Alternative Parameterizations

[vincenzocoia]

One issue that distplyr hasn't considered yet is that of alternative parameterizations of a distribution.

For example, dst_norm() accepts mean and variance as the parameters, but other parameterizations would be useful, such as specifying, say, the 0.05- and 0.95-quantiles (use case: "I'm pretty sure this expense will be between $100 and $200", just set these as lower and upper quantiles). Even mean and stdev would be useful.

I have two ideas for achieving this, but they shouldn't be implemented until after the package is submitted to CRAN.

1. A parametric distribution is still only specified using its "canonical" parameters (mean and variance for Normal; alpha and beta for Beta; etc.), but changed downstream.

Something like:

dst_norm() %>%
    set_quantiles(0.05 ~ 100, 0.95 ~ 200)

This would operate rlang's tidy evaluation machinery, leaving unevaluated parameters as unknown variables (quosures, probably), until they are evaluated downstream. Bonus: this would also allow for scenarios like dst_unif(a, a + 1), where a is unknown.

2. Specify parameters by name in the dst_ call.

I'm thinking along the lines of ggplot2's aes() function. The aes() function "parameterizes" a plot according to some aesthetics (see ?geom_linerange for an example of being able to parameterize in more than one way). Except having a special function like aes() might not be useful here, because the parameters can be specified right in the dst_*() call.

So, instead of:

dst_norm <- function(mean, variance) { ... }

we would have something like:

dst_norm <- function(mean, variance, ...)

akin to aes <- function(x, y, ...).

Calls would look something like:

dst_norm(0, 1)  # For mean and variance
dst_norm(0, stdev = 5) # For mean and standard deviation
dst_norm(0.05 ~ 100, 0.95 ~ 200) # For the 0.05- and 0.95-quantiles.