Wider applications of {epiparameter} extraction

[adamkucharski]

I'm finding the functions in {epiparameter} very helpful for defining plausible distributions based on summary statistics in the literature, e.g. for a prior in H5N1 cluster analysis or @sarahdicksonuoft's work that required distribution of measles vaccine coverage.

Could be nice to have a post outlining other use cases if anyone has specific examples of this having been done manually in the past (e.g. papers converting literature estimate with 95% CrI into a parameteric prior).

[TimTaylor]

That looks good! Very out of the loop with things at the moment so maybe this has already been discussed but I wonder if densities/distributions/quantiles/random generation functions could be an output from epiparameter?
E.g.

# Get prior for R based on Aditama et al, PLOS ONE, 2012
get_prior <- extract_param(
  type = "percentiles",
  values = c(0.009, 0.315),
  distribution = "gamma",
  percentiles = c(0.025,0.975)
)

h5_prior_r <- function(x){dgamma(x,shape = get_prior[["shape"]], scale = get_prior[["scale"]])}

could become

h5_prior_r <- get_prior[["dgamma"]]

I'm not sure though. Maybe this would be too opaque?

[joshwlambert]

I'm not sure. My immediate thought is it is overcomplicating the output (which is currently a numeric vector) and is somewhat duplicating the functionality of <epiparameter> by having distribution methods.

We currently depend on {distributional} so we can utilise their classes and methods for this, so it wouldn't add dependencies unless there was clear reasoning to go with {distributions3}.

I'm happy to be convinced of making the change. One possible benefit with some upcoming changes in {epiparameter} would be to do something like this:

dist <- extract_param(...)
epiparameter(
  disease = "<disease>",
  epi_dist = "<epi_dist>",
  prob_distribution = dist
)

[TimTaylor]

Yeah - think you're right. I need to be more familiar with what the package. I misunderstood the example.

[TimTaylor]

@joshwlambert - I'm still pondering. It feels like extract_param() is almost an extract_distribution() type thing so you could have the case you have above

dist <- extract_param(...)
epiparameter(
  disease = "<disease>",
  epi_dist = "<epi_dist>",
  prob_distribution = dist
)

and also

dist <- extract_param(...)
density(dist, ...)
cdf(dist, ...)

etc, etc

No preference as to either package (just liked the initial feel when I saw distributions3). Was unaware of distributional. Looking at the available options it's only a matter of time until some one makes a distributions7 (with s7) 😅

[joshwlambert]

Thanks for the input @TimTaylor, I'm still unsure. Please could you open this as an issue in {epiparameter} and we can discuss the details there.

For me the decider would be if there is a use case where extract_param() would need to be used directly used as a distribution.

Some other options:

• add an extract_distribution() function do what you're suggesting
• add an argument to extract_param() called something like as_distribution or vector which could take a logical and then the output could be changed accordingly.

[TimTaylor]

• add an extract_distribution() function do what you're suggesting

This would probably be the best approach if desired/needed.

[joshwlambert]

@adamkucharski thanks for posting, I like the idea of having a post showcasing examples of where the conversion and extraction functions can help users for specific tasks. We'd need a list of examples and it might be a fair amount of work to search the literature for cases. I'd be happy to collaborate on the post.

[adamkucharski]

Good point – potentially better to keep this post as a placeholder for now, so can add examples as they crop up (and solicit from others), rather than focusing on getting out ASAP.

[adamkucharski]

In case useful, also highlighting a similar tool in python: https://preliz.readthedocs.io/en/latest/examples/param_space_1d_examples.html#from-intervals-to-maximum-entropy-distributions