priors, utilities, elicitation & pharmaceutical r&d
DESCRIPTION
Priors, Utilities, Elicitation & Pharmaceutical R&D. Andy Grieve Statistical Research Centre Pfizer Global R&D. Outline. Use of Bayesian Methods in Pharmaceutical R&D Three Prior Elicitation Examples Acute toxicity – LD50 Sample Sizing & Confidence Intervals - PowerPoint PPT PresentationTRANSCRIPT
Priors, Utilities, Elicitation & Pharmaceutical R&D
Andy GrieveStatistical Research Centre
Pfizer Global R&D
2
Outline
• Use of Bayesian Methods in Pharmaceutical R&D
• Three Prior Elicitation Examples• Acute toxicity – LD50• Sample Sizing & Confidence Intervals• Counting Tablets in Dosing Dogs
• Elicitation for Internal Company Decision Making – Portfolio Management
04/21/23 CS4NS 3
Are Bayesian Methods Acceptable in Drug Development?
Not Forbidden by Regulation
04/21/23 CS4NS 4
Extracts from Drug Regulations
E4 : Dose Response 1993
Agencies should be open to the use of various statistical and pharmacometric techniques such as Bayesian and population methods, modelling, and pharmacokinetic-pharmacodynamic approaches.
= International Conference on Harmonisation
04/21/23 CS4NS 5
Extracts from Drug Regulations
CPMP Biostatistical Guidelines 1994
.. the use of Bayesian or other approaches may be considered when the reasons for their use are clear and when the resulting conclusions are sufficiently robust to alternative assumptions
= European Medicines
Evaluation Agency
04/21/23 CS4NS 6
Extracts from Drug Regulations
E9 : Statistical Principles in Clinical Trials 1998
Essentially same as CPMP Guidelines
.. the use of Bayesian or other approaches may be considered when the reasons for their use are clear and when the resulting conclusions are sufficiently robust (to alternative assumptions)
LUKEWARM !!!
04/21/23 CS4NS 7
Why ?
Clinical_Trials List (1996)
• Background• Hair-thinning • Researcher Bias
“If I hadn’t believed it, I wouldn’t have seen it with my own eyes”
- Trust
04/21/23 CS4NS 8
• Feeling that use of “subjective priors may allow unscrupulous companies and/or statisticians to attempt to pull the wool over the regulators eyes.” (Greg Campbell – FDA Centre for Devices & Radiological Health)
• If it were that easy they are not very good and we probably need new regulators
Trust
04/21/23 CS4NS 9
• Stephen Senn - “nowhere is the discipline of statistics conducted with greater discipline than in the pharmaceutical industry”
• Nowhere will Bayesian statistics be conducted with more discipline than in the pharmaceutical industry
• Document
Trust
04/21/23 CS4NS 10
• Document• Where did the prior come from ?• Is it based on data? Is it subjective ?
• “to present a Bayesian analysis in which the company’s own prior beliefs are used to augment the trial data will in general not be acceptable to a regulatory agency” (O’Hagan & Stevens, 2001)
• “the frequentist approach is less assumption dependent and can provide the statistical strength of evidence required for a confirmatory trial that may be lacking in a more assumption dependent Bayesian approach” (Chi, Hung & O’Neill – Biopharmaceutical Report, Vol. 9, No.2, 2001)
• SENSITIVITY ANALYSIS
Trust
11
• We work in a Frequentist World
Remember Acceptance of Bayesian methods is Lukewarm
• We will be asked about false positive rates• We will be asked about the impact of multiple
looks at the data• We need to be calibrated
Trust
12
Assessing a Prior in Acute toxicity
13
Motivating Example
Dose (mg/kg)
# of Animals
# of Deaths
500 5 1
1000 5 2
2500 5 3
5000 5 2
Based on these data we wish to determine the LD50 to classify the drug according
to the following classification scheme (Swiss Poison Regs.)Toxicity
Class1 2 3 4 5
Range of LD50
(mg/kg)
< 5 5-50 50-500 500-2000 2000-5000
14
Model
• Data triplets • { di , ni , ri } : i=1,..,k
• Probabilities of response• i : i=1,…,k
• Logistic Model • log [ i / (1-i)] = + log(di)
• Median Lethal dose (LD50)• log(LD50) =
Probit Model : i =( + log(di))
15
Bayesian Solutions
• Likelihood Function
• Prior distribution - p() ( > 0 )
• Define : - log(LD50)
• Inference
ii
i
rnk
ii
ri dGdGXL
1
))log((1))log(()|,(
0
)|,()|( dXpXp
U
L
dXpXP UL )|()|(
16
Likelihood Contours - Motivating Example
Dose (mg/kg
)
# of Animal
s
# of Deaths
500 5 1
1000 5 2
2500 5 3
5000 5 2
17
Likelihood Function - Hypothetical Example
Normal AnalyticApproximation
Dose (mg/kg
)
# of Animal
s
# of Deaths
100 3 1
1000 3 2
18
Motivating Example
Dose (mg/kg)
# of Animals
# of Deaths
500 5 1
1000 5 2
2500 5 3
5000 5 2
Experienced toxicologists will know that they need to span the LD50
with the doses they choose.
The choice of doses contains information concerning the toxicologist’s beliefs about the likely value of the LD50.
19
Choice of Prior 1) Tsutakawa (1975) : logit
• Choose doses d1 & d2
s.t. P(d1<LD50<d2)=0.5
0 10
1
1
2
2 1• Implies 1 and 2 uniform over the half square
• p) : logit n.c.p. probit BN (truncated)
(Grieve , 1988)
• Implies knowledge of i : i ≠1,2
20
Choice of Prior 2) Tsutakawa (1975) - logit
• Choose doses d1 & d2
iii
lmllml
ml
ˆ)2(1
)1()1( 12
12
11
11
222111
• p) : logit n.c.p. probit BN (truncated)
(Grieve, 1988)
• Implies knowledge of i : i ≠1,2
• Assume n.c.p. for 1 and 2
1̂ 2̂• Specify modal responses probabilities and
21
Choice of Prior 3) Grieve (1988) - probit
Toxicity Class
1 2 3 4 5
Probability 0.05 0.15 0.40 0.30 0.05
2
2
0
0 ,
BN• Suppose p() is bivariate normal :
• Can the parameters be determined ? • Not uniquely !!!• The c.d.f. of –depends only on :
430
20
1 ,,, cccc• Implying any 4 probabilities are sufficient to determine c1,c2,c3 and c4
• Any one of the 5 parameters is also needed • Modal slope ? How about median ?• Feedback•
22
Determining a Prior for Sample Sizing
04/21/23 CS4NS 23
Sample Sizing CIs : Simon Day (Lancet, 1988)
2n patients :
(1-)% CI :
Width :
Acceptable Width =
x x s1 22, ,
x x t sn
x x t sn
1 2 1 22 2
,
w t sn
22
w0 nt s
w
8 2
02
04/21/23 CS4NS 24
Alternative Approach : Grieve - Lancet, 1989
Required :
Solve by search
P w w
P t sn
w
Ps w n
t
( )
0
0
2
202
2 2
1
22
1
81
2
04/21/23 CS4NS 25
Simon Day’s Example
Two Anti-HypertensivesDifference in Diastolic BP - 95% CI = 10 mm Hg, w0=10
Grieve - Lancet , 1989
1- n 32 37 39 41
n=32
04/21/23 CS4NS 26
Never be absolutely certain of Never be absolutely certain of anythinganything
Bertrand RussellBertrand Russell
A Bayesian approach is an A Bayesian approach is an unconditional approach accounting unconditional approach accounting for uncertainty in parametersfor uncertainty in parameters
04/21/23 CS4NS 27
Beal - Biometrics , 1989
“ “ A prior estimate of A prior estimate of …… is …… is needed. This clearly introduces needed. This clearly introduces some uncertainty regarding the some uncertainty regarding the required sample size “required sample size “
ConditionalityConditionality
04/21/23 CS4NS 28
Relation Between and n
7 8 9 10 11 12 13 14
n 21 27 33 39 46 54 62 71
Suppose we have some idea about the likely value of through a probabilitydistribution
04/21/23 CS4NS 29
Unconditional Approach
P w w P w w p d( ) ( | ) ( ) 0 02 2 2
2
04/21/23 CS4NS 30
Where do we get p(2) from ?
• Previous studies• Expert opinion - subjective ?
• Estimate of 2 : based on 0 d.f.
• Inverse-Gamma prior
s02
2220
2200
22002
0
0
2
22/)(
/
)(/
)]/(sexp[)/s()(p
04/21/23 CS4NS 31
Conditional Formula
Unconditional Formula – Unconditional Formula – Grieve(1991)Grieve(1991)
P w w Pw n
t( )
0
2 02
2 281
P w w P Fw n
t s( ) ,
0
02
2020 8
1
32
Elicitation of Inverse-Gamma
• Expert provides and s.t.
• Not enough information – assume upper and lower limits are (1-p0)/2 percentiles
• Solve directly or modify algorithm in Martz and Waller(1982 – Bayesian Reliability Analysis), Grieve (1987,1991)
2L
2U
02
2220
2200
22002
2
2 0
0
2
22pd
)(/
)]/(sexp[)/s()(p
U
L/)(
/
04/21/23 CS4NS 33
Illustrative Example
Probability ( 8 < < 13 ) = 0.8
Implies
0=14.66 , =95.55
20s
04/21/23 CS4NS 34
Relation Between P(w<w0) and n
n 51 52 53 54 55 56 57
P(w<w0) 0.873 0.882 0.892 0.899 0.906 0.912 0.918
In this example accounting for uncertaintyincreases the sample size by 40 %
35
Elaborating a Prior for Tablet Counting
36
Checking the Dosing of Dogs
• dogs dosed on mg/kg basis• adjusted weekly
• Example• Unit Dose : 36 mg/kg• Weight : 19.2 kg• Required dose : 691 mg
37
Pre-Manufactured Tablet Strengths
300 mg300 mg 25 mg25 mg 5 mg5 mg 0.5 mg
691 mg
2 23 3
4 - 5 0 -11 0 - 4 0 - 4
38
Dog Dosing
• Tablets placed in a gelatine capsule
• Are the correct number of tablets in the capsule ?
39
Possible Approaches• Do Nothing
• Hope - No : Inspection• Acceptance Sampling
• too few samples - 308 capsules/wk• checking creates errors
• Check Everything• checking creates errors
• Weigh Capsules & Contents
40
Tablet /Capsule Weights
Tablet Strength (mg)
Mean (g) St. Dev.
300 0.602 0.0036
25 0.298 0.0035
5 0.150 0.0013
0.5 0.075 0.0008
Capsules 0.701 0.0410
41
Statistical Model
• T tablet sizes• tablet weights are : • capsule weights are :
• Ni tablets of each size chosen
• Total weight w is distributed as
42
Hypothetical Example
2 X 300 mg = 600 mg3 X 25 mg = 75 mg3 X 5 mg = 15 mg2 x 0.5 mg = 1 mg
691 mg
• Given a total weight of 3.397g (simulated)• What can we say about the likely numbers of
tablets?
43
Dog Dosing - Solution (1)
• Co-primal Weights• 3 : 7 : 13 : 23 instead of 1 : 2 : 4 : 8
44
Dog Dosing - Solution (2)
• Co-primal Weights• 3 : 7 : 13 : 23 instead of 1 : 2 : 4 : 8
• Pre-Weighing of Capsules
45
Dog Dosing - Weights
Tablet Strength (mg)
CV (%)
300 0.6
25 1.2
5 0.9
0.5 1.1
Capsules 5.8
46
Solution (3)
• Co-primal Weights• 3 : 7 : 13 : 23 instead of 1 : 2 : 4 : 8
• Pre-Weighing of Capsules• Prior distribution
belief in ability to count to 5
greater than
belief in ability to count to 19
47
Elaborating a Prior Grieve et al (1994)• Suppose a technician tries to count to M
tablets of a given strength• A model of the process could be :
• The total of M tablets is “achieved” by M individual operations each attempting to count to 1
• An error can be made in either direction : xj=0,1 or 2
• The total count is : x1+x2+ …. + xM=N
48
Elaborating a Prior
• Suppose the probability distribution of results from a single count is given by :
x 0 1 2
P(xj=x) q r pwith p.g.f. - P(t)=q+rt+pt2
212121
1
1
2
22
0 0 2121
2
)!(!
!
)()()(
kkMkkMkkM
k
cM
k
NNN
tprqkkMkk
M
ptrtqtPtP
•Assuming independent counts the p.g.f. of N is :
Mcn]/)n[(M
)nM,max(k
knMk prq)!Mkn()!knM(!k
!M)nN(P
21
0
22
22
• Giving :
49
Elaborating a Prior
ADVANTAGES
• A prior distribution need not be elicited for every M
• Elaboration ensures consistency• If Mk=Mj+1 then P(Nk=Mk) < P(Nj=Mj)
DISADVANTAGES
• Need to elicit p,q (r=1-p-q)
• Assumptions
50
Feeding Back
MValues of r ( p=q=(1-r)/2)
0.9 0.95 0.99 0.995
0.999
2P(N=M) 0.81
50.904 0.98
00.99
00.99
8
P(N=M1) 0.090
0.048 0.010
0.005
0.001
4P(N=M) 0.68
00.821 0.96
10.98
00.99
6
P(N=M1) 0.147
0.086 0.019
0.010
0.002
6P(N=M) 0.58
10.750 0.94
20.97
10.99
4
P(N=M1) 0.183
0.117 0.029
0.015
0.003
8P(N=M) 0.50
70.689 0.92
40.96
10.99
2
P(N=M1) 0.204
0.141 0.037
0.019
0.004
51
Dog Dosing - Conclusions
• Such a scheme is practicable• Computations trivial• Pre-weighing essential• Prior distribution essential
• Perfect for robotification