8. 階層モデル
• この論論⽂文では、次のモデルに議論論を絞る
• データ yij は正規分布に従うが、平均値は
グループごとに異異なる。
• グループごとの平均値の分散 σα
2
e basic hierarchical model
ork with a simple two-level normal model of data yij with group
yij ∼ N(µ + αj, σ2
y), i = 1, . . . , nj, j = 1, . . . , J
αj ∼ N(0, σ2
α), j = 1, . . . , J.
discuss other hierarchical models in Section 7.2.
(1) has three hyperparameters—µ, σy, and σα—but in this paper
nly with the last of these. Typically, enough data will be avail
d σy that one can use any reasonable noninformative prior distri
(µ, σy) ∝ 1 or p(µ, log σy) ∝ 1.
noninformative prior distributions for σα have been suggested8
27. 3. 理理論論的考察
• 階層ベイズモデルの階層分散パラメータ
σα に対して、どんな無情報事前分布を
使⽤用したらいいかについて考察する。
lly-conjugate family. We propose a half-t model and demonstra
nformative prior distribution and as a component in a hierarchic
arameters.
e basic hierarchical model
ork with a simple two-level normal model of data yij with group
yij ∼ N(µ + αj, σ2
y), i = 1, . . . , nj, j = 1, . . . , J
αj ∼ N(0, σ2
α), j = 1, . . . , J.
discuss other hierarchical models in Section 7.2.
(1) has three hyperparameters—µ, σy, and σα—but in this paper
nly with the last of these. Typically, enough data will be avail
d σy that one can use any reasonable noninformative prior distri
(µ, σ ) ∝ 1 or p(µ, log σ ) ∝ 1.
27
35. 4. 実際のデータに適⽤用
• 8-schools データ
• 8 つの学校で⾏行行われた共通テストの点数
• 階層モデルにより学校間の得点差をモデル化
• σα に対して無情報事前分布を適⽤用してみる
lly-conjugate family. We propose a half-t model and demonstra
nformative prior distribution and as a component in a hierarchic
arameters.
e basic hierarchical model
ork with a simple two-level normal model of data yij with group
yij ∼ N(µ + αj, σ2
y), i = 1, . . . , nj, j = 1, . . . , J
αj ∼ N(0, σ2
α), j = 1, . . . , J.
discuss other hierarchical models in Section 7.2.
(1) has three hyperparameters—µ, σy, and σα—but in this paper
nly with the last of these. Typically, enough data will be avail
d σy that one can use any reasonable noninformative prior distri35
47. 3-schools 半コーシー分布
• σα 〜~ HalfCauchy(25)
• 半コーシーでは、右裾が抑えられる
variance parameters in hierarchical models
00
n
σα
0 50 100 150 200
3 schools: posterior on σα given
half−Cauchy (25) prior on σα
ulations of the between-school standard deviation,47