Accounting for correlations among measurements
In mashr: Multivariate Adaptive Shrinkage

knitr::opts_chunk$set(echo = TRUE,comment = "#",fig.width = 5,
                      fig.height = 4,fig.align = "center",
                      eval = TRUE)

Introduction

In some settings measurements and tests in different conditions may be correlated with one another. For example, in eQTL applications this can occur due to sample overlap among the different conditions.

Failure to deal with such correlations can cause false positives in a mashr analysis.

To deal with these correlations mashr allows the user to specify a correlation matrix $V$ when setting up the data in mash_set_data. We introduce two methods to estimate this correlation matrix. The first method is simple and fast. It estimates the correlation matrix using estimate_null_correlation_simple, which, as its name suggests, uses the null tests (specifically, tests without a strong $z$ score) to estimate the correlations. The second method may provide a better mash fit. It estimates the correlations using mash_estimate_corr_em, which uses an ad hoc EM algorithm.

Method 1

The method is described in Urbut et al.

Here we simulate data with correlations.

library(mashr)
set.seed(1)
simdata = simple_sims(500,5,1)
V = matrix(0.5,5,5)
diag(V) = 1
simdata$Bhat = simdata$B + mvtnorm::rmvnorm(2000, sigma = V)

Read in the data, and estimate correlations:

data   = mash_set_data(simdata$Bhat, simdata$Shat)
V.simple = estimate_null_correlation_simple(data)
data.Vsimple = mash_update_data(data, V=V.simple)

Now we have two mash data objects, one (data.Vsimple) with correlations specified, and one without (data). So analyses using data.Vsimple will allow for correlations, whereas analyses using data will assume measurements are independent.

Here, for illustration purposes, we proceed to analyze the data with correlations, using just the simple canonical covariances as in the initial introductory vignette.

U.c = cov_canonical(data.Vsimple) 
m.Vsimple = mash(data.Vsimple, U.c) # fits with correlations because data.V includes correlation information 
print(get_loglik(m.Vsimple),digits=10) # log-likelihood of the fit with correlations set to V

We can also compare with the original analysis. (Note that the canonical covariances do not depend on the correlations, so we can use the same U.c here for both analyses. If we used data-driven covariances we might prefer to estimate these separately for each analysis as the correlations would affect them.)

m.orig = mash(data, U.c) # fits without correlations because data object was set up without correlations
print(get_loglik(m.orig),digits=10)

loglik = c(get_loglik(m.orig), get_loglik(m.Vsimple))
significant = c(length(get_significant_results(m.orig)), length(get_significant_results(m.Vsimple)))
false_positive = c(sum(get_significant_results(m.orig) < 501), 
                   sum(get_significant_results(m.Vsimple) < 501))
tb = rbind(loglik, significant, false_positive)
colnames(tb) = c('without cor', 'V simple')
row.names(tb) = c('log likelihood', '# significance', '# False positive')
tb

The log-likelihood with correlations is higher than without correlations. The false positives reduce.

Method 2

The method is described in Yuxin Zou's thesis.

To estimate the residual correlations using EM method, it requires covariance matrices for the signals. We proceed with the simple canonical covariances.

With details = TRUE in mash_estimate_corr_em, it returns the estimates residual correlation matrix with the mash fit.

V.em = mash_estimate_corr_em(data, U.c, details = TRUE)
m.Vem = V.em$mash.model
print(get_loglik(m.Vem),digits=10) # log-likelihood of the fit

loglik = c(get_loglik(m.orig), get_loglik(m.Vsimple), get_loglik(m.Vem))
significant = c(length(get_significant_results(m.orig)), length(get_significant_results(m.Vsimple)),
                length(get_significant_results(m.Vem)))
false_positive = c(sum(get_significant_results(m.orig) < 501), 
                   sum(get_significant_results(m.Vsimple) < 501),
                   sum(get_significant_results(m.Vem) < 501))
tb = rbind(loglik, significant, false_positive)
colnames(tb) = c('without cor', 'V simple', 'V EM')
row.names(tb) = c('log likelihood', '# significance', '# False positive')
tb

Comparing with Method 1, the log likelihood from Method 2 is higher.

The EM updates in mash_estimate_corr_em needs some time to converge. There are several things we can do to reduce the running time. First of all, we can set the number of iterations to a small number. Because there is a large improvement in the log-likelihood within the first few iterations, running the algorithm with small number of iterations provides estimates of correlation matrix that is better than the initial value. Moreover, we can estimate the correlation matrix using a random subset of genes, not the whole observed genes.