When writing research code I do test my code and results, but until recently I've only been doing this informally and not in any systematic way. I decided it was time to change my testing habits when I noticed I had recoded a variable incorrectly despite my informal tests suggesting nothing was wrong. When I went back and corrected this error this made a small, but noticeable, difference to my model.
I didn't want to write formal unit tests that required setting up new folders and files, etc., as I would use this as an excuse not to switch tabs and write the tests. My other concern is that I am often testing large dataframes that I remove from memory once checked and merged into something else, so they don't persist for external testing (like a function). For my own needs I just wanted to write a test inline with the script, which would halt the execution of the script if the test did not pass. I would suggest this approach would work for many researchers writing code as scripts in environments like R or Python.
Writing informal tests formally
For R users a great tool I found is the
testthat package by Hadley Wickham. It's available on CRAN so is as easy as any other package to install. Stick near the top of your script:
Really the package is written for package authors who want to automate their testing, but by dropping a few bits it can easily be integrated into a script and the tests are performed when sourceing the file.
The testthat documentation outlines how to write a test, and the syntax to use. As an example, the following test checks to see that the right number of rows and columns are present, assuming we know our data should have 1,000 respondents and 20 variables, respectively:
Clearly this is a trivial example and you probably wouldn't test to this detail, but this does show you the outline of the test format. I use a new context for each group of tests focused around an individual object or function. For example, I would write a new context when no longer testing
df. Within each context you can then have as many test_that() calls, and within each test_that() calls you can have as many expect_that() calls as you like. This effectively allows you to give your test groups ‘headings’ to remind you of their purpose and structure. These headings are useful when the tests fail because they form part of the error message returned to you so you can easily spot what's wrong.
The advantage of using testthat in this way is that it stops execution of the script if any tests fail. Assuming you write good tests, you can be confident that if the script finishes that there are no errors.
Writing informal tests systematically
Think about your unit tests
Of course, tests don't write themselves so you need to cover this yourself. This takes practice, am I don't claim to have mastered writing unit testing yet. The bits of advice I have picked up through trial and error and reading others’ work, though, are:
- Test systematically. Test everything. Test every time you create a new object (function, data frame, variable, vector, etc).
- Of course, some objects are pretty straightforward, so use your judgement when applying suggestion 1! If in doubt, err on the side of testing.
- If writing a function, write some tests that compare the output with what you expect the output to be.
- If testing a data frame check you have the right number of respondents (rows) and variables (columns) if necessary. Test to see that there are no responses that do not have an incorrect value. For example, if your variable should be binary (0/1) only check there are no 2s.
- If the class of an object is important, test to see if it is that class.
- When you recode a variable, test these have been recoded correctly.
- Prefer tests that are comprehensive over tests that sample, if possible. For example, write tests that check a whole vector rather than a selection of values in that vector.
- Test what you can, and you'll get better with practice. Testing some things is better than testing no things.
Why bother testing?
Well, clearly it's up to you if you test. There is an investment upfront in learning to test, but it's not that great. You can be up and running writing testthat tests within half a day, tops. In my case, that investment has saved me two things:
- Time. When I come to re-use a function (or copy+paste a section of code) I have the tests in place so I can run the new code immediately and know if it's working or not without having to scrutinise every line of code or test a number of outputs from the function. They're written, and they'll fail if there's a problem.
- Worry. I am now much more confident in my results than ever before, because I know that I've made sensible tests along the way. Crucially I also know that if I ever need to change anything the tests will inform me if there's a problem, rather than having to check everything carefully or hope for the best_._