PCR's notepad » statistics

Computing the standard normal cumulative distribution with R

David Dernoncourt — Mon, 21 Nov 2011 15:36:35 +0000

I recently had to compute the Bayes error for a Gaussian model.
Let x_i be a predictive variable, following a Gaussian distribution of variance 1 and mean T_i in a class, and -T_i in the other class.
The Bayes error of the model is e=1-Φ(sqrt(sum(T_i^2)).
In R, the standard normal cumulative distribution function is computed using pnorm().

So, here’s the Bayes error if for instance we have 10 variables with mean 0.5 in a class and -0.5 in the other class:
Bayes_error=1-pnorm(1*sqrt(10*0.5^2));

If we had more complicated things, like 5 variables with mean 0.5,0.6,0.7,0.8,0.9 in a class and -0.5,-0.6,-0.7,-0.8,-0.9 in the other, we could do something like:
mus=c(0.5,0.6,0.7,0.8,0.9); Bayes_error=1-pnorm(1*sqrt(t(mus)%*%mus));
or not vectorized: Bayes_error=1-pnorm(1*sqrt(sum(mus*mus)));

Quick statistics competence assessment

David Dernoncourt — Thu, 18 Mar 2010 16:32:58 +0000

Do you understand the following comic?

If you don’t, your statistical skills require an emergency treatment (unless you actually don’t use statistics) because you are likely to draw erroneous excessive conclusions.
If you do, well, I can’t conclude.

R-project Portable

David Dernoncourt — Thu, 21 Jan 2010 08:21:33 +0000

The guys at R-project tell you that you can run R from a CD or USB drive (so it’s portable), but nowhere they provide you with a really portable version. I.e., you have to actually run the installer then you can make a portable copy of the R folder on your USB key. That’s not exactly practical because if for instance you want to “install” it at work on a restricted computer where you can’t run an installer, you need to start with some homework first…
So for the lazy like me, here is a portable version of R 2.10.1. I made it from a custom installation (as far as I remember it’s like the standard installation without language packs and with only some of the manuals) and it’s in English, of course

Also, note that automatically installing additional packages should work fine on this portable version (I did so with my previous 2.10.0 portable, so no reason it wouldn’t work with that one).