README.md

In lionel-/gsim: Gorgeous manipulations of (posterior) simulations

gsim

gsim is a DSL for statistical simulations. It seeks to:

• Provide tools to setup your data for Stan or Jags.

• Automate loops over simulations in an efficient way to let you focus on the mathematics of your computations rather than the programming technicalities.

• Help you create numerical and graphical summaries of parameters and fitted models.

gsim is designed to work with the magrittr pipe %>% to create clean and readable pipelines of objects transformations. To install the package:

install.packages("devtools")
devtools::install_github("lionel-/gsim")

Setup data

• define() is like dplyr's mutate() method but it works on lists and discards any variables that are not mentioned. This last property is useful with samplers such as Jags that will complain when they find unused variables in the input.

• design() creates design matrices. The .unjoin argument allows to reduce repeated group-level data to one value per group.

# Fetch Radon dataset from Stan's repo
radon_data <- radon_data()

# Create the list of data, including design matrices
radon_data <- radon_data %>%
  define(
    y, county,
    X = design(1, x),
    U = design(1, u, .unjoin = county),
    N = nrow(.),
    J = n_distinct(county),
    P_X = ncol(X),
    P_U = ncol(U)
  )

# Get example model code
radon_model <- radon_model(language = "stan", variant = "centered")

# Fit the model
stanfit <- rstan::stan(model_code = radon_model, data = radon_data)

dplyr-like verbs

You can work directly from a Stan or Jags object with gsim functions. They will be converted to an object of class sims which is simply a list of simulations arrays with some additional information.

• select() parameters from the list. You can select directly by name or use one of the matchers from dplyr. Here we select sigma and all parameters whose name contains Beta.

• mutate() applies a transformation datawise.

• summarise() applies a transformation simulationwise. In other words, it creates posterior summaries.

Note that unlike the method in dplyr, summarise() does not discard the other variables: sims lists can contain both sims_array and normal objects, with the only constraint to keep all the sims_array objects consistent (i.e., same number of simulations).

sims <- stanfit %>%
  select(sigma, contains("Beta")) %>%
  mutate(Beta_res = Beta - cols_mean(Beta)) %>%
  summarise(
    post_sd = sd(Beta_res),
    post_mean = mean(Beta_res)
  )

sims
#> 100 simulations
#>
#>       sigma: [1:1]
#>        Beta: [1:85, 1:2]
#>  sigma_Beta: [1:2]
#>  Sigma_Beta: [1:2, 1:2]
#>    Beta_res: [1:85, 1:2]
#>
#> Other objects: post_sd, post_mean

lionel-/gsim documentation built on May 21, 2019, 6:41 a.m.