Paper 11 Bayesian Modeling of Joint Regressions for the Mean and Covariance Matrix

这三篇应该都是很类似的,所以会比较简短。

Cepeda and Gamerman (2004)

11.1 Abstract

Present the Bayesian approach used to fit the models, based on a generalization of the Metropolis-Hastings algorithm of (Cepeda and Gamerman 2000). Real data is about growth and development of a group of deaf childeren. Then concluded with a few proposed extensions.

11.2 Introduction

Since the covariance matrix structure does not depend on the ordering of observations we can apply this model for the variance-covariance matrix in the analysis of spatial data.

However, this parameters do not have a practical interpretation, but estimation of the covariance matrix may lead to better estimates of the parameters of the mean model.

This paper will follow the modelling strategy proposed by Pourahmadi 1999, with similar approach presented by Pourhmadi and Daniels (2002).

Next section will present the models and reviews some topics of ante-dependence models (Daniel那篇),used in our proposed models for the covariance matrix. Section 3 presented the Bayesian proposal based on Fisher scoring for the MCMC algorithm. 这个proposal后面Xu,Liu用过,效果看起来还很好的样子。

11.3 The model

Ante-dependence model:

\[ Y_{i}-\mu_{i}=\sum_{j=1}^{i-1} \phi_{i j}\left(Y_{j}-\mu_{j}\right)+v_{i}, \quad v_{i} \sim N\left(0, \sigma_{i}^{2}\right), \quad i=1, \ldots, n \]

Then can be re-write as

\[ v=T(Y-\mu), \quad v \sim N(0, D) \quad \text { and } \quad D=\operatorname{diag}\left(\sigma_{i}^{2}\right) \] \(T=(t_{ij})\) with \[ t_{i j}=\left\{\begin{array}{cl}{1} & {\text { if } j=i} \\ {-\phi_{i j}} & {\text { if } j<i} \\ {0} & {\text { elsewhere }}\end{array}\right. \] and \[ \operatorname{Var}(v)=D=T \operatorname{Var}(Y-\mu) T^{\prime}=T \Sigma T^{\prime} \] 然后有 \[ \tilde{Y}=\left(I_{n}-T\right) \tilde{Y}+v \] 以及 \[ \phi_{i j}=w_{i j}^{\prime} \lambda, \quad 1 \leq j<i \leq n \] Since we have \(\phi_{i j}=\Sigma_{l=1}^{r} w_{i j, l} \lambda_{l}\), the matrix \(I_{n}-T\) can be expressed as the linear combination \[ I_{n}-T=\lambda_{1} W_{1}+\ldots+\lambda_{r} W_{r} \] As a consequence, model can be expressed as \[ \begin{aligned} \tilde{Y} &=\lambda_{1} W_{1} \tilde{Y}+\ldots+\lambda_{r} W_{r} \tilde{Y}+v \\ &=\lambda_{1} V_{1}+\ldots+\lambda_{r} V_{r}+v \\ &=V \lambda+v \end{aligned} \] That is, \(\tilde{Y} \sim N(V \lambda, D)\).

Then the full model is \[ h\left(\mu_{i}\right)=x_{i}^{\prime} \beta, \quad g\left(\sigma_{i}^{2}\right)=z_{i}^{\prime} \gamma, \quad f\left(\phi_{i j}\right)=w_{i j}^{\prime} \lambda \]

11.4 Bayesian Methodology

The likelihood function is given by

\[ L(\beta, \gamma, \lambda | \mathrm{Y}) \propto|D|^{-1 / 2} \exp \left\{-\frac{1}{2}(Y-\mu)^{\prime} \Sigma^{-1}(Y-\mu)\right\} \] since \(|\Sigma|=\left|T^{\prime}\right||D||T|=|D|\).

A prior must be assigned to obtain the posterior distribution. For simplicity, assume \(\theta \sim N\left(\theta_{0}, \Sigma_{0}\right)\) where \(\theta_{0}=\left(b_{0}, g_{0}, l_{0}\right)^{\prime}\). The values of \((b,g,l)\) and \((B,G,L)\) can be evaluated from \(\theta_0\) and \(\Sigma_0\).

The posterior is \[ \pi(\beta, \gamma, \lambda) \propto|D|^{-1 / 2} \exp \left\{-\frac{1}{2}(Y-X \beta)^{\prime} \Sigma^{-1}(Y-X \beta)-\frac{1}{2}\left(\theta-\theta_{0}\right)^{\prime} \Sigma_{0}^{-1}\left(\theta-\theta_{0}\right)\right\} \]

\[ (\beta | \gamma, \lambda, Y) \sim N\left(b^{*}, B^{*}\right) \]

\[ (\beta | \gamma, \lambda, Y) \sim N\left(b^{*}, B^{*}\right) \] Let \[ Q(Y)=\tilde{Y}^{\prime} T^{\prime} D^{-1} T \tilde{Y}=(\tilde{Y}-V \lambda)^{\prime} D^{-1}(\tilde{Y}-V \lambda) \]

Therefore, it is easy to obtain that

\[ (\lambda | \beta, \gamma, Y) \sim N\left(l^{*}, L^{*}\right) \] with \(l^{*}=L^{*}\left(L^{-1} l+V^{\prime} D^{-1} \tilde{Y}\right)\) and \(L^{*}=\left(L^{-1}+V^{\prime} D^{-1} V\right)^{-1}\).

The full conditional distribution of \(\gamma\) given by \[ \pi(\gamma | \beta, \lambda) \propto|D|^{-1 / 2} \exp \left\{-\frac{1}{2}(\tilde{Y}-V \lambda)^{\prime} D^{-1}(\tilde{Y}-V \lambda)-\frac{1}{2}(\gamma-g)^{\prime} G^{-1}(\gamma-g)\right\} \] is intractable analytically and not easily generated from, then construct suitable proposals for a MH step.

11.5 Centering and Order Methods

References

Cepeda, Edilberto, and Dani Gamerman. 2004. “Bayesian modeling of joint regressions for the mean and covariance matrix.” Biometrical Journal, April.

Cepeda, Edilberto, and Dani Gamerman. 2000. “Bayesian modeling of variance heterogeneity in normal regression models.” Brazilian Journal of Probability and Statistics 14 (2): 207–21.