View source for R1

[[category:Pragmatic Programming]]
'''Open Source Statistics with R'''

=Introduction=

R is a mature, open-source (i.e. free!) statistics package, with an intuitive interface, excellent graphics and a vibrant community constantly adding new methods for the statistical investigation of your data to the library of packages available.

The goal of this tutorial is to introduce you to the R package, and not to be an introductory course in statistics.

Some excellent examples of using R can also be found at: http://msenux.redwoods.edu/math/R/ and http://www.r-tutor.com/

If you are working on a Linux system, you will typically start R from the command line.  On a Windows machine, or a Mac, you will typically start up R in some form of GUI.  However you get R started, you will have access to an R command prompt.  The good news is that the examples below will all work at the R command prompt, however you gained access to it.

=Getting Started=

The very simplest thing we can do with R is to perform some arithmetic at the command prompt:

<source>
> phi <- (1+sqrt(5))/2
> phi
[1] 1.618034
</source>

Parentheses are used to modify the usual order of precedence of the operators ('''/''' will typically be evaluated before '''+''').  Note the '''[1]''' accompanying the returned value.  All numbers entered at the console are interpreted as a vector.  The '[1]' indicates that the line in question is displaying the vector of values starting at first index.  We can use the handy sequence function to create a vector containing more than a single element:

<source>
> odds <- seq(from=1, to=67, by=2)
> odds
 [1]  1  3  5  7  9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49
[26] 51 53 55 57 59 61 63 65 67
</source>

From the above example, we can see that both the '''<-''' and '''=''' operators can be used for assignment.

Vectors are commonly used data structures in R:

<source>
coords.bris <- c(51.5, 2.6)
</source>

As are matrices:

<source>
> magic <- matrix(data=c(2,7,6,9,5,1,4,3,8),nrow=3,ncol=3)
> magic
     [,1] [,2] [,3]
[1,]    2    9    4
[2,]    7    5    3
[3,]    6    1    8
</source>

Where the '''c''' function combines the arguments given in the parentheses.  We can access portions of the array using the syntax shown in the square brackets.  For example, we can access the first row using the '''[1,]''' notation, and similarly the second column using '''[,2]'''.  Since the square is 3x3 magic, the numbers in both slices should sum to 15:

<source>
> sum(magic[1,])
[1] 15
> sum(magic[,2])
[1] 15
</source>

Single elements and ranges can also accessed:

<source>
> magic[2,2]
[1] 5
> magic[2:3,2:3]
     [,1] [,2]
[1,]    5    3
[2,]    1    8
</source>

R also provides '''arrays''', which have more than two dimensions, and '''lists''' to hold heterogeneous collections.

A very commonly used data structure is the '''data frame''', which R uses to store tabular data.  Given several vectors of equal length, we can collate them into a data frame:

<source>
> country <- c("USA", "China", "GB")
> gold <- c(46, 38, 29)
> silver <- c(29, 27, 17)
> bronze <- c(29, 23, 19)
> medals.2012 <- data.frame(country, gold, silver, bronze)
> medals.2012
    country  gold  silver  bronze
1       USA    46      29      29
2     China    38      27      23
3        GB    29      17      19
</source>

=Standard Graphics: A taster=

An aspect which makes R popular are it's graphing functions.  R also has some very handy built-in data sets--we'll use this to demonstrate just a small fraction of R's graphing abilities.

First up is the humble '''plot()''' function.  Given a data frame of points, such as one charting the relationship between temperature and the vapour pressure of mercury, it will give us a (handily labelled) scatter plot: 

<source>
> plot(pressure)
</source>

See the gallery below for all the plots created in this section.

The plot function will also accept a time-series (another class of object recognised by R) and will sensibly join the points with a line:

<source>
> plot(co2)
> class(co2)
[1] "ts"
</source>

Pie charts are easily constructed.  In this case, to show the relative proportions of electricity generated from different sources in the UK in 2011 (source: https://www.gov.uk/government/.../5942-uk-energy-in-brief-2012.pdf‎):

<source>
> uk.electricty.sources.2011 <- c(41,29,18,5,4,2,1)
> names(uk.electricty.sources.2011) <- ("Gas", "Coal", "Nuclear", "Hydro & other", "Wind", "Imports", "Oil")
> pie(uk.electricty.sources.2011, main="UK Electricty Generating Mix, 2011", col=rainbow(7))
</source>

Next, let's create a bar chart of monthly average precipitation falling here in the fair city of Bristol (source: http://www.worldweatheronline.com):

<source>
> bristol.precip <- c(82.9, 56.1, 59.2, 69, 50.8, 50.9, 50.8, 74.8, 74.7, 91.1, 94.5, 93.6)
> names(bristol.precip) <- c("Jan", "Feb", "Mar", "Apr", "May", "Jun", "Jul", "Aug", "Sep", "Oct", "Nov", "Dec")
> barplot(bristol.precip,
+ main="Average Monthly Precipitation in Bristol",
+ ylab="Mean precipitation (mm)",
+ ylim=c(0,100),
+ col=c("darkblue"))
</source>

[http://en.wikipedia.org/wiki/Box_plot 'Box and whisker' plots] are useful ways to graph the quartiles of some data.  In this case, the fuel efficiencies of various US cars, circa 1974:

<source>
> boxplot(mpg~cyl,data=mtcars, main="Car Milage Data",
+    xlab="Number of Cylinders", ylab="Miles Per Gallon")
</source>

R includes a very useful help facility.  In the case of the '''filled.contour()''' plotting function, the help page includes an example of it's use to plot the topology of a volcano in Auckland, NZ:

<source>
> ?filled.countour
</source>

<gallery widths=300px heights=300px perrow=3>
File:Vapour-pressure.png|Vapour pressure of mercury against temperature
File:Mauna-loa.png|CO2 concentrations measured at Mauna-Loa between 1959 and 1997
File:Pie.png|The UK's electricity generating mix, 2011
File:Barplot.png|Average monthly precipitation in Bristol
File:Boxplot.png|Range of fuel efficiencies for different engine sizes 
File:Maunga-Whau.png|Topology of Maunga Whau volcano in Auckland
</gallery>

There are many more example plots--complete with the R code required to create the plots (at the bottom of the page, after the comments)--on the following web page:
* http://gallery.r-enthusiasts.com/thumbs.php

=Functions=

You can define your own functions in R, using the '''function''' keyword.  For example, Pythagoras' Theorem:

<source>
> hypotenuse <- function(x, y) {sqrt(x^2 + y^2)}
</source>

The braces ({}) are optional, but add clarity.

To call the function:

<source>
> hypotenuse(3,4)
[1] 5
</source>

We can provide default values for the arguments, which can be overridden for any given invocation of the function:

<source>
> hypot2 <- function(x=3 ,y=4) {sqrt(x^2 + y^2)}
> hypot2()
[1] 5
> hypot2(x=12, y=16)
[1] 20
</source>

=Packages=

Listed at http://cran.r-project.org/

Let's install the '''multicore''' package, that will give us access to functions within R which will run on the multiple processors which we often find in our computers these days:   

<source>
> install.packages("multicore")
</source>

Et voila!  It is done.

We can check which packages are currently loaded into the library available from our workspace:

<source>
> library()
</source>

If we need to add one, we type e.g.:

<source>
> library(multicore)
</source>

Now, an example of using a function from the multicore package.  The '''lapply''' function, which is included in the standard R core, will map a given function over a list inputs, giving a list of the function outputs in return.  For example, we can map a squaring function over the list of integers from 1 to 3:

<source>
> lapply(1:3, function(x) x^2)
</source>

which gives us the list:

<pre>
[[1]]
[1] 1

[[2]]
[1] 4

[[3]]
[1] 9
</pre>

Now, we can do the same work in parallel using:

<source>
> mclapply(1:3, function(x) x^2)
</source>

=Reading Data from File=

R provides some very useful functions for reading and writing data from/to file.

==Text Files==

Let's start with text files.  If your data is organised into a file such that it looks like a table with column headings:  


Perhaps the simplest one is '''read.table()'''.  If I have a text file with the following contents:

<pre>
country              gold silver bronze
"USA"                46   29     29
"China"              38   27     23
"Great Britain"      29   17     19
"Russian Federation" 24   26     32
"Republic of Korea"  13   8      7
"Germany"            11   19     14
</pre>

It will be a simple matter to use the '''read.table()''' function to load the data into R:

<source>
> medals.2012 <- read.table("medals.txt", header=TRUE)
> medals.2012
             country gold silver bronze
1                USA   46     29     29
2              China   38     27     23
3      Great Britain   29     17     19
4 Russian Federation   24     26     32
5  Republic of Korea   13      8      7
6            Germany   11     19     14
</source>

There is a corresponding '''write.table()''' function to export the contents of a data frame into a text file.

CSV files can be easily handled by specifying '''sep=","''' as an argument to read.table().  However, for convenience, there are also '''read.csv()''' and '''write.csv()''' functions defined.  For example:

<source>
> write.csv(medals.2012,"medals.csv")
</source>

Gives us the file, '''medals.csv''', with the contents:

<pre>
"","country","gold","silver","bronze"
"1","USA",46,29,29
"2","China",38,27,23
"3","Great Britain",29,17,19
"4","Russian Federation",24,26,32
"5","Republic of Korea",13,8,7
"6","Germany",11,19,14
</pre>

==Binary Files==

The '''save()''' function will store an R data structure in binary form:

<source>
> save(medals.2012,file="medals.RData")
</source>

<pre>
gethin@gethin-desktop:~$ file medals.RData 
medals.RData: gzip compressed data, from Unix
</pre>

There is, of course, a corresponding function to load such data:

<source>
> load("medals.RData")
</source>

==Databases==

If you would like to read and write data directly from/to a database, there are several packages to help you.  See http://cran.r-project.org/doc/manuals/r-release/R-data.html#Relational-databases for more information.

==NetCDF==

The [http://cran.r-project.org/web/packages/ncdf/index.html '''ncdf''' package] provides an interface to NetCDF files.  Before installing the package, you will need the Unidata NetCDF libraries installed on your system.  On Linux, the standard package managers conveniently provide this.  Note that you will need the 'development' packages.  Once the prerequisites are satisfied, you can use the standard R command to install the package from CRAN:

<source>
> install.packages("ncdf")
</source>

=Examples of Common Tasks=

==Linear Regression==

<source>
> plot(cars)
> res=lm(dist ~ speed, data=cars)
> abline(res)
</source>

[[Image:R-lm(cars)-abline.png|400px|thumbnail|center|linear regression of stopping distance against speed from the built-in data set, cars]]

----

'''Exercise'''
* Weighted least squares.  The '''lm''' function will accept a vector of weights, '''lm(... weights=...)'''.  If given, the function will optimise the line of best fit according a the equation of weighted least squares.  Experiment with different linear model fits, given different weighting vectors.  Some handy hints for creating a vector of weights:
** '''w1<-rep(0.1,50)''' will give you a vector, length 50, where each element has a value of 0.1.  W1[1]<-10 will give the first element of the vector a value of 10.
** '''w2<-seq(from=0.02, to=1.0, by=0.02)''' provides a vector containing a sequence of values from 0.02 to 1.0 in steps of 0.02 (handily, again 50 in total).

==Significance Testing==

<source>
> boys_2=c(90.2, 91.4, 86.4, 87.6, 86.7, 88.1, 82.2, 83.8, 91, 87.4)
> girls_2=c(83.8, 86.2, 85.1, 88.6, 83, 88.9, 89.7, 81.3, 88.7, 88.4)
> res=var.test(boys_2,girls_2)
> res

	F test to compare two variances

data:  boys_2 and girls_2 
F = 1.0186, num df = 9, denom df = 9, p-value = 0.9786
alternative hypothesis: true ratio of variances is not equal to 1 
95 percent confidence interval:
 0.2529956 4.1007126 
sample estimates:
ratio of variances 
          1.018559 
> res=t.test(boys_2, girls_2, var.equal=TRUE, paired=FALSE)
> res

	Two Sample t-test

data:  boys_2 and girls_2 
t = 0.8429, df = 18, p-value = 0.4103
alternative hypothesis: true difference in means is not equal to 0 
95 percent confidence interval:
 -1.656675  3.876675 
sample estimates:
mean of x mean of y 
    87.48     86.3
</source>

=R and HPC=

* http://cran.r-project.org/web/views/HighPerformanceComputing.html

==Multicore==

mutlicore.r:
<source>
library(multicore)
# how many cores are present?
multicore:::detectCores()
# Create a 10 x 10,000 matrix of random numbers
test <- lapply(1:10,function(x) rnorm(10000))
# Map a function over the matrix.  First in serial..
system.time(x <- lapply(test,function(x) loess.smooth(x,x)))
# .. and secondly in parallel (using multicore, within a node)
system.time(x <- mclapply(test,function(x) loess.smooth(x,x)))
</source>

Submission script:
<source>
#!/bin/bash 

#PBS -l nodes=1:ppn=8,walltime=00:00:05

#! Ensure that we have the correct version of R loaded
module add languages/R-2.15.1

#! change the working directory (default is home directory)
cd $PBS_O_WORKDIR

#! Run the R script
R CMD BATCH multicore.r
</source>


Output:
<pre>
> library(multicore)
> # how many cores are present?
> multicore:::detectCores()
[1] 8
> # Create a 10 x 10,000 matrix of random numbers
> test <- lapply(1:10,function(x) rnorm(10000))
> # Map a function over the matrix.  First in serial..
> system.time(x <- lapply(test,function(x) loess.smooth(x,x)))
   user  system elapsed 
  0.651   0.003   0.654 
> # .. and secondly in parallel (using multicore, within a node)
> system.time(x <- mclapply(test,function(x) loess.smooth(x,x)))
   user  system elapsed 
  0.293   0.063   0.163
</pre>

==Rmpi==

Rmpi.r:
<source>
library(Rmpi)
# spawn as many slaves as possible
mpi.spawn.Rslaves()
mpi.remote.exec(mpi.get.processor.name())
mpi.remote.exec(runif(1))
mpi.close.Rslaves()
mpi.quit()
</source>

Submission script:
<source>
#!/bin/bash 

#PBS -l nodes=4:ppn=1,walltime=00:00:05

#! Ensure that we have the correct version of R loaded
module add languages/R-2.15.1

#! change the working directory (default is home directory)
cd $PBS_O_WORKDIR

#! Create a machine file (used for multi-node jobs)
cat $PBS_NODEFILE > machine.file.$PBS_JOBID

#! Run the R script
mpirun -np 1 -machinefile machine.file.$PBS_JOBID R CMD BATCH Rmpi.r
</source>

output:
<pre>
> library(Rmpi)
> # spawn as many slaves as possible
> mpi.spawn.Rslaves()
        4 slaves are spawned successfully. 0 failed.
master (rank 0, comm 1) of size 5 is running on: u03n074 
slave1 (rank 1, comm 1) of size 5 is running on: u03n098 
slave2 (rank 2, comm 1) of size 5 is running on: u04n029 
slave3 (rank 3, comm 1) of size 5 is running on: u04n030 
slave4 (rank 4, comm 1) of size 5 is running on: u03n074 
> mpi.remote.exec(mpi.get.processor.name())
$slave1
[1] "u03n098"

$slave2
[1] "u04n029"

$slave3
[1] "u04n030"

$slave4
[1] "u03n074"

> mpi.remote.exec(runif(1))
         X1        X2        X3        X4
1 0.5154871 0.5154871 0.5154871 0.5154871
> mpi.close.Rslaves()
[1] 1
> mpi.quit()
</pre>

==Snow==

snow.r:
<source>
library(snow)
# request a cluster of 3 worker nodes
cl <- makeCluster(3)
clusterCall(cl, function() Sys.info()[c("nodename","machine")])
clusterCall(cl, runif, 1)
stopCluster(cl)
</source>

==Parallel==

parallel.r:
<source>
library(parallel)
# how many cores are present?
parallel:::detectCores()
# Create a 10 x 10,000 matrix of random numbers
test <- lapply(1:10,function(x) rnorm(10000))
# Map a function over the matrix.  First in serial..
system.time(x <- lapply(test,function(x) loess.smooth(x,x)))
# .. and secondly in parallel (using multicore, within a node)
system.time(x <- mclapply(test,function(x) loess.smooth(x,x)))
</source>

I have not been able to get a distributed memory cluster working on BCp2 using the parallel package.

=Further Reading=

* [http://shop.oreilly.com/product/9780596801717.do| R in a Nutshell]