 R : T ðd Þx21;ð1 ¡ tÞ g. Then, using score test, a 100(1 ¡ t)% conﬁdence interval for BMD is ½Tmin ; Tmax . Note that the above conﬁdence intervals are by nature two-sided. To get one-sided conﬁdence interval (BMDL), some researchers (Nitcheva et al. 2005; Buckley et al. 2009) proposed an adjustment by doubling the signiﬁcance level of the test and then ignoring the upper limit. So, we construct 100(1 ¡ 2t)% two-sided conﬁdence intervals for BMD using the above methods and then the lower limits of that intervals are taken as the one-sided 100(1 ¡ t)% lower conﬁdence bound for BMD. Since all of the above conﬁdence intervals are constructed using asymptotic results, here we provide a conﬁdence interval using a bootstrap technique. 4. A bootstrap lower conﬁdence bound: for constructing bootstrap lower conﬁdence bound, we generate l responses yk D [y1k, y2k, y3k, y4k]0 using the ﬁtted model br i D h 0 b where yik has binomial distribution with parameter b; b ½a h ðdi Þ D f ðdi Þb, h ðdi Þ with b ðn;br i Þ for i D 1, …, 4 and k D 1, …, l. From the generated l data sets, we estimate BMDs using the proposed method to have samples fBMD1 ; BMD2 ; …; BMDl g for BMD. The 100(1 ¡ t)% bootstrap lower conﬁdence limit for BMD is given by the tth quantile of the sample fBMD1 ; BMD2 ; …; BMDl g. Let us denote ML, LR, ST, and BT for the methods of estimating BMDL using ML estimates, LR test, score test, and bootstrap technique, respectively. Example and simulation studies are provided to illustrate and test the performance of the proposed methods in next sections. Example: Experiment on rats exposed to 1-bromopropane For illustrating the proposed method, we present an example of estimating BMD using a real data set on lung cancer incidence of rats exposed to 1-Bromopropane given in the NTP Technical Report TR-569 (Program et al. 2011). In this study, four groups of rats with each group contains 50 rats are exposed to four dose levels of 1-Bromopropane. After 2 years of 10 I. DAS studies, the observed lung cancer incidence of rats at four dose levels 0 ppm, 62.5 ppm, 125 ppm, and 250 ppm is recorded as 1/50, 9/50, 8/50, and 14/50, respectively. Wheeler and Bailer (Wheeler and Bailer 2012) also analyzed the same data set noting that “the data exhibit a linear or supra-linear response indicating that MA may not be able to capture the true D-R relationship.” Here, we use the proposed method of using a family of response functions (FR) to estimate BMD and corresponding BMDLs using ML estimates (ML), likelihood ratio test (LR), score test (ST), and bootstrap technique (BT) as described in the section “Methods”. Wheeler and Bailer (2012) provide the estimates of BMD and corresponding BMDLs using semi-parametric (diffuse), semi-parametric (historical controls), model-averaging, and quantal-linear. In Table 1, we report the estimated values of BMD using FR and estimated values of BMDLs using four methods ML, LR, ST, and BT within bracket for BMR = 0.01, and 0.1. Also, we report the estimated values of BMD and corresponding BMDLs in bracket using semiparametric (diffuse), semi-parametric (historical controls), model-averaging, and quantal-linear for BMR = 0.01, and 0.1 derived from Wheeler and Bailer (2012) in Table 1. From Table 1, we observe that the estimated values of BMD using FR are consistent with the estimated values of BMD using semi-parametric (diffuse), semi-parametric (historical controls), and quantal-linear for all values of BMR. As mentioned in Wheeler and Bailer (2012), the estimated values of BMD using MA diverge and smaller than those using other methods. We observe that the estimated values of BMDLs using FR are higher than those reported by Wheeler and Bailer (2012). So, the proposed methods may provide better estimates of BMDL if the estimated conﬁdence intervals have the expected coverage probabilities for small samples. So, we need to do simulation studies for verifying the coverage probabilities of the proposed conﬁdence intervals for small samples. Simulation studies In this section, we conduct simulation studies for testing the performance of the proposed methods of estimating BMD and BMDL considering all types of possible cases of generating datasets. Let us denote the proposed method of estimating BMD using the family of response functions as FR. The model averaging method is usually denoted as MA. For testing the performance of FR compare to MA, we provide a simulation study by estimating BMD using Table 1. Estimated values of BMD (ppm) and the corresponding BMDL (ppm) within bracket by four methods calculated using the family of response functions, and also the same derived from Wheeler and Bailer (2012) by semi-parametric (diffuse), semi-parametric (historical controls), model-averaging, and quantallinear. BMR Method Family of response functions Semi-parametric (Diffuse) Semi-parametric (Historical controls) Model-averaging Quantal-linear 0.01 0.1 8.6 (6.3, 6.2, 6.2, 7.4) 6.1 (2.1) 68.9 (61.8, 50.0, 49.8, 57.8) 56.6 (17.5) 6.6 (1.6) 97.1 (23.1) 1.1 (0.14) 7.8 (5.2) 51.1 (17.2) 81.5 (55.0) HUMAN AND ECOLOGICAL RISK ASSESSMENT 11 Table 2. Six scenarios for the dose–response curves with true values for d, the probabilities of adverse events at dose levels, and true BMDs corresponding to BMR = 0.01 and 0.1. BMR d D ½b0 ; b1 ; a1 ; a2 Scenario 1 2 3 4 5 6 ’ [ ¡ 4.5031, 4.9075, 0.1170, 1.5162]0 [ ¡ 2.9252, 4.9961, 1.9078, ¡1.1403]0 [ ¡ 1.3677, 2.4678, 1.6912, ¡0.8872]0 [ ¡ 0.7784, 3.9106, 1.6554, ¡0.8438]0 [ ¡ 0.3852, 4.7828, 1.9908, ¡0.0870]0 [1.9190, 3.9682, 0.9064, 0.6930]0 R(d1) R(d2) R(d3) R(d4) 0.01 0.1 0.0000 0.0176 0.1067 0.1374 0.0905 0.1909 0.0015 0.0276 0.1474 0.2060 0.2229 0.7202 0.0149 0.0760 0.2330 0.3829 0.5058 0.9000 0.0224 1.0000 0.9856 1.0000 1.0000 0.9999 0.4200 0.2475 0.0690 0.0430 0.0260 0.0039 0.8535 0.5418 0.4195 0.2914 0.1912 0.0365 FR and MA considering different simulation scenarios with varying sample sizes. Simulation studies are also conducted for testing the performance of the proposed methods of estimating BMDL with respect to their coverage probabilities for small samples. Comparison between FR and MA We compare the proposed method FR with MA considering the following simulation set up with experimental design consists of four dose levels as d1 D 0, d2 D 0.25, d3 D 0.5, and d4 D 1.00 mimicking the design considered by West et al. (2012). Six scenarios for dose–response relationships have been considered to represent all types of possibilities of having probabilities of adverse events at dose levels varying from shallow to steep curves. The scenarios with true parameter values and the probabilities of adverse events at dose levels are given in Table 2. For each scenario, responses (yi) are generated from binomial distribution with parameter [n, R(di)], where n is the number of patients administered the dose level di, i 2 {1, …, 4}. For testing the performance of the methods with varying sample sizes and BMR, we consider d MA . Table 3. Eight standard models used in MA for computing BMD M 1 Name Logistic 2 Probit 3 Quantal-linear R(d) 1 1 C expð ¡ b0 ¡ b1 dÞ 1 b1 F(b0 C b1)d 1 ¡ exp ( ¡ b0 ¡ b1d) 4 Quantal-quadratic g 0 C (1 ¡ g 0)(1 ¡ exp [b1d ]) 5 Two-stage 1 ¡ exp ( ¡ b0 ¡ b1d ¡ b2d2) 2 BMD ¡b 0 BMR log 1 C1 e¡ BMR Notes None F ¡ 1 ½BMRð1 ¡ ’0 Þ C ’0 ¡ b0 b1 f0 D F(b0) ¡ logð1 ¡ BMRÞ b1 b0 0, b1 0 qﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ¡ logð1 ¡ BMRÞ b1 ¡ b1 C pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ ﬃ 2 b1 C 4b2 T 2b2 0 4 g 0 4 1, b1 0 bj 0, j D 0, 1, 2 T D ¡ logð1 ¡ BMRÞ 6 Log-logistic 7 Log-probit 8 Weibull g0 C 1 ¡ g0 1 C expð ¡ b0 ¡ b1 log½dÞ g 0 C (1 ¡ g 0)F[b0 C b1log (x)] b0 b1 g 0 C ð1 ¡ g 0 Þ½1 ¡ expð ¡ e d Þ exp L ¡ b0 b1 h ¡1 i Þ ¡ b0 exp F ðBMR b1 h i exp logðTbÞ ¡ b0 1 0 4 g 0 4 1, b1 0 L D log 1 ¡BMR BMR 0 4 g 0 4 1, b1 0 0 4 g 0 4 1, b1 0 T D ¡ logð1 ¡ BMRÞ 12 I. DAS two different values of BMR D 0:01; and 0:1 and three different sample sizes n D 25, 50, 100 for each dose–response curve. This provides in total 6 curves £ 2 values of BMR £ 3 sample sizes = 36 different cases. For getting model averaging estimates of BMD, eight standard models (West et al. 2012) given in Table 3 are considered. The expression for model d MA is given by averaging estimates of BMD, denoted as BMD d MA D BMD 8 X bk; wk B (19) kD1 b k is the estimate of BMD using model k, and wk D P8exp ð ¡ 0:5Ak Þ where B kD1 exp ð ¡ 0:5Ak Þ with Ak is the L k C 2pk , where b L k is the maxAkaike Information Criteria (Akaike 1973) given by Ak D ¡ 2b imized log-likelihood value and pk is the number of parameters in model k. We generate l D 2000 data sets for each simulation set up. For some cases the simulated responses produce virtually ﬂat dose–response curve (Wheeler and Bailer 2009) which does not give any ﬁnite estimate of BMD. So, we mimic the methodology given in Wheeler and Bailer (2009) of screening the data sets using the Kendall correlation test (Kendall 1955). We regenerate responses until the responses exhibit the Kendall p-value less than or equal to 0.15. For each case described above, we estimate BMDs by FR and MA using 2000 simulated data sets. The proposed method FR is compared with MA on the basis of observed absolute relative d B MD ¡ BMD (Wheeler and Bailer 2007). median bias deﬁned by the absolute value of median BMD The estimated values of BMD are used as a sample of size 2000 for computing absolute relative median bias by FR and MA. Smaller values of absolute relative median bias’ are desirable for having better performance by a BMD estimation method. The absolute relative median bias (ARMB) values by FR and MA for each simulation set-up are reported in Table 4. Table 4. Comparison between FR and MA for accounting model uncertainty in BMD estimation. The observed values of absolute relative median bias’ for FR and MA for different cases are reported in the columns FR and MA, respectively. Sc n BMR FR MA Sc n BMR FR MA 1 25 0.01 0.1 0.01 0.1 0.01 0.1 0.01 0.1 0.01 0.1 0.01 0.1 0.01 0.1 0.01 0.1 0.01 0.1 0.2043 0.0691 0.1083 0.0298 0.0433 0.0460 0.0600 0.0520 0.0997 0.0212 0.0336 0.0193 0.1059 0.0729 0.0773 0.0428 0.1487 0.0248 0.3149 0.1814 0.1384 0.1321 0.1392 0.1228 0.6192 0.0247 1.6802 0.4306 1.7606 0.4790 5.7751 0.4777 5.8792 0.5366 5.4006 0.4850 4 25 0.01 0.1 0.01 0.1 0.01 0.1 0.01 0.1 0.01 0.1 0.01 0.1 0.01 0.1 0.01 0.1 0.01 0.1 0.2480 0.1420 0.2925 0.1563 0.3092 0.1834 0.5322 0.3872 0.4079 0.2450 0.2973 0.2243 0.4644 0.3822 0.2919 0.2101 0.1474 0.1618 7.9541 0.8534 10.1225 1.2207 10.4497 1.2917 5.2402 0.7904 7.9412 1.3338 11.8273 2.4060 4.3147 2.1070 4.3990 2.1520 4.3084 2.5693 50 100 2 25 50 100 3 25 50 100 50 100 5 25 50 100 6 25 50 100 HUMAN AND ECOLOGICAL RISK ASSESSMENT 13 From Table 4, we see that the observed values of ARMB by FR are very close to those by MA for all values of n and BMR in Scenario 1. So, FR and MA have comparable performance with respect to their observed ARMB values for Scenario 1. Note that the chosen curve for generating data sets in Scenario 1 has very slowly increasing probabilities of adverse events at dose levels with R(d1) D 0, and R(d4) D 0.0224. So, it can be concluded that FR and MA provide comparable performance for extremely shallow dose–response curves. If we move toward less shallow dose–response curves (Scenarios 2–6), we see that the observed values of ARMB by FR are smaller than those by MA. For example, in Scenario 4 with BMR D 0:01, the values of ARMB are 0.5322, 0.4079, and 0.2973 by method FR, and 5.2402, 7.9412, and 11.8273 by method MA for sample sizes n D 25, 50, and 100, respectively. Also, for the same Scenario with BMR = 0.1, the values of ARMB are 0.3872, 0.2450, and 0.2243 by method FR, and 0.7904, 1.3338, and 2.4060 by method MA for sample sizes n D 25, 50, and 100, respectively. This shows that the values of ARMB by FR are smaller than those of ARMB by MA for these cases. Hence, FR performs better than MA with respect to their observed ARMB values for Scenarios 2–6. Also, it is noted that the values of ARMB by MA increase with sample sizes for some scenarios. This shows that the estimates by MA are asymptotically biased when the true models are not included in the model space of MA to estimate BMD. Comparison among four BMDL estimation methods Here, we conduct simulation studies to compare four methods of estimating BMDL using ML estimates (ML), likelihood ratio test (LR), score test (ST), and bootstrap technique (BT) with respect to their observed coverage probabilities for small samples. We choose similar simulation set-up considered in the section “Comparison between FR and MA” with the experimental design d D [0.0, 0.25, 0.5, 1.0]0 and scenarios given in Table 2 to generate data sets. We also consider two values of BMR D 0:01 & 0.1, and three sample sizes n D 25, 50, and 100 for each scenario. For each simulation set up, we generate l D 1000 data sets which are also screened by the Kendall correlation test (Kendall 1955) as discussed in the section “Comparison between FR and MA”. The simulated data sets are used to estimate 95% BMDL using the four methods ML, LR, ST, and BT. After estimating BMDL using a method, we ﬁnd an approximate value of coverage probability given by Nl l , where Nl is the number of times the estimated values of BMDL are less than or equal to BMD out of l datasets generated. The coverage probabilities by four methods ML, LR, ST, and BT for each simulation set-up are reported in Table 5. From Table 5, we see that the observed coverage probabilities by LR and ST are greater than 0.95 for all scenarios and BMR values with all sample sizes. The method BT fails to provide the expected coverage probabilities for Scenario 2 with n D 25, 50 and Scenarios 5 and 6 for all sample sizes when BMR = 0.01. The observed coverage probabilities by ML also exceed the expected probability 0.95 for all the cases except for Scenario 5 with n D 50, 100, when BMR = 0.1. We also studied the observed average lengths of the one-sided conﬁdence d intervals (average of [BMD-BMDL]) by all four methods. it was observed that BT provides d the smallest values and ML and ST provide largest values for the average of (BMD-BMDL) for all the cases considered. Hence, we conclude that LR is best among all the methods of estimating BMDL with respect to coverage probabilities and lengths of the conﬁdence intervals. 14 I. DAS Table 5. Comparison of four methods of estimating BMDL with respect to their coverage probabilities for different simulation set-up. The observed coverage probabilities of ML, LR, ST, and BT are given against the column ML, LR, ST, and BT respectively. Methods Methods Sc n BMR ML LR ST BT Sc n BMR ML LR ST BT 1 25 0.01 0.1 0.01 0.1 0.01 0.1 0.01 0.1 0.01 0.1 0.01 0.1 0.01 0.1 0.01 0.1 0.01 0.1 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 0.99 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 0.99 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 0.95 0.68 1.00 0.91 1.00 0.98 1.00 1.00 1.00 1.00 1.00 1.00 1.00 4 25 0.01 0.1 0.01 0.1 0.01 0.1 0.01 0.1 0.01 0.1 0.01 0.1 0.01 0.1 0.01 0.1 0.01 0.1 1.00 1.00 1.00 1.00 1.00 0.99 1.00 0.97 1.00 0.94 1.00 0.91 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 0.99 0.99 0.99 0.99 0.98 0.97 0.99 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 0.98 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 0.88 1.00 0.93 1.00 0.95 1.00 0.76 1.00 0.77 1.00 0.83 1.00 50 100 2 25 50 100 3 25 50 100 50 100 5 25 50 100 6 25 50 100 Computer programs used for computations A language and environment for statistical computing called R (R Core Team 2017) are used for all computations. The MLEs are computed by minimizing the negative of the log-likelihood using an R function “optim.” The Fisher information matrix is found by computing gradient of a function using an R-package “numDeriv” (Gilbert and Varadhan 2016). An Rpackage “Kendall” (McLeod 2011) is used for computing Kendall rank correlation of a simulated data. An R-function “optimize” is also used to compute a minimum of a function within a restricted domain. Conclusions For accounting model uncertainty in BMD estimation, a family of response functions for binary response models is used to develop a method for estimating BMD. The family of response functions provides local orthogonality between response and regression parameters to reduce the variance inﬂations of the estimated regression parameters. An inﬁnite number of response functions including some standard response functions are the members of the family. For accounting model uncertainty in BMD estimation, the family of response functions provides a better approach than model averaging method as the model space considered in MA to get model-averaged estimate usually contains only a ﬁnite number of models. Methods of estimating BMDL are also provided using the family of response functions. The proposed method is illustrated by an example with a real data set observing that FR is consistent with the existing results in the literature. By comparing FR with MA using simulation studies considering different simulation scenarios, we see that FR outperforms MA for most of the scenarios. Simulation studies are also conducted to compare the four methods of HUMAN AND ECOLOGICAL RISK ASSESSMENT 15 estimating BMDL and we see that the LR is best among the four methods of estimating BMDL considering both the coverage probability as well as the length of the conﬁdence intervals. If the true model generating a data cannot be approximated by any member of the proposed inﬁnite family of models, then the results using the proposed method may not be good. For this, we need to diagnose the model ﬁtting with respect to some measures such as deviance. In such cases, a non-parametric approach may be a better alternative. There are other methods exist in the literature using Bayesian and non-parametric approach for accounting model uncertainty in BMD estimation. However, the frequentist methods are usually easy to implement and require less time for computations than the non-frequentist approaches. We have compared FR with MA as both the methods are based on frequentist approach to deal with the model uncertainty problems. In future, the proposed method may be compared with other approaches such as non-parametric, Bayesian in estimating BMD to test the performance of FR. Funding This work was supported by the University Grants Commission (F.NO.5-181/2016(IC)). References Akaike, H. 1973. Information theory and an extension of the maximum likelihood principle. In: Petrov, BN, Csaki, B (eds), Proceedings of the Second International Symposium on Information Theory, pp. 267–81. Akademiai Kiado, Budapest Bailer, AJ, Noble, RB, and Wheeler, MW. 2005. Model uncertainty and risk estimation for experimental studies of quantal responses. Risk Anal 25(2):291–9 Buckley, BE, Piegorsch, WW, and West, RW. 2009. Conﬁdence limits on one-stage model parameters in benchmark risk assessment. Environ Ecol Stat 16(1):53–62 Cox, DR and Reid, N. 1987. Parameter orthogonality and approximate conditional inference. J R Stat Soc 49:1–39 Crump, KS. 1984. A new method for determining allowable daily intakes. Fundam Appl Toxicol 4 (5):854–71 Czado, C. 1989. Link misspeciﬁcation and data selected transformations in binary regression models. Technical Report. Ph.D. Thesis. School of Operations Research and Industrial Engineering, Cornell University, Ithaca, NY Czado, C. 1997. On selecting parametric link transformation families in generalized linear models. J Stat Plan Infer 61:125–39 Das, I and Mukhopadhyay, S. 2014. On generalized multinomial models and joint percentile estimation. J Stat Plan Infer 145:190–203 EFSA. 2017. Guidance on the use of the benchmark dose approach in risk assessment. EFSA J 15:4658–99 Fahrmeir, L and Tutz, G. 2001. Multivariate Statistical Modelling Based on Generalized Linear Models, 2nd ed. Springer, New York Gilbert, P and Varadhan, R. 2016. numDeriv: Accurate Numerical Derivatives. R Package Version 2016.8-1. Kang, SH, Kodell, RL, and Chen, JJ. 2000. Incorporating model uncertainties along with data uncertainties in microbial risk assessment. Regulat Toxicol Pharmacol 32(1):68–72 Kendall, MG. 1955. Rank Correlation Methods. Hafner Publishing Co, New York McLeod, A. 2011. Kendall: Kendall Rank Correlation and Mann-Kendall Trend Test. R Package Version 2.2. 16 I. DAS Morales, KH, Ibrahim, JG, Chen, CJ, et al. 2006. Bayesian model averaging with applications to benchmark dose estimation for arsenic in drinking water. J Am Stat Assoc 101(473):9–17 Nitcheva, DK, Piegorsch, WW, Webster West R, et al. 2005. Multiplicity-adjusted inferences in risk assessment: Benchmark analysis with quantal response data. Biometrics 61(1):277–86 OECD. 2008. Draft guidance document on the performance of chronic toxicity and carcinogenicity studies, supporting tg 451, 452 and 453. Technical Report. Organisation For Economic Co-Operation and Development, Paris Piegorsch, WW, An L, Wickens AA, et al. 2013. Information-theoretic model-averaged benchmark dose analysis in environmental risk assessment. Environmetrics 24(3):143–57 Program, NT et al. 2011. Toxicology and carcinogenesis studies of 1-bromopropane (cas no. 106-94-5) in f344/n rats and b6c3f1 mice (inhalation studies). National Toxicology Program Technical Report Series, (564):1 R Core Team. 2017. R: A Language and Environment for Statistical Computing. R Foundation for Statistical Computing, Vienna, Austria Shao, K and Small, MJ. 2011. Potential uncertainty reduction in model-averaged benchmark dose estimates informed by an additional dose study. Risk Anal 31(10):1561–75 Shao, K and Small, MJ. 2012. Statistical evaluation of toxicological experimental design for Bayesian model averaged benchmark dose estimation with dichotomous data. Hum Ecol Risk Assess Int J 18 (5):1096–119 Simmons, SJ, Chen, C, Li, X, et al. 2015. Bayesian model averaging for benchmark dose estimation. Environ Ecol Stat 22(1):5–16 Stukel, TA. 1988. Generalized logistic models. J Am Stat Assoc 83:426–31 Taylor, JMG. 1988. The cost of generalized logistic regression. J Am Stat Assoc 83:1078–83 US EPA. 2012. Benchmark dose technical guidance document. Technical Report #EPA/100/R-12/001, U.S. Environmental Protection Agency, Washington, DC West, RW, Piegorsch, WW, Pe~ na EA, et al. 2012. The impact of model uncertainty on benchmark dose estimation. Environmetrics 23(8):706–16 Wheeler, MW and Bailer, AJ. 2007. Properties of model-averaged BMDLs: A study of model averaging in dichotomous response risk estimation. Risk Anal 27(3):659–70 Wheeler, MW and Bailer, AJ. 2009. Comparing model averaging with other model selection strategies for benchmark dose estimation. Environ Ecol Stat 16(1):37–51 Wheeler, M and Bailer, AJ. 2012. Monotonic Bayesian semiparametric benchmark dose analysis. Risk Anal 32(7):1207–18