Robust P-Value Combination Methods.
metapro
Introduction
The meta-analysis is conducted to increase the statistical power by combining evidences (e.g., effect sizes and p-values) obtained from multiple experiments. The p-value combination has been widely used for meta-analysis when effect sizes are not available. metapro is a CRAN R package that provides functions for p-value combinations. There are four functions including
- wFisher : weighted Fisher's method
- wZ : weighted Z-method
- lancaster : Lancaster's procedure
The original Fisher's method and (weighted) Z-method have been commonly used for p-value combination. In particular, the Z-method is effective when a feature is associated with most of the cohorts. However, this attributes is disadvantageous detecting partially associated patterns (e.g., detection of African-specific features in trans-ethnic analysis).Therefore, in this study, we designed the ordmeta and wFisher method that work effectively in those situations. The detailed formulas are described in our paper.
Installation
Prerequisites
- JAVA : If not installed yet, please install it first. Download: https://www.java.com/en/
- Rtools (for Windows): Rtools are required for Windows users. Download: https://cran.r-project.org/bin/windows/Rtools/
After all prerequisites are installed, open R and install the metapro package by typing
install.packages('devtools') # install 'devtools'
library(devtools)
install_github('unistbig/metapro', INSTALL_opts=c("--no-multiarch"))
library(metapro) # Load metapro package
or
install.packages('metapro', INSTALL_opts=c("--no-multiarch"))
library(metapro)
Tip: Dealing with Errors
In most cases, installation errors come from 1) encoding problem and 2) JAVA environment variable settings.
So, please try followings.<p>
Encoding error
On Windows,
- System locale setting modification: Click Start -> Control panel-> Clock and Region -> Region -> Administrative -> Change system locale... -> set 'Current system locale' as English (United States) and check the box in the below (Beta: Use Unicode UTF-8 for worldwide language support).
- Reboot the PC, and execute Rstudio.
- In the R console, type
Sys.setlocale('LC_ALL','C')
and install the package again.
Error from JAVA setting
Most error will occur with rJAVA. In this case, please
- Check whether both JAVA and R are 64-bit, and
- Set proper environment variable for JAVA_HOME, CLASSPATH, and RPATH.
- Tip: System environment variable can be easily set by using Set.setenv() function in R. For example,
Sys.setenv(JAVA_HOME='C:\\Program Files\\Java\\jdk1.8.0_211")
Sys.setenv(CLASSPATH="C:\\Program Files\\Java\\jdk1.8.0_211\\jre\\lib\\ext")
Sys.setenv(RPATH="C:\\Program Files\\R\\R-3.6.0\\bin\\x64")
If you are Linux or Mac OSX user, type following on terminal and re-open the R console after setting the environment variables.
sudo R CMD javareconf
Next, check whether rJAVA is installed correctly, and try installing metapro again.
install.packages('rJava')
library(rJava)
Usage
1. wFisher
wFisher is designed to assign weights to each experiment based on the sample size.
Example
wFisher(p=c(0.01, 0.02, 0.8, 0.25), weight = c(200, 500, 100, 80), is.onetail=FALSE, eff.sign = c(1,1,1,-1))
will return
$p
[1] 0.003060523
$overall.eff.direction
[1] "+"
Input Arguments and Values
Arguments | |
---|---|
p | A numeric vector of p-values |
is.onetail | Logical. If set TRUE, p-values are combined without considering the direction of effect, and vice versa. Default: TRUE. |
weight | A numeric vector of weight or sample size for each experiment. Note! If no weight option is given, the original Fisher method is performed. |
eff.sign | A vector of signs of effects. It works when is.onetail = FALSE |
Value | |
---|---|
p | Combined p-value |
overall.eff.direction | The direction of combined effects. |
2. wZ
Weighted Z-method. This function has been modified from sumz function in metap package.
Example
wZ(p=c(0.01, 0.02, 0.8, 0.25), weight = c(200, 500, 100, 80), is.onetail=FALSE, eff.sign = c(1,1,1,-1))
will return
[1] 0.001798156
$overall.eff.direction
[1] "+"
$sumz
[1] 2.911558
Input Arguments and Values
Arguments | |
---|---|
p | A numeric vector of p-values |
is.onetail | Logical. If set TRUE, p-values are combined without considering the direction of effect, and vice versa. Default: TRUE. |
weight | A numeric vector of weight or sample size for each experiment. Note! If no weight option is given, the Stouffer's method is performed. |
eff.sign | A vector of signs of effects. It works when is.onetail = FALSE |
Value | |
---|---|
p | Combined p-value |
overall.eff.direction | The direction of combined effects. |
sumz | Transformed sum of z-values |
3. lancaster
Lancaster's procedure - the generalized version Fisher's method
Example
lancaster(p=c(0.01, 0.02, 0.8, 0.25), weight = c(200, 500, 100, 80), is.onetail=FALSE, eff.sign = c(1,1,1,-1))
will return
$p
[1] 0.005694935
$overall.eff.direction
[1] "+"
Input Arguments and Values
Arguments | |
---|---|
p | A numeric vector of p-values |
is.onetail | Logical. If set TRUE, p-values are combined without considering the direction of effect, and vice versa. Default: TRUE. |
weight | A numeric vector of weight or sample size for each experiment. Required! |
eff.sign | A vector of signs of effects. It works when is.onetail = FALSE |
Value | |
---|---|
p | Combined p-value |
overall.eff.direction | The direction of combined effects. |
Contact
Sora Yoon: [email protected]
Department of Biological Sciences, UNIST.