Introduction

Compare 1. smashgen-ash.identity 2. smashgen-ash.log 3. smashgen-ash.identity.zero 4. smashgen-ash.log.zero (known and unknown nugget effect) with smash-anscombe.

Settings: spike mean function, mean function range (0.1,6) and (20,50)

Note:

1. idk: expand around ash posterior mean using identity link in lik_pois; given nugget effect(use sigma=sqrt(nugget^2+s^2) in smash.gaus).
2. idu: same as 1; unknown nugget effect(use sigma=NULL in smash.gaus).
3. id0k: expand around ash posterior mean using identity link only for 0 \(x\)s and around \(x\) for nonzero \(x\)s; given nugget effect.
4. id0u: same as 3; unkown nugget effect
5. logk: expand around ash posterior mean using log link in lik_pois; given nugget effect
6. logu: same as 5; unkown nuggect effect
7. log0k: expand around ash posterior mean using log link only for 0 \(x\)s and around \(x\) for nonzero \(x\)s; given nugget effect.
8. logu: same as 7; unkown nugget effect
9. ans: use anscombe transormation on poisson data, variance around 1/4; use 0 variance for 0 \(x\)s.

vst_smooth=function(x,method,ep=1e-5){
  n=length(x)
  if(method=='sr'){
    x.t=sqrt(x)
    x.var=rep(1/4,n)
    x.var[x==0]=0
    mu.hat=(smashr::smash.gaus(x.t,sigma=sqrt(x.var)))^2
    
  }
  if(method=='anscombe'){
    x.t=sqrt(x+3/8)
    x.var=rep(1/4,n)
    x.var[x==0]=0
    mu.hat=(smashr::smash.gaus(x.t,sigma=sqrt(x.var)))^2-3/8
  }
  if(method=='log'){
    x.t=x
    x.t[x==0]=ep
    x.var=1/x.t
    x.t=log(x.t)
    mu.hat=exp(smashr::smash.gaus(x.t,sigma=sqrt(x.var)))
  }
  return(mu.hat)
}

smash_gen_all=function(x,sigma,method){
  n=length(x)
  if(method=='identity'){
    x.ash=ash(rep(0,n),1,lik=lik_pois(x,link='identity'))$result$PosteriorMean
  }
  if(method=='log'){
    x.ash=ash(rep(0,n),1,lik=lik_pois(x,link='log'))$result$PosteriorMean
  }
  if(method=='identity.zero'){
    x.ash=ash(rep(0,n),1,lik=lik_pois(x,link='identity'))$result$PosteriorMean
    x.ash[x!=0]=x[x!=0]
  }
  if(method=='log.zero'){
    x.ash=ash(rep(0,n),1,lik=lik_pois(x,link='log'))$result$PosteriorMean
    x.ash[x!=0]=x[x!=0]
  }
  y=log(x.ash)+(x-x.ash)/x.ash
  s2=1/x.ash
  mu.sigk=exp(smash.gaus(y,sigma=sqrt(sigma^2+s2)))
  mu.sigu=exp(smash.gaus(y))
  return(list(mu.sigk=mu.sigk,mu.sigu=mu.sigu))
}

simu_study=function(m,sigma=0,nsimu=100,seed=12345){
  set.seed(12345)
  idk=c()
  idu=c()
  id0k=c()
  id0u=c()
  logk=c()
  logu=c()
  log0k=c()
  log0u=c()
  ans=c()
  for (i in 1:nsimu) {
    lambda=exp(log(m)+rnorm(n,0,sigma))
    x=rpois(n,lambda)
    id=smash_gen_all(x,sigma,'identity')
    id0=smash_gen_all(x,sigma,'identity.zero')
    logg=smash_gen_all(x,sigma,'log')
    log0=smash_gen_all(x,sigma,'log.zero')
    
    idk=rbind(idk,id$mu.sigk)
    idu=rbind(idu,id$mu.sigu)
    id0k=rbind(id0k,id0$mu.sigk)
    id0u=rbind(id0u,id0$mu.sigu)
    logk=rbind(logk,logg$mu.sigk)
    logu=rbind(logu,logg$mu.sigu)
    log0k=rbind(log0k,log0$mu.sigk)
    log0u=rbind(log0u,log0$mu.sigu)
    ans=rbind(ans,vst_smooth(x,'anscombe'))
  }
  return(list(idk=idk,idu=idu,id0k=id0k,id0u=id0u,logk=logk,logu=logu,log0k=log0k,log0u=log0u,ans=ans))
}

First we compare all the methods mentioned above using spike mean function whose mean range is around (0.1,6) so there are a number of zero counts in the sequence. This would be a challenge for smashgen since we are using log transformation.

library(ashr)
library(smashr)

spike.f = function(x) (0.75 * exp(-500 * (x - 0.23)^2) + 1.5 * exp(-2000 * (x - 0.33)^2) + 3 * exp(-8000 * (x - 0.47)^2) + 2.25 * exp(-16000 * 
    (x - 0.69)^2) + 0.5 * exp(-32000 * (x - 0.83)^2))
n = 512
t = 1:n/n
m = spike.f(t)

m=m*2+0.1
range(m)

[1] 0.100000 6.076316

result=simu_study(m,sigma=0,nsimu = 30)

mses=lapply(result, function(x){apply(x, 1, function(y){mean((y-m)^2)})})

unlist(lapply(mses, mean))

         idk          idu         id0k         id0u         logk 
6.805100e+05 7.167519e+04 3.837905e-01 2.060559e-01 4.102731e+51 
        logu        log0k        log0u          ans 
1.042183e+32 1.171031e+00 5.780713e-01 1.192800e-01 

boxplot(mses[-c(1,2,5,6)],main='nugget=0',ylab='MSE')

When there is no nugget effect, id0k, id0u, log0u have relatively smaller mean square error(mse) while anscombe transformation outperforms all smashgen methods and achieves smaller mse.

We plot the estimated mean function of id0u, log0u and ans for comparison. log0u seems to result in underestimations of mean function. id0u overestimates small means and underestimates large means. So when there are a number of zero observations, it’s very crucial to choose where to expand for 0 \(x\)s.

par(mfrow=c(2,2))

for (j  in c(1,2,3,4)) {
  plot(m,type='l',main='nugget=0')
  lines(result$id0u[j,],col=2)
  lines(result$log0u[j,],col=3)
  lines(result$ans[j,],col=4)
  legend('topleft',c('mean','ash identity link','ash log link','anscombe'),lty=c(1,1,1,1),col=c(1,2,3,4))
}

Now we increase nugget effect to \(\sigma=1\). Obviously, using anscombe transformation, we are estimating \(exp(\log(\mu)+N(0,\sigma^2))\) so its mse is large and gives spiky fit.

result=simu_study(m,sigma=1,nsimu = 30)

mses=lapply(result, function(x){apply(x, 1, function(y){mean((y-m)^2)})})

unlist(lapply(mses, mean))

         idk          idu         id0k         id0u         logk 
2.345193e+36 2.710025e+35 6.004389e-01 3.633385e-01 2.281759e+48 
        logu        log0k        log0u          ans 
2.059211e+10 1.167105e+00 7.626772e-01 4.053297e+00 

boxplot(mses[-c(1,2,5,6)],main='nugget=1',ylab='MSE')

boxplot(mses[-c(1,2,5,6,9)],main='nugget=1',ylab='MSE')

par(mfrow=c(2,2))

for (j  in c(1,2,3,4)) {
  plot(m,type='l',main='nugget=1')
  lines(result$id0u[j,],col=2)
  lines(result$log0u[j,],col=3)
  lines(result$ans[j,],col=4)
  legend('topleft',c('mean','ash identity link','ash log link','anscombe'),lty=c(1,1,1,1),col=c(1,2,3,4))
}

How about a larger mean function? Increase the range to (20,50). Some observations from the plot: 1. Now, known nugget effect gives smaller mse than unkown ones(e.g idk

Plots compare idk, logk and ans:

par(mfrow=c(2,2))

for (j  in c(1,2,3,4)) {
  plot(m,type='l',main='nugget=0')
  lines(result$idk[j,],col=2)
  lines(result$logk[j,],col=3)
  lines(result$ans[j,],col=4)
  legend('topleft',c('mean','ash identity link','ash log link','anscombe'),lty=c(1,1,1,1),col=c(1,2,3,4))
}

Summary

Maybe can develop a version of anscombe to deal with nugget effect? Also is nugget effect necessarily defined as \(exp(\log(\mu)+\sigma^2)\)?

sessionInfo()

R version 3.5.1 (2018-07-02)
Platform: x86_64-apple-darwin15.6.0 (64-bit)
Running under: macOS High Sierra 10.13.6

Matrix products: default
BLAS: /Library/Frameworks/R.framework/Versions/3.5/Resources/lib/libRblas.0.dylib
LAPACK: /Library/Frameworks/R.framework/Versions/3.5/Resources/lib/libRlapack.dylib

locale:
[1] en_US.UTF-8/en_US.UTF-8/en_US.UTF-8/C/en_US.UTF-8/en_US.UTF-8

attached base packages:
[1] stats     graphics  grDevices utils     datasets  methods   base     

other attached packages:
[1] smashr_1.2-0 ashr_2.2-7  

loaded via a namespace (and not attached):
 [1] Rcpp_0.12.18      knitr_1.20        whisker_0.3-2    
 [4] magrittr_1.5      workflowr_1.1.1   REBayes_1.3      
 [7] MASS_7.3-50       pscl_1.5.2        doParallel_1.0.14
[10] SQUAREM_2017.10-1 lattice_0.20-35   foreach_1.4.4    
[13] stringr_1.3.1     caTools_1.17.1.1  tools_3.5.1      
[16] parallel_3.5.1    grid_3.5.1        data.table_1.11.6
[19] R.oo_1.22.0       git2r_0.23.0      htmltools_0.3.6  
[22] iterators_1.0.10  assertthat_0.2.0  yaml_2.2.0       
[25] rprojroot_1.3-2   digest_0.6.17     Matrix_1.2-14    
[28] bitops_1.0-6      codetools_0.2-15  R.utils_2.7.0    
[31] evaluate_0.11     rmarkdown_1.10    wavethresh_4.6.8 
[34] stringi_1.2.4     compiler_3.5.1    Rmosek_8.0.69    
[37] backports_1.1.2   R.methodsS3_1.7.1 truncnorm_1.0-8