Continuous time likelihood calculation
Purpose
This document outlines a new form of the likelihood for an epidemic model, using a continuous-time approximation to a discrete Markov process.
State transition model
Epidemic models are represented as State Transition Models, where at any time in an epidemic individuals are assumed to exist in one of a number of states s = 1, \dots, S
connected by directed edges, or transitions, r=1,\dots,R
. For example:
graph LR;
s1 -- r1 --> s2;
s2 -- r2 --> s3;
s2 -- r3 --> s4;
Notation
Let us assume a y_{tmr}
represents the number of events occurring at timepoint t
in epidemiological unit m=1,\dots,M
of type r=1,\dots,R
, where r
represents a transition (or edge) in a State Transition Model graph. For convenience, we refer to the event tensor Y
with shape [T, M, R]
.
We represent the number of individuals in each epidemiological unit m
at time t
in state s
as x_{tms}
, and refer to a state tensor X
with shape [T, M, S]
.
Likelihood computation
Consider a timepoint t
and state s
. Under the (approximate) continuous-time model, the probability of y_{tmr}
events transitioning out of s
via transition r
at the end of a time interval \delta t
is
f(y_{tmr} | \lambda_{mr}(t), x_{tms}) = (\lambda_{mr}(t))^{y_{tmr}} \exp \left( -x_{tms}\sum_{r \in o(s)} \lambda_{mr}(t) \delta t \right)
where \lambda_{mr}(t)
is the hazard rate for the r
th transition, and o(s)
is the set of outgoing transitions from state s
.
Remarks:
- This is essentially survival analysis, where the lifespan function is a product of a hazard rate (
\lambda_{mr}(t)
) and survivor functionS(T > t) = 1 - F(T \leq t) = \exp (-\lambda_{mr}(t) \delta t