statistical methodology for evaluating a cell mediated immunity-based hiv vaccine
DESCRIPTION
Statistical Methodology for Evaluating a Cell Mediated Immunity-Based HIV Vaccine. Devan V. Mehrotra* and Xiaoming Li Merck Research Laboratories, Blue Bell, PA *e-mail: [email protected] Biostat 578A Lecture 4 - PowerPoint PPT PresentationTRANSCRIPT
Statistical Methodology for Evaluating a Cell Mediated Immunity-Based HIV
Vaccine
Devan V. Mehrotra* and Xiaoming LiMerck Research Laboratories, Blue Bell, PA
*e-mail: [email protected]
Biostat 578A Lecture 4Adapted from Devan’s presentation at the ASA/Northeastern Illinois Chapter Meeting
October 14, 2004
2
Outline
• Science behind the numbers• Merck’s HIV vaccine project• Proof of concept (POC) efficacy study• Statistical methods• Simulation study• Concluding remarks
3
Worldwide Distribution of HIV-1 Clades (Subtypes)*
Note: *Dominant clades are bolded above; All regions have multiple clades in their populations
B B, BC
B
C
A, B, AB, Other
G
B, F, Other
B, FB, AE
B, Other
AE, B, Other B, AE, Other
B, Other
B
O
B, O
A
A
All
C
C, Other
B, Other
A, Other
G, Other
G, OtherAG
A,C,D
Legend
B dominant + Another
C
O
A
All
B, AE
B, G, OtherC
B, C
F
Other
B
4
T Cell Recognition of Infected Cells
5
HIV Infection: CD4 cell count and Viral Load
6
Merck’s HIV Vaccine Project
• Lead vaccine is an Adenovirus type 5 (Ad5) vector encoding HIV-1 gag, pol and nef genes
• Goal: to induce broad cell mediated immune (CMI) responses against HIV that provide at least one of the following:
Protection from HIV infection: acquisition or sterilizing immunity.
Protection from disease: if infected, low HIV RNA “set point”, preservation of CD4 cells, long term non-progressor (LTNP)-like clinical state.
7
Proof of Concept (POC) Efficacy Study
• Design- Randomized, double-blind, placebo-controlled- Subjects at high risk of acquiring HIV infection- HIV diagnostic test every 6 mos. (~ 3 yrs. f/up)
• Co-Primary Endpoints- HIV infection status (infected/uninfected)- Viral load set-point (vRNA at ~ 3 months after diagnosis of HIV infection)
• Secondary/exploratory endpoints: vRNA at 6-18 months, rate of CD4 decline, time to initiation of antiretroviral therapy, etc., for infected subjects
8
POC Efficacy Study (continued)
• Vaccine Efficacy (VE) =
• Null Hypothesis: Vaccine is same as Placebo Same HIV infection rates (VE = 0) and Same distribution of viral load among infected subjs.
• Alternative Hypothesis: Vaccine is better than Placebo Lower HIV infection rate (VE > 0) and/or Lower viral load for infected subjects who got vaccine
• Proof of Concept: reject above composite null hypothesis with at least 95% confidence
PLACEBO}|infectionPr{HIV VACCINE}|infectionPr{HIV
1
9
Notation for Statistical Methodology
Vaccine Placebo
Number Randomized vN cN
Number I nf ected vn cn
Proportion I nfected v
vv N
np
c
cc N
np
log10(vRNA) of I nf ected Subjects
vnx
x
1
cny
y
1
Let cvv
c ppDNN
r , ,
vn
iixRankS
1
)(
(ranking done af ter pooling x’s and y’s together)
10
Notation (cont’d)
C o m p a r i s o n o f H I V i n f e c t i o n r a t e s
L e t p 1 = o n e - t a i l e d p - v a l u e u s i n g a n e x a c t c o n d i t i o n a l b i n o m i a l t e s t
))1
1,(~|Pr(1 r
nnBinomialBnBp cvv
L e t Z 1 = t e s t s t a t i s t i c f o r c o m p a r i n g i n f e c t i o n r a t e s u s i n g a n a p p r o x i m a t e c o n d i t i o n a l t e s t
)]()1/[(
)1/(1)/(
),|(
,|2
0
01
cv
cvv
cv
cv
nnrr
rnnn
HnnDVar
HnnDEDZ
U n d e r t h e n u l l , Z 1 i s a p p r o x i m a t e l y N ( 0 , 1 )
11
Notation (cont’d)
Comparison of Viral Load (among infected subjects)
Let p2 = one-tailed p-value using an exact conditional Wilcoxon Rank Sum test (PROC TWOSAMPL in Proc-StatXact)
Let Z2 = approximate conditional Wilcoxon Rank Sum test
),,|(
),,|(
0
02 HnnnSVar
HnnnSESZ
vcv
vcv
Under the null, Z2 is approximately N(0,1)
Note: p1 and p2 are stochastically independent
12
Competing Methods for Establishing POC
2- part- model (χ2) (Lachenbruch, 2001): Reject null if 2
1,222
21
2 ZZ
Note: Suitable f or 2-tailed test, but not f or our 1-tailed test of interest
2- part- model (Z) (O’Brien, 1984):
Reject null if 121
2Z
ZZZ
Note: both endpoints get equal weight
Weighted- Z (Follmann, 1995):
Reject null if 12
221
2211 Zww
ZwZwZ
where w1 and w2 are pre-specifi ed weights
Optimal weights: 1221
11 1,
)()()(
wwZEZE
ZEw
13
Optimal Weights for Viral Load Component of Composite Test (w2) under Different Scenarios
True VE (%)
δ = true mean diff. (placebo – vaccine) in log10(vRNA) among infected subjects
.5 .6 .8 1.0 1.3 1.5 2.0
0% ~1 ~1 ~1 ~1 ~1 ~1 ~1
15% .78 .81 .84 .86 .88 .89 .91
30% .62 .65 .71 .74 .78 .79 .81
45% .49 .53 .59 .63 .67 .69 .72
60% .38 .42 .48 .52 .57 .59 .62
75% .28 .31 .37 .41 .45 .48 .51
90% .17 .19 .23 .27 .31 .32 .36
14
Simes test (Simes, 1986): Reject null if 2/),min(),max( 2121 pporpp
Weighted- Simes (Hochberg and Liberman, 1994):
Reject null if 2/),min( ),max(2
2
1
1
2
2
1
1 vp
vp
orvp
vp
where ,2 ,2 2211 wvwv and w1 and w2 are pre- specifi ed weights (same as before)
Methods for Establishing POC (cont’d)
15
Methods for Establishing POC (cont’d)
Fisher’s combined p- value method (Fisher, 1932): Test is based on 2
12
1
21 ppq 4,log41log4Pr 2
)4( q--PROBCHIqvaluep ee
Reject null if valuep Note: both endpoints get equal weight
Weighted- Fisher’s method (Good, 1955): Test is based on 21
21~ ww ppq
12
2
21
12
11
1 ~~
wwqw
wwqw
valuepww
with 1w and 2w as before 21 ww Reject null if valuep
16
Illustration of Simes, Weighted-Simes, Fisher’s, Weighted-Fisher’s Methods (Hypothetical Examples)
Note: w1 = .14, w2 = .86 for weighted-Simes’ and weighted-Fisher’s methods
Composite p-value (Reject composite null hypothesis?)
p-value f or inf ection
endpoint (p1)
p-value f or viral load
endpoint (p2) Simes’ W-Simes’ Fisher’s W-Fisher’s 0.040
0.047 0.012 0.028
0.040 0.040 (Y) (Y) (Y) (Y)
0.080
0.047 0.037 0.035 0.150 0.040
(N) (Y) (Y) (Y) 0.048 0.028 0.065 0.026
0.500 0.024 (Y) (Y) (N) (Y)
0.040 0.071 0.014 0.063 (N)
0.020 0.100 (Y) (N) (Y) (N)
17
Critical Boundaries: Simes, Weighted-Simes, Fisher’s, Weighted-Fisher’s
Note: w1 = .14, w2 = .86 for weighted Fisher’s method. Boundaries are shown assuming p2 p1
p-value for the Viral Load Endpoint (p2)
p-v
alu
e f
or
the
HIV
-1 I
nfe
ctio
n E
nd
po
int
(p1
)
0.0 0.025 0.050 0.075 0.100 0.125 0.150
0.0
0.1
0.2
0.3
0.4
0.5
Simes (S)Fisher (F)Weighted-Simes (WS)Weighted-Fisher (WF)
F S WFWS
18
Additional Notation for Two Other MethodsBasic Idea: Plug in viral load = 0 for uninfected
subjects Vaccine Placebo
log10(vRNA) of I nf ected Subjects
vnx
x
1
cny
y
1
log10(vRNA) of Uninfected Subjects
vv nN
0
0
cc nN
0
0
Overall average log10(vRNA)
“Burden of illness” per subject (Chang, Guess, Heyse, 1994) v
n
ii
N
xv
1
c
n
ii
N
yc
1
19
Additional Notation for Two Other Methods (cont’d)
Note that
v
n
ii
vv
n
ii
n
xp
N
xvv
11 , where
v
vv N
np
c
n
ii
cc
n
ii
n
yp
N
ycv
11 , where
c
cc N
np
Let cv
cv
NNnn
p = overall HI V infection rate
20
Methods for Establishing POC (cont’d) Burden-of -I llness (BOI ) (Chang, Guess, Heyse, 1994): The diff erence in BOI per subject:
c
n
ii
v
n
ii
N
y
N
xT
cv
11
Let
),|(,|
0
0
HnnTVarHnnTET
Zcv
cvBOI
(see appendix f or details)
Reject null if ZZBOI
21
Methods for Establishing POC (cont’d)
Overall Wilcoxon Rank Sum Test (af ter plugging in log10(vRNA) = 0 f or all uninfected subjects):
I mplicitly assigns “best rank” (BR) to all uninfected subjects. From Chen, Gould, and Nessly (in press):
22*12 1212
13Z
pp
ppZ
pp
pZ cvBR
11
*1
1
cv
cv
nnpp
ppZ , 2Z is as before (viral load statistic)
*1Z is a score statistic f or comparing two independent
proportions. But it is invalid if the no. of inf ections is fixed in advance since the proportions are correlated!
When both cv pp and (and hence p) are small (e.g. < 5%),
the bulk of the weight in BRZ goes to *1Z .
22
Illustrative Example: Hypothetical Data Placebo Vaccine
Randomized 750 750 I nfected (HI V+) 16 14
297 2,964 52 55
3,281 3,617 275 533
6,098 6,612 556 719
22,641 25,070 1,881 2,181
39,535 49,351 7,187 11,271
72,194 132,388 15,263 39,273
218,419 239,210 53,179 58,534
Observed vRNA at 3 months post HI V+
diagnosis
256,844 266,901
vRNA G. Mean (c/ ml) 24,419 2,280 Median (c/ ml) 32,303 2,031 log10 vRNA Mean 4.3877 3.3580 Median 4.5092 3.3077
INFECTION endpoint: Z1=- 0.3688, p1=0.3561, w1=0.14, VEobs=13% VIRAL LOAD endpoint: Z2=- 2.5150, p2=0.0060, w2=0.86
23
Illustrative Example: Hypothetical Data (cont’d)
Method p-value* POC established? 2-part-model (Z) .0209 Yes
Weighted-Z .0055 Yes Simes .0119 Yes
Weighted-Simes .0069 Yes Fisher’s .0177 Yes
Weighted-Fisher’s .0062 Yes BOI .1463 No
Best-Rank .3453 No * f or the composite null hypotheis
24
Simulation Study
Assumed vRNA distributions f or inf ected subjects: Placebo: log10(vRNA) ~ Normal (log10(30,000), 0.75) Vaccine: Null: log10(vRNA) ~ Normal (log10(30,000), 0.75) Alt.: log10(vRNA) ~ Mixture of normals 20% Normal (log10(27,200) + b, 0.65) 24.3% Normal (log10(27,200) + b - 0.5 , 0.65) 55.7% Normal (log10(27,200) + b – 1.0, 0.65), f or diff erent choices of b; overall SD = 0.91
Overall average diff erence (placebo – vaccine) in log10(vRNA): δ = 1 - b
Pre-specified weights: w1 = 0.14 and w2 = 0.86, assuming VE=15% and δ = 1 log10 copies/ ml (i.e., b = 0)
25
Assumed Distributions for log10(viral laod)
Placebo
μ - δViral Load Set-Point (log10 copies/ml)
SD = 0.75
SD = 0.91
Vaccine
μ
Note: Assumed VL distribution for vaccine is asymmetric and more variable (mixture of vaccine “non-responders” and “responders”)
26
Simulation Study (cont’d)
Total enrollment N = Nv + Nc = 1500, r = Nc:Nv = 1:1 and the number of infections fixed at n = nv + nc
Let = (1 - VE)/ (1 + r - VE); r = 1 [VE = 0 iff = ½]
For a given number of infections n, we drew the number of vaccine inf ections nv f rom Binomial(n, ), and set the number of placebo infections to nc = n - nv
We drew nc and nv viral loads f rom the assumed placebo and vaccine viral load distributions, respectively
For each method, we flagged if the null was rejected
Repeated 5,000 times, calculated type-I error rate and power (under diff erent scenarios) at α = 0.05 (1-sided)
27
Simulation Results: Type-I Error Rate (=5%)T
yp
e-I
Err
or
Ra
te (
%)
012345678
10 20 30 40 50 60 70 80 90
Total Number of Infections
Two-Part Model (Z)
one-tailed 5% level
Typ
e-I
Err
or
Ra
te (
%)
012345678
10 20 30 40 50 60 70 80 90
Total Number of Infections
Weighted Two-Part Model (Z)
+2 S.E. (5,000 iterations)
Typ
e-I
Err
or
Ra
te (
%)
012345678
10 20 30 40 50 60 70 80 90
Total Number of Infections
Simes
Typ
e-I
Err
or
Ra
te (
%)
012345678
10 20 30 40 50 60 70 80 90
Total Number of Infections
Weighted-Simes
28
Simulation Results: Type-I Error (nominal =5%)
Typ
e-I
Err
or
Ra
te (
%)
012345678
10 20 30 40 50 60 70 80 90
Total Number of Infections
Fisher's
one-tailed 5% level
Typ
e-I
Err
or
Ra
te (
%)
012345678
10 20 30 40 50 60 70 80 90
Total Number of Infections
Weighted-Fisher's
+2 S.E. (5,000 iterations)
Typ
e-I
Err
or
Ra
te (
%)
012345678
10 20 30 40 50 60 70 80 90
Total Number of Infections
Wilcoxon (Best Rank)
Typ
e-I
Err
or
Ra
te (
%)
012345678
10 20 30 40 50 60 70 80 90
Total Number of Infections
BOI (log Scale)
29
Simulation Results: Power ( = 5%, 1-tailed)VE=0%, δ=0.5 VE=0%, δ=1.0
Number of Infections
Po
we
r (%
)
0
10
20
30
40
50
60
70
80
90
100
10 20 30 40 50 60 70 80 90 100
Weighted Fisher'sWeighted-ZWeighted-SimesFisher'sTwo-part (Z)SimesBOIWilcoxon (best rank)
Number of Infections
Po
we
r (%
)
0
10
20
30
40
50
60
70
80
90
100
10 20 30 40 50 60 70 80 90 100
Weighted Fisher'sWeighted-ZWeighted-SimesFisher'sTwo-part (Z)SimesBOIWilcoxon (best rank)
30
Simulation Results: Power ( = 5%, 1-tailed)VE=30%, δ=0.5 VE=30%, δ=1.0
Number of Infections
Po
we
r (%
)
0
10
20
30
40
50
60
70
80
90
100
10 20 30 40 50 60 70 80 90 100
Weighted Fisher'sWeighted-ZWeighted-SimesFisher'sTwo-part (Z)SimesBOIWilcoxon (best rank)
Number of Infections
Po
we
r (%
)
0
10
20
30
40
50
60
70
80
90
100
10 20 30 40 50 60 70 80 90 100
Weighted Fisher'sWeighted-ZWeighted-SimesFisher'sTwo-part (Z)SimesBOIWilcoxon (best rank)
31
Simulation Results: Power ( = 5%, 1-tailed)VE=60%, δ=0.5 VE=60%, δ=1.0
Number of Infections
Po
we
r (%
)
0
10
20
30
40
50
60
70
80
90
100
10 20 30 40 50 60 70 80 90 100
Weighted Fisher'sWeighted-ZWeighted-SimesFisher'sTwo-part (Z)SimesBOIWilcoxon (best rank)
Number of Infections
Po
we
r (%
)
0
10
20
30
40
50
60
70
80
90
100
10 20 30 40 50 60 70 80 90 100
Weighted Fisher'sWeighted-ZWeighted-SimesFisher'sTwo-part (Z)SimesBOIWilcoxon (best rank)
32
Number of Infections Required for Establishing POC*
Simes’, Fisher’s, Weighted-Fisher’s methods80% power, =5% (1-tailed)
True (log10 copies/ ml)
0.0 0.5 0.7 0.9 1.0 Vaccine Effi cacy
(%) S, F, WF S, F, WF S, F, WF S, F, WF S, F, WF
0% > 100 93, >100, 76 51, 56, 43 33, 36, 28 28, 31, 23
10% > 100 92, 93, 77 50, 52, 42 33, 36, 27 28, 30, 23
20% > 100 87, 81, 74 49, 48, 42 33, 33, 27 27, 28, 23
30% > 100 77, 67, 71 47, 43, 41 30, 30, 27 27, 26, 23
40% > 100 63, 56, 67 42, 37, 40 30, 27, 27 25, 23, 23
50% 77, 80, > 100 49, 41, 64 35, 30, 39 28, 23, 27 23, 21, 23
60% 47, 49, > 100 35, 30, 58 30, 26, 38 23, 21, 27 23, 18, 23
70% 28, 30, > 100 25, 21, 52 23, 19, 37 20, 16, 27 17, 16, 23
* POC is established if the composite null hypothesis is rejected
33
Challenge for the Merck Vaccine
• Pre-existing immunity to Adenovirus Type 5 may prevent or dampen the T cell response to the HIV proteins
• In the U.S., ~30-50% of people have neutralizing antibodies to Ad-5 virus
• In Southern Africa, ~75-95% of people neutralize Ad-5
• Summary of data from Phase I-II trials– Ad-5 Neut Titers < 18: ~80% vaccinees have a CD8+
ELISpot response– Ad-5 Neut Titers > 1000: ~40% have a response– In responders, geometric mean titer ~200 for
vaccinees with Ad-5 Neut Titers < 18; ~100 for vaccinees with Ad-5 Neut Titers > 1000
34
Concluding Remarks
• For a POC trial of a CMI-based HIV vaccine, Fisher’s (and Simes’) methods are good choices.
• If the composite null hypothesis is rejected at the 5% level, the p-values for the two endpoints can each be assessed separately at the 5% level.
• Challenges for the viral load analysis:- Initiation of antiretroviral therapy < 3
months after HIV+ diagnosis (“missing” vRNA data)
- Important to add “sensitivity analyses” to safeguard against potential selection
bias (e.g., Gilbert et al, 2003).- Estimating causal effect of vaccine on post-
infection viral load (ongoing research)
35
Appendix Conditional variance f or BOI statistic (Chang, Guess, Heyse, 1994):
Let cv
n
i
n
iii
nn
yxa
v c
1 1
1
)( 2
12
v
n
ii
x n
xxs
v
and 1
)( 2
12
c
n
ii
y n
yys
c
Then )//
)((),|(ˆ222
0cv
cyvx
cvcvcv NN
NsNs
NNa
nnHnnTV
36
References• Chang MN, Guess HA, Heyse JF (1994). Reduction in the burden of
illness: a new efficacy measure for prevention trials. Statistics in Medicine, 13, 1807-1814.
• Chen J, Gould AL, Nessly ML. Comparing two treatments by using a biomarker with assay limit. Statistics in Medicine, in press.
• Fisher RA (1932). Statistical methods for research workers. Oliver and Boyd, Edinburgh and London.
• Follman D (1995). Multivariate tests for multiple endpoints in clinical trials. Statistics in Medicine, 14, 1163-1175.
• Gilbert PB, Bosch RJ, Hudgens MG. Sensitivity analysis for the assessment of causal vaccine effects on viral load in HIv vaccine clinical trials. Biometrics, 59, 531-541.
• Good IJ (1955). On the weighted combination of significance tests. Biometrika, 264-265.
• Hochberg Y, Liberman U (1994). An extended Simes’ test. Statistics & Probability Letters, 21, 101-105.
• Lachenbruch PA (1976). Analysis of data with clumping at zero. Biometrische Zeitschrift, 18, 351-356.
• O’Brien PC (1984). Procedures for comparing samples with multiple endpoints. Biometrics, 40, 1079-1087.