1
Immunisation with a Partially Effective Vaccine
Niels G Becker National Centre for Epidemiology and Population Health Australian National University
Ideally vaccines immunise individuals completely. In practice some vaccinees are infected, making it very important to determine the efficacy of a vaccine as well as what happens when a partially effective vaccine is used in mass vaccination.
This lectures looks at this issue, in part to demonstrate that infectious disease models also play an important role in the analysis of infectious disease data.
2
BackgroundHow is the efficacy of a vaccine usually measured?nV vaccinated individuals are observed over interval [0,T].Of these CV become cases.
nU unvaccinated individuals are observed over interval [0,T].Of these CU become cases.
Classical measure of vaccine efficacy
UU
VV//1 VEnCnC
This sort of measure is sensible for assessing interventions in chronic diseases, but less so for infectious diseases. It does not account for transmission, and so can seriously mislead us.
3
To illustrate this weakness consider two types of vaccine effects on susceptibility (protection).
1. “Complete/none” vaccine effect The vaccine gives complete protection with probability
1f
and gives no protection with probability f.
2. “Partial/uniform” vaccine effect The vaccine gives the same partial protection to every
vaccinee. The probability of a contact close enough to infect an unvaccinated individual has only probability a, where 0 < a < 1, of infecting a vaccinated person.
These two vaccine effects were first considered by Smith, Rodrigues and Fine (1984). Int. J. Epidemiol. 13, 87-96.
4
Example 1(a) Consider a community of homogeneous individuals who mix uniformly.For a vaccine with a “complete/none” effect it can be shown that estimates 1f,which is a sensible estimate of vaccine efficacy for this vaccine effect.
(b) Consider now a vaccine with a “partial/uniform” effect. When the infection intensity acting on individuals is t for unvaccinated individuals and a.t for a vaccinated ones, a sensible measure of vaccine efficacy is 1 a. Then VE estimates estimates
UU
VV//1 VEnCnC
)exp(1)exp(11
T
TaVE
5
0 5 100.0
0.1
0.2
0.3
0.4
0.5
Lambda_T
V EVE as a function of duration of exposure (a=0.5)
)exp(1)exp(11
T
TaVE
The graph of is
Therefore the value of VE depends on• the size of the ‘epidemic’ over the observation period• the duration of the observation period• the vaccination coverage
6
Example 2 Suppose (a) the study consists of households of size two, with
1 vaccinated and 1 unvaccinated in each household;
(b) the vaccine reduces susceptibility a little, but reduces the infectivity of the vaccinee a lot.
Then, for a large studyUU
VV//1 VEnCnC
gives a value that depends on • the magnitude of the reduction in infectivity (1-b),
and• the secondary attack rate (SAR).
7
8
Heuristic explanation
vaccinated
unvaccinated
Can we formulate a measure of vaccine efficacy that accounts for transmission?
9
The trouble with
is that it estimates a different quantity in different settings.What we need is a measure (a parameter) with the same interpretation in every setting. The cost is that we then need to find an appropriate estimate of this parameter for different vaccine trials and different observational settings. In fact we need more than one measure, because we are interested in the reduction in susceptibility, the reduction in infectivity and the reduction in community transmission.
UU
VV//1 VEnCnC
10
Let us begin by proposing suitable measures for the vaccine effects.
Reduction in susceptibilitySuppose that vaccination reduces an individual's probability of disease transmission, per close contact, by a random factor A. That is, a force of infection (t) acting on an unvaccinated individual at time t reduces to A.(t) for a vaccinated individual.
A=0 implies complete protectionA=1 implies no vaccine effectPr(A=1) = f, Pr(A=0) = 1- f implies complete protection apart from a fraction f of failures.
11
02468
0 5 10 15 20
Reduction in infectivityx = infectiousness function indicates how infectious an unvaccinated individual is x time units after being infected.
X
days
12
The effect of the vaccine on infectiousness, in the event that a vaccinee is infected, might be a shorter duration of illness, shorter infectious period, a lower rate of shedding pathogen, etc. than they would have if not vaccinated.From public health point of view there is interest in the reduction in transmission. The potential for an infective to infect others is the area under x • BU when infective unvaccinated• BV when the infective is vaccinated. Relative infection potential B =BV/BU is random
13
We call (A,B) the vaccine response (or vaccine effect)We allow A and B have any probability distribution(Expect them to lie between 0 and 1, and be correlated)
Summary measures of vaccine effects
1. Define VES = 1 E(A) (protective vaccine efficacy)
• For the “complete/none” response Pr(A = 1) = f , Pr(A = 0) = 1 f giving VES = 1 - f, which is sensible.
• For the “partial/uniform” response Pr(A = a) = 1 giving VES = 1 - a, which is also sensible.
14
2. Define VEI = 1 E(AB) / E(A)
• When Pr(A=a)=1, then VEI =1-E(B)• E(A) = 0 when Pr(A=0), but this is not of concern as no vaccinee becomes infected• VEI is more difficult to estimate than VES
3. Define VEIS = 1 E(AB)
To demonstrate that this is an important parameter, consider a community of homogeneous individuals who mix uniformly. Vaccinate a fraction v of them. Then
RV = [1 v + v.E(AB)] R0
15
Consider first VES= 1 - E(A)A SINGLE OUTBREAKSuppose that every individual has the same exposure to infectious individuals. Then we can show
1 - cV/cU VES 1- log(1cV) / log(1cU) irrespective of the distribution of A
cU = proportion of unvaccinated participants who become casescV = proportion of vaccinated participants who become cases
The bounds are parameters, written in a way that makes estimates obvious.[ B & Utev (2002) Biometrical Journal, 44, 29-42.]
16
((i) LH bound is attained for ‘complete/no protection’ response.
(ii) RH bound is attained for ‘partial and uniform protection’
response.(iii) The above two responses are extremes in the sense
that, with a common mean E(A) , the CN response has the
most variation for A and the PU response has the least
variation.(iv) Estimate cU by CU/nU and cV by CV/nV.(v) We have standard errors.(vi) VES is not identifiable when only CU and CV are
observed.(vii) These bounds are useful when cU and cV are small ?
1 - cV/cU VES 1 - log(1cV) / log(1cU)
17
Unvaccinated Vaccinated
Cases 96 8 Not infected 174 57 270 65
ApplicationOutbreak of mumps in school of Ashtabula County, Ontario
The estimatesCU/nU = 96/270 = .356, CV/nV = 8/65 = .123give 0.654 VES 0.702
(.118) (.111)These bounds are of practical value!
1cV/cU VES = 1 E(A) 1log(1cV)/log(1cU)
18
Can we estimate VEI and VEIS?OUTBREAKS in HOUSEHOLD PAIRSP = probability of an unvaccinated susceptible individual being infected from outside the pair = 1 - Qp = probability of a susceptible individual being infected by an infected partner = 1 - qTo illustrate, assume uniform (a,b) response and hshlds size 2, with 1 vaccinated and 1 unvaccinated member.
0 cases U case V case 2 cases Q1+a Q(1-Qa)qb PQaqa balance n00 n10 n01 n11
Likelihood inference for a and b is straightforward.
19
Illustration with reference to smallpox dataEpidemic of variola minor in Braganca Paulista County (Brazil), 1956A total of 338 householdsHousehold sizes from 1 to 12 (mean 4.6)809 vaccinated and 733 unvaccinated 85 and 425 were infectedVaccine response modelThree vaccine responses
c1 – f – c fProbability
(0,•)(a,b)(1,1)Response (A,B)
B, O’Neill & Britton (2003),
Biometrics.
20
Transmission modelUnvaccinated are homogeneous.3 types of vaccine responses, BUT immunes can be ignored; and vaccine failures are like the unvaccinated, so we need to deal with 2 types only.Numerical computation of u,v(i,j) = Pr [ (i,j) out of (u,v) are infected ]is manageable. Numerical computation of the likelihood function is manageable.Inferences: Bayesian inferences via MCMC methods
Uniform (0,1) priors for a, b, c, f, q, Q.
21
22
a = partial reduction in susceptibility c = Pr(complete protection) = Pr(A=0)
23
Posterior distribution for c
24
Posterior distribution for a
25
The data are not compatible with small c and large
a
No points here
26
If c is large, then the data are compatible with most values
of a
27
If a is small, then the data are
compatible with most values of c
28
It is difficult to distinguish between (a 0 c 0) and (a
1 c .8) (both indicate low
susceptibility)
29
Conclusions
• Precise estimation of VES = 1 E(A) is possible.
• Estimation of VEIS = 1 E(AB) is surprisingly good.
• Estimation of the reduction in infectivity, per se, was not particularly effective. It may be more effective from data on smaller households.
Further work
30
Collaborators1. Becker NG, Starczak DN (1998). The effect of random vaccine
response on the vaccination coverage required to prevent epidemics. Mathematical Biosciences 154, 117-135.
2. Becker NG, Utev S (2002). Protective vaccine efficacy when vaccine response is random. Biometrical Journal 44, 29-42.
3. Becker NG, Britton T, O’Neill PD (2003). Estimating vaccine effects on transmission of infection from household data. Biometrics 59, 467-475.
4. Becker NG, Britton T (2004). Estimating vaccine efficacy from small outbreaks. Biometrika 91, 363-382.
5. Becker NG, Lefevre C, Utev S (2005). Estimating protective vaccine efficacy from large trials with recruitment. Journal of Statistical Planning and Inference (in press).
6. Becker NG, Britton T, O’Neill PD (2005). Estimating vaccine effects from studies of outbreaks in household pairs. Statistics in Medicine (in press). The EndThe End