Chapter 17 Danger Zone

little things to include somewhere: - the danger of misspecified arguments disappearing into ...

don’t give your model lm class and count on other stuff to magically work. your should implement methods specifically for your class, and if that is just wrapping lm internals, great, but take explicit control

provide an example of extensive formula tests: include a log term in tests to catch formula edge cases

17.1 Warnings / places to use care

model.frame() is not the be all end all

double evaluation issue in

fit <- lm(hp ~ log(mpg), mtcars)
predict(fit, newdata = model.frame(fit))

## Error in eval(predvars, data, env): object 'mpg' not found

don’t do predict(object, newdata = model.frame(object)) since this will breaaaaaaaak

17.2 Anti-patterns

17.2.1 Using the default method of a generic

i.e. funneling everything into confint.default. Default methods for new generics should throw an error.

What do to about existing generics?

Key principle here: want to guarantee to the user that they are getting the right numbers

don’t funnel everything through augment_columns and add special cases slowly - v hard to maintain. much better to have individualized S3 methods with consistent behavior abstracted out into small helpers.

related idea to the belong: have enough classes! using inheritance appropriately.

17.2.2 the documentation that isn’t documentation and doesn’t feature an actual use case

• cough ?MASS::predict.rlm cough
• misisng doc ?MASS:::predict.polr()

17.2.3 Never use missing arguments

because people have to write more code to pass the objects they want

predict(m) and predict(m, newdata = NULL) should do the same thing you should test this

17.2.4 special casing everything through one workhorse function instead of using S3 methods

basically augment_columns

don’t funnel everything through augment_columns and add special cases slowly - v hard to maintain. much better to have individualized S3 methods with consistent behavior abstracted out into small helpers.

17.3 Things to be aware of

If you call things data or df and users fail to specify data arguments, R will try and perform dataframe operations on the functions data and df. The resulting error messages for this can be cryptic. In this case you may wish to write an informative error message with a hint:

Error: Can't *do data frame thing* on a function.
Are you sure you passed a tibble to *argument*?