Reece Goding
- 1 Introduction
- 2 General Feelings
- 3 What R Does Right
- 4 What R Does Wrong
- 5 The Tidyverse
- 6 Conclusion
- 7 Feedback
What follows is an account of my experiences from about one year of roughly daily R usage. It began as a list of things that I liked and disliked about the language, but grew to be something huge. Once the list exceeded ten thousand words, I knew that it must be published. By the time it was ready, it had nearly tripled in length. It took five months of weekends just to get it all in R Markdown.
This isn’t an attack on R or a pitch for anything else. It is only an account of what I’ve found to be right and wrong with the language. Although the length of my list of what is wrong far exceeds that of what is right, that may be my failing rather than R’s. I suspect that my list of what R does right will grow as I learn other languages and begin to miss some of R’s benefits. I welcome any attempts to correct this or any other errors that you find. Some major errors will have slipped in somewhere or other.
To start, I must issue a warning: This document is huge. I have tried to keep everything contained in small sections, such that the reader has plenty of points where they can pause and return to the document later, but the word count is still far higher than I’m happy with. I have tried to not be too petty, but every negative point in here comes from an honest position of frustration. There are some things that I really love about R. I’ve even devoted an entire section to them. However, if there is one point that I really want this document to get across, it’s that R is filled to the brim with small madnesses. Although I can name a few major issues with R, its ultimate problem is the sum of its little problems. This document couldn’t be short.
Also, on the topic of the sections in this document, watch out for all of the internal links. Nothing in R Markdown makes them look distinct from external ones, so you might lose your place if you don’t take care to open all of your links in a new tab/window.
Before I say anything nasty about R, a show of good faith is in order. In my year with R, I have done the following:
- Added almost 100 R solutions to Rosetta Code.
- Asked over 100 Stack Overflow R questions.
- Read both editions of Advanced R from cover to cover. I didn’t do the exercises, but I’d recommend the books to any serious R user.
- Read the first edition of R for Data
Science from cover to cover. It’s a good
enough non-technical introduction to the Tidyverse and a handful of
other popular parts of R’s ecosystem. However, I can’t give it a
strong recommendation for a variety of reasons:
- A lot of the exercises didn’t specify what they wanted from your answer. This made checking your solutions against anyone else’s quite difficult.
- It deliberately avoids the fundamentals of programming – e.g. making functions, loops, and if statements – until the second half. I therefore suspect that any non-novice would be better off finding an introduction to the relevant packages with their favourite search engine.
- Despite my efforts, I can find no “Tidyverse for Programmers” book. When one is inevitably written, it will make this book redundant for many potential readers.
- Read The R
Inferno and
some other well-known PDFs and manuals, such as Rtips. Revival
2014! and the official An
Introduction to
R,
R Language
Definition,
and R FAQ manuals.
Out of all of these, I must recommend The R
Inferno. The
page count may be intimidating, but it’s a delightfully fast read that
mirrors many of my points. In many cases I have pointed the reader
straight to its relevant section. Its only true fault is its age. I
wish that I could claim that this document is a sequel to it, but I’m
writing to review rather than advise.
- Update: After publishing this review, I skimmed a handful of books by John Chambers. There were some gems in them and I’ve mentioned them where needed, but I don’t expect that I will ever read those books closely. I read them far too quickly for me to be able to say anything insightful, but I will confess that I feel fundamentally opposed to any programming textbooks that lack exercises.
- Made minor contributions to open source R projects.
At minimum, I can say with confidence that unless I happen to pick up an R-focused statistics textbook – the R FAQ has some tempting items – I’ve already done all of the R-related reading that I ever plan to do. All that is left for me is to use the language more and more. I hope that this section shows that I’ve given it a good chance before writing this review of it.
I am not an R expert. I freely admit that I am lacking in the following regards:
- You can never have done enough statistics with R. I’ve mostly used R as a programming language rather than a statistics tool. My arguments would certainly be stronger if I had some published stats work to back them up, even just blogs. I might correct this at some point.
- The above point makes me more ignorant of formulae objects
(e.g. expressions like
foo ~ log(bar) * bar^2
), theplot()
function, and factor variables than I ought to be. I saw a lot of them during my degree, but have long since forgotten them and have never needed to really pick them back up. For similar reasons, I have nothing to say on how hard it can sometimes be to read data in to R. - I haven’t used enough of the community’s favourite libraries. My
biggest regret is my near-total ignorance of
data.table
. From what little I’ve seen, it’s a real pleasure. More practice withggplot2
, the wider Tidyverse, and R Markdown is also in order. If I continue to use R, I will gradually master these. For now, it suffices to say that my experience with base R far exceeds my knowledge of both the Tidyverse and many other well-loved packages. If I’ve missed any gems, let me know. - I know almost nothing about Shiny, but it appears to be far better than Power BI.
- My experience with R’s competitors is minimal. In particular, I have virtually no experience with Python or Julia. Most of my points on R are about R on its own merits, rather than comparing it to its competition. I plan to pick up Python soon, but Julia is in my distant future.
- Although I have used SQL professionally, how it compares to R has
rarely crossed my mind. This suggests that I’m missing something about
both languages. I plan to one day read a SQL book while having
dplyr
loaded. - R’s functional aspects make me wish that I knew more Lisp. All that I’ve done is finish reading Structure and Interpretation of Computer Programs. I will learn more, but R’s clear Scheme inspiration makes Lisp books a lot less fun to read. It’s like I’ve already been spoiled on some of the best bits.
- I haven’t done enough OOP in R. My only real experience is with S3. S4 looks enough like CLOS that I expect that I will revisit it at some point after picking up Common Lisp, but that will just be to play around.
- I have never made a package for R and have no experience with the
ecosystem surrounding that (e.g.
roxygen2
). I have no plans for this. - I have no experience in developing large projects in R. This is likely a part of why I have never felt the need to make significant use of its OOP. I do not expect this to change.
The above list is unlikely to be exhaustive. I’m not against reading another book about R as a programming language, but Advanced R seems to be the only one that anyone ever mentions. For the foreseeable future, the main thing that I plan to do to improve my evaluation of R is to learn Python. I’ll probably read a book on it.
You’d be a fool to read this without some experience of R. I don’t think
that I’ve written anything that requires an expert level of
understanding, but you’re unlikely to get much out of this document
without at least a basic idea of R. I’ve also mentioned the Tidyverse a
few times without giving it much introduction, particularly its tibble
package. If you care enough about R to consider reading this document,
then you really ought to be familiar with the most popular parts of the
Tidyverse. It’s rare for any discussion of R to go long without some
mention of purrr
, dplyr
or magrittr
.
This document started out as personal notes that I had no intention of publishing. There’s a good chance that I might have copy and pasted someone’s example from somewhere and totally forgot that it wasn’t my own. If you spot any plagiarism, let me know.
My overall feelings about R are tough to quantify. As I mentioned near the start, its ultimate problem is the sum of its little problems. However, if I must speak generally, then I think that the problem with R is that it’s always some mix of the following:
- A statistics language with countless useful libraries and an excellent collection of mathematical tools.
- A Scheme-inspired language that tries to be functional while maintaining a C-like syntax.
- Decades of haphazard patches for S.
- A collection of semantic semtex that is powerful in the hands of a master and crippling in the hands of a novice.
When it’s anything but #3, R is great. Statisticians and mathematicians love it for #1 and programmers love it for #2 and #4. If it weren’t for #3, R would be an amazing – albeit, domain-specific – language, but #3 is such a big factor that it makes the language unpredictable, inconsistent, and infuriating. Mixed with #4, it makes being an R novice hellish. It gives me little doubt that R is not the ideal tool for many of the jobs that it wants to do, but #1 and #2 leave me with equally little doubt that R can be a very good tool.
As a final show of good faith, here is what I think R does right. In summary, along with having some great functional programming toys, R has some domain-specific tools that can work excellently when they’re in their element. Whatever the faults of R, it’s always going to be my first choice for some problems.
R wants to be a mathematics and statistics tool. Many of its fundamental design choices support this. For example, vectors are primitive types and R isn’t at all shy about giving you a table or matrix as output. Similarly, the base libraries are packed with maths and stats functions that are usually a good combination of relevant, generic, and helpful. Some examples:
-
Lots of stats is made easy. Commands like
boxplot(data)
orquantile(data)
just work and there are lots of handy functions likecolSums()
,table()
,cor()
, orsummary()
. -
R is the language of research-level statistics. If it’s stats, R either has it built-in or has a library for it. It’s impossible to visit a statistics Q&A website and not see R code. For this reason alone, R will never truly die.
-
The generic functions in the base stats library work magic. Whenever you try to print or summarise a model from there, you’re going to get all of the details that you could ever realistically ask for and you’re going to get them presented in a very helpful way. For example
model <- lm(mpg ~ wt, data = mtcars) print(model) ## ## Call: ## lm(formula = mpg ~ wt, data = mtcars) ## ## Coefficients: ## (Intercept) wt ## 37.285 -5.344 summary(model) ## ## Call: ## lm(formula = mpg ~ wt, data = mtcars) ## ## Residuals: ## Min 1Q Median 3Q Max ## -4.5432 -2.3647 -0.1252 1.4096 6.8727 ## ## Coefficients: ## Estimate Std. Error t value Pr(>|t|) ## (Intercept) 37.2851 1.8776 19.858 < 2e-16 *** ## wt -5.3445 0.5591 -9.559 1.29e-10 *** ## --- ## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1 ## ## Residual standard error: 3.046 on 30 degrees of freedom ## Multiple R-squared: 0.7528, Adjusted R-squared: 0.7446 ## F-statistic: 91.38 on 1 and 30 DF, p-value: 1.294e-10
shows us plenty of useful information and works just as well even if we change to another type of model. Your mileage may vary with packages, but it usually works as expected. Other examples are easy to come by, e.g.
plot(model)
. -
The rules for subsetting data, although requiring mastery, are extremely expressive. Coupled with sub-assignment tricks like
result[result < 0.5] <- 0
, which often do exactly what you think they will, you can really save yourself a lot of work. Being able to demand precisely what parts of your data that you want to see or change is a really great feature. -
The factor and ordered data types are definitely the sort of tools that I want to have in a stats language. They’re a bit unpredictable, but they’re great when they work.
-
It’s no surprise that an R terminal has fully replaced my OS’s built-in calculator. It’s my first choice for any arithmetical task. When checking a gaming problem, I once opened R and used
(0.2 * seq(1000, 1300, 50) + 999) / seq(1000, 1300, 50)
. That would’ve been several lines in many other languages. Furthermore, a general-purpose language that was capable of the same would’ve had a call to something long-winded likemath.vec.seq()
rather than justseq()
. I find the cumulative functions, e.g.cumsum()
andcummax()
, similarly enjoyable. -
How many other language have matrix algebra fully built-in? Solving systems of linear equations is just
solve()
. -
The
rep()
function is outstandingly versatile. I’d give examples, but those found in its documentation are more than sufficient. Open up R and runexample(rep)
if you want to see them. If tricks likecbind(rep(1:6, each = 6), rep(1:6, times = 6))
have yet to become second nature, then you’re really missing out. -
On top of replacing your computer’s calculator, R can replace your graphing calculator as well. Unless you need to tinker with the axes or stop the asymptotes causing you problems – problems that your graphing calculator would give you anyway – functions like
curve(x / (x^3 + 9), -10, 10)
(output below) do exactly what you would expect and exactly how.
These seem like trivial features, but the language’s deep integration of them is extremely beneficial for manipulating and presenting your data. They assist subsetting, variable creation, plotting, printing, and even metaprogramming.
-
The ability to name the components of vectors, e.g.
c(Fizz=3, Buzz=5)
, is a nice trick for toy programs. The same syntax is used to much greater effect with lists, data frames, and S4 objects. However, it’s good to show how far you can get with even the most basic case. Here’s my submission for a General FizzBuzz task:namedGenFizzBuzz <- function(n, namedNums) { factors <- sort(namedNums)#Required by the task: We must go from least factor to greatest. for(i in 1:n) { isFactor <- i %% factors == 0 print(if(any(isFactor)) paste0(names(factors)[isFactor], collapse = "") else i) } } namedNums <- c(Fizz=3, Buzz=5, Baxx=7)#Notice that we can name our inputs without a function call. namedGenFizzBuzz(105, namedNums)
I’ve little doubt that an R guru could improve this, but the amount of expressiveness in each line is already impressive. A lot of that is owed to R’s love for names.
-
Having a tabular data type in your base library – the data frame – is very handy for when you want a nice way to present your results without having to bother importing anything. Due to this and the aforementioned ability to name vectors, my output in coding challenges often looks nicer than most other people’s.
-
I like how data frames are constructed. Even if you don’t know any R at all, it’s pretty obvious what
data.frame(who = c("Alice", "Bob"), height = c(1.2, 2.3))
produces and what adding therow.names = c("1st subject", "2nd subject")
argument would do. -
As a non-trivial example of how far these features can get you, I’ve had some good fun making alists out of syntactically valid expressions and using only those alists to build a data frame where both the expressions and their evaluated values are shown:
expressions <- alist(-x ^ p, -(x) ^ p, (-x) ^ p, -(x ^ p)) x <- c(-5, -5, 5, 5) p <- c(2, 3, 2, 3) output <- data.frame(x, p, setNames(lapply(expressions, eval), sapply(expressions, deparse)), check.names = FALSE) print(output, row.names = FALSE) ## x p -x^p -(x)^p (-x)^p -(x^p) ## -5 2 -25 -25 25 -25 ## -5 3 125 125 125 125 ## 5 2 -25 -25 25 -25 ## 5 3 -125 -125 -125 -125
(stolen from my submission here). Did you notice that the output knew the names of
x
andp
without being told them? Did you also notice that a similar thing happened in after our call tocurve()
earlier on? Finally, did you notice how easy it was to get such neat output?
I’ve already admitted a great deal of ignorance of this topic, but there are some parts of R’s ecosystem that I’m happy to call outstanding. The below are all things that I’m sure to miss in other languages.
corrplot
: It has less than ten functions, but it only needed one to blow my mind. Once you’ve even as much as read the introduction, you will never try to read a correlation matrix again.ggplot2
: I’m not experienced enough to know what faults it has, but it’s fun to use. That single fact makes it blow any other graphing software that I’ve used out of the water: It’s fun.magrittr
: It sold me on pipes. I’d say that any package that makes you consider changing your programming style is automatically outstanding. However, the real reason why I love it is because whenever I’ve runbigLongExpression()
in my console and decided that I really wantedfoo()
of it, it’s so much easier to press the up arrow and type CTRL+SHIFT+M+“foo” than it is to do anything that results infoo(bigLongExpression())
appearing. Maybe there’s a keyboard shortcut that I never learned, but this isn’t the only reason why I lovemagrittr
. I’ll say more about it much later.R Markdown
has served me well in writing this document. It’s buggier than I’d like, rarely has helpful error messages, and does things that I can’t explain or fix even after setting a bounty on Stack Overflow, but it’s still a great way to make a document from R. It’s the closest thing that I know of to an R user’s LaTeX. I had to wait on this bug fix before I could start numbering my sections. Hopefully it didn’t break anything.
When it’s not causing you problems, the vectorization can be the best thing about the language:
-
The vector recycling rules are powerful when mastered. Expressions like
c("x", "y")[rep(c(1, 2), times = 4)]
let you do a lot with only a little work. My favourite ever FizzBuzz could well bex <- paste0(rep("", 100), c("", "", "Fizz"), c("", "", "", "", "Buzz")) cat(ifelse(x == "", 1:100, x), sep = "\n")
I wish that I could claim credit for that, but I stole it from an old version of this page and improved it a little.
-
Basically everything is a vector, so R comes with some great vector-manipulation tools like
ifelse()
(seen above) and makes it very easy to use a function on a collection. Can you believe thatmtcars / 20
actually works? -
Tricks like
array / seq_along(array)
save a lot of loop writing. -
Even simple things like being able to subtract a vector from a constant (e.g.
10 - 1:5
) and get a sensible result are a gift when doing mathematics. -
Vectorization of functions is sometimes very useful, particularly when it lets you do what should’ve been two loops worth of work in one line. You’d be amazed by how often you can get away with calling
foo(1:100)
without needing to vectorizefoo()
yourself.
R’s done a good job of harnessing the power of functional languages while maintaining a C-like syntax. It makes no secret of being inspired by Scheme and has reaped many of its benefits.
It’s impossible to not notice that functions are first-class in R. You’re almost forced to learn functional programming idioms like mapping functions, higher-order functions, and anonymous functions. This is a good thing. Where else do you find a language with enough useful higher-order functions for the community to be able to discourage new users from writing loops? Some examples:
- All of the functional programming toys that you could want are easily
found in R, e.g. closures, anonymous functions, and higher-order
functions like
Map()
,Filter()
, andReduce()
. Once you’re used to them, you can write some very expressive code. - The apply family of functions is basically a set of DSL mapping
functions for stats. Both
apply()
andtapply()
can produce some very concise code, as can related functions likeby()
. - Where else can you write functions that are both anonymous and recursive? Not that you should, of course.
- First-class functions sometimes interact with R’s vectorization
obsession in a very entertaining way. In how many other languages do
you see somebody take a list of functions and, in a single line, call
them all with a vector as a single argument to each function? Code
like
lapply(listOfFuns, function(f) f(1:10))
is entirely valid. It calls each function inlistOfFuns
with the entire vector1:10
as their first argument. - Code like
Vectorize(foo)(1:100)
is not particularly hard to understand, but I’d struggle to name another language that lets me do the same thing with so much ease.
Not only are functions first-class in R, environments are too. You therefore have lots of control over what environment an expression is evaluated in. This is an amazing source of power that tends to scare off beginners, but I cannot overstate how much of an asset it can be. If you’re not familiar with the below, look it up. You will not regret it.
- Because R’s environments are first-class, functions like
with()
andwithin()
can generate them on the fly. I’ve seen this called “data masking”. Advanced R has a whole chapter on it. It lets you do things like “treat this list of functions as if it were a namespace, so I can write code that uses function names that I wouldn’t dare use elsewhere”. This can also be used with data. For example,tapply(mtcars$mpg, list(mtcars$cyl, mtcars$gear), mean)
usesmtcars
far too many times, butwith(mtcars, tapply(mpg, list(cyl, gear), mean))
gives us an easy fix. Ad-hoc namespaces are an amazing thing to have, particularly when using functions that don’t have adata
argument (e.g.plot()
). - Modelling functions like
lm()
use the data-masking facilities that I’ve just described, as do handy functions likesubset()
. This saves incredible amounts of typing and massively increases the readability of your stats code. For example,aggregate(mpg ~ cyl + gear, mtcars, mean)
returns very similar output to my above calls totapply()
without needing the complexity of usingwith()
. It also allows for ridiculously concise code likeaggregate(. ~ cyl + gear, mtcars, mean)
. - You can write your own data-masking functions. Doing so relies on
controlling the non-standard evaluation of some of your arguments and
is the closest thing that R has to metaprogramming. The names
mechanisms do a lot to remove any ambiguity from your attempts at
this. Stealing an example from the documentation, do I even need to
explain what
transform(airquality, new = -Ozone, Temp = (Temp-32)/1.8)
does? Being able to do all of that in one line is outstanding. Without R allowing developers to add new functions like this, the Tidyverse would’ve been impossible.
You might have spotted a pattern by now. R often lets you do very much with very little.
Generic function OO is pretty nice to have, even if I wouldn’t use
anything more complicated than S3. Being able to call foo(whatever)
and be confident that it’s going to do what I mean is always nice. Some
positives of R’s approach are:
-
As mentioned earlier on, S3 is used excellently in the base R stats library. Functions like
print()
,plot()
, andsummary()
almost always tell me everything that I wanted to know and tell me them with great clarity. -
When you’re not trapped by the technicalities, S3 is an outstandingly simple tool that does exactly what R needs it to do. Have a look at all of the methods that the pre-loaded libraries define for
plot()
methods(plot) ## [1] plot.acf* plot.data.frame* plot.decomposed.ts* ## [4] plot.default plot.dendrogram* plot.density* ## [7] plot.ecdf plot.factor* plot.formula* ## [10] plot.function plot.hclust* plot.histogram* ## [13] plot.HoltWinters* plot.isoreg* plot.lm* ## [16] plot.medpolish* plot.mlm* plot.ppr* ## [19] plot.prcomp* plot.princomp* plot.profile.nls* ## [22] plot.raster* plot.spec* plot.stepfun ## [25] plot.stl* plot.table* plot.ts ## [28] plot.tskernel* plot.TukeyHSD* ## see '?methods' for accessing help and source code
because a statistician often only need to dispatch on the type of model being used, S3 is the perfect tool to make functions like
plot()
easy to extend, meaning that it’s easy to make it give your users exactly what they want. This isn’t just theoretical either. The output formethods(plot)
gets a lot longer if I go through my list of packages and start loading some random number of them. Go try it yourself! -
S3 generics and objects are very easy to write. The trade-off is that they don’t do anything to protect you from yourself. However, being able to tell R to shut up and do what I want it to a nice part of S3.
-
I like the idea of S3’s group generics, but I don’t like not being able to make my own. However, I think that you can do it for S4.
-
I have it on good authority that biology people often need to dispatch on more than one type of model at a time. This means that they shower the S4 object system with greater praise than what I’ve just given S3. Apparently, the
bioconductor
package is the outstanding example of their love of it. -
S4 has multiple inheritance and multiple dispatch. I’m not going to say that multiple inheritance is a good thing, but it’s not always found in other OOP systems.
-
RC and the
R6
package are about as close as you’re ever going to get to having Java-like OOP in a mostly function language.
Some of the syntax is nice:
- It can cause you problems, but the
:
operator is handy for things likefor(i in 1:20){...}
. - The
for
loop syntax is always the same:for(element in vector){...}
. This means that there is no difference between the typical “do n times” case likefor(i in 1:n)
and the “for every member of this collection” case likefor(v in sample(20))
. I appreciate the consistency. - The
...
notation has a very nice “do what I mean” feel, particularly when you’re playing around with anonymous functions. - Because of
repeat
loops, you never need to writewhile(TRUE)
. - Although I have major issues with them, the rules
for accessing elements sometimes give nice results. For example
array[c(i, j)] <- array[c(j, i)]
swaps elementsi
andj
in a very clean way. - It’s nice to be able to do many variable assignments in one line
e.g.
Alice <- Bob <- "Married"
. The best examples are when you do something likelastElement <- output[lastIndex <- lastIndex + 1] <- foo
, letting you avoid having to do anything twice. - The syntax for manipulating environments makes sense. You have to
learn the difference between
<-
and<<-
, but having environments use a subset of the list syntax was a very good idea. It was a similarly good idea to have a lot of R’s internals (e.g. quoted function calls) be pairlists. This lets them be manipulated in exactly the same way as lists. The similarities between lists, pairlists, environments, and data frames go deeper than you may expect. For example, theeval()
function lets you evaluate an expression in the specified environment, but it’s happy to take any of the data types that I’ve just listed in place of an environment. At times, R almost lets you forget that lists and environments aren’t the same thing. - The function names for making and manipulating S4 objects and
functions are generally what you would expect them to be. For example,
once you know
setClass()
andsetGeneric()
, you can probably guess what the corresponding function for methods is called.
- The built-in vectors
letters
andLETTERS
come in handy surprisingly often. You’ll see me use them a lot. - The base library surprises me from time to time. It’s always worth
putting what you want in to a search engine; Sometimes, you’ll find
it. My most recent surprises were
findInterval()
andcut()
. - The
na.print
argument toprint()
, trivial as it is, can be a thing of beauty. - R’s condition handling and recovery system is build atop S3, making it
extremely customisable by letting you add and handle custom metadata
pretty much however you want. It also has some nice built-in types of
conditions like errors, warnings, and messages, as well as having
finally
blocks intryCatch()
. The only real oddity of the system is that its conditionals are treated as functions of the error, meaning that you will have to write strange code liketryCatch(code, error = function(unused) "error", warning = function(unused) "warning")
. However, this is the price that you pay for being able to use code liketryCatch(code, myError = function(e) paste0(e$message, " was triggered by ",e$call,". Try ",e$suggestion)
. As a final point of interest, I’ve heard that R’s condition handling system is one of the best copies of Common Lisp’s, which I’ve heard awesome things about. - Speaking of Lisp, statistical Lisps used to be a thing. I’ve heard rumours of them still being used in Japan, but I can’t find anything to back that up. Everything that I’ve found says that R killed them off. As far as I know, nobody’s tried to make another such Lisp since the 90’s. The fact that R can claim to have eradicated an entire category of language design is a great point in its favour. It’s also possible evidence that I’m correct to say that R resembling C is to its benefit. However, I’d be overjoyed to hear of such Lisps making a comeback. Imagine if we’re just one good Clojure library away from R surrendering to its Scheme roots and birthing a modern statistical Lisp.
- Base R is quite stable. Breaking changes are almost unheard of. I don’t agree that they should be trying so hard to maintain compatibility with S, but this is an undeniable benefit of that decision.
- The fact that the Tidyverse is just an R library, rather than an entirely separate language, is a testament to R’s metaprograming. I like being able to define new infix operators and replacement functions, but the Tidyverse went above and beyond. Where else do you see an entire library of pipes? Until very recently, base R didn’t even have pipes!
- The Tidyverse is proof that people are trying to fix R. Although that comes with the implication that R is broken, the fact that people are both willing and able to fix it definitely says something nice about R.
- I generally like the RStudio IDE. When Emacs is the only alternative that anyone takes seriously, you know that you’ve done a good job.
- There is only one implementation of R that anyone’s ever heard of, so you never need to worry about undefined behaviour.
- There’s something hilarious about R being a language where dangerous
forbidden techniques actually run. In most other languages, comments
that read
# FORBIDDEN
would indicate that the code produces some sort of error. Not R.
This is where this documents starts to get long. Brace yourself. I really don’t want to give off the impression that I hate R, but there are just too many things wrong with it. Again, R’s ultimate problem is the sum of its small madnesses. No language is perfectly consistent or without compromises, but R’s choices of compromises and inconsistencies are utterly unpredictable. I could deal with a handful of problems like the many that will follow, but this is far more than a handful.
March 2023 update: It occurs to me that I don’t really have a summary of where R goes wrong. This is consistent with my above claim that R’s ultimate problem is the sum of its little problems. However, one pattern has become obvious both from learning other languages and from reading the section headings: R doesn’t have the right data structures; Roughly half of the subsections here are dedicated to complaining about them. This is no small complaint. One of the big rules in the Unix Philosophy is that data structures are central to programming. If your data structures are wrong, then finding the correct algorithm becomes much harder. It’s little wonder that a focus of the Tidyverse is to clean up one of R’s primary data structures (the data frame) and then stick to it as much as possible.
We’ll start gentle. R’s list type is an unavoidable part of the language, but it’s very strange. As the following examples show, it’s frequently a special case that you can rarely avoid.
-
https://stackoverflow.com/questions/2050790/ does a good job of demonstrating that the list type is not like anything that another language would prepare you for. It and its many answers are very much worth a read.
-
Lists are the parent class of data frames. Data frames are mandatory for anyone who wants to do stats in R and most of the problems with lists are inherited by data frames. This makes the oddities of lists unavoidable.
-
Particularly when extracting single elements of lists, you need to be vigilant for whether R is going to give what you wanted or the list containing what you wanted. Most of this comes down to learning the distinction between
[
and[[
andsapply()
andlapply()
. It’s not too bad, but it’s a complication. -
Because they won’t attempt to coerce your inputs to a common type and because, unless you count matrices, you cannot nest vectors (e.g.
c(c(1, 2), c(3, 4))
is1:4
), lists are what you’re most likely to use when you want to put two or more vectors in one place. A lot of your lists will therefore be nested structures. This is not inherently a problem, but extracting elements from nested structures is hard, both in a general sense and specifically for R’s lists. R does little to help you with this. Give https://stackoverflow.com/q/9624169/ and some of its answers a read. Why does this simple question get seven different answers? Do we really need libraries, anonymous functions, or knowing that[
is a function, just for what ought to be a commonplace operation? -
Some common R functions do not work properly with lists. Some functions like
sort()
andorder()
will not work at all, even if you list only contains numbers, and other will work but misbehave. For example, what do you expect to get fromc(someList, someVectorWithManyElements)
? You might expect a list that is now one item longer. Instead, you get your original list with every element of the vector appended to it in new slots, i.e. a list that islength(someVectorWithManyElements)
longer.c(list(1, 2, 3), LETTERS[1:5]) ## [[1]] ## [1] 1 ## ## [[2]] ## [1] 2 ## ## [[3]] ## [1] 3 ## ## [[4]] ## [1] "A" ## ## [[5]] ## [1] "B" ## ## [[6]] ## [1] "C" ## ## [[7]] ## [1] "D" ## ## [[8]] ## [1] "E"
The same output is given by
append()
. To getlist(1, 2, 3, LETTERS[1:5])
, you must do something likex <- list(1, 2, 3); x[[4]] <- LETTERS[1:5]
. -
Note the use of
[[4]]
and not[4]
above. Using[4]
gets you a warning and the outputlist(1, 2, 3, "A")
. The[
version is intended for cases likex[4:8] <- LETTERS[1:5]
, which gives the same output asc()
did above. The[
/[[
distinction is a beginner’s nightmare, as is R’s tendency to give you many ways to do the same thing. -
Primarily due to the commonality of data frames, R has a handful of functions that are essentially “foo, but the list version”.
lapply()
is the most obvious example. -
A few functions, such as
strsplit()
, can catch you off guard by returning a list when there’s no obvious reason why a vector or matrix wouldn’t have done. Forstrsplit()
in particular, I think that the idea is that it’s designed to be used on character vectors of lengths greater than one. However, in my experience, I almost always want a length-one version. I’d far rather have that function andlapply()
/sapply()
/whatever it as need be than have to constantly usestrsplit("foo")[[1]]
. Similarly, some functions, e.g.merge()
, insist on returning data frames even when the inputs were matrices. Coercing these unwanted outputs in to what you actually wanted is often harder than it has any right to be.
I think that the ultimate problem with lists is that the right way to use them is not easy to guess from your knowledge of the language’s other constructs. If everything in R worked like lists do, or if lists weren’t so common, then you wouldn’t really mind. As it is, you’ll often make mistakes with lists and have to guess your way through correcting them. This isn’t terrible. It’s just annoying.
R’s strings suck. The overarching problem is that because there is no
language-level distinction between characters vectors and their
individual elements, R’s vectorization means that almost everything that
you want to do with a string needs to be done by knowing the right
function to use (rather than by using R’s ordinary syntax). I find that
the correct functions can be hard to find and use. Although it doesn’t
fix many of these issues, the common sentiment of “just use
stringr
/stringi
” is difficult to dismiss.
-
Technically, R doesn’t even have a type for strings. You would want a string to be a vector of characters, but R’s characters are already vectors, so R can’t have a normal string type. Despite this, the documentation often uses the word “string”. The language definition will tell you how to make sense of that, but I don’t think that information is found anywhere in the sorts of documentation that you’ll find in your IDE.
-
It’s a pain to have to account for how R has two types of empty string:
character(0)
and""
. -
Character vectors aren’t trivially split in to the characters that make each element. For example,
"dog"[1]
is"dog"
because"dog"
is a vector of length one. The idiomatic way to split up a string in to its characters –strsplit("dog", "")
– returns a list, so rather than just getting the"d"
from"dog"
by doing"dog"[1]
, you have to do something likeunlist(strsplit("dog", ""))[1]
orstrsplit("dog", "")[[1]][1]
. Thesubstr()
function can serve you better for trivial jobs, but you often needstrsplit()
. -
Here’s a challenge: Find the function that checks if
"es"
is in"test"
. You’ll be on for a while. -
Many of R’s best functions for handling strings expect you to know regex and are all documented in the same place (grep {base}, titled “Pattern Matching and Replacement”). If you don’t know regex – exactly what I’d expect of R’s target audience – then you’re thrice damned:
- Firstly, you’re going to have a very hard time figuring out that the
functions in
?grep
are probably what you need. A glance at their documentation suggests that they’re difficult materials and therefore presumably not required for your task. - Secondly, you’re going to struggle to find the correct function in that documentation because the information that you need is surrounded by concepts that are not familiar to you. This reinforces your initial impression that you’re in the wrong place, letting you stumble over the first point again.
- Thirdly and finally, the function names will be meaningless to you –
“what the heck does
regexpr()
mean and how does it relate toregexec()
?” – leaving you with no straws to clutch at.
- Firstly, you’re going to have a very hard time figuring out that the
functions in
-
The right function for the job can still be tough to use. Compare the
stringr
answer to this question to the base R answers. Or better yet, usegregexpr()
orgregexec()
for any task and then tell me with a straight face that you both understand their outputs and find them easy to work with.gregexpr("a", c("greatgreat", "magic", "not")) ## [[1]] ## [1] 4 9 ## attr(,"match.length") ## [1] 1 1 ## attr(,"index.type") ## [1] "chars" ## attr(,"useBytes") ## [1] TRUE ## ## [[2]] ## [1] 2 ## attr(,"match.length") ## [1] 1 ## attr(,"index.type") ## [1] "chars" ## attr(,"useBytes") ## [1] TRUE ## ## [[3]] ## [1] -1 ## attr(,"match.length") ## [1] -1 ## attr(,"index.type") ## [1] "chars" ## attr(,"useBytes") ## [1] TRUE
-
The most useful function for printing strings seem to counter-intuitively be
cat()
rather thanprint()
orformat()
. For example,print()
ignores your\n
characters. The only time whereprint()
comes in really handy for string-related stuff is when your inputs are either quoted or lists. In both cases,print()
accepts these butcat()
does not. Without significant coercion (mostlydeparse()
), I’ve yet to find a way to mix\n
with quoted input. Most of my attempts to do anything clever that mixes printing and lists end with me giving up and using a data frame. -
Without defining a new operator, you can’t add strings in the way that other languages have taught you to, i.e.
"a"+"b"
. John Chambers is against fixing this. I’m not convinced that he’s wrong, but it is annoying. -
If you’re converting numbers to characters, or using a function like
nchar()
that’s meant for characters but accepts numbers, a shocking number of things will break when your numbers get big enough for R to automatically start using scientific notation.nchar(10000) ## [1] 5 nchar(100000) ## [1] 5 a <- 10000 nchar(a) == nchar(a * 100) ## [1] TRUE
You’re supposed to use
format()
to coerce these sorts of numbers in to characters, but you won’t know about that until something breaks andnchar()
’s documentation doesn’t seem to mention it (try?as.character
). Theformat()
function also has a habit of requiring you to set the right flags to get what you want.trim = TRUE
comes up a lot. If you’re using a package or unfamiliar function, you’re forced to check to see if the author dealt with these issues before you use their work. I’d rather just have a genericnchar()
-like function that does what I mean. Would you believe thatnchar()
’s documentation says it’s a generic? It’s not lying and it later tells you thatnchar()
coerces non-characters to characters, but R sure does know how to mess with your expectations.
R has some problems with its general facilities for manipulating variables. Some of the following will be seen every time that you use R.
-
It lacks both
i++
andx += i
. It also lacks anything that would make these unnecessary, such as Python’senumerate
. -
One day, you’ll be tripped up by R’s hierarchy of how it likes to simplify mixed types outside of lists. The basics are documented with the
c()
function. For example,c(2, "2")
returnsc("2", "2")
. An exercise from Advanced R presents a few troubling cases:- “Why is
1 == "1"
true?” - “Why is
-1 < FALSE
true?” - “Why is
"one" < 2
false?”.
- “Why is
-
To get complete information about the typing and structure of something, you will almost certainly need to call several functions. For example, do any of the following tell you everything about
x
?x <- diag(3) x ## [,1] [,2] [,3] ## [1,] 1 0 0 ## [2,] 0 1 0 ## [3,] 0 0 1 typeof(x) ## [1] "double" class(x) ## [1] "matrix" "array" attributes(x) ## $dim ## [1] 3 3 str(x) ## num [1:3, 1:3] 1 0 0 0 1 0 0 0 1 dput(x) #Dirty trick, don't use in practice. ## structure(c(1, 0, 0, 0, 1, 0, 0, 0, 1), dim = c(3L, 3L))
Among these,
str()
is the closest. However, you can see that it doesn’t give you all of the class information. This doesn’t improve if you add non-implicit classes tox
, but I’m avoiding that topic for as long as I can. -
R likes to use “double” and “numeric” almost interchangeably. You’ve just seen one such example (
str(x)
vstypeof(x)
). -
Integers are almost second class.
?integer
suggests that they’re mostly for talking to other languages, but the problem seems to go deeper than that. It’s as if R tries to avoid integers unless you tell it not to. For example,4
is a double, not an integer. Why? Unless you’re very careful, any integer that you give to R will eventually be coerced to a double. -
There’s no trivial way to express a condition like 1 < x < 5. In a maths language, I’d expect that exact syntax to work. There’s probably a good reason why it doesn’t, and it’s not at all hard to build an equivalent condition, but it still annoys me from time to time. I suspect that the
<-
syntax is to blame. -
The distinction between
<-
and=
is something that you’d have to look up. I’d try to explain the difference, but from what I’ve gathered, the difference only matters when using=
rather than<-
causes bugs. Like most R users, I’ve picked up the habit of “use=
only for the arguments of functions and use<-
everywhere else”. -
<-
was designed for keyboards that don’t exist any more. It’s a pain to type on a modern system. IDEs can fix this. -
The day that you accidentally have
<
rather than<-
without it throwing an error will be an awful one. The reverse can also happen. For example, there are two things that you could have meant byif(x<-2)
. -
Y<--2
is a terrible way to have to say “setY
to be equal to negative two”.Y<<--2
is even worse. -
<<-
is probably the only good thing about the convention of using<-
, but it’s only useful if you either know what a closure is and have reason to use one or if you’re some sort of guru with R’s first-class environments. You can sometimes use<<-
to great effect without deliberately writing a closure, but it always feels like a hack because you’re implicitly using one. For example,replicate(5, x <- x+1)
andreplicate(5, x <<- x+1)
are very different, with the<<-
case being a very cool trick,x <- 1 replicate(5, x <- x+1) ## [1] 2 2 2 2 2 x ## [1] 1 replicate(5, x <<- x+1) ## [1] 2 3 4 5 6 x ## [1] 6
but it only works because
replicate()
quietly wraps its second argument in an anonymous function. -
The idiomatic way to add an item to the end of a collection is
a[length(a) + 1] <- "foo"
. This is rather verbose and a bit unpredictable when adding a collection to a list. -
A quote from the language definition: “supplied arguments and default arguments are treated differently”. This usually doesn’t trip you up, but you’re certain to discover it on the first day that you use
eval()
. It hasparent.frame()
as one of its default arguments, but callingeval()
with that argument supplied manually will produce different results than letting it be supplied by default.x <- 1 (function(x) eval(quote(x + 1)))(100) ## [1] 101 (function(x) eval(quote(x + 1), envir = parent.frame()))(100) ## [1] 2
An easier-to-discover example can be found in section 8.3.19 of The R Inferno.
-
Argument names can be partially matched. See this link for some examples. I can’t tell if it’s disgusting or awesome, but it’s definitely dangerous. If I called
f(n = 1)
, I probably didn’t meanf(nukeEarth = 1)
! At least it throws an error if it fails to partially match (e.g. if there were multiple valid partial matches). More on that when I cover the$
operator. -
The
...
argument doesn’t make its users throw errors when they’ve been called with arguments that they don’t have or, even worse, those you misspelled. Advanced R has a great example in its Functions chapter. Would you have guessed thatsum(1, 2, 3, na.omit = TRUE)
returns7
, not6
? Similarly, the Functionals chapter shows how this can lead to baffling errors and what strange things you have to do help your users avoid them. -
NaN
,NULL
, andNA
have been accused of inconsistencies and illogical outputs, making it impossible to form a consistent mental model of them. -
For many other examples, see section 8 of The R Inferno.
R has some strange ideas about switch statements:
-
It’s not a special form of any kind; Its syntax is that of a normal function call. If I’m being consistent in my formatting, then I should be calling it “R’s
switch()
function”. -
It’s only about 70 lines of C code, suggesting that it can’t be all that optimised.
-
R doesn’t have one switch statement, it has two. There is one where it switches on a numeric input and another for characters. The numeric version makes the strange assumption that the first argument (i.e. the argument being switched on) can be safely translated to a set of cases that must follow an ordering like “if input is 1, do the first option, if 2, do the second…”. There is no flexibility like letting you start at 2, having jumps higher than 1, or letting you supply a default case. Reread that last one: R has a switch without defaults! It’s frankly the worst switch that I’ve ever seen. The other version, the one that switches on characters, is more sensible. I’d give examples, but I don’t know how to demonstrate a non-feature.
-
As is a trend in R, both versions of switch are capable of silently doing nothing. For example, these do nothing:
switch(3, "foo", "bar") switch("c", a = "foo", b = "bar") print(switch("c", a = "foo", b = "bar")) #Showing off the return value. ## NULL
and they do it silently, returning
NULL
. I’d expect a warning message informing me of this, but there is no such thing. If you want that behaviour, then you have to write it yourself e.g.switch("c", a = "foo", b = "bar", stop("Invalid input"))
orswitch("c", a = "foo", b = "bar", warning("Invalid input"))
. You can’t do that with the numeric version, because R has a switch without defaults.
Now for the nasty stuff. R’s rules for selecting elements, deleting elements, and any other sort of subsetting require mastery. They’re extremely powerful once mastered, but until that point, they make using R a nightmare. For a stats language, this is unforgivable. Before mentioning any points that are best put in their own subsections, I’ll cover some more general points:
-
You never quite know whether you want to use
x
, the name ofx
, or a function likesubset()
,which()
,Find()
,Position()
, ormatch()
. R’s operators make this even more of mess. You either want$
,[
,@
or even[[
. Making the wrong choice of[x]
,[x,]
,[,x]
or[[x]]
is another frequent source of error. You will get used to it eventually, but your hair will not survive the journey. Similar stories can be found about the apply family. -
The
[[
operator has been accused of inconsistent behaviour. Advanced R covers this better than I could. The short version is that it sometimes returnsNULL
and other times throws an error. Personally, I’ve never noticed these because I’ve rarely tried to subsetNULL
and I don’t see any reason why you would use[[
on an atomic vector. As far as I know,[
does the same job. The only exception that I can think of is if your atomic vector was named. For example:a <- c(Alice = 1, Bob = 2) a["Alice"] ## Alice ## 1 a[["Alice"]] ## [1] 1
-
When doing variable assignment on anything more than one-dimensional,
object
andobject[]
behave differently when you try to assign variables to them. Compare:(c <- b <- diag(3)) ## [,1] [,2] [,3] ## [1,] 1 0 0 ## [2,] 0 1 0 ## [3,] 0 0 1 b[] <- 5 c <- 5 b ## [,1] [,2] [,3] ## [1,] 5 5 5 ## [2,] 5 5 5 ## [3,] 5 5 5 c ## [1] 5
This kind of makes sense, but it will trip you up.
-
The syntax for deleting elements of collections by index can be rather verbose. You can’t just pop out an element, you have to write
vect <- vect[-index]
orvect <- vect[-c(index, nextIndex, ...)]
. -
R is 1-indexed, but accessing element 0 of a vector gives an empty vector rather than an error. This probably makes sense considering that index -1 deletes element 1, but it’s a clear source of major errors.
-
With the sole exception of environments, every named object in R is allowed to have duplicate names. I guarantee that will one day break your subsetting (e.g. see section 8.1.19 of The R Inferno). Fortunately, the constructor for data frames has a
check.names
argument that corrects duplicates by default. Unfortunately, it does this silently, so you may not notice that some of your column names have been changed. Another oddity is that many functions that work on data frames, most notably[
, will silently correct duplicated names even if you told the original data frame to not do so. Why even let me have duplicated names if you’re going to make it so hard to keep them?data.frame(x = 1:3, x = 11:13) ## x x.1 ## 1 1 11 ## 2 2 12 ## 3 3 13 #Notice the x.1? You didn't ask for that. To get x twice, you need this: correctNames <- data.frame(x = 1:3, x = 11:13, check.names = FALSE) correctNames ## x x ## 1 1 11 ## 2 2 12 ## 3 3 13 correctNames[1:3, ]#As expected. ## x x ## 1 1 11 ## 2 2 12 ## 3 3 13 correctNames[1:2]#What? ## x x.1 ## 1 1 11 ## 2 2 12 ## 3 3 13
Not only is this behaviour inconsistent, it is silent; No warnings or errors are thrown by the above code. Tibbles are much better about this:
library(tibble) #We can't repeat our original first line, because tibble(x = 1:5, x = 11:15) throws an error: ## > tibble(x = 1:5, x = 11:15) ## Error: Column name `x` must not be duplicated. ## Use .name_repair to specify repair. #We follow the error's advice. #The .name_repair argument provides a few useful options, so we must pick one. correctNames <- tibble(x = 1:5, x = 11:15, .name_repair = "minimal") correctNames ## # A tibble: 5 × 2 ## x x ## <int> <int> ## 1 1 11 ## 2 2 12 ## 3 3 13 ## 4 4 14 ## 5 5 15 correctNames[1:3,]#Good ## # A tibble: 3 × 2 ## x x ## <int> <int> ## 1 1 11 ## 2 2 12 ## 3 3 13 correctNames[1:2]#Still good! ## # A tibble: 5 × 2 ## x x ## <int> <int> ## 1 1 11 ## 2 2 12 ## 3 3 13 ## 4 4 14 ## 5 5 15
This may seem like an isolated example. It isn’t. A related example is that the
check.names
argument indata.frame()
is very insistent on silently doing things, even to the point of overruling arguments that you explicitly set. For example, these column names aren’t what I asked for.as.data.frame(list(1, 2, 3, 4, 5), col.names = paste("foo=bar", 6:10)) ## foo.bar.6 foo.bar.7 foo.bar.8 foo.bar.9 foo.bar.10 ## 1 1 2 3 4 5 as.data.frame(list(1, 2, 3, 4, 5), col.names = paste("foo=bar", 6:10), check.names = FALSE)#The fix. ## foo=bar 6 foo=bar 7 foo=bar 8 foo=bar 9 foo=bar 10 ## 1 1 2 3 4 5
I think R does this to ensure your names are suitable for subsetting. Subsetting with non-unique or non-syntactic column names could be a pain, but the decision to not inform the user of this correction is baffling. Even if you’re fortunate enough to notice the silent changes, the lack of a warning message will leave you with no idea how to correct them. You could perhaps argue that duplicated names are the user’s fault and they deserve what they get, but that argument falls apart for non-syntactic names. Who hasn’t put a space or an equals sign in their column names before? Mind, tibbles aren’t much better when it comes to non-syntactic names. Neither
tibble("Example col" = 4)
nordata.frame("Example col" = 4)
warn you of the name change. -
For what I believe to be memory reasons, objects of greater than one dimension are stored in column order rather than row order. Quick, what output do you expect from
matrix(1:9, 3, 3)
?matrix(1:9, 3, 3) ## [,1] [,2] [,3] ## [1,] 1 4 7 ## [2,] 2 5 8 ## [3,] 3 6 9
This gives us a matrix with first row
c(1, 4, 7)
. This goes against the usual English convention of reading from left to right. It is also inconsistent with functions likeapply()
, whereMARGIN = 1
corresponds to their by-row version andMARGIN = 2
is for by-column (if R privileges columns, shouldn’t they be the= 1
case?). This means that you can never really be sure if R is working in column order or row order. This is bad enough on its own, but it can also be a source of subtle bugs when working with matrices. Many mathematical functions don’t see any difference between a matrix and its transpose. -
There is no nice way to access the last element of a vector. The idiomatic way is
x[length(x)]
. The only good part of this is thatx[length(x) - 0:n]
is a very nice way to get the lastn + 1
elements. You could usetail()
, but Stack Overflow tells me it’s very slow. -
The
sort()
andorder()
functions are the main ways to sort stuff in R. If you’re trying to sort some data by a particular variable, then R counter-intuitively wants you to useorder()
rather thansort()
. The syntax fororder()
doesn’t help matters. It returns a permutation, so rather thanorder(x, params)
, you will wantx[order(params),]
. My only explanation for this is that it makesorder()
much easier to use with thewith()
function. For example,data[with(data, order(col1, col2, col3)),]
is perhaps more pleasant to write than the hypotheticalorder(data, data$col1, data$col2, data$col3)
. The Tidyverse’sdplyr
solves these problems:dplyr::arrange(data, col1, col2, col3)
does what you think. I’d much rather usearrange(mtcars, cyl, disp)
thanmtcars[with(mtcars, order(cyl, disp)),]
. -
The
order()
case above illustrates another frequent annoyance with subsetting. Rather than asking for what you want, you often need to generate a vector that matches up to it. A collection of booleans (R calls these “logical vectors”) is one of the most common ways to do this, withduplicated()
being a typical example.head(Nile) ## [1] 1120 1160 963 1210 1160 1160 duplicated(head(Nile)) ## [1] FALSE FALSE FALSE FALSE TRUE TRUE head(Nile)[duplicated(head(Nile))] ## [1] 1160 1160
This means that you will usually be asking for
items[bools]
(and maybe[,bools]
or[bools,]
…) in order to get the items that you want. There is great power in being able to do this, but having to do it is annoying and can catch you off guard. For example, what do you expectlower.tri()
to return when called on a matrix? What you wanted fromlower.tri(mat)
is probably what you get frommat[lower.tri(mat)]
. Also, don’t expect a helpful error message if your construction ofbools
is wrong. As I’ll discuss later on, the vector recycling rules will often make an incorrect construction give sensible-looking output. -
For reasons that I cannot explain,
aperm(x, params)
is the correct syntax, notx[aperm(params)]
or anything like it. I think that it’s trying to be consistent with R’s ideas of how to manipulate matrices, but it’s yet another source of confusion. I don’t want to have to think about if I’m treating my data like a matrix or like a data frame. -
Good luck trying to figure out how to find a particular sequence of elements within a vector. For example, try finding if/where the unbroken vector
1:3
has occurred insample(6, 100, replace = TRUE)
. You’re best off just writing thefor
loop.
This one isn’t too bad, but it’s worth a mention. Combining operations can lead to some counter-intuitive results:
-
If
a <- 1:5
, what do you expect to get froma[-1] <- 12:15
? Do you expecta[1]
to be removed or not? This is great once you know how it works, but it’s confusing to a beginner.a <- 1:5 a[-1] <- 12:15 a ## [1] 1 12 13 14 15
-
Because
data[-index]
can be used to remove elements anddata["colName"]
can be used to select elements, you might expectdata[-"colName"]
ordata[-c("colName1", "colName2")]
to work. You would be wrong. Both throw errors.## > mtcars[-"wt"] ## Error in -"wt" : invalid argument to unary operator
-
Attempting to remove both by index and by name at the same time will never work. For example,
mtcars[-c(1, "cyl")]
is an error andmtcars[c(1, "cyl")] <- NULL
will only remove thecyl
variable. Weirdly enough, I can’t actually show thismtcars[c(1, "cyl")] <- NULL
example. R is perfectly happy to show it, but R Markdown isn’t. What happens is thatc(1, "cyl")
is coerced toc("1", "cyl")
. After this, R does not inform you that there is no1
column to remove.
Now for the serious stuff…
This issue is notorious: R likes to remove unnecessary dimensions from your data in ways that are not easily predicted, forcing you to waste time preventing them. Rumour has it that this can be blamed on S being designed as a calculator rather than as a programming language. I can’t cite that, but it’s easy to believe. No programmer would include any of the below in a programming language.
-
Unless you add
, drop=FALSE
to all of your data selection/deletion lines, you run the risk of having all of your code that expects your data to have a particular structure unexpectedly break. This gives no errors or warnings. Compare:(mat <- cbind(1:4, 4:1)) ## [,1] [,2] ## [1,] 1 4 ## [2,] 2 3 ## [3,] 3 2 ## [4,] 4 1 mat[, -1] ## [1] 4 3 2 1 mat[, -1, drop=FALSE] ## [,1] ## [1,] 4 ## [2,] 3 ## [3,] 2 ## [4,] 1
and you will see that one of these is not a matrix. Data frames have the same issue unless you do all of your subsetting in a 1D form.
mat <- cbind(1:4, 4:1) (frame <- as.data.frame(mat)) ## V1 V2 ## 1 1 4 ## 2 2 3 ## 3 3 2 ## 4 4 1 frame[, -1] ## [1] 4 3 2 1 frame[, -1, drop=FALSE] ## V2 ## 1 4 ## 2 3 ## 3 2 ## 4 1 frame[-1]#1D subsetting ## V2 ## 1 4 ## 2 3 ## 3 2 ## 4 1
The Tidyverse, specifically
tibble
, does its best to remove this.library(tibble) mat <- cbind(1:4, 4:1) (tib <- as_tibble(mat)) ## Warning: The `x` argument of `as_tibble.matrix()` must have unique column names if ## `.name_repair` is omitted as of tibble 2.0.0. ## ℹ Using compatibility `.name_repair`. ## # A tibble: 4 × 2 ## V1 V2 ## <int> <int> ## 1 1 4 ## 2 2 3 ## 3 3 2 ## 4 4 1 tib[, -1] ## # A tibble: 4 × 1 ## V2 ## <int> ## 1 4 ## 2 3 ## 3 2 ## 4 1 tib[, -1, drop=FALSE] ## # A tibble: 4 × 1 ## V2 ## <int> ## 1 4 ## 2 3 ## 3 2 ## 4 1 tib[-1] ## # A tibble: 4 × 1 ## V2 ## <int> ## 1 4 ## 2 3 ## 3 2 ## 4 1 tib[, -1, drop=TRUE] ## [1] 4 3 2 1
You can think of tibbles as having
drop=FALSE
as their default. I can’t explain why base R doesn’t do the same. It’s got to either be some sort of compromise for matrix algebra or for making working in your console nicer.- Update: Section 6.8 of the Software for Data Analysis: Programming
with R book by John Chambers offers a partial explanation: “The
default is, and always has been,
drop=TRUE
; probably an unwise decision on our part long ago, but now one of those back-compatibility burdens that are unlikely to be changed.”
- Update: Section 6.8 of the Software for Data Analysis: Programming
with R book by John Chambers offers a partial explanation: “The
default is, and always has been,
-
The
drop
argument is even stranger than I’m letting on. Its defaults differ depending on whether there may only be one column remaining or if there may only be one row. To quote the documentation (?"[.data.frame"
): “The default is to drop if only one column is left, but not to drop if only one row is left”. Unlike the previous point, I can sort of make sense of this. For example, a single column can only ever be one type (even if that may be a container for mixed types, such as a list) but a single row could easily be a mix of types. Dropping on a row of mixed types will just give you a really ugly list, so you’d much rather have a data frame. With a column, it’s only with years of experience that the community has realised that they probably still want the data frame; It’s nowhere near as obvious that the vector is not preferable. -
As you can tell by taking a close look at the documentation for
[
and that of[.data.frame
, thedrop
argument does not do the same thing for arrays and matrices as it does for data frames. This means that my earlier example could be dishonest. However, the confusion that you would need to overcome in order to check for if I’ve been dishonest is so great that it proves that there’s definitely something wrong with thedrop
argument. -
You may think that
object
andobject[,]
are the same thing. They are not. You would expect and get an error ifobject
is one-dimensional. However, if it’s a data frame or matrix with one of its dimensions having size 1, then you do not get an error and bothobject
andobject[,]
are very different.library(tibble) colMatrix <- matrix(1:3) colMatrix ## [,1] ## [1,] 1 ## [2,] 2 ## [3,] 3 colMatrix[,] ## [1] 1 2 3 rowMatrix <- matrix(1:3, ncol = 3) rowMatrix ## [,1] [,2] [,3] ## [1,] 1 2 3 rowMatrix[,] ## [1] 1 2 3 colFrame <- as.data.frame(colMatrix) colFrame ## V1 ## 1 1 ## 2 2 ## 3 3 colFrame[,] ## [1] 1 2 3 rowFrame <- as.data.frame(rowMatrix) rowFrame ## V1 V2 V3 ## 1 1 2 3 rowFrame[,] ## V1 V2 V3 ## 1 1 2 3 colTib <- as_tibble(colMatrix) colTib ## # A tibble: 3 × 1 ## V1 ## <int> ## 1 1 ## 2 2 ## 3 3 colTib[,] ## # A tibble: 3 × 1 ## V1 ## <int> ## 1 1 ## 2 2 ## 3 3 rowTib <- as_tibble(rowMatrix) rowTib ## # A tibble: 1 × 3 ## V1 V2 V3 ## <int> <int> <int> ## 1 1 2 3 rowTib[,] ## # A tibble: 1 × 3 ## V1 V2 V3 ## <int> <int> <int> ## 1 1 2 3
Can you guess why? It’s because the use of
[
makes R check if it should be dropping dimensions. This makesobject
andobject[,,drop=FALSE]
equivalent, whereasobject[,]
is a vector rather than whatever it was originally. Tibbles, of course, don’t have this issue. -
If you’ve struggled to read this section, then you’re probably now aware of another point: It’s really easy to get the commas for
drop=FALSE
mixed up. What do you thinkdata[4, drop=FALSE]
is? Ifdata
is a data frame, you get column 4 and a warning that thedrop
argument was ignored. Did you expect row 4? Whether you did or not, you should be able to see why somebody may come to the opposite answer. Although I see no sensible alternative, thedrop
argument needing its own comma is terrible syntax for a language where a stray comma is the difference between your data’s life and death. This is made even worse by the syntax for[
occasionally needing stray commas. Expressions likedata[4,]
are commonplace in R, so it’s far too easy to forget that you needed the extra comma for thedrop
argument.
The $
operator is both silently hazardous and redundant:
-
As an S3 generic, you can never be certain that
$
does what you want it to when you use it on a class from a package. For example, it’s common knowledge that base R’s$
and the Tidyverse’s$
are not the same thing. In fact,$
does not even behave consistently in base R. Compare the following partial matching behaviour:library(tibble) list(Bob = 5, Dobby = 7)$B ## [1] 5 env <- list2env(list(Bob = 5, Dobby = 7)) env$B ## NULL data.frame(Bob = 5, Dobby = 7)$B ## [1] 5 tibble(Bob = 5, Dobby = 7)$B ## Warning: Unknown or uninitialised column: `B`. ## NULL
For what it’s worth, replacing
Dobby
withBobby
gives more consistent results.library(tibble) list(Bob = 5, Bobby = 7)$B ## NULL env <- list2env(list(Bob = 5, Bobby = 7)) env$B ## NULL data.frame(Bob = 5, Bobby = 7)$B ## NULL tibble(Bob = 5, Bobby = 7)$B ## Warning: Unknown or uninitialised column: `B`. ## NULL
In theory, I should note that
[
and[[
are also S3 generics and therefore should share this issue. Aside from thedrop
issues above, I rarely notice such misbehaviour in practice. -
Consistency aside, partial matching is inherently dangerous.
data$Pen
might give thePenetration
column if you forgot that you removed thePen
column. By default, R does not give you any warnings when partial matches happen, so you won’t have any idea that you got the wrong column. -
The documentation for
$
points out its redundancy in base R: “x$name
is equivalent tox[["name", exact = FALSE]]
”. In other words, even if I want the behaviour of$
, I can get it with[[
. Another benefit of[[
is that it will only partially match if you tell it to (useexact = FALSE
). That matters because… -
The partial matching of
$
can be even worse than I’ve just described. If there are multiple valid partial matches, rather than get any of them, you getNULL
. This is what happened with the Bob/Bobby example above. To give another example,mtcars$di
andmtcars$dr
both give sensible output because there is only one valid partial match, butmtcars$d
is justNULL
. I’m largely okay with this behaviour, but you don’t even get a warning!mtcars$di ## [1] 160.0 160.0 108.0 258.0 360.0 225.0 360.0 146.7 140.8 167.6 167.6 275.8 ## [13] 275.8 275.8 472.0 460.0 440.0 78.7 75.7 71.1 120.1 318.0 304.0 350.0 ## [25] 400.0 79.0 120.3 95.1 351.0 145.0 301.0 121.0 mtcars$dr ## [1] 3.90 3.90 3.85 3.08 3.15 2.76 3.21 3.69 3.92 3.92 3.92 3.07 3.07 3.07 2.93 ## [16] 3.00 3.23 4.08 4.93 4.22 3.70 2.76 3.15 3.73 3.08 4.08 4.43 3.77 4.22 3.62 ## [31] 3.54 4.11 mtcars$d ## NULL
-
Tibbles try to fix the partial-matching issues of
$
by completely disallowing partial matching. They will not partially match even if you tell them to with[[, exact=FALSE]]
. If you try to partially match anyway, it will give you a warning and returnNULL
. I sometimes wonder if it should be an error.library(tibble) mtTib <- as_tibble(mtcars) mtTib$di ## Warning: Unknown or uninitialised column: `di`. ## NULL mtTib$dr ## Warning: Unknown or uninitialised column: `dr`. ## NULL mtTib$d ## Warning: Unknown or uninitialised column: `d`. ## NULL mtcars[["d", exact = FALSE]] ## NULL mtTib[["d", exact = FALSE]] ## Warning: `exact` ignored. ## NULL
-
On the base R side, there is a global option that makes
$
give you warnings whenever partial matching happens. It’s disabled by default. Common sense suggests it should be otherwise. -
The
$
operator is another case of R quietly changing your data structures. For example, I would callmtcars$mpg
unreadable.mtcars$mpg ## [1] 21.0 21.0 22.8 21.4 18.7 18.1 14.3 24.4 22.8 19.2 17.8 16.4 17.3 15.2 10.4 ## [16] 10.4 14.7 32.4 30.4 33.9 21.5 15.5 15.2 13.3 19.2 27.3 26.0 30.4 15.8 19.7 ## [31] 15.0 21.4 typeof(mtcars$mpg) ## [1] "double"
You probably wanted
mtcars["mpg"]
mtcars["mpg"] ## mpg ## Mazda RX4 21.0 ## Mazda RX4 Wag 21.0 ## Datsun 710 22.8 ## Hornet 4 Drive 21.4 ## Hornet Sportabout 18.7 ## Valiant 18.1 ## Duster 360 14.3 ## Merc 240D 24.4 ## Merc 230 22.8 ## Merc 280 19.2 ## Merc 280C 17.8 ## Merc 450SE 16.4 ## Merc 450SL 17.3 ## Merc 450SLC 15.2 ## Cadillac Fleetwood 10.4 ## Lincoln Continental 10.4 ## Chrysler Imperial 14.7 ## Fiat 128 32.4 ## Honda Civic 30.4 ## Toyota Corolla 33.9 ## Toyota Corona 21.5 ## Dodge Challenger 15.5 ## AMC Javelin 15.2 ## Camaro Z28 13.3 ## Pontiac Firebird 19.2 ## Fiat X1-9 27.3 ## Porsche 914-2 26.0 ## Lotus Europa 30.4 ## Ford Pantera L 15.8 ## Ferrari Dino 19.7 ## Maserati Bora 15.0 ## Volvo 142E 21.4 typeof(mtcars["mpg"]) ## [1] "list"
and you definitely did not want
mtcars[, "mpg"]
ormtcars[["mpg"]]
, which both give the same output as using$
.mtcars[, "mpg"] ## [1] 21.0 21.0 22.8 21.4 18.7 18.1 14.3 24.4 22.8 19.2 17.8 16.4 17.3 15.2 10.4 ## [16] 10.4 14.7 32.4 30.4 33.9 21.5 15.5 15.2 13.3 19.2 27.3 26.0 30.4 15.8 19.7 ## [31] 15.0 21.4 mtcars[["mpg"]] ## [1] 21.0 21.0 22.8 21.4 18.7 18.1 14.3 24.4 22.8 19.2 17.8 16.4 17.3 15.2 10.4 ## [16] 10.4 14.7 32.4 30.4 33.9 21.5 15.5 15.2 13.3 19.2 27.3 26.0 30.4 15.8 19.7 ## [31] 15.0 21.4
Would you have guessed that? Tibbles share the above behaviour with
$
and[[
, but keep["name"]
and[, "name"]
identical due to their promise to not drop dimensions with[
. -
The
$
operator does not have any uses beyond selection. For example, there is no way to combine$
with operators like-
and there’s no way to pass arguments likedrop=FALSE
to it. -
$
is not allowed for atomic vectors likec(fizz=3, buzz=5)
, unlike[
and[[
. This is particularly annoying when dealing with named matrices because you end up having to usemat[, "x"]
wheremat$x
should have done. -
Section 8.1.21 of The R Inferno: There exists a
$<-
operator. You hardly ever see it used. The R Inferno points out that it does not do partial matching, even for lists, unlike$
. This is actually documented behaviour – in fact,?Extract
mentions it twice – but I challenge you to find it. I can see why it would be difficult to make a$<-
with partial matching, but making$<-
inconsistent with$
is just laughable.
In conclusion, once you know the difference between ["colname"]
and
[, "colname"]
, $
is only useful if it’s making your code cleaner,
saving you typing, or if you actually want the partial matching.
Personally, I’m uncomfortable with the inherent risks of partial
matching, so $
is only really useful for interactive use and my IDE’s
auto-completion. That might even be its intended job. But if that is the
case, nobody warns you of it.
When dealing with any sort of collection, any of the following mistakes can give indistinguishable results. This can make your debugging so messy that by the time that you’re done, you don’t know what was broken.
-
Trying to select an incorrect sequence of elements. This can be caused by
:
orseq()
misbehaving or by simple user error. A tiny bit more on that later -
The vector recycling rules silently causing the vector that you used to select elements to be recycled in an undesired way. More on that later.
-
Selecting an out-of-bounds value. You almost always don’t get any error or warning when you do this. For example, both out-of-bounds positive numbers and logical vectors that are longer than the vector that you’re subsetting silently return
NA
for the inappropriate values.length(LETTERS) ## [1] 26 LETTERS[c(1, 5, 20, 100)] ## [1] "A" "E" "T" NA LETTERS[rep(TRUE, 100)] ## [1] "A" "B" "C" "D" "E" "F" "G" "H" "I" "J" "K" "L" "M" "N" "O" "P" "Q" "R" ## [19] "S" "T" "U" "V" "W" "X" "Y" "Z" NA NA NA NA NA NA NA NA NA NA ## [37] NA NA NA NA NA NA NA NA NA NA NA NA NA NA NA NA NA NA ## [55] NA NA NA NA NA NA NA NA NA NA NA NA NA NA NA NA NA NA ## [73] NA NA NA NA NA NA NA NA NA NA NA NA NA NA NA NA NA NA ## [91] NA NA NA NA NA NA NA NA NA NA
Again, as with many of the issues that we’ve mentioned recently, this happens silently.
-
Accessing/subsetting a collection in the wrong way. For example, wrongly using any of
[c(x, y)]
,[x, y]
, or[cbind(x, y)]
, selecting[x]
rather than[x, ]
,[[x]]
, or[, x]
, using the wrongrbind()/cbind()
, or an error in your call to anything likesubset()
orwithin()
. -
Selecting element 0.
-
Any sort of off-by-one errors, e.g. a modulo mistake of any sort, genuine off-by-one errors, or R’s 1-indexing causing you to trip up.
-
Misuse of searching functions like
which()
,duplicated()
, ormatch()
.
This list also reveals another issue with subsetting: There’s too many ways to do it…
…and they don’t all work everywhere. For example, there’s a wide range of tools for using names to work with lists and data frames, but very few of them work for named atomic vectors (which includes named matrices).
-
The
$
operator simply does not work. -
Although
namedVector["name"]
can be used for subsetting and subassignment,namedVector["name"] <- NULL
throws an error. For a list or data frame, this would have deleted the selected data points.typeof(letters) ## [1] "character" named <- setNames(letters, LETTERS) tail(named) ## U V W X Y Z ## "u" "v" "w" "x" "y" "z" named["Z"] ## Z ## "z" named["Z"] <- "Super!" tail(named) ## U V W X Y Z ## "u" "v" "w" "x" "y" "Super!" #So subsetting and subassignment work just fine. However, for NULL... ## > named["Z"] <- NULL ## Error in named["Z"] <- NULL : replacement has length zero #But for a data frame, this is just fine. (data <- data.frame(A = 1, B = 2, Z = 3)) ## A B Z ## 1 1 2 3 data["Z"] <- NULL data ## A B ## 1 1 2
Incidentally,
anyAtomicVector[index] <- NULL
is also an error. e.g.LETTERS[22] <- NULL
. -
Sorry, did I say that
namedVector["name"]
works for subsetting?a <- diag(3) colnames(a) <- LETTERS[1:3] a ## A B C ## [1,] 1 0 0 ## [2,] 0 1 0 ## [3,] 0 0 1 a["A"] ## [1] NA a["Z"] ## [1] NA
Long story short, named atomic vectors make a distinction between names and colnames that data frames do not.
a <- diag(3) colnames(a) <- LETTERS[1:3] colnames(a) ## [1] "A" "B" "C" names(a) ## NULL names(mtcars) ## [1] "mpg" "cyl" "disp" "hp" "drat" "wt" "qsec" "vs" "am" "gear" ## [11] "carb" colnames(mtcars) ## [1] "mpg" "cyl" "disp" "hp" "drat" "wt" "qsec" "vs" "am" "gear" ## [11] "carb" identical(names(mtcars), colnames(mtcars)) ## [1] TRUE
So what happens when you give an atomic vector plain old names rather than colnames? For a non-matrix, it works fine (see the
named <- setNames(letters, LETTERS)
example above). For a matrix - and presumably for any array, but let’s not get in to that distinction - it’s a little bit more complicated. Look closely at this output before reading further.a <- diag(3) (a <- setNames(a, LETTERS[1:3])) ## [,1] [,2] [,3] ## [1,] 1 0 0 ## [2,] 0 1 0 ## [3,] 0 0 1 ## attr(,"names") ## [1] "A" "B" "C" NA NA NA NA NA NA a["A"] ## A ## 1 a["Z"]#For a data frame, this would be an error... ## <NA> ## NA
When you try to give an atomic vector ordinary names, R will only try to name it element-by-element (even if said vector has dimensions). Data frames, on the other hand, treat names as colnames. R ultimately sees named matrices as named atomic vectors that happen to have a second dimension. This means that you can subset them with both
["name"]
and[, "name"]
and get different results.a <- setNames(diag(3), LETTERS[1:3]) colnames(a) <- LETTERS[1:3] a ## A B C ## [1,] 1 0 0 ## [2,] 0 1 0 ## [3,] 0 0 1 ## attr(,"names") ## [1] "A" "B" "C" NA NA NA NA NA NA a["A"] ## A ## 1 a["Z"] ## <NA> ## NA a[, "A"] ## [1] 1 0 0 #I'd love to show a[, "Z"], but it throws the error "Error in a[, "Z"] : subscript out of bounds". #This is clearly consistent with a["Z"] and my earlier bits on out-of-bounds stuff not throwing errors.
Of course,
["name"]
and[, "name"]
aren’t identical for data frames either, but let’s not get back in to talking about thedrop
argument. Starting to see what I mean about R being inconsistent? -
You cannot use named atomic vectors to generate environments. This means that awesome tricks like
within(data, remove(columnIDoNotWant, anotherColumn))
work for lists and data frames but not for named atomic vectors.#Data frames are fine. head(within(mtcars, remove("mpg"))) ## cyl disp hp drat wt qsec vs am gear carb ## Mazda RX4 6 160 110 3.90 2.620 16.46 0 1 4 4 ## Mazda RX4 Wag 6 160 110 3.90 2.875 17.02 0 1 4 4 ## Datsun 710 4 108 93 3.85 2.320 18.61 1 1 4 1 ## Hornet 4 Drive 6 258 110 3.08 3.215 19.44 1 0 3 1 ## Hornet Sportabout 8 360 175 3.15 3.440 17.02 0 0 3 2 ## Valiant 6 225 105 2.76 3.460 20.22 1 0 3 1 #Named atmomic vectord are not. ## > within(setNames(letters, LETTERS), remove("Z")) ## Error in UseMethod("within") : ## no applicable method for 'within' applied to an object of class "character"
-
When you want to work with the names of named atomic vectors, you probably want to access their names directly and use expressions like
namedVect[!names(namedVect) %in% c("remove", "us")]
.namedVect <- setNames(letters, LETTERS) namedVect[!names(namedVect) %in% c("A", "Z")] ## B C D E F G H I J K L M N O P Q R S T U ## "b" "c" "d" "e" "f" "g" "h" "i" "j" "k" "l" "m" "n" "o" "p" "q" "r" "s" "t" "u" ## V W X Y ## "v" "w" "x" "y"
However, this is a bad habit for non-atomic vectors because, unless you take the precautions mentioned earlier,
[
likes to remove duplicated names and unnecessary dimensions from your data. -
Don’t think that functional programming will save you from my previous point. The base library’s higher-order functions don’t play nice with the
names()
function. I think it’s got something to do withlapply()
usingX[[i]]
under the hood (see its documentation).namedVect <- setNames(letters, LETTERS) Filter(function(x) names(x) == "A", namedVect) ## named character(0) head(lapply(namedVect, function(x) names(x) == "A")) ## $A ## logical(0) ## ## $B ## logical(0) ## ## $C ## logical(0) ## ## $D ## logical(0) ## ## $E ## logical(0) ## ## $F ## logical(0) head(sapply(namedVect, function(x) names(x) == "A")) ## $A ## logical(0) ## ## $B ## logical(0) ## ## $C ## logical(0) ## ## $D ## logical(0) ## ## $E ## logical(0) ## ## $F ## logical(0)
Did you notice that
Filter
andlapply
’s arguments are in inconsistent orders? A little bit more on that much later.
From the above few points, you can see that it’s hard to find a way to
manipulate named atomic vectors by their names that is both safe for
them and for other named objects. The only one that comes to mind is to
use [
with the aforementioned precautions. That’s bad enough on its
own – it makes R feel unsafe and inconsistent – but it also makes named
atomic vectors feel like an afterthought. I find that most of my code
that makes extended use of named atomic vectors comes out looking
disturbingly unidiomatic. A little bit more on that when I talk about
matrices.
I’ve already given a few examples of R either silently doing nothing or silently doing what you don’t want. Let’s have a few more:
-
Again, much of what I’ve listed in the Indistinguishable Errors and Removing Dimensions sections occur silently.
-
As documented here, negative out-of-bounds values are silently disregarded when deleting elements. For example, if you have
x <- 1:10
, thenx[-20]
returns an unmodified version ofx
without warning or error.x <- 1:10 x[20] ## [1] NA x ## [1] 1 2 3 4 5 6 7 8 9 10 x[-20] ## [1] 1 2 3 4 5 6 7 8 9 10 identical(x, x[-20]) ## [1] TRUE
Given that
x[20]
isNA
– a questionable decision in of itself – is this the behaviour that you expected? -
Subassigning
NULL
to a column that your data does not have does not give a warning or error. For example, trying to accessmtcars["weight"]
is an error, butmtcars["weight"] <- NULL
silently does nothing.$
and$<-
have the same issue. -
Using
within()
to remove unwanted columns from your data, e.g.within(data, rm(colName1, colName2))
, does nothing to any columns with duplicated names. Again, no warning or error…dupe <- cbind(mtcars, foo = 3, foo = 4) head(dupe) ## mpg cyl disp hp drat wt qsec vs am gear carb foo foo ## Mazda RX4 21.0 6 160 110 3.90 2.620 16.46 0 1 4 4 3 4 ## Mazda RX4 Wag 21.0 6 160 110 3.90 2.875 17.02 0 1 4 4 3 4 ## Datsun 710 22.8 4 108 93 3.85 2.320 18.61 1 1 4 1 3 4 ## Hornet 4 Drive 21.4 6 258 110 3.08 3.215 19.44 1 0 3 1 3 4 ## Hornet Sportabout 18.7 8 360 175 3.15 3.440 17.02 0 0 3 2 3 4 ## Valiant 18.1 6 225 105 2.76 3.460 20.22 1 0 3 1 3 4 head(within(dupe, rm(carb, foo))) ## mpg cyl disp hp drat wt qsec vs am gear foo foo ## Mazda RX4 21.0 6 160 110 3.90 2.620 16.46 0 1 4 4 4 ## Mazda RX4 Wag 21.0 6 160 110 3.90 2.875 17.02 0 1 4 4 4 ## Datsun 710 22.8 4 108 93 3.85 2.320 18.61 1 1 4 4 4 ## Hornet 4 Drive 21.4 6 258 110 3.08 3.215 19.44 1 0 3 4 4 ## Hornet Sportabout 18.7 8 360 175 3.15 3.440 17.02 0 0 3 4 4 ## Valiant 18.1 6 225 105 2.76 3.460 20.22 1 0 3 4 4
By the way,
cbind()
doesn’t silently correct duplicated column names. By now, you probably expected otherwise. This is documented behaviour, but I don’t think that anyone ever bothered to read the docs forcbind()
. -
Using
subset()
rather thanwithin()
is sometimes suggested for operations like what I was trying to do in the previous point. For example, you can remove columns withsubset(data, select = -c(colName1, colName2))
. However, for duplicated names, I’d argue thatsubset()
is even weirder thanwithin()
. Withsubset()
, attempting to remove a duplicated column by name will only remove the first such column and removing any non-duplicated column will change the names of your duplicated columns.#First, I'll show subset() working as normal and save us some space. mtcars2 <- subset(mtcars, mpg > 25, select = -c(cyl, disp, hp, wt)) mtcars2 ## mpg drat qsec vs am gear carb ## Fiat 128 32.4 4.08 19.47 1 1 4 1 ## Honda Civic 30.4 4.93 18.52 1 1 4 2 ## Toyota Corolla 33.9 4.22 19.90 1 1 4 1 ## Fiat X1-9 27.3 4.08 18.90 1 1 4 1 ## Porsche 914-2 26.0 4.43 16.70 0 1 5 2 ## Lotus Europa 30.4 3.77 16.90 1 1 5 2 dupe <- cbind(mtcars2, foo = 3, foo = 4, foo = 5) dupe ## mpg drat qsec vs am gear carb foo foo foo ## Fiat 128 32.4 4.08 19.47 1 1 4 1 3 4 5 ## Honda Civic 30.4 4.93 18.52 1 1 4 2 3 4 5 ## Toyota Corolla 33.9 4.22 19.90 1 1 4 1 3 4 5 ## Fiat X1-9 27.3 4.08 18.90 1 1 4 1 3 4 5 ## Porsche 914-2 26.0 4.43 16.70 0 1 5 2 3 4 5 ## Lotus Europa 30.4 3.77 16.90 1 1 5 2 3 4 5 subset(dupe, select = -foo)#Names have silently changed and only one foo was dropped. ## mpg drat qsec vs am gear carb foo foo.1 ## Fiat 128 32.4 4.08 19.47 1 1 4 1 4 5 ## Honda Civic 30.4 4.93 18.52 1 1 4 2 4 5 ## Toyota Corolla 33.9 4.22 19.90 1 1 4 1 4 5 ## Fiat X1-9 27.3 4.08 18.90 1 1 4 1 4 5 ## Porsche 914-2 26.0 4.43 16.70 0 1 5 2 4 5 ## Lotus Europa 30.4 3.77 16.90 1 1 5 2 4 5 subset(dupe, select = -c(foo, foo))#Identical to previous. ## mpg drat qsec vs am gear carb foo foo.1 ## Fiat 128 32.4 4.08 19.47 1 1 4 1 4 5 ## Honda Civic 30.4 4.93 18.52 1 1 4 2 4 5 ## Toyota Corolla 33.9 4.22 19.90 1 1 4 1 4 5 ## Fiat X1-9 27.3 4.08 18.90 1 1 4 1 4 5 ## Porsche 914-2 26.0 4.43 16.70 0 1 5 2 4 5 ## Lotus Europa 30.4 3.77 16.90 1 1 5 2 4 5 subset(dupe, select = -carb)#Foo's names have silently changed, despite us not touching foo! ## mpg drat qsec vs am gear foo foo.1 foo.2 ## Fiat 128 32.4 4.08 19.47 1 1 4 3 4 5 ## Honda Civic 30.4 4.93 18.52 1 1 4 3 4 5 ## Toyota Corolla 33.9 4.22 19.90 1 1 4 3 4 5 ## Fiat X1-9 27.3 4.08 18.90 1 1 4 3 4 5 ## Porsche 914-2 26.0 4.43 16.70 0 1 5 3 4 5 ## Lotus Europa 30.4 3.77 16.90 1 1 5 3 4 5 subset(dupe, select = -c(carb, foo))#Names have silently changed and only one foo was dropped. ## mpg drat qsec vs am gear foo foo.1 ## Fiat 128 32.4 4.08 19.47 1 1 4 4 5 ## Honda Civic 30.4 4.93 18.52 1 1 4 4 5 ## Toyota Corolla 33.9 4.22 19.90 1 1 4 4 5 ## Fiat X1-9 27.3 4.08 18.90 1 1 4 4 5 ## Porsche 914-2 26.0 4.43 16.70 0 1 5 4 5 ## Lotus Europa 30.4 3.77 16.90 1 1 5 4 5
I think that the worst example here is
subset(dupe, select = -carb)
. I didn’t touchfoo
, so why change it? I’d rather havewithin()
’s silent inaction thansubset()
’s silent sabotage.
Needless to say, there will be more examples of R silently misbehaving later on in this document. This was just a good place to throw in a few that are specific to subsetting.
This should be easy, shouldn’t it? Go through the data and only give me
the bits that have the property that I’m asking for. What could possibly
go wrong? Turns out, it’s quite a lot. Even predicates as simple as
“does the element equal x
?” are a minefield. I understand why these
examples are the way that they are – really, I do – but how to delete
unwanted elements is one of the first things that you’re going to want
to learn in a stats language. For something that you’re going to want to
be able to do on day one of using R, there are far too many pitfalls.
-
You might think that
setdiff()
is sufficient for removing data – it’s certainly the first thing tool that a mathematician would reach for – but it has the side-effect of removing duplicate entries from the original vector and destroying your data structures by applyingas.vector()
to them.Nile ## Time Series: ## Start = 1871 ## End = 1970 ## Frequency = 1 ## [1] 1120 1160 963 1210 1160 1160 813 1230 1370 1140 995 935 1110 994 1020 ## [16] 960 1180 799 958 1140 1100 1210 1150 1250 1260 1220 1030 1100 774 840 ## [31] 874 694 940 833 701 916 692 1020 1050 969 831 726 456 824 702 ## [46] 1120 1100 832 764 821 768 845 864 862 698 845 744 796 1040 759 ## [61] 781 865 845 944 984 897 822 1010 771 676 649 846 812 742 801 ## [76] 1040 860 874 848 890 744 749 838 1050 918 986 797 923 975 815 ## [91] 1020 906 901 1170 912 746 919 718 714 740 setdiff(Nile, 1160)#Not a time series any more. ## [1] 1120 963 1210 813 1230 1370 1140 995 935 1110 994 1020 960 1180 799 ## [16] 958 1100 1150 1250 1260 1220 1030 774 840 874 694 940 833 701 916 ## [31] 692 1050 969 831 726 456 824 702 832 764 821 768 845 864 862 ## [46] 698 744 796 1040 759 781 865 944 984 897 822 1010 771 676 649 ## [61] 846 812 742 801 860 848 890 749 838 918 986 797 923 975 815 ## [76] 906 901 1170 912 746 919 718 714 740 setdiff(Nile, 0)#Hey, where did the other 1160s go? ## [1] 1120 1160 963 1210 813 1230 1370 1140 995 935 1110 994 1020 960 1180 ## [16] 799 958 1100 1150 1250 1260 1220 1030 774 840 874 694 940 833 701 ## [31] 916 692 1050 969 831 726 456 824 702 832 764 821 768 845 864 ## [46] 862 698 744 796 1040 759 781 865 944 984 897 822 1010 771 676 ## [61] 649 846 812 742 801 860 848 890 749 838 918 986 797 923 975 ## [76] 815 906 901 1170 912 746 919 718 714 740
It’s safer when you’re dealing with names, e.g.
data[setdiff(names(data), "nameOfThingToDelete")]
head(mtcars) ## mpg cyl disp hp drat wt qsec vs am gear carb ## Mazda RX4 21.0 6 160 110 3.90 2.620 16.46 0 1 4 4 ## Mazda RX4 Wag 21.0 6 160 110 3.90 2.875 17.02 0 1 4 4 ## Datsun 710 22.8 4 108 93 3.85 2.320 18.61 1 1 4 1 ## Hornet 4 Drive 21.4 6 258 110 3.08 3.215 19.44 1 0 3 1 ## Hornet Sportabout 18.7 8 360 175 3.15 3.440 17.02 0 0 3 2 ## Valiant 18.1 6 225 105 2.76 3.460 20.22 1 0 3 1 head(mtcars[setdiff(names(mtcars), "wt")]) ## mpg cyl disp hp drat qsec vs am gear carb ## Mazda RX4 21.0 6 160 110 3.90 16.46 0 1 4 4 ## Mazda RX4 Wag 21.0 6 160 110 3.90 17.02 0 1 4 4 ## Datsun 710 22.8 4 108 93 3.85 18.61 1 1 4 1 ## Hornet 4 Drive 21.4 6 258 110 3.08 19.44 1 0 3 1 ## Hornet Sportabout 18.7 8 360 175 3.15 17.02 0 0 3 2 ## Valiant 18.1 6 225 105 2.76 20.22 1 0 3 1
but anything that’s only sometimes safe doesn’t fill me with confidence.
-
Because
which()
is an extremely intuitive function for extracting/changing subsets of your data and for dealing with missing values (see The R Inferno, section 8.1.12), it is one of the first things that a beginner will learn about. However, although your intuition is screaming for you to do it, you almost never want to usedata <- data[-which(data==thingToDelete)]
. Whenwhich()
finds no matches, it evaluates to something of length 0. This makesdata[-which(data==thingToDelete)]
also returns something of length 0, deleting your data.Nile ## Time Series: ## Start = 1871 ## End = 1970 ## Frequency = 1 ## [1] 1120 1160 963 1210 1160 1160 813 1230 1370 1140 995 935 1110 994 1020 ## [16] 960 1180 799 958 1140 1100 1210 1150 1250 1260 1220 1030 1100 774 840 ## [31] 874 694 940 833 701 916 692 1020 1050 969 831 726 456 824 702 ## [46] 1120 1100 832 764 821 768 845 864 862 698 845 744 796 1040 759 ## [61] 781 865 845 944 984 897 822 1010 771 676 649 846 812 742 801 ## [76] 1040 860 874 848 890 744 749 838 1050 918 986 797 923 975 815 ## [91] 1020 906 901 1170 912 746 919 718 714 740 Nile[-which(Nile==1160)]#This is fine. ## [1] 1120 963 1210 813 1230 1370 1140 995 935 1110 994 1020 960 1180 799 ## [16] 958 1140 1100 1210 1150 1250 1260 1220 1030 1100 774 840 874 694 940 ## [31] 833 701 916 692 1020 1050 969 831 726 456 824 702 1120 1100 832 ## [46] 764 821 768 845 864 862 698 845 744 796 1040 759 781 865 845 ## [61] 944 984 897 822 1010 771 676 649 846 812 742 801 1040 860 874 ## [76] 848 890 744 749 838 1050 918 986 797 923 975 815 1020 906 901 ## [91] 1170 912 746 919 718 714 740 which(Nile==11600) ## integer(0) Nile[-which(Nile==11600)]#This is not. ## numeric(0)
What you probably expected was
which()
leaving your data unchanged when it has not found a match. You might also have expected a warning or error, but surely you’ve learned your lesson by now? Anyway, section 8.1.13 of The R Inferno offers some ways to get this behaviour, but the only practical-looking suggestion isdata[!(data %in% thingToDelete)]
. I think that you can get away with removing the curly brackets there.Nile[!Nile %in% 1160] ## [1] 1120 963 1210 813 1230 1370 1140 995 935 1110 994 1020 960 1180 799 ## [16] 958 1140 1100 1210 1150 1250 1260 1220 1030 1100 774 840 874 694 940 ## [31] 833 701 916 692 1020 1050 969 831 726 456 824 702 1120 1100 832 ## [46] 764 821 768 845 864 862 698 845 744 796 1040 759 781 865 845 ## [61] 944 984 897 822 1010 771 676 649 846 812 742 801 1040 860 874 ## [76] 848 890 744 749 838 1050 918 986 797 923 975 815 1020 906 901 ## [91] 1170 912 746 919 718 714 740 Nile[!Nile %in% 11600] ## [1] 1120 1160 963 1210 1160 1160 813 1230 1370 1140 995 935 1110 994 1020 ## [16] 960 1180 799 958 1140 1100 1210 1150 1250 1260 1220 1030 1100 774 840 ## [31] 874 694 940 833 701 916 692 1020 1050 969 831 726 456 824 702 ## [46] 1120 1100 832 764 821 768 845 864 862 698 845 744 796 1040 759 ## [61] 781 865 845 944 984 897 822 1010 771 676 649 846 812 742 801 ## [76] 1040 860 874 848 890 744 749 838 1050 918 986 797 923 975 815 ## [91] 1020 906 901 1170 912 746 919 718 714 740
That’s mostly okay. However,
identical(Nile, Nile[!Nile %in% 11600])
isFALSE
. Can you guess why? It’s like R has no always safe ways to subset. -
At least removing elements that are equal to a particular number is simple for vectors. Even for lists, it’s just
data[data!=x]
. It’s maybe not what a beginner would guess (“I have to writedata
twice?”), but it’s simple enough. -
For removing a vector from a list of vectors, you’re going to want to learn some functional programming idioms. Not hard if you’re a programmer, but shouldn’t this be easier in a stats and maths tool? Anyway, you probably want
Filter(function(x) all(x!=vectorToDelete), data)
. You can also do it with the apply family, but I don’t see why you would. -
Removing what you don’t want from a data frame largely comes down to mastering the subsetting rules, a nightmare that I’ve spent the previous few thousand words covering. I often end up with very ugly lines like
outcomes[outcomes$playerChoice == playerChoice & outcomes$computerChoice == computerChoice, "outcome"]
-
Before you ask,
subset()
,with()
, andwithin()
aren’t good enough either. I’ve already mentioned some of their issues, but more on them later.
Overall, it’s like R has no safe ways to subset. What is safe for one
job is often either unsafe, invalid, or inconsistent with another. R’s
huge set of subsetting tools is useful – maybe even good – once
mastered, but until then you’re forced to adopt a guess-and-check style
of programming and pray that you get a useful error/warning message when
you get something wrong. Worse still, these prayers are rarely answered
and, in the cases where R silently does something that you didn’t want,
they’re outright mocked. Do you understand how damning that is for a
stats language? I can’t stress this point enough. Subsetting in R should
be easy and intuitive. Instead, it’s something that I’ve managed to
produce thousands of words of complaints about and it still trips me up
with alarming
regularity,
despite my clear knowledge of the correct way to do things. If I want a
vector of consonants, you can bet that I’m going to write
letters[-c("a", "e", "i", "o", "u")]
,
letters[-which(letters == c("a", "e", "i", "o", "u"))]
, and
letters[c("a", "e", "i", "o", "u") %in% letters]
before remembering
the right way to do it. If I’m still making those mistakes for something
simple, then I can only imagine what it’s like for a true beginner doing
something complicated.
You’ve heard the good, now for the bad. R’s vectorization is probably the best thing about the language and it will work miracles when you’re doing mathematics. However, it will trip you up in other areas. A lot of these points are minor, but when they cause you problems their source can be tough to track down. This is because R is working as intended and therefore not giving you any warnings or errors (spotting a pattern?). Furthermore, if you have correctly identified that you have a vectorization problem, then pretty much any function in R could be to blame, because most of R’s functions are vectorized.
-
The commonality of vectors leads to some new syntax that must be memorised. For example,
if(x|y)
andif(x||y)
are very different and using&&
rather than&
can be fatal. Compare the following:mtcars[mtcars$mpg < 20 && mtcars$hp > 150,] ## Warning in mtcars$mpg < 20 && mtcars$hp > 150: 'length(x) = 32 > 1' in coercion ## to 'logical(1)' ## [1] mpg cyl disp hp drat wt qsec vs am gear carb ## <0 rows> (or 0-length row.names) mtcars[mtcars$mpg < 20 & mtcars$hp > 150,] ## mpg cyl disp hp drat wt qsec vs am gear carb ## Hornet Sportabout 18.7 8 360.0 175 3.15 3.440 17.02 0 0 3 2 ## Duster 360 14.3 8 360.0 245 3.21 3.570 15.84 0 0 3 4 ## Merc 450SE 16.4 8 275.8 180 3.07 4.070 17.40 0 0 3 3 ## Merc 450SL 17.3 8 275.8 180 3.07 3.730 17.60 0 0 3 3 ## Merc 450SLC 15.2 8 275.8 180 3.07 3.780 18.00 0 0 3 3 ## Cadillac Fleetwood 10.4 8 472.0 205 2.93 5.250 17.98 0 0 3 4 ## Lincoln Continental 10.4 8 460.0 215 3.00 5.424 17.82 0 0 3 4 ## Chrysler Imperial 14.7 8 440.0 230 3.23 5.345 17.42 0 0 3 4 ## Camaro Z28 13.3 8 350.0 245 3.73 3.840 15.41 0 0 3 4 ## Pontiac Firebird 19.2 8 400.0 175 3.08 3.845 17.05 0 0 3 2 ## Ford Pantera L 15.8 8 351.0 264 4.22 3.170 14.50 0 1 5 4 ## Ferrari Dino 19.7 6 145.0 175 3.62 2.770 15.50 0 1 5 6 ## Maserati Bora 15.0 8 301.0 335 3.54 3.570 14.60 0 1 5 8
Personally, I find that it’s easy to remember to use
&
forif
but I often forget to use&
for subsetting. It looks like version 4.1.4 is going to make||
and&&
throw warnings. -
The
if
statements accept vectors of length greater than 1 as their predicate, but will only pay attention to the very first element. This throws a warning and there is a global option to make it an error instead, but I can’t see why R accepts such predicates at all. Why would I ever useif(c(TRUE, FALSE))
to mean “if the first element of my vector is true, then…”? This is also what the&&
and||
syntax is for (e.g.c(TRUE, FALSE) && c(TRUE, FALSE)
isTRUE
), but I still don’t see why anyone would use several logical vectors and only be interested in their first elements.- It appears that version 4.1.4 is going to do something about this. Specifically, replace this warning with an error.
-
When dealing with anything 2D, you need to be very careful to not mix up any of
length()
,lengths()
,nrow()
, orncol()
. In particular,length()
is so inconsistent that I’m unsure why they let it work for 2D structures (probably something to do with it being an internal generic). For example, the length of a data frame is its number of columns and the length of a matrix is its number of elements.(a <- diag(4)) ## [,1] [,2] [,3] [,4] ## [1,] 1 0 0 0 ## [2,] 0 1 0 0 ## [3,] 0 0 1 0 ## [4,] 0 0 0 1 (b <- as.data.frame(a)) ## V1 V2 V3 V4 ## 1 1 0 0 0 ## 2 0 1 0 0 ## 3 0 0 1 0 ## 4 0 0 0 1 length(a) ## [1] 16 length(b) ## [1] 4
-
Vectors are collections and therefore inherit the previous section’s issues about selecting elements.
-
Because virtually everything is already a vector, you never know what to use when you want a collection or anything nested. Lists? Arrays?
c()
? Data frames? One ofcbind()
/rbind()
? Matrices? You get used to it eventually, but it takes a while to understand the differences. -
Some functions are vectorized in such a way that you’re forced to remember the difference between how they behave for n length-one vectors and and how they behave for the corresponding single vector of length n. For example,
paste("Alice", "Bob", "Charlie")
is not the same aspaste(c("Alice", "Bob", "Charlie"))
.paste("Alice", "Bob", "Charlie") ## [1] "Alice Bob Charlie" paste(c("Alice", "Bob", "Charlie")) ## [1] "Alice" "Bob" "Charlie" paste("Alice", "Bob", "Charlie", collapse = "") ## [1] "Alice Bob Charlie" paste(c("Alice", "Bob", "Charlie"), collapse = "") ## [1] "AliceBobCharlie"
I’m not saying that this doesn’t make sense, but it is a source of unpredictability.
-
Another unpredictable example: What does
max(100:200, 250:350, 276)
return? You might be surprised to discover that the output is the single number350
, rather than a vector of many outputs.max(100:200, 250:350, 276) ## [1] 350
The fix for this isn’t some
collapse
-like argument like it is forpaste()
, it’s an entirely different function:pmax()
. Why?pmax(100:200, 250:350, 276) ## [1] 276 276 276 276 276 276 276 276 276 276 276 276 276 276 276 276 276 276 ## [19] 276 276 276 276 276 276 276 276 276 277 278 279 280 281 282 283 284 285 ## [37] 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 ## [55] 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 ## [73] 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 ## [91] 340 341 342 343 344 345 346 347 348 349 350
-
A further annoyance comes from how many things behave differently on vectors of length one. For example,
sample(1:5)
is exactly the same assample(5)
, which is bound to give you bugs when you usesample(5:n)
for changingn
. -
R has rules for recycling vector elements when you try to get it to do something with several vectors that don’t all have the same length. You saw this abused when I gave the
x <- paste0(rep("", 100), c("", "", "Fizz"), c("", "", "", "", "Buzz"))
FizzBuzz example. When recycling occurs, R only throws a warning if the longest vector’s length is not a multiple of the others. For example, neitherMap(sum, 1:6, 1:3)
nor that FizzBuzz line warn you that recycling has occurred, butMap(sum, 1:6, 1:4)
will.Map(sum, 1:6, 1:3) ## [[1]] ## [1] 2 ## ## [[2]] ## [1] 4 ## ## [[3]] ## [1] 6 ## ## [[4]] ## [1] 5 ## ## [[5]] ## [1] 7 ## ## [[6]] ## [1] 9 Map(sum, 1:6, 1:4) ## Warning in mapply(FUN = f, ..., SIMPLIFY = FALSE): longer argument not a ## multiple of length of shorter ## [[1]] ## [1] 2 ## ## [[2]] ## [1] 4 ## ## [[3]] ## [1] 6 ## ## [[4]] ## [1] 8 ## ## [[5]] ## [1] 6 ## ## [[6]] ## [1] 8
The first case – where no warnings are given – can be an unexpected source of major error. The authors of the Tidyverse seem to agree with me. For example, you’re only allowed to recycle vectors of length 1 when constructing a tibble, so
tibble(1:4, 1:2)
will throw a clear error message whereasdata.frame(1:4, 1:2)
silently recycles the second argument. Similarly,map2(1:6, 1:3, sum)
is an error, butmap2(1:6, 1, sum)
is not.library(tibble) ## > tibble(1:4, 1:2) ## Error: Tibble columns must have compatible sizes. ## * Size 4: Existing data. ## * Size 2: Column at position 2. ## ℹ Only values of size one are recycled. ## Run `rlang::last_error()` to see where the error occurred. data.frame(1:4, 1:2) ## X1.4 X1.2 ## 1 1 1 ## 2 2 2 ## 3 3 1 ## 4 4 2 library(purrr) ## > map2(1:6, 1:3, sum) ## Error: Mapped vectors must have consistent lengths: ## * `.x` has length 6 ## * `.y` has length 3 map2(1:6, 1, sum) ## [[1]] ## [1] 2 ## ## [[2]] ## [1] 3 ## ## [[3]] ## [1] 4 ## ## [[4]] ## [1] 5 ## ## [[5]] ## [1] 6 ## ## [[6]] ## [1] 7
-
Section 8.1.6 of The R Inferno: The recycling of vectors lets you attempt to do things that look correct to a novice and make sense to a master, but are almost certainly not what was wanted. For example,
c(4, 6) == 1:10
isTRUE
only in its sixth element. The recycling rules turn it in toc(4, 6, 4, 6, 4, 6, 4, 6, 4, 6) == 1:10
. Again, there is no warning given to the user unless the longest vector’s length is not a multiple of the other’s. In this case, what you wanted was probablyc(4, 6) %in% 1:10
, maybe with a call toall()
.c(4, 6) == 1:10 ## [1] FALSE FALSE FALSE FALSE FALSE TRUE FALSE FALSE FALSE FALSE c(4, 6, 4, 6, 4, 6, 4, 6, 4, 6) == 1:10 ## [1] FALSE FALSE FALSE FALSE FALSE TRUE FALSE FALSE FALSE FALSE c(4, 6) %in% 1:10 ## [1] TRUE TRUE all(c(4, 6) %in% 1:10) ## [1] TRUE
-
Some functions don’t recycle in the way that you would expect. For example, read the documentation for
strsplit()
and ask yourself if you expectstrsplit("Alice", c("l", "c"))
andstrsplit("Alice", "l")
to give the same output. If you think that they don’t, you’re wrong. If you expected the first option to warn you about the"c"
part not being used, you’re sane, but wrong. If you want to see how the second argument is supposed to work, re-run the earlier code withc("Alice", "Boblice")
as your first argument.strsplit("Alice", c("l", "c")) ## [[1]] ## [1] "A" "ice" strsplit("Alice", "l") ## [[1]] ## [1] "A" "ice" strsplit(c("Alice", "Boblice"), c("l", "c")) ## [[1]] ## [1] "A" "ice" ## ## [[2]] ## [1] "Bobli" "e"
-
Remember what I said about needing to generate the correct logical vector when you want to subset a collection? Logical vectors are also recycled when subsetting collections. Because this vector recycling does not always throw warnings or errors, it’s a new Hell. I’m honestly not sure if the exact rules for when this does/doesn’t throw warnings/errors are documented anywhere. The language definition claims that using a logical vector to subset a longer vector follows the same rules as when you’re using two such vectors for arithmetic (i.e. you get a warning if the larger of the two’s length isn’t a multiple of the smaller’s). However, I know this to be false.
a <- 1:10 a + rep(1, 9) #Arithmetic; Gives a warning. ## Warning in a + rep(1, 9): longer object length is not a multiple of shorter ## object length ## [1] 2 3 4 5 6 7 8 9 10 11 a[rep(TRUE, 9)] #Logical subsetting; 10 results without warning. ## [1] 1 2 3 4 5 6 7 8 9 10 a[c(TRUE, FALSE, TRUE)] #Again, 10 results. Shouldn't it be either 10 with a warning or just 3? ## [1] 1 3 4 6 7 9 10
I’ll take this chance to repeat my claim that this is extremely powerful if used correctly, but the potential for errors slipping through unnoticed is huge. This toy example isn’t so bad, but wait until these errors creep in to your dataset with 50 rows and columns, leaving you with no idea where it all went wrong. The first time where this really caught me out was when I used the same logical vector for two similar datasets of slightly different sizes. I had hoped that if anything went wrong, I’d get an error. Because I didn’t, I continued on without knowing that half of my data was now ruined.
-
Logical vectors also recycle
NA
without warning. I can’t point to any documentation that contradicts this, but it will always catch you off guard. On the bright side, this is consistent with the addition and subsetting rules for numeric vectors withNA
s.arithmetic <- c(2, NA) arithmetic + c(11, 12, 13, 14) #Keeps NA and recycles. ## [1] 13 NA 15 NA logic <- c(TRUE, FALSE, TRUE, NA) LETTERS[logic] ## [1] "A" "C" NA "E" "G" NA "I" "K" NA "M" "O" NA "Q" "S" NA "U" "W" NA "Y" LETTERS[arithmetic] #Keeps NA and recycling is not expected. ## [1] "B" NA
-
You sometimes have to tell R that you wanted to work on the entire vector rather than its elements. For example,
rep(matrix(1:4, nrow = 2, ncol = 2), 5)
will not repeat the matrix 5 times, it will repeat its elements 5 times. The fix is to userep(list(matrix(1:4, nrow = 2, ncol = 2)), 5)
instead.m <- matrix(1:4, nrow = 2, ncol = 2) rep(m, 5) ## [1] 1 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4 rep(list(m), 5) ## [[1]] ## [,1] [,2] ## [1,] 1 3 ## [2,] 2 4 ## ## [[2]] ## [,1] [,2] ## [1,] 1 3 ## [2,] 2 4 ## ## [[3]] ## [,1] [,2] ## [1,] 1 3 ## [2,] 2 4 ## ## [[4]] ## [,1] [,2] ## [1,] 1 3 ## [2,] 2 4 ## ## [[5]] ## [,1] [,2] ## [1,] 1 3 ## [2,] 2 4
Similarly, you might think that
vect %in% listOfVectors
will work, but it will instead check if the elements ofvect
are elements oflistOfVectors
. Again, the solution is to wrap the vector in a list. For example, you wantlist(1:4) %in% list(5:10, 10:15, 1:4)
not1:4 %in% list(5:10, 10:15, 1:4)
.list(1:4) %in% list(5:10, 10:15, 1:4) ## [1] TRUE 1:4 %in% list(5:10, 10:15, 1:4) ## [1] FALSE FALSE FALSE FALSE
You might be surprised that the last result was entirely
FALSE
. After all, some of1:4
is in the last element of the list. I’ll leave that one to you.
Again, for the most part, these aren’t major issues. I don’t
particularly like the inconsistency between functions like paste()
and
max()
, but the only true minefield is the vector recycling rules. When
they silently do things that you don’t want, you’re screwed.
R makes no secret of being essentially half a century of patches for S. Many things disagree, lack any clear conventions, or are just plain bad, but show no signs of changing. Because so many packages depend on these inconsistencies, I don’t think that they will ever be removed from base R. R could be salvaged if its means of helping you manage the inconsistency were up to scratch – e.g. the documentation, the function/argument names, or the warning/error messages – but they’re not. It’s therefore hard to guess about anything or to help yourself when you’ve guessed wrong. These sounds like minor complaints, but R can be so poor in these regards that it becomes a deal-breaker for the entire language. If there’s one thing that will make you quit R forever, it’s this. It may sound like I’m being harsh, but I’m not alone in saying it. Both Advanced R and The R Inferno can barely go a section without pointing out an inconsistency in R.
Really, this is R’s biggest issue. You can get used to the arcane laws powering R’s subsetting and vectorization, the abnormalities of its variable manipulations, and it’s tendency to do dangerous things without warning you. However, this is the one thing that you can never learn to live with. R is openly, dangerously, and eternally inconsistent and also does a poor job of helping you live with that. In the very worst cases, you can’t find the relevant documentation, the thing that’s conceptually close to what you’re after doesn’t link to it, the examples are as poor as they are few, the documentation is simultaneously incomplete and filled with irrelevant information while assuming familiarity with something alien, the error messages don’t tell you what line threw the errors that your inevitable misunderstandings caused, the dissimilarity between what you’re working with and the rest of the language makes it impossible to guess where you’ve slipped up, there’s undocumented behaviour that you need to look at the C code to discover, and you know that none of this will ever be fixed!
These issues tend to overlap, but I’ve done my best to split this up in to sections that cover each aspect of this problem. All in all, this section came out to be shorter than I expected. However, I hope that I have made the magnitude of some of these points clear.
If R had outstanding documentation, then I could live with its inconsistencies. Sadly, it doesn’t. The documentation does almost nothing to help you in this regard and has more than its fair share of issues:
-
Some of the docs are ancient and therefore have examples that are either terrible, few in number, or non-existent. The references in these docs suggests that this is a disease inherited from S, but sometimes it’s really unforgivable:
- The Examples section in the documentation for control flow shows no
examples of how to use
if
,while
,repeat
,break
, ornext
. They’re explained in the actual text, but I expect the Examples section to give examples! - The documentation for the list data type has five examples, many of which are for the other seven functions that it shares it documentation with, despite it being an absolutely fundamental data type. For some reason, that documentation also includes a library call. And yes, that does mean that some of the functions don’t have examples.
- For stats functions, we typically see documentation for a set of
algorithms. However, said documentation will often have far fewer
examples than there are members in said set. The
quantile()
function’s docs are an extreme examples of this. A similar sin can be found in the docs forlm()
andglm()
. However, their See Also sections link to a lot of functions that use them in their own examples, so I can just barely forgive this.
The Tidyverse seems to be far better in this regard, with the examples often taking up almost as much room as the actual documentation. However, I don’t like how its docs often don’t have a Value section, like a lot of base R’s docs do.
- The Examples section in the documentation for control flow shows no
examples of how to use
-
Some of the docs have no examples at all e.g.
UseMethod()
,vcov()
, andxtfrm()
. -
Some of the docs will document many seemingly identical things and not tell you how they differ. For example, can you tell from the documentation if there’s a difference between
rm()
andremove()
? An even worse case is trying to figure out the difference betweenresid()
andresiduals()
. The documentation correctly tells you that one is an alias for another, but then it tells you thatresid()
is intended to encourage you to not do a certain thing. This implies thatresiduals()
does not have that same intention, incorrectly hinting that they might have different behaviour. -
In some of the standard libraries, you can find functions without any documentation. For example,
MASS::as.fraction()
is totally undocumented. -
The R Language Definition is incomplete. I imagine that this will really bother some people on principle alone. Personally, I would be satisfied if it were incomplete in the sense of “each section is complete and correct, but the document is missing many key sections”. However, it’s really more like a rough draft. It has sentences that stop mid-word, prompts for where to write something latter, and lots of information that is either clearly incomplete or very out of date.
-
A lot of R’s base functions are not written in R, so if you really want to understand how an R function works, you need to learn an extra language. I find that a lot of the power users have gotten used to reading the C source code for a lot of R. That wouldn’t be so bad, but…
-
For a long time, I didn’t know why many of my technical questions on Stack Overflow were answered by direct reference to R’s code, without any mention of its documentation. I eventually learned that R’s functions occasionally have undocumented behaviour, meaning that you can’t trust anything other than the code. For example:
-
Where do the docs tell you that the
expr
argument inreplicate()
gets wrapped in an anonymous function, meaning that you can’t use it to do<-
variable assignment to its calling environment (e.g. code liken <- 0; replicate(5, n <- n + 1)
does not changen
)? You might just spot this if you check the R code, but even then it’s not clear.replicate ## function (n, expr, simplify = "array") ## sapply(integer(n), eval.parent(substitute(function(...) expr)), ## simplify = simplify) ## <bytecode: 0x55e8793a81c0> ## <environment: namespace:base>
-
Where do
rep()
’s docs tell you that it’s a special kind of generic where your extensions to it won’t dispatch properly? Even the R code –function (x, ...) .Primitive("rep")
– won’t help you here. -
Where do
lapply()
andFilter()
’s docs tell you that they don’t play nice with thenames()
function? Again, even the R code won’t help here.lapply ## function (X, FUN, ...) ## { ## FUN <- match.fun(FUN) ## if (!is.vector(X) || is.object(X)) ## X <- as.list(X) ## .Internal(lapply(X, FUN)) ## } ## <bytecode: 0x55e8772de9b8> ## <environment: namespace:base>
You can sometimes find parts of the documentation that very vaguely hint to this misbehaviour, but such things are rarely specific or said at a non-expert level: Their meaning is only obvious in retrospect. On the rare occasion that the documentation is specific about the misbehaviour, it can be incomplete. For example, the documentation for
choose()
tells you that it behaves differently for smallk
, but what is “smallk
”? I think that it’s 29 or less, but that assumes that I’ve found the correct C code (I think it’s this?) and read it correctly. -
-
In the same vein to the
choose()
example, functions in the base stats library do not always tell you which calculation method they used. This can make you falsely assume that a figure was calculated exactly. For example,prop.test()
computes an approximation, but the only mention of this in its documentation is the See Also section saying “binom.test()
for an exact test of a binomial hypothesis”. Not only is this in a terrible place, it only suggests that an approximation has been used inprop.test()
. The details of the approximation are left for the reader to guess. -
Some functions act very strangely because they’re designed with S compatibility in mind. This issue goes on to damage the documentation for said functions. For example, have a look at the docs for the
seq()
function. It won’t tell you whatseq_along()
does, but it will tell you what to useseq_along()
instead of! I’ll let Stack Overflow explainseq.int()
’s documentation issues. Said documentation is so poor that I’ve been scared out of using the function. I really don’t know why R pays this price: Who is still using S? Another example is the**
operator. I’ll let the Arithmetic Operators documentation (try?'**'
) speak for itself on that one. Its three sentences on the topic are**
’s only documentation. Given that you shouldn’t use it, it would be harsh for me to say more. For further reading, I will only give this. -
As the previous example shows, backwards compatibility is a priority for R. This means that its inconsistencies will almost certainly never be fixed. Things would be better if the docs did a better job of helping you, but this section demonstrates ad nauseam that they do not. One wonders if there’s ever been any real interest in fixing it.
-
Some docs assume stats knowledge even when there should be no need to. If you don’t know what “sweeping out” is, you will never understand the docs for
sweep()
. I findrmultinom()
’s docs to be similarly lacking. It talks about “the typical multinomial experiment” as if you’ll know what that is. Its Details section tells you the mathematical technicalities, but if I wanted that then I would’ve gone to Wikipedia. All that they had to do was give an example about biased die and that would’ve told the reader all that they will need to know. A similar case can be made aboutrbinom()
, but I can forgive that on the grounds of “who uses R without knowing at least that much stats?”. -
The docs often do a bad job of linking to other relevant functions. For example,
match()
’s doesn’t tell you aboutPosition()
,subset()
,which()
, or the various grep things,mapply()
’s doesn’t tell you aboutMap()
, andrbinom()
’s doesn’t tell you aboutrmultinom()
. -
I sometimes can’t understand how to search for functions in the documentation. For example,
Filter()
’s docs are in the “funprog {base}” category, but putting?funprog
in to R won’t return those docs. Another oddity is that it’s sometimes case sensitive. For example,?Extract
works but?extract
doesn’t. In case you missed it, there is noExtract()
orextract()
function. -
I find that the documentation tries to cover too many functions at once. For example, in order to understand any particular function in the funprog or grep documentation, you’re probably going to have to go as far as understanding all of them. The worst case is the Condition Handling and Recovery documentation (
?tryCatch
), which lists about 30 functions, forever dooming me to never really understand any more of R’s exception system thanstop()
andstopifnot()
. A much smaller example is that bothabs()
andsqrt()
are documented in the same place, despite barely having anything in common and not sharing this documentation with anything else. This issue also compromises the quality of the examples that are given. For example, the funprog documentation gives no examples of how to useMap()
,Find()
, orPosition()
, something that never would have happened if they were alone in their own documentation pages. Then again,which()
andarrayInd()
are the only functions in their documentation, andarrayInd()
has no examples, so maybe I’m giving R too much credit. After all, like I hinted at earlier, even totally fundamental stuff like lists have more functions in their documentation than examples. -
The docs sometimes spend a distracting amount of time comparing their subjects to other languages that you might not know. The best example is the funprog docs, which are needlessly cluttered with mentions of Common Lisp. A close second to this is the documentation for pairlists, which even in the language definition have little more description than “Pairlist objects are similar to Lisp’s dotted-pair lists”. My favourite example is probably “regexpr and gregexpr with perl = TRUE allow Python-style named captures”, if only because it manages to mention two languages in a totally unexpected way. I should also mention that I’ve already complained about how some functions are so obsessed with S compatibility that both their documentation and functionality are compromised. As a final but forgiveable case,
sprintf()
is deliberately about C-style stuff and therefore never shuts up about C, making the R documentation pretty difficult for anyone who doesn’t know C. -
If pairlists are not really intended for use by normal users, why are they documented in the exact same place as normal lists, which are critical to normal R usage?
-
Guidelines for unusual operators, such as using
[
as a function, are rather hard to find in the documentation. One example that I found particularly annoying is in thenames()
documentation. It can’t make its mind up about whether it wants to talk about thenames(x) <- value
version or the"names<-"(x, value)
version. The only place where it’s apparent that there’s a meaningful difference between the two is in the second part of the Values section, which says:- “For
names<-
, the updated object. (Note that the value ofnames(x) <- value
is that of the assignment,value
, not the return value from the left-hand side.)”
…Wasn’t that helpful? You’ll only really be able to understand it if you understand the abstract notion of R’s replacement functions, but nowhere in the
names()
documentation will point you to that. In fact, unless you find the correct section of the language definition, you’re never going to find it at all (I’m not linking to that, go prove my point and find it yourself!). - “For
Don’t get me wrong, R’s documentation isn’t terrible. Its primary issue is that it does a poor job of helping you navigate R’s inconsistencies. If the examples were plentiful and the docs for each function linked to plenty of other related functions without themselves being cluttered with mentions of other functions and languages, then it would go a long way towards stopping R from tripping people up.
There are several inconsistencies in R’s functions and how you use them. This means that you either have to adopt a guess-and-check style of coding or constantly double-check the documentation before using a lot of R’s functions. Neither are satisfactory.
-
There are a few too many functions that have names synonymous with “do more than once”. There’s
replicate()
,repeat
loops, andrep()
. Good luck remembering which does what. -
Why do we have both
structure()
andstr()
orseq()
andsequence()
, all of which are different, while havingrm()
/remove()
andresiduals()
/resid()
, which are not? The potential for confusion is obvious: If I were to write a new function,Pos()
, should you or should you not assume that it’s an alias forPosition()
? -
There is no consistent convention for function names in the base libraries, even for related functions. I struggle to think of a function-naming scheme that isn’t found somewhere in R. For example, the documentation for
mean()
links to bothcolMeans()
andweighted.mean()
. Similarly, theseq()
documentation contains bothseq.int()
andseq_len()
. I also don’t like how there’s bothreadline()
andreadLines()
ornrow()
andNROW()
. Or how aboutall.equal()
andanyDuplicated()
? There’s even all of those functions with leading capitals likeVectorize()
or the funprog stuff. I could go on… -
The above issue gets even worse if we discuss functions that you’d expect to exist but don’t. For example, we have
write()
but notread()
(the equivalent is probablyscan()
). -
Argument names are also inconsistent. Most of the apply family calls its function argument
FUN
, butrapply()
and the funprog stuff usef
. -
Even when argument names are consistent, their behaviour may not be. For example,
complex(real = 1, imaginary = 2, length.out = 0)
andrep_len(complex(real = 1, imaginary = 2), length.out = 0)
do not have the same return value. If you ask me, it’scomplex()
that has the wrong behaviour here. I can’t see anywhere in its documentation mentioning that the other arguments can overrule thelength.out = 0
argument and give you vectors larger than what you asked for. At least throw a warning! -
Related functions sometimes expect their arguments to be given in a different order. For example, except for
mapply()
, the entire apply family wants the data to come before the function, whereas all of the funprog functions (e.g.Map()
,Filter()
, etc), want the reverse. When you realise that you picked the wrong function for a job, this makes rewriting your code infuriating. -
Functions that should be related in theory are not always related in practice. For example,
subset()
is not documented with the Set Operations (union()
,setdiff()
, etc) and works on completely different principles. The Set Operations are the extremely dangerous functions that remove duplicates from their inputs and applyas.vector()
to them. Thesubset()
function is a non-standard evaluation tool likewithin()
, making it completely different and dangerous in a different way. Finally, despite it being documented with the Set Operations, none of these warnings apply foris.element()
. I regret every time that I wrote off someone’s advice to usesubset()
because of my (entirely reasonable!) assumption that it would be a (dangerous) Set Operation. -
Functions with related names sometimes have different effects. For example, here is a damning quote from section 3.2.4 of Advanced R:
- “Generally, you can test if a vector is of a given type with an
is.*()
function, but these functions need to be used with care.is.logical()
,is.integer()
,is.double()
, andis.character()
do what you might expect: they test if a vector is a character, double, integer, or logical. Avoidis.vector()
,is.atomic()
, andis.numeric()
: they don’t test if you have a vector, atomic vector, or numeric vector; you’ll need to carefully read the documentation to figure out what they actually do.”
Another example is that
any()
,all()
, andidentical()
are all predicate functions, butall.equal()
andanyDuplicated()
are definitely not. - “Generally, you can test if a vector is of a given type with an
-
Similar to the above, from the solutions to Advanced R:
- “Note that
as.vector()
andis.vector()
use different definitions of ”vector!””.
The above quote is then followed by showing that
is.vector(as.vector(mtcars))
returnsFALSE
. I’ve found similar issues withas.matrix()
andis.matrix()
. - “Note that
-
The language can’t really decide if it wants you to be using lambdas. The apply family has arguments like
...
andMoreArgs
to make it so you don’t always have to do so, but the funprog stuff gives you no such choice. I almost always find that I want the lambdas, so the apply family’s tools to help you avoid them only serve to complicate the documentation. -
As an enjoyable example of how these inconsistencies can ruin your time with R, read the documentation for
Vectorize()
. It’s packed with tips for avoiding these pitfalls.
Let’s talk about matrices. I’ve already discussed some oddities like how
functions like [
, $
and length()
treat them in ways that seem
inconsistent with either the rest of the language or your expectations,
but let’s go deeper:
-
As covered earlier, matrices want to have rownames and colnames rather than names. This gives us a few more inconsistencies to deal with that I didn’t mention at the time. The rest of the language has trained you to use
setNames(data, names)
. When you do this,data
is returned with its column names changed without any changes todata
. However, matrices wantcolnames(data) <- names
and the obvious equivalent forrownames()
. This modifiesdata
and does not return it.a <- b <- diag(3) (colnames(a) <- c("I", "Return", "Me")) ## [1] "I" "Return" "Me" a#Changed ## I Return Me ## [1,] 1 0 0 ## [2,] 0 1 0 ## [3,] 0 0 1 setNames(b, c("I", "Return", "b")) ## [,1] [,2] [,3] ## [1,] 1 0 0 ## [2,] 0 1 0 ## [3,] 0 0 1 ## attr(,"names") ## [1] "I" "Return" "b" NA NA NA NA NA ## [9] NA b#Not changed ## [,1] [,2] [,3] ## [1,] 1 0 0 ## [2,] 0 1 0 ## [3,] 0 0 1
Not only are the function names inconsistent (why not
colNames()
?), the syntax is wildly so. Also, take a look at the incomprehensible error message thatcolnames()
gives if you usediag(3)
directly rather than assigning it to a variable beforehand.a <- diag(3) colnames(a) <- c("Not", "A", "Problem") ## > colnames(diag(3)) <- c("Big", "Bad", "Bug") ## Error in colnames(diag(3)) <- c("Big", "Bad", "Bug") : ## target of assignment expands to non-language object ## > colnames(a <- diag(3)) <- c("Has", "Similar", "Problem") ## Error in colnames(a <- diag(3)) <- c("Has", "Similar", "Problem") : ## object 'a' not found
setNames()
has no such issue.setNames(diag(3), c("Works", "Just", "Fine")) ## [,1] [,2] [,3] ## [1,] 1 0 0 ## [2,] 0 1 0 ## [3,] 0 0 1 ## attr(,"names") ## [1] "Works" "Just" "Fine" NA NA NA NA NA NA setNames(a <- diag(3), c("Works", "Just", "Fine")) ## [,1] [,2] [,3] ## [1,] 1 0 0 ## [2,] 0 1 0 ## [3,] 0 0 1 ## attr(,"names") ## [1] "Works" "Just" "Fine" NA NA NA NA NA NA
In truth, I don’t mind either
colnames()
orsetNames()
. I just wish that R would pick one way of handling names and stick to it. -
Unlike anything else in R that I can think of, matrices are happy to let you work by row and even have dedicated functions for it, with
rowSums()
andapply(..., MARGIN = 1)
being the obvious examples. There is a good reasons for this difference – matrices are always one type, unlike things like data frames – but it’s still an inconsistency. This inconsistency leads to code that is tough to justify. For instance, I frequently find that I want to treat the output ofexpand.grid()
as a matrix.unique(t(apply(expand.grid(1:4, 1:4, 1:4, 1:4), 1, sort)))
is one of my recent examples. This isn’t too bad, but I honestly have no idea why I needed thet()
. Experience has taught me not to question it, which is pretty bad in of itself. R’s inconsistency eventually makes you either fall in to the habit of not questioning sudden transformations of your data or forces you to become completely paralysed when trying to understand what ought to be trivial operations in your code. Doubts like “is there really no better way? R is supposed to be good with this sort of stuff” become frequent when wanting to work by row.- March 2023 update: I don’t think that there’s anywhere in R’s docs
that tell you that
expand.grid()
is used to make Cartesian products. This compares poorly to languages such as Racket, which calls the practically equivalent functioncartesian-product
. Similarly, Python just calls itproduct
. In both cases, they return a collection of lists/tuples where each list/tuple would be a row inexpand.grid()
’s data frame. Nested collections aren’t the easiest things to deal with, but they seem to produce more intuitive code than the messes that you get from trying to treat a data frame like a matrix. Maybe the lesson here is that you should iterate through a data frame’s rows the hard way rather than using functions that let you think of it as a matrix? I think that this explains the doubt that I’ve mentioned above; R’s data structures have a habit of making the right way the hard way.
- March 2023 update: I don’t think that there’s anywhere in R’s docs
that tell you that
-
So what happens if, when manipulating a matrix, you write the
sapply()
that the rest of the language has taught you to expect? At best, it gets treated like a vector in column-order.(mat <- matrix(1:9, nrow = 3, byrow = TRUE)) ## [,1] [,2] [,3] ## [1,] 1 2 3 ## [2,] 4 5 6 ## [3,] 7 8 9 sapply(mat, max) ## [1] 1 4 7 2 5 8 3 6 9
At worst, it doesn’t do anything like what you wanted.
mat <- matrix(1:9, nrow = 3, byrow = TRUE) sapply(mat, sum) ## [1] 1 4 7 2 5 8 3 6 9
The trick for avoiding this is to use numbers as your data argument and let subsetting be the function.
mat <- matrix(1:9, nrow = 3, byrow = TRUE) sapply(1:3, function(x) sum(mat[x, ])) ## [1] 6 15 24 sapply(1:3, function(x) max(mat[x, ])) ## [1] 3 6 9
Better yet, just use
apply()
.mat <- matrix(1:9, nrow = 3, byrow = TRUE) apply(mat, MARGIN = 1, sum) ## [1] 6 15 24 apply(mat, MARGIN = 1, max) ## [1] 3 6 9
But why did we have the learn any of this in the first place?
-
Your turn: What does
seq_along(diag(3))
return?1:3
or1:9
? What if you added a row? What if you added a column? Or is the name of that functionseq.along()
? Are you sure? Tempted to check the docs? Which docs? Feeling helpless? You should! -
Many functions that are designed for matrices should be forgotten about everywhere else. Several guides warn against using
apply()
on non-matrices and I wouldn’t dare uset()
on a non-matrix. Tryt(iris)
. -
I always expect
c()
of a matrix to work in row-order. It doesn’t. However, that’s probably more the fault ofc()
and I than it is of matrices. There are times when I can’t explainc(mtcars)
to myself. -
Named matrices are named atomic vectors, so they break in the ways discussed earlier. This puts you in a dilemma when you’re using data that’s essentially only one type: Do you keep it as a matrix and lose the awesome subsetting powers of a data frame? Or do you make it in to a data frame and lose the power to work by row that matrices give you? At times, I’m tempted to forget that I named the matrix in the first place and just manipulate it like a mathematician. None of these solutions are good.
Overall, matrices are so inconsistent with the rest of the language that your matrix-manipulation code never looks right. It leaves you with an awful sense of unease.
Something to mention while we’ve still got some bad error messages fresh
in our minds: People often say that R’s error messages aren’t very good
and I’m starting to agree. Errors like “dim(X)
must have a positive
length” are useless when you’re not told which function in the line
that threw the error had that error, what X
is, or in the very worst
cases, what line the error was even in. This means that almost any error
that R throws is going to require you looking through both the result of
traceback()
(to find where the error happened) and the documentation
(to identify the problematic argument). It seems that this issue gets
even worse when you try to do statistics. Warnings like “Warning
message: In ks.test(foo, bar)
: ties should not be present for the
Kolmogorov-Smirnov test” don’t even tell you where the tie was. Was it
in one of my arguments? Is it some technical detail of the test?
Somewhere safe to ignore? You don’t know and R won’t tell you unless you
study the documentation. Worst come to worst, you have to read the code
or learn the secret for getting traceback()
to work on warning
messages. And yes, that last bit is something that you have to learn. It
makes warnings messages a lot harder to debug than errors.
Of course, the more worrying (and frequent?) issue is when R gives you no warnings/errors at all. I’d much rather have a bad error message than none at all, but a bad error message is still annoying.
Maybe you think I’m clutching at straws? I admit, I sometimes wonder if my outrage is unjustified. Let’s settle this with a challenge. If you win, then by all means close this document and write me off as a madman. If you lose, then maybe I’ve got a point.
CHALLENGE
Taking in to account R’s vector recycling rules, figure out how
mapply()
’s MoreArgs
and ...
arguments differ and when you would
want to pass something as a MoreArgs
argument rather than in the ...
argument. No cheating by going online (trust me, it won’t help). Solve
this without leaving your R IDE. You’re encouraged to check the
documentation.
If my criticisms are true, you will find that mapply()
’s documentation
is of little help and that your confidence in your R knowledge is too
small to make an educated guess.
HINT 1
Don’t try to cheat by looking at mapply()
’s code; Most of it is in C
and therefore will be of no help to you.
HINT 2
You might think that the documentation for sapply()
will help you, but
it’ll actually mislead you because mapply()
’s ...
is essentially
sapply()
’s X
and sapply()
’s ...
is most like mapply()
’s
MoreArgs
.
Solution below. Time to stop scrolling.
SOLUTION
.
.
.
.
.
.
.
How do MoreArgs
and ...
differ?
It’s tough to explain. mapply()
uses the default vector recycling
rules for the ...
arguments but reuses every element of MoreArgs
for
each call. Because the MoreArgs
argument must be a list and R recycles
the elements of lists (e.g. using a length one list as a ...
argument
will have the element of that list reused for each call), the difference
is subtle to the point of near invisibility. Ultimately,
MoreArgs = list(a, b, c)
is equivalent to using list(a)
, list(b)
,
and list(c)
as three separate ...
arguments. The answer is therefore
that MoreArgs
only exists as syntactic sugar for this ...
case.
When use MoreArgs
rather than ...
?
Beyond what I’ve already said, I barely have any idea. If you want to
keep some function arguments fixed for each call, then just use an
anonymous function. I struggle to invent a useful example of where I’d
even consider using MoreArgs
, never mind one that doesn’t look
tailor-made to make the anonymous function option look better. The one
and only example that the documentation gives for using MoreArgs
does
not help here. Their example of
mapply(rep, times = 1:4, MoreArgs = list(x = 42))
is identical to
mapply(rep, times = 1:4, list(x = 42))
. Read that again: You can get
identical functionality by deleting the thing that they’re trying to
demonstrate!
Bonus
Did you notice that the documentation for mapply()
has a notable
omission? It doesn’t mention this, but you can call mapply()
without
the ...
argument, e.g. mapply(rep, MoreArgs = list(1:4))
. You won’t
get sensible output, but you also don’t get any warnings or errors.
If I’ve won this challenge, then allow me to take a victory lap by
making the following point: By giving you the options of using ...
,
MoreArgs
, or an anonymous function to do the same task, R gives you
plenty of room to confuse yourself without providing any help in its
documentation. Either provide fewer options, document them better, or
make them so commonplace and consistent within the language that I only
need to understand it once in order to understand it everywhere!
On top of many of the things that I’ve already said about the apply
family, fans of the Tidyverse, particularly purrr
, often point out of
the following inconsistencies. They’ve never bothered me, but they’re
undeniably correct. It makes me wonder why we can’t just give all of the
apply family a simplify
argument that takes either TRUE
, FALSE
, or
whatever vapply()
would consider a valid FUN.VALUE
argument.
- Probably due to its use of
as.vector()
,apply()
has nosimplify
argument. Version 4.0.6 did something about this, but I’ve yet to get my head around it. - There is no equivalent to
vapply()
for any member of the apply family other thansapply()
. Among others, neithertapply()
normapply()
have one. - You can argue that some functions are missing in base R. For example,
if we think of
lapply()
as “list in, data frame out” andby()
as “data frame in, list out”, and so on forsapply()
and others, then where is the “array in, list out” function? - This might be a repeat of my complaints about the documentation, but
there are several apply family functions that hardly anyone uses or
understands. Few understand
eapply()
and even fewer use it. Just about everyone who uses R for stats has had to invest a few hours getting their head aroundtapply()
, but at least that’s worth it. As for the other obscure ones – e.g.simplify2array()
,rapply()
– I honestly cannot recall ever using them or seeing them used.
There are some community issues that make R harder to learn and work with. Put together with the earlier issues, it means that help can often neither be found inside nor outside of R.
- The community can’t agree on if we should be using base R’s data
frames, the Tidyverse’s tibbles, or the
data.table
package. Even if you find one that you like, you can bet that you will someday want to use a package that requires another. For example, if your friendly neighbourhood package author made use of the defaultdrop = TRUE
argument when manipulating data frames, you’re not going to be allowed to use tibbles. Protecting the user from this isn’t easy, because both data.tables and tibbles returnTRUE
foris.data.frame()
. - The community is also split on which OOP system to use. For example,
R6
gets about as much mention as RC, even though they both do the same job. Depending on when they were made, you also see some popular libraries that are fully committed to some particular OOP system. For example, you will see a lot of S3 in base R, but theBioconductor
package sticks to S4. Fortunately, all of this only becomes a problem when you want to contribute to these packages. If all goes well, you will never really notice which OOP system has been used; You will just have polymorphic code and not need to question it. - A lot of people seem to treat R as secondary to the Tidyverse
packages. Many books essentially ignore base R or go out of their way
to tell you why the Tidyverse is better – Advanced R is a good
example of the latter, particularly the second edition, so much so
that I needed to read both editions – and many R Q&A sites will give
you a Tidyverse solution to pretty much any problem, even if it can be
solved almost as well in base R. You sometimes see the same with
data.table
, but it’s much less common. - The popularity of the Tidyverse is a major blow to your motivation to learn R. Why would anyone want to learn a language that is treated as secondary to some packages? Worse still, if that turns out to be the best way to use R, then you’re forced to admit that R is a polished turd with a fragmented community. Why would you ever knowingly use a polished turd? The popularity of the Tidyverse is possibly the strongest piece of evidence that not only does R suck, the community knows it.
- If base R is best ignored in favour of packages, or at least if the community thinks so, then how are you expected to actually learn R? When do you stop learning the base library and move on?
- Data science people have a strong preference for Python, so a lot of good tutorials that should be in R are only available in Python. I’ve also noticed at least one case where the R package documentation is clearly inferior to the Python equivalent, despite the gap in functionality not being as wide.
- R’s functional programming aspects are so strong that loops,
particularly
for
loops, are considered a code smell. This goes double in the Tidyverse, with the R for Data Science book not even introducing them until chapter 21 (of 30). There are practical reasons for this, mostly in relation to the apply family’s code being written in C and therefore being faster than most R loops. However, it encourages you to do some silly things. For example, you have to make a judgement call between writing afor
loop that is inherently fast but slowed down by being written in R or writing ansapply()
that ought to be slow but is speeded up due tosapply()
calling C code. This issue also affects how you present your code. Calls to the apply family are inherently one-liners, so it’s difficult to find the right way to present/comment them when they become complex. You either end up introducing unnecessary variables in to your code or indenting it in unconventional ways. The Tidyverse’s solution – piping – often does the trick, but it openly admits to not being a universal solution. The new|>
base R pipe doesn’t do the trick either. As any advocate of the Tidyverse will tell you, base R just isn’t designed for piping.
R’s generic function OOP systems are yet another source of unpredictability and internal inconsistency. They’re very cool and I must admit that I’ve not used them much, but what I’ve seen when trying to use them has discouraged me. Most of what I’m about to say is about S3, but you’ll rarely find much said about R’s OOP systems at all. It’s not really any surprise. S3, S4, RC, and any of the OOP systems that come from packages are all openly admitted to being bolted-on to R rather than something that was part of its design from the early days. Points like the below make discovering this fact unavoidable. Presumably, this is what the Julia fans are talking about when they say that they’re the only ones who have a generic function OOP system that is baked-in to their language. I’ve never used Julia or enough S4 or RC to be able to really comment, but I bet they’re right.
The class system is a mess and the docs do a poor job of explaining it.
Good luck understanding it without a book like Advanced R and a
package like sloop
. I believe that this problem is mostly isolated to
S3, but I’ve not used enough S4 to be able to say that with any
certainty. Here are some problems that you’re likely to encounter early
on:
-
Functions like
mode()
andstorage.mode()
exist only for compatibility with S. As an R user, they exist only to increase your confusion. This is particularly common when reading the language definition; It never shuts up about the modes of things. -
Advanced R makes a strong case for
is.numeric()
being inconsistent, particularly regarding its interaction with S3. -
The documentation for
class()
uses vague statements like “method dispatch may use more classes than are returned byclass(x)
”. May? MAY??? What am I supposed to do with that? Where do I look for more info? It mentions.class2()
, but warns you against using it. Why? It doesn’t say! Did you think that.class2()
would have its own documentation somewhere else? It doesn’t! All of the documentation for.class2()
is in the docs forclass()
and most of that is a warning to not use it! -
Call
class()
on a matrix and you will see that they have a few classes. However,is.object()
, which has docs that correctly state that it will returnTRUE
for anything with a class attribute, returnsFALSE
.a <- diag(3) class(a) ## [1] "matrix" "array" is.object(a) ## [1] FALSE
Why? Because
class()
also returns implicit classes – as detailed in its docs – whichis.object()
ignores because implicit classes aren’t part of the class attribute. The documentation foris.object()
does not mention this fact and theclass()
function’s output does not tell you which classes are implicit. Can you see the potential for confusion? The docs never lied or even mislead, but they make naivety fatal. Maybe that’s starting to become a theme. -
So what base R function actually returns the non-implicit classes? I think that you have to use
attr(foo, "class")
. I say “I think” because the documentation forclass()
does not offer any help. -
Don’t ask what determines the implicit classes of an object or how S3 dispatch occurs with them. It’s far too complicated and not clearly documented in any place that I know of. It’s also what’s used for dispatching on anything without a class attribute, such as matrices. Good luck with that!
-
Don’t ask what an object is either. The community will tell you that
is.object()
is poorly named and the language definition will tell you that pretty much everything in R is an object. However, there are functions likenames(x)<-
that will not work on several types of objects (e.g. anything anonymous) despite their documentation saying thatx
can be “an R object”. I’d give examples, but you really don’t want to think too hard about this.- Update: I’ve recently skimmed some books by John Chambers. He also
claims that everything in R is an object. I think that what this
really means is that
class()
always returns something?
- Update: I’ve recently skimmed some books by John Chambers. He also
claims that everything in R is an object. I think that what this
really means is that
-
The
[
and[[
functions like to drop the attributes from your S3 objects, meaning that you almost always have to write a[
and[[
method for them. On the bright side, this is documented behaviour.
The generic functions that you’ll find in the base and other common libraries have a few surprises:
- Some functions appear to be generic, but aren’t. For example, both
abline()
andsample()
can behave differently depending on what sort of input you gave them, but that’s because they’re hard-coded to do so. If you expect to find someabline.lm()
function or be able to write your ownabline.myClass()
method, you’ll be disappointed. - In stats libraries, I’ve seen the ability to overload function names
abused. For example, why is the function to return a one-hot encoded
dataset in the
caret
librarypredict()
? I’m pretty sure that my Bayesian Statistics lecturer also showed me a few cases whereanova()
definitely does not do an ANOVA. I’m out of practice, but I think that the R FAQ gives one such example. - Although you can usually write your own functions to fix this, it’s
annoying when you expect a generic function to behave in a sensible
way on your input, only to discover that it has no special case for
your type of input and treats it like any other vector. To repeat an
earlier example, why does
rep(matrix(1:4, nrow = 2, ncol = 2), 5)
treat the input matrix like it’s a normal vector? I can’t imagine anyone callingrep()
on a matrix and wanting to work element-by-element rather than repeating the matrix. - Because of S3, it’s now considered bad practice to write function
names in the form
foo.bar()
because they look like extensions tofoo()
. Open up any standard R library and you will see countless functions written in this form that are not extensions to anything.t()
andt.test()
are the most cited example.
This is tough to explain, but I’ll try. If you consult
?"internal generic"
, you will find a big list of functions that you
cannot extend properly with S3. Specifically, anything on that list
cannot be extended to dispatch on any object for which is.object()
returns FALSE
. For example, writing a rep.matrix()
function that
does what my earlier example wanted is easy, but because rep()
is on
that list and matrices are not objects in this sense, rep()
will not
dispatch to rep.matrix()
when given a matrix. The documentation for
rep()
and other functions that share this misbehaviour do not do much
to help the reader discover this fact. Advanced R has the only good
explanation that I’ve found for this, but it’s the sort of thing where
you either have to read two
chapters or the first edition’s OO
Field Guide. The short
explanation is “internal generics are written in C and therefore only
understand non-implicit classes and whatever internal R type the C code
ultimately gets fed”.
-
Did you notice that
length()
is on the list mentioned above? This explains the inconsistency in how it behaves on data frames and matrices. You cannot properly extend internal generics with S3, so you cannot change howlength()
behaves on implicitly classed input. Data frames are non-implicitly lists and matrices are non-implicitly atomic vectors, so that’s howlength()
treats them. This issue isn’t unique to just data frames and matrices. Advanced R has a comical example in its S3 chapter: Linear models, which are non-implicitly lists, have a length of about 12! My personal favourite is giving it quoted input:length(quote(5^30)) ## [1] 3 length(quote(5^30 + 1)) ## [1] 3 length(quote(5^30 + 12)) ## [1] 3 length(quote(1)) ## [1] 1 length(quote(length(1))) ## [1] 2
You will find issues like this – i.e. unexpected and tough to explain output – whenever using or extending most internal generics;
length()
is just the easiest example to show. -
Did you notice that I lied about data frames? The careful reader will notice that due to having the real
data.frame
class, data frames don’t have any implicit classes. That’s a thing, by the way, having a real class means not having implicit classes.attr(mtcars, "class") ## [1] "data.frame" class(mtcars) ## [1] "data.frame" is.object(mtcars) ## [1] TRUE
This means that
length(someDataFrame)
cannot possibly dispatch to somelength.list()
internal method. Further inspection reveals that there is no S3 (i.e. non-internal)length.data.frame()
method. What actually happens is that R tries to findlength.data.frame()
, fails, and then tries to findlength.default()
, only to fail again and get pointed to the internal C code that presumably treats data frames just like lists. This happens even though data frames do not have implicit classes. Enjoying the complexity? -
So what happens if you try to write a
length.data.frame()
method – something that is totally allowed because data frames returnTRUE
foris.object()
andlength()
is an internal generic function – and havelength()
dispatch to it? You’ll probably break R. I once redefined the length of a data frame to be its number of rows and I got a stack usage error. Please, take a few seconds to appreciate all of the complexity that we’ve had to work through just for R’s most basic object system.
Much of the above makes the class system – and therefore S3 dispatch – impossible to clearly explain. Any real explanation would be so full of exceptions that it would become incomprehensible. The only way to explain it is to ignore the contradictions for as long as possible, meaning that you must be given incorrect information until you’re ready to read about the exceptions. This ultimately means that you cannot even find a good reference manual for the class system, because you never know if you’re reading the whole truth or not. Furthermore, if this is at all representative of the complexity of S3, how can anyone be expected to have the patience to even begin learning S4? I know that it’s dishonest to blame S4 for the sins of S3, but I wouldn’t blame any newcomer to R’s OOP for doing so. One wonders if we should start newcomers on S4 and leave S3 until much later. The books by John Chambers take this approach and generally say to stick to S4.
Let’s talk about S4. I promise that this will be an easier read than the earlier sections. I’m quite ignorant of S4, as I’ve already admitted to, so I’ve got very little to say. Regardless, the following seems clear:
-
Everything that I’ve read about S4 gives me the impression that it has far fewer stupid technicalities than S3. If I’m right, then I find that laughable. How have we managed to make S3 more complicated than S4? S3 should be extremely simple, but the technicalities of the previous few sections are too easy to stumble upon.
-
If Advanced R’s chapter on S4 is to be trusted, then the official documentation for S4 contains a lot of bad advice. I’ve not looked closely, but I have noticed that it shares R’s tendency to put many functions in one page of documentation and then not give examples for many of them. For example,
?getMethod
documents five functions, but only gives examples for two. Similarly,@
has no examples in its documentation. -
S4 has some strange semantics. Why call something that is sometimes not a predicate function
is()
? Why does it use an@
operator to do what the rest of R would use$
for?is(mtcars) ## [1] "data.frame" "list" "oldClass" "vector"
-
As far as I can tell, S4 doesn’t inform you if there was some ambiguity in your dispatch, such as if it had to pick one option from two equally appropriate potential dispatches. I think that unless there is no appropriate method to dispatch to, it has some internal rules that silently handle these cases, meaning that there is no ambiguity even when there probably should be. In other words, it may misbehave by silently resolving the developer’s ambiguities. Without being too spiteful, by now I find it quite easy to believe that R has an OOP system that silently misbehaves.
Section 8.2 of The R Inferno calls the factor and ordered variables “chimeras”. This is exactly the right criticism. Under the hood, they’re S3 objects with integers as their base type and a character vector – the levels – as an attribute. When using these variables, it is difficult to predict if R will treat them as their integer base type, as their character vector levels attribute, or as a factor object. And that’s not even mentioning how the labels come in to it. The R Inferno has said more than I will and gives some examples of their unpredictable behaviour, but here are some points from my own experience:
-
There is no base R function for extracting the original object from its corresponding factor. To extract your original set of numbers (assuming that they were numbers, if not, you get nonsense) from a factor variable called
f
, the documentation tells you to use eitheras.numeric(levels(f))[f]
or the sloweras.numeric(as.character(f))
. Let’s use a bit more code than usual and show off what each of these functions do before and after composition:(withoutLabels <- factor(rep(seq(from = 2, by = 2, to = 10), 3))) ## [1] 2 4 6 8 10 2 4 6 8 10 2 4 6 8 10 ## Levels: 2 4 6 8 10 (withLabels <- factor(rep(seq(from = 2, by = 2, to = 10), 3), labels = LETTERS[1:5])) ## [1] A B C D E A B C D E A B C D E ## Levels: A B C D E fList <- list(withoutLabels, withLabels) #Just to make sure that we're on the same page, here's the output of str(). #The internal integers are in plain sight. lapply(fList, str) ## Factor w/ 5 levels "2","4","6","8",..: 1 2 3 4 5 1 2 3 4 5 ... ## Factor w/ 5 levels "A","B","C","D",..: 1 2 3 4 5 1 2 3 4 5 ... ## [[1]] ## NULL ## ## [[2]] ## NULL #Nothing surprising to start: lapply(fList, levels) ## [[1]] ## [1] "2" "4" "6" "8" "10" ## ## [[2]] ## [1] "A" "B" "C" "D" "E" #as.character() returns the non-attribute part of what you get when you print the factor #i.e. the result of mapping its internal integers to its character vector of levels. #Notice that these are characters. It's not obvious from printing your factors that #the non-attribute part becomes a character. lapply(fList, as.character) ## [[1]] ## [1] "2" "4" "6" "8" "10" "2" "4" "6" "8" "10" "2" "4" "6" "8" "10" ## ## [[2]] ## [1] "A" "B" "C" "D" "E" "A" "B" "C" "D" "E" "A" "B" "C" "D" "E" #Calling `as.numeric()` on a factor does not return the original numbers. #It returns the underlying integers. #Why would you ever want or expect these? lapply(fList, as.numeric) ## [[1]] ## [1] 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 ## ## [[2]] ## [1] 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 #Subsetting with factors always treats them as their integer base type. #If the factor fundamentally has nothing to do with integers #(e.g. if you made the factor from something that was originally a set of characters), #then you can expect nonsense. #If the factor did originally have something to do with integers, #then you're probably going to be very confused because it hasn't subsetted with #the numbers that you get from printing the factor. #In short, it's almost never a good idea, but R lets you do it anyway. #Now ask yourself: What is the point of having a categorical data type if #it's not practical to subset with? lapply(fList, function(f) levels(f)[f]) ## [[1]] ## [1] "2" "4" "6" "8" "10" "2" "4" "6" "8" "10" "2" "4" "6" "8" "10" ## ## [[2]] ## [1] "A" "B" "C" "D" "E" "A" "B" "C" "D" "E" "A" "B" "C" "D" "E" lapply(fList, function(f) as.numeric(levels(f))[f]) ## Warning in FUN(X[[i]], ...): NAs introduced by coercion ## [[1]] ## [1] 2 4 6 8 10 2 4 6 8 10 2 4 6 8 10 ## ## [[2]] ## [1] NA NA NA NA NA NA NA NA NA NA NA NA NA NA NA lapply(fList, function(f) as.numeric(as.character(f))) ## Warning in FUN(X[[i]], ...): NAs introduced by coercion ## [[1]] ## [1] 2 4 6 8 10 2 4 6 8 10 2 4 6 8 10 ## ## [[2]] ## [1] NA NA NA NA NA NA NA NA NA NA NA NA NA NA NA
-
As you’ve probably noticed by now, factor variables are inherently complex enough that they need you to either carefully read their documentation or be an R master before you can use them with confidence. You cannot tell me that
as.numeric(levels(f))[f]
made perfect sense when you read it or that you would have come up with it yourself. It’s arcane. Half of the reason why I let the code speak for itself above, rather than adopting my usual bullet point style, is because I hardly even trust myself to describe them. In fact, even whoever wrote the R FAQ seems to have not mastered the art. In section 7.10, they suggestas.numeric(levels(f))[as.integer(f)]
for the same task as what we’ve covered above. Can you see the redundant function call? -
When writing example code, factors want to be called
f
, just like functions do. This offends me. -
On the bright side, it looks like R version 4 is steadily trying to fix factors. Every few updates, we see a minor change. For example, before version 4, you had to pass
stringsAsFactors = FALSE
to a lot of functions. This was to stop R creating factor when you hadn’t asked for them. It was widely considered extremely annoying because there is nothing in the way that data frames print that signals to the reader that they’re looking at a factor variable. For all you knew, you were looking at a character vector. You often would not discover your mistake until you had a serious error.
Personally, I’m afraid to use factor variables. Their unpredictability makes any code that uses them dramatically more complex, even if you’re confident that you know their rules.
The syntactic sugar is a source of problems, often to such a great degree that your best solution is to completely avoid the sugar. I’ll start with some small cases before splitting some of the bigger ones in to sections.
-
You usually only see this when dealing with
names()
, but having a function that is both a setter and getter is a guaranteed source of confusion and found more than once in R. For example,names(output)
will give you the names ofoutput
, butnames(output) <- c("Alice", "Bob")
will changeoutput
’s names (it’s sugar for some complicated"names<-"
nonsense).names(mtcars) ## [1] "mpg" "cyl" "disp" "hp" "drat" "wt" "qsec" "vs" "am" "gear" ## [11] "carb" names(mtcars) <- LETTERS[1:11] head(mtcars, 2) ## A B C D E F G H I J K ## Mazda RX4 21 6 160 110 3.9 2.620 16.46 0 1 4 4 ## Mazda RX4 Wag 21 6 160 110 3.9 2.875 17.02 0 1 4 4
Now what do you think
names(foo) <- names(bar)
does? Seriously, can you guess? I can think of roughly four realistic guesses. Is it even valid syntax? Here’s the truth:names(mtcars) <- LETTERS[1:11] a <- rep(c("example", "text"), length.out = 11) names(a) <- LETTERS[12:22] a ## L M N O P Q R S ## "example" "text" "example" "text" "example" "text" "example" "text" ## T U V ## "example" "text" "example" names(a) <- names(mtcars) a ## A B C D E F G H ## "example" "text" "example" "text" "example" "text" "example" "text" ## I J K ## "example" "text" "example"
A lot of people seem to make the correct guess here, but syntax shouldn’t leave you guessing. Where possible, I try to stick to
setNames()
. -
The syntactic sugar sometimes leads to surprising syntax. For example
names(output[2]) <- "foo"
doesn’t work, butnames(output)[2] <- "foo"
does. -
As is extremely well documented,
T
andF
can be used in place ofTRUE
andFALSE
, but you should never do this becauseT
andF
are just variables that can be overwritten in your code. Why let us do something that we never should? To my surprise, there’s actually a sensible answer. Section 3.3.3 of the R FAQ says that S hadT
andF
as reserved words, but R changed that to allow variables called"T"
and"F"
to appear in your datasets. I can see the reasoning behind both S’s approach and R’s change to it, but I still think that R’s approach of “you can do this, but never do” is obviously wrong. My suspicion is that the third option of “just makeT
andF
not mean anything until they’re assigned to” will never be taken, because the current (and dangerous) approach helps with backwards compatibility. I don’t think that it’s a good trade. -
Although I like R’s many functional programming tools, the temptation to try to use them to solve every problem is very strong. I’ve wasted countless hours trying to pick the right one of
sapply()
/lapply()
/mapply()
/Filter()
/Map()
… (not to mention their various arguments) when I really should’ve just written thefor
loop. This is more my fault than it is R’s, but it’s a curse that every intermediate R user will suffer from. It’s a price that any R expert will tell you was worth it in the end. However, it’s still a price that I don’t enjoy paying. It wouldn’t be so bad if R had less such functions, better error messages, or more consistency between these functions, but we’ve already discussed that can of worms. Don’t think that I’m advocating forpurrr
here. It has so many functional programming tools that it arguably makes the situation worse. I’ll cover its costs and benefits later.
The :
operator is absolutely lovely… until it screws you. The solution
is to prefer the seq()
functions to using :
. Some quick examples:
-
Stuff like
i in 1:n
is great, but if you accidentally haven <= 0
, it silently gives behaviours that you probably don’t want.1:2 ## [1] 1 2 1:1 ## [1] 1 1:0 ## [1] 1 0 1:-1 ## [1] 1 0 -1
seq_len()
is better behaved, so I try to stick to it.seq_len(2) ## [1] 1 2 seq_len(1) ## [1] 1 seq_len(0) ## integer(0) #seq_len(-1) is an error.
-
:
has operator precedence issues. You might expect stuff like-6/2:3
to generate(-(6/2)):3
i.e.-3:3
. It doesn’t.-6/2:3 #Treated as -6/(2:3) ## [1] -3 -2
I’ll leave
-6/2:6/2
as an exercise for the reader. I’d like to keep things simple and say that the trick is that:
is always evaluated first, but that’s actually not true. Even if we’re only talking about arithmetical operations, exponentiation is done before:
is applied.3^1:5 #Treated as (3^1):5 ## [1] 3 4 5
-
Can you guess what
data[-1:5]
returns? I can’t either, so don’t ever try it. If you must know, it’s actually an error.
As I’ve said, seq()
and its related functions usually fix this issue.
The only real disappointment with seq()
itself is that its
documentation warns against not naming its arguments, so you’re forced
to write the long-winded seq(from = 0, to = 100, by = 6)
rather than
just seq(0, 100, 6)
.
The documentation for several functions with non-standard evaluation,
e.g. with()
and subset()
, explicitly warns the user to not use them
when programming. This is a source of a number of problems, both
practically and in a meta sense:
-
First of all, the existence of functions that are not for programming use is abhorrent.
-
They’re excellent syntactic sugar, so I hate not being able to use them! For example, I’d argue that
within(data, rm(colName1, colName2))
is the best way to remove unwanted columns from my data: It does not require me to quote or escape my column names, does not require me to putdata$
before everything, does not require me to pass that annoyingdrop = FALSE
argument, warns me if I am trying to remove a column that is not in my data, and reads almost like English. All in all, that’s some major advantages over using[
. They both have some misbehaviour if your column names are duplicated, but that’s not very relevant here. -
The documentation for a lot of the functions that use non-standard evaluation warn you to take care with them, but they do very little to tell you how or why. I’ve searched high and low, but I sincerely believe that there is almost nothing in R’s documentation that tells you what can go wrong with these functions. Seriously, if you can find it, let me know. I’ve even read the Thomas Lumley article that some of the docs tell you to check and I still can’t find much of relevance. As with
.class2()
, I find R’s habit of putting unexplained warnings in its documentation deeply maddening. -
Because you don’t know when it’s safe to use these functions and when it isn’t, you feel incredible anger when the perfect solution to your problem is to use one of them. You get in situations like “I could either easily do this with
with()
or write out anmapply()
with a long-winded anonymous function…” and always have to choose to do things the hard way. It makes you want to never use R outside of the REPL. This is one part where the Tidyverse completely destroys base R. -
So what can actually go wrong with them? To be honest, I don’t really know. A lot of these functions internally rely on a function called
substitute()
, which has special behaviour when it tries to interact with anything defined in the global environment, so it’s slightly difficult to invent easy-to-type examples of these functions misbehaving. All that I’ve managed to find is:- I’ve been told that if something exists in the calling environment of your function with non-standard evaluation, but not in the data that you’re trying to work on, then you’re in trouble. I’ve yet to be able to invent an example that shows unexpected behaviour.
- Code like
subset(data, exampleCol > x)
will misbehave ifx
is a column indata
but you intended it to come from the calling environment.
head(airquality) ## Ozone Solar.R Wind Temp Month Day ## 1 41 190 7.4 67 5 1 ## 2 36 118 8.0 72 5 2 ## 3 12 149 12.6 74 5 3 ## 4 18 313 11.5 62 5 4 ## 5 NA NA 14.3 56 5 5 ## 6 28 NA 14.9 66 5 6 head(subset(airquality, Wind * 5 > Temp)) ## Ozone Solar.R Wind Temp Month Day ## 5 NA NA 14.3 56 5 5 ## 6 28 NA 14.9 66 5 6 ## 8 19 99 13.8 59 5 8 ## 9 8 19 20.1 61 5 9 ## 15 18 65 13.2 58 5 15 ## 18 6 78 18.4 57 5 18 Temp <- 90000000 head(subset(airquality, Wind * 5 > Temp))#Identical to the previous call. ## Ozone Solar.R Wind Temp Month Day ## 5 NA NA 14.3 56 5 5 ## 6 28 NA 14.9 66 5 6 ## 8 19 99 13.8 59 5 8 ## 9 8 19 20.1 61 5 9 ## 15 18 65 13.2 58 5 15 ## 18 6 78 18.4 57 5 18
I think that this is the case that
with()
’s documentation is trying to warn you about.subset()
’s documentation does not appear to contain any such warning, unless you’re generous enough to count “the non-standard evaluation of argument subset can have unanticipated consequences”. How hard would it have been to give an example like the one that I’ve just given?- Advanced R tries to explain some other issues, but the extent of them has never been completely clear to me.
Of all the problems that I’ve written about, this section’s probably
bother me the most. So many of R’s problems could be sidestepped if we
could fearlessly use with()
and subset()
at all times, but R’s nasty
habit of not explaining the dangers that it warms you of leaves me in
constant paranoia.
Some things seems obviously missing from R:
-
For a Scheme-inspired language, the lack of any tail call optimisation or any macro system is strange. Then again, being a heavily functional language that looks like C is one of the best things about R. If it had tail call optimisation or Lisp-like macros, it’d probably start to look more like a weird statistical version of Lisp.
-
You can only break out of the innermost loop. Unless you refactor, there’s no way to be many loops deep and break out of them all with one command.
-
R has no
do-while
loop. It’s never bothered me, but I think that’s because I’ve never used one in any language. I can see it bothering others, but if I need one, then I’m pretty sure that they’re trivial to make from arepeat
loop. -
Without crude
if(FALSE){}
workarounds, there’s no way to comment out blocks. IDEs can fix this. -
Outside of packages, R lacks any real dictionary, associative array, or linked list type. The closest that we can get is matching elements to their names like this. I’ve always thought that it seems like a hacky way to get what other languages have built in. You can also do it with environments, which apparently has O(1) lookup, but I’ve never seen anyone do it. That may have something to do with how the base R syntax for creating environments from scratch isn’t as nice as its syntax for creating lists. You have to name and assign each element individually, e.g.
e <- new.env(); e$a <- 1; e$b <- 2; e$c <- 3
, rather than justl <- list(a = 1, b = 2, c = 3)
. And if you’re going to use a package to fix this syntax issue, then you might as well just use one that gives you actual hash tables.- It looks like version 4.1.4 might be fixing this?
-
Given that R is a maths/stats language, I find the follow omissions surprising:
- There’s no base function for counting the number of possible permutations of a collection of objects.
- Despite there being a function for finding the combinations that you
can make from the elements of a vector, there’s no function that
does that with repetitions. For example,
combn(1:3, 2)
can’t be convinced to includec(1, 1)
,c(2, 2)
, andc(3, 3)
.expand.grid(1:3, 1:3)
comes close, but that trick generates permutations rather than combinations. - There’s no built-in big integer class.
- Despite supporting equations, R offers no obvious way to simplify them.
- Although R has some matrix functionality built in, there’s no
is.square()
function. - There is no base R function to check if a number is a whole number.
is.integer()
checks for integer typing rather than if the input is in of itself an integer. The docs even show thatis.integer(1)
isFALSE
. Worse still, these docs actually show you the code for a goodis.wholenumber()
function! Why couldn’t that be in the base library?
-
Once you’re aware of it, the previous issue starts coming up in weird places. This suggests that R’s missing something in its error checks. Take a look:
seq_len(4.8) #Not an error ## [1] 1 2 3 4 1:4.8 ## [1] 1 2 3 4 a <- 1:10 a[4.8] ## [1] 4 a[-4.8] ## [1] 1 2 3 5 6 7 8 9 10 sample(4.8) ## [1] 3 2 1 4
The pattern is that R silently truncates the numeric index of choice towards 0.
-
The base libraries have no obvious dedicated functions for pivoting. You can do it with
tapply()
, but nothing in the docs would make you guess that. In fact, virtually every occurrence of the word “pivot” in R’s docs is talking aboutchol()
. I think that you can pivot/unpivot withstack()
/unstack()
, but the only time I’ve ever seen those functions mentioned was in this SQL article.
Admittedly, few if any of these are major, but they’re a bit annoying.
And now for everything that I’ve got left in the bag.
-
The two language problem: Sooner or later, you’ll run in to a memory issue, go to Stack Overflow, and be told that the solution is to use a package that lets R talk to C++. Julia claims to have solved this. I don’t know if I believe it.
-
The index in a
for
loop uses the same environment as its caller, so loops likefor(i in 1:10)
will overwrite any variable calledi
in the parent environment and set it to10
when the loop finishes.i <- 2000 for(i in 1:10){} i ## [1] 10
This sounds awful, but I’ve never encountered it in practice. After all, it sounds like bad practice to use the same variable name for two different things. Apparently the
for
loops also like to strip attributes, breaking S3 objects, but again, I’ve never encountered this. After all, idiomatic R it to prefer functions likesapply()
tofor
loops. -
Advanced R claims that R is a great language to metaprogram. I cannot deny that the Tidyverse is very strong evidence for that, but who would dare metaprogram a language as poorly documented and as inconsistent as I’ve claimed R is? Certainly not me. I can’t even predict R’s behaviour when I’m programming it, never mind metaprogramming! I’ve regretted most of my attempts at doing so. I usually get tripped up by some quirk of R’s string-manipulation facilities and how the strings get parsed as expressions.
-
For a language that was inspired by Scheme, R’s metaprogramming feels very limited. As far as I can tell, aside from the typical operation of building code from text that I’d expect any language to be capable of, it is only used to facilitate the creation of functions that evaluate their arguments in a non-standard way. Usually, this doesn’t go any further than creating an ad-hoc environment where the function’s arguments make sense, despite said arguments having no meaning in the calling environment. Typical examples are
with()
and modelling functions likelm()
, which let you write code likelm(mpg ~ wt, mtcars)
. Being able to say “let me tell you what data I want you to treat like an environment, so I can refer to its variables as if they were objects in the calling environment” is great, but it’s nowhere near what a Lisp user would expect. -
The
plot()
function has some strange defaults. For example, you need to have a plot before you can plot individual points, and it often doesn’t know what to do in terms of how long/wide its axes should be. I also don’t like how “predictmpg
fromwt
” isfoo(mpg~wt)
, but “plotmpg
on the y-axis andwt
on the x-axis” isplot(wt, mpg)
. I understand why both options are the way that they are, but it creates unpredictability. -
I seem to have terrible luck with the documentation for R’s libraries. Even when using popular packages that have been around for years, I often find documentation errors that are so basic that I can’t explain how they’ve gone unnoticed. I’ve seen documentation that reports the wrong return types, imports unnecessary libraries in its example code, and completely fails to mention significant parameters! I try to fix these when I find them, so I can no longer name names, but it’s a source of significant annoyance.
-
Advanced R points out that good R code is rare, but I have a different take on it that I think explains my poor luck with R libraries: Statisticians don’t want to write code or learn GitHub and programmers don’t want to use any more R than they strictly need to. This means that nobody is really doing any bug fixing or even reporting. On the bright side, this makes it very easy to improve other people’s R code and get accepted pull requests.
-
For obscure and maybe machine-dependent C reasons, you will not be able to guess the return value of
log(complex(real = Inf, imaginary = Inf))
. I also don’t quite like howcomplex(real = Inf, imaginary = Inf)
prints. SeeingInf+Infi
makes me think that I’ve forgotten some variable called “Infi”. Luckily, I’ve never had a reason to really bother with R’s complex numbers.
As I’ve already admitted, my knowledge of the Tidyverse is much less
than my knowledge of base R. However, no critique of R is complete
without at least giving this a mention. Its popularity, along with R
version 4.0.6. adopting some its ideas (pipes and a shorter anonymous
function syntax), are clear evidence that it’s on to something. Before
going in to the specific libraries, I’ve given some general thoughts
below. You may also be interested in Hadley Wickham’s comments on this
section.
Particularly with regards to purrr
, I was surprised by how much we
agree (look inside the changes made in the pull request).
- I can’t back down from the “polished turd” point that I made
earlier. No matter how good the Tidyverse is, any
attempt to fix R’s inconsistencies by making new libraries is doomed
to fail. Base R is inconsistent, so the only way to be consistent with
it is to be inconsistent. For example,
as.foo()
is inconsistent with R’s S3 system, but it’s what I’d expect to find with a new class calledfoo
in a library. The only solution to this problem is to somehow write code that completely ignores base R, but that becomes impossible as soon as you try to load anyone else’s packages. - I’m sure that I’ve seen the main author of the Tidyverse quoted as
saying that he didn’t want it to be a monolith. However, it undeniably
is one. Tidyverse packages will throw errors with
rlang
, be made specifically to work with other Tidyverse packages (see the first paragraph of the manifesto), and deprecate their own functions in favour of functions from different Tidyverse packages. - For the reasons explained earlier, the
authors are very scared of the
...
argument’s ability to pass arguments to where they should not have gone. To counter this, most of the Tidyverse functions use names that you would never type. This means that without an IDE prompting you, you’re going to get a lot of the argument names wrong. I’ve slipped up withtibble
’s.name_repair
argument a few times. Get it wrong and R probably won’t let you know! - I really understand the “consistent interface” selling point of the
Tidyverse. Because of the
[x]
/[x,]
/[,x]
business, I often guess wrong with functions likebase::order()
, but I almost never guess wrong withdplyr::arrange()
. - The API of the Tidyverse is rarely stable; It constantly deprecates
functions and owns up to its willingness to do
so.
I understand the value of not being tied to your past mistakes – see
my many complaints about base R’s interest in supporting S – but it’s
rare that I look through the documentation for a Tidyverse package and
not see a note saying that something either is deprecated or soon will
be. It’s even worse when I see a Stack Overflow answer with just the
function that I need, only to find that my newer version of the
package doesn’t even have the old function. However, the worst example
by far is when the R for Data Science book can’t keep up with the
deprecation. For example, when I read chapter 25, the code in the book
was spitting out warnings about the
.drop
argument tounnest()
being deprecated. Importing a package when making your own package is already a risky proposition, but issues like this would have me do all of my work in base R even if there was a perfect Tidyverse function for the job. - The Tidyverse is undeniably designed around piping (see the second
point of the
manifesto).
It’s not obvious that piping is a good framework to build around. To
keep pipes simple, you must build functions that are easy to compose.
This means that you will design your functions to have the minimum
number of arguments that you can get away with. If you need more
arguments, then you will instead make more functions wherever
possible. This is a significant increase in complexity. Surely
arguments are less complicated than functions? I dread to think what
it takes to replicate a function like
aggregate()
in the Tidyverse. Even something as simple asdplyr::select()
has about 10 helper functions in its documentation. I’m willing to be proven wrong here, but everything that I’ve just said strikes me as obviously true to anyone who has useddplyr
orpurrr
.- Hadley’s comments, linked above, point to the
tidyr
package as a strong counterexample to my claim that the argument count must be minimised in a pipe-based design. They also mention that there’s no obvious better way to designdplyr::select()
. On all counts, I have no counterargument. However, I’m confident that I’m still on to something here, even if my original points are wrong. Pipes must come at a cost, but it appears that I’ve incorrectly identified what that cost is.
- Hadley’s comments, linked above, point to the
Overall, I’m more than happy to use Tidyverse functions when I’m writing some run-once code or messing around in the REPL, but the unstable API point is a real killer for anything else. In terms of how it compares to base R, I’d rather use quite a few of its packages than their base R equivalents. However, that doesn’t mean that it can actually replace base R. I see it as nothing more than a handy set of libraries.
Now for the specific libraries. Assume that I’m ignorant of any that I’ve skipped.
- I don’t like how it has name conflicts with the base R stats library. It just seems rude.
- Remember all of my complaints about base R’s
subsetting and how I’d rather use the non-standard
evaluation functions if it weren’t for
all of their vague warnings?
dplyr
completely nullifies most of these complains, for data frames at least. This is a huge win for the Tidyverse. - R’s factor variables are scary.
dplyr::group_by()
takes a more SQL-like approach to the problem and feels a lot safer to work with. - Compared to base R, the knowledge that
dplyr
will only output a tibble is a relief. There’s no need to consider if I needtapply()
,by()
, oraggregate()
for a job or if I need to coerce my input in order to go in/out of functions liketable()
. I therefore need to do a lot less guessing. This link demonstrates it better than I can, although the formula solution withaggregate()
is in base R’s favour. dplyr::mutate()
is just plain better than base R’stransform()
. In particular, it allow you to refer to columns that you’ve just created.- This
link
shows a comparison between base R and
dplyr
. It’s rather persuasive. In particular, it gives you the sense that you can make safe guesses about thedplyr
functions. - I need to play around with it more, but I think that
pivot_wider()
’svalues_fn
argument makesdplyr
the only tool that I’ve ever seen that allows arbitrary functions in a pivot table. - I don’t like how the
dplyr
functions only accept data frame or objects derived from them. If I’m doing some work with something likestringr
, I instinctively want to use a Tidyverse solution to problems like subsetting my inputs. However, if I reach fordplyr::filter()
, I get errors due to character vector not being data frames. This isn’t reallydplyr
’s fault and they shouldn’t try to fix it, but it’s still a minor annoyance.
- To repeat my earlier praise for this library, it’s fun. That’s a huge win.
- It has amazingly sane defaults. Whenever I make the same graph in both
this and base R’s
plot()
,ggplot2
’s is much better. You can tell R to do stuff like include a useful legend or grid, butggplot2
does it by default. - I like how graphs are made by what amounts to composing functions. It makes it very easy to focus on one specific element of your plots at a time. I’d even go as far as say that it’s fun to see what happens when you replace a component with another valid but strange one. You can discover entirely new categories of graphs by accident.
- I miss the genericness of base R’s
plot()
. When I can’t be bothered to think about what sort of plot I need,plot()
can save me the trouble by making a correct guess. There is no such facility inggplot2
. Hadley’s comments have pointed outautoplot()
, but I’ve never gotten it to work. There are no examples in its documentation and I’ve not found all that much help online.
I don’t use dates much, so I don’t have much to say. In fact, I don’t
think that I’ve mentioned them before now. lubridate
appears to be
much easier to use than base R dates and times, but I know neither very
well. I don’t like how base R seems to force you to do any time
arithmetic in seconds and date arithmetic in days. I also don’t like how
it’s hard to do arithmetic with times in base R without a date being
attached. However, I could be wrong about all of that. I really don’t
know any of them too well and I’ve never found much reason to learn. I’d
really like to see
https://rosettacode.org/wiki/Convert_seconds_to_compound_duration
solved in both base R and lubridate
. I’m not saying that it would be
hard, but I’d love to see the comparison. Overall, all that I can really
say for certain is that experience has shown that when the day comes,
I’ll have a much easier time learning this package than what base R
offers for the same jobs.
Pipes come in very handy, but I’ve never been completely sold on them,
even when viewing teaching examples that are supposed to demonstrate
their superiority. I’ll admit that there is a time and a place for them
– e.g. printing and graphing code – but I think that they only really
shine when you’ve abandoned base R in favour of purrr
. After all, base
R wasn’t built for pipes. I’d even go as far as to say that the people
who swear by magrittr
and purrr
have adopted a completely different
paradigm to those who don’t, so they end up using totally different
tools. For example, a master of the Tidyverse finds Advanced R
chapter nine’s
by_cyl %>%
map(~ lm(mpg ~ wt, data = .x)) %>%
map(coef) %>%
map_dbl(2)
just as informative as my
lapply(by_cyl, function(x) lm(mpg ~ wt, data = x)$coef[[2]])
(or its
equivalent sapply()
or vapply()
, if you really insist). Overall, I
think that I can’t evaluate magrittr
without also evaluating purrr
.
As a side-note, the claim that foo %>% bar()
is equivalent to
bar(foo)
appears to be a white lie. Try it with a plotting function
that cares about the variable name of its argument. Spot the difference:
plot(Nile)
library(magrittr)
##
## Attaching package: 'magrittr'
## The following object is masked from 'package:purrr':
##
## set_names
Nile %>% plot()
Nile |> plot() #The same as plot(Nile)
Don’t get me wrong, I like pipes a lot. When you’re dealing with data,
there’s sometimes no way to avoid “do foo()
to my data and then do
bar()
” code. However, you’d be mad to use them all of the time. For
people that do use them, all that I can say is that you should take the
time to learn all of them and that said time really isn’t much. None of
them are much more complicated than %>%
and %$%
is a handy
replacement for with()
.
As a final point, I don’t like how much trouble base R’s new |>
pipe
causes me. You can’t do x |> foo
. You instead need x |> foo()
. Also,
to use a function where the target argument isn’t first, you need to use
some nonsense like x |> (function(x) foo(bar, x))()
. For example,
mtcars |> (function(x) Map(max, x))()
. I don’t like all of those extra
brackets. magrittr
can do it with just
mtcars %>% (function(x) Map(max, x))
or even mtcars %>% Map(max, .)
.
Regardless, base R’s pipe is still new, so perhaps I’m judging it too
early. It appears that future versions will expand it.
Unlike base R, where I can point to a specific example and explain to
you why it’s silly, my objections to purrr
are mostly philosophical.
It certainly does some things much better than base R. For example, I
really like the consistency in the map functions. They’re a breath of
fresh air compared to base R’s apply family vs funprog mess. My code
would probably be a lot easier to read and modify if I replaced all of
my apply family and funprog calls with their purrr
equivalents.
However, when writing the code in the first place, I’d much rather have
the flexibility that the base R functions offer. I also like pluck()
and it being built in to the map functions, but I’ve yet to get used to
it. Overall, I wouldn’t mind using purrr
, but I have some major
objections:
- It takes the idea of “a function should only do one thing” to a
pathological level. I see no reason why
map_lgl()
,map_int()
,map_dbl()
, andmap_chr()
are separate functions. They do exactly the same thing, but will throw an error if they don’t get the return type in their name. Why isn’t this the job of some general mapping function that takes the desired output type as an argument (e.g. like base R’svapply()
)? This same issue is found in the entire library. There is no need for themap2()
andpmap()
functions or their countless _type variants. Just make a generalmap()
function! To steal a point from the TidyverseSkeptic essay,purrr
has 178 functions, and 52 are maps. What would you rather learn: a handful of complex map functions (like base R) or 52 simple ones? The only defences that I’ve seen for thepurrr
approach is that base R can be a bit verbose, e.g.vapply()
’s arguments likeFUN.VALUE = logical(1)
, and that using the most restrictive possible tool for any given job increases the readability of your code. - It makes base R’s
~
operator able to form anonymous functions, at least withinpurrr
(it’s some funky parsing). I could get used to it, but I don’t like how it robs the user of the ability to give the arguments to their anonymous function any meaningful names. This is because thepurrr
authors thought that the normal anonymous function syntax was too verbose, but I’d argue that they’ve gone too far and made their syntax too terse.Map(function(x) runif(1), 1:3)
is not long or particularly obscure, butmap(1:3, ~ runif(1))
crosses the line for me, as doesmap(data, ~ .x * 2)
. My example in the previous section, which includedmap(~ lm(mpg ~ wt, data = .x))
, demonstrates another problem: It overloads the~
operator in a dangerous way. The~
inside themap()
is very different from the~
in the call tolm()
. - I suspect that my above two points interact. Could it be that
purrr
users don’t use a generalised map function because they’ve written off base R’s anonymous function syntax and replaced it with a variant that is so terse that their code becomes unreadable without the names of their functions telling the reader what they’re doing?
Overall, I could probably be convinced that purrr
’s way is better than
base R’s, but I doubt that purrr
’s way is the best way.
For me, these both fall in to the same box. They’re not particularly
outstanding, but they’re clearly much better than their base R
equivalents. I’ve praised tibble
enough already and have said plenty
about the state of R’s strings. The only thing that I’ve
got left to say is that once you’ve noticed that tibbles let you use the
columns that you’ve just defined to define other columns, you really
start to hate how many extra lines of code you have to write when using
data frames for the same task.
If I were being generous, I would say that R teaches you some great lessons about functional programming while being a useful DSL and that its biggest fault is that it tries to do too much, ultimately becoming brutally inconsistent. I’d also say that the Tidyverse is a useful set of packages that, while unable to fix R and certainly not a panacea, do a lot to improve it within their specific domains. However, I’m not that generous.
The most damning thing about R is that much of The R Inferno still holds true. The fact that a nearly decade-old document that credibly compared R to a journey in to Hell is still a useful reference manual speaks volumes about the language’s attitude to change. To put it plainly, R won’t change. If something about R frustrates you today, it always will. That’s what kills the language for me. The popularity of the Tidyverse proves that R is broken and the continuing validity of the The R Inferno proves that it will stay that way. You may be able to put a blanket of sanity on top of it, as the best packages try to, but you won’t fix it. Unless you find said packages so useful that they make R worth it, I find it impossible to argue against jumping ship. My ultimate conclusion on R is that it’s good, but doomed by the unshifting weight of the countless little problems that I’ve documented here. Personally, I’m going to give Python a shot and I wouldn’t blame you for doing the same. Let’s hope that I don’t end up writing a document of this size complaining about that.
All that being said, I have no intention of uninstalling R or going out of my way to avoid it. I’d gladly use it professionally and I’ve learned enough of its semantic semtex to get really damn good at using R to do in few lines what other languages would do in many. I wasn’t joking when I said that it’s the best desktop calculator that I’ve ever used. But would I recommend learning it to anyone else? Absolutely not. We can do so much better.
In late March 2022, this article suddenly exploded overnight. It was briefly in the top 10 on Hacker News and got about 20,000 views in one day. This came as quite a shock to me, given that I wasn’t even done proofreading yet. I’ve read through as much of the online commentary on this article as I can find. The comments on the Hacker News page are by far the most in-depth. You can find a fair bit on Twitter and Reddit as well, but you’d have to go looking. I found that Twitter had the most positive reception, Reddit was more negative and Hacker News was mixed.
In April, I finished the proofreading and sent this off to R-devel. I am very thankful for their comments and their sincere efforts to help me. I hope that my replies didn’t come off as half hearted. I was simply unequipped to handle to mass of feedback.
I won’t single out any particular commenters, but there are some ideas and trends that I feel are worth addressing:
-
From the negative feedback, I can’t help but wonder if some of my examples were too trivial or too petty. This document would have been a lot easier to read and write if I only mentioned big issues. I’ve made some minor edits to address this, but it’s hard to judge. A master would never make some of the mistakes that my subsetting section warns against, but does that mean I shouldn’t even mention those issues?
-
I find it interesting to note what hasn’t been criticised. The following examples stand out to me:
- Although many said they found my ‘
"es"
in"test"
’ challenge a bit too easy (I think they missed my point), I could only find one person who made any attempt at mymapply()
challenge. - If my complaints about R not telling you what the dangers of non-standard evaluation are had simple resolutions (e.g. a hyperlink), then I’d expect someone to have provided them. I’ve yet to see any feedback on this topic. The same is true of what I’ve said of generic functions.
- Despite some fair criticism of my section on subsetting, I don’t think that anyone mentioned any disagreements with what I’ve said about the the vector recycling.
- I don’t think that anyone disagreed with what I said about R’s documentation and error messages.
It could just be luck that these parts weren’t mentioned anywhere that I found, but you’ll forgive me for concluding that the lack of criticism implies my points were very strong.
- Although many said they found my ‘
-
A very common objection was that my Ignorance section invalidates much of my commentary. Of course, said ignorance makes me unable to know if they’re right or not. The two most common criticisms were that my lack of expertise in the Tidyverse and/or
data.table
mean that I’ve got nothing worthwhile to say and that using R as a programming language rather than a statistics tool is fundamentally wrong. All of these criticisms are partly correct. Using R for interactive data analysis is very different from trying to program with it, so such users simply won’t encounter many of the issues that I’ve mentioned. Similarly, swapping base R for the Tidyverse automatically nullifies many of my complaints. You can even go through my table of contents and cross sections off.dplyr
and its tibble-focus already knock off most of my complaints about base R’s variable manipulations, data types, subsetting rules, and vector rules. Don’t get me wrong, the Tidyverse has its own problems. For example, I’d hate to develop anything reliant on the Tidyverse’s unstable API. However, if you’re doing a run-once piece of analysis, then it’s probably great. It’s just a shame to see so much of R replaced by its packages.- December 2022 update: I’m further in to my career as a software
developer than I was when I wrote this and I’m starting to put more
and more weight on the above point. Every R developer I’ve
encountered professionally has been an exclusive user of the
Tidyverse. I remember one person whose only memory of her R training
was that you “need to type
library(tidyverse)
before anything works”. I have a growing suspicion that using R for anything other than the Tidyverse is steadily – and perhaps even correctly – becoming seen as simply wrong. It’s unfortunate that I sincerely love many parts of it.
- December 2022 update: I’m further in to my career as a software
developer than I was when I wrote this and I’m starting to put more
and more weight on the above point. Every R developer I’ve
encountered professionally has been an exclusive user of the
Tidyverse. I remember one person whose only memory of her R training
was that you “need to type
-
I’ve perhaps undersold just how good R can be at what it’s specialised for. This chain of Hacker News comments seems to get across something that I haven’t. I’ve certainly said that R is a large mathematics and statistics tool that is easy to extend and has clear Scheme inspiration, but the sum of those comments seems to say it better. As for the idea that R is a ‘Worse is Better’ language, I find it appealing but I don’t feel qualified to judge. If anything was ‘Worse is Better’, then it was probably S (which would make R “almost the right thing”, in that essay’s terms). However, I’m not historically knowledgeable enough to know key factors like how simple S’s early implementations were. I hear that it was very easy to get running on Unix?
-
I never made it clear that I understand why backwards compatibility is a priority for R. For example, R code appears in a lot of science papers and you don’t want such code to become unrunnable or to change meaning.
As a final point, making the changes to this document to reflect the changes coming in what I presume to be R version 4.1.4 has forced me to question my points about R being unable to change. I’ve not changed my mind yet, but time will tell. They certainly prove that R can change, but I think the real issue might be that it can’t fundamentally change.