Page 23 - Contributed Paper Session (CPS) - Volume 1
P. 23

CPS653 Chang-Yun L.
            The resulting values of the latent variables which have higher marginal or joint
            posterior  probability  are  then  used  to  identify  active  factors  and  select
            promising  models  for  further  consideration.  This  procedure  is  called  the
            stochastic search variable selection (SSVS) by George and McCulloch (1993).

            4.  Examples: the SSVS method for SPDS designs
            We apply the proposed SSVS method to analyze the data for the eight-factor
            SPSD design in Lin and Yang (2015). The SPDS design and the responses are
            listed in Table 1. This design was constructed by arranging a 17×8 definitive
            screening design into nine unbalanced whole plots (  =  17,   =  8, and   =
             9), in which   and   are  whole-plot  factors  and   ,· · · ,   are  subplot
                                 2
                          1
                                                                         8
                                                                  3
            factors.  The  responses  were  generated  from  the  true  model ( ) =  50 −
                                                2
                                       2
            4  + 3.5  +    + 3.5  − 4.5  and the covariance matrix   =  ′  +
                       3
               1
                                                3
                             1 3
                                       1
             . We choose the tuning parameters   =  10 and   = .5. To ensure that the
             
            selected models follow the effect heredity principle, we set   =   = .01
                                                                               00
                                                                        0
            and  =  =  11  = .75 To reduce the uncertainty, we use   =   =   =
                                                                             1
                                                                                   11
                      1
             .75 (lower probability increases the model uncertainty). The Gibbs sampling
            with  11000  iterations  are  conducted  by  using  the  WinBUGS  code.  The
            algorithm  starts  with   =  0 for  the  intercept,   =  0 and   =  0 for  the
                                   0
                                                             
                                                                          
            fixed  effects,    =  1  for  the  random  effects,  and    =   =  1  for  the
                                                                         
                            
                                                                   
            variances of the whole-plot and random errors. A burn-in of the first 1000
            iterations is taken for eliminating the non-stationary portion of the chain at
            beginning and then only every 5th value from the Gibbs sampling

                           Table 1: The SPDS design and data in Lin and Yang (2015).
                   Whole plot     X1  X2    X3  X4  X5  X6  X7  X8           Y
                       1          -1   -1     1   -1    1    0    -1    1  55.073
                       1          -1   -1     1    1   -1    1     0   -1  56.359
                       1          -1   -1    -1    1    1    -1    1    0  50.529
                       2          -1    0    -1   -1   -1    1     1    1  50.349
                       3          -1    1     0    1   -1    -1   -1    1  58.019
                       3          -1    1    -1    0    1    1    -1   -1  49.619
                       3          -1    1     1   -1    0    -1    1   -1  55.049
                       4           0    1     1    1    1    1     1    1  48.806
                       5           0    0     0    0    0    0     0    0  48.955
                       6           0   -1    -1   -1   -1    -1   -1   -1  42.708
                       7           1   -1     0   -1    1    1     1   -1  50.650
                       7           1   -1     1    0   -1    -1    1    1  51.752
                       7           1   -1    -1    1    0    1    -1    1  41.401
                       8           1    0     1    1    1    -1   -1   -1  48.219
                       9           1    1    -1    1   -1    0     1   -1  41.676
                       9           1    1    -1   -1    1    -1    0    1  42.943
                       9           1    1     1   -1   -1    1    -1    0  52.089

                                                                12 | I S I   W S C   2 0 1 9
   18   19   20   21   22   23   24   25   26   27   28