HAL Id: hal-02807182
https://hal.inrae.fr/hal-02807182
Submitted on 6 Jun 2020
HAL is a multi-disciplinary open access
archive for the deposit and dissemination of
sci-entific research documents, whether they are
pub-lished or not. The documents may come from
teaching and research institutions in France or
abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est
destinée au dépôt et à la diffusion de documents
scientifiques de niveau recherche, publiés ou non,
émanant des établissements d’enseignement et de
recherche français ou étrangers, des laboratoires
publics ou privés.
Statistical inference for epidemic models approximated
by diffusion processes
Romain Guy, Catherine Laredo, Elisabeta Vergu
To cite this version:
Romain Guy, Catherine Laredo, Elisabeta Vergu. Statistical inference for epidemic models
approxi-mated by diffusion processes. Journées de l’ANR MANEGE, Jan 2013, Université Paris 13
Villeta-neuse, France. 33 diapos. �hal-02807182�
Inference for discrete observations of diusion or Gaussian processes with small diusion coecient Epidemics incompletely observed: partially and integrated diusion processes (Work in progress)
Statistical Inference for epidemic models approximated by
diusion processes
Romain GUY
1,2Joint work with C. Larédo
1,2and E. Vergu
11UR 341, MIA, INRA, Jouy-en-Josas 2 UMR 7599, LPMA, Université Paris Diderot
ANR MANEGE, Université Paris 13
30 janvier 2013
Inference for discrete observations of diusion or Gaussian processes with small diusion coecient Epidemics incompletely observed: partially and integrated diusion processes (Work in progress)
Outline
Provide a framework for estimating key parameters of epidemics
1
Characteristics of the epidemic process
Constraints imposed by the observation of the epidemic process
Simple mechanistic models
2
Various mathematical approaches for epidemic spread
Natural approach: Markov jump process
First approximation by ODEs
Gaussian approximation of the Markov jump process
Diusion approximation of the Markov jump process
3
Inference for discrete observations of diusion or Gaussian processes with small
diusion coecient
Contrast processes for xed or large number of observations
Correction of a non asymptotic bias
Comparison of estimators on simulated epidemics
4
Epidemics incompletely observed: partially and integrated diusion processes (Work
in progress)
Back to epidemic data
Inference approach: Work in progress
Inference for discrete observations of diusion or Gaussian processes with small diusion coecient Epidemics incompletely observed: partially and integrated diusion processes (Work in progress)Simple mechanistic models
Outline
1
Characteristics of the epidemic process
Constraints imposed by the observation of the epidemic process
Simple mechanistic models
2
Various mathematical approaches for epidemic spread
Natural approach: Markov jump process
First approximation by ODEs
Gaussian approximation of the Markov jump process
Diusion approximation of the Markov jump process
3
Inference for discrete observations of diusion or Gaussian processes with small
diusion coecient
Contrast processes for xed or large number of observations
Correction of a non asymptotic bias
Comparison of estimators on simulated epidemics
4
Epidemics incompletely observed: partially and integrated diusion processes (Work
in progress)
Back to epidemic data
Inference approach: Work in progress
Inference for discrete observations of diusion or Gaussian processes with small diusion coecient Epidemics incompletely observed: partially and integrated diusion processes (Work in progress)Simple mechanistic models
Incomplete Data
Imperfect data
Incomplete observations
Temporally aggregated
Sampling & reporting error
Unobserved cases
Dierent dynamics framework
Ex.: Inuenza like illness cases (Sentinelles
surveillance network)
One outbreak study
Recurrent oubreaks study
Main goal: key parameter estimation
Basic reproduction number, R0
(nb. of
secondary cases generated by one primary
case in an entirely susceptible population)
Average infectious time period (d)
Inference for discrete observations of diusion or Gaussian processes with small diusion coecient Epidemics incompletely observed: partially and integrated diusion processes (Work in progress)Simple mechanistic models
Compartmental representation of the population dynamics
Dene: Nb. of health states, possible transitions and associated rates Notations N :
population size, λ : transmission rate, γ : recovery rate
S, I , R : numbers of susceptible, infected, removed individuals
One of the simplest model: SIR
Closed population ⇒ N=S + I + R
Well-mixing population
⇒ (
S, I )
λSI /N→
(
S − 1, I + 1)
Convenient to study one epidemics
Key parameters: R0
=
λγ, d =
γ1Summary: coecients α
L(
S, I ) → (S − 1, I + 1) = (S, I ) + (−1, 1) at rate α(−1,1)
(
S, I ) = λS
NIand
(
S, I ) → (S, I − 1) = (S, I ) + (0, −1) at rate α(
0,−1)(
S, I ) = γI
Inference for discrete observations of diusion or Gaussian processes with small diusion coecient Epidemics incompletely observed: partially and integrated diusion processes (Work in progress)Simple mechanistic models
Natural extensions of the SIR model
Increase the number of health states : e.g. Exposed Class ⇒ SEIR model
Additionnal transitions: e.g. (S, I ) → (S − 1, I ) (vaccination)
Temporal dependence: SIRS with seasonality in transmission and demography
δ
: waning immunity rate (years)
µ
: demographic renewal rate (decades)
λ(t) = λ0
(1 + λ1sin(2π
tTper
))
λ
1=
0 ⇒ oscillations vanishes
Suited to study recurrent epidemics
Key parameters: R
0Moy=
γ+µλ0, d =
1γSummary:
α
(−1,1)(
t
,
S, I ) = λ(
t
)
S
NI,
α
(1,0)(
S, I ) = Nµ + δ(N − S − I ),
α
(−1,0)(
S, I ) = µS
α
(0,−1)(
S, I ) = (µ + γ)I
Inference for discrete observations of diusion or Gaussian processes with small diusion coecient
Epidemics incompletely observed: partially and integrated diusion processes (Work in progress)Gaussian approximation of the Markov jump processDiusion approximation of the Markov jump process
Outline
1
Characteristics of the epidemic process
Constraints imposed by the observation of the epidemic process
Simple mechanistic models
2
Various mathematical approaches for epidemic spread
Natural approach: Markov jump process
First approximation by ODEs
Gaussian approximation of the Markov jump process
Diusion approximation of the Markov jump process
3
Inference for discrete observations of diusion or Gaussian processes with small
diusion coecient
Contrast processes for xed or large number of observations
Correction of a non asymptotic bias
Comparison of estimators on simulated epidemics
4
Epidemics incompletely observed: partially and integrated diusion processes (Work
in progress)
Back to epidemic data
Inference approach: Work in progress
Inference for discrete observations of diusion or Gaussian processes with small diusion coecient
Epidemics incompletely observed: partially and integrated diusion processes (Work in progress)Gaussian approximation of the Markov jump processDiusion approximation of the Markov jump process
Markov jump process
α(−
1,1)(
S, I ) = λS
NI, α(
0,−1)(
S, I ) = γI
Notations:
E = {0, .., N}
d∀
L ∈ E
−= {−
N, .., N}
d, we dene α
L
(·) :
E → [0, +∞[
We dene (Z
t)
the Markov jump process on E with Q-matrix: q
X ,Y= α
Y −X(
X )
Assume α(X ) =
X
L∈E−
α
L(
X ) < +∞ ⇒ Sojourn time Exp(α(X ))
Easily simulated (using Gillespie algorithm)
3 realizations for N = 10000, λ = 0.5, γ = 1/3, (S0
,
I0
) = (
9990, 10)
Inference for discrete observations of diusion or Gaussian processes with small diusion coecient
Epidemics incompletely observed: partially and integrated diusion processes (Work in progress)Gaussian approximation of the Markov jump processDiusion approximation of the Markov jump process
Interest of the deterministic approach
λ(
t) = λ0
(
1 + λ1sin(2πt/Tper
)
Some trajectories of the SIRS Markov proc. :
N = 10
5,
R
0
=
1.5, d = 3,
δTper1=
2,
µTper1=
50
Drawbacks of the Markov jump approach
N = 10
7:more than 10
5events in one
week (MLE: observation of all the jumps
required)
Extinction probability non negligible
SIRS ODE solution
ds dtdi= µ(
1 − s) + δ(1 − s − i) − λ(t)si
dt= λ(
t)si − (µ + γ)i
λ(
t) = λ0
(
1 + λ1sin(2πt/T )
(
s(0), i(0)) =
ZN0ODE trajectory
Link between the two approaches
As N → +∞ we have
ZtN N→∞
−→
x(t), where
x(t) is the deterministic solution of the ODE:
dx(t)
dt
=
b(x(t))
Function b is explicit
Inference for discrete observations of diusion or Gaussian processes with small diusion coecient
Epidemics incompletely observed: partially and integrated diusion processes (Work in progress)Gaussian approximation of the Markov jump processDiusion approximation of the Markov jump process
Beyond deterministic limit : Gaussian process
Additionnal assumption: smooth version of αL, βL
We have α
L:
E → (0, +∞) transition rate : X
αL(X)→
X + l
Assume β
L: [
0, 1]
d→ [
0, +∞[ well dene and regular :
∀
x ∈ [0, 1]
d,
1 Nα
L(b
Nxc) −→
N→∞β
L(
x)
SIR: α
(−1,1)(
S, I ) = λS
NI⇒ β
(−1,1)(
x) = λx
1x
2, α
(0,1)(
S, I ) = γI ⇒ β
(0,−1)(
x) = γx
2Denition of function b (
dx(t) dt=
b(x(t)))
b(x) =
X
L∈E−Lβ
L(
x)
, SIR: b((λ, γ), (s, i)) =
−
1
1
λ
si +
0
1
γ
i =
−λ
si
λ
si + γi
ODE approximation : no longer dependance w.r.t. N
Asymptotic expansion w.r.t. N : Gaussian process
√
N
ZtN−
x(t)
−→
N→∞
g(t) where g(t) centered Gaussian process:
dg(t) =
∂b∂x
(
x(t))g(t)dt + σ(x(t))dB
t, where σ
tσ(
x) =
Σ(
x) =
X
L∈E−
L
tLβ
L(
x)
SIR: Σ((λ, γ), (s, i)) = λsi
−
1
1
−
1
1 + γi
0
1
0
1 =
λ
−λ
si
si
λ
−λ
si + γi
si
Cholesky algorithm ⇒ σ(x) =
√
λ
si
0
−
√
λ
si
√
γ
i
Inference for discrete observations of diusion or Gaussian processes with small diusion coecient
Epidemics incompletely observed: partially and integrated diusion processes (Work in progress)Gaussian approximation of the Markov jump processDiusion approximation of the Markov jump process
Innitesimal generator approach: diusion approximation
Renormalized version of the Markov jump process (˜Zt
)
:
( ˜
Z
t)
Markov jump process on E with transition rates q
X ,Y=
Nβ
Y −X(
X )
¯
Z
t=
Zt˜Nnormalized process
Asymptotic development of the innitesimal generator of Z
t(Ethier & Kurtz (86))
Generator of ˜
Z
t: Af (x) =
X
L∈E−NLβ
L(
x) (f (x + L) − f (x))
⇒
Generator of ¯
Z
t: ¯
Af (x) = b(x). 5 f (x) +
N1 dX
i ,j=1Σ
i ,j(
x)
∂
2f
∂
x
i∂
x
j(
x) + O(
1
N
2)
Dropping negligible terms leads to the generator of a diusion process:
dX
t=
b(X
t)
dt +
√1 Nσ(
X
t)
dB
t, where b(x) =
X
l ∈E−Lβ
L(
x), Σ(x) =
X
L∈E−L
tLβ
L(
x)
Temporal dependence βL
(
t
,
x): generator approach no longer available
Decomposition of the diusion process using Gaussian process ( =
√1N
)
Taylor's stochastic formula (Wentzell-Freidlin(79), Azencott (82))
Let dX
t=
b(X
t)
dt + σ(X
t)
dB
t,
X
0=
x
0Then, under regularity assumptions, X
t=
x(t) + g(t) + O
P(
2)
Inference for discrete observations of diusion or Gaussian processes with small diusion coecient
Epidemics incompletely observed: partially and integrated diusion processes (Work in progress)Gaussian approximation of the Markov jump processDiusion approximation of the Markov jump process
Links between approximations
Zt
Markov jump process on E with transition qX ,Y
= α
Y −X(
X )
¯
Zt
Markov jump process on E/N with transition qx,y
=
Nβb
Nxc−bNyc(
x)
¯
Zt
∼
x(t) +
√1N
g(t)
with x(·) the ODE solution and g a Gaussian process
Xt
: dXt
=
b(x(t))dt +
√1N
σ(
x(t))dBt
diusion with small diusion coecient
and
P{ sup
0≤t≤Tk ¯
Z
t−
X
tk >
C
Tlog(N) N} −→
N→∞0
X
t=
x(t) +
√1 Ng(t) + O
P(
N1)
→
Diusion approximation
→
Expansion in N of the process
→
Taylor's stochastic expansion
Important : All mathematical representations completely dened by (αL
)
Inference for discrete observations of diusion or Gaussian processes with small diusion coecient
Epidemics incompletely observed: partially and integrated diusion processes (Work in progress)
Correction of a non asymptotic bias
Comparison of estimators on simulated epidemics
Outline
1
Characteristics of the epidemic process
Constraints imposed by the observation of the epidemic process
Simple mechanistic models
2
Various mathematical approaches for epidemic spread
Natural approach: Markov jump process
First approximation by ODEs
Gaussian approximation of the Markov jump process
Diusion approximation of the Markov jump process
3
Inference for discrete observations of diusion or Gaussian processes with small
diusion coecient
Contrast processes for xed or large number of observations
Correction of a non asymptotic bias
Comparison of estimators on simulated epidemics
4
Epidemics incompletely observed: partially and integrated diusion processes (Work
in progress)
Back to epidemic data
Inference approach: Work in progress
Inference for discrete observations of diusion or Gaussian processes with small diusion coecient
Epidemics incompletely observed: partially and integrated diusion processes (Work in progress)
Correction of a non asymptotic bias
Comparison of estimators on simulated epidemics
Classical Estimators
Maximum likelihood estimation for the SIR Markov Jump process (Andersson &
Britton (00))
Observation of all jumps
Analytic expression of the estimators for SIR model:
ˆ
λ =
N
Z
TS(0)−S(T ) 0S(t)I (t)dt
, ˆ
γ =
S(0)+I(0)−S(T )−I(T )Z
T 0I (t)dt
√
N(ˆθ
MLE− θ
0) −→
N→∞N
0, I
−1 b(θ
0)
, where
I
−1 b((λ
0, γ
0)) =
λ20 s(0)−s(T)0
0
γ20 s(0)+i (0)−s(T )−i (T )
Maximum likelihood estimation for homoscedastic observations of the ODE
n observations at n discrete times tk
of xθ(
tk
) + ξ
k, with ξk
∼ N (
0, CN
(θ
0)
Id
)
MLE=LSE
√
n
ˆ
θ
LSE− θ
0−→
n→∞
N (
0, I
N(θ
0))
Inference for discrete observations of diusion or Gaussian processes with small diusion coecient
Epidemics incompletely observed: partially and integrated diusion processes (Work in progress)
Correction of a non asymptotic bias
Comparison of estimators on simulated epidemics
Specicities of the statistical framework
Model:
We dene X
t: dX
t=
b(θ
1,
X
t)
dt + σ(θ
2,
X
t)
dB
t,
X
0=
x
0∈ R
d⇒
separation of the parameters θ
1, θ
2required
(not estimated at the same rate)
Continuous observation of the diusion on [0, T ] (Kutoyants (80))
MLE :
−1θ
1MLE− θ
01→ N
0, I
b(θ
01, θ
20)
−1Existing discrete observation results : estimation of θ
2at rate
√
n
Observations:
We observe X
tkfor t
k=
k∆, k ∈ {0, .., n}, t
k∈ [
0, T ] (n∆ = T ), T is xed
Two dierent asymptotics: → 0 & n (∆) is xed // → 0 & n → ∞ (∆ → 0)
Notation: η = (θ
1, θ
2)
SIR: =
√1N
, θ
1= θ
2= θ = (λ, γ)
, and I
b(θ
01, θ
02)
equals the Markov jump process Fisher
Information matrix
Inference for discrete observations of diusion or Gaussian processes with small diusion coecient
Epidemics incompletely observed: partially and integrated diusion processes (Work in progress)
Correction of a non asymptotic bias
Comparison of estimators on simulated epidemics
Main idea: study of g
η(
t) (Multidimensionnal generalization of Genon-Catalot(90))
Gaussian process: Yt
=
xθ
1(
t) + g
η(t), n obs. at regular time intervals tk
=
k∆, for
k = 1, .., n.
Denition : Resolvent matrix of the linearized ODE system Φθ
1Let Φθ
1be the invertible matrix solution of
dΦθ1
dt
(
t, t0
) =
∂∂bx(
xθ
1(
t))Φθ
1(
t, t0
),
with Φθ
1(
t0
,
t0
) =
Id
.
Important property of g
ηgη(
tk
) = Φθ
1(
tk
,
tk−1
)
gη(
tk−1
) +
√
∆
Z
kη(
Z
kη)
k∈{1,..,n}independent Gaussian vectors with covriance matrix S
kηS
η k=
∆1Z
tktk−1
Φ
θ1(
t
k,
s)Σ(θ
2,
x
θ1(
s))
tΦ
θ1(
t
k,
s)ds
Function of the observations
Let y ∈ C [0, T ], R
dNk
(θ
1,
y) = y(tk
) −
xθ
1(
tk
) − Φθ
1(
tk
,
tk−1
)
y(tk−1
) −
xθ
1(
tk−1
) (=
√
∆
Z
kη)
Inference for discrete observations of diusion or Gaussian processes with small diusion coecient
Epidemics incompletely observed: partially and integrated diusion processes (Work in progress)
Correction of a non asymptotic bias
Comparison of estimators on simulated epidemics
Back to the diusion process: dX
t=
b(θ
1,
X
t)
dt + σ(θ
2,
X
t)
dB
t(
Z
kη)
Gaussian familly ⇒ Likelihood tractable:
−
L∆,(η) =
2 nX
k=1log[det(S
η k)] +
∆
1
nX
k=1 tNk
(θ
1,
Y )(S
kη)
−1Nk
(θ
1,
Y )
ˆ
θ
2has good properties as → 0, only if ∆ → 0 (n → +∞)
1. n xed, → 0: General case (low frequency contrast with θ
2unknown)
¯
U(θ
1) =
∆1 nX
k=1
t
Nk
(θ
1,
X )Nk
(θ
1,
X ) ⇒ Associated MCE ¯θ
1,=
argmin
θ1∈Θ¯
U(θ
1)
2. n xed, → 0 : Case θ
2=
f (θ
1)
(low frequency contrast with information on θ
2)
˜
U
(θ1) =
∆1 nX
k=1 tNk
(θ
1,
X )(˜S
kθ1,f (θ1))
−1Nk
(θ
1,
X ) ⇒ ˜θ1,
=
argmin
θ1∈Θ˜
U
(θ1)
3. n → ∞, → 0 (high frequency contrast)
ˇ
U
∆,(θ
1, θ
2) =
2 nX
k=1log[det(Σ(θ
2,
X
tk−1))] +
1
∆
nX
k=1 tN
k(θ
1,
X )Σ
−1(θ
2,
X
tk−1)
N
k(θ
1,
X )
⇒ ˇ
θ
1,,∆, ˇ
θ
2,,∆=
argmin
η∈Θ¯
U,∆(η)
Inference for discrete observations of diusion or Gaussian processes with small diusion coecient
Epidemics incompletely observed: partially and integrated diusion processes (Work in progress)
Correction of a non asymptotic bias
Comparison of estimators on simulated epidemics
What kind of distance is minimized: comparison with Least squares
Nk
(θ
1,
y) = y(tk
) −
xθ
1(
tk
) − Φθ
1(
tk
,
tk−1
)
y(tk−1
) −
xθ
1(
tk−1
)
N = 1000, R0
=
1.5, d = 3 days, 1 obs/day, T = 50 days
Figure:
Diusion (blue), x
θ1(
t) (green)
Inference for discrete observations of diusion or Gaussian processes with small diusion coecient
Epidemics incompletely observed: partially and integrated diusion processes (Work in progress)
Correction of a non asymptotic bias
Comparison of estimators on simulated epidemics
What kind of distance is minimized: comparison with Least squares
Zoom between t = 5 and t = 6
g
η(tk
) = Φθ
1(
tk
,
tk−1
)
g
η(tk−1
) +
√
∆
Z
kηInference for discrete observations of diusion or Gaussian processes with small diusion coecient
Epidemics incompletely observed: partially and integrated diusion processes (Work in progress)
Correction of a non asymptotic bias
Comparison of estimators on simulated epidemics
What kind of distance is minimized: comparison with Least squares
Zoom between t = 5 and t = 6
g
η(tk
) = Φθ
1(
tk
,
tk−1
)
g
η(tk−1
) +
√
∆
Z
kηInference for discrete observations of diusion or Gaussian processes with small diusion coecient
Epidemics incompletely observed: partially and integrated diusion processes (Work in progress)
Correction of a non asymptotic bias
Comparison of estimators on simulated epidemics
What kind of distance is minimized: comparison with Least square
Figure:
Distance to the deterministic model
Figure:
Comparison: N
k(
X , θ
1)
(blue) and X
tk−
x
θ1(
t
k)
(green)
Inference for discrete observations of diusion or Gaussian processes with small diusion coecient
Epidemics incompletely observed: partially and integrated diusion processes (Work in progress)
Correction of a non asymptotic bias
Comparison of estimators on simulated epidemics
Results about dX
t=
b(θ
1,
X
t)
dt + σ(θ
2,
X
t)
dB
tUnder classical regularity assumptions on b and σ, we proved
n xed, → 0, low frequency contrast
Identiability assumption: θ1
6= θ
01
⇒ {∃
k, 1 ≤ k ≤ n, xθ
1(
tk
) 6=
xθ
01(
tk
)}.
1. General case (no information on θ
2)
−1θ
¯
1− θ
01−→
→0
N (
0, J
−1 ∆(θ
01, θ
02))
2. case θ
2=
f (θ
1)
(with information on θ
2)
−1˜
θ
1− θ
01−→
→0
N (
0, I
−1 ∆(θ
01, θ
02))
3. n → ∞ → 0: high frequency contrast
−1( ˇ
θ
1,∆− θ
01)
√
n( ˇ
θ
2,∆− θ
02)
−→
n→∞,→0N
0,
I
−1 b(θ
01, θ
02)
0
0
I
−1 σ(θ
10, θ
20)
Remarks
Epidemics: =
√1N
, θ = θ
1= θ
2, then for contrast 3: Ibis the same as for the
Markov jump process (all jumps observed)
J∆
is not optimal, but I∆
is, in the sense that I∆(θ
1, θ
2) −→
∆→0
Ib
(θ
1, θ
2)
Inference for discrete observations of diusion or Gaussian processes with small diusion coecient
Epidemics incompletely observed: partially and integrated diusion processes (Work in progress)
Correction of a non asymptotic bias
Comparison of estimators on simulated epidemics
Results on SIR for N = 10000, empirical mean estimators on 1000 runs and 95% theoretical CI
(R
0=
1.2, d = 3)
Figure:
0: MLE, 1: ¯
θ
1low frequency MCE (general case), 2: ˜
θ
1low frequency MCE
(θ
1= θ
2), 3: ˇ
θ
1,∆high frequency MCE
About unpresented results
Good results even for N = 100 (our methods seem more robust than MLE)
Similar performance (w.r.t. MLE) on more sophisticated models
Inference for discrete observations of diusion or Gaussian processes with small diusion coecient
Epidemics incompletely observed: partially and integrated diusion processes (Work in progress)
Correction of a non asymptotic bias
Comparison of estimators on simulated epidemics
About ˜
θ
1(Low frequency with information on θ
2)
θ
2unknown
1. ¯
U
(θ
1) =
∆1 nX
k=1 tN
k(θ
1,
X )N
k(θ
1,
X )
Figure:
Comparison between Data and ODE
(x
θ1˜(
t))
Not good t of the data
With information on θ
22. ˜
U
(θ
1) =
∆1 nX
k=1 tN
k(θ
1,
X )
(˜
S
kθ1)
−1N
k(θ
1,
X )
Figure:
Evolution of det
Σ
−1(θ
01,
x
θ01
(
t))
Too much weight on the boundaries
Inference for discrete observations of diusion or Gaussian processes with small diusion coecient
Epidemics incompletely observed: partially and integrated diusion processes (Work in progress)
Correction of a non asymptotic bias
Comparison of estimators on simulated epidemics
About ˜
θ
1(Low frequency with information on θ
2)
Comparing to high frequency contrast
3. ˇ
U
∆,(θ
1, θ
2))
=
2 nX
k=1log [det (Σ
k)]
+
∆1 nX
k=1 tN
k(θ
1)Σ
−k1N
k(θ
1)
where
Σ
k= Σ(θ
2,
X
tk−1)
Figure:
Previous results
Corrected contrast with information on θ
22
0. ˜
U
cor(α)
=
2 nX
k=1log
hdet(˜S
θ1 k)
i
+
∆1 nX
k=1 tN
k(θ
1)
˜
S
θ1 k −1N
k(θ
1)
Figure:
Corrected results
Inference for discrete observations of diusion or Gaussian processes with small diusion coecient
Epidemics incompletely observed: partially and integrated diusion processes (Work in progress)
Correction of a non asymptotic bias
Comparison of estimators on simulated epidemics
Dierent shapes of the condence ellipsoids
SIR Model condence ellipsoids for the corrected contrast
N = 1000, (R0
,d) = {(1.5, 3), (1.5, 7), (5, 3), (5, 7)}
Number of observations:
n = 10 (blue)
,
n = 1obs/day ≈40 d.(green)
, n = 2000
(black),
MLE CR (red) (Theoretical limit)
Inference for discrete observations of diusion or Gaussian processes with small diusion coecient
Epidemics incompletely observed: partially and integrated diusion processes (Work in progress)
Correction of a non asymptotic bias
Comparison of estimators on simulated epidemics
Bias of MLE (N = 400; R
0=
1.5; d = {3, 7}; (s
0,
i
0) = (
0.99, 0.01))
Too much variability
Figure:
Trajectories for d=3
Need of an empiric threshold: 5% of infected
Estimation results (d = 3)
Other values of d were investigated
Other thresholds: time of extinction, number max of
infected
Results (d = 7, time of extinction)
Zoom on the red trajectory (see Figure)
(
R0
,
d)MLE
= (
0.8749, 2.9945)
(
R0
,
d)LSE
= (
0.94, 2.5794)
(
R0
,
d)cont
= (
1.01, 3.0973)
Inference for discrete observations of diusion or Gaussian processes with small diusion coecient
Epidemics incompletely observed: partially and integrated diusion processes (Work in progress)
Correction of a non asymptotic bias
Comparison of estimators on simulated epidemics
Temporal dependence (SIRS) : λ
1dicult to estimate
SIRS with constant immigration in I class
λ(
t) = λ0
(
1 + λ1sin(2πt/T ))
Values
R0
=
1.5; d = 3d;
δT1per
=
2y,
λ
1= {
0.05, 0.15}
Fixed: Tper
=
365, µ = 1/50Tper, ζ =
10N,
N = 10
7, 1 obs/day(week) for 20 years
Term for λ1
in Ib
(θ
0)
very small ⇒
N > 10
5for satisfactory CI
λ
1bifurcation parameter for the ODE
λ
1=
0.05 (weak seasonality)
λ
1=
0.15 (stronger seasonality)
Detailed results not shown
Main idea: R0
,d, δ well estimated
λ
1: biased (often estimated to 0)Inference for discrete observations of diusion or Gaussian processes with small diusion coecient
Epidemics incompletely observed: partially and integrated diusion processes (Work in progress)Inference approach: Work in progress
Outline
1
Characteristics of the epidemic process
Constraints imposed by the observation of the epidemic process
Simple mechanistic models
2
Various mathematical approaches for epidemic spread
Natural approach: Markov jump process
First approximation by ODEs
Gaussian approximation of the Markov jump process
Diusion approximation of the Markov jump process
3
Inference for discrete observations of diusion or Gaussian processes with small
diusion coecient
Contrast processes for xed or large number of observations
Correction of a non asymptotic bias
Comparison of estimators on simulated epidemics
4
Epidemics incompletely observed: partially and integrated diusion processes (Work
in progress)
Back to epidemic data
Inference approach: Work in progress
Inference for discrete observations of diusion or Gaussian processes with small diusion coecient
Epidemics incompletely observed: partially and integrated diusion processes (Work in progress)Inference approach: Work in progress
Incidence for SIR models
Discrete observation of all the coordinates
Conned studies
Childhood diseases
SIR Model:
Incidence at t
2:
Z
t2 t1λ
S(t)
I (t)
N
dt
High frequency data ≈ λS(t
2)
I (t2)Nd small : new infected ≈ new removed,
Z
t2t1
γ
I (t)dt = R(t
2) −
R(t
1)
Diusion perspectives
Partial and discrete obs. of the diusion
process (Itô's formula)
Partial and Integrated discrete obs.
Inference for discrete observations of diusion or Gaussian processes with small diusion coecient
Epidemics incompletely observed: partially and integrated diusion processes (Work in progress)Inference approach: Work in progress
Partially observed diusion process: initial idea
Previous main idea not directly applicable:
i(t) s(t) S_5 S_5 S_6 I_5 I_6 εgi(t5) εgi(t6) εgs(t6) εgs(t5) Φi , i(t6,t5)εgi(t5) Φs ,i(t6,t5)εgs(t5) ε√ΔZ6i
Inference for discrete observations of diusion or Gaussian processes with small diusion coecient
Epidemics incompletely observed: partially and integrated diusion processes (Work in progress)Inference approach: Work in progress
Partially observed diusion process: initial idea
No good properties for n xed, → 0.
Inference for discrete observations of diusion or Gaussian processes with small diusion coecient
Epidemics incompletely observed: partially and integrated diusion processes (Work in progress)Inference approach: Work in progress