本⽇日紹介する論論⽂文• “Prior distributionsfor varianceparameters in hierarchical models”• (階層モデルの分散パラメータの事前分布)• by Andrew Gelman• Bayesian Analysis 2006https://projecteuclid.org/euclid.ba/13403710482
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 σα2e basic hierarchical modelork with a simple two-level normal model of data yij with groupyij ∼ N(µ + αj, σ2y), i = 1, . . . , nj, j = 1, . . . , Jαj ∼ N(0, σ2α), j = 1, . . . , J.discuss other hierarchical models in Section 7.2.(1) has three hyperparameters—µ, σy, and σα—but in this papernly with the last of these. Typically, enough data will be availd 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
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 demonstranformative prior distribution and as a component in a hierarchicarameters.e basic hierarchical modelork with a simple two-level normal model of data yij with groupyij ∼ N(µ + αj, σ2y), i = 1, . . . , nj, j = 1, . . . , Jαj ∼ N(0, σ2α), j = 1, . . . , J.discuss other hierarchical models in Section 7.2.(1) has three hyperparameters—µ, σy, and σα—but in this papernly with the last of these. Typically, enough data will be availd σy that one can use any reasonable noninformative prior distri(µ, σ ) ∝ 1 or p(µ, log σ ) ∝ 1.27
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 demonstranformative prior distribution and as a component in a hierarchicarameters.e basic hierarchical modelork with a simple two-level normal model of data yij with groupyij ∼ N(µ + αj, σ2y), i = 1, . . . , nj, j = 1, . . . , Jαj ∼ N(0, σ2α), j = 1, . . . , J.discuss other hierarchical models in Section 7.2.(1) has three hyperparameters—µ, σy, and σα—but in this papernly with the last of these. Typically, enough data will be availd σy that one can use any reasonable noninformative prior distri35
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 models00nσα0 50 100 150 2003 schools: posterior on σα givenhalf−Cauchy (25) prior on σαulations of the between-school standard deviation,47