Page 22 - Contributed Paper Session (CPS) - Volume 1
P. 22

CPS653 Chang-Yun L.
                  that Σ is  a  biased  estimator  of Σ (Kenward  and  Roger,  1997),   and Vare
                                                                                  ̃
                                                                                        ̃
                       ̂
                  commonly used for hypothesis tests in the regression methods.
                     For SPDS and BDS designs, the full second-order model is inestimable due
                  to insufficient degrees of freedom. Hence, the process for model selections
                  must  be  implemented. Two  regression  methods,  MSR  and  FSR,  have  been
                  proposed in the literature for performing model selections for SPDS and BDS
                  designs. The two methods have the following limitations: (i) the active effects
                  cannot be estimated if the number of them is greater than the rank of the
                  model matrix and (ii) ̂ =  0 may be obtained by using the REML method. To
                                        
                  circumvent these problems, we propose a Bayesian variable selection method
                  as described in the next section.

                  3.  Stochastic search variable selection
                      Let  = ( ,  ,  ,  ,  ) denote  the  vector  of  the  model  parameters,
                           ′
                                 ′
                                    ′
                                       ′
                                            
                                         
                  where   = ( ,· · · ,  ,   ,· · · ,  −1, ,  ,· · · ,    )′  .  By  the  Baye’s
                                                         11
                               1
                                          12
                                      
                  Theorem  (see  DeGroot,  1970,  p28,  and  Robinson  et  al.,  2012),  the  joint
                  posterior density of the model parameters is proportional to the product of
                  the likelihood (y|Θ) and the joint prior density π(Θ), which is

                                                                 (|) ∝  (|)().                                          (4)

                  The  responses  conditioned  on  the  model  parameters  are  assumed
                  independent and hence the likelihood in (4) has the form

                                                      
                                         (y|Θ) = ∏ ∏ ( | ,  ),
                                                                
                                                                    
                                                             
                                                  −1  =1

                  where  µ  =   + ∑      + ∑ −1 ∑        + ∑     2  +  .  The
                                      =1
                                                   =1
                                            ,
                                                                                      
                                                                              ,
                                                                         =1
                           
                                                        =+1
                                                              , ,
                                 0
                  joint prior density of the model parameters can be obtained by

                                             (Θ) ∝ ∏  |[] ()
                                                     ∈Θ

                                         −1                             
                    ∝  ( ) ∏    ( ) ∏ ∏         ( ) ∏      ( ) ∏     ( )
                         0  0    |         |       |       |   
                              =1         =1  =+1        =1           =1

                                   −1                 
                        × ∏  ( ) ∏ ∏            ( ) ∏     ( ) ( ) ( ).
                                           | ,        |             
                           =1      −1  =+1         =1
                  To obtain an approximated distribution, we apply the Markov chain Monte
                  Carlo method (Gelfand and Smith, 1990, Casella and George, 1992, and Chib
                  and Greenberg, 1995). The Gibbs sampling introduced in the next section is a
                  widely  used  algorithm  for  simulating  Markov  chains.  The  algorithm
                  sequentially generates samples from the full conditional posterior distribution
                  for each parameter conditioned on the current values of all other parameters.
                                                                      11 | I S I   W S C   2 0 1 9
   17   18   19   20   21   22   23   24   25   26   27