Decide on standard format for PSA results

[evaenns]

How should we require PSA results be formatted for input into dampack functions? We need to balance ease of use for the user (e.g. how will they typically have a PSA stored), flexibility, and clarity in structure of the functions. The format needs to be able to handle:

• A large number of parameters being varied
• An arbitrary number of outcomes of interest (not just costs and QALYs)
• Arbitrary number of strategies

[feralaes]

I suggest that we require the PSA outputs to be formatted as two different types of objects:

• Matrices of outcomes. For example, for costs, m.c and for effectiveness, m.e, of dimensions n x k, where n is the number of PSA simulations and k the number of strategies. If there are more (i.e., other) outcomes, we can ask the user to specify them as m.o.
• A matrix of parameters of dimensions n x p, where p is the number of parameters

Thoughts?

[caleb-easterly]

I think that sounds good! I actually just changed the ceac() function to take 2 matrices: costs and effectiveness, so that would work with your proposal, Fernando.

[caleb-easterly]

Do you think we should have a psa class? The function to create it could enforce some of the conditions (parameters, cost, effectiveness, etc., should all have the same number of rows, for example).

[feralaes]

Excellent!

Yeah, I think that could work. Would you generate a psa object from the cost, effectiveness and input parameter matrices? Or would you ask the user to create a psa matrix ex-ante and then the function will just make the matrix a psa class?

[caleb-easterly]

It could be something like make_psa_obj <- function(list_of_outcomes, cost, param), and then it could return a psa object. I don't know yet how many functions make use of the PSA results, but it could standardize it across the functions.

[feralaes]

It could be something like make_psa_obj <- function(list_of_outcomes, cost, param), and then it could return a psa object. I don't know yet how many functions make use of the PSA results, but it could standardize it across the functions.

Makes sense, although I can see a trade off here by adding a new function but also, one the psa object is created, then it's easier to use later on. Right of the bat, can think of at least 5 core functions that use psa: CEAC, EVPI, Expected loss curves (ELC), EVPPI, EVSI and also other ones like OWSA, TWSA, and TORNADO using regression metamodeling.

[evaenns]

I think a PSA object makes sense, given that it is the backbone of so many of these other functions. I think the key is to have the make_psa_obj() function have really clear error messages and that the PSA object have some nice summary methods so it is easy for a user to view what they have created.

[caleb-easterly]

see PSA object in #15