dynamite: Estimate a Bayesian Dynamic Multivariate Panel Model
In santikka/dynamite: Bayesian Modeling and Causal Inference for Multivariate Longitudinal Data

dynamite

R Documentation

Estimate a Bayesian Dynamic Multivariate Panel Model

Description

Fit a Bayesian dynamic multivariate panel model (DMPM) using Stan for Bayesian inference. The dynamite package supports a wide range of distributions and allows the user to flexibly customize the priors for the model parameters. The dynamite model is specified using standard R formula syntax via dynamiteformula(). For more information and examples, see 'Details' and the package vignettes.

The formula method returns the model definition as a quoted expression.

Information on the estimated dynamite model can be obtained via print() including the following: The model formula, the data, the smallest effective sample sizes, largest Rhat and summary statistics of the time-invariant and group-invariant model parameters.

The summary() method provides statistics of the posterior samples of the model; this is an alias of as.data.frame.dynamitefit() with summary = TRUE.

Usage

dynamite(
  dformula,
  data,
  time,
  group = NULL,
  priors = NULL,
  backend = "rstan",
  verbose = TRUE,
  verbose_stan = FALSE,
  stanc_options = list("O0"),
  threads_per_chain = 1L,
  grainsize = NULL,
  custom_stan_model = NULL,
  debug = NULL,
  ...
)

## S3 method for class 'dynamitefit'
formula(x, ...)

## S3 method for class 'dynamitefit'
print(x, full_diagnostics = FALSE, ...)

## S3 method for class 'dynamitefit'
summary(object, ...)

Arguments

`dformula`	[`dynamiteformula`] The model formula. See `dynamiteformula()` and 'Details'.
`data`	[`data.frame`, `tibble::tibble`, or `data.table::data.table`] The data that contains the variables in the model in long format. Supported column types are `integer`, `logical`, `double`, and `factor`. Columns of type `character` will be converted to factors. Unused factor levels will be dropped. The `data` can contain missing values which will simply be ignored in the estimation in a case-wise fashion (per time-point and per channel). Input `data` is converted to channel specific matrix representations via `stats::model.matrix.lm()`.
`time`	[`character(1)`] A column name of `data` that denotes the time index of observations. If this variable is a factor, the integer representation of its levels are used internally for defining the time indexing.
`group`	[`character(1)`] A column name of `data` that denotes the unique groups or `NULL` corresponding to a scenario without any groups. If `group` is `NULL`, a new column `.group` is created with constant value `1L` is created indicating that all observations belong to the same group. In case of name conflicts with `data`, see the `group_var` element of the return object to get the column name of the new variable.
`priors`	[`data.frame`] An optional data frame with prior definitions. See `get_priors()` and 'Details'.
`backend`	[`character(1)`] Defines the backend interface to Stan, should be either `"rstan"` (the default) or `"cmdstanr"`. Note that `cmdstanr` needs to be installed separately as it is not on CRAN. It also needs the actual `CmdStan` software. See https://mc-stan.org/cmdstanr/ for details.
`verbose`	[`logical(1)`] All warnings and messages are suppressed if set to `FALSE`. Defaults to `TRUE`. Setting this to `FALSE` will also disable checks for perfect collinearity in the model matrix.
`verbose_stan`	[`logical(1)`] This is the `verbose` argument for `rstan::sampling()`. Defaults to `FALSE`.
`stanc_options`	[`list()`] This is the `stanc_options` argument passed to the compile method of a `CmdStanModel` object via `cmdstan_model()` when `backend = "cmdstanr"`. Defaults to `list("O0")`. To enable level one compiler optimizations, use `list("O1")`. See https://mc-stan.org/cmdstanr/reference/cmdstan_model.html for details.
`threads_per_chain`	[`integer(1)`] A Positive integer defining the number of parallel threads to use within each chain. Default is `1`. See `rstan::rstan_options()` and https://mc-stan.org/cmdstanr/reference/model-method-sample.html for details.
`grainsize`	[`integer(1)`] A positive integer defining the suggested size of the partial sums when using within-chain parallelization. Default is number of time points divided by `threads_per_chain`. Setting this to `1` leads the workload division entirely to the internal scheduler. The performance of the within-chain parallelization can be sensitive to the choice of `grainsize`, see Stan manual on reduce-sum for details.
`custom_stan_model`	[`character(1)`] An optional character string that either contains a customized Stan model code or a path to a `.stan` file that contains the code. Using this will override the generated model code. For expert users only.
`debug`	[`list()`] A named list of form `name = TRUE` indicating additional objects in the environment of the `dynamite` function which are added to the return object. Additionally, values `no_compile = TRUE` and `no_sampling = TRUE` can be used to skip the compilation of the Stan code and sampling steps respectively. This can be useful for debugging when combined with `model_code = TRUE`, which adds the Stan model code to the return object.
`...`	For `dynamite()`, additional arguments to `rstan::sampling()` or the `⁠$sample()⁠` method of the `CmdStanModel` object (see https://mc-stan.org/cmdstanr/reference/model-method-sample.html), such as `chains` and `cores` (`chains` and `parallel_chains` in `cmdstanr`). For `summary()`, additional arguments to `as.data.frame.dynamitefit()`. For `print()`, further arguments to the print method for tibbles (see tibble::formatting). Not used for `formula()`.
`x`	[`dynamitefit`] The model fit object.
`full_diagnostics`	By default, the effective sample size (ESS) and Rhat are computed only for the time- and group-invariant parameters (`full_diagnostics = FALSE`). Setting this to `TRUE` computes ESS and Rhat values for all model parameters, which can take some time for complex models.
`object`	[`dynamitefit`] The model fit object.

Details

The best-case scalability of dynamite in terms of data size should be approximately linear in terms of number of time points and and number of groups, but as wall-clock time of the MCMC algorithms provided by Stan can depend on the discrepancy of the data and the model (and the subsequent shape of the posterior), this can vary greatly.

Value

dynamite returns a dynamitefit object which is a list containing the following components:

stanfit
A stanfit object, see rstan::sampling() for details.
dformulas
A list of dynamiteformula objects for internal use.
data
A processed version of the input data.
data_name
Name of the input data object.
stan
A list containing various elements related to Stan model construction and sampling.
group_var
Name of the variable defining the groups.
time_var
Name of the variable defining the time index.
priors
Data frame containing the used priors.
backend
Either "rstan" or "cmdstanr" indicating which package was used in sampling.
permutation
Randomized permutation of the posterior draws.
call
Original function call as an object of class call.

formula returns a quoted expression.

print returns x invisibly.

summary returns a data.frame.

References

Santtu Tikka and Jouni Helske (2024). dynamite: An R Package for Dynamic Multivariate Panel Models. arXiv preprint, doi:10.48550/arXiv.2302.01607.

Jouni Helske and Santtu Tikka (2022). Estimating Causal Effects from Panel Data with Dynamic Multivariate Panel Models. Advances in Life Course Research, 60, 100617. doi:10.1016/j.alcr.2024.100617.

Examples

data.table::setDTthreads(1) # For CRAN

# Please update your rstan and StanHeaders installation before running
# on Windows
if (!identical(.Platform$OS.type, "windows")) {
  fit <- dynamite(
    dformula = obs(y ~ -1 + varying(~x), family = "gaussian") +
      lags(type = "varying") +
      splines(df = 20),
    gaussian_example,
    "time",
    "id",
    chains = 1,
    refresh = 0
  )
}


data.table::setDTthreads(1) # For CRAN
formula(gaussian_example_fit)

data.table::setDTthreads(1) # For CRAN
print(gaussian_example_fit)

data.table::setDTthreads(1) # For CRAN
summary(gaussian_example_fit,
  types = "beta",
  probs = c(0.05, 0.1, 0.9, 0.95)
)

santikka/dynamite documentation built on April 17, 2025, 11:47 a.m.