Always call `as.matrix` or `as.array` on the first argument

[bgoodri]

If I have a stanreg object called post, then

bayesplot::mcmc_acf(post)

should just work, but it doesn't because the first argument is neither a matrix nor an array. But when no specific method is defined for stanreg objects, the bayesplot functions should just call as.matrix or as.array on the first argument internally, like apply and things like that do. That way we can accept anything that can be coerced into a matrix or an array without having to write methods for everything.

[jgabry]

Coming back to this now that I have some time to work on the next bayesplot release. @tjmahr, this seems like a good idea to me. Any reason you can think of not to do this?

[tjmahr]

So if x is not an array or matrix, we try calling as.array() on it when we run prepare_mcmc_array()? And something similar for when we need matrices? Can we some additional check that tells us whether the result of as.array() is appropriate for later work?

Like, both of these return something after as.array(), but only one of them will has the right structure for plotting.

fit <- rstanarm::stan_glm(mpg ~ ., data = mtcars, iter = 50)
fit2 <- lm(mpg ~ ., data = mtcars)

[jgabry]

Good point. Maybe we can just check the result of calling dim on the object returned by as.array(...). We would need the resulting object to have either 2 or 3 dimensions (2 if all chains combined, 3 if separate). In the example above dim(as.array(fit)) returns the vector c(25, 4, 12), which is fine, but dim(as.array(fit2)) returns just 12, so it's just a 1-D array and we could reject it.

[tjmahr]

We would need the resulting object to have either 2 or 3 dimensions (2 if all chains combined, 3 if separate).

I like that.

[bgoodri]

If some package has a specialized method, then it should definitely dispatch to that. The default method should just allow the function to work without a specialized method whenever that makes sense. In cases where it can't make sense, it doesn't work currently so this is just an opportunity to throw a better error message.

…

On Thu, Apr 11, 2019 at 2:57 PM TJ Mahr ***@***.***> wrote: We would need the resulting object to have either 2 or 3 dimensions (2 if all chains combined, 3 if separate). I like that. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#175 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ADOrqvCiLSlhybzlnVs_HT7LJKtZuqcBks5vf4WFgaJpZM4bRCA3> .

[tjmahr]

Right.

• as.array() on a valid array... safe
• as.array() on an object with an as.array() method... safe.
• as.array() on anything else... as.array.default() adds a dim attribute. we reject if it doesn't have valid dimensions.