Skip to content

jhelvy/stanTuneR

Repository files navigation

stanTuneR

This code uses the algebra solver in Stan to find the parameters of a distribution that produce a desired tail behavior. This can be a useful tool when choosing parameters for prior distributions. Here’s how to use it:

  1. Choose a distribution.
  2. Define the quantile boundaries and the amount of probability density that you wish to have above and below those boundaries.
  3. Let Stan go find the parameters that produce the desired distribution.

Currently supported distributions:

  • Normal
  • Log-Normal
  • Beta
  • Gamma
  • Inverse Gamma

Required libraries:

Obviously you need to have Stan installed on your machine.

If you don’t want to use the shiny app interface, you only need to have the rstan library installed:

install.packages('rstan')

To use the shiny app, you will also need to install the shiny and shinycssloaders libraries:

install.packages('shiny')
install.packages('shinycssloaders')

Using the Shiny app

To use the Shiny app, just run the following code in R:

library(shiny)
runGitHub('jhelvy/stanTuneR')

The interface should look like this:

screenshot

Example without the Shiny app

(see the examples.R file for more examples)

Let’s say I want to find the parameters of a normal distribution such that P[x < -2.0] ~ 0.01 and P[x > 2.0] ~ 0.01. That is, I want a normal distribution where 98% of the probability density is between (-2, 2).

First, load the rstan library, tweak some settings, and source the functions.R file (all functions are loaded in a new environment called funcs):

# Load libraries
library(shiny)
library(rstan)

# Stan settings
rstan_options(auto_write = TRUE)
options(mc.cores = parallel::detectCores())

# Load the functions
funcs <- new.env()
source('functions.R', local=funcs)

Then use the targets argument to set the desired tail properties:

targets = list(
    bound_L = -2,   # LOWER quantile boundary
    bound_U = 2,    # UPPER quantile boundary
    dens_L  = 0.01, # Target density below LOWER quantile boundary
    dens_U  = 0.01) # Target density above UPPER quantile boundary

Then use the tuneParams function to find the parameters:

results = funcs$tuneParams(distribution='normal', targets)
## 
## SAMPLING FOR MODEL 'model' NOW (CHAIN 1).
## Chain 1: Iteration: 1 / 1 [100%]  (Sampling)
## Chain 1: 
## Chain 1:  Elapsed Time: 0 seconds (Warm-up)
## Chain 1:                0.001128 seconds (Sampling)
## Chain 1:                0.001128 seconds (Total)
## Chain 1:

View the resulting parameters and verify that the quantiles of 10,000 draws from the resulting distribution match your criteria:

results$params
## $mu
## [1] 0
## 
## $sigma
## [1] 0.859717
results$quantiles
##        1%       99% 
## -2.027173  2.066950

Finally, view a histogram of the resulting distribution:

results$histogram

Explanation of the backend

The meat of this app is in the functions.R code. The main function is the tuneParams() function. After the user defines the distribution and the targets for the quantiles and density, tuneParams() calls the generateStanCode() function to generate the Stan code for the model, which is written to the file model.stan. This file is always overwritten every time the generateStanCode() function is called. It then calls the stan function to fit the model. Finally, once the model is fit, it calls the summarizeResults() function to extract the results of the fit model. The output of tuneParams() is a list with the following values:

  • params : The fit parameters
  • draws : 10,000 draws from the resulting distribution using the fit parameters
  • quantiles : The quantiles of the draws at the desired upper and lower quantile boundaries
  • histogram : A histogram of the draws

Author and License

  • Author: John Paul Helveston (www.jhelvy.com)
  • Date First Written: Tuesday, April 30, 2019
  • License: GPL-3

About

This code uses the algebra solver in Stan (https://mc-stan.org/) to find the parameters of a distribution that produce a desired tail behavior.

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published