Chapter 18 Write your own R functions, part 1

18.1 What and why?

My goal here is to reveal the process a long-time useR employs for writing functions. I also want to illustrate why the process is the way it is. Merely looking at the finished product, e.g. source code for R packages, can be extremely deceiving. Reality is generally much uglier … but more interesting!

Why are we covering this now, smack in the middle of data aggregation? Powerful machines like dplyr, purrr, and the built-in “apply” family of functions, are ready and waiting to apply your purpose-built functions to various bits of your data. If you can express your analytical wishes in a function, these tools will give you great power.

18.3 Max - min

Say you’ve got a numeric vector, and you want to compute the difference between its max and min. lifeExp or pop or gdpPercap are great examples of a typical input. You can imagine wanting to get this statistic after we slice up the Gapminder data by year, country, continent, or combinations thereof.

18.4 Get something that works

First, develop some working code for interactive use, using a representative input. I’ll use Gapminder’s life expectancy variable.

R functions that will be useful: min(), max(), range(). (Read their documentation: here and here)

Internalize this “answer” because our informal testing relies on you noticing departures from this.

18.4.1 Skateboard >> perfectly formed rear-view mirror

This image widely attributed to the Spotify development team conveys an important point.

Build that skateboard before you build the car or some fancy car part. A limited-but-functioning thing is very useful. It also keeps the spirits high.

This is related to the valuable Telescope Rule:

It is faster to make a four-inch mirror then a six-inch mirror than to make a six-inch mirror.

18.5 Turn the working interactive code into a function

Add NO new functionality! Just write your very first R function.

Check that you’re getting the same answer as you did with your interactive code. Test it eyeball-o-metrically at this point.

18.6 Test your function

18.6.1 Test on new inputs

Pick some new artificial inputs where you know (at least approximately) what your function should return.

I know that 10 minus 1 is 9. I know that random uniform [0, 1] variates will be between 0 and 1. Therefore max - min should be less than 1. If I take LOTS of them, max - min should be pretty close to 1.

It is intentional that I tested on integer input as well as floating point. Likewise, I like to use valid-but-random data for this sort of check.

18.6.2 Test on real data but different real data

Back to the real world now. Two other quantitative variables are lying around: gdpPercap and pop. Let’s have a go.

Either check these results “by hand” or apply the “does that even make sense?” test.

18.6.3 Test on weird stuff

Now we try to break our function. Don’t get truly diabolical (yet). Just make the kind of mistakes you can imagine making at 2am when, 3 years from now, you rediscover this useful function you wrote. Give your function inputs it’s not expecting.

How happy are you with those error messages? You must imagine that some entire script has failed and that you were hoping to just source() it without re-reading it. If a colleague or future you encountered these errors, do you run screaming from the room? How hard is it to pinpoint the usage problem?

18.6.4 I will scare you now

Here are some great examples STAT545 students devised during class where the function should break but it does not.

In both cases, R’s eagerness to make sense of our requests is unfortunately successful. In the first case, a data.frame containing just the quantitative variables is eventually coerced into numeric vector. We can compute max minus min, even though it makes absolutely no sense at all. In the second case, a logical vector is converted to zeroes and ones, which might merit an error or at least a warning.

18.7 Check the validity of arguments

For functions that will be used again – which is not all of them! – it is good to check the validity of arguments. This implements a rule from the Unix philosophy:

Rule of Repair: When you must fail, fail noisily and as soon as possible.

18.7.2 if then stop

stopifnot() doesn’t provide a very good error message. The next approach is very widely used. Put your validity check inside an if() statement and call stop() yourself, with a custom error message, in the body.

In addition to a gratuitous apology, the error raised also contains two more pieces of helpful info:

  • Which function threw the error.
  • Hints on how to fix things: expected class of input vs actual class.

If it is easy to do so, I highly recommend this template: “you gave me THIS, but I need THAT”.

The tidyverse style guide has a very useful chapter on how to construct error messages.

18.8 Wrap-up and what’s next?

Here’s the function we’ve written so far:

What we’ve accomplished:

  • We’ve written our first function.
  • We are checking the validity of its input, argument x.
  • We’ve done a good amount of informal testing.

Where to next? In part 2 we generalize this function to take differences in other quantiles and learn how to set default values for arguments.

18.9 Resources

Packages for runtime assertions (the last 3 seem to be under more active development than assertthat):

  • assertthat on CRAN and GitHub - the Hadleyverse option
  • ensurer on CRAN and GitHub - general purpose, pipe-friendly
  • assertr on CRAN and GitHub - explicitly data pipeline oriented
  • assertive on CRAN and Bitbucket - rich set of built-in functions

Hadley Wickham’s book Advanced R (2015):