search

chjackson / voi

Methods to calculate the Expected Value of Information
https://chjackson.github.io/voi/
5 stars 5 forks source link

question: Which methods are suitable for noisy outcomes? #5

Open BlackEdder opened 3 weeks ago

BlackEdder commented 3 weeks ago

We are using an Individual Based Model (IBM) to model transmission of AMR. As such we end up with multiple net monetary benefit values per (transmission) parameter set. As far as I understand this is no problem when calculating EVP(P)I, because it is based around the expected value, but supposedly some methods will fail? For example, Gaussian Processes need to be adjusted to be able to take this into account (e.g. https://cran.r-project.org/web/packages/GPM/GPM.pdf)

Do you happen to know which methods are suitable for such model outcomes and which would not be?

chjackson commented 3 weeks ago

Any of the methods can be used if you can estimate the expected net benefit (over a population of individuals) given specific parameter values. For individual-based models, the obvious way is to take an average of the individual-level outputs for each set of parameter values.

Though I don't know what the most computationally efficient approach would be (e.g. how many parameter samples vs how many individual samples), or if there are more efficient VoI estimators designed for microsimulation.