Elena Belogolovsky writes:
Congratulations to the R Core Team on receiving the 2026 Rousseeuw Prize for Statistics.
R has made creative, open-ended statistical analysis and graphics accessible to generations of statisticians and applied researchers. It has also been central to statistical research, methodology, and applications during decades when statistics became more computational and more important across science, engineering, business, and public health.
One of the great strengths of R is that it is not just a software platform. It is also a community. The system of R packages allows anyone to implement a new method and share it with the world, helping make statistical research more open, useful, and alive. R has also been the medium for major developments in statistical graphics, transforming applied statistics and the way people work with data.
The volunteers who have developed, guided, and maintained R and the R community are richly deserving of this major award.
I agree with the committee that the R team is an excellent recipient of this award. I say this for several reasons:
– Most obviously, R is super-useful and it’s changed statistics, both by enabling more complicated and reliable analysis and by establishing a common language for statistical coding.
– R integrates statistical modeling with graphics, which traditionally (but, in my opinion, mistakenly) have been thought of as in opposition.
– R is open source. This might sound like no big deal, but its predecessor was Splus, which was a commercial package. Before that came S, which was open but was not set up to expand in a scalable way.
– With its system of packages, R became modular: different groups of users (including me!) could write their own packages and develop new and useful tools without needing to get tangled in core R issues. For example, we have cmdstanr, which lets you run Stan programs from R. This is super-useful for Bayesian workflow.
– R is a programming language, not a menu-based set of commands. This is no big deal now, given that the natural comparison to R is Python, but, back in the day, when R’s competitors were Sas, Spss, Stata, etc., it was a big deal that with R you write programs, you don’t just push buttons. A big deal for workflow in statistics and data science.
– Regarding the R community . . . ok, this gets complicated. Still and all, the R core team is very helpful to outsiders and has been a clear net benefit to the communities of developers, statisticians, and users.
I’m sure I’m missing a few things. My only disagreement with the award citation is that it doesn’t mention S, the statistical software environment developed by John Chambers and others at Bell Labs back in the 1980s. R is a rewrite of S. With lots of improvements, but I do think the S team deserves credit for setting up the template.
The website of the Rousseeuw prize for statistics mentions “Five members of the R Core Team have been awarded the Rousseeuw Prize for Statistics for their decades of work building and maintaining the R Project.”, one of which is professor Brian Ripley. In his bio on the Rousseeuw prize page the following can also be read:
“He is known for pioneering work in spatial statistics, simulation-based inference, pattern recognition and statistical computing, the subjects of his six books. ‘Modern Applied Statistics with S’ (1994) has been very widely adopted and influential in the take-up of S and R. He contributed extensively to R before becoming a member of the R Core Team in 1999.”
As someone who has mentioned some criticism regarding handing out individual scientific awards on this blog (e.g. see https://statmodeling.stat.columbia.edu/2025/05/05/charles-margossian-and-lawrence-saul-win-aistats-25-best-paper/#comment-2396870), this award and especially the specific phrasing of things seems much better at first glance to me.
For instance, on the website it is stated that the prize is awarded to “the R-project”, and the following can also be read on the site: “Half of the prize money goes to the five laureates because they are deemed to have made the longest sustained contributions, and half goes to the other members of the R Core Team.”.
The prize was especially given for Ripley’s thoughtful and generous contributions to the R help forums over the years.
Having experienced being Ripleyed, I think the ease of compilation of R and packages, on multiple platforms, has a lot to do with his contribution – ‘works for me’ does not count when you need to compile on obscure unix variants or processors. So far, I’ve found compiling R just works, while trying to compile open source Rstudio is dependency hell.
Brian responded to the only question I will ever ask on the R forums, which came with a long explanation of where we were stuck in the 250 pages of C integration doc and contained a reproducible example. Brian wrote back that we (it was me and Jiqiang Guo, the original RStan author) were too stupid to understand the answer and shouldn’t bother asking for more help. I use “Brian” rather than “Professor Ripley” because his response to one of the postdocs at Columbia when he responded to Brian was, “How dare you address a full professor by his first name.” followed by a rant on how he wasn’t going to answer the question.
We currently can’t update RStan on CRAN because it doesn’t run on Brian’s home machine, which we do not have access to. Even the public releases of the C++ interfaces make the strange decisions to rely on intermediate and alpha versions (the reason we couldn’t use C++11 in Stan for the longest time). The restrictive 1980s-era policies on package size and complete lack of dependency management make maintaining an R package that’s bigger than 1MB a nightmare. It’s why, despite huge efforts on our part, we haven’t updated RStan in several years and why we don’t want to put cmdstanr on CRAN.
This is not a problem in Python, by the way. They actually had people involved in that project who understood how computer languages should work and how packages should be managed.
Bob,
Wow–I’ve never heard Ripley referred to as “Brian” before. Isn’t that the nickname of the former Prince Charles?
“He’s not the messiah, he’s a very naughty boy!”
Re the “this gets complicated” link
“So, I appreciate the years of unpaid service that these volunteers have put into the R help list, and as far as I’m concerned, they can be as crabby as they want for as long as they want. I hope we can work out a better system so they can be even more effective and so that they don’t feel they need to waste so much of their time on the easy questions.”
I guess that better system was LLMs… and it also ended up killing the likes of Stack Overflow! Granted, I think LLMs were the nail in the coffin; from my perspective, the nastiness (and other issues like bad moderation/”this has already been asked even though it actually hasn’t”/etc) had already already reached a degree that few people wanted to post there anymore anyway.
SAS technically works as a programming language too, but of course it is an very old and inelegant one. I think the modularity is the biggest draw for R.
Am I wrong in thinking that SAS itself is not Turing complete, and that its generality comes from its macro language being Turing complete? Basically to write any real amount of SAS you have to write a macro that potentially writes an infinite series of base SAS instructions? That was what my impression was back when I worked with SAS people (1990’s)
Unless I am mistaken, data steps are already Turing complete. For one, you could simulate a simple CPU in one.
It’s been decades so I don’t know, but is it possible for a data step to loop back and reread the same dataset modifying it and repeating until some condition is met? I was under the impression each datastep ran through the dataset once and you had to have the macro language to decide if you wanted to process the data again. Like I said, decades since I’ve used it, so maybe I’m wrong.
My understanding is it is actually difficult to avoid making your software turing complete. Like you may want to for security reasons, but after continuing to add features it will just happen on accident.
I have had a lot of just pure fun with R. As long as you pay attention to your Ps and Qs the ability to combine a set of packages with all kinds of capabilities, particularly with statistics and graphics, gives a lot of power to try to solve problems. Modern Applied Statistics with S’ (1994) is on my bookshelf and has coffee stains all over it.
+1 for using “As long as you pay attention to your Ps and Qs” in a post about R whilst also referring to S.
A prize for these R contributors is well deserved. I started using R before the 1.0 release (must have been 1997 or 1998?). At that time machines were 32 bit, and you had to give a command line to estimate the size of the cons cell area and the heap separately, so that you didn’t run out of space during your analysis. Reading a CSV file required the file to be small enough that it could be read entirely into the heap and parsed out into structures… A large amount of RAM was like 128 MEGA bytes, so anyway, it’s come a long way and its absolutely great. That being said, after using it for 20 or so years, I moved to Julia in 2019 and I haven’t looked back. Julia’s big advantage is that it compiles directly to machine code while still being a high level interactive language like R. It has a good DataFrames package, it has MixedModels.jl written by the author of lmer it has GLM for generalized linear models, both Turing.jl and Stan.jl for Bayesian modeling, and quite honestly the single best Differential Equations library of any modern language period.
Anyway, I think Julia is under-appreciated and people should take a look at it. But I’ve been admonished to keep the Julia boosterism down to like 1 per week so this is my current quota. ;-)
I stopped using R and switched to Python the day I stopped working for Andrew.
Neither Stan nor Julia are on the cutting edge of statistical computing. For that, read Matt, Colin, and Pavel’s paper in the latest MCMC Handbook, Running MCMC on modern hardware and software (hint: it’s not R!). Lots of folks have followed Matt et al. into JAX because of its GPU and parallelization performance. For example, Elizaveta Semenova, who we invited to StanCon for a plenary session started her talk saying she used to use Stan, but it’s just not performant enough for their current uses (epidemiology) and they’ve switched to JAX. The Dodgers baseball team reported the same thing. As have many other people non-publicly to me.
With JAX, you don’t need something like Stan—it’s already a differentiable, compiled programming language. But it’s not just programming languages. I suspect diffusion models or simpler normalizing flows will dethrone MCMC as the go-to method for Bayesian inference as soon as SIMD compute gets another order of magnitude cheaper. On the other hand, the parallel algorithms for MCMC from Zoltowski et al. are very tantalizing, and we’re also exploring those. They might rescue MCMC from near-certain obsolescence in the face of parallel compute.
“as soon as SIMD compute gets another order of magnitude cheaper. ”
You may have a long wait. I’m running a 3080. Two generations later, the 5080 is the corresponding device.
I asked a relatively computer-literate friend to guess how much faster the 5080 is than the 3080.
HE.WAS.WRONG.
What’s your guess???
The new Macbooks have pretty decent GPUs and are very cheap. Their flop throughput obviously isn’t better than a 3080, but their enormous L2 cache and unified memory access may make them even faster for certain classes of algorithms.
Replying to Somebody:
Yes. The Mac trick of unified memory speeded things up (but it’s pretty old news by now). The new ‘Doze Spark/RTX ‘notebooks will do that as well.
The blurb I just read claimed those new things will have “5070” level performance, so they won’t be noticeably beating a 3080. For the nonce, notebooks will be power limited, so they won’t be where “order of magnitude cheaper/faster” SIMD machines appear. For the nonce.
I’m also amused by the claimed specs for the next gen i86 chips: an insane number of processors in three levels of power efficiency (since instructions per watt is now more important than instructions per second and high instruction/Watt cores take up less real estate.). But that won’t amuse most folks here.
Bob:
I can’t speak for Julia, but I disagree with you when you say that Stan is not “at the cutting edge of statistical computing.” There are different aspects of “the cutting edge.” I’m sure you’re right that Stan is not at the cutting edge of using parallel processing for GPUs, but that’s not the only cutting edge of statistical computing. Being able to express a model in a way that is understandable, to fit the model and simulate it, and to be able to alter it, that’s on the cutting edge of statistical computing in its own way.
To put it another way, one part of the cutting edge of statistical computing is Bayesian workflow.
It’s fine for you to switch from R to Python, and I know you have good reasons for it. Also, a lot of things can be done well in R, and there’s a reason that many people continue to use R. I complain about R as much as anyone, and in some ways it’s not at the cutting edge. In other ways it is.
I’m not a computer scientist and I’m not much of a software developer. I’m speaking here as a user and as a developer of statistical methods.
“– R is a programming language, not a menu-based set of commands. This is no big deal now, given that the natural comparison to R is Python, but, back in the day, when R’s competitors were Sas, Spss, Stata, etc., it was a big deal that with R you write programs, you don’t just push buttons.”
This is not an accurate characterization of SAS, SPSS, and Stata. Other commenters have also pointed out that SAS has a programming language, if not necessarily a very good one. The same is true of SPSS and Stata.
For Stata, we can go even further: Stata version 1 was entirely a programming language–there were no buttons to push. I don’t remember at what version Stata introduced menu-driven commands. But those are mostly used by beginners or infrequent users. Professional Stata users rely primarily, if not exclusively, on the programming language, which, by the way, is Turing complete and has more than a passing resemblance to C.
Clyde:
I’ve used Stata, and, yes, it’s a programming language, but in practice the way it is used is as a set of fixed routines. People load their data into Stata and run regressions. It’s very much focused on particular models–all sorts of things you can do with a rectangular dataset. The Stata analyses I’ve seen have not been flexible and exploratory the way that R sessions can be. So, even though Stata is not literally a point-and-click menu-based language, I think that operating in Stata, or Sas, or Spss, have a much different feel than operating in open-ended languages such as R and Python.
I would guess that the vast majority of R users use it for its fixed set of routines. Certainly that’s how all the biologists I’ve worked with use it.
Even the fixed set of routines in brms and posterior (two Stan packages in R) are way more popular than writing Stan and writing your own plotting and analysis code. That may change with LLMs, because it’s just so easy to use them to generate whatever graphics code you want.
I remember visiting John Carlin (BDA co-author) around a dozen years ago and asking him what kinds of models his students built. He laughed and said his students were in epidemiology, so their limit was usually running ANOVA and trying to interpret the output. Having said that, he was working with some stats students on some really cool models related to Andrew’s serial dilution example (the first talk I ever saw Andrew give, long before we worked together).
It depends who you see using these things. There are people using presets in R and people writing quite low-level programs in Mata within Stata. Personally, I only use LAPACK for its fixed routines. :-)
re: “My only disagreement with the award citation is that it doesn’t mention S, the statistical software environment developed by John Chambers and others at Bell Labs back in the 1980s. R is a rewrite of S. With lots of improvements, but I do think the S team deserves credit for setting up the template.”
A few people might be interested in oral history I did with John for Computer History Museum:
https://www.youtube.com/watch?v=CcA4aHKL_bE video
https://www.computerhistory.org/collections/catalog/102792762/ transcript