---
title: "The posterior R package"
author: "Paul Bürkner, Jonah Gabry, Matthew Kay, and Aki Vehtari"
output:
rmarkdown::html_vignette:
toc: true
toc_depth: 3
vignette: >
%\VignetteIndexEntry{The posterior R package}
%\VignetteEngine{knitr::rmarkdown}
%\VignetteEncoding{UTF-8}
---
## Introduction
The posterior R package is intended to provide useful tools for both users
and developers of packages for fitting Bayesian models or working with output
from Bayesian models. The primary goals of the package are to:
* Efficiently convert between many different useful formats of
draws (samples) from posterior or prior distributions.
* Provide consistent methods for operations commonly performed on draws,
for example, subsetting, binding, or mutating draws.
* Provide various summaries of draws in convenient formats.
* Provide lightweight implementations of state of the art posterior inference
diagnostics.
## Installation
You can install the latest official release version via
```{r install, eval=FALSE}
install.packages("posterior")
```
or the latest development version from GitHub via
```{r install_github, eval=FALSE}
# install.packages("remotes")
remotes::install_github("stan-dev/posterior")
```
## Example
```{r setup}
library("posterior")
```
To demonstrate how to work with the posterior package, throughout the rest of
this vignette we will use example posterior draws obtained from the eight
schools hierarchical meta-analysis model described in Gelman et al. (2013). The
variables are an estimate per school (`theta[1]` through `theta[8]`) as well as
an overall mean (`mu`) and standard deviation across schools (`tau`).
```{r example-drawss}
eight_schools_array <- example_draws("eight_schools")
print(eight_schools_array, max_variables = 3)
```
The structure of this object is explained in the next section.
## Draws formats
### Available formats
Because different formats are preferable in different situations, posterior
supports multiple formats and easy conversion between them. The currently
supported formats are:
* `draws_array`: An iterations by chains by variables array.
* `draws_matrix`: A draws (iterations x chains) by variables array.
* `draws_df`: A draws by variables data frame with addition meta columns
`.chain`, `.iteration`, `.draw`.
* `draws_list`: A list with one sublist per chain. Each sublist is a named list
with one vector of iterations per variable.
* `draws_rvars`: A list of random variable `rvar` objects, one per variable. See
`vignette("rvar")` for an introduction to this new data type.
These formats are essentially base R object classes and can be used as such. For
example, a `draws_matrix` object is just a `matrix` with a little more
consistency (e.g., no dropping of dimensions with one level when indexing) and
additional methods. The exception to this is the `draws_rvars` format, which
contains `rvar` objects that behave somewhat like arrays but are really a unique
data type. See the separate vignette on the `rvar` and `draws_rvars` data types
for details.
The draws for our example come as a `draws_array` object with
`r niterations(eight_schools_array)` iterations,
`r nchains(eight_schools_array)` chains, and
`r nvariables(eight_schools_array)` variables:
```{r draws_array-structure}
str(eight_schools_array)
```
### Converting between formats
Each of the formats has a method `as_draws_<format>` (e.g., `as_draws_list()`)
for creating an object of the class from any of the other formats. As a
demonstration we can convert the example `draws_array` to a `draws_df`,
a data frame with additional meta information. To convert to a `draws_df` we
use `as_draws_df()`.
```{r draws_df}
eight_schools_df <- as_draws_df(eight_schools_array)
str(eight_schools_df)
print(eight_schools_df)
```
### Converting regular R objects to `draws` formats
The example draws already come in a format natively supported by posterior, but
we can of course also import the draws from other sources like common base R
objects.
#### Example: create draws_matrix from a matrix
In addition to converting other `draws` objects to the `draws_matrix` format,
the `as_draws_matrix()` function will convert a regular matrix to a
`draws_matrix`.
```{r draws_matrix-from-matrix}
x <- matrix(rnorm(50), nrow = 10, ncol = 5)
colnames(x) <- paste0("V", 1:5)
x <- as_draws_matrix(x)
print(x)
```
Because the matrix was converted to a `draws_matrix`, all of the methods for
working with `draws` objects described in subsequent sections of this vignette
will now be available.
Instead of `as_draws_matrix()` we also could have just used `as_draws()`, which
attempts to find the closest available format to the input object. In this case
the result would be a `draws_matrix` object either way.
#### Example: create draws_matrix from multiple vectors
In addition to the `as_draws_matrix()` converter function there is also a
`draws_matrix()` constructor function that can be used to create draws matrix
from multiple vectors.
```{r draws_matrix-from-vectors}
x <- draws_matrix(alpha = rnorm(50), beta = rnorm(50))
print(x)
```
Analogous functions exist for the other draws formats and are used similarly.
## Manipulating `draws` objects
The posterior package provides many methods for manipulating draws objects in
useful ways. In this section we demonstrate several of the most commonly used
methods. These methods, like the other methods in posterior, are available for
every supported draws format.
### Subsetting
Subsetting `draws` objects can be done according to various aspects of the draws
(iterations, chains, or variables). The posterior package provides a convenient
interface for this purpose via `subset_draws()`. For example, here is the
code to extract the first five iterations of the first two chains of the
variable `mu`.
```{r subset-df}
sub_df <- subset_draws(eight_schools_df, variable = "mu", chain = 1:2, iteration = 1:5)
str(sub_df)
```
The same call to `subset_draws()` can be used regardless of the draws format.
For example, here is the same code except replacing the `draws_df` object with
the `draws_array` object.
```{r subset-array}
sub_arr <- subset_draws(eight_schools_array, variable = "mu", chain = 1:2, iteration = 1:5)
str(sub_arr)
```
We can check that these two calls to `subset_draws()` (the first with the data
frame, the second with the array) produce the same result.
```{r subset-compare, results='hold'}
identical(sub_df, as_draws_df(sub_arr))
identical(as_draws_array(sub_df), sub_arr)
```
It is also possible to use standard R subsetting syntax with `draws` objects.
The following is equivalent to the use of `subset_draws()` with the array above.
```{r subset-standard}
eight_schools_array[1:5, 1:2, "mu"]
```
The major difference between how posterior behaves when indexing and how base R
behaves is that posterior will _not_ drop dimensions with only one level. That
is, even though there is only one variable left after subsetting, the result of
the subsetting above is still a `draws_array` and not a `draws_matrix`.
### Mutating (transformations of variables)
The magic of having obtained draws from the joint posterior (or prior)
distribution of a set of variables is that these draws can also be used to
obtain draws from any other variable that is a function of the original
variables.
That is, if we are interested in the posterior distribution of, say,
`phi = (mu + tau)^2` all we have to do is to perform the transformation for each
of the individual draws to obtain draws from the posterior distribution of the
transformed variable. This procedure is handled by `mutate_variables()`.
```{r mutate}
x <- mutate_variables(eight_schools_df, phi = (mu + tau)^2)
x <- subset_draws(x, c("mu", "tau", "phi"))
print(x)
```
### Renaming
To rename variables use `rename_variables()`. Here we rename the scalar `mu` to
`mean` and the vector `theta` to `alpha`.
```{r rename}
# mu is a scalar, theta is a vector
x <- rename_variables(eight_schools_df, mean = mu, alpha = theta)
variables(x)
```
In the call to `rename_variables()` above, `mu` and `theta` can be quoted or
unquoted.
It is also possible to rename individual elements of non-scalar parameters,
for example we can rename just the first element of `alpha`:
```{r rename-element}
x <- rename_variables(x, a1 = `alpha[1]`)
variables(x)
```
### Binding
The `bind_draws()` method can be used to combine `draws` objects along different
dimensions. As an example, suppose we have several different `draws_matrix`
objects:
```{r objects-to-bind}
x1 <- draws_matrix(alpha = rnorm(5), beta = rnorm(5))
x2 <- draws_matrix(alpha = rnorm(5), beta = rnorm(5))
x3 <- draws_matrix(theta = rexp(5))
```
We can bind `x1` and `x3` together along the `'variable'` dimension to get
a single `draws_matrix` with the variables from both `x1` and `x3`:
```{r bind-variable}
x4 <- bind_draws(x1, x3, along = "variable")
print(x4)
```
Because `x1` and `x2` have the same variables, we can bind them along the
`'draw'` dimension to create a single `draws_matrix` with more draws:
```{r bind-draw}
x5 <- bind_draws(x1, x2, along = "draw")
print(x5)
```
As with all posterior methods, `bind_draws()` can be used with all draws formats
and depending on the format different dimensions are available to bind on. For
example, we can bind `draws_array` objects together by `iteration`, `chain`, or
`variable`, but a 2-D `draws_matrix` with the chains combined can only by bound
by `draw` and `variable`.
## Summaries and diagnostics
### summarise_draws() basic usage
Computing summaries of posterior or prior draws and convergence diagnostics for
posterior draws are some of the most common tasks when working with Bayesian
models fit using Markov Chain Monte Carlo (MCMC) methods. The posterior
package provides a flexible interface for this purpose via `summarise_draws()`
(or `summarize_draws()`), which can be passed any of the formats supported by
the package.
```{r summary}
# summarise_draws or summarize_draws
summarise_draws(eight_schools_df)
```
The result is a data frame with one row per variable and one column per
summary statistic or convergence diagnostic. The summaries `rhat`, `ess_bulk`,
and `ess_tail` are described in Vehtari et al. (2020). We can choose which
summaries to compute by passing additional arguments, either functions or names
of functions. For instance, if we only wanted the mean and its corresponding
Monte Carlo Standard Error (MCSE) we could use either of these options:
```{r summary-with-measures}
# the function mcse_mean is provided by the posterior package
s1 <- summarise_draws(eight_schools_df, "mean", "mcse_mean")
s2 <- summarise_draws(eight_schools_df, mean, mcse_mean)
identical(s1, s2)
print(s1)
```
### Changing column names
The column names in the output can be changed by providing the functions
as name-value pairs, where the name is the name to use in the output and the
value is a function name or definition. For example, here we change the names
`mean` and `sd` to `posterior_mean` and `posterior_sd`.
```{r change-summary-names}
summarise_draws(eight_schools_df, posterior_mean = mean, posterior_sd = sd)
```
### Using custom functions
For a function to work with `summarise_draws()`, it needs to take a vector or
matrix of numeric values and return a single numeric value or a named vector of
numeric values. Additional arguments to the function can be specified in a list
passed to the `.args` argument.
```{r summary-.args}
weighted_mean <- function(x, wts) {
sum(x * wts)/sum(wts)
}
summarise_draws(
eight_schools_df,
weighted_mean,
.args = list(wts = rexp(ndraws(eight_schools_df)))
)
```
### Specifying functions using lambda-like syntax
It is also possible to specify a summary function using a one-sided formula that
follows the conventions supported by `rlang::as_function()`. For example, the
function
```{r standard-quantile, eval = FALSE}
function(x) quantile(x, probs = c(0.4, 0.6))
```
can be simplified to
```{r lambda-quantile, eval = FALSE}
# for multiple arguments `.x` and `.y` can be used, see ?rlang::as_function
~quantile(., probs = c(0.4, 0.6))
```
Both can be used with `summarise_draws()` and produce the same output:
```{r lambda-syntax}
summarise_draws(eight_schools_df, function(x) quantile(x, probs = c(0.4, 0.6)))
summarise_draws(eight_schools_df, ~quantile(.x, probs = c(0.4, 0.6)))
```
See `help("as_function", "rlang")` for details on specifying these functions.
### Other diagnostics
In addition to the default diagnostic functions used by `summarise_draws()`
(`rhat()`, `ess_bulk()`, `ess_tail()`), posterior also provides additional
diagnostics like effective sample sizes and Monte Carlo standard errors for
quantiles and standard deviations, an experimental new diagnostic called R*, and
others. For a list of available diagnostics and links to their individual help
pages see `help("diagnostics", "posterior")`.
If you have suggestions for additional diagnostics that should be implemented in
posterior, please open an issue at
<https://github.com/stan-dev/posterior/issues>.
## Other methods for working with `draws` objects
In addition to the methods demonstrated in this vignette, posterior has various
other methods available for working with `draws` objects. The following is a
(potentially incomplete) list.
|**Method**|**Description**|
|:----------|:---------------|
| `order_draws()` | Order `draws` objects according to iteration and chain number |
| `repair_draws()`| Repair indices of `draws` objects so that iterations chains, and draws are continuously and consistently numbered |
|`resample_draws()` | Resample `draws` objects according to provided weights |
| `thin_draws()` | Thin `draws` objects to reduce size and autocorrelation |
| `weight_draws()`| Add weights to draws objects, with one weight per draw, for use in subsequent weighting operations |
| `extract_variable()` | Extract a vector of draws of a single variable |
| `extract_variable_matrix()` | Extract an iterations x chains matrix of draws of a single variable |
| `merge_chains()` | Merge chains of `draws` objects into a single chain. |
| `split_chains()` | Split chains of `draws` objects by halving the number of iterations per chain and doubling the number of chains. |
If you have suggestions for additional methods that would be useful for working
with `draws` objects, please open an issue at
<https://github.com/stan-dev/posterior/issues>.
## References
Gelman A., Carlin J. B., Stern H. S., David B. Dunson D. B., Vehtari, A.,
& Rubin D. B. (2013). *Bayesian Data Analysis, Third Edition*. Chapman and
Hall/CRC.
Vehtari A., Gelman A., Simpson D., Carpenter B., & Bürkner P. C. (2020).
Rank-normalization, folding, and localization: An improved Rhat for assessing
convergence of MCMC. *Bayesian Analysis*, 16(2):667-718.