- 2010-2015 PhD in Statistics at JYU
- State space models (time series),
prediction problems, statistical software
- State space models (time series),
Jouni Helske (with Miikka Voutilainen and Harri Högmander)
By Jouni Helske, based on paper by Voutilainen, Helske, Högmander (2019) by the same name, conditionally accepted to Demography
Given the initial population estimate, full birth and death records:
μt=μt−1+Bt−Dt,t=1648,…,1850,
where
μt is the population
Bt=∑197i=1bt,i is the total number of births
Dt=∑197i=1dt,i is the total number of deaths at year t.
But we don’t know μ1647 nor yearly births and deaths…
Three types of missingness:
Census data Ct is likely not accurate either, but we’ll assume that
Ct∼N(μt,σ2C),t∈T,
μ1647∼Gamma(215,0.0005),μt∼Gamma(ψμ(μt−1+βt−δt−st),ψμ),t≠1697
βt and δt are annual birth and death estimates, st are the military casualties at year t
Prior for first year translates to prior mean 430,000 and prior sd of 30,000
Special treatment of year 1697 is needed due to the Great Famine∗
mean of μt is μt−1+βt−δt−st, sd is √μt/ψμ
We consider 177 parishes of 197 which have at least 10 years of baptism and burial records
βt=1λbt(∑i∈Ωbtbt,i+∑j∉Ωbtˆbt,j),
~λjt=[1+exp(−rj(t−mj))]−1,λjt=˜λjt−˜λj1648˜λj1850−˜λj1648(λ1850−λj1648)+λj1648,j=b,d.
bt,i∼Gamma(ψbdexp(νbt,i),ψbd),ˆbt,j∼Gamma(ψbdexp(νbt,j),ψbd),νbt,i∼N(νbt−1,i+ηbt,σ2b,ν),ηbt∼N(ηbt−1,σ2b,η),
- (using posterior samples of hyperparameters, βt, and δt)
Some references: