64
• φ = (y>1,..., φn) is the vector of indicator variables indexing the value of
(β↑,δ↑),..., (βκ,δ*κ) related with (βi,δi) for г = l,...,n. Thatis, φi = к,
if (∕¾Λ) = G¾*Λ)∙
This configuration determines the following set of variables:
• Sk := {i : ψi — k} is the set of sample indices of (βi,δβ related to (∕0⅛,5⅛) for
k = 1,...,M.
• nk := #Sk.
We will denote as K~l, nβ't and Sβτ for к — 1,..., К~г, (β*,δ*)~t = (^ɪ, <5j,)-∖ ...,
to the configuration corresponding to (β,δ')~l (βι,δι),..., (βi-l,δi+β),
(∕¾+1,⅝+ι),... ,(∕3n, δn). We now describe the corresponding steps (a)-(e) of Section
3.2 and extra step (f) for the proposed model.
(a) Given the current imputed values (K, (β*, δ*), φ), generate a new configuration
by simulating φl,..., φn from the complete conditional posterior distribution,
P{φi = k∖Ni, (β, δ)-i, φ~i, K~i) = ¾fc, for к = 0,..., K~l
where,
(∙⅝⅛)sg Γ(sja + М2)
<J^ + S0t0^+N" Γ(sz3)
{sβββi Γ(,s⅛ + M3 )
{μi + s6t0)ss+Ni3 r(ʤ) ]’
qik = cn~kiβ*k m2δ*k Ni3eXp{-^k + δ*k)}, for j = 1,..., K~i,
and c is a normalization constant such that ¾o ÷ ■ ∙ ∙ ÷ — 1. Whenever we
sample φi = 0, we generate a new observation (βi, δi) from
Gfα(∕¾∣sj3 + Ni2,Sβtβ ÷ μβ × Ga(¾∣s⅛ + Ni2, sβs + μi),
and update, accordingly, the new configuration by K — K + 1 and φi = n + 1.