Transcript Modeling and Estimation of Benchmark Dose (BMD)

Modeling and Estimation of
Benchmark Dose (BMD)
for Binary Response Data
 Wei Xiong
1
Outline
 Benchmark dose (BMD) and datasets
 Statistical models
 logistic

probit

multi–stage

gamma multi–hit
 Model fitting and analyses
 Conclusions
2
In environment risk assessment,
NOAEL (no-observed-adverse-effect level)
is used to derive a safe dose,
NOAEL
ADI 
SF
where,
ADI: acceptable daily intake
SF: safety factor
3
Problem:
(Filipsson et al., 2003)
4
Benchmark dose (BMD)
 BMD: Point estimate of the dose which induces a
given response (e.g. 10%) above unexposed
controls
 BMDL: 1–sided 95% confidence lower limit
for BMD
5
Benchmark dose (BMD)
• Fit a model to all data
• Estimate the BMD
from a given BMR (10%)
• Derive “safe dose”
from BMD
Advantage: BMD uses all the
data information by fitting a model
(Filipsson et al., 2003)
6
Non–cancer data
Ryan and Van (1981)
i
1
2
3
di
24
27
30
ri
0
0
4
ni
30
30
30
4
5
6
34
37
40
11
10
16
30
30
30
7
8
45
50
26
26
30
30
30 mice in each
dose group
drug: botulinum toxin
in 10–15 gram
response:
death (Y or N)
within 24 hrs
7
Plot of non-cancer Data
0.6
0.4
0.2
0.0
Propn of Mice Mortality
0.8
Plot of Non-cancer Data
3.2
3.4
3.6
log(dose)
3.8
8
Cancer data
Bryan and Shimkin (1943)
i
di
ri
ni
1
3.9
0
19
2
7.8
3
17
3
15.6
6
18
4
31
13
20
5
62
17
21
6
125
21
21
17 to 21 mice in each
dose group
drug: carcinogenic
methylcholanthrene
in 10–6 gram
response:
tumor (Y or N)
9
Plot of cancer data
0.6
0.4
0.2
0.0
Propn of Mice Bearing Tumors
0.8
1.0
Plot of Cancer Data
2
3
4
log(dose)
10
How to estimate BMD ?
 What models to be used
 ? Need to use different models for the cancer
and non-cancer data
 How to fit the model curve
11
Statistical models
•
•
•
•
Logistic
Probit
Multi–stage
Gamma multi–hit
Model form:
P(d )    (1   ) F (d ;  ,  )
where,
1> >=0 is the background response as dose0
F is the cumulative dist’n function
12
Probit model
1 
P(d )   
2

   loge d

e
 x2 / 2
dx
Assuming:
log(d) is approx. normally distributed
13
Logistic model
P (d )   
1 
1 e
 (   log e d )
Assuming:
log(d) has a logistic distribution
14
Multi–stage model
(Crump, 1981)
P(d )    (1   )[1  exp( j 1  j d )]
n
j
Assuming:
1. Ordered stages of mutation, initiation or transformation
for a cell to become a tumor
2. Probability of tumor occurrence at jth stage is
proportional to dose by jd j
15
Gamma multi–hit model
(Rai and Van, 1981)
d

P(d )    (1   )

0

0
t
 1  t
e dt
t  1et dt
Assuming: a tumor incidence is induced by at least   1 hits
of units of dose and  follows a Poisson distribution
The gamma model is derived from the Poisson dist’n of 
16
Model fitting
 Models are fit by maximum likelihood method
 Model fitting tested by Pearson’s 2 statistic

m
2  
i 1
(ri  ni P i ) 2


ni P i (1  Pi )
~  n2 p

where,
Pi is estimated from the fitted model
 If p-value  10%, the model fits the data well and
the mle of BMD is obtained from the fitted model
17
BMDL by LRT
(Crump and Howe, 1985)



2[ (  ,  )  ( P ,  )]



=D(P ,  )  D(  ,  )  
2
1 2*0.05,1
where,
 and  are model parameters
 P is the log(BMD) at response = p
18
 The BMDL is the value P, which is lower than the

mle , so that,
P



2[ (  ,  )  ( P ,  )]
=
2
1 2*0.05,1
19
BMDL by Fieller’s Theorem
(Morgan, 1992)
Fieller’s Theom constructs CI for the ratio of R.V.
For logistic model,


    ~ N (0,V11  2V12   2V22 )
the BMDL is derived as,


V12
c
P(
)( P 
)
1 c
V22


2
V

V11  2 P V12   P 2 V22  c(V11  12 )
V22
 (1  c)
Z10.05
where,

2
c  Z 2(10.05)V22 / 
20
BMDL computation
 BMDS (benchmark dose software, US EPA)
provides the 4 models for BMDL using LRT
 S–Plus calculates BMDL using LRT and
Fieller’s Theorem
21
BMDS logistic modeling for non–cancer data
(Pearson’s 2, p = 0.325 > 0.1)
Log-Logistic Model with 0.95 Confidence Level
1
Log-Logistic
0.8
0.6
0.4
0.2
0
BMDL
25
BMD
30
35
40
45
50
dose
05:48 07/02 2005
22
BMDS multi–stage modeling for non–cancer data
(Pearson’s 2, p = 0.0000)
Multistage Model with 0.95 Confidence Level
1
Multistage
0.8
0.6
0.4
0.2
0
BMDL
BMD
10
20
30
40
50
dose
05:39 07/02 2005
23
BMDS two–stage modeling for cancer data
(Pearson’s 2, p = 0.556)
Multistage Model with 0.95 Confidence Level
Multistage
1
0.8
0.6
0.4
0.2
0
BMDL BMD
0
20
40
60
80
100
120
dose
07:02 07/01 2005
24
BMDL=1.536 by LRT
Lower CL= 1.5361391518069
3
2
1
0
Profile Deviance
4
5
(Probit model for cancer data)
1.4
1.6
1.8
2.0
25
x0 values
MLE of BMD
(non–cancer data)
Software
BMDS
S–plus
Model
Logistic
Probit
30.042
30.039
(0.325)
(0.386)
30.042
30.039
( p–value by
Pearson’s 2 )
26
Summary of BMDL
(non–cancer data)
Methods Software
BMDS
Model
Logistic Probit
28.143
28.296
S–plus
28.139
28.293
Fieller’s S–plus
Theorem
27.991
28.218
LRT
27
MLE of BMD
(cancer data)
Software
Logistic
BMDS
7.168
# 0.585
S–plus
7.171
Model
Probit
Two–
stage
7.203
4.867
# 0.666
# 0.556
Multi–
hit
6.334
# 0.602
7.199
# p–value by Pearson’s 2
28
Summary of BMDL
(cancer data)
Model
Methods Software
Logistic Probit
Two–
stage
Multi–
hit
BMDS
4.434
4.647
3.087
3.290
S–plus
4.434
4.647
Fieller’s S–plus
Theorem
4.181
4.519
LRT
29
Conclusions
 Non–cancer data, BMD = 30.042 (logistic) and 30.039
(probit) in 10–15 gram; cancer data, BMD = 7.168
(logistic), 7.203 (probit), 4.867 (multi–stage) and 6.334
(multi–hit).
 Logistic and probit model fit both data sets well, multi–
stage and multi–hit fit only the cancer data well.
 BMDL obtained by Fieller’s Theorem seems to be smaller
than that by LRT, why ?
30
Questions ?
31
A note on qchisq( ) of 1–sided 95%
 > (qnorm(1 - 0.05))^2
 [1] 2.705543
 > qchisq(1 - 2 * 0.05, 1)
 [1] 2.705543
32
95% CI for proportion in slides 21 & 22
 When n is large, nP  5 and n(1-P)  5, the sample
proportion p is used to infer underlying proportion P.
 p is approximately normal with mean P and
s.e.=sqrt(P(1-P)/n)
 Solving the following equation,
| p  P | 1/(2n)
 Z10.05/ 2
PQ / n
33
Fitted and re–parameterized model
Fitted logistic model
P
log e (
)   d
1 P
Re-parameterized logistic model
P
log e (
)  c   (d   P )
1 P
where,

c  log e (
)
1 
34
Abbott’s Formula
P  c  (1  c) BMR
where,
P – observed response
c – response at dose zero
BMR – benchmark response with
default value 10%
35