What is your favorite free tool on Linux for multivariate logistic regression?
Possibilities I've seen:
- R (see paper). This question says use design.
- Can you use SciPy?
Other choices?
Do people have experience with large data?
$\begingroup$
$\endgroup$
-
5$\begingroup$ What is the purpose of your question? Why do you need such a list? Doesn't google search on statistical packages on linux answer your question? $\endgroup$ – mpiktas Jan 20 '11 at 16:49
-
$\begingroup$ Purpose: to elicit actual experiences. Need: I think I may have large multivariate logistic regression in my near future. Google searches: no, it doesn't. Have you done the searches? Plenty of non-free, plenty of abandoned projects, etc. $\endgroup$ – dfrankow Jan 21 '11 at 4:25
-
$\begingroup$ what is large? Do you have any special needs? The multivariate logistic regression has pretty standard algorithm for solving it (iterative reweighted least squares), so every software is simply reimplementing it. In this case you usually go with the most popular implementation. Or you can write the algorithm yourself, it is not so hard. $\endgroup$ – mpiktas Jan 21 '11 at 7:03
-
1$\begingroup$ No special needs. When you say "simply reimplementing" .. that is work. I want easy. $\endgroup$ – dfrankow Jan 25 '11 at 17:28
$\begingroup$
$\endgroup$
If you want to use Python, you can use the scikit-learn, which relies on Scipy. See the documentation for an example of the logistic regression:
http://scikit-learn.sourceforge.net/auto_examples/linear_model/plot_logistic_path.html
This implementation is based on liblinear, thus it scales reasonably well. In addition, it implements L1 and L2 penalization, for sparse or shrunk regression, when dealing with high-dimensional data.
-
$\begingroup$ Does it report fit statistics well (e.g., p values, variance inflation factors, ...)? $\endgroup$ – dfrankow Jan 25 '11 at 17:29
-
$\begingroup$ I don't know logistic regression terribly well, so I cannot give a precise answer. However, I can see that the object implements a predict_proba method: scikit-learn.sourceforge.net/modules/generated/… . With regards to variance inflation factors, I am pretty sure that they are not implemented, though they could be computed as a post-processing step. $\endgroup$ – Gael Varoquaux Jan 25 '11 at 22:19
$\begingroup$
$\endgroup$
R is a full statistical computing environment, featuring a programming language especially designed and optimized to this purpose and an enormous an high quality library tightly covering the whole are of data science. SciPy is just a BLAS for Python with some support to basic statistics.
-
$\begingroup$ Scipy has much more than BLAS. In terms of Fortran packs, it binds optimization packs such as minpack, fitpack, and other linear algebra packs. It also has signal or image processing tools, optimization. I personally find that SciPy is much more versatile than R. It addition, you seem to be comparing the complete set of R packages to the single Python package 'scipy'. To be fair, you should take in account the huge variety of Python packages, a lot of them useful for science. Now I will easily acknowledge that for the specific task of statistical data processing, R is much better equipped. $\endgroup$ – Gael Varoquaux Jan 21 '11 at 16:42
$\begingroup$
$\endgroup$
I'm not sure, but i think PSPP does logistic regression. Not entirely certain if it can handle multivariate logistic regression.
