Open marciomacielbastos opened 5 years ago
Here we can go to a stochastic SIR model or just add lognormal noise like so:
y = odeint(SIR, t=times, y0=[0.99, 0.01], args=((beta, gamma),), rtol=1e-8)
# Simulando dados Assumindo uma distribuição log-normal com média igual às séries simuladas
yobs = np.random.lognormal(mean=np.log(y[1::]), sigma=[0.2, 0.3])
I like the log normal alternative. It's simpler and is in line with the literature. Moreover, we would then have a correctly specified error model. Using a stochastic SIR would be nice as an extra, to try and study what happens when the error model is misspecified.
There's a slight problem with this:
os valores da distribuição lognormal extrapolam o limite superior do domínio do modelo (0,1) we'd have to truncate it
One way to combat this is to generate a truncated log normal directly, instead of truncating post facto.
Randomness must maintain the model representative of an actual epidemic.