Last updated: 2019-05-05

Checks: 5 1

Knit directory: drift-workflow/analysis/

This reproducible R Markdown analysis was created with workflowr (version 1.2.0). The Report tab describes the reproducibility checks that were applied when the results were created. The Past versions tab lists the development history.


The R Markdown is untracked by Git. To know which version of the R Markdown file created these results, you’ll want to first commit it to the Git repo. If you’re still working on the analysis, you can ignore this warning. When you’re finished, you can run wflow_publish to commit the R Markdown file and build the HTML.

Great job! The global environment was empty. Objects defined in the global environment can affect the analysis in your R Markdown file in unknown ways. For reproduciblity it’s best to always run the code in an empty environment.

The command set.seed(20190211) was run prior to running the code in the R Markdown file. Setting a seed ensures that any results that rely on randomness, e.g. subsampling or permutations, are reproducible.

Great job! Recording the operating system, R version, and package versions is critical for reproducibility.

Nice! There were no cached chunks for this analysis, so you can be confident that you successfully produced the results during this run.

Great! You are using Git for version control. Tracking code development and connecting the code version to the results is critical for reproducibility. The version displayed above was the version of the Git repository at the time these results were generated.

Note that you need to be careful to ensure that all relevant files for the analysis have been committed to Git prior to generating the results (you can use wflow_publish or wflow_git_commit). workflowr only checks the R Markdown file, but you know if there are other scripts or data files that it depends on. Below is the status of the Git repository when the results were generated:


Ignored files:
    Ignored:    .Rhistory
    Ignored:    analysis/.Rhistory
    Ignored:    analysis/flash_cache/
    Ignored:    data.tar.gz
    Ignored:    data/datasets/
    Ignored:    data/raw/
    Ignored:    output.tar.gz
    Ignored:    output/

Untracked files:
    Untracked:  analysis/ebnm_bimodal.Rmd
    Untracked:  code/ebnm_bimodal.R
    Untracked:  docs/figure/ebnm_bimodal.Rmd/

Unstaged changes:
    Modified:   analysis/index.Rmd
    Modified:   analysis/simpler_tree_simulation.Rmd

Note that any generated files, e.g. HTML, png, CSS, etc., are not included in this status report because it is ok for generated content to have uncommitted changes.


There are no past versions. Publish this analysis with wflow_publish() to start tracking its development.


Here I explore the idea of “hacking” ashr to solve the Empirical Bayes Normal Means problem with a bimodal prior, specifically with the modes of the prior at 0 and 1. The idea is we’d like to “penalize” against estimating intermediate effects i.e. we shrink the effects to 1 if there large enough and 0 if their small enough, accounting for the precision of the estimate and learning the right level to shrink.

Imports

library(ggplot2)
library(dplyr)
library(tidyr)
source("../code/ebnm_bimodal.R")

Functions

Here are some helper function for simulation, fitting, and plotting.

sim = function(n0, n1, sigma_e){
  n = n0+n1
  beta = c(rep(0, n0), rep(1, n1))
  s = abs(rnorm(n, 0, sigma_e))
  betahat = rnorm(n, beta, s)
  
  return(list(betahat=betahat, s=s, beta=beta, n=n))
}


fit = function(betahat, s, beta){
  
  n = length(betahat)
  fit_res = ebnm_bimodal(betahat, s, list())
  betapm = fit_res$postmean
  df = data.frame(betahat=betahat, beta=beta, betapm=betapm, s=s, idx=1:n)  

  return(df)
}

plot_sim = function(df, title){
  
  gath_df = df %>% gather(variable, value, -idx, -s)
  p0 = ggplot(gath_df, aes(x=idx, y=value, 
                        color=factor(variable, levels=c("beta", "betahat", "betapm")))) + 
      geom_point() + 
      theme_bw() +
      labs(color="") +
      xlab("Variable") + 
      ylab("Value") +
      theme(legend.position="bottom")
  
  min_betahat = min(df$betahat)
  max_betahat = max(df$betahat)
  p1 = ggplot(df, aes(betahat, betapm, color=s)) + 
       geom_point() + viridis::scale_color_viridis() + 
       theme_bw() + 
       theme(legend.position="bottom") +
       xlim(c(min_betahat, max_betahat)) +
       ylim(c(min_betahat, max_betahat)) + 
       geom_abline() 

  p = cowplot::plot_grid(p0, p1, nrow=1) 
  title = cowplot::ggdraw() + cowplot::draw_label(title)
  print(cowplot::plot_grid(title, p, ncol=1, rel_heights=c(0.1, 1)))

}

Approach

The approach I took was to setup two grids:

  1. A “positive” mixture of uniforms ash prior with mode 0 and grange from 0 to 1
  2. A “negative” mixture of uniforms ash prior with mode 1 and grange from 0 to 1

I adapted code from ashr to extract the grids in ../code/ebnm_bimodal.R. Below is a sanity check that my helper function extracts the “correct” grids:

# simulate data
sim_res = sim(40, 40, .1)
betahat = sim_res$betahat
s = sim_res$s
data = ashr:::set_data(as.vector(betahat), as.vector(s), ashr:::add_etruncFUN(ashr::lik_normal()), 0)

# get grid for mode 0 (my helper function)
grid0 = get_bimodal_grid(data, 0, "+uniform")

# get grid for mode 0 (ashr)
res = ashr::ash(betahat, s, mixcompdist="+uniform", mode=0, grange=c(0, 1))
res$fitted_g$a == grid0$a
 [1] TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE
[15] TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE
[29] TRUE TRUE TRUE
res$fitted_g$b == grid0$b
 [1] TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE
[15] TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE
[29] TRUE TRUE TRUE
# get grid for mode 1 (my helper function)
grid1 = get_bimodal_grid(data, 1, "-uniform")

# get grid for mode 1 (ashr)
res = ashr::ash(betahat, s, mixcompdist="-uniform", mode=1, grange=c(0, 1))
res$fitted_g$a == grid1$a
 [1] TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE
[15] TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE
[29] TRUE TRUE TRUE
res$fitted_g$b == grid1$b
 [1] TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE
[15] TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE
[29] TRUE TRUE TRUE

It looks like the grids match the ashr implementation. I next wrote an additional helper function that estimates the prior \(g\) using the the grid of “intervals” defined above (see get_bimodal_g in ../code/ebnm_bimodal.R):

# setup bimodal grid
a = c(grid0$a, grid1$a)
b = c(grid0$b, grid1$b)

# estimate prior
ghat = get_bimodal_g(data, a, b)
idx_c = which(ghat$pi > 1e-5)
print(a[idx_c])
[1] 0.0000000 1.0000000 0.9724091
print(b[idx_c])
[1] 0 1 1

We can see for this example only a few components get weight i.e. those with mass close to 0 and 1. Finally given this \(g\) I wrote a ebnm function that computes the posterior with the fixed \(g\) (see ebnm_bimodal in ../code/ebnm_bimodal.R):

df = fit(betahat, s, sim_res$beta)
plot_sim(df, "")

Pretty cool! Effects close to 0 are shrunk to 0 while effects close to 1 are shrunk to 1.

Simulations

Next I simulated a bunch of normal means scenarios where the true \(\beta\)s are set to 0 or 1. In each simulation I specify the number of zeros n0 the number of ones n1 and standard deviation used to simulate std. errors.

n0 = c(rep(40, 3), rep(25, 3), rep(10, 3), rep(0, 3))
n1 = c(rep(40, 3), rep(55, 3), rep(70, 3), rep(80, 3))
sigma_e = rep(c(.05, .1, .25), 4)

for(i in 1:length(n0)){
  sim_res = sim(n0[i], n1[i], sigma_e[i])
  betahat = sim_res$betahat
  s = sim_res$s
  beta = sim_res$beta
  df = fit(betahat, s, beta)
  title = paste0("n0=",n0[i], ",n1=", n1[i], ",sigma_e=", sigma_e[i])
  plot_sim(df, title)
}

I think the idea roughly works! The most interesting scenarios to compare are when the std. errors of the estmates are high but the number of zeros and ones are different. Maybe we can define a term “bimodality” which I’m thinking is how bimodal the distribution is. When the bimodality is low (i.e. the prior distribution is closer to unimodal) the effects seem to be more correctly estimated. As we can see estimating more bimodal effects is a more difficult problem than unimodal effects.

I’d like to think more about the setting the prior grid but this was a relatively easily implementable first pass. It would also be interesting to think more about how to weight the prior mixture proportions if that would be helpful.

sessionInfo()
R version 3.5.1 (2018-07-02)
Platform: x86_64-apple-darwin13.4.0 (64-bit)
Running under: macOS  10.14.2

Matrix products: default
BLAS/LAPACK: /Users/jhmarcus/miniconda3/lib/R/lib/libRblas.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] tidyr_0.8.2   dplyr_0.8.0.1 ggplot2_3.1.0

loaded via a namespace (and not attached):
 [1] Rcpp_1.0.0        compiler_3.5.1    pillar_1.3.1     
 [4] git2r_0.23.0      plyr_1.8.4        workflowr_1.2.0  
 [7] viridis_0.5.1     iterators_1.0.10  tools_3.5.1      
[10] digest_0.6.18     viridisLite_0.3.0 evaluate_0.12    
[13] tibble_2.0.1      gtable_0.2.0      lattice_0.20-38  
[16] pkgconfig_2.0.2   rlang_0.3.1       foreach_1.4.4    
[19] Matrix_1.2-15     parallel_3.5.1    yaml_2.2.0       
[22] xfun_0.4          gridExtra_2.3     withr_2.1.2      
[25] stringr_1.4.0     knitr_1.21        fs_1.2.6         
[28] cowplot_0.9.4     rprojroot_1.3-2   grid_3.5.1       
[31] tidyselect_0.2.5  glue_1.3.0        R6_2.4.0         
[34] rmarkdown_1.11    mixsqp_0.1-115    purrr_0.3.0      
[37] ashr_2.2-37       magrittr_1.5      MASS_7.3-51.1    
[40] codetools_0.2-16  backports_1.1.3   scales_1.0.0     
[43] htmltools_0.3.6   assertthat_0.2.0  colorspace_1.4-0 
[46] labeling_0.3      stringi_1.2.4     pscl_1.5.2       
[49] doParallel_1.0.14 lazyeval_0.2.1    munsell_0.5.0    
[52] truncnorm_1.0-8   SQUAREM_2017.10-1 crayon_1.3.4