Sensitivity analysis is a critical component of a good modeling workflow. Yet as the number and power of Bayesian computational tools has increased, the options for sensitivity analysis have remained largely the same: compute importance sampling weights manually, or fit a large number of similar models, dramatically increasing computation time. Neither option is satisfactory for most applied modeling.

adjustr is an R package which aims to make sensitivity analysis faster and easier, and works with Bayesian models fitted with Stan. Users provide a series of alternate sampling specifications, and the package uses Pareto-smoothed importance sampling to estimate the posterior under each specification. The package also provides functions to summarize and plot how posterior quantities change across specifications.

The package provides simple interface that makes it as easy as possible for researchers to try out various adjustments to their Stan models, without needing to write any specific Stan code or even recompile or rerun their model.

The package works by parsing Stan model code, so everything works best if the model was written by the user. Models made using brms may in principle be used as well. Models made using rstanarm are constructed using more complex model templates, and cannot be used.

## Getting Started

The basic adjustr workflow is as follows:

1. Use make_spec to specify the set of alternative model specifications you’d like to fit.

2. Use adjust_weights to calculate importance sampling weights which approximate the posterior of each alternative specification.

3. Use summarize and spec_plot to examine posterior quantities of interest for each alternative specification, in order to assess the sensitivity of the underlying model.

To illustrate, the package lets us do the following:

extract_samp_stmts(eightschools_m)
#> Sampling statements for model 2c8d1d8a30137533422c438f23b83428:
#>   parameter   eta ~ std_normal()
#>   data        y ~ normal(theta, sigma)

make_spec(eta ~ student_t(0, 1, df), df=1:10) %>%
summarize(wasserstein(mu))
#> # A tibble: 11 x 5
#>       df .samp                     .weights      .pareto_k wasserstein(mu)
#>    <int> <chr>                     <list>            <dbl>             <dbl>
#>  1     1 eta ~ student_t(df, 0, 1) <dbl [1,000]>     1.02              0.928
#>  2     2 eta ~ student_t(df, 0, 1) <dbl [1,000]>     1.03              0.736
#>  3     3 eta ~ student_t(df, 0, 1) <dbl [1,000]>     0.915             0.534
#>  4     4 eta ~ student_t(df, 0, 1) <dbl [1,000]>     0.856             0.411
#>  5     5 eta ~ student_t(df, 0, 1) <dbl [1,000]>     0.826             0.341
#>  6     6 eta ~ student_t(df, 0, 1) <dbl [1,000]>     0.803             0.275
#>  7     7 eta ~ student_t(df, 0, 1) <dbl [1,000]>     0.782             0.234
#>  8     8 eta ~ student_t(df, 0, 1) <dbl [1,000]>     0.753             0.195
#>  9     9 eta ~ student_t(df, 0, 1) <dbl [1,000]>     0.736             0.166
#> 10    10 eta ~ student_t(df, 0, 1) <dbl [1,000]>     0.721             0.151
#> 11    NA <original model>          <dbl [1,000]>  -Inf                 0

The tutorial vignette walks through a full sensitivity analysis for this 8-schools example. Smaller examples are also included in the package documentation.

## Installation

if (!require("devtools")) {
devtools::install_github("corymccartan/adjustr@*release")