1、采用MCMC(Markov Chain Monte Carlo)参数估计思想,具体的抽样方法选择吉布斯抽样方法(Gibbs sampling approach)在随意给定待估参数一个初始值之后,开始生成参数的新数值,并根据新数值生成其他参数的新数值,如此往复,对每一个待估参数,将得到一组生成的数值,根据该组数值,计算其均值,即为待估参数的贝叶斯估计值。三、贝叶斯空间计量模型的类型空间自回归模型 far_g()空间滞后模型(空间回归自回归混合模型) sar_g()空间误差模型 sem_g()广义空间模型(空间自相关模型) sac_g()四、贝叶斯空间模型与普通空间模型的选择标准首先按照参数显著性
2、,以及极大似然值,确定普通空间计量模型的具体类型,之后对于该确定的类型,再判断是否需要进一步采用贝叶斯估计方法。标准一:对普通空间计量模型的残差项做图,观察参数项是否是正态分布,若非正态分布,则考虑使用贝叶斯方法估计。 技巧:r=30的贝叶斯估计等价于普通空间计量模型估计,此时可以做出v的分布图,观察其是否基本等于1,若否,则应采用贝叶斯估计方法。标准二:若按标准一发现存在异方差,采用贝叶斯估计后,如果参数结果与普通空间计量方法存在较大差异,则说明采用贝叶斯估计是必要的。例1:选举 投票率 普通SAR与贝叶斯SAR对比: load elect.dat; load ford.dat; y=ele
3、ct(:,7)./elect(:,8); x1=elect(:,9)./elect(: x2=elect(:,10)./elect(: x3=elect(:,11)./elect(: w=sparse(ford(:,1),ford(:,2),ford(:,3); x=ones(3107,1) x1 x2 x3; res1=sar(y,x,w); res2=sar_g(y,x,w,2100,100); Vnames=strvcat(voter,const, educ, home, income); prt(res1);prt(res2);Spatial autoregressive Model
4、Estimates Dependent Variable = voter R-squared = 0.4605 Rbar-squared = 0.4600 sigma2 = 0.0041 Nobs, Nvars = 3107, 4 log-likelihood = 5091.6196 # of iterations = 11 min and max rho = -1.0000, 1.0000 total time in secs = 1.0530 time for lndet = 0.2330 time for t-stats = 0.0220 time for x-impacts = 0.7
5、380 # draws x-impacts = 1000 Pace and Barry, 1999 MC lndet approximation used order for MC appr = 50 iter for MC appr = 30 Variable Coefficient Asymptot t-stat z-probability const -0.100304 -8.406299 0.000000 educ 0.335704 21.901099 0.000000 home 0.754060 28.212211 0.000000 income -0.008135 -8.53521
6、2 0.000000 rho 0.527962 335.724359 0.000000 检验是否存在异方差-是否存在遗漏变量:贝叶斯-对列向量做柱状图。bar(res.vmean);Bayesian spatial autoregressive model Heteroscedastic model R-squared = 0.4425 Rbar-squared = 0.4419 mean of sige draws = 0.0023 sige, epe/(n-k) = 0.0065 r-value = 4 ndraws,nomit = 2100, 100 total time in secs
7、 = 20.6420 time for lndet = 0.2370 time for sampling = 19.2790 Posterior Estimates Variable Coefficient Std Deviation p-level const -0.107863 0.012729 0.000000 educ 0.348416 0.018072 0.000000 home 0.727799 0.026416 0.000000 income -0.009603 0.001050 0.000000 rho 0.561054 0.013313 0.000000 对遗漏变量的测量:
8、lat=elect(:,5);lon=elect(:,6); lons li=sort(lon); lats=lat(li,1); elects=elect(li,:); y=elects(:,7)./elects(: x1=elects(:,9)./elects(: x2=elecrs(:,10)./elects(: x2=elects(: x3=elects(:,11)./elects(: w1 w w2=xy2cont(lons,lats); vnames=strvcat(voters,consteduchomeincome res=sar(y,x,w,2100,100); res=sa
9、r_g(y,x,w,2100,100); prt(res,vnames);Dependent Variable = voters R-squared = 0.4402 Rbar-squared = 0.4396 mean of sige draws = 0.0022 total time in secs = 20.3230 time for lndet = 0.2460 time for sampling = 18.9770 *const -0.133182 0.012633 0.000000 educ 0.300653 0.017986 0.000000 home 0.725202 0.025944 0.000000 income -0.008219 0.001009 0.000000 rho 0.628407 0.014116 0.000000 例2:elect数据 2个权重矩阵-W1 W2W2=slag(W1,2) bres sar(sem/sac)_gSAR(2个) SEM(2个) SAC(4个)普通*贝叶斯 共计16个模型(注:可对变量统一取对数)
copyright@ 2008-2022 冰豆网网站版权所有
经营许可证编号:鄂ICP备2022015515号-1