R in The New York Times

The link to r-project.org appears to be defunct. There is just a web search squatting on the domain.

@2: Try it with a leading www, like this: http://www.r-project.org/

Well, that was an unfortunate ratio of posts to information. And I’m just making it better, aren’t I?

This sounds like a European adult film: “The Erotic Flowering of R”

And I’m just making it better, aren’t I?

Not really, I’d say.

I agree with you Kieran but I find R awesomely annoying to work with. Yes the language is pithy and that’s good but it has a lot of idiosyncracies that can be annoying. The worst one to date so far has been the Japanese-language R’s platform dependence. A program I wrote on my computer failed on my colleague’s until we removed the comments.

Also, the mailing list helpy-forum is completely evil. Everyone on there is nasty.

I find Stata really horrible to use compared to R. It manages to somehow mangle the flexible vector/matrix approach of R by merging it with the sequential approach of a 60s program. You have to have the linear model as the last modelling operation in order to extract residuals, right? This is very painful. And the way of looping through objects has to be the most painful thing I have ever seen. Applying those weird names in quotes to a programming object which itself has a funny name, and has to be manipulated in a funny way (it should be clear from this sentence I haven’t used Stata a lot).

Both of these packages have huge positives too, but the way they really fall over (and the only reason SAS remains popular IMHO) is that they just can’t handle big data. For my data sets I would need a computer with at least 8Gb of RAM if I wanted to use stata, even after crunching the data set to maximum possible efficiency. If I want to do anything exploratory and have a few more variables, I need 16Gb. And there’s no guarantee that even then I’ll be able to actually complete the analysis in my own lifetime.

APL? Pray tell me more!

I use both Stata and R but prefer Stata (or SAS) for large datasets.

You mean to tell me that SAS Institute peeps haven’t seen the Linux Penguin on their in flight entertainment centers? My oh my.

Yes, I use it. Basically, non-statisticians, undergraduate statisticians, and several working statisticians that I know of.

I don’t think that’s necessarily true Walt. To the best of my knowledge the implementation of GEEs in R is pretty crappy (at least it was last time I checked) and its ARIMA functions are annoyingly incomplete. I don’t know much about the topic, but I’m not under the impression that its odbc support or sql is not at all good. Plus of course the help is crap, and full of nasty rude people misunderstanding your question and being rude. I think these kinds of problems come up with open source software rather a lot.

Eszter, hundreds of thousands of observations aren’t the challenge. I have 10s of millions, and stata can’t handle any data set it can’t load. This makes a lot of trouble. For ANOVA-type problems this isn’t a big deal, but for something like a random effects model or a survival analysis it’s impossible. It’s really annoying when you have millions of observations but you essentially have to sample from them so viciously that they might as well not exist, because your software can’t handle the size.

What would be really nice would be a stats package you can switch between vector-style analysis for small data sets, and SAS-style chunk-by-chunk stuff for large data sets. Which I suppose I have, since I use both…

sg, I can see the problem there, definitely.

MH, tons of undergraduates and even graduate students use SPSS (not to mention professors in some of the fields with which I’m affiliated). My impression is also that people at marketing firms and such use SPSS often.

A truly scary quote in the quoted article:

“We have customers who build engines for aircraft. I am happy they are not using freeware when I get on a jet.”

I’ll interpret “freeware” as “free software,” as I doubt she was using the term in the original sense (free but not open source). This is a horribly thought out comment, in the same vein as thinking closed source software voting is a good idea. What would you prefer?: flying on a plane running code 10 people had read and tested, or flying on a plane that had had thousands of eyes read and testing the code? It’s a clear answer to me.

Any company that refers to itself as an “Institute” is bound to be prone to fits of stuffy huffiness.

mds, the idea of ‘many eyes’ isn’t because people are stupid; it’s because many eyes tend to spot errors. I just spent an afternoon fixing something that I had done carefully, in methods I had set up to minimize errors, then checked, and then was checked by a co-worker who’s very careful. Then our boss checked it,because he was presenting to upper management, and didn’t like embarrassment.

Somebody in the 100 people to whom it was sent spotted the errors.

sg: Eventually they won’t suck, and eventually RAM will get so big that 99% of users will have datasets that R or Stata can load. At that point the jig is up.

SAS has the worst programming language known to man, and has a fantastically steep learning curve. These will eventually prove fatal, just similar matters proved fatal to Cobol.

Just yesterday a colleague announced that he made a really cool simulation where Excel called out to R code, efficiently and seamlessly. Until that tale I was expecting R to be as unstable as, say, ten-years-ago linux, but if it cooperates with Office, it must be pretty robust, to say the least.

SAS isn’t going away because it has the commercial data-mining market captivated (if not captive). It handles enormous datasets easily, and is backward compatible. But things that take one line in Stata can take several paragraphs worth of SAS code. (And, like Eszter, I’ve never used either language’s GUI.)

SAS is a pain for some things and most models are more easily done in Stata (for me at least). But, if you need to combine data from different sources and to aggregate across records, Stata becomes very awkward very quickly. I use Stata frequently, but 99% of the time, its on datasets that were built in SAS.

SAS’s various attempts to make a user friendly interface have all been massively unworkable, even worse than finding the information you need in their documentation. Though SAS 6 did have a workable poker game on it. On the other hand, Stata’s pull-down analysis menus are actually a good way to learn about some feature that is new to you. Particularly if you learned in SPSS and like that format.

watson, in my experience sas doesn’t have any real overlap with Matlab, which is a matrix programming language and has limited (but nice) statistical functionality.

Also, regarding writing your own code for statistics functions – that way lies madness. There is an enormous amount of work behind even the R code for a linear model, and reproducing that yourself is a huge amount of work and carries huge risk of error. Also, if you have to do a diverse range of tasks, every new task carries a lot of startup costs. If that’s what APL requires then I think it has no use for me.

Another aspect of SAS I think is often overlooked is its voluminous online help, which is impossible to search (very correct MH) but extremely detailed in the mathematics.

I don’t use SPSS routinely but there is one aspect of SPSS which I really like – the sample survey functions. In SPSS you generate a sample plan file, which you have to open before you can do any analysis on complex survey data. This is great if you have a lot of people working on the same survey across an organisation, who know how to do basic statistics in a GUI but don’t understand the survey stats. One expert generates the plan file and then everyone in the organisation can be told to use it. It reduces the risk that one of your psychology grads will use frequency weights on their survey data, or use the wrong set of probability weights, or miss out the finite sample correction, etc. so you can farm out the simple tasks without having to teach the complex code.

I also agree with Walt at 34 but stata and R have programming languages that are much less intuitive than SAS, IMO. Particularly R. This will slow down its general acceptance, I feel.

I think after all that, my question was still unclear. Sorry, I’m tired.

Does SAS, like Matlab, have a core base of users in industry (and less so academia) that are unlikely to bother learning a replacement because a) they know it and b) it does most of what they want, even if a bit painfully?

“We have customers who build engines for aircraft. I am happy they are not using freeware when I get on a jet.”

I am happy inform you that while I do not know about the engines, I do know that wings are designed using open source software.

@ 41, 42 – No, SAS is not like Matlab in that respect; it’s far, far huger. SAS is really a data processing system with a statistical analysis system. A large corporation/gov’t agency could do massive data processing with SAS, while only a few people use SAS for any statistical use beyond descr. stats and frequency tables.

Back in the 1990’s, one pharmaceutical company (Merck?) went from SAS to S-plus, and that was big news; they’re locked into SAS.

Sorry, sg–I originally thought your question was a joke.

I think it’s been answered; APL is an array-based programming language dating back to the 1950’s, with its good points and its bad. I don’t know about statistical libraries, but there’s semi-publicly-available code to do almost anything data-related.

If you want a plug-and-play statistical system, APL is NOT what you want.

And a question in turn to Watson Aname–what’s older than APL?

It depends on the scale. Ideally, the data goes into databases to start, (possibly pre-processed by SAS), with proper data entry forms, scanning, validation, etc. Then, one could extract the data using report-writing software (e.g., Crystal Reports), either add hoc or pre-set batch jobs. SAS has modules to pull data from databases as if they were native SAS, from the viewpoint of the SAS user. At that point, a SAS user can ‘post-process’ in SAS for the desired analyses and reports.

Depending on the set-up and one’s position, one might be using SAS-processed data, or SAS-generated reports, or SAS-based input screens, without knowing that SAS stands for anything other than some Brit SF guys, or a shoe company in San Antonio.

In my background, SAS is usually used as the data processing/analysis software for research projects (engineering, social science and medical). It’s convenient to use one package for a lot of stuff. The trick is data entry and validation; I’ve seen a lot of people figure that incorrectly-entered data will be ‘taken care of later’, possibly using Statistical Methodology :(

And a question in turn to Watson Aname—what’s older than APL?

For various things, I use both Fortran (for libraries) and Lisp (for flexibility, metaprogramming). Of course, “older” is a bit of a misnomer in all cases, as it’s not like we’re using the original variants of any of these.

Barry @ 46: Thanks, I knew it was used in industry, but had no real feel for the userbase.

You’re welcome, Watson. As a statistician, it feels like somebody asked if MS Office is used much in the corporate world :)

Watson, I think the others have answered your question. As an example of what SAS is used for, the New South Wales Health Department (now NSWHealth) built a huge data extraction and analysis system for health professionals, called HOIST, mostly in SAS. I use SAS for partially legacy reasons, but also because it’s a very convenient way to implement sql for large datasets and do statistical analysis on those large sets.

I think there is a large body of data miners, and anyone who works with statistical analysis on large datasets, who use SAS.

Also, I have been waiting for a decent implementation of GEEs in R for at least 3-5 years, so I’m not so sure it’s coming soon. It must be something that the R community don’t need. In my experience, if you want to do a decent log-linear model with clustering and/or serial dependence, SAS is the easiest way to put it together. And writing that by yourself is a lot of work. I don’t just do plug and play stats, but I am not crazy enough to pretend I can do a better job of writing a whole GEE analysis package by myself – that’s what we have smart people for!

In physics, the trend has been moving to ROOT, an open source analysis system developed at CERN (see root.cern.ch). ROOT is based in C++, and most new collider physics experiments are using ROOT as the underlying foundation for their custom made analysis framework. We have insanely large datasets, and ROOT seems to do well. In addition, new, powerful features are continually being added.

The Times article was also discussed by Andrew Gelman, who says “I just hate SAS. It’s not just a matter of it having poor capabilities; SAS also makes people into worse statisticians, I think.” He explains in the comments, “SAS spews out pages and pages of output for any analysis. The output isn’t easy to post-process; as a result, people stare at the output and pick out numbers. R more easily allows graphical and other postprocessing of inferences.” He’s right about post-processing, at least. The R list (similar to a Ruby or Python hash) is a more straightforward way to access results than ODS. R beats Stata here too.

Statisticians are fond of blaming computer packages for making people “worse statisticians”, but in my experience the people they are thinking of didn’t know what they were doing anyway.

Just so. And there is often handwringing about packages like SPSS enabling the heathens to analyze data. I’ve always felt that such concerns were misplaced, because the heathens would otherwise be making decisions without any acquaintance with the data.

you are unlikely to be doing HLMs anyway unless you have direct access to God

I know the deity.interface group has stuff in progress, but it isn’t on CRAN yet.