Last updated: 2018-05-12

workflowr checks: (Click a bullet for more information)
  • R Markdown file: up-to-date

    Great! Since the R Markdown file has been committed to the Git repository, you know the exact version of the code that produced these results.

  • Environment: empty

    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.

  • Seed: set.seed(12345)

    The command set.seed(12345) 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.

  • Session information: recorded

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

  • Repository version: ddf9062

    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:    .DS_Store
        Ignored:    .Rhistory
        Ignored:    .Rproj.user/
        Ignored:    analysis/.DS_Store
        Ignored:    analysis/BH_robustness_cache/
        Ignored:    analysis/FDR_Null_cache/
        Ignored:    analysis/FDR_null_betahat_cache/
        Ignored:    analysis/Rmosek_cache/
        Ignored:    analysis/StepDown_cache/
        Ignored:    analysis/alternative2_cache/
        Ignored:    analysis/alternative_cache/
        Ignored:    analysis/ash_gd_cache/
        Ignored:    analysis/average_cor_gtex_2_cache/
        Ignored:    analysis/average_cor_gtex_cache/
        Ignored:    analysis/brca_cache/
        Ignored:    analysis/cash_deconv_cache/
        Ignored:    analysis/cash_fdr_1_cache/
        Ignored:    analysis/cash_fdr_2_cache/
        Ignored:    analysis/cash_fdr_3_cache/
        Ignored:    analysis/cash_fdr_4_cache/
        Ignored:    analysis/cash_fdr_5_cache/
        Ignored:    analysis/cash_fdr_6_cache/
        Ignored:    analysis/cash_plots_cache/
        Ignored:    analysis/cash_sim_1_cache/
        Ignored:    analysis/cash_sim_2_cache/
        Ignored:    analysis/cash_sim_3_cache/
        Ignored:    analysis/cash_sim_4_cache/
        Ignored:    analysis/cash_sim_5_cache/
        Ignored:    analysis/cash_sim_6_cache/
        Ignored:    analysis/cash_sim_7_cache/
        Ignored:    analysis/correlated_z_2_cache/
        Ignored:    analysis/correlated_z_3_cache/
        Ignored:    analysis/correlated_z_cache/
        Ignored:    analysis/create_null_cache/
        Ignored:    analysis/cutoff_null_cache/
        Ignored:    analysis/design_matrix_2_cache/
        Ignored:    analysis/design_matrix_cache/
        Ignored:    analysis/diagnostic_ash_cache/
        Ignored:    analysis/diagnostic_correlated_z_2_cache/
        Ignored:    analysis/diagnostic_correlated_z_3_cache/
        Ignored:    analysis/diagnostic_correlated_z_cache/
        Ignored:    analysis/diagnostic_plot_2_cache/
        Ignored:    analysis/diagnostic_plot_cache/
        Ignored:    analysis/efron_leukemia_cache/
        Ignored:    analysis/fitting_normal_cache/
        Ignored:    analysis/gaussian_derivatives_2_cache/
        Ignored:    analysis/gaussian_derivatives_3_cache/
        Ignored:    analysis/gaussian_derivatives_4_cache/
        Ignored:    analysis/gaussian_derivatives_5_cache/
        Ignored:    analysis/gaussian_derivatives_cache/
        Ignored:    analysis/gd-ash_cache/
        Ignored:    analysis/gd_delta_cache/
        Ignored:    analysis/gd_lik_2_cache/
        Ignored:    analysis/gd_lik_cache/
        Ignored:    analysis/gd_w_cache/
        Ignored:    analysis/knockoff_10_cache/
        Ignored:    analysis/knockoff_2_cache/
        Ignored:    analysis/knockoff_3_cache/
        Ignored:    analysis/knockoff_4_cache/
        Ignored:    analysis/knockoff_5_cache/
        Ignored:    analysis/knockoff_6_cache/
        Ignored:    analysis/knockoff_7_cache/
        Ignored:    analysis/knockoff_8_cache/
        Ignored:    analysis/knockoff_9_cache/
        Ignored:    analysis/knockoff_cache/
        Ignored:    analysis/knockoff_var_cache/
        Ignored:    analysis/marginal_z_alternative_cache/
        Ignored:    analysis/marginal_z_cache/
        Ignored:    analysis/mosek_reg_2_cache/
        Ignored:    analysis/mosek_reg_4_cache/
        Ignored:    analysis/mosek_reg_5_cache/
        Ignored:    analysis/mosek_reg_6_cache/
        Ignored:    analysis/mosek_reg_cache/
        Ignored:    analysis/pihat0_null_cache/
        Ignored:    analysis/plot_diagnostic_cache/
        Ignored:    analysis/poster_obayes17_cache/
        Ignored:    analysis/real_data_simulation_2_cache/
        Ignored:    analysis/real_data_simulation_3_cache/
        Ignored:    analysis/real_data_simulation_4_cache/
        Ignored:    analysis/real_data_simulation_5_cache/
        Ignored:    analysis/real_data_simulation_cache/
        Ignored:    analysis/rmosek_primal_dual_2_cache/
        Ignored:    analysis/rmosek_primal_dual_cache/
        Ignored:    analysis/seqgendiff_cache/
        Ignored:    analysis/simulated_correlated_null_2_cache/
        Ignored:    analysis/simulated_correlated_null_3_cache/
        Ignored:    analysis/simulated_correlated_null_cache/
        Ignored:    analysis/simulation_real_se_2_cache/
        Ignored:    analysis/simulation_real_se_cache/
        Ignored:    analysis/smemo_2_cache/
        Ignored:    data/LSI/
        Ignored:    docs/.DS_Store
        Ignored:    docs/figure/.DS_Store
        Ignored:    output/fig/
    
    Unstaged changes:
        Deleted:    analysis/cash_plots_fdp.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.
Expand here to see past versions:
    File Version Author Date Message
    rmd cc0ab83 Lei Sun 2018-05-11 update
    html 0f36d99 LSun 2017-12-21 Build site.
    html 853a484 LSun 2017-11-07 Build site.
    html 1ea081a LSun 2017-07-03 sites
    rmd b7aaa28 LSun 2017-03-29 debug
    html b7aaa28 LSun 2017-03-29 debug
    rmd c803b43 LSun 2017-03-27 fitting and write-up
    html c803b43 LSun 2017-03-27 fitting and write-up

Problem setting

With all the assumptions, we formulate a convex optimization as follows.

\[ \begin{array}{rl} \max\limits_{w_1, \ldots, w_K} & \sum\limits_{i = 1}^n\log\left(1 +\sum\limits_{k = 1}^K w_kh_k(z_i)\right)\\ \text{s.t.} & 1 + \sum\limits_{k = 1}^K w_kh_k(z_i) \geq0 \end{array} \] It can also be written as

\[ \begin{array}{rl} \max\limits_{w} & \sum\log\left(1 +Hw\right)\\ \text{s.t.} & 1 +Hw \geq0 \end{array} \]

where \(H_{ik} = h_k(z_i)\).

Choosing \(K\)

With finite \(K\) Gaussian derivatives, the fitted log-likelihood \(\log\prod\limits_{i = 1}^nf_0(z_i) = \sum\limits_{i = 1}^n\log \varphi(z_i) + \sum\limits_{i = 1}^n\log\left(1 +\sum\limits_{k = 1}^K \hat w_kh_k(z_i)\right)\). Note that \(\hat w_k \equiv 0\) is a feasible solution, so the optimal solution will always have a log-likelihood no less than \(\sum\limits_{i = 1}^n\log \varphi(z_i)\), the log-likelihood of \(N(0, 1)\). Similarly, let \(\hat w_K\) be the optimal solution with \(K\) Gaussian derivatives, then \([\hat w_K, 0]\) will be a feasible solution with \(K + 1\) Gaussian derivatives, with the same objective value, which should be no larger than the optimal objective value for \(K + 1\). This fact implies that the fitted log-likelihood should be non-decreasing with respect to \(K\).

This property implies that for the same data set, multiple models with increasing \(K\)’s can be fitted, and the searching stops at a sufficiently large \(K\). Let \(\hat g_K = \max\limits_w\sum\limits_{i = 1}^n\log\left(1 +\sum\limits_{k = 1}^K w_kh_k(z_i)\right) = \sum\limits_{i = 1}^n\log\left(1 +\sum\limits_{k = 1}^K \hat w_kh_k(z_i)\right)\), the optimal objective value for \(K\), drawing on insight from the likelihood ratio test, the optimal \(\hat K = \inf\left\{K: 2(\hat g_{K + m} - \hat g_K) \leq \chi_{m, 1-\alpha}^2, m = 1, \ldots, M\right\}\) with pre-specified \(\alpha\) and \(M\).

\(M\) should be greater than \(1\) because odd order Gaussian derivatives and even order ones have distinct properties, especially in lower orders. Odd order Gaussian derivatives are odd functions, more associated with the skewness of the empirical distribution, whereas even order ones are even functions, more associated with its kurtosis. Therefore, it’s not at all impossible that the difference in the log-likelihood objective between \(K\) and \(K + 1\) is small, yet that between \(K\) and \(K + 2\) is large.

The function ecdfz.optimal in the script ecdfz.R is using this rule with the default setting \(\alpha = 0.05, M = 2\).

Fitting with cvxr

The script ecdfz.R is using cvxr to find the maximum likelihood estimate of the observed density of correlated null \(z\) scores.

source("../code/ecdfz.R")
z = read.table("../output/z_null_liver_777.txt")
p = read.table("../output/p_null_liver_777.txt")

Examples

Several selected ash-hostile and/or BH-hostile data sets are fitted as follows. More detailed information of these selected data sets are here and here.

For each of these selected data sets, we plot the histogram, the density of \(N(0, 1)\) in red line, and that fitted by Gaussian derivatives in blue. We also give the information of the number of false discoveries by Benjamini-Hochberg, and \(\hat\pi_0\) estimated by ash. Note that BH’s false discoveries suggests the inflation of the most extreme observations, whereas ash’s \(\hat\pi_0\) the empirical distribution’s general deviation from \(N(0, 1)\).

We also plot the optimal objective \(\hat g_K\) for all the fitted \(K\), and indicate the optimal \(\hat K\) found according to the aforementioned rule. It appears the second order derivative is usually the most important, followed by the fourth.

library(ashr)
DataSet = c(32, 327, 355, 483, 778)
res_DataSet = list()
for (i in 1:length(DataSet)) {
  zscore = as.numeric(z[DataSet[i], ])
  fit.ecdfz = ecdfz.optimal(zscore)
  fit.ash = ash(zscore, 1, method = "fdr")
  fit.ash.pi0 = get_pi0(fit.ash)
  pvalue = as.numeric(p[DataSet[i], ])
  fd.bh = sum(p.adjust(pvalue, method = "BH") <= 0.05)
  res_DataSet[[i]] = list(DataSet = DataSet[i], fit.ecdfz = fit.ecdfz, fit.ash = fit.ash, fit.ash.pi0 = fit.ash.pi0, fd.bh = fd.bh, zscore = zscore, pvalue = pvalue)
}
Data Set 32 : Number of BH's False Discoveries: 0 ; ASH's pihat0 = 0.184423 ; Chosen number of Gaussian derivatives K = 4 
Optimal weights of Gaussian derivatives w =
1 : -0.0365352616939748 ; 2 : 0.1999283988026 ; 3 : 0.0104709711917952 ; 4 : -0.0201268873245717 ;

Data Set 327 : Number of BH's False Discoveries: 489 ; ASH's pihat0 = 0.1522796 ; Chosen number of Gaussian derivatives K = 9 
Optimal weights of Gaussian derivatives w =
1 : 0.0339429040246409 ; 2 : 0.73456039139193 ; 3 : -0.153251399156623 ; 4 : 0.188391233293932 ; 5 : -0.0550489496726788 ; 6 : 0.0229717983852805 ; 7 : -0.00690839817896393 ; 8 : 0.00121590052122215 ; 9 : -0.000314826111681546 ;

Data Set 355 : Number of BH's False Discoveries: 639 ; ASH's pihat0 = 0.04750946 ; Chosen number of Gaussian derivatives K = 9 
Optimal weights of Gaussian derivatives w =
1 : 0.0226254633548931 ; 2 : 0.921396261289049 ; 3 : 0.0218042845603059 ; 4 : 0.17609829567166 ; 5 : -0.0137967389059908 ; 6 : 0.00396812175677823 ; 7 : -0.00490495150100676 ; 8 : -0.000608489154565857 ; 9 : -0.000298931322154747 ;

Data Set 483 : Number of BH's False Discoveries: 1 ; ASH's pihat0 = 0.9998824 ; Chosen number of Gaussian derivatives K = 4 
Optimal weights of Gaussian derivatives w =
1 : 0.045465618357444 ; 2 : -0.127037015386377 ; 3 : 0.00972556942208551 ; 4 : 0.0105165371329065 ;

Data Set 778 : Number of BH's False Discoveries: 1 ; ASH's pihat0 = 0.07619716 ; Chosen number of Gaussian derivatives K = 4 
Optimal weights of Gaussian derivatives w =
1 : 0.00594319611757542 ; 2 : 0.398070908599464 ; 3 : -0.00922245793412396 ; 4 : 0.026306961083798 ;

Conclusion

Gaussian derivatives can deal with a variety of empirical null distribution, especially those not close to normal.

Session information

sessionInfo()
R version 3.4.3 (2017-11-30)
Platform: x86_64-apple-darwin15.6.0 (64-bit)
Running under: macOS High Sierra 10.13.4

Matrix products: default
BLAS: /Library/Frameworks/R.framework/Versions/3.4/Resources/lib/libRblas.0.dylib
LAPACK: /Library/Frameworks/R.framework/Versions/3.4/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] CVXR_0.95     EQL_1.0-0     ttutils_1.0-1

loaded via a namespace (and not attached):
 [1] Rcpp_0.12.16      knitr_1.20        whisker_0.3-2    
 [4] magrittr_1.5      workflowr_1.0.1   bit_1.1-12       
 [7] lattice_0.20-35   R6_2.2.2          stringr_1.3.0    
[10] tools_3.4.3       grid_3.4.3        R.oo_1.21.0      
[13] git2r_0.21.0      scs_1.1-1         htmltools_0.3.6  
[16] bit64_0.9-7       yaml_2.1.18       rprojroot_1.3-2  
[19] digest_0.6.15     gmp_0.5-13.1      Matrix_1.2-12    
[22] ECOSolveR_0.4     R.utils_2.6.0     evaluate_0.10.1  
[25] rmarkdown_1.9     stringi_1.1.6     Rmpfr_0.6-1      
[28] compiler_3.4.3    backports_1.1.2   R.methodsS3_1.7.1



This reproducible R Markdown analysis was created with workflowr 1.0.1