Fells Stats Know Your Data


Great Infographic

This is a really great exposition on an
infographic. Note that the design elements and "chart junk" serve
to better connect and communicate the data to the viewer. The
choice not to go with pie charts for the first set of plots is a
good one. The drawbacks of polar representations of proportions is
very well know to statisticians, but designers often feel that
circular elements are more pleasing in their infographics. This
bias toward circular representations is sometimes justifiable.
After all, if your graphic is pretty then no one will take the time
to look at it.
Take a look:


Filed under: Uncategorized No Comments

R for Dummies

The book R for Dummies was released recently, and was just reviewed by Dirk Eddelbuettel in the Journal of Statistical Software.

Dirk is an R luminary, creating such fantastic works as Rcpp. R for Dummies seems to have beaten Dirk's natural disinclination to like anything with "for Dummies" appended to it, receiving a pretty positive review. Here is the last bit:

"R for Dummies may well break new ground, and introduce R to new audiences. The pricing is aggressive and should help the book to find its way into the hands of a large number of students, data analysis practitioners as well as researchers. They will find a well-written and easy-to-read introduction to the language and environment - if they can overcome any initial bias against a Dummies title as this reviewer did."

I haven't had a chance to read the text, but I have worked (albeit briefly) with Andie de Vries who is one of the authors. He is the author of several R packages up on CRAN, is a regular contributor to the R tag on Stack Overflow, and has a nuanced understanding of the language.

Filed under: Uncategorized No Comments

Eclipse + Rcpp + RInside = Magic


I've been doing R/Java development for some time, creating packages both large and small with this tool chain. I have used Eclipse as my package development environment almost exclusively, but I didn't realize how much I relied on the IDE before I had to do some serious R/C++ package development. My first Rcpp package (wordcloud) only used a little bit of complied code, so just using a standard text editor was enough to get the job done, however when I went on to a more complex project, TextWrangler just wasn't cutting it anymore. Where was my code completion? Where were my automatically detected syntax errors?

I found myself spending more time looking up class APIs, fixing stupid errors, and waiting for R CMD INSTALL to finish than I did coding. Unfortunately the internet did not come to the rescue. I searched and searched but couldn't find anyone who used a modern IDE for package development with compiled code. In particular I wanted the following features.

  • Code completion. What methods does that class have again? What arguments does that function take? Perhaps everyone is just smarter than me, but I can't seem to keep the whole R/Rcpp API in my head. I just want to hit ctrl+space and see what my options are:

  • Syntax error detection. Oh, I meant hasAttribute. Whoops, forgot that semicolon again.

  • Incremental builds. It takes a long time for R CMD INSTALL to complete. The package that I am currently working on takes 2 minutes and 30 seconds to install. I don't want to have to wait for a full install every time I make a small change to a function I am playing with.
  • R and C++ syntax highlighting. Editing good looking code makes for a better coding experience.

The solution I found is a combination of Eclipse, Rcpp and RInside. Rcpp is undeniably the best way to embed high performance C++ code within R. It makes writing C++ code with native R objects a breeze. Even John Chambers is unhappy with the lack of elegance in the R/C interface. Rcpp uses some of the more advanced language features in C++ to abstract away all of the ugliness. RInside on the other hand embeds R within a C++ program.

The steps below will show you how to link up Eclipse with Rcpp to get code completion and syntax error detection, then RInside will be used to link up R with Eclipses build and run systems allowing for incremental building and rapid prototyping.



Step 1: Download Eclipse


You will need to download the CDT version of eclipse for c/c++ development. It can be obtained at http://www.eclipse.org/cdt/.


Step 2: Install statet


statet is an eclipse plug-in supporting R integration. Follow the installation instructions at http://www.walware.de/goto/statet. R will need to be configured, which is covered in section 3 of Longhow Lam's user guide.


Step 3: Install Rcpp and Rinside


In R run:



Step 4: Create a skeleton for a new C++ project


  • Right click on the Navigator area and select New -> C++ Project.

  • Name the project MyCppPackage and click Finish.

  • In R create a cpp package skeleton with the following code:


  • In the file system navigate to your R working directory. There you should find a new folder labeled MyCppPackage. Drag the contents of that folder into your project in eclipse.

Step 5: Making the package installable


Open the external tools configurations and create a new statet R CMD Tools configuration for R CMD INSTALL. Change the directory from resource_loc to project_loc. Note, you will probably want to also add the command line argument "--clean" so that temporary files created during the install process are removed. If you want the fastest possible installation, use "--no-test-load --no-multiarch --clean".

Your shiny new Rcpp project should now be installable with the click of a button. Run the configuration (you should an install log in the console) and check that the package is now loadable by running library(MyRcppPackage) in R.

Step 6: Setting up links and properties


Now eclipse knows how to install the package, but it doesn't know what any of the symbols in your C++ files mean because it can't see the Rcpp package. If you look at the rcpp_hello_world.cpp file you should see a bunch of bugs.
In this step we will set up the project properties necessary for Eclipse to link up with R, Rcpp and RInside. There may be a few things that you need to change for your system. Notably, the directories listed below all start with /Library/Frameworks/R.framework/Resources. This is the default R_HOME for Mac OS X. If your R home is located in a different place, or you have a user library directory where Rcpp and RInside are installed, you will need to edit the directories accordingly. Also, the architecture for my R is x86_64, so if your architecture is different, you will need to change this.
  • Right click on MyCppPackage and select properties.
  • Add references to the header files for R, Rcpp and RInside to the G++ compiler includes.
(note: these may be different on your system depending on your R_HOME and library paths)
(edit: I've removed the include directive for the /Rcpp/include/Rcpp directory as it is not needed and can lead to conflicting references for string.h on case-insensitive operating systems).

  • Add a reference to the R library to G++ linker - Libraries.
(note: if you are not using the x86_64 arch, replace this with the appropriate architecture.)
  • Add linkages for the RInside library to the G++ linker - Miscellaneous

(edit: I've removed the reference to Rcpp/lib/x86_64/libRcpp.a as Rcpp is header only (as of 0.11.0))

  • Add the "-arch x86_64" architecture flag for g++ compiler. This option is found in the Miscellaneous settings. If your computer's architecture is different, set the flag accordingly.

  • Set the g++ Compiler - Preprocessor symbol INSIDE. This is used in the main.cpp file in the next section.



Step 7: Enabling building and launching

  • Create a new c++ file main.cpp in the src folder with the following code
#ifdef INSIDE
#include <Rcpp.h>
#include <RInside.h>                    // for the embedded R via RInside
#include "rcpp_hello_world.h"
using namespace Rcpp;
using namespace std;
int main(int argc, char *argv[]) {
    RInside R(argc, argv);              // create an embedded R instance
    SEXP s = rcpp_hello_world();
    Language call("print",s);
    return 0;



  • Execute Project - "Build All". You should see the project building in the eclipse console, ending with **** Build Finished ****.
  • Next, go to Run - "Run Configurations".

  • Create a new C++ run configuration set to launch the binary Debug/MyCppPackage

  • Hit run. You should then see the binary building and running. The output of main.cpp will be emitted into the Eclipse console. This allows you to do rapid prototyping of c++ functions even if they require the full functionality of R/Rcpp.

You can also make use of Eclipse's extensive and very useful code sense to both detect errors in your code, and to assist in finding functions and methods. ctrl+space will trigger code completion


Final thoughts


Rcpp has made R package development with C++ code orders of magnitude easier than earlier APIs. However, without the features modern IDE, the developers life is made much harder. Perhaps all the old school kids with Emacs/ESS have trained their brains to not need the features I outlined above, but I am certainly not that cool. I found that I was at least 4-5 times more productive with the full Eclipse set-up as I was prior.

After following the steps above, you should have a template project all set up with Eclipse and are ready build it out into your own custom Rcpp based package. These steps are based on what worked for me, so your milage may vary. If you have experiences, tips, suggestions or corrections, please feel free to post them in the comments. I will update the post with any relevant changes that come up.

Update: Windows users may find this SO helpful 



Fellows Statistics provides experienced, professional statistical advice and analysis for the corporate and academic worlds.




Filed under: R, Rcpp, Uncategorized 1 Comment

Reading Excel data is easy with JGR and XLConnect

Despite the fact that Excel is the most widespread application for data manipulation and (perhaps) analysis, R's support for the xls and xlsx file formats has left a lot to be desired. Fortunately, the XLConnect package has been created to fill this void, and now JGR 1.7-8 includes integration with XLConnect package to load .xls and .xlsx documents into R.

Not fancy, but very useful.


Normality tests don’t do what you think they do

Last week a question came up on Stack Overflow about determining whether a variable is distributed normally. Some of the answers reminded me of a common and pervasive misconception about how to apply tests against normality. I felt the topic was general enough to reproduce my comments here (with minor edits).

Misconception: If your statistical analysis requires normality, it is a good idea to use a preliminary hypothesis test to screen for departures from normality.

Shapiro's test, Anderson Darling, and others are null hypothesis tests against the the assumption of normality. These should not be used to determine whether to use normal theory statistical procedures. In fact they are of virtually no value to the data analyst. Under what conditions are we interested in rejecting the null hypothesis that the data are normally distributed? I, personally, have never come across a situation where a normal test is the right thing to do. The problem is that when the sample size is small, even big departures from normality are not detected, and when your sample size is large, even the smallest deviation from normality will lead to a rejected null.

Let's look at a small sample example:

> set.seed(100)
> x <- rbinom(15,5,.6)
> shapiro.test(x)
Shapiro-Wilk normality test
data: x
W = 0.8816, p-value = 0.0502
> x <- rlnorm(20,0,.4)
> shapiro.test(x)
Shapiro-Wilk normality test
data: x
W = 0.9405, p-value = 0.2453

In both these cases (binomial and lognormal variates) the p-value is > 0.05 causing a failure to reject the null (that the data are normal). Does this mean we are to conclude that the data are normal? (hint: the answer is no). Failure to reject is not the same thing as accepting. This is hypothesis testing 101.

But what about larger sample sizes? Let's take the case where there the distribution is very nearly normal.

> library(nortest)
> x <- rt(500000,200)
> ad.test(x)
Anderson-Darling normality test
data: x
A = 1.1003, p-value = 0.006975
> qqnorm(x)
> hist(x,breaks=100)

Here we are using a t-distribution with 200 degrees of freedom. The qq and histogram plots show the distribution is closer to normal than any distribution you are likely to see in the real world, but the test rejects normality with a very high degree of confidence.

Does the significant test against normality mean that we should not use normal theory statistics in this case? (Another hint: the answer is no 🙂 )

Filed under: R, Uncategorized 3 Comments

Installing rgdal on a Mac

So, installing rgdal, which is an important R package for spatial data analysis can be a bit of a pain on the mac. Here are two ways to make it happen.


The Easy Way
  1. In R run: install.packages('rgdal',repos="http://www.stats.ox.ac.uk/pub/RWin")
The Hard Way
  1. Download and install GDAL 1.8 Complete and  PROJ framework v4.7.0-2   from: http://www.kyngchaos.com/software/frameworks%29
  2. Download the latest version of rgdal from CRAN. Currently 0.7-1
  3. Run Terminal.app, and in the same directory as rgdal_0.7-1.tar.gz run:


R CMD INSTALL --configure-args='--with-gdal-config=/Library/Frameworks/GDAL.framework/Programs/gdal-config --with-proj-include=/Library/Frameworks/PROJ.framework/Headers --with-proj-lib=/Library/Frameworks/PROJ.framework/unix/lib' rgdal_0.7-1.tar.gz






Tagged as: , , 6 Comments

A Titan

Another titan of computer science died yesterday. Dennis Richie, father of C and UNIX passed away at 70. I enjoyed this post by Herb Sutter which gives an idea of the scale of cognitive leap that he had to make in order to think that he could make a universal programming language competitive with platform specific assembly.

For me it begs the modern question: Is it possible to build a dynamic modern programming language (ala Python, Ruby, R, etc.) that performs competitively with C? Most examples that I can think off are of by two orders of magnitude.

Filed under: Uncategorized No Comments

Open Street maps

There have been some exciting developments in the Deducer ecosystem over the summer which should go into CRAN release in the next few months. Today I'm going to give a quick sneak peek at an Open Street Map - R connection with accompanying GUI. This post will just show the non-GUI components.

The first part of the project was to create a way to download and plot Open Street Map data from either Mapnik or Bing in a local R instance. Before we can do that however, we need to install DeducerSpatial.


#get development versions

Note that you will need rgdal. And you will need your development tools to install the development versions of JGR, Deducer and DeducerSpatial.

Plot an Open Street Map Image

We are going to take a look at the median age of households in the 2000 california census survey. First, lets see if we can get the open street map image for that area.


#load package

#create an open street map image
lat <- c(43.834526782236814,30.334953881988564)
lon <- c(-131.0888671875  ,-107.8857421875)
southwest <- openmap(c(lat[1],lon[1]),c(lat[2],lon[2]),zoom=5,'osm')

Note that plot has an argument 'raster' which determines if the image is plotted as a raster image. 'zoom' controls the level of detail in the image. Some care needs to be taken in choosing the right level of zoom, as you can end up trying to pull street level images for the entire world if you are not careful.


Make a Choropleth Plot

Next, we can add a choropleth to the plot. We bring in the census data from the UScensus2000 package, and transform it to the mercator projection using spTransform.


#load in califonia data and transform coordinates to mercator
california.tract <- spTransform(california.tract,osm())

#add median age choropleth
choro_plot(california.tract,dem = california.tract@data[,'med.age'],
	legend.title = 'Median Age')


Use Aerial Imagery

We can also easily use bing satellite imagery instead.

southwest <- openmap(c(lat[1],lon[1]),c(lat[2],lon[2]),5,'bing')
choro_plot(california.tract,dem = california.tract@data[,'med.age'],alpha=.8,
	legend.title = 'Median Age')



One other fun thing to note is that the image tiles are cached so they do not always need to be re-downloaded.


> system.time(southwest <- openmap(c(lat[1],lon[1]),c(lat[2],lon[2]),zoom=7,'bing'))
   user  system elapsed
  9.502   0.547  17.166
> system.time(southwest <- openmap(c(lat[1],lon[1]),c(lat[2],lon[2]),zoom=7,'bing'))
   user  system elapsed
  9.030   0.463   9.169


Notice how the elapsed time in the second call is half that of the first one, this is due to caching.

This just scratches the surface of what is possible with the new package. You can plot any type of spatial data (points, lines, etc.) supported by the sp package so long as it is in the mercator projection. Also, there is a full featured GUI to help you load your data and make plots, but I'll talk about that in a later post.

Tagged as: , 2 Comments

Blog open

The blog is now up and running. This self referential post is the first one.

Filed under: Uncategorized 1 Comment