T Statistics HW-5 (Arefin-2824475)

Bayesian Statistics
3470: 589-001
HW - 05 Mir Shahnewaz Arefin
Fall 2016 Student ID No: 2824475
Figure 3.4 (A first course in Bayesian methods)
> par(mfrow=c(2,2))
> theta<-seq(0,1,length=100)
> a<-1; b<-1
> n<-5 ; y<-1
> plot(theta,dbeta(theta,a+y,b+n-y),type="l",ylab=expression(paste(italic("p("),theta,"|y)",sep="")),
xlab=expression(theta), lwd=2)
> mtext(expression(paste("beta(1,1) prior, ", italic("n"),"=5 ",italic(sum(y[i])),"=1",sep="")), side=3,line=.1)
> #abline(v=c((a+y-1)/(a+b+n-2),(a+y)/(a+b+n)),col=c("black","gray"),lty=c(2,2))
> lines(theta,dbeta(theta,a,b),type="l",col="gray",lwd=2)
> legend(.45,2.4,legend=c("prior","posterior"),lwd=c(2,2),col=c("gray","black"), bty="n")
> a<-3; b<-2
> n<-5 ; y<-1
> #expression(italic(paste("p(",theta,"|y)",sep=""))), xlab=expression(theta),lwd=2)
> #abline(v=c((a+y-1)/(a+b+n-2), (a+y)/(a+b+n)) , col=c("green","red") )
> a<-1 ; b<-1
> n<-100; y<-20
> #expression(italic(paste("p(",theta,"|y)",sep=""))), xlab=expression(theta),lwd=2)
> a<-3 ; b<-2
> n<-100; y<-20
> #expression(italic(paste("p(",theta,"|y)",sep=""))),xlab=expression(theta),
> #lwd=2)
>
Page 1 of 22
Bayesian Statistics
3470: 589-001
> theta.support<-seq(0,1,length=5000)
> plot(theta.support, dbeta(theta.support, a+y, b+n-y),
type="l",xlab=expression(theta),ylab=expression(paste(italic("p("),theta,"|y)")))
> pth<-dbeta(theta.support, a+y, b+n-y)
> pth<-pth
> ord<- order(-pth)
> xpx<-cbind(theta.support[ord], pth[ord])
> xpx<-cbind(xpx,cumsum(xpx[,2])/sum(xpx[,2]))
> hpd<-function(x,dx,p){
+ md<-x[dx==max(dx)]
+ px<-dx/sum(dx)
+ pxs<--sort(-px)
+ ct<-min(pxs[cumsum(pxs)< p])
Page 2 of 22
Bayesian Statistics
3470: 589-001
+ list(hpdr=range(x[px>=ct]),mode=md) }
> tmp<-hpd(xpx[,1],xpx[,2],.5)$hpdr
> lines( x=c(tmp[1],tmp[1],tmp[2],tmp[2]),
+ y=dbeta(c(0,tmp[1],tmp[2],0),a+y,b+n-y) ,col=gray(.75),lwd=2 )
+ y=dbeta(c(0,tmp[1],tmp[2],0),a+y,b+n-y) ,col=gray(.5),lwd=2 )
+ y=dbeta(c(0,tmp[1],tmp[2],0),a+y,b+n-y) ,col=gray(0),lwd=2 )
> tmp<-qbeta( c(.025,.975), a+y,b+n-y)
+ y=dbeta(c(0,tmp[1],tmp[2],0),a+y,b+n-y) ,col=gray(0),lwd=2 ,lty=2 )
> legend(.5, 2.75, c("50% HPD","75% HPD","95% HPD","95% quantile-
based"),col=c(gray(.75),gray(.5),gray(0),gray(0)),lty=c(1,1,1,2),lwd=c(2,2,2,2),bty="n")
Page 3 of 22
Bayesian Statistics
3470: 589-001
> par(mfrow=c(2,3))
> a<-1 ; b<-1
> x<-seq(.001,10,length=100)
> plot(x, dgamma(x,a,b),type="l",
+ xlab=expression(theta), ylab=expression(italic(paste("p(",theta,")",sep=""))))
> mtext(expression(italic(paste("a=",1," b=",1,sep=""))),side=3,line=.12,cex=.8)
> a<-2 ; b<-2
> x<-seq(.001,10,length=100)
> a<-4 ; b<-4
> x<-seq(.001,10,length=100)
> a<-2 ; b<-1
> x<-seq(.001,10,length=100)
> a<-8 ; b<-4
> x<-seq(.001,10,length=100)
> a<-32 ; b<-16
> x<-seq(.001,10,length=100)
Page 4 of 22
Bayesian Statistics
3470: 589-001
> par(mfrow=c(1,2))
> a<-2
> b<-1
> xtheta<-seq(0,5,length=1000)
plot(xtheta,dgamma(xtheta,a+s1,b+n1),type="l",col=gray(.5),xlab=expression(theta),ylab=expression(paste(italic("p
("),theta,"|",y[1],"...",y[n],")",sep="")))
> lines(xtheta,dgamma(xtheta,a+s2,b+n2),col=gray(0),lwd=2)
> lines(xtheta,dgamma(xtheta,a,b),type="l",lty=2,lwd=2)
> abline(h=0,col="black")
> y<-(0:12)
> plot(y-.1, dnbinom(y, size=(a+s1), mu=(a+s1)/(b+n1)) , col=gray(.5)
,type="h",ylab=expression(paste(italic("p("),y[n+1],"|",y[1],"...",y[n],")",sep="")),xlab=expression(italic(y[n+1])),yl
im=c(0,.35),lwd=3)
> points(y+.1, dnbinom(y, size=(a+s2), mu=(a+s2)/(b+n2)) , col=gray(0) ,type="h",lwd=3)
> legend(1,.375,legend=c("Less than bachelor's","Bachelor's or
higher"),bty="n",lwd=c(3,3),col=c(gray(.5),gray(0)))
Page 5 of 22
Bayesian Statistics
3470: 589-001
> par(mfrow=c(1,3))
> set.seed(1)
> a<-2 ; b<-1
> sy<-66 ; n<-44
> nsim<-1000
> theta.sim<-rgamma(nsim,a+sy,b+n)
> #cumulative mean
> cmean<-cumsum(theta.sim)/(1:nsim)
> cvar<- cumsum(theta.sim^2)/(1:nsim) - cmean^2
> ccdf<- cumsum(theta.sim<1.75)/ (1:nsim)
> cq<-NULL
> for(j in 1:nsim){ cq<-c(cq,quantile(theta.sim[1:j],probs=0.975)) }
> sseq<- c(1,(1:100)*(nsim/100))
> cmean<-cmean[sseq]
> cq<-cq[sseq]
> ccdf<-ccdf[sseq]
> plot(sseq,cmean,type="l",xlab="# of Monte Carlo samples",ylab="cumulative mean",
+ col="black")
> abline(h= (a+sy)/(b+n),col="gray",lwd=2)
> plot(sseq,ccdf,type="l",xlab="# of Monte Carlo samples",ylab="cumulative cdf at 1.75",col="black")
> abline(h= pgamma(1.75,a+sy,b+n),col="gray",lwd=2)
Page 6 of 22
Bayesian Statistics
3470: 589-001
> plot(sseq,cq,type="l",xlab="# of Monte Carlo samples",ylab="cumulative 97.5% quantile",col="black")

> abline(h= qgamma(.975,a+sy,b+n),col="gray",lwd=2)
> set.seed(1)
> a<-1 ; b<-1
> theta.prior.sim<-rbeta(10000,a,b)
> gamma.prior.sim<- log( theta.prior.sim/(1-theta.prior.sim) )
> n0<-860-441 ; n1<-441
> theta.post.sim<-rbeta(10000,a+n1,b+n0)
> gamma.post.sim<- log( theta.post.sim/(1-theta.post.sim) )
> #pdf("fig4_3.pdf",family="Times",height=3.5,width=7)
> #par(mar=c(3,3,1,1),mgp=c(1.75,.75,0))
> par(mfrow=c(2,3))
> par(cex=.8)
> par(mfrow=c(1,2),mar=c(3,3,1,1), mgp=c(1.75,.75,.0))
> plot(density(gamma.prior.sim,adj=2),xlim=c(-5,5),main="", xlab=expression(gamma),
+ ylab=expression(italic(p(gamma))),col="gray")
> plot(density(gamma.post.sim,adj=2),xlim=c(-5,5),main="",xlab=expression(gamma),
+ # ylab=expression(italic(paste("p(",gamma,"|",y[1],"...",y[n],")"))))
+ ylab=expression(paste(italic("p("),gamma,"|",y[1],"...",y[n],")",
+ sep="")) )
> lines(density(gamma.prior.sim,adj=2),col="gray")
Page 7 of 22
Bayesian Statistics
3470: 589-001
> set.seed(1)
> a<-2 ; b<-1
> sy1<-217 ; n1<-111
> sy2<-66 ; n2<-44
> theta1.mc<-rgamma(10000,a+sy1, b+n1)
> theta2.mc<-rgamma(10000,a+sy2, b+n2)
> y1.mc<-rpois(10000,theta1.mc)
> y2.mc<-rpois(10000,theta2.mc)
> mean(theta1.mc>theta2.mc)
[1] 0.9708
> mean(y1.mc>y2.mc)
[1] 0.4846
> #par(mar=c(3,3,1,1),mgp=c(1.75,.75,0))
> par(mfrow=c(1,1))
> plot(density(theta1.mc/theta2.mc,adj=2),main="",xlim=c(.75,2.25),
+ xlab=expression(gamma==theta[1]/theta[2]),
+ #ylab=expression(italic(paste("p(",gamma,"|",bold(y[1]),",",bold(y[2]),")",
+ # sep="")) ))
+ ylab=expression(paste(italic("p("),gamma,"|",bold(y[1]),",",bold(y[2]),")",
+ sep="")) )
> #par(mar=c(3,3,1,1),mgp=c(1.75,.75,0))
> par(mfrow=c(1,1))
Page 8 of 22
Bayesian Statistics
3470: 589-001
> diff.mc<-y1.mc-y2.mc
> ds<- -11:11
> plot(ds,(table(c(diff.mc,ds))-1)/length(diff), type="h",lwd=3,
+ xlab=expression(italic(D==tilde(Y)[1]-tilde(Y)[2])),
+ ylab=expression(paste(italic("p(D"),"|",bold(y[1]),",",bold(y[2]),")",sep="")))
Problem 2.8 (Bayesian data analysis)
1 n
2
180+ 2 150
40 20 1
a) |y N ( 1 n
,
1 n
)
+ +
402 202 402 202
1 n
2
180+ 2 150
40 20 1
b)  ~y |y N ( , +202 )
1 n 1 n
2
+ 2 2
+ 2
40 20 40 20
c) 95% posterior interval for θ | y = 150, n = 10: 150.7 ± 1.96(6.25) = [138, 163]
95% posterior interval for ˜y | y = 150, n = 10: 150.7 ± 1.96(20.95) = [110, 192]
d) 95% posterior interval for θ | y = 150, n = 100: [146, 154]
Page 9 of 22
Bayesian Statistics
3470: 589-001
95% posterior interval for ˜y | y = 150, n = 100: [111, 189]
a)
> theta <- seq(0,1,.001)

> dens <- dbeta(theta,1,.67)
> plot (theta, dens, xlim=c(0,1), ylim=c(0,3),
+ type="l", xlab="theta", ylab="", xaxs="i",
+ yaxs="i", yaxt="n", bty="n", cex=2)
> lines (c(1,1),c(0,3),col=0)
> lines (c(1,1),c(0,3),lty=3)
The density blows up at = 1 but has a finite integral.
b) n = 1000, y = 650, Posterior distribution is p (θ|y) = Beta (α+650, b+350) = Beta (651, 350.67).
The data dominate the prior distribution. E (θ|y) = 0.6499, sd (θ |y) = 0.015
> theta <- seq(0,1,.001)

> dens <- dbeta(theta,651,350.67)
> cond <- dens/max(dens) > 0.001
> plot (theta[cond], dens[cond],
+ type="l", xlab="theta", ylab="", xaxs="i",
+ yaxs="i", yaxt="n", bty="n", cex=2)
>
2.10
Page 10 of 22
Bayesian Statistics
3470: 589-001
a)
b)
Page 11 of 22
Bayesian Statistics
3470: 589-001
c)
Page 12 of 22
Bayesian Statistics
3470: 589-001
Page 13 of 22
Bayesian Statistics
3470: 589-001
Page 14 of 22
Bayesian Statistics
3470: 589-001
Problem 3.3 (Introduction Monte Carlo Methods with R)
> h=function(x){ 1/(x^2*sqrt(2*pi)*exp(1/(2*x^2)))}

> par(mfrow=c(2,1))
> curve(h,from=0,to=1/20,xlab="x",ylab="h(x)",lwd="2")
> I=1/20*h(runif(10^4)/20)
Page 15 of 22
Bayesian Statistics
3470: 589-001
> estint=cumsum(I)/(1:10^4)
> esterr=sqrt(cumsum((I-estint)^2))/(1:10^4)
> plot(estint,xlab="Iterations",ty="l",lwd=2,
+ ylim=mean(I)+20*c(-esterr[10^4],esterr[10^4]),ylab="")
> lines(estint+2*esterr,col="gold",lwd=2)
> lines(estint-2*esterr,col="gold",lwd=2)
> integrate(h,0,1/20)
2.759158e-89 with absolute error < 5.4e-89
> pnorm(-20)
[1] 2.753624e-89
>
Page 16 of 22
Bayesian Statistics
3470: 589-001
b)
> h=function(x){ exp(-(x-3)^2/2)+exp(-(x-6)^2/2)}

> smpl=matrix(h(rnorm(100*10^3)),ncol=100)
> cvrg=apply(smpl,2,cumsum)/(1:10^3)
> matplot(cvrg,type="l",col="grey56",lty=1,xlab="",ylab="")
> abline(h=0.07461577,lty=2,col="sienna")
c)
> h=function(t){
+ x=tan(pi*(t+4.5)/7)
+ dnorm(x)*pi*(exp(-.5*(x-3)^2)+exp(-.5*(x-6)^2))/cos(t)^2}
> smpl=matrix(h(runif(100*10^3,min=-8,max=-1)),ncol=100)
Page 17 of 22
Bayesian Statistics
3470: 589-001
> matplot(cvrg,type="l",lty=1,xlab="",ylab="",ylim=c(0,.15))
> h=function(x){ 7*dnorm(x)*(exp(-(x-3)^2/2)+exp(-(x-6)^2/2))}

> smpl=matrix(h(runif(100*10^3,min=-8,max=-1)),ncol=100)
> matplot(cvrg,type="l",col="grey56",lty=1,xlab="",ylab="")
Page 18 of 22
Bayesian Statistics
3470: 589-001
Solution:
> h=function(x){ exp(-x)*sqrt(x)/gamma(3/2)}

> X = rexp(10^4,1) + 12.5
> I=exp(-12.5)*sqrt(X)/gamma(3/2)
Page 19 of 22
Bayesian Statistics
3470: 589-001
> plot(estint,xlab="Iterations",ty="l",lwd=2,
> integrate(h,12.5,Inf)
1.544033e-05 with absolute error < 3.4e-06
> pchisq(25,3,low=F)
[1] 1.544050e-05
> h=function(x){ 1/sqrt(5*pi)*gamma(3)/gamma(2.5)*1/(1+x^2/5)^3}

> integrate(h,50,Inf)
3.023564e-08 with absolute error < 2e-08
> X = rexp(10^4,1) + 50
> I=1/sqrt(5*pi)*gamma(3)/gamma(2.5)*1/(1+X^2/5)^3*1/exp(-X+50)
> plot(estint,xlab="Mean and error range",type="l",lwd=2,
Page 20 of 22
Bayesian Statistics
3470: 589-001
As seen on the graph, this method induces jumps in the convergence patterns. Those jumps are indicative
of variance problems, as should be since the estimator does not have a finite variance in this case. The
value returned by this approach differs from alternatives evaluations:
> mean(I)
[1] 1.529655e-08
> sd(I)/10^2
[1] 9.328338e-10
> integrate(h,50,Inf)
3.023564e-08 with absolute error < 2e-08
> pt(50,5,low=F)
[1] 3.023879e-08
Page 21 of 22
Bayesian Statistics
3470: 589-001
c)
> x=rexp(100)
> mean(dnorm(x+4.5)*exp(x))
[1] 2.817864e-06
> var(dnorm(x+4.5)*exp(x))/100
[1] 1.544983e-13
Page 22 of 22

T Statistics HW-5 (Arefin-2824475)

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

T Statistics HW-5 (Arefin-2824475)

Uploaded by

Copyright:

Available Formats

Bayesian Statistics

Figure 3.4 (A first course in Bayesian methods)

Figure 3.6 (A first course in Bayesian methods)

Figure 3.8 (A first course in Bayesian methods)

Figure 3.10 (A first course in Bayesian methods)

Figure 4.2 (A first course in Bayesian methods)

> plot(sseq,cq,type="l",xlab="# of Monte Carlo samples",ylab="cumulative 97.5% quantile",col="black")

Figure 4.3 (A first course in Bayesian methods)

Figure 4.5 (A first course in Bayesian methods)

Problem 2.8 (Bayesian data analysis)

d) 95% posterior interval for θ | y = 150, n = 100: [146, 154]

95% posterior interval for ˜y | y = 150, n = 100: [111, 189]

Problem 2.9 (Bayesian data analysis)

> theta <- seq(0,1,.001)

The density blows up at = 1 but has a finite integral.

> theta <- seq(0,1,.001)

Problem 2.10 (Bayesian data analysis)

Problem 2.13 (Bayesian data analysis)

Problem 3.3 (Introduction Monte Carlo Methods with R)

> h=function(x){ 1/(x^2*sqrt(2*pi)*exp(1/(2*x^2)))}

Problem 3.4 (Introduction Monte Carlo Methods with R)

> h=function(x){ exp(-(x-3)^2/2)+exp(-(x-6)^2/2)}

> h=function(x){ 7*dnorm(x)*(exp(-(x-3)^2/2)+exp(-(x-6)^2/2))}

Problem 3.10 (Introduction Monte Carlo Methods with R)

Problem 3.11 (Introduction Monte Carlo Methods with R)

> h=function(x){ exp(-x)*sqrt(x)/gamma(3/2)}

> h=function(x){ 1/sqrt(5*pi)*gamma(3)/gamma(2.5)*1/(1+x^2/5)^3}

You might also like

> h=function(x){ 1/(x^2sqrt(2pi)exp(1/(2x^2)))}

> h=function(x){ 7dnorm(x)(exp(-(x-3)^2/2)+exp(-(x-6)^2/2))}

> h=function(x){ 1/sqrt(5pi)gamma(3)/gamma(2.5)*1/(1+x^2/5)^3}