mess_g.m

计量工具箱

                 alpha = alpha2;
              end;

          rtmp(iter,1) = alpha;

      
     % evaulate the likelihood using current draws
     if lflag == 0
                like = -(n/2)*log(2*pi*sige) - (e'*e)/(2*sige);
     end;
                       
     % update rval
     if mm ~= 0           
     rval = gamm_rnd(1,1,mm,kk);  
     end;
              
    if iter > nomit % if we are past burn-in, save the draws
    bsave(iter-nomit,:) = bhat';
    ssave(iter-nomit,1) = sige;
    asave(iter-nomit,1) = alpha; 
     if lflag == 0
       lsave = lsave + like;
     else
       lsave = lsave + 0;
     end;
    end;
                    
    if iter == nomit % update cc based on initial draws
         tst = 2*std(rtmp(1:nomit,1));
         if tst > 0.1
         cc = tst;
         end;
    end;

iter = iter + 1; 
waitbar(iter/ndraw);         
end; % end of sampling loop
close(hwait);

stime = etime(clock,t0);

% compute posterior means
if lflag == 0
lmean = lsave/(ndraw-nomit);
else 
   lmean = 0;
end;
amean = mean(asave);
bmean = mean(bsave);
astd = std(asave);
bstd = std(bsave);
smean = mean(ssave);

% find acceptance rate
results.accept = 1 - cnta/(iter+cnta);
% NOTE: this could be interpreted as the
% probability that alpha is in the mesh grid

% do the expensive calculation here
% rather than lookup
tmp = rho.^(0:neigh-1);
tmp = tmp/sum(tmp);

wy = y;
Y = y(:,ones(1,q));
for i=2:q;
wy = wy(nnlist)*tmp';
Y(:,i) = wy;
end;
[junk nq] = size(Y);
nq1 = nq-1;
v = ones(nq,1);
for i=2:nq;
v(i,1) = amean.^(i-1);
end;
W = (1./[1 cumprod(1:nq1)]);
sy = Y*diag(W)*v;

e = sy - x*bmean';
yhat = y - e;

sigu = e'*e;
ym = y - mean(y);
rsqr1 = sigu;
rsqr2 = ym'*ym;
rsqr = 1.0 - rsqr1/rsqr2; % r-squared
rsqr1 = rsqr1/(n-k);
rsqr2 = rsqr2/(n-1.0);
rbar = 1 - (rsqr1/rsqr2); % rbar-squared

time = etime(clock,timet);


results.meth  = 'mess_g';
results.bdraw = bsave;
results.adraw = asave;
results.bmean = bmean';
results.bstd  = bstd';
results.amean = amean;
results.astd  = astd;
results.smean = smean;
results.sdraw = ssave;
results.rho   = rho;
results.lmean = lmean;
results.bprior = c;
results.bpstd  = sqrt(diag(T));
results.nobs  = n;
results.nvar  = k;
results.ndraw = ndraw;
results.nomit = nomit;
results.time  = time;
results.stime = stime;
results.ntime = gtime;
results.nu = nu;
results.d0 = d0;
results.tflag = 'plevel';
results.aflag = pflag;
results.palpha = palpha;
results.acov = S;
results.y = y;
results.yhat = yhat;
results.resid = e;
results.rsqr = rsqr;
results.rbar = rbar;
results.neigh = neigh;
results.q     = q;
results.nobs = n;
results.nvar = k;
results.xflag = xflag;
results.nflag = nflag;

case{1} % case of x-variables transformed
   
   xone = x(:,1);
   if all(xone == 1)
      xsub = x(:,2:k);
   else
      xsub = x;
   end;

t1 = clock;   % time this operation

tmp = rho.^(0:neigh-1);
tmp = tmp/sum(tmp);

% we have to construct the weight matrix using neighbors
% find index into nearest neighbors
if nflag == 0
nnlist = find_nn(latt,long,neigh);
elseif nflag == 1
nnlist = find_nn2(latt,long,neigh);
else
error('mess_g: bad nflag option');
end;

% check for empty nnlist columns
chk = find(nnlist == 0);
if length(chk) > 0; 
 if nflag == 1 % no saving the user here
 error('mess_g: trying too many neighbors, some do not exist');
 else % we save the user here
 nnlist = find_nn2(latt,long,neigh);
 end;
end;

% construct and save Sy
wy = y;
Y = y(:,ones(1,q));
for i=2:q;
wy = wy(nnlist)*tmp';
Y(:,i) = wy;
end;

% create and save Sx
[junk nk] = size(xsub);
xout = x;
for i=1:nk;
xi = xsub(:,i);
tmpp = xi(nnlist)*tmp';
xout = [xout tmpp];
end;
xmat = xout;

% end of up front stuff with Sy saved 
gtime = etime(clock,t1);

% ====== initializations
% compute this stuff once to save time
[junk kk] = size([x xsub]); % need to add diffuse priors 
                          % to the spatial lags of x-variables
Tnew = eye(kk)*1e+12;
Tnew(1:k,1:k) = T;
TI = inv(Tnew);
ctmp = zeros(kk,1);
ctmp(1:k,1) = c;
c = ctmp;
TIc = TI*c;
cc=0.2; % initial metropolis value
cntr = 0; 
iter = 1;
alpha = astart;
rho = 1;
in = ones(n,1);
sige = sig0;

% storage for draws
          bsave = zeros(ndraw-nomit,kk);
          asave = zeros(ndraw-nomit,1);
          ssave = zeros(ndraw-nomit,1);
           lsave = 0;
          rtmp = zeros(nomit,1);



hwait = waitbar(0,'MCMC sampling ...');
t0 = clock;                  
iter = 1;
          while (iter <= ndraw); % start sampling;

          [junk nq] = size(Y);
          nq1 = nq-1;
          v = ones(nq,1);
          for i=2:nq;
          v(i,1) = alpha.^(i-1);
          end;
          W = (1./[1 cumprod(1:nq1)]);
          Sy = Y*diag(W)*v;

          % update beta   
          AI = inv(xmat'*xmat + sige*TI);
          
          b = xmat'*Sy + sige*TIc;
          b0 = AI*b;
          bhat = norm_rnd(sige*AI) + b0;  
         
          % update sige
          nu1 = n + 2*nu; 
          e = (Sy - xmat*bhat);
          d1 = 2*d0 + e'*e;
          chi = chis_rnd(1,nu1);
          sige = d1/chi;

          % metropolis step to get alpha update
          if pflag == 0
          alphax = c_mess(alpha,y,xmat,Y,bhat,sige); 
          elseif pflag == 1
          alphax = c_mess(alpha,y,xmat,Y,bhat,sige,palpha,S); 
          end;
          
          accept = 0;
          alpha2 = alpha + cc*randn(1,1);
          while accept == 0
           if alpha2 <=0 
           accept = 1;  
           else
           alpha2 = alpha + cc*randn(1,1);
           cntr = cntr+1; % counts accept rate
           end; 
          end; 
           if pflag == 0
           alphay = c_mess(alpha2,y,xmat,Y,bhat,sige);
           elseif pflag == 1
           alphay = c_mess(alpha2,y,xmat,Y,bhat,sige,palpha,S);
           end;
          
          ru = unif_rnd(1,0,1);
          if ((alphay - alphax) > exp(1)),
          p = 1;
          else,          
          ratio = exp(alphay-alphax);
          p = min(1,ratio);
          end;
              if (ru < p)
                 alpha = alpha2;
              end;

          rtmp(iter,1) = alpha;
      
     % evaulate the likelihood using current draws
     if lflag == 0
                like = -(n/2)*log(2*pi*sige) - (e'*e)/(2*sige);
     end;
                       
     % update rval
     if mm ~= 0           
     rval = gamm_rnd(1,1,mm,kk);  
     end;
              
    if iter > nomit % if we are past burn-in, save the draws
    bsave(iter-nomit,:) = bhat';
    ssave(iter-nomit,1) = sige;
    asave(iter-nomit,1) = alpha; 
    if lflag == 0
       lsave = lsave + like;
    else
       lsave = lsave + 0;
    end;
    
    
    end;

       if iter == nomit % update cc based on initial draws
         tst = 2*std(rtmp(1:nomit,1));
         if tst > 0.05
         cc = tst;
         end;
    end;
                 

iter = iter + 1; 
waitbar(iter/ndraw);         
end; % end of sampling loop
close(hwait);

stime = etime(clock,t0);

% compute posterior means
if lflag == 0
lmean = lsave/(ndraw-nomit);
else 
   lmean = 0;
end;
amean = mean(asave);
bmean = mean(bsave);
astd = std(asave);
bstd = std(bsave);
smean = mean(ssave);

% find acceptance rate
results.accept = 1 - cntr/(iter+cntr);

tmp = rho.^(0:neigh-1);
tmp = tmp/sum(tmp);

% construct Sy, Sx based on posterior means
wy = y;
Y = y(:,ones(1,q));
for i=2:q;
wy = wy(nnlist)*tmp';
Y(:,i) = wy;
end;
[junk nq] = size(Y);
nq1 = nq-1;
v = ones(nq,1);
for i=2:nq;
v(i,1) = amean.^(i-1);
end;
W = (1./[1 cumprod(1:nq1)]);
sy = Y*diag(W)*v;

% create  Sx based on posterior mean of rho
xmat = x;
for i=1:nk;
xi = xsub(:,i);
tmpp = xi(nnlist)*tmp';
xmat = [xmat tmpp];
end;

e = Sy - xmat*bmean';
yhat = y - e;

sigu = e'*e;
ym = y - mean(y);
rsqr1 = sigu;
rsqr2 = ym'*ym;
rsqr = 1.0 - rsqr1/rsqr2; % r-squared
rsqr1 = rsqr1/(n-k);
rsqr2 = rsqr2/(n-1.0);
rbar = 1 - (rsqr1/rsqr2); % rbar-squared

time = etime(clock,timet);


results.meth  = 'mess_g';
results.bdraw = bsave;
results.adraw = asave;
results.bmean = bmean';
results.bstd  = bstd';
results.amean = amean;
results.astd  = astd;
results.smean = smean;
results.sdraw = ssave;
results.lmean = lmean;
results.bprior = c;
results.bpstd  = sqrt(diag(Tnew));
results.nobs  = n;
results.nvar  = k;
results.ndraw = ndraw;
results.nomit = nomit;
results.time  = time;
results.stime = stime;
results.nu = nu;
results.d0 = d0;
results.tflag = 'plevel';
results.aflag = aflag;
results.palpha = palpha;
results.acov = S;
results.y = y;
results.yhat = yhat;
results.resid = e;
results.rsqr = rsqr;
results.rbar = rbar;
results.rho   = rho;
results.neigh = neigh;
results.q     = q;
results.ntime = gtime;
results.nobs = n;
results.nvar = k;
results.xflag = xflag;


otherwise
   
end; % end of switch


   
function cout = c_mess(alpha,ys,xs,Symat,beta,sige,a,B);
% PURPOSE: evaluate the conditional distribution of alpha 
%          for the Bayesian mess_g model
% ---------------------------------------------------
%  USAGE: cout = c_mess2(alpha,y,x,Symat,beta,alpha,rho)
%  where:  alpha  = matrix exponential alpha spatial parameter
%            y    = dependent variable vector
%            x    = explanatory variables matrix
%          Symat  = matrix from mess_g 
%            beta = kx1 current bhat vector
%            sige = current value of sige
%               a = prior mean for alpha     (optional input)
%               B = prior variance for alpha (optional input)
% ---------------------------------------------------
%  RETURNS: a conditional used in Metropolis-Hastings sampling
%  NOTE: called only by mess_g
%  --------------------------------------------------
%  SEE ALSO: mess_g, mess_gd
% ---------------------------------------------------

% written by: James P. LeSage 1/2000
% University of Toledo
% Department of Economics
% Toledo, OH 43606
% [e-mail]


n = length(ys);

[junk nq] = size(Symat);
nq1 = nq-1;
v = ones(nq,1);
for i=2:nq;
v(i,1) = alpha.^(i-1);
end;
W = (1./[1 cumprod(1:nq1)]);
Sy = Symat*diag(W)*v;

e = Sy - xs*beta;
epe = e'*e;

if nargin == 6
cout =  -epe/(2*sige);
elseif nargin == 8
B = B*sige;
cout = -epe/(2*sige) - 0.5*(((alpha-a)^2)/B);
else
error('c_mess: Wrong # of inputs arguments');
end;