Don't test assumptions that are not required

[mattansb]

By model type.

Here are some examples that should not work. (@bwiernik Can you double check me?)

m <- glm(am ~ mpg, mtcars,
         family = binomial())

performance::check_normality(m)
#> OK: residuals appear as normally distributed (p = 0.059).


m <- glm(gear ~ mpg, mtcars,
         family = poisson())

performance::check_normality(m)
#> Warning: Non-normality of residuals detected (p = 0.001).

performance::check_heteroscedasticity(m) # fails properly
#> Error in check_heteroscedasticity.default(m): This Breusch-Pagan Test currently only works Gaussian models.

^{Created on 2021-04-08 by the reprex package (v1.0.0)}

[bwiernik]

Yeah, this is a good point.

performance::check_normality should return a message indicating that no normality assumption is needed when family isn't normal.

For heteroscedasticity, we could direct the user toward more appropriate checks, such as overdispersion.

It would be good to make a list of families each check is relevant for.

I don't know that family-mismatch problems like this necessarily warrant errors--messages are probably more appropriate. Similarly, are warnings really needed when the assumption is violated? That might produce weirdness if someone has set warnings = error. Perhaps messages would again be better?

[strengejacke]

when family isn't normal.

Does this include/exclude inverse-Gaussian and similar? And what about istrumental/2SLS, or truncated/censored?

[bwiernik]

Need to think through the various models/families that we support. Not sure off hand about inverse-Gaussian. IV/2SLS doesn't have a normality assumption I don't think. Tobit models have a normality assumption on the latent variable, so a different test is needed.

[strengejacke]

We could have a vignette that simply lists the available check-functions depending on model type, like:

• tests for linear regresion: check_..., ...
• tests for logistic regresion: binned_residuals(), ...

Anyone willing to do this? 😄

[IndrajeetPatil]

I think this is a good idea. I will make an entry for this here and will work on it as I find time.

[strengejacke]

Closing in favour of #376